Merge branch 'md-raid6-accel' into ioat3.2
Dan Williams [Wed, 9 Sep 2009 00:42:29 +0000 (17:42 -0700)]
Conflicts:
include/linux/dmaengine.h

23 files changed:
Documentation/crypto/async-tx-api.txt
arch/arm/include/asm/hardware/iop3xx-adma.h
arch/arm/mach-iop13xx/include/mach/adma.h
arch/arm/mach-iop13xx/setup.c
arch/arm/plat-iop/adma.c
crypto/async_tx/Kconfig
crypto/async_tx/Makefile
crypto/async_tx/async_memcpy.c
crypto/async_tx/async_memset.c
crypto/async_tx/async_pq.c [new file with mode: 0644]
crypto/async_tx/async_raid6_recov.c [new file with mode: 0644]
crypto/async_tx/async_tx.c
crypto/async_tx/async_xor.c
crypto/async_tx/raid6test.c [new file with mode: 0644]
drivers/dma/Kconfig
drivers/dma/dmaengine.c
drivers/dma/dmatest.c
drivers/dma/iop-adma.c
drivers/md/Kconfig
drivers/md/raid5.c
drivers/md/raid5.h
include/linux/async_tx.h
include/linux/dmaengine.h

index 9f59fcb..ba046b8 100644 (file)
@@ -54,20 +54,23 @@ features surfaced as a result:
 
 3.1 General format of the API:
 struct dma_async_tx_descriptor *
-async_<operation>(<op specific parameters>,
-                 enum async_tx_flags flags,
-                 struct dma_async_tx_descriptor *dependency,
-                 dma_async_tx_callback callback_routine,
-                 void *callback_parameter);
+async_<operation>(<op specific parameters>, struct async_submit ctl *submit)
 
 3.2 Supported operations:
-memcpy       - memory copy between a source and a destination buffer
-memset       - fill a destination buffer with a byte value
-xor          - xor a series of source buffers and write the result to a
-              destination buffer
-xor_zero_sum - xor a series of source buffers and set a flag if the
-              result is zero.  The implementation attempts to prevent
-              writes to memory
+memcpy  - memory copy between a source and a destination buffer
+memset  - fill a destination buffer with a byte value
+xor     - xor a series of source buffers and write the result to a
+         destination buffer
+xor_val - xor a series of source buffers and set a flag if the
+         result is zero.  The implementation attempts to prevent
+         writes to memory
+pq     - generate the p+q (raid6 syndrome) from a series of source buffers
+pq_val  - validate that a p and or q buffer are in sync with a given series of
+         sources
+datap  - (raid6_datap_recov) recover a raid6 data block and the p block
+         from the given sources
+2data  - (raid6_2data_recov) recover 2 raid6 data blocks from the given
+         sources
 
 3.3 Descriptor management:
 The return value is non-NULL and points to a 'descriptor' when the operation
@@ -80,8 +83,8 @@ acknowledged by the application before the offload engine driver is allowed to
 recycle (or free) the descriptor.  A descriptor can be acked by one of the
 following methods:
 1/ setting the ASYNC_TX_ACK flag if no child operations are to be submitted
-2/ setting the ASYNC_TX_DEP_ACK flag to acknowledge the parent
-   descriptor of a new operation.
+2/ submitting an unacknowledged descriptor as a dependency to another
+   async_tx call will implicitly set the acknowledged state.
 3/ calling async_tx_ack() on the descriptor.
 
 3.4 When does the operation execute?
@@ -119,30 +122,42 @@ of an operation.
 Perform a xor->copy->xor operation where each operation depends on the
 result from the previous operation:
 
-void complete_xor_copy_xor(void *param)
+void callback(void *param)
 {
-       printk("complete\n");
+       struct completion *cmp = param;
+
+       complete(cmp);
 }
 
-int run_xor_copy_xor(struct page **xor_srcs,
-                    int xor_src_cnt,
-                    struct page *xor_dest,
-                    size_t xor_len,
-                    struct page *copy_src,
-                    struct page *copy_dest,
-                    size_t copy_len)
+void run_xor_copy_xor(struct page **xor_srcs,
+                     int xor_src_cnt,
+                     struct page *xor_dest,
+                     size_t xor_len,
+                     struct page *copy_src,
+                     struct page *copy_dest,
+                     size_t copy_len)
 {
        struct dma_async_tx_descriptor *tx;
+       addr_conv_t addr_conv[xor_src_cnt];
+       struct async_submit_ctl submit;
+       addr_conv_t addr_conv[NDISKS];
+       struct completion cmp;
+
+       init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST, NULL, NULL, NULL,
+                         addr_conv);
+       tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len, &submit)
 
-       tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len,
-                      ASYNC_TX_XOR_DROP_DST, NULL, NULL, NULL);
-       tx = async_memcpy(copy_dest, copy_src, 0, 0, copy_len,
-                         ASYNC_TX_DEP_ACK, tx, NULL, NULL);
-       tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len,
-                      ASYNC_TX_XOR_DROP_DST | ASYNC_TX_DEP_ACK | ASYNC_TX_ACK,
-                      tx, complete_xor_copy_xor, NULL);
+       submit->depend_tx = tx;
+       tx = async_memcpy(copy_dest, copy_src, 0, 0, copy_len, &submit);
+
+       init_completion(&cmp);
+       init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST | ASYNC_TX_ACK, tx,
+                         callback, &cmp, addr_conv);
+       tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len, &submit);
 
        async_tx_issue_pending_all();
+
+       wait_for_completion(&cmp);
 }
 
 See include/linux/async_tx.h for more information on the flags.  See the
index 83e6ba3..26eefea 100644 (file)
@@ -756,13 +756,14 @@ static inline void iop_desc_set_block_fill_val(struct iop_adma_desc_slot *desc,
        hw_desc->src[0] = val;
 }
 
-static inline int iop_desc_get_zero_result(struct iop_adma_desc_slot *desc)
+static inline enum sum_check_flags
+iop_desc_get_zero_result(struct iop_adma_desc_slot *desc)
 {
        struct iop3xx_desc_aau *hw_desc = desc->hw_desc;
        struct iop3xx_aau_desc_ctrl desc_ctrl = hw_desc->desc_ctrl_field;
 
        iop_paranoia(!(desc_ctrl.tx_complete && desc_ctrl.zero_result_en));
-       return desc_ctrl.zero_result_err;
+       return desc_ctrl.zero_result_err << SUM_CHECK_P;
 }
 
 static inline void iop_chan_append(struct iop_adma_chan *chan)
index 5722e86..1cd31df 100644 (file)
@@ -428,18 +428,20 @@ static inline void iop_desc_set_block_fill_val(struct iop_adma_desc_slot *desc,
        hw_desc->block_fill_data = val;
 }
 
-static inline int iop_desc_get_zero_result(struct iop_adma_desc_slot *desc)
+static inline enum sum_check_flags
+iop_desc_get_zero_result(struct iop_adma_desc_slot *desc)
 {
        struct iop13xx_adma_desc_hw *hw_desc = desc->hw_desc;
        struct iop13xx_adma_desc_ctrl desc_ctrl = hw_desc->desc_ctrl_field;
        struct iop13xx_adma_byte_count byte_count = hw_desc->byte_count_field;
+       enum sum_check_flags flags;
 
        BUG_ON(!(byte_count.tx_complete && desc_ctrl.zero_result));
 
-       if (desc_ctrl.pq_xfer_en)
-               return byte_count.zero_result_err_q;
-       else
-               return byte_count.zero_result_err;
+       flags = byte_count.zero_result_err_q << SUM_CHECK_Q;
+       flags |= byte_count.zero_result_err << SUM_CHECK_P;
+
+       return flags;
 }
 
 static inline void iop_chan_append(struct iop_adma_chan *chan)
index bee42c6..faaef95 100644 (file)
@@ -478,7 +478,7 @@ void __init iop13xx_platform_init(void)
                        dma_cap_set(DMA_MEMCPY, plat_data->cap_mask);
                        dma_cap_set(DMA_XOR, plat_data->cap_mask);
                        dma_cap_set(DMA_DUAL_XOR, plat_data->cap_mask);
-                       dma_cap_set(DMA_ZERO_SUM, plat_data->cap_mask);
+                       dma_cap_set(DMA_XOR_VAL, plat_data->cap_mask);
                        dma_cap_set(DMA_MEMSET, plat_data->cap_mask);
                        dma_cap_set(DMA_MEMCPY_CRC32C, plat_data->cap_mask);
                        dma_cap_set(DMA_INTERRUPT, plat_data->cap_mask);
@@ -490,7 +490,7 @@ void __init iop13xx_platform_init(void)
                        dma_cap_set(DMA_MEMCPY, plat_data->cap_mask);
                        dma_cap_set(DMA_XOR, plat_data->cap_mask);
                        dma_cap_set(DMA_DUAL_XOR, plat_data->cap_mask);
-                       dma_cap_set(DMA_ZERO_SUM, plat_data->cap_mask);
+                       dma_cap_set(DMA_XOR_VAL, plat_data->cap_mask);
                        dma_cap_set(DMA_MEMSET, plat_data->cap_mask);
                        dma_cap_set(DMA_MEMCPY_CRC32C, plat_data->cap_mask);
                        dma_cap_set(DMA_INTERRUPT, plat_data->cap_mask);
@@ -502,13 +502,13 @@ void __init iop13xx_platform_init(void)
                        dma_cap_set(DMA_MEMCPY, plat_data->cap_mask);
                        dma_cap_set(DMA_XOR, plat_data->cap_mask);
                        dma_cap_set(DMA_DUAL_XOR, plat_data->cap_mask);
-                       dma_cap_set(DMA_ZERO_SUM, plat_data->cap_mask);
+                       dma_cap_set(DMA_XOR_VAL, plat_data->cap_mask);
                        dma_cap_set(DMA_MEMSET, plat_data->cap_mask);
                        dma_cap_set(DMA_MEMCPY_CRC32C, plat_data->cap_mask);
                        dma_cap_set(DMA_INTERRUPT, plat_data->cap_mask);
-                       dma_cap_set(DMA_PQ_XOR, plat_data->cap_mask);
+                       dma_cap_set(DMA_PQ, plat_data->cap_mask);
                        dma_cap_set(DMA_PQ_UPDATE, plat_data->cap_mask);
-                       dma_cap_set(DMA_PQ_ZERO_SUM, plat_data->cap_mask);
+                       dma_cap_set(DMA_PQ_VAL, plat_data->cap_mask);
                        break;
                }
        }
index 3c127aa..da1dd0d 100644 (file)
@@ -198,7 +198,7 @@ static int __init iop3xx_adma_cap_init(void)
        dma_cap_set(DMA_INTERRUPT, iop3xx_aau_data.cap_mask);
        #else
        dma_cap_set(DMA_XOR, iop3xx_aau_data.cap_mask);
-       dma_cap_set(DMA_ZERO_SUM, iop3xx_aau_data.cap_mask);
+       dma_cap_set(DMA_XOR_VAL, iop3xx_aau_data.cap_mask);
        dma_cap_set(DMA_MEMSET, iop3xx_aau_data.cap_mask);
        dma_cap_set(DMA_INTERRUPT, iop3xx_aau_data.cap_mask);
        #endif
index d8fb391..e5aeb2b 100644 (file)
@@ -14,3 +14,12 @@ config ASYNC_MEMSET
        tristate
        select ASYNC_CORE
 
+config ASYNC_PQ
+       tristate
+       select ASYNC_CORE
+
+config ASYNC_RAID6_RECOV
+       tristate
+       select ASYNC_CORE
+       select ASYNC_PQ
+
index 27baa7d..d1e0e6f 100644 (file)
@@ -2,3 +2,6 @@ obj-$(CONFIG_ASYNC_CORE) += async_tx.o
 obj-$(CONFIG_ASYNC_MEMCPY) += async_memcpy.o
 obj-$(CONFIG_ASYNC_MEMSET) += async_memset.o
 obj-$(CONFIG_ASYNC_XOR) += async_xor.o
+obj-$(CONFIG_ASYNC_PQ) += async_pq.o
+obj-$(CONFIG_ASYNC_RAID6_RECOV) += async_raid6_recov.o
+obj-$(CONFIG_ASYNC_RAID6_TEST) += raid6test.o
index ddccfb0..98e15bd 100644 (file)
  * async_memcpy - attempt to copy memory with a dma engine.
  * @dest: destination page
  * @src: src page
- * @offset: offset in pages to start transaction
+ * @dest_offset: offset into 'dest' to start transaction
+ * @src_offset: offset into 'src' to start transaction
  * @len: length in bytes
- * @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK,
- * @depend_tx: memcpy depends on the result of this transaction
- * @cb_fn: function to call when the memcpy completes
- * @cb_param: parameter to pass to the callback routine
+ * @submit: submission / completion modifiers
+ *
+ * honored flags: ASYNC_TX_ACK
  */
 struct dma_async_tx_descriptor *
 async_memcpy(struct page *dest, struct page *src, unsigned int dest_offset,
-       unsigned int src_offset, size_t len, enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_param)
+            unsigned int src_offset, size_t len,
+            struct async_submit_ctl *submit)
 {
-       struct dma_chan *chan = async_tx_find_channel(depend_tx, DMA_MEMCPY,
+       struct dma_chan *chan = async_tx_find_channel(submit, DMA_MEMCPY,
                                                      &dest, 1, &src, 1, len);
        struct dma_device *device = chan ? chan->device : NULL;
        struct dma_async_tx_descriptor *tx = NULL;
 
        if (device) {
                dma_addr_t dma_dest, dma_src;
-               unsigned long dma_prep_flags = cb_fn ? DMA_PREP_INTERRUPT : 0;
+               unsigned long dma_prep_flags;
 
+               dma_prep_flags = submit->cb_fn ? DMA_PREP_INTERRUPT : 0;
                dma_dest = dma_map_page(device->dev, dest, dest_offset, len,
                                        DMA_FROM_DEVICE);
 
@@ -67,13 +67,13 @@ async_memcpy(struct page *dest, struct page *src, unsigned int dest_offset,
 
        if (tx) {
                pr_debug("%s: (async) len: %zu\n", __func__, len);
-               async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
+               async_tx_submit(chan, tx, submit);
        } else {
                void *dest_buf, *src_buf;
                pr_debug("%s: (sync) len: %zu\n", __func__, len);
 
                /* wait for any prerequisite operations */
-               async_tx_quiesce(&depend_tx);
+               async_tx_quiesce(&submit->depend_tx);
 
                dest_buf = kmap_atomic(dest, KM_USER0) + dest_offset;
                src_buf = kmap_atomic(src, KM_USER1) + src_offset;
@@ -83,26 +83,13 @@ async_memcpy(struct page *dest, struct page *src, unsigned int dest_offset,
                kunmap_atomic(dest_buf, KM_USER0);
                kunmap_atomic(src_buf, KM_USER1);
 
-               async_tx_sync_epilog(cb_fn, cb_param);
+               async_tx_sync_epilog(submit);
        }
 
        return tx;
 }
 EXPORT_SYMBOL_GPL(async_memcpy);
 
-static int __init async_memcpy_init(void)
-{
-       return 0;
-}
-
-static void __exit async_memcpy_exit(void)
-{
-       do { } while (0);
-}
-
-module_init(async_memcpy_init);
-module_exit(async_memcpy_exit);
-
 MODULE_AUTHOR("Intel Corporation");
 MODULE_DESCRIPTION("asynchronous memcpy api");
 MODULE_LICENSE("GPL");
index 5b5eb99..b896a6e 100644 (file)
  * @val: fill value
  * @offset: offset in pages to start transaction
  * @len: length in bytes
- * @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
- * @depend_tx: memset depends on the result of this transaction
- * @cb_fn: function to call when the memcpy completes
- * @cb_param: parameter to pass to the callback routine
+ *
+ * honored flags: ASYNC_TX_ACK
  */
 struct dma_async_tx_descriptor *
-async_memset(struct page *dest, int val, unsigned int offset,
-       size_t len, enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_param)
+async_memset(struct page *dest, int val, unsigned int offset, size_t len,
+            struct async_submit_ctl *submit)
 {
-       struct dma_chan *chan = async_tx_find_channel(depend_tx, DMA_MEMSET,
+       struct dma_chan *chan = async_tx_find_channel(submit, DMA_MEMSET,
                                                      &dest, 1, NULL, 0, len);
        struct dma_device *device = chan ? chan->device : NULL;
        struct dma_async_tx_descriptor *tx = NULL;
 
        if (device) {
                dma_addr_t dma_dest;
-               unsigned long dma_prep_flags = cb_fn ? DMA_PREP_INTERRUPT : 0;
+               unsigned long dma_prep_flags;
 
+               dma_prep_flags = submit->cb_fn ? DMA_PREP_INTERRUPT : 0;
                dma_dest = dma_map_page(device->dev, dest, offset, len,
                                        DMA_FROM_DEVICE);
 
@@ -64,38 +61,25 @@ async_memset(struct page *dest, int val, unsigned int offset,
 
        if (tx) {
                pr_debug("%s: (async) len: %zu\n", __func__, len);
-               async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
+               async_tx_submit(chan, tx, submit);
        } else { /* run the memset synchronously */
                void *dest_buf;
                pr_debug("%s: (sync) len: %zu\n", __func__, len);
 
-               dest_buf = (void *) (((char *) page_address(dest)) + offset);
+               dest_buf = page_address(dest) + offset;
 
                /* wait for any prerequisite operations */
-               async_tx_quiesce(&depend_tx);
+               async_tx_quiesce(&submit->depend_tx);
 
                memset(dest_buf, val, len);
 
-               async_tx_sync_epilog(cb_fn, cb_param);
+               async_tx_sync_epilog(submit);
        }
 
        return tx;
 }
 EXPORT_SYMBOL_GPL(async_memset);
 
-static int __init async_memset_init(void)
-{
-       return 0;
-}
-
-static void __exit async_memset_exit(void)
-{
-       do { } while (0);
-}
-
-module_init(async_memset_init);
-module_exit(async_memset_exit);
-
 MODULE_AUTHOR("Intel Corporation");
 MODULE_DESCRIPTION("asynchronous memset api");
 MODULE_LICENSE("GPL");
diff --git a/crypto/async_tx/async_pq.c b/crypto/async_tx/async_pq.c
new file mode 100644 (file)
index 0000000..108b21e
--- /dev/null
@@ -0,0 +1,388 @@
+/*
+ * Copyright(c) 2007 Yuri Tikhonov <yur@emcraft.com>
+ * Copyright(c) 2009 Intel Corporation
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License as published by the Free
+ * Software Foundation; either version 2 of the License, or (at your option)
+ * any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 59
+ * Temple Place - Suite 330, Boston, MA  02111-1307, USA.
+ *
+ * The full GNU General Public License is included in this distribution in the
+ * file called COPYING.
+ */
+#include <linux/kernel.h>
+#include <linux/interrupt.h>
+#include <linux/dma-mapping.h>
+#include <linux/raid/pq.h>
+#include <linux/async_tx.h>
+
+/**
+ * scribble - space to hold throwaway P buffer for synchronous gen_syndrome
+ */
+static struct page *scribble;
+
+static bool is_raid6_zero_block(struct page *p)
+{
+       return p == (void *) raid6_empty_zero_page;
+}
+
+/* the struct page *blocks[] parameter passed to async_gen_syndrome()
+ * and async_syndrome_val() contains the 'P' destination address at
+ * blocks[disks-2] and the 'Q' destination address at blocks[disks-1]
+ *
+ * note: these are macros as they are used as lvalues
+ */
+#define P(b, d) (b[d-2])
+#define Q(b, d) (b[d-1])
+
+/**
+ * do_async_gen_syndrome - asynchronously calculate P and/or Q
+ */
+static __async_inline struct dma_async_tx_descriptor *
+do_async_gen_syndrome(struct dma_chan *chan, struct page **blocks,
+                     const unsigned char *scfs, unsigned int offset, int disks,
+                     size_t len, dma_addr_t *dma_src,
+                     struct async_submit_ctl *submit)
+{
+       struct dma_async_tx_descriptor *tx = NULL;
+       struct dma_device *dma = chan->device;
+       enum dma_ctrl_flags dma_flags = 0;
+       enum async_tx_flags flags_orig = submit->flags;
+       dma_async_tx_callback cb_fn_orig = submit->cb_fn;
+       dma_async_tx_callback cb_param_orig = submit->cb_param;
+       int src_cnt = disks - 2;
+       unsigned char coefs[src_cnt];
+       unsigned short pq_src_cnt;
+       dma_addr_t dma_dest[2];
+       int src_off = 0;
+       int idx;
+       int i;
+
+       /* DMAs use destinations as sources, so use BIDIRECTIONAL mapping */
+       if (P(blocks, disks))
+               dma_dest[0] = dma_map_page(dma->dev, P(blocks, disks), offset,
+                                          len, DMA_BIDIRECTIONAL);
+       else
+               dma_flags |= DMA_PREP_PQ_DISABLE_P;
+       if (Q(blocks, disks))
+               dma_dest[1] = dma_map_page(dma->dev, Q(blocks, disks), offset,
+                                          len, DMA_BIDIRECTIONAL);
+       else
+               dma_flags |= DMA_PREP_PQ_DISABLE_Q;
+
+       /* convert source addresses being careful to collapse 'empty'
+        * sources and update the coefficients accordingly
+        */
+       for (i = 0, idx = 0; i < src_cnt; i++) {
+               if (is_raid6_zero_block(blocks[i]))
+                       continue;
+               dma_src[idx] = dma_map_page(dma->dev, blocks[i], offset, len,
+                                           DMA_TO_DEVICE);
+               coefs[idx] = scfs[i];
+               idx++;
+       }
+       src_cnt = idx;
+
+       while (src_cnt > 0) {
+               submit->flags = flags_orig;
+               pq_src_cnt = min(src_cnt, dma_maxpq(dma, dma_flags));
+               /* if we are submitting additional pqs, leave the chain open,
+                * clear the callback parameters, and leave the destination
+                * buffers mapped
+                */
+               if (src_cnt > pq_src_cnt) {
+                       submit->flags &= ~ASYNC_TX_ACK;
+                       dma_flags |= DMA_COMPL_SKIP_DEST_UNMAP;
+                       submit->cb_fn = NULL;
+                       submit->cb_param = NULL;
+               } else {
+                       dma_flags &= ~DMA_COMPL_SKIP_DEST_UNMAP;
+                       submit->cb_fn = cb_fn_orig;
+                       submit->cb_param = cb_param_orig;
+                       if (cb_fn_orig)
+                               dma_flags |= DMA_PREP_INTERRUPT;
+               }
+
+               /* Since we have clobbered the src_list we are committed
+                * to doing this asynchronously.  Drivers force forward
+                * progress in case they can not provide a descriptor
+                */
+               for (;;) {
+                       tx = dma->device_prep_dma_pq(chan, dma_dest,
+                                                    &dma_src[src_off],
+                                                    pq_src_cnt,
+                                                    &coefs[src_off], len,
+                                                    dma_flags);
+                       if (likely(tx))
+                               break;
+                       async_tx_quiesce(&submit->depend_tx);
+                       dma_async_issue_pending(chan);
+               }
+
+               async_tx_submit(chan, tx, submit);
+               submit->depend_tx = tx;
+
+               /* drop completed sources */
+               src_cnt -= pq_src_cnt;
+               src_off += pq_src_cnt;
+
+               dma_flags |= DMA_PREP_CONTINUE;
+       }
+
+       return tx;
+}
+
+/**
+ * do_sync_gen_syndrome - synchronously calculate a raid6 syndrome
+ */
+static void
+do_sync_gen_syndrome(struct page **blocks, unsigned int offset, int disks,
+                    size_t len, struct async_submit_ctl *submit)
+{
+       void **srcs;
+       int i;
+
+       if (submit->scribble)
+               srcs = submit->scribble;
+       else
+               srcs = (void **) blocks;
+
+       for (i = 0; i < disks; i++) {
+               if (is_raid6_zero_block(blocks[i])) {
+                       BUG_ON(i > disks - 3); /* P or Q can't be zero */
+                       srcs[i] = blocks[i];
+               } else
+                       srcs[i] = page_address(blocks[i]) + offset;
+       }
+       raid6_call.gen_syndrome(disks, len, srcs);
+       async_tx_sync_epilog(submit);
+}
+
+/**
+ * async_gen_syndrome - asynchronously calculate a raid6 syndrome
+ * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
+ * @offset: common offset into each block (src and dest) to start transaction
+ * @disks: number of blocks (including missing P or Q, see below)
+ * @len: length of operation in bytes
+ * @submit: submission/completion modifiers
+ *
+ * General note: This routine assumes a field of GF(2^8) with a
+ * primitive polynomial of 0x11d and a generator of {02}.
+ *
+ * 'disks' note: callers can optionally omit either P or Q (but not
+ * both) from the calculation by setting blocks[disks-2] or
+ * blocks[disks-1] to NULL.  When P or Q is omitted 'len' must be <=
+ * PAGE_SIZE as a temporary buffer of this size is used in the
+ * synchronous path.  'disks' always accounts for both destination
+ * buffers.
+ *
+ * 'blocks' note: if submit->scribble is NULL then the contents of
+ * 'blocks' may be overridden
+ */
+struct dma_async_tx_descriptor *
+async_gen_syndrome(struct page **blocks, unsigned int offset, int disks,
+                  size_t len, struct async_submit_ctl *submit)
+{
+       int src_cnt = disks - 2;
+       struct dma_chan *chan = async_tx_find_channel(submit, DMA_PQ,
+                                                     &P(blocks, disks), 2,
+                                                     blocks, src_cnt, len);
+       struct dma_device *device = chan ? chan->device : NULL;
+       dma_addr_t *dma_src = NULL;
+
+       BUG_ON(disks > 255 || !(P(blocks, disks) || Q(blocks, disks)));
+
+       if (submit->scribble)
+               dma_src = submit->scribble;
+       else if (sizeof(dma_addr_t) <= sizeof(struct page *))
+               dma_src = (dma_addr_t *) blocks;
+
+       if (dma_src && device &&
+           (src_cnt <= dma_maxpq(device, 0) ||
+            dma_maxpq(device, DMA_PREP_CONTINUE) > 0)) {
+               /* run the p+q asynchronously */
+               pr_debug("%s: (async) disks: %d len: %zu\n",
+                        __func__, disks, len);
+               return do_async_gen_syndrome(chan, blocks, raid6_gfexp, offset,
+                                            disks, len, dma_src, submit);
+       }
+
+       /* run the pq synchronously */
+       pr_debug("%s: (sync) disks: %d len: %zu\n", __func__, disks, len);
+
+       /* wait for any prerequisite operations */
+       async_tx_quiesce(&submit->depend_tx);
+
+       if (!P(blocks, disks)) {
+               P(blocks, disks) = scribble;
+               BUG_ON(len + offset > PAGE_SIZE);
+       }
+       if (!Q(blocks, disks)) {
+               Q(blocks, disks) = scribble;
+               BUG_ON(len + offset > PAGE_SIZE);
+       }
+       do_sync_gen_syndrome(blocks, offset, disks, len, submit);
+
+       return NULL;
+}
+EXPORT_SYMBOL_GPL(async_gen_syndrome);
+
+/**
+ * async_syndrome_val - asynchronously validate a raid6 syndrome
+ * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
+ * @offset: common offset into each block (src and dest) to start transaction
+ * @disks: number of blocks (including missing P or Q, see below)
+ * @len: length of operation in bytes
+ * @pqres: on val failure SUM_CHECK_P_RESULT and/or SUM_CHECK_Q_RESULT are set
+ * @spare: temporary result buffer for the synchronous case
+ * @submit: submission / completion modifiers
+ *
+ * The same notes from async_gen_syndrome apply to the 'blocks',
+ * and 'disks' parameters of this routine.  The synchronous path
+ * requires a temporary result buffer and submit->scribble to be
+ * specified.
+ */
+struct dma_async_tx_descriptor *
+async_syndrome_val(struct page **blocks, unsigned int offset, int disks,
+                  size_t len, enum sum_check_flags *pqres, struct page *spare,
+                  struct async_submit_ctl *submit)
+{
+       struct dma_chan *chan = async_tx_find_channel(submit, DMA_PQ_VAL,
+                                                     NULL, 0,  blocks, disks,
+                                                     len);
+       struct dma_device *device = chan ? chan->device : NULL;
+       struct dma_async_tx_descriptor *tx;
+       enum dma_ctrl_flags dma_flags = submit->cb_fn ? DMA_PREP_INTERRUPT : 0;
+       dma_addr_t *dma_src = NULL;
+
+       BUG_ON(disks < 4);
+
+       if (submit->scribble)
+               dma_src = submit->scribble;
+       else if (sizeof(dma_addr_t) <= sizeof(struct page *))
+               dma_src = (dma_addr_t *) blocks;
+
+       if (dma_src && device && disks <= dma_maxpq(device, 0)) {
+               struct device *dev = device->dev;
+               dma_addr_t *pq = &dma_src[disks-2];
+               int i;
+
+               pr_debug("%s: (async) disks: %d len: %zu\n",
+                        __func__, disks, len);
+               if (!P(blocks, disks))
+                       dma_flags |= DMA_PREP_PQ_DISABLE_P;
+               if (!Q(blocks, disks))
+                       dma_flags |= DMA_PREP_PQ_DISABLE_Q;
+               for (i = 0; i < disks; i++)
+                       if (likely(blocks[i])) {
+                               BUG_ON(is_raid6_zero_block(blocks[i]));
+                               dma_src[i] = dma_map_page(dev, blocks[i],
+                                                         offset, len,
+                                                         DMA_TO_DEVICE);
+                       }
+
+               for (;;) {
+                       tx = device->device_prep_dma_pq_val(chan, pq, dma_src,
+                                                           disks - 2,
+                                                           raid6_gfexp,
+                                                           len, pqres,
+                                                           dma_flags);
+                       if (likely(tx))
+                               break;
+                       async_tx_quiesce(&submit->depend_tx);
+                       dma_async_issue_pending(chan);
+               }
+               async_tx_submit(chan, tx, submit);
+
+               return tx;
+       } else {
+               struct page *p_src = P(blocks, disks);
+               struct page *q_src = Q(blocks, disks);
+               enum async_tx_flags flags_orig = submit->flags;
+               dma_async_tx_callback cb_fn_orig = submit->cb_fn;
+               void *scribble = submit->scribble;
+               void *cb_param_orig = submit->cb_param;
+               void *p, *q, *s;
+
+               pr_debug("%s: (sync) disks: %d len: %zu\n",
+                        __func__, disks, len);
+
+               /* caller must provide a temporary result buffer and
+                * allow the input parameters to be preserved
+                */
+               BUG_ON(!spare || !scribble);
+
+               /* wait for any prerequisite operations */
+               async_tx_quiesce(&submit->depend_tx);
+
+               /* recompute p and/or q into the temporary buffer and then
+                * check to see the result matches the current value
+                */
+               tx = NULL;
+               *pqres = 0;
+               if (p_src) {
+                       init_async_submit(submit, ASYNC_TX_XOR_ZERO_DST, NULL,
+                                         NULL, NULL, scribble);
+                       tx = async_xor(spare, blocks, offset, disks-2, len, submit);
+                       async_tx_quiesce(&tx);
+                       p = page_address(p_src) + offset;
+                       s = page_address(spare) + offset;
+                       *pqres |= !!memcmp(p, s, len) << SUM_CHECK_P;
+               }
+
+               if (q_src) {
+                       P(blocks, disks) = NULL;
+                       Q(blocks, disks) = spare;
+                       init_async_submit(submit, 0, NULL, NULL, NULL, scribble);
+                       tx = async_gen_syndrome(blocks, offset, disks, len, submit);
+                       async_tx_quiesce(&tx);
+                       q = page_address(q_src) + offset;
+                       s = page_address(spare) + offset;
+                       *pqres |= !!memcmp(q, s, len) << SUM_CHECK_Q;
+               }
+
+               /* restore P, Q and submit */
+               P(blocks, disks) = p_src;
+               Q(blocks, disks) = q_src;
+
+               submit->cb_fn = cb_fn_orig;
+               submit->cb_param = cb_param_orig;
+               submit->flags = flags_orig;
+               async_tx_sync_epilog(submit);
+
+               return NULL;
+       }
+}
+EXPORT_SYMBOL_GPL(async_syndrome_val);
+
+static int __init async_pq_init(void)
+{
+       scribble = alloc_page(GFP_KERNEL);
+
+       if (scribble)
+               return 0;
+
+       pr_err("%s: failed to allocate required spare page\n", __func__);
+
+       return -ENOMEM;
+}
+
+static void __exit async_pq_exit(void)
+{
+       put_page(scribble);
+}
+
+module_init(async_pq_init);
+module_exit(async_pq_exit);
+
+MODULE_DESCRIPTION("asynchronous raid6 syndrome generation/validation");
+MODULE_LICENSE("GPL");
diff --git a/crypto/async_tx/async_raid6_recov.c b/crypto/async_tx/async_raid6_recov.c
new file mode 100644 (file)
index 0000000..0c14d48
--- /dev/null
@@ -0,0 +1,448 @@
+/*
+ * Asynchronous RAID-6 recovery calculations ASYNC_TX API.
+ * Copyright(c) 2009 Intel Corporation
+ *
+ * based on raid6recov.c:
+ *   Copyright 2002 H. Peter Anvin
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License as published by the Free
+ * Software Foundation; either version 2 of the License, or (at your option)
+ * any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 51
+ * Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ */
+#include <linux/kernel.h>
+#include <linux/interrupt.h>
+#include <linux/dma-mapping.h>
+#include <linux/raid/pq.h>
+#include <linux/async_tx.h>
+
+static struct dma_async_tx_descriptor *
+async_sum_product(struct page *dest, struct page **srcs, unsigned char *coef,
+                 size_t len, struct async_submit_ctl *submit)
+{
+       struct dma_chan *chan = async_tx_find_channel(submit, DMA_PQ,
+                                                     &dest, 1, srcs, 2, len);
+       struct dma_device *dma = chan ? chan->device : NULL;
+       const u8 *amul, *bmul;
+       u8 ax, bx;
+       u8 *a, *b, *c;
+
+       if (dma) {
+               dma_addr_t dma_dest[2];
+               dma_addr_t dma_src[2];
+               struct device *dev = dma->dev;
+               struct dma_async_tx_descriptor *tx;
+               enum dma_ctrl_flags dma_flags = DMA_PREP_PQ_DISABLE_P;
+
+               dma_dest[1] = dma_map_page(dev, dest, 0, len, DMA_BIDIRECTIONAL);
+               dma_src[0] = dma_map_page(dev, srcs[0], 0, len, DMA_TO_DEVICE);
+               dma_src[1] = dma_map_page(dev, srcs[1], 0, len, DMA_TO_DEVICE);
+               tx = dma->device_prep_dma_pq(chan, dma_dest, dma_src, 2, coef,
+                                            len, dma_flags);
+               if (tx) {
+                       async_tx_submit(chan, tx, submit);
+                       return tx;
+               }
+       }
+
+       /* run the operation synchronously */
+       async_tx_quiesce(&submit->depend_tx);
+       amul = raid6_gfmul[coef[0]];
+       bmul = raid6_gfmul[coef[1]];
+       a = page_address(srcs[0]);
+       b = page_address(srcs[1]);
+       c = page_address(dest);
+
+       while (len--) {
+               ax    = amul[*a++];
+               bx    = bmul[*b++];
+               *c++ = ax ^ bx;
+       }
+
+       return NULL;
+}
+
+static struct dma_async_tx_descriptor *
+async_mult(struct page *dest, struct page *src, u8 coef, size_t len,
+          struct async_submit_ctl *submit)
+{
+       struct dma_chan *chan = async_tx_find_channel(submit, DMA_PQ,
+                                                     &dest, 1, &src, 1, len);
+       struct dma_device *dma = chan ? chan->device : NULL;
+       const u8 *qmul; /* Q multiplier table */
+       u8 *d, *s;
+
+       if (dma) {
+               dma_addr_t dma_dest[2];
+               dma_addr_t dma_src[1];
+               struct device *dev = dma->dev;
+               struct dma_async_tx_descriptor *tx;
+               enum dma_ctrl_flags dma_flags = DMA_PREP_PQ_DISABLE_P;
+
+               dma_dest[1] = dma_map_page(dev, dest, 0, len, DMA_BIDIRECTIONAL);
+               dma_src[0] = dma_map_page(dev, src, 0, len, DMA_TO_DEVICE);
+               tx = dma->device_prep_dma_pq(chan, dma_dest, dma_src, 1, &coef,
+                                            len, dma_flags);
+               if (tx) {
+                       async_tx_submit(chan, tx, submit);
+                       return tx;
+               }
+       }
+
+       /* no channel available, or failed to allocate a descriptor, so
+        * perform the operation synchronously
+        */
+       async_tx_quiesce(&submit->depend_tx);
+       qmul  = raid6_gfmul[coef];
+       d = page_address(dest);
+       s = page_address(src);
+
+       while (len--)
+               *d++ = qmul[*s++];
+
+       return NULL;
+}
+
+static struct dma_async_tx_descriptor *
+__2data_recov_4(size_t bytes, int faila, int failb, struct page **blocks,
+             struct async_submit_ctl *submit)
+{
+       struct dma_async_tx_descriptor *tx = NULL;
+       struct page *p, *q, *a, *b;
+       struct page *srcs[2];
+       unsigned char coef[2];
+       enum async_tx_flags flags = submit->flags;
+       dma_async_tx_callback cb_fn = submit->cb_fn;
+       void *cb_param = submit->cb_param;
+       void *scribble = submit->scribble;
+
+       p = blocks[4-2];
+       q = blocks[4-1];
+
+       a = blocks[faila];
+       b = blocks[failb];
+
+       /* in the 4 disk case P + Pxy == P and Q + Qxy == Q */
+       /* Dx = A*(P+Pxy) + B*(Q+Qxy) */
+       srcs[0] = p;
+       srcs[1] = q;
+       coef[0] = raid6_gfexi[failb-faila];
+       coef[1] = raid6_gfinv[raid6_gfexp[faila]^raid6_gfexp[failb]];
+       init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+       tx = async_sum_product(b, srcs, coef, bytes, submit);
+
+       /* Dy = P+Pxy+Dx */
+       srcs[0] = p;
+       srcs[1] = b;
+       init_async_submit(submit, flags | ASYNC_TX_XOR_ZERO_DST, tx, cb_fn,
+                         cb_param, scribble);
+       tx = async_xor(a, srcs, 0, 2, bytes, submit);
+
+       return tx;
+
+}
+
+static struct dma_async_tx_descriptor *
+__2data_recov_5(size_t bytes, int faila, int failb, struct page **blocks,
+             struct async_submit_ctl *submit)
+{
+       struct dma_async_tx_descriptor *tx = NULL;
+       struct page *p, *q, *g, *dp, *dq;
+       struct page *srcs[2];
+       unsigned char coef[2];
+       enum async_tx_flags flags = submit->flags;
+       dma_async_tx_callback cb_fn = submit->cb_fn;
+       void *cb_param = submit->cb_param;
+       void *scribble = submit->scribble;
+       int uninitialized_var(good);
+       int i;
+
+       for (i = 0; i < 3; i++) {
+               if (i == faila || i == failb)
+                       continue;
+               else {
+                       good = i;
+                       break;
+               }
+       }
+       BUG_ON(i >= 3);
+
+       p = blocks[5-2];
+       q = blocks[5-1];
+       g = blocks[good];
+
+       /* Compute syndrome with zero for the missing data pages
+        * Use the dead data pages as temporary storage for delta p and
+        * delta q
+        */
+       dp = blocks[faila];
+       dq = blocks[failb];
+
+       init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+       tx = async_memcpy(dp, g, 0, 0, bytes, submit);
+       init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+       tx = async_mult(dq, g, raid6_gfexp[good], bytes, submit);
+
+       /* compute P + Pxy */
+       srcs[0] = dp;
+       srcs[1] = p;
+       init_async_submit(submit, ASYNC_TX_XOR_DROP_DST, tx, NULL, NULL,
+                         scribble);
+       tx = async_xor(dp, srcs, 0, 2, bytes, submit);
+
+       /* compute Q + Qxy */
+       srcs[0] = dq;
+       srcs[1] = q;
+       init_async_submit(submit, ASYNC_TX_XOR_DROP_DST, tx, NULL, NULL,
+                         scribble);
+       tx = async_xor(dq, srcs, 0, 2, bytes, submit);
+
+       /* Dx = A*(P+Pxy) + B*(Q+Qxy) */
+       srcs[0] = dp;
+       srcs[1] = dq;
+       coef[0] = raid6_gfexi[failb-faila];
+       coef[1] = raid6_gfinv[raid6_gfexp[faila]^raid6_gfexp[failb]];
+       init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+       tx = async_sum_product(dq, srcs, coef, bytes, submit);
+
+       /* Dy = P+Pxy+Dx */
+       srcs[0] = dp;
+       srcs[1] = dq;
+       init_async_submit(submit, flags | ASYNC_TX_XOR_DROP_DST, tx, cb_fn,
+                         cb_param, scribble);
+       tx = async_xor(dp, srcs, 0, 2, bytes, submit);
+
+       return tx;
+}
+
+static struct dma_async_tx_descriptor *
+__2data_recov_n(int disks, size_t bytes, int faila, int failb,
+             struct page **blocks, struct async_submit_ctl *submit)
+{
+       struct dma_async_tx_descriptor *tx = NULL;
+       struct page *p, *q, *dp, *dq;
+       struct page *srcs[2];
+       unsigned char coef[2];
+       enum async_tx_flags flags = submit->flags;
+       dma_async_tx_callback cb_fn = submit->cb_fn;
+       void *cb_param = submit->cb_param;
+       void *scribble = submit->scribble;
+
+       p = blocks[disks-2];
+       q = blocks[disks-1];
+
+       /* Compute syndrome with zero for the missing data pages
+        * Use the dead data pages as temporary storage for
+        * delta p and delta q
+        */
+       dp = blocks[faila];
+       blocks[faila] = (void *)raid6_empty_zero_page;
+       blocks[disks-2] = dp;
+       dq = blocks[failb];
+       blocks[failb] = (void *)raid6_empty_zero_page;
+       blocks[disks-1] = dq;
+
+       init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+       tx = async_gen_syndrome(blocks, 0, disks, bytes, submit);
+
+       /* Restore pointer table */
+       blocks[faila]   = dp;
+       blocks[failb]   = dq;
+       blocks[disks-2] = p;
+       blocks[disks-1] = q;
+
+       /* compute P + Pxy */
+       srcs[0] = dp;
+       srcs[1] = p;
+       init_async_submit(submit, ASYNC_TX_XOR_DROP_DST, tx, NULL, NULL,
+                         scribble);
+       tx = async_xor(dp, srcs, 0, 2, bytes, submit);
+
+       /* compute Q + Qxy */
+       srcs[0] = dq;
+       srcs[1] = q;
+       init_async_submit(submit, ASYNC_TX_XOR_DROP_DST, tx, NULL, NULL,
+                         scribble);
+       tx = async_xor(dq, srcs, 0, 2, bytes, submit);
+
+       /* Dx = A*(P+Pxy) + B*(Q+Qxy) */
+       srcs[0] = dp;
+       srcs[1] = dq;
+       coef[0] = raid6_gfexi[failb-faila];
+       coef[1] = raid6_gfinv[raid6_gfexp[faila]^raid6_gfexp[failb]];
+       init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+       tx = async_sum_product(dq, srcs, coef, bytes, submit);
+
+       /* Dy = P+Pxy+Dx */
+       srcs[0] = dp;
+       srcs[1] = dq;
+       init_async_submit(submit, flags | ASYNC_TX_XOR_DROP_DST, tx, cb_fn,
+                         cb_param, scribble);
+       tx = async_xor(dp, srcs, 0, 2, bytes, submit);
+
+       return tx;
+}
+
+/**
+ * async_raid6_2data_recov - asynchronously calculate two missing data blocks
+ * @disks: number of disks in the RAID-6 array
+ * @bytes: block size
+ * @faila: first failed drive index
+ * @failb: second failed drive index
+ * @blocks: array of source pointers where the last two entries are p and q
+ * @submit: submission/completion modifiers
+ */
+struct dma_async_tx_descriptor *
+async_raid6_2data_recov(int disks, size_t bytes, int faila, int failb,
+                       struct page **blocks, struct async_submit_ctl *submit)
+{
+       BUG_ON(faila == failb);
+       if (failb < faila)
+               swap(faila, failb);
+
+       pr_debug("%s: disks: %d len: %zu\n", __func__, disks, bytes);
+
+       /* we need to preserve the contents of 'blocks' for the async
+        * case, so punt to synchronous if a scribble buffer is not available
+        */
+       if (!submit->scribble) {
+               void **ptrs = (void **) blocks;
+               int i;
+
+               async_tx_quiesce(&submit->depend_tx);
+               for (i = 0; i < disks; i++)
+                       ptrs[i] = page_address(blocks[i]);
+
+               raid6_2data_recov(disks, bytes, faila, failb, ptrs);
+
+               async_tx_sync_epilog(submit);
+
+               return NULL;
+       }
+
+       switch (disks) {
+       case 4:
+               /* dma devices do not uniformly understand a zero source pq
+                * operation (in contrast to the synchronous case), so
+                * explicitly handle the 4 disk special case
+                */
+               return __2data_recov_4(bytes, faila, failb, blocks, submit);
+       case 5:
+               /* dma devices do not uniformly understand a single
+                * source pq operation (in contrast to the synchronous
+                * case), so explicitly handle the 5 disk special case
+                */
+               return __2data_recov_5(bytes, faila, failb, blocks, submit);
+       default:
+               return __2data_recov_n(disks, bytes, faila, failb, blocks, submit);
+       }
+}
+EXPORT_SYMBOL_GPL(async_raid6_2data_recov);
+
+/**
+ * async_raid6_datap_recov - asynchronously calculate a data and the 'p' block
+ * @disks: number of disks in the RAID-6 array
+ * @bytes: block size
+ * @faila: failed drive index
+ * @blocks: array of source pointers where the last two entries are p and q
+ * @submit: submission/completion modifiers
+ */
+struct dma_async_tx_descriptor *
+async_raid6_datap_recov(int disks, size_t bytes, int faila,
+                       struct page **blocks, struct async_submit_ctl *submit)
+{
+       struct dma_async_tx_descriptor *tx = NULL;
+       struct page *p, *q, *dq;
+       u8 coef;
+       enum async_tx_flags flags = submit->flags;
+       dma_async_tx_callback cb_fn = submit->cb_fn;
+       void *cb_param = submit->cb_param;
+       void *scribble = submit->scribble;
+       struct page *srcs[2];
+
+       pr_debug("%s: disks: %d len: %zu\n", __func__, disks, bytes);
+
+       /* we need to preserve the contents of 'blocks' for the async
+        * case, so punt to synchronous if a scribble buffer is not available
+        */
+       if (!scribble) {
+               void **ptrs = (void **) blocks;
+               int i;
+
+               async_tx_quiesce(&submit->depend_tx);
+               for (i = 0; i < disks; i++)
+                       ptrs[i] = page_address(blocks[i]);
+
+               raid6_datap_recov(disks, bytes, faila, ptrs);
+
+               async_tx_sync_epilog(submit);
+
+               return NULL;
+       }
+
+       p = blocks[disks-2];
+       q = blocks[disks-1];
+
+       /* Compute syndrome with zero for the missing data page
+        * Use the dead data page as temporary storage for delta q
+        */
+       dq = blocks[faila];
+       blocks[faila] = (void *)raid6_empty_zero_page;
+       blocks[disks-1] = dq;
+
+       /* in the 4 disk case we only need to perform a single source
+        * multiplication
+        */
+       if (disks == 4) {
+               int good = faila == 0 ? 1 : 0;
+               struct page *g = blocks[good];
+
+               init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+               tx = async_memcpy(p, g, 0, 0, bytes, submit);
+
+               init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+               tx = async_mult(dq, g, raid6_gfexp[good], bytes, submit);
+       } else {
+               init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+               tx = async_gen_syndrome(blocks, 0, disks, bytes, submit);
+       }
+
+       /* Restore pointer table */
+       blocks[faila]   = dq;
+       blocks[disks-1] = q;
+
+       /* calculate g^{-faila} */
+       coef = raid6_gfinv[raid6_gfexp[faila]];
+
+       srcs[0] = dq;
+       srcs[1] = q;
+       init_async_submit(submit, ASYNC_TX_XOR_DROP_DST, tx, NULL, NULL,
+                         scribble);
+       tx = async_xor(dq, srcs, 0, 2, bytes, submit);
+
+       init_async_submit(submit, 0, tx, NULL, NULL, scribble);
+       tx = async_mult(dq, dq, coef, bytes, submit);
+
+       srcs[0] = p;
+       srcs[1] = dq;
+       init_async_submit(submit, flags | ASYNC_TX_XOR_DROP_DST, tx, cb_fn,
+                         cb_param, scribble);
+       tx = async_xor(p, srcs, 0, 2, bytes, submit);
+
+       return tx;
+}
+EXPORT_SYMBOL_GPL(async_raid6_datap_recov);
+
+MODULE_AUTHOR("Dan Williams <dan.j.williams@intel.com>");
+MODULE_DESCRIPTION("asynchronous RAID-6 recovery api");
+MODULE_LICENSE("GPL");
index 06eb6cc..60615fe 100644 (file)
@@ -42,16 +42,21 @@ static void __exit async_tx_exit(void)
        async_dmaengine_put();
 }
 
+module_init(async_tx_init);
+module_exit(async_tx_exit);
+
 /**
  * __async_tx_find_channel - find a channel to carry out the operation or let
  *     the transaction execute synchronously
- * @depend_tx: transaction dependency
+ * @submit: transaction dependency and submission modifiers
  * @tx_type: transaction type
  */
 struct dma_chan *
-__async_tx_find_channel(struct dma_async_tx_descriptor *depend_tx,
-       enum dma_transaction_type tx_type)
+__async_tx_find_channel(struct async_submit_ctl *submit,
+                       enum dma_transaction_type tx_type)
 {
+       struct dma_async_tx_descriptor *depend_tx = submit->depend_tx;
+
        /* see if we can keep the chain on one channel */
        if (depend_tx &&
            dma_has_cap(tx_type, depend_tx->chan->device->cap_mask))
@@ -59,17 +64,6 @@ __async_tx_find_channel(struct dma_async_tx_descriptor *depend_tx,
        return async_dma_find_channel(tx_type);
 }
 EXPORT_SYMBOL_GPL(__async_tx_find_channel);
-#else
-static int __init async_tx_init(void)
-{
-       printk(KERN_INFO "async_tx: api initialized (sync-only)\n");
-       return 0;
-}
-
-static void __exit async_tx_exit(void)
-{
-       do { } while (0);
-}
 #endif
 
 
@@ -83,8 +77,8 @@ static void
 async_tx_channel_switch(struct dma_async_tx_descriptor *depend_tx,
                        struct dma_async_tx_descriptor *tx)
 {
-       struct dma_chan *chan;
-       struct dma_device *device;
+       struct dma_chan *chan = depend_tx->chan;
+       struct dma_device *device = chan->device;
        struct dma_async_tx_descriptor *intr_tx = (void *) ~0;
 
        /* first check to see if we can still append to depend_tx */
@@ -96,11 +90,11 @@ async_tx_channel_switch(struct dma_async_tx_descriptor *depend_tx,
        }
        spin_unlock_bh(&depend_tx->lock);
 
-       if (!intr_tx)
+       /* attached dependency, flush the parent channel */
+       if (!intr_tx) {
+               device->device_issue_pending(chan);
                return;
-
-       chan = depend_tx->chan;
-       device = chan->device;
+       }
 
        /* see if we can schedule an interrupt
         * otherwise poll for completion
@@ -134,6 +128,7 @@ async_tx_channel_switch(struct dma_async_tx_descriptor *depend_tx,
                        intr_tx->tx_submit(intr_tx);
                        async_tx_ack(intr_tx);
                }
+               device->device_issue_pending(chan);
        } else {
                if (dma_wait_for_async_tx(depend_tx) == DMA_ERROR)
                        panic("%s: DMA_ERROR waiting for depend_tx\n",
@@ -144,13 +139,14 @@ async_tx_channel_switch(struct dma_async_tx_descriptor *depend_tx,
 
 
 /**
- * submit_disposition - while holding depend_tx->lock we must avoid submitting
- *     new operations to prevent a circular locking dependency with
- *     drivers that already hold a channel lock when calling
- *     async_tx_run_dependencies.
+ * submit_disposition - flags for routing an incoming operation
  * @ASYNC_TX_SUBMITTED: we were able to append the new operation under the lock
  * @ASYNC_TX_CHANNEL_SWITCH: when the lock is dropped schedule a channel switch
  * @ASYNC_TX_DIRECT_SUBMIT: when the lock is dropped submit directly
+ *
+ * while holding depend_tx->lock we must avoid submitting new operations
+ * to prevent a circular locking dependency with drivers that already
+ * hold a channel lock when calling async_tx_run_dependencies.
  */
 enum submit_disposition {
        ASYNC_TX_SUBMITTED,
@@ -160,11 +156,12 @@ enum submit_disposition {
 
 void
 async_tx_submit(struct dma_chan *chan, struct dma_async_tx_descriptor *tx,
-       enum async_tx_flags flags, struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_param)
+               struct async_submit_ctl *submit)
 {
-       tx->callback = cb_fn;
-       tx->callback_param = cb_param;
+       struct dma_async_tx_descriptor *depend_tx = submit->depend_tx;
+
+       tx->callback = submit->cb_fn;
+       tx->callback_param = submit->cb_param;
 
        if (depend_tx) {
                enum submit_disposition s;
@@ -220,30 +217,29 @@ async_tx_submit(struct dma_chan *chan, struct dma_async_tx_descriptor *tx,
                tx->tx_submit(tx);
        }
 
-       if (flags & ASYNC_TX_ACK)
+       if (submit->flags & ASYNC_TX_ACK)
                async_tx_ack(tx);
 
-       if (depend_tx && (flags & ASYNC_TX_DEP_ACK))
+       if (depend_tx)
                async_tx_ack(depend_tx);
 }
 EXPORT_SYMBOL_GPL(async_tx_submit);
 
 /**
- * async_trigger_callback - schedules the callback function to be run after
- * any dependent operations have been completed.
- * @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
- * @depend_tx: 'callback' requires the completion of this transaction
- * @cb_fn: function to call after depend_tx completes
- * @cb_param: parameter to pass to the callback routine
+ * async_trigger_callback - schedules the callback function to be run
+ * @submit: submission and completion parameters
+ *
+ * honored flags: ASYNC_TX_ACK
+ *
+ * The callback is run after any dependent operations have completed.
  */
 struct dma_async_tx_descriptor *
-async_trigger_callback(enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_param)
+async_trigger_callback(struct async_submit_ctl *submit)
 {
        struct dma_chan *chan;
        struct dma_device *device;
        struct dma_async_tx_descriptor *tx;
+       struct dma_async_tx_descriptor *depend_tx = submit->depend_tx;
 
        if (depend_tx) {
                chan = depend_tx->chan;
@@ -262,14 +258,14 @@ async_trigger_callback(enum async_tx_flags flags,
        if (tx) {
                pr_debug("%s: (async)\n", __func__);
 
-               async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
+               async_tx_submit(chan, tx, submit);
        } else {
                pr_debug("%s: (sync)\n", __func__);
 
                /* wait for any prerequisite operations */
-               async_tx_quiesce(&depend_tx);
+               async_tx_quiesce(&submit->depend_tx);
 
-               async_tx_sync_epilog(cb_fn, cb_param);
+               async_tx_sync_epilog(submit);
        }
 
        return tx;
@@ -295,9 +291,6 @@ void async_tx_quiesce(struct dma_async_tx_descriptor **tx)
 }
 EXPORT_SYMBOL_GPL(async_tx_quiesce);
 
-module_init(async_tx_init);
-module_exit(async_tx_exit);
-
 MODULE_AUTHOR("Intel Corporation");
 MODULE_DESCRIPTION("Asynchronous Bulk Memory Transactions API");
 MODULE_LICENSE("GPL");
index 95fe2c8..56b5f98 100644 (file)
 /* do_async_xor - dma map the pages and perform the xor with an engine */
 static __async_inline struct dma_async_tx_descriptor *
 do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list,
-            unsigned int offset, int src_cnt, size_t len,
-            enum async_tx_flags flags,
-            struct dma_async_tx_descriptor *depend_tx,
-            dma_async_tx_callback cb_fn, void *cb_param)
+            unsigned int offset, int src_cnt, size_t len, dma_addr_t *dma_src,
+            struct async_submit_ctl *submit)
 {
        struct dma_device *dma = chan->device;
-       dma_addr_t *dma_src = (dma_addr_t *) src_list;
        struct dma_async_tx_descriptor *tx = NULL;
        int src_off = 0;
        int i;
-       dma_async_tx_callback _cb_fn;
-       void *_cb_param;
-       enum async_tx_flags async_flags;
+       dma_async_tx_callback cb_fn_orig = submit->cb_fn;
+       void *cb_param_orig = submit->cb_param;
+       enum async_tx_flags flags_orig = submit->flags;
        enum dma_ctrl_flags dma_flags;
        int xor_src_cnt;
        dma_addr_t dma_dest;
@@ -63,23 +60,23 @@ do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list,
        }
 
        while (src_cnt) {
-               async_flags = flags;
+               submit->flags = flags_orig;
                dma_flags = 0;
-               xor_src_cnt = min(src_cnt, dma->max_xor);
+               xor_src_cnt = min(src_cnt, (int)dma->max_xor);
                /* if we are submitting additional xors, leave the chain open,
                 * clear the callback parameters, and leave the destination
                 * buffer mapped
                 */
                if (src_cnt > xor_src_cnt) {
-                       async_flags &= ~ASYNC_TX_ACK;
+                       submit->flags &= ~ASYNC_TX_ACK;
                        dma_flags = DMA_COMPL_SKIP_DEST_UNMAP;
-                       _cb_fn = NULL;
-                       _cb_param = NULL;
+                       submit->cb_fn = NULL;
+                       submit->cb_param = NULL;
                } else {
-                       _cb_fn = cb_fn;
-                       _cb_param = cb_param;
+                       submit->cb_fn = cb_fn_orig;
+                       submit->cb_param = cb_param_orig;
                }
-               if (_cb_fn)
+               if (submit->cb_fn)
                        dma_flags |= DMA_PREP_INTERRUPT;
 
                /* Since we have clobbered the src_list we are committed
@@ -90,7 +87,7 @@ do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list,
                                              xor_src_cnt, len, dma_flags);
 
                if (unlikely(!tx))
-                       async_tx_quiesce(&depend_tx);
+                       async_tx_quiesce(&submit->depend_tx);
 
                /* spin wait for the preceeding transactions to complete */
                while (unlikely(!tx)) {
@@ -101,11 +98,8 @@ do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list,
                                                      dma_flags);
                }
 
-               async_tx_submit(chan, tx, async_flags, depend_tx, _cb_fn,
-                               _cb_param);
-
-               depend_tx = tx;
-               flags |= ASYNC_TX_DEP_ACK;
+               async_tx_submit(chan, tx, submit);
+               submit->depend_tx = tx;
 
                if (src_cnt > xor_src_cnt) {
                        /* drop completed sources */
@@ -124,23 +118,27 @@ do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list,
 
 static void
 do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset,
-           int src_cnt, size_t len, enum async_tx_flags flags,
-           dma_async_tx_callback cb_fn, void *cb_param)
+           int src_cnt, size_t len, struct async_submit_ctl *submit)
 {
        int i;
        int xor_src_cnt;
        int src_off = 0;
        void *dest_buf;
-       void **srcs = (void **) src_list;
+       void **srcs;
+
+       if (submit->scribble)
+               srcs = submit->scribble;
+       else
+               srcs = (void **) src_list;
 
-       /* reuse the 'src_list' array to convert to buffer pointers */
+       /* convert to buffer pointers */
        for (i = 0; i < src_cnt; i++)
                srcs[i] = page_address(src_list[i]) + offset;
 
        /* set destination address */
        dest_buf = page_address(dest) + offset;
 
-       if (flags & ASYNC_TX_XOR_ZERO_DST)
+       if (submit->flags & ASYNC_TX_XOR_ZERO_DST)
                memset(dest_buf, 0, len);
 
        while (src_cnt > 0) {
@@ -153,61 +151,70 @@ do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset,
                src_off += xor_src_cnt;
        }
 
-       async_tx_sync_epilog(cb_fn, cb_param);
+       async_tx_sync_epilog(submit);
 }
 
 /**
  * async_xor - attempt to xor a set of blocks with a dma engine.
- *     xor_blocks always uses the dest as a source so the ASYNC_TX_XOR_ZERO_DST
- *     flag must be set to not include dest data in the calculation.  The
- *     assumption with dma eninges is that they only use the destination
- *     buffer as a source when it is explicity specified in the source list.
  * @dest: destination page
- * @src_list: array of source pages (if the dest is also a source it must be
- *     at index zero).  The contents of this array may be overwritten.
- * @offset: offset in pages to start transaction
+ * @src_list: array of source pages
+ * @offset: common src/dst offset to start transaction
  * @src_cnt: number of source pages
  * @len: length in bytes
- * @flags: ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DEST,
- *     ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
- * @depend_tx: xor depends on the result of this transaction.
- * @cb_fn: function to call when the xor completes
- * @cb_param: parameter to pass to the callback routine
+ * @submit: submission / completion modifiers
+ *
+ * honored flags: ASYNC_TX_ACK, ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DST
+ *
+ * xor_blocks always uses the dest as a source so the
+ * ASYNC_TX_XOR_ZERO_DST flag must be set to not include dest data in
+ * the calculation.  The assumption with dma eninges is that they only
+ * use the destination buffer as a source when it is explicity specified
+ * in the source list.
+ *
+ * src_list note: if the dest is also a source it must be at index zero.
+ * The contents of this array will be overwritten if a scribble region
+ * is not specified.
  */
 struct dma_async_tx_descriptor *
 async_xor(struct page *dest, struct page **src_list, unsigned int offset,
-       int src_cnt, size_t len, enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_param)
+         int src_cnt, size_t len, struct async_submit_ctl *submit)
 {
-       struct dma_chan *chan = async_tx_find_channel(depend_tx, DMA_XOR,
+       struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR,
                                                      &dest, 1, src_list,
                                                      src_cnt, len);
+       dma_addr_t *dma_src = NULL;
+
        BUG_ON(src_cnt <= 1);
 
-       if (chan) {
+       if (submit->scribble)
+               dma_src = submit->scribble;
+       else if (sizeof(dma_addr_t) <= sizeof(struct page *))
+               dma_src = (dma_addr_t *) src_list;
+
+       if (dma_src && chan) {
                /* run the xor asynchronously */
                pr_debug("%s (async): len: %zu\n", __func__, len);
 
                return do_async_xor(chan, dest, src_list, offset, src_cnt, len,
-                                   flags, depend_tx, cb_fn, cb_param);
+                                   dma_src, submit);
        } else {
                /* run the xor synchronously */
                pr_debug("%s (sync): len: %zu\n", __func__, len);
+               WARN_ONCE(chan, "%s: no space for dma address conversion\n",
+                         __func__);
 
                /* in the sync case the dest is an implied source
                 * (assumes the dest is the first source)
                 */
-               if (flags & ASYNC_TX_XOR_DROP_DST) {
+               if (submit->flags & ASYNC_TX_XOR_DROP_DST) {
                        src_cnt--;
                        src_list++;
                }
 
                /* wait for any prerequisite operations */
-               async_tx_quiesce(&depend_tx);
+               async_tx_quiesce(&submit->depend_tx);
 
-               do_sync_xor(dest, src_list, offset, src_cnt, len,
-                           flags, cb_fn, cb_param);
+               do_sync_xor(dest, src_list, offset, src_cnt, len, submit);
 
                return NULL;
        }
@@ -222,104 +229,90 @@ static int page_is_zero(struct page *p, unsigned int offset, size_t len)
 }
 
 /**
- * async_xor_zero_sum - attempt a xor parity check with a dma engine.
+ * async_xor_val - attempt a xor parity check with a dma engine.
  * @dest: destination page used if the xor is performed synchronously
- * @src_list: array of source pages.  The dest page must be listed as a source
- *     at index zero.  The contents of this array may be overwritten.
+ * @src_list: array of source pages
  * @offset: offset in pages to start transaction
  * @src_cnt: number of source pages
  * @len: length in bytes
  * @result: 0 if sum == 0 else non-zero
- * @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
- * @depend_tx: xor depends on the result of this transaction.
- * @cb_fn: function to call when the xor completes
- * @cb_param: parameter to pass to the callback routine
+ * @submit: submission / completion modifiers
+ *
+ * honored flags: ASYNC_TX_ACK
+ *
+ * src_list note: if the dest is also a source it must be at index zero.
+ * The contents of this array will be overwritten if a scribble region
+ * is not specified.
  */
 struct dma_async_tx_descriptor *
-async_xor_zero_sum(struct page *dest, struct page **src_list,
-       unsigned int offset, int src_cnt, size_t len,
-       u32 *result, enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_param)
+async_xor_val(struct page *dest, struct page **src_list, unsigned int offset,
+             int src_cnt, size_t len, enum sum_check_flags *result,
+             struct async_submit_ctl *submit)
 {
-       struct dma_chan *chan = async_tx_find_channel(depend_tx, DMA_ZERO_SUM,
+       struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR_VAL,
                                                      &dest, 1, src_list,
                                                      src_cnt, len);
        struct dma_device *device = chan ? chan->device : NULL;
        struct dma_async_tx_descriptor *tx = NULL;
+       dma_addr_t *dma_src = NULL;
 
        BUG_ON(src_cnt <= 1);
 
-       if (device && src_cnt <= device->max_xor) {
-               dma_addr_t *dma_src = (dma_addr_t *) src_list;
-               unsigned long dma_prep_flags = cb_fn ? DMA_PREP_INTERRUPT : 0;
+       if (submit->scribble)
+               dma_src = submit->scribble;
+       else if (sizeof(dma_addr_t) <= sizeof(struct page *))
+               dma_src = (dma_addr_t *) src_list;
+
+       if (dma_src && device && src_cnt <= device->max_xor) {
+               unsigned long dma_prep_flags;
                int i;
 
                pr_debug("%s: (async) len: %zu\n", __func__, len);
 
+               dma_prep_flags = submit->cb_fn ? DMA_PREP_INTERRUPT : 0;
                for (i = 0; i < src_cnt; i++)
                        dma_src[i] = dma_map_page(device->dev, src_list[i],
                                                  offset, len, DMA_TO_DEVICE);
 
-               tx = device->device_prep_dma_zero_sum(chan, dma_src, src_cnt,
-                                                     len, result,
-                                                     dma_prep_flags);
+               tx = device->device_prep_dma_xor_val(chan, dma_src, src_cnt,
+                                                    len, result,
+                                                    dma_prep_flags);
                if (unlikely(!tx)) {
-                       async_tx_quiesce(&depend_tx);
+                       async_tx_quiesce(&submit->depend_tx);
 
                        while (!tx) {
                                dma_async_issue_pending(chan);
-                               tx = device->device_prep_dma_zero_sum(chan,
+                               tx = device->device_prep_dma_xor_val(chan,
                                        dma_src, src_cnt, len, result,
                                        dma_prep_flags);
                        }
                }
 
-               async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
+               async_tx_submit(chan, tx, submit);
        } else {
-               unsigned long xor_flags = flags;
+               enum async_tx_flags flags_orig = submit->flags;
 
                pr_debug("%s: (sync) len: %zu\n", __func__, len);
+               WARN_ONCE(device && src_cnt <= device->max_xor,
+                         "%s: no space for dma address conversion\n",
+                         __func__);
 
-               xor_flags |= ASYNC_TX_XOR_DROP_DST;
-               xor_flags &= ~ASYNC_TX_ACK;
+               submit->flags |= ASYNC_TX_XOR_DROP_DST;
+               submit->flags &= ~ASYNC_TX_ACK;
 
-               tx = async_xor(dest, src_list, offset, src_cnt, len, xor_flags,
-                       depend_tx, NULL, NULL);
+               tx = async_xor(dest, src_list, offset, src_cnt, len, submit);
 
                async_tx_quiesce(&tx);
 
-               *result = page_is_zero(dest, offset, len) ? 0 : 1;
+               *result = !page_is_zero(dest, offset, len) << SUM_CHECK_P;
 
-               async_tx_sync_epilog(cb_fn, cb_param);
+               async_tx_sync_epilog(submit);
+               submit->flags = flags_orig;
        }
 
        return tx;
 }
-EXPORT_SYMBOL_GPL(async_xor_zero_sum);
-
-static int __init async_xor_init(void)
-{
-       #ifdef CONFIG_DMA_ENGINE
-       /* To conserve stack space the input src_list (array of page pointers)
-        * is reused to hold the array of dma addresses passed to the driver.
-        * This conversion is only possible when dma_addr_t is less than the
-        * the size of a pointer.  HIGHMEM64G is known to violate this
-        * assumption.
-        */
-       BUILD_BUG_ON(sizeof(dma_addr_t) > sizeof(struct page *));
-       #endif
-
-       return 0;
-}
-
-static void __exit async_xor_exit(void)
-{
-       do { } while (0);
-}
-
-module_init(async_xor_init);
-module_exit(async_xor_exit);
+EXPORT_SYMBOL_GPL(async_xor_val);
 
 MODULE_AUTHOR("Intel Corporation");
 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api");
diff --git a/crypto/async_tx/raid6test.c b/crypto/async_tx/raid6test.c
new file mode 100644 (file)
index 0000000..98c83ca
--- /dev/null
@@ -0,0 +1,241 @@
+/*
+ * asynchronous raid6 recovery self test
+ * Copyright (c) 2009, Intel Corporation.
+ *
+ * based on drivers/md/raid6test/test.c:
+ *     Copyright 2002-2007 H. Peter Anvin
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ */
+#include <linux/async_tx.h>
+#include <linux/random.h>
+
+#undef pr
+#define pr(fmt, args...) pr_info("raid6test: " fmt, ##args)
+
+#define NDISKS 16 /* Including P and Q */
+
+static struct page *dataptrs[NDISKS];
+static struct page *data[NDISKS+3];
+static struct page *spare;
+static struct page *recovi;
+static struct page *recovj;
+
+static void callback(void *param)
+{
+       struct completion *cmp = param;
+
+       complete(cmp);
+}
+
+static void makedata(int disks)
+{
+       int i, j;
+
+       for (i = 0; i < disks; i++) {
+               for (j = 0; j < PAGE_SIZE/sizeof(u32); j += sizeof(u32)) {
+                       u32 *p = page_address(data[i]) + j;
+
+                       *p = random32();
+               }
+
+               dataptrs[i] = data[i];
+       }
+}
+
+static char disk_type(int d, int disks)
+{
+       if (d == disks - 2)
+               return 'P';
+       else if (d == disks - 1)
+               return 'Q';
+       else
+               return 'D';
+}
+
+/* Recover two failed blocks. */
+static void raid6_dual_recov(int disks, size_t bytes, int faila, int failb, struct page **ptrs)
+{
+       struct async_submit_ctl submit;
+       addr_conv_t addr_conv[disks];
+       struct completion cmp;
+       struct dma_async_tx_descriptor *tx = NULL;
+       enum sum_check_flags result = ~0;
+
+       if (faila > failb)
+               swap(faila, failb);
+
+       if (failb == disks-1) {
+               if (faila == disks-2) {
+                       /* P+Q failure.  Just rebuild the syndrome. */
+                       init_async_submit(&submit, 0, NULL, NULL, NULL, addr_conv);
+                       tx = async_gen_syndrome(ptrs, 0, disks, bytes, &submit);
+               } else {
+                       struct page *blocks[disks];
+                       struct page *dest;
+                       int count = 0;
+                       int i;
+
+                       /* data+Q failure.  Reconstruct data from P,
+                        * then rebuild syndrome
+                        */
+                       for (i = disks; i-- ; ) {
+                               if (i == faila || i == failb)
+                                       continue;
+                               blocks[count++] = ptrs[i];
+                       }
+                       dest = ptrs[faila];
+                       init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
+                                         NULL, NULL, addr_conv);
+                       tx = async_xor(dest, blocks, 0, count, bytes, &submit);
+
+                       init_async_submit(&submit, 0, tx, NULL, NULL, addr_conv);
+                       tx = async_gen_syndrome(ptrs, 0, disks, bytes, &submit);
+               }
+       } else {
+               if (failb == disks-2) {
+                       /* data+P failure. */
+                       init_async_submit(&submit, 0, NULL, NULL, NULL, addr_conv);
+                       tx = async_raid6_datap_recov(disks, bytes, faila, ptrs, &submit);
+               } else {
+                       /* data+data failure. */
+                       init_async_submit(&submit, 0, NULL, NULL, NULL, addr_conv);
+                       tx = async_raid6_2data_recov(disks, bytes, faila, failb, ptrs, &submit);
+               }
+       }
+       init_completion(&cmp);
+       init_async_submit(&submit, ASYNC_TX_ACK, tx, callback, &cmp, addr_conv);
+       tx = async_syndrome_val(ptrs, 0, disks, bytes, &result, spare, &submit);
+       async_tx_issue_pending(tx);
+
+       if (wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000)) == 0)
+               pr("%s: timeout! (faila: %d failb: %d disks: %d)\n",
+                  __func__, faila, failb, disks);
+
+       if (result != 0)
+               pr("%s: validation failure! faila: %d failb: %d sum_check_flags: %x\n",
+                  __func__, faila, failb, result);
+}
+
+static int test_disks(int i, int j, int disks)
+{
+       int erra, errb;
+
+       memset(page_address(recovi), 0xf0, PAGE_SIZE);
+       memset(page_address(recovj), 0xba, PAGE_SIZE);
+
+       dataptrs[i] = recovi;
+       dataptrs[j] = recovj;
+
+       raid6_dual_recov(disks, PAGE_SIZE, i, j, dataptrs);
+
+       erra = memcmp(page_address(data[i]), page_address(recovi), PAGE_SIZE);
+       errb = memcmp(page_address(data[j]), page_address(recovj), PAGE_SIZE);
+
+       pr("%s(%d, %d): faila=%3d(%c)  failb=%3d(%c)  %s\n",
+          __func__, i, j, i, disk_type(i, disks), j, disk_type(j, disks),
+          (!erra && !errb) ? "OK" : !erra ? "ERRB" : !errb ? "ERRA" : "ERRAB");
+
+       dataptrs[i] = data[i];
+       dataptrs[j] = data[j];
+
+       return erra || errb;
+}
+
+static int test(int disks, int *tests)
+{
+       addr_conv_t addr_conv[disks];
+       struct dma_async_tx_descriptor *tx;
+       struct async_submit_ctl submit;
+       struct completion cmp;
+       int err = 0;
+       int i, j;
+
+       recovi = data[disks];
+       recovj = data[disks+1];
+       spare  = data[disks+2];
+
+       makedata(disks);
+
+       /* Nuke syndromes */
+       memset(page_address(data[disks-2]), 0xee, PAGE_SIZE);
+       memset(page_address(data[disks-1]), 0xee, PAGE_SIZE);
+
+       /* Generate assumed good syndrome */
+       init_completion(&cmp);
+       init_async_submit(&submit, ASYNC_TX_ACK, NULL, callback, &cmp, addr_conv);
+       tx = async_gen_syndrome(dataptrs, 0, disks, PAGE_SIZE, &submit);
+       async_tx_issue_pending(tx);
+
+       if (wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000)) == 0) {
+               pr("error: initial gen_syndrome(%d) timed out\n", disks);
+               return 1;
+       }
+
+       pr("testing the %d-disk case...\n", disks);
+       for (i = 0; i < disks-1; i++)
+               for (j = i+1; j < disks; j++) {
+                       (*tests)++;
+                       err += test_disks(i, j, disks);
+               }
+
+       return err;
+}
+
+
+static int raid6_test(void)
+{
+       int err = 0;
+       int tests = 0;
+       int i;
+
+       for (i = 0; i < NDISKS+3; i++) {
+               data[i] = alloc_page(GFP_KERNEL);
+               if (!data[i]) {
+                       while (i--)
+                               put_page(data[i]);
+                       return -ENOMEM;
+               }
+       }
+
+       /* the 4-disk and 5-disk cases are special for the recovery code */
+       if (NDISKS > 4)
+               err += test(4, &tests);
+       if (NDISKS > 5)
+               err += test(5, &tests);
+       err += test(NDISKS, &tests);
+
+       pr("\n");
+       pr("complete (%d tests, %d failure%s)\n",
+          tests, err, err == 1 ? "" : "s");
+
+       for (i = 0; i < NDISKS+3; i++)
+               put_page(data[i]);
+
+       return 0;
+}
+
+static void raid6_test_exit(void)
+{
+}
+
+/* when compiled-in wait for drivers to load first (assumes dma drivers
+ * are also compliled-in)
+ */
+late_initcall(raid6_test);
+module_exit(raid6_test_exit);
+MODULE_AUTHOR("Dan Williams <dan.j.williams@intel.com>");
+MODULE_DESCRIPTION("asynchronous RAID-6 recovery self tests");
+MODULE_LICENSE("GPL");
index 3b3c01b..912a51b 100644 (file)
@@ -4,7 +4,7 @@
 
 menuconfig DMADEVICES
        bool "DMA Engine support"
-       depends on !HIGHMEM64G && HAS_DMA
+       depends on HAS_DMA
        help
          DMA engines can do asynchronous data transfers without
          involving the host CPU.  Currently, this framework can be
index 5a87384..9659847 100644 (file)
@@ -644,8 +644,12 @@ int dma_async_device_register(struct dma_device *device)
                !device->device_prep_dma_memcpy);
        BUG_ON(dma_has_cap(DMA_XOR, device->cap_mask) &&
                !device->device_prep_dma_xor);
-       BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) &&
-               !device->device_prep_dma_zero_sum);
+       BUG_ON(dma_has_cap(DMA_XOR_VAL, device->cap_mask) &&
+               !device->device_prep_dma_xor_val);
+       BUG_ON(dma_has_cap(DMA_PQ, device->cap_mask) &&
+               !device->device_prep_dma_pq);
+       BUG_ON(dma_has_cap(DMA_PQ_VAL, device->cap_mask) &&
+               !device->device_prep_dma_pq_val);
        BUG_ON(dma_has_cap(DMA_MEMSET, device->cap_mask) &&
                !device->device_prep_dma_memset);
        BUG_ON(dma_has_cap(DMA_INTERRUPT, device->cap_mask) &&
@@ -939,49 +943,24 @@ EXPORT_SYMBOL(dma_async_tx_descriptor_init);
 
 /* dma_wait_for_async_tx - spin wait for a transaction to complete
  * @tx: in-flight transaction to wait on
- *
- * This routine assumes that tx was obtained from a call to async_memcpy,
- * async_xor, async_memset, etc which ensures that tx is "in-flight" (prepped
- * and submitted).  Walking the parent chain is only meant to cover for DMA
- * drivers that do not implement the DMA_INTERRUPT capability and may race with
- * the driver's descriptor cleanup routine.
  */
 enum dma_status
 dma_wait_for_async_tx(struct dma_async_tx_descriptor *tx)
 {
-       enum dma_status status;
-       struct dma_async_tx_descriptor *iter;
-       struct dma_async_tx_descriptor *parent;
+       unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
 
        if (!tx)
                return DMA_SUCCESS;
 
-       WARN_ONCE(tx->parent, "%s: speculatively walking dependency chain for"
-                 " %s\n", __func__, dma_chan_name(tx->chan));
-
-       /* poll through the dependency chain, return when tx is complete */
-       do {
-               iter = tx;
-
-               /* find the root of the unsubmitted dependency chain */
-               do {
-                       parent = iter->parent;
-                       if (!parent)
-                               break;
-                       else
-                               iter = parent;
-               } while (parent);
-
-               /* there is a small window for ->parent == NULL and
-                * ->cookie == -EBUSY
-                */
-               while (iter->cookie == -EBUSY)
-                       cpu_relax();
-
-               status = dma_sync_wait(iter->chan, iter->cookie);
-       } while (status == DMA_IN_PROGRESS || (iter != tx));
-
-       return status;
+       while (tx->cookie == -EBUSY) {
+               if (time_after_eq(jiffies, dma_sync_wait_timeout)) {
+                       pr_err("%s timeout waiting for descriptor submission\n",
+                               __func__);
+                       return DMA_ERROR;
+               }
+               cpu_relax();
+       }
+       return dma_sync_wait(tx->chan, tx->cookie);
 }
 EXPORT_SYMBOL_GPL(dma_wait_for_async_tx);
 
index fb7da51..58e49e4 100644 (file)
@@ -43,6 +43,11 @@ module_param(xor_sources, uint, S_IRUGO);
 MODULE_PARM_DESC(xor_sources,
                "Number of xor source buffers (default: 3)");
 
+static unsigned int pq_sources = 3;
+module_param(pq_sources, uint, S_IRUGO);
+MODULE_PARM_DESC(pq_sources,
+               "Number of p+q source buffers (default: 3)");
+
 /*
  * Initialization patterns. All bytes in the source buffer has bit 7
  * set, all bytes in the destination buffer has bit 7 cleared.
@@ -227,6 +232,7 @@ static int dmatest_func(void *data)
        dma_cookie_t            cookie;
        enum dma_status         status;
        enum dma_ctrl_flags     flags;
+       u8                      pq_coefs[pq_sources];
        int                     ret;
        int                     src_cnt;
        int                     dst_cnt;
@@ -243,6 +249,11 @@ static int dmatest_func(void *data)
        else if (thread->type == DMA_XOR) {
                src_cnt = xor_sources | 1; /* force odd to ensure dst = src */
                dst_cnt = 1;
+       } else if (thread->type == DMA_PQ) {
+               src_cnt = pq_sources | 1; /* force odd to ensure dst = src */
+               dst_cnt = 2;
+               for (i = 0; i < pq_sources; i++)
+                       pq_coefs[i] = 1;
        } else
                goto err_srcs;
 
@@ -310,6 +321,15 @@ static int dmatest_func(void *data)
                                                      dma_dsts[0] + dst_off,
                                                      dma_srcs, xor_sources,
                                                      len, flags);
+               else if (thread->type == DMA_PQ) {
+                       dma_addr_t dma_pq[dst_cnt];
+
+                       for (i = 0; i < dst_cnt; i++)
+                               dma_pq[i] = dma_dsts[i] + dst_off;
+                       tx = dev->device_prep_dma_pq(chan, dma_pq, dma_srcs,
+                                                    pq_sources, pq_coefs,
+                                                    len, flags);
+               }
 
                if (!tx) {
                        for (i = 0; i < src_cnt; i++)
@@ -446,6 +466,8 @@ static int dmatest_add_threads(struct dmatest_chan *dtc, enum dma_transaction_ty
                op = "copy";
        else if (type == DMA_XOR)
                op = "xor";
+       else if (type == DMA_PQ)
+               op = "pq";
        else
                return -EINVAL;
 
@@ -501,6 +523,10 @@ static int dmatest_add_channel(struct dma_chan *chan)
                cnt = dmatest_add_threads(dtc, DMA_XOR);
                thread_count += cnt > 0 ?: 0;
        }
+       if (dma_has_cap(DMA_PQ, dma_dev->cap_mask)) {
+               cnt = dmatest_add_threads(dtc, DMA_PQ);
+               thread_count += cnt > 0 ?: 0;
+       }
 
        pr_info("dmatest: Started %u threads using %s\n",
                thread_count, dma_chan_name(chan));
index 2f05226..4496bc6 100644 (file)
@@ -660,9 +660,9 @@ iop_adma_prep_dma_xor(struct dma_chan *chan, dma_addr_t dma_dest,
 }
 
 static struct dma_async_tx_descriptor *
-iop_adma_prep_dma_zero_sum(struct dma_chan *chan, dma_addr_t *dma_src,
-                          unsigned int src_cnt, size_t len, u32 *result,
-                          unsigned long flags)
+iop_adma_prep_dma_xor_val(struct dma_chan *chan, dma_addr_t *dma_src,
+                         unsigned int src_cnt, size_t len, u32 *result,
+                         unsigned long flags)
 {
        struct iop_adma_chan *iop_chan = to_iop_adma_chan(chan);
        struct iop_adma_desc_slot *sw_desc, *grp_start;
@@ -906,7 +906,7 @@ out:
 
 #define IOP_ADMA_NUM_SRC_TEST 4 /* must be <= 15 */
 static int __devinit
-iop_adma_xor_zero_sum_self_test(struct iop_adma_device *device)
+iop_adma_xor_val_self_test(struct iop_adma_device *device)
 {
        int i, src_idx;
        struct page *dest;
@@ -1002,7 +1002,7 @@ iop_adma_xor_zero_sum_self_test(struct iop_adma_device *device)
                PAGE_SIZE, DMA_TO_DEVICE);
 
        /* skip zero sum if the capability is not present */
-       if (!dma_has_cap(DMA_ZERO_SUM, dma_chan->device->cap_mask))
+       if (!dma_has_cap(DMA_XOR_VAL, dma_chan->device->cap_mask))
                goto free_resources;
 
        /* zero sum the sources with the destintation page */
@@ -1016,10 +1016,10 @@ iop_adma_xor_zero_sum_self_test(struct iop_adma_device *device)
                dma_srcs[i] = dma_map_page(dma_chan->device->dev,
                                           zero_sum_srcs[i], 0, PAGE_SIZE,
                                           DMA_TO_DEVICE);
-       tx = iop_adma_prep_dma_zero_sum(dma_chan, dma_srcs,
-                                       IOP_ADMA_NUM_SRC_TEST + 1, PAGE_SIZE,
-                                       &zero_sum_result,
-                                       DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
+       tx = iop_adma_prep_dma_xor_val(dma_chan, dma_srcs,
+                                      IOP_ADMA_NUM_SRC_TEST + 1, PAGE_SIZE,
+                                      &zero_sum_result,
+                                      DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
 
        cookie = iop_adma_tx_submit(tx);
        iop_adma_issue_pending(dma_chan);
@@ -1072,10 +1072,10 @@ iop_adma_xor_zero_sum_self_test(struct iop_adma_device *device)
                dma_srcs[i] = dma_map_page(dma_chan->device->dev,
                                           zero_sum_srcs[i], 0, PAGE_SIZE,
                                           DMA_TO_DEVICE);
-       tx = iop_adma_prep_dma_zero_sum(dma_chan, dma_srcs,
-                                       IOP_ADMA_NUM_SRC_TEST + 1, PAGE_SIZE,
-                                       &zero_sum_result,
-                                       DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
+       tx = iop_adma_prep_dma_xor_val(dma_chan, dma_srcs,
+                                      IOP_ADMA_NUM_SRC_TEST + 1, PAGE_SIZE,
+                                      &zero_sum_result,
+                                      DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
 
        cookie = iop_adma_tx_submit(tx);
        iop_adma_issue_pending(dma_chan);
@@ -1192,9 +1192,9 @@ static int __devinit iop_adma_probe(struct platform_device *pdev)
                dma_dev->max_xor = iop_adma_get_max_xor();
                dma_dev->device_prep_dma_xor = iop_adma_prep_dma_xor;
        }
-       if (dma_has_cap(DMA_ZERO_SUM, dma_dev->cap_mask))
-               dma_dev->device_prep_dma_zero_sum =
-                       iop_adma_prep_dma_zero_sum;
+       if (dma_has_cap(DMA_XOR_VAL, dma_dev->cap_mask))
+               dma_dev->device_prep_dma_xor_val =
+                       iop_adma_prep_dma_xor_val;
        if (dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask))
                dma_dev->device_prep_dma_interrupt =
                        iop_adma_prep_dma_interrupt;
@@ -1249,7 +1249,7 @@ static int __devinit iop_adma_probe(struct platform_device *pdev)
 
        if (dma_has_cap(DMA_XOR, dma_dev->cap_mask) ||
                dma_has_cap(DMA_MEMSET, dma_dev->cap_mask)) {
-               ret = iop_adma_xor_zero_sum_self_test(adev);
+               ret = iop_adma_xor_val_self_test(adev);
                dev_dbg(&pdev->dev, "xor self test returned %d\n", ret);
                if (ret)
                        goto err_free_iop_chan;
@@ -1257,12 +1257,12 @@ static int __devinit iop_adma_probe(struct platform_device *pdev)
 
        dev_printk(KERN_INFO, &pdev->dev, "Intel(R) IOP: "
          "( %s%s%s%s%s%s%s%s%s%s)\n",
-         dma_has_cap(DMA_PQ_XOR, dma_dev->cap_mask) ? "pq_xor " : "",
+         dma_has_cap(DMA_PQ, dma_dev->cap_mask) ? "pq " : "",
          dma_has_cap(DMA_PQ_UPDATE, dma_dev->cap_mask) ? "pq_update " : "",
-         dma_has_cap(DMA_PQ_ZERO_SUM, dma_dev->cap_mask) ? "pq_zero_sum " : "",
+         dma_has_cap(DMA_PQ_VAL, dma_dev->cap_mask) ? "pq_val " : "",
          dma_has_cap(DMA_XOR, dma_dev->cap_mask) ? "xor " : "",
          dma_has_cap(DMA_DUAL_XOR, dma_dev->cap_mask) ? "dual_xor " : "",
-         dma_has_cap(DMA_ZERO_SUM, dma_dev->cap_mask) ? "xor_zero_sum " : "",
+         dma_has_cap(DMA_XOR_VAL, dma_dev->cap_mask) ? "xor_val " : "",
          dma_has_cap(DMA_MEMSET, dma_dev->cap_mask)  ? "fill " : "",
          dma_has_cap(DMA_MEMCPY_CRC32C, dma_dev->cap_mask) ? "cpy+crc " : "",
          dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask) ? "cpy " : "",
index 36e0675..09c0c6e 100644 (file)
@@ -124,6 +124,8 @@ config MD_RAID456
        select MD_RAID6_PQ
        select ASYNC_MEMCPY
        select ASYNC_XOR
+       select ASYNC_PQ
+       select ASYNC_RAID6_RECOV
        ---help---
          A RAID-5 set of N drives with a capacity of C MB per drive provides
          the capacity of C * (N - 1) MB, and protects against a failure
@@ -152,9 +154,33 @@ config MD_RAID456
 
          If unsure, say Y.
 
+config MULTICORE_RAID456
+       bool "RAID-4/RAID-5/RAID-6 Multicore processing (EXPERIMENTAL)"
+       depends on MD_RAID456
+       depends on SMP
+       depends on EXPERIMENTAL
+       ---help---
+         Enable the raid456 module to dispatch per-stripe raid operations to a
+         thread pool.
+
+         If unsure, say N.
+
 config MD_RAID6_PQ
        tristate
 
+config ASYNC_RAID6_TEST
+       tristate "Self test for hardware accelerated raid6 recovery"
+       depends on MD_RAID6_PQ
+       select ASYNC_RAID6_RECOV
+       ---help---
+         This is a one-shot self test that permutes through the
+         recovery of all the possible two disk failure scenarios for a
+         N-disk array.  Recovery is performed with the asynchronous
+         raid6 recovery routines, and will optionally use an offload
+         engine if one is available.
+
+         If unsure, say N.
+
 config MD_MULTIPATH
        tristate "Multipath I/O support"
        depends on BLK_DEV_MD
index bb37fb1..0a5cf21 100644 (file)
@@ -47,7 +47,9 @@
 #include <linux/kthread.h>
 #include <linux/raid/pq.h>
 #include <linux/async_tx.h>
+#include <linux/async.h>
 #include <linux/seq_file.h>
+#include <linux/cpu.h>
 #include "md.h"
 #include "raid5.h"
 #include "bitmap.h"
@@ -499,11 +501,14 @@ async_copy_data(int frombio, struct bio *bio, struct page *page,
        struct page *bio_page;
        int i;
        int page_offset;
+       struct async_submit_ctl submit;
 
        if (bio->bi_sector >= sector)
                page_offset = (signed)(bio->bi_sector - sector) * 512;
        else
                page_offset = (signed)(sector - bio->bi_sector) * -512;
+
+       init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
        bio_for_each_segment(bvl, bio, i) {
                int len = bio_iovec_idx(bio, i)->bv_len;
                int clen;
@@ -525,15 +530,14 @@ async_copy_data(int frombio, struct bio *bio, struct page *page,
                        bio_page = bio_iovec_idx(bio, i)->bv_page;
                        if (frombio)
                                tx = async_memcpy(page, bio_page, page_offset,
-                                       b_offset, clen,
-                                       ASYNC_TX_DEP_ACK,
-                                       tx, NULL, NULL);
+                                                 b_offset, clen, &submit);
                        else
                                tx = async_memcpy(bio_page, page, b_offset,
-                                       page_offset, clen,
-                                       ASYNC_TX_DEP_ACK,
-                                       tx, NULL, NULL);
+                                                 page_offset, clen, &submit);
                }
+               /* chain the operations */
+               submit.depend_tx = tx;
+
                if (clen < len) /* hit end of page */
                        break;
                page_offset +=  len;
@@ -592,6 +596,7 @@ static void ops_run_biofill(struct stripe_head *sh)
 {
        struct dma_async_tx_descriptor *tx = NULL;
        raid5_conf_t *conf = sh->raid_conf;
+       struct async_submit_ctl submit;
        int i;
 
        pr_debug("%s: stripe %llu\n", __func__,
@@ -615,22 +620,34 @@ static void ops_run_biofill(struct stripe_head *sh)
        }
 
        atomic_inc(&sh->count);
-       async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
-               ops_complete_biofill, sh);
+       init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
+       async_trigger_callback(&submit);
 }
 
-static void ops_complete_compute5(void *stripe_head_ref)
+static void mark_target_uptodate(struct stripe_head *sh, int target)
 {
-       struct stripe_head *sh = stripe_head_ref;
-       int target = sh->ops.target;
-       struct r5dev *tgt = &sh->dev[target];
+       struct r5dev *tgt;
 
-       pr_debug("%s: stripe %llu\n", __func__,
-               (unsigned long long)sh->sector);
+       if (target < 0)
+               return;
 
+       tgt = &sh->dev[target];
        set_bit(R5_UPTODATE, &tgt->flags);
        BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
        clear_bit(R5_Wantcompute, &tgt->flags);
+}
+
+static void ops_complete_compute(void *stripe_head_ref)
+{
+       struct stripe_head *sh = stripe_head_ref;
+
+       pr_debug("%s: stripe %llu\n", __func__,
+               (unsigned long long)sh->sector);
+
+       /* mark the computed target(s) as uptodate */
+       mark_target_uptodate(sh, sh->ops.target);
+       mark_target_uptodate(sh, sh->ops.target2);
+
        clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
        if (sh->check_state == check_state_compute_run)
                sh->check_state = check_state_compute_result;
@@ -638,16 +655,24 @@ static void ops_complete_compute5(void *stripe_head_ref)
        release_stripe(sh);
 }
 
-static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
+/* return a pointer to the address conversion region of the scribble buffer */
+static addr_conv_t *to_addr_conv(struct stripe_head *sh,
+                                struct raid5_percpu *percpu)
+{
+       return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
+}
+
+static struct dma_async_tx_descriptor *
+ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
 {
-       /* kernel stack size limits the total number of disks */
        int disks = sh->disks;
-       struct page *xor_srcs[disks];
+       struct page **xor_srcs = percpu->scribble;
        int target = sh->ops.target;
        struct r5dev *tgt = &sh->dev[target];
        struct page *xor_dest = tgt->page;
        int count = 0;
        struct dma_async_tx_descriptor *tx;
+       struct async_submit_ctl submit;
        int i;
 
        pr_debug("%s: stripe %llu block: %d\n",
@@ -660,17 +685,207 @@ static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
 
        atomic_inc(&sh->count);
 
+       init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
+                         ops_complete_compute, sh, to_addr_conv(sh, percpu));
        if (unlikely(count == 1))
-               tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
-                       0, NULL, ops_complete_compute5, sh);
+               tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
        else
-               tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
-                       ASYNC_TX_XOR_ZERO_DST, NULL,
-                       ops_complete_compute5, sh);
+               tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 
        return tx;
 }
 
+/* set_syndrome_sources - populate source buffers for gen_syndrome
+ * @srcs - (struct page *) array of size sh->disks
+ * @sh - stripe_head to parse
+ *
+ * Populates srcs in proper layout order for the stripe and returns the
+ * 'count' of sources to be used in a call to async_gen_syndrome.  The P
+ * destination buffer is recorded in srcs[count] and the Q destination
+ * is recorded in srcs[count+1]].
+ */
+static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
+{
+       int disks = sh->disks;
+       int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
+       int d0_idx = raid6_d0(sh);
+       int count;
+       int i;
+
+       for (i = 0; i < disks; i++)
+               srcs[i] = (void *)raid6_empty_zero_page;
+
+       count = 0;
+       i = d0_idx;
+       do {
+               int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
+
+               srcs[slot] = sh->dev[i].page;
+               i = raid6_next_disk(i, disks);
+       } while (i != d0_idx);
+       BUG_ON(count != syndrome_disks);
+
+       return count;
+}
+
+static struct dma_async_tx_descriptor *
+ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
+{
+       int disks = sh->disks;
+       struct page **blocks = percpu->scribble;
+       int target;
+       int qd_idx = sh->qd_idx;
+       struct dma_async_tx_descriptor *tx;
+       struct async_submit_ctl submit;
+       struct r5dev *tgt;
+       struct page *dest;
+       int i;
+       int count;
+
+       if (sh->ops.target < 0)
+               target = sh->ops.target2;
+       else if (sh->ops.target2 < 0)
+               target = sh->ops.target;
+       else
+               /* we should only have one valid target */
+               BUG();
+       BUG_ON(target < 0);
+       pr_debug("%s: stripe %llu block: %d\n",
+               __func__, (unsigned long long)sh->sector, target);
+
+       tgt = &sh->dev[target];
+       BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
+       dest = tgt->page;
+
+       atomic_inc(&sh->count);
+
+       if (target == qd_idx) {
+               count = set_syndrome_sources(blocks, sh);
+               blocks[count] = NULL; /* regenerating p is not necessary */
+               BUG_ON(blocks[count+1] != dest); /* q should already be set */
+               init_async_submit(&submit, 0, NULL, ops_complete_compute, sh,
+                                 to_addr_conv(sh, percpu));
+               tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
+       } else {
+               /* Compute any data- or p-drive using XOR */
+               count = 0;
+               for (i = disks; i-- ; ) {
+                       if (i == target || i == qd_idx)
+                               continue;
+                       blocks[count++] = sh->dev[i].page;
+               }
+
+               init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
+                                 ops_complete_compute, sh,
+                                 to_addr_conv(sh, percpu));
+               tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
+       }
+
+       return tx;
+}
+
+static struct dma_async_tx_descriptor *
+ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
+{
+       int i, count, disks = sh->disks;
+       int syndrome_disks = sh->ddf_layout ? disks : disks-2;
+       int d0_idx = raid6_d0(sh);
+       int faila = -1, failb = -1;
+       int target = sh->ops.target;
+       int target2 = sh->ops.target2;
+       struct r5dev *tgt = &sh->dev[target];
+       struct r5dev *tgt2 = &sh->dev[target2];
+       struct dma_async_tx_descriptor *tx;
+       struct page **blocks = percpu->scribble;
+       struct async_submit_ctl submit;
+
+       pr_debug("%s: stripe %llu block1: %d block2: %d\n",
+                __func__, (unsigned long long)sh->sector, target, target2);
+       BUG_ON(target < 0 || target2 < 0);
+       BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
+       BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
+
+       /* we need to open-code set_syndrome_sources to handle to the
+        * slot number conversion for 'faila' and 'failb'
+        */
+       for (i = 0; i < disks ; i++)
+               blocks[i] = (void *)raid6_empty_zero_page;
+       count = 0;
+       i = d0_idx;
+       do {
+               int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
+
+               blocks[slot] = sh->dev[i].page;
+
+               if (i == target)
+                       faila = slot;
+               if (i == target2)
+                       failb = slot;
+               i = raid6_next_disk(i, disks);
+       } while (i != d0_idx);
+       BUG_ON(count != syndrome_disks);
+
+       BUG_ON(faila == failb);
+       if (failb < faila)
+               swap(faila, failb);
+       pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
+                __func__, (unsigned long long)sh->sector, faila, failb);
+
+       atomic_inc(&sh->count);
+
+       if (failb == syndrome_disks+1) {
+               /* Q disk is one of the missing disks */
+               if (faila == syndrome_disks) {
+                       /* Missing P+Q, just recompute */
+                       init_async_submit(&submit, 0, NULL, ops_complete_compute,
+                                         sh, to_addr_conv(sh, percpu));
+                       return async_gen_syndrome(blocks, 0, count+2,
+                                                 STRIPE_SIZE, &submit);
+               } else {
+                       struct page *dest;
+                       int data_target;
+                       int qd_idx = sh->qd_idx;
+
+                       /* Missing D+Q: recompute D from P, then recompute Q */
+                       if (target == qd_idx)
+                               data_target = target2;
+                       else
+                               data_target = target;
+
+                       count = 0;
+                       for (i = disks; i-- ; ) {
+                               if (i == data_target || i == qd_idx)
+                                       continue;
+                               blocks[count++] = sh->dev[i].page;
+                       }
+                       dest = sh->dev[data_target].page;
+                       init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
+                                         NULL, NULL, to_addr_conv(sh, percpu));
+                       tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
+                                      &submit);
+
+                       count = set_syndrome_sources(blocks, sh);
+                       init_async_submit(&submit, 0, tx, ops_complete_compute,
+                                         sh, to_addr_conv(sh, percpu));
+                       return async_gen_syndrome(blocks, 0, count+2,
+                                                 STRIPE_SIZE, &submit);
+               }
+       }
+
+       init_async_submit(&submit, 0, NULL, ops_complete_compute, sh,
+                         to_addr_conv(sh, percpu));
+       if (failb == syndrome_disks) {
+               /* We're missing D+P. */
+               return async_raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE,
+                                              faila, blocks, &submit);
+       } else {
+               /* We're missing D+D. */
+               return async_raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE,
+                                              faila, failb, blocks, &submit);
+       }
+}
+
+
 static void ops_complete_prexor(void *stripe_head_ref)
 {
        struct stripe_head *sh = stripe_head_ref;
@@ -680,12 +895,13 @@ static void ops_complete_prexor(void *stripe_head_ref)
 }
 
 static struct dma_async_tx_descriptor *
-ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
+ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
+              struct dma_async_tx_descriptor *tx)
 {
-       /* kernel stack size limits the total number of disks */
        int disks = sh->disks;
-       struct page *xor_srcs[disks];
+       struct page **xor_srcs = percpu->scribble;
        int count = 0, pd_idx = sh->pd_idx, i;
+       struct async_submit_ctl submit;
 
        /* existing parity data subtracted */
        struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
@@ -700,9 +916,9 @@ ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
                        xor_srcs[count++] = dev->page;
        }
 
-       tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
-               ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
-               ops_complete_prexor, sh);
+       init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST, tx,
+                         ops_complete_prexor, sh, to_addr_conv(sh, percpu));
+       tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 
        return tx;
 }
@@ -742,17 +958,21 @@ ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
        return tx;
 }
 
-static void ops_complete_postxor(void *stripe_head_ref)
+static void ops_complete_reconstruct(void *stripe_head_ref)
 {
        struct stripe_head *sh = stripe_head_ref;
-       int disks = sh->disks, i, pd_idx = sh->pd_idx;
+       int disks = sh->disks;
+       int pd_idx = sh->pd_idx;
+       int qd_idx = sh->qd_idx;
+       int i;
 
        pr_debug("%s: stripe %llu\n", __func__,
                (unsigned long long)sh->sector);
 
        for (i = disks; i--; ) {
                struct r5dev *dev = &sh->dev[i];
-               if (dev->written || i == pd_idx)
+
+               if (dev->written || i == pd_idx || i == qd_idx)
                        set_bit(R5_UPTODATE, &dev->flags);
        }
 
@@ -770,12 +990,12 @@ static void ops_complete_postxor(void *stripe_head_ref)
 }
 
 static void
-ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
+ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
+                    struct dma_async_tx_descriptor *tx)
 {
-       /* kernel stack size limits the total number of disks */
        int disks = sh->disks;
-       struct page *xor_srcs[disks];
-
+       struct page **xor_srcs = percpu->scribble;
+       struct async_submit_ctl submit;
        int count = 0, pd_idx = sh->pd_idx, i;
        struct page *xor_dest;
        int prexor = 0;
@@ -809,18 +1029,36 @@ ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
         * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
         * for the synchronous xor case
         */
-       flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
+       flags = ASYNC_TX_ACK |
                (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
 
        atomic_inc(&sh->count);
 
-       if (unlikely(count == 1)) {
-               flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
-               tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
-                       flags, tx, ops_complete_postxor, sh);
-       } else
-               tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
-                       flags, tx, ops_complete_postxor, sh);
+       init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
+                         to_addr_conv(sh, percpu));
+       if (unlikely(count == 1))
+               tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
+       else
+               tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
+}
+
+static void
+ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
+                    struct dma_async_tx_descriptor *tx)
+{
+       struct async_submit_ctl submit;
+       struct page **blocks = percpu->scribble;
+       int count;
+
+       pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
+
+       count = set_syndrome_sources(blocks, sh);
+
+       atomic_inc(&sh->count);
+
+       init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
+                         sh, to_addr_conv(sh, percpu));
+       async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
 }
 
 static void ops_complete_check(void *stripe_head_ref)
@@ -835,63 +1073,115 @@ static void ops_complete_check(void *stripe_head_ref)
        release_stripe(sh);
 }
 
-static void ops_run_check(struct stripe_head *sh)
+static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
 {
-       /* kernel stack size limits the total number of disks */
        int disks = sh->disks;
-       struct page *xor_srcs[disks];
+       int pd_idx = sh->pd_idx;
+       int qd_idx = sh->qd_idx;
+       struct page *xor_dest;
+       struct page **xor_srcs = percpu->scribble;
        struct dma_async_tx_descriptor *tx;
-
-       int count = 0, pd_idx = sh->pd_idx, i;
-       struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
+       struct async_submit_ctl submit;
+       int count;
+       int i;
 
        pr_debug("%s: stripe %llu\n", __func__,
                (unsigned long long)sh->sector);
 
+       count = 0;
+       xor_dest = sh->dev[pd_idx].page;
+       xor_srcs[count++] = xor_dest;
        for (i = disks; i--; ) {
-               struct r5dev *dev = &sh->dev[i];
-               if (i != pd_idx)
-                       xor_srcs[count++] = dev->page;
+               if (i == pd_idx || i == qd_idx)
+                       continue;
+               xor_srcs[count++] = sh->dev[i].page;
        }
 
-       tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
-               &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
+       init_async_submit(&submit, 0, NULL, NULL, NULL,
+                         to_addr_conv(sh, percpu));
+       tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
+                          &sh->ops.zero_sum_result, &submit);
+
+       atomic_inc(&sh->count);
+       init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
+       tx = async_trigger_callback(&submit);
+}
+
+static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
+{
+       struct page **srcs = percpu->scribble;
+       struct async_submit_ctl submit;
+       int count;
+
+       pr_debug("%s: stripe %llu checkp: %d\n", __func__,
+               (unsigned long long)sh->sector, checkp);
+
+       count = set_syndrome_sources(srcs, sh);
+       if (!checkp)
+               srcs[count] = NULL;
 
        atomic_inc(&sh->count);
-       tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
-               ops_complete_check, sh);
+       init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
+                         sh, to_addr_conv(sh, percpu));
+       async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
+                          &sh->ops.zero_sum_result, percpu->spare_page, &submit);
 }
 
-static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
+static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
 {
        int overlap_clear = 0, i, disks = sh->disks;
        struct dma_async_tx_descriptor *tx = NULL;
+       raid5_conf_t *conf = sh->raid_conf;
+       int level = conf->level;
+       struct raid5_percpu *percpu;
+       unsigned long cpu;
 
+       cpu = get_cpu();
+       percpu = per_cpu_ptr(conf->percpu, cpu);
        if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
                ops_run_biofill(sh);
                overlap_clear++;
        }
 
        if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
-               tx = ops_run_compute5(sh);
-               /* terminate the chain if postxor is not set to be run */
-               if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
+               if (level < 6)
+                       tx = ops_run_compute5(sh, percpu);
+               else {
+                       if (sh->ops.target2 < 0 || sh->ops.target < 0)
+                               tx = ops_run_compute6_1(sh, percpu);
+                       else
+                               tx = ops_run_compute6_2(sh, percpu);
+               }
+               /* terminate the chain if reconstruct is not set to be run */
+               if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
                        async_tx_ack(tx);
        }
 
        if (test_bit(STRIPE_OP_PREXOR, &ops_request))
-               tx = ops_run_prexor(sh, tx);
+               tx = ops_run_prexor(sh, percpu, tx);
 
        if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
                tx = ops_run_biodrain(sh, tx);
                overlap_clear++;
        }
 
-       if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
-               ops_run_postxor(sh, tx);
+       if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
+               if (level < 6)
+                       ops_run_reconstruct5(sh, percpu, tx);
+               else
+                       ops_run_reconstruct6(sh, percpu, tx);
+       }
 
-       if (test_bit(STRIPE_OP_CHECK, &ops_request))
-               ops_run_check(sh);
+       if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
+               if (sh->check_state == check_state_run)
+                       ops_run_check_p(sh, percpu);
+               else if (sh->check_state == check_state_run_q)
+                       ops_run_check_pq(sh, percpu, 0);
+               else if (sh->check_state == check_state_run_pq)
+                       ops_run_check_pq(sh, percpu, 1);
+               else
+                       BUG();
+       }
 
        if (overlap_clear)
                for (i = disks; i--; ) {
@@ -899,6 +1189,7 @@ static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
                        if (test_and_clear_bit(R5_Overlap, &dev->flags))
                                wake_up(&sh->raid_conf->wait_for_overlap);
                }
+       put_cpu();
 }
 
 static int grow_one_stripe(raid5_conf_t *conf)
@@ -948,6 +1239,28 @@ static int grow_stripes(raid5_conf_t *conf, int num)
        return 0;
 }
 
+/**
+ * scribble_len - return the required size of the scribble region
+ * @num - total number of disks in the array
+ *
+ * The size must be enough to contain:
+ * 1/ a struct page pointer for each device in the array +2
+ * 2/ room to convert each entry in (1) to its corresponding dma
+ *    (dma_map_page()) or page (page_address()) address.
+ *
+ * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
+ * calculate over all devices (not just the data blocks), using zeros in place
+ * of the P and Q blocks.
+ */
+static size_t scribble_len(int num)
+{
+       size_t len;
+
+       len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
+
+       return len;
+}
+
 static int resize_stripes(raid5_conf_t *conf, int newsize)
 {
        /* Make all the stripes able to hold 'newsize' devices.
@@ -976,6 +1289,7 @@ static int resize_stripes(raid5_conf_t *conf, int newsize)
        struct stripe_head *osh, *nsh;
        LIST_HEAD(newstripes);
        struct disk_info *ndisks;
+       unsigned long cpu;
        int err;
        struct kmem_cache *sc;
        int i;
@@ -1041,7 +1355,7 @@ static int resize_stripes(raid5_conf_t *conf, int newsize)
        /* Step 3.
         * At this point, we are holding all the stripes so the array
         * is completely stalled, so now is a good time to resize
-        * conf->disks.
+        * conf->disks and the scribble region
         */
        ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
        if (ndisks) {
@@ -1052,10 +1366,30 @@ static int resize_stripes(raid5_conf_t *conf, int newsize)
        } else
                err = -ENOMEM;
 
+       get_online_cpus();
+       conf->scribble_len = scribble_len(newsize);
+       for_each_present_cpu(cpu) {
+               struct raid5_percpu *percpu;
+               void *scribble;
+
+               percpu = per_cpu_ptr(conf->percpu, cpu);
+               scribble = kmalloc(conf->scribble_len, GFP_NOIO);
+
+               if (scribble) {
+                       kfree(percpu->scribble);
+                       percpu->scribble = scribble;
+               } else {
+                       err = -ENOMEM;
+                       break;
+               }
+       }
+       put_online_cpus();
+
        /* Step 4, return new stripes to service */
        while(!list_empty(&newstripes)) {
                nsh = list_entry(newstripes.next, struct stripe_head, lru);
                list_del_init(&nsh->lru);
+
                for (i=conf->raid_disks; i < newsize; i++)
                        if (nsh->dev[i].page == NULL) {
                                struct page *p = alloc_page(GFP_NOIO);
@@ -1594,258 +1928,13 @@ static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
 }
 
 
-
-/*
- * Copy data between a page in the stripe cache, and one or more bion
- * The page could align with the middle of the bio, or there could be
- * several bion, each with several bio_vecs, which cover part of the page
- * Multiple bion are linked together on bi_next.  There may be extras
- * at the end of this list.  We ignore them.
- */
-static void copy_data(int frombio, struct bio *bio,
-                    struct page *page,
-                    sector_t sector)
-{
-       char *pa = page_address(page);
-       struct bio_vec *bvl;
-       int i;
-       int page_offset;
-
-       if (bio->bi_sector >= sector)
-               page_offset = (signed)(bio->bi_sector - sector) * 512;
-       else
-               page_offset = (signed)(sector - bio->bi_sector) * -512;
-       bio_for_each_segment(bvl, bio, i) {
-               int len = bio_iovec_idx(bio,i)->bv_len;
-               int clen;
-               int b_offset = 0;
-
-               if (page_offset < 0) {
-                       b_offset = -page_offset;
-                       page_offset += b_offset;
-                       len -= b_offset;
-               }
-
-               if (len > 0 && page_offset + len > STRIPE_SIZE)
-                       clen = STRIPE_SIZE - page_offset;
-               else clen = len;
-
-               if (clen > 0) {
-                       char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
-                       if (frombio)
-                               memcpy(pa+page_offset, ba+b_offset, clen);
-                       else
-                               memcpy(ba+b_offset, pa+page_offset, clen);
-                       __bio_kunmap_atomic(ba, KM_USER0);
-               }
-               if (clen < len) /* hit end of page */
-                       break;
-               page_offset +=  len;
-       }
-}
-
-#define check_xor()    do {                                              \
-                               if (count == MAX_XOR_BLOCKS) {            \
-                               xor_blocks(count, STRIPE_SIZE, dest, ptr);\
-                               count = 0;                                \
-                          }                                              \
-                       } while(0)
-
-static void compute_parity6(struct stripe_head *sh, int method)
-{
-       raid5_conf_t *conf = sh->raid_conf;
-       int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
-       int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
-       struct bio *chosen;
-       /**** FIX THIS: This could be very bad if disks is close to 256 ****/
-       void *ptrs[syndrome_disks+2];
-
-       pd_idx = sh->pd_idx;
-       qd_idx = sh->qd_idx;
-       d0_idx = raid6_d0(sh);
-
-       pr_debug("compute_parity, stripe %llu, method %d\n",
-               (unsigned long long)sh->sector, method);
-
-       switch(method) {
-       case READ_MODIFY_WRITE:
-               BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
-       case RECONSTRUCT_WRITE:
-               for (i= disks; i-- ;)
-                       if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
-                               chosen = sh->dev[i].towrite;
-                               sh->dev[i].towrite = NULL;
-
-                               if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
-                                       wake_up(&conf->wait_for_overlap);
-
-                               BUG_ON(sh->dev[i].written);
-                               sh->dev[i].written = chosen;
-                       }
-               break;
-       case CHECK_PARITY:
-               BUG();          /* Not implemented yet */
-       }
-
-       for (i = disks; i--;)
-               if (sh->dev[i].written) {
-                       sector_t sector = sh->dev[i].sector;
-                       struct bio *wbi = sh->dev[i].written;
-                       while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
-                               copy_data(1, wbi, sh->dev[i].page, sector);
-                               wbi = r5_next_bio(wbi, sector);
-                       }
-
-                       set_bit(R5_LOCKED, &sh->dev[i].flags);
-                       set_bit(R5_UPTODATE, &sh->dev[i].flags);
-               }
-
-       /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
-
-       for (i = 0; i < disks; i++)
-               ptrs[i] = (void *)raid6_empty_zero_page;
-
-       count = 0;
-       i = d0_idx;
-       do {
-               int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
-
-               ptrs[slot] = page_address(sh->dev[i].page);
-               if (slot < syndrome_disks &&
-                   !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
-                       printk(KERN_ERR "block %d/%d not uptodate "
-                              "on parity calc\n", i, count);
-                       BUG();
-               }
-
-               i = raid6_next_disk(i, disks);
-       } while (i != d0_idx);
-       BUG_ON(count != syndrome_disks);
-
-       raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
-
-       switch(method) {
-       case RECONSTRUCT_WRITE:
-               set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
-               set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
-               set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
-               set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
-               break;
-       case UPDATE_PARITY:
-               set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
-               set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
-               break;
-       }
-}
-
-
-/* Compute one missing block */
-static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
-{
-       int i, count, disks = sh->disks;
-       void *ptr[MAX_XOR_BLOCKS], *dest, *p;
-       int qd_idx = sh->qd_idx;
-
-       pr_debug("compute_block_1, stripe %llu, idx %d\n",
-               (unsigned long long)sh->sector, dd_idx);
-
-       if ( dd_idx == qd_idx ) {
-               /* We're actually computing the Q drive */
-               compute_parity6(sh, UPDATE_PARITY);
-       } else {
-               dest = page_address(sh->dev[dd_idx].page);
-               if (!nozero) memset(dest, 0, STRIPE_SIZE);
-               count = 0;
-               for (i = disks ; i--; ) {
-                       if (i == dd_idx || i == qd_idx)
-                               continue;
-                       p = page_address(sh->dev[i].page);
-                       if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
-                               ptr[count++] = p;
-                       else
-                               printk("compute_block() %d, stripe %llu, %d"
-                                      " not present\n", dd_idx,
-                                      (unsigned long long)sh->sector, i);
-
-                       check_xor();
-               }
-               if (count)
-                       xor_blocks(count, STRIPE_SIZE, dest, ptr);
-               if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
-               else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
-       }
-}
-
-/* Compute two missing blocks */
-static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
-{
-       int i, count, disks = sh->disks;
-       int syndrome_disks = sh->ddf_layout ? disks : disks-2;
-       int d0_idx = raid6_d0(sh);
-       int faila = -1, failb = -1;
-       /**** FIX THIS: This could be very bad if disks is close to 256 ****/
-       void *ptrs[syndrome_disks+2];
-
-       for (i = 0; i < disks ; i++)
-               ptrs[i] = (void *)raid6_empty_zero_page;
-       count = 0;
-       i = d0_idx;
-       do {
-               int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
-
-               ptrs[slot] = page_address(sh->dev[i].page);
-
-               if (i == dd_idx1)
-                       faila = slot;
-               if (i == dd_idx2)
-                       failb = slot;
-               i = raid6_next_disk(i, disks);
-       } while (i != d0_idx);
-       BUG_ON(count != syndrome_disks);
-
-       BUG_ON(faila == failb);
-       if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
-
-       pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
-                (unsigned long long)sh->sector, dd_idx1, dd_idx2,
-                faila, failb);
-
-       if (failb == syndrome_disks+1) {
-               /* Q disk is one of the missing disks */
-               if (faila == syndrome_disks) {
-                       /* Missing P+Q, just recompute */
-                       compute_parity6(sh, UPDATE_PARITY);
-                       return;
-               } else {
-                       /* We're missing D+Q; recompute D from P */
-                       compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
-                                            dd_idx2 : dd_idx1),
-                                       0);
-                       compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
-                       return;
-               }
-       }
-
-       /* We're missing D+P or D+D; */
-       if (failb == syndrome_disks) {
-               /* We're missing D+P. */
-               raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
-       } else {
-               /* We're missing D+D. */
-               raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
-                                 ptrs);
-       }
-
-       /* Both the above update both missing blocks */
-       set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
-       set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
-}
-
 static void
-schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
+schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
                         int rcw, int expand)
 {
        int i, pd_idx = sh->pd_idx, disks = sh->disks;
+       raid5_conf_t *conf = sh->raid_conf;
+       int level = conf->level;
 
        if (rcw) {
                /* if we are not expanding this is a proper write request, and
@@ -1858,7 +1947,7 @@ schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
                } else
                        sh->reconstruct_state = reconstruct_state_run;
 
-               set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
+               set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
 
                for (i = disks; i--; ) {
                        struct r5dev *dev = &sh->dev[i];
@@ -1871,17 +1960,18 @@ schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
                                s->locked++;
                        }
                }
-               if (s->locked + 1 == disks)
+               if (s->locked + conf->max_degraded == disks)
                        if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
-                               atomic_inc(&sh->raid_conf->pending_full_writes);
+                               atomic_inc(&conf->pending_full_writes);
        } else {
+               BUG_ON(level == 6);
                BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
                        test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
 
                sh->reconstruct_state = reconstruct_state_prexor_drain_run;
                set_bit(STRIPE_OP_PREXOR, &s->ops_request);
                set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
-               set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
+               set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
 
                for (i = disks; i--; ) {
                        struct r5dev *dev = &sh->dev[i];
@@ -1899,13 +1989,22 @@ schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
                }
        }
 
-       /* keep the parity disk locked while asynchronous operations
+       /* keep the parity disk(s) locked while asynchronous operations
         * are in flight
         */
        set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
        clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
        s->locked++;
 
+       if (level == 6) {
+               int qd_idx = sh->qd_idx;
+               struct r5dev *dev = &sh->dev[qd_idx];
+
+               set_bit(R5_LOCKED, &dev->flags);
+               clear_bit(R5_UPTODATE, &dev->flags);
+               s->locked++;
+       }
+
        pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
                __func__, (unsigned long long)sh->sector,
                s->locked, s->ops_request);
@@ -1986,13 +2085,6 @@ static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, in
 
 static void end_reshape(raid5_conf_t *conf);
 
-static int page_is_zero(struct page *p)
-{
-       char *a = page_address(p);
-       return ((*(u32*)a) == 0 &&
-               memcmp(a, a+4, STRIPE_SIZE-4)==0);
-}
-
 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
                            struct stripe_head *sh)
 {
@@ -2133,9 +2225,10 @@ static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
                        set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
                        set_bit(R5_Wantcompute, &dev->flags);
                        sh->ops.target = disk_idx;
+                       sh->ops.target2 = -1;
                        s->req_compute = 1;
                        /* Careful: from this point on 'uptodate' is in the eye
-                        * of raid5_run_ops which services 'compute' operations
+                        * of raid_run_ops which services 'compute' operations
                         * before writes. R5_Wantcompute flags a block that will
                         * be R5_UPTODATE by the time it is needed for a
                         * subsequent operation.
@@ -2174,61 +2267,104 @@ static void handle_stripe_fill5(struct stripe_head *sh,
        set_bit(STRIPE_HANDLE, &sh->state);
 }
 
-static void handle_stripe_fill6(struct stripe_head *sh,
-                       struct stripe_head_state *s, struct r6_state *r6s,
-                       int disks)
+/* fetch_block6 - checks the given member device to see if its data needs
+ * to be read or computed to satisfy a request.
+ *
+ * Returns 1 when no more member devices need to be checked, otherwise returns
+ * 0 to tell the loop in handle_stripe_fill6 to continue
+ */
+static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
+                        struct r6_state *r6s, int disk_idx, int disks)
 {
-       int i;
-       for (i = disks; i--; ) {
-               struct r5dev *dev = &sh->dev[i];
-               if (!test_bit(R5_LOCKED, &dev->flags) &&
-                   !test_bit(R5_UPTODATE, &dev->flags) &&
-                   (dev->toread || (dev->towrite &&
-                    !test_bit(R5_OVERWRITE, &dev->flags)) ||
-                    s->syncing || s->expanding ||
-                    (s->failed >= 1 &&
-                     (sh->dev[r6s->failed_num[0]].toread ||
-                      s->to_write)) ||
-                    (s->failed >= 2 &&
-                     (sh->dev[r6s->failed_num[1]].toread ||
-                      s->to_write)))) {
-                       /* we would like to get this block, possibly
-                        * by computing it, but we might not be able to
+       struct r5dev *dev = &sh->dev[disk_idx];
+       struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
+                                 &sh->dev[r6s->failed_num[1]] };
+
+       if (!test_bit(R5_LOCKED, &dev->flags) &&
+           !test_bit(R5_UPTODATE, &dev->flags) &&
+           (dev->toread ||
+            (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
+            s->syncing || s->expanding ||
+            (s->failed >= 1 &&
+             (fdev[0]->toread || s->to_write)) ||
+            (s->failed >= 2 &&
+             (fdev[1]->toread || s->to_write)))) {
+               /* we would like to get this block, possibly by computing it,
+                * otherwise read it if the backing disk is insync
+                */
+               BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
+               BUG_ON(test_bit(R5_Wantread, &dev->flags));
+               if ((s->uptodate == disks - 1) &&
+                   (s->failed && (disk_idx == r6s->failed_num[0] ||
+                                  disk_idx == r6s->failed_num[1]))) {
+                       /* have disk failed, and we're requested to fetch it;
+                        * do compute it
                         */
-                       if ((s->uptodate == disks - 1) &&
-                           (s->failed && (i == r6s->failed_num[0] ||
-                                          i == r6s->failed_num[1]))) {
-                               pr_debug("Computing stripe %llu block %d\n",
-                                      (unsigned long long)sh->sector, i);
-                               compute_block_1(sh, i, 0);
-                               s->uptodate++;
-                       } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
-                               /* Computing 2-failure is *very* expensive; only
-                                * do it if failed >= 2
-                                */
-                               int other;
-                               for (other = disks; other--; ) {
-                                       if (other == i)
-                                               continue;
-                                       if (!test_bit(R5_UPTODATE,
-                                             &sh->dev[other].flags))
-                                               break;
-                               }
-                               BUG_ON(other < 0);
-                               pr_debug("Computing stripe %llu blocks %d,%d\n",
-                                      (unsigned long long)sh->sector,
-                                      i, other);
-                               compute_block_2(sh, i, other);
-                               s->uptodate += 2;
-                       } else if (test_bit(R5_Insync, &dev->flags)) {
-                               set_bit(R5_LOCKED, &dev->flags);
-                               set_bit(R5_Wantread, &dev->flags);
-                               s->locked++;
-                               pr_debug("Reading block %d (sync=%d)\n",
-                                       i, s->syncing);
+                       pr_debug("Computing stripe %llu block %d\n",
+                              (unsigned long long)sh->sector, disk_idx);
+                       set_bit(STRIPE_COMPUTE_RUN, &sh->state);
+                       set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
+                       set_bit(R5_Wantcompute, &dev->flags);
+                       sh->ops.target = disk_idx;
+                       sh->ops.target2 = -1; /* no 2nd target */
+                       s->req_compute = 1;
+                       s->uptodate++;
+                       return 1;
+               } else if (s->uptodate == disks-2 && s->failed >= 2) {
+                       /* Computing 2-failure is *very* expensive; only
+                        * do it if failed >= 2
+                        */
+                       int other;
+                       for (other = disks; other--; ) {
+                               if (other == disk_idx)
+                                       continue;
+                               if (!test_bit(R5_UPTODATE,
+                                     &sh->dev[other].flags))
+                                       break;
                        }
+                       BUG_ON(other < 0);
+                       pr_debug("Computing stripe %llu blocks %d,%d\n",
+                              (unsigned long long)sh->sector,
+                              disk_idx, other);
+                       set_bit(STRIPE_COMPUTE_RUN, &sh->state);
+                       set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
+                       set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
+                       set_bit(R5_Wantcompute, &sh->dev[other].flags);
+                       sh->ops.target = disk_idx;
+                       sh->ops.target2 = other;
+                       s->uptodate += 2;
+                       s->req_compute = 1;
+                       return 1;
+               } else if (test_bit(R5_Insync, &dev->flags)) {
+                       set_bit(R5_LOCKED, &dev->flags);
+                       set_bit(R5_Wantread, &dev->flags);
+                       s->locked++;
+                       pr_debug("Reading block %d (sync=%d)\n",
+                               disk_idx, s->syncing);
                }
        }
+
+       return 0;
+}
+
+/**
+ * handle_stripe_fill6 - read or compute data to satisfy pending requests.
+ */
+static void handle_stripe_fill6(struct stripe_head *sh,
+                       struct stripe_head_state *s, struct r6_state *r6s,
+                       int disks)
+{
+       int i;
+
+       /* look for blocks to read/compute, skip this if a compute
+        * is already in flight, or if the stripe contents are in the
+        * midst of changing due to a write
+        */
+       if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
+           !sh->reconstruct_state)
+               for (i = disks; i--; )
+                       if (fetch_block6(sh, s, r6s, i, disks))
+                               break;
        set_bit(STRIPE_HANDLE, &sh->state);
 }
 
@@ -2362,114 +2498,61 @@ static void handle_stripe_dirtying5(raid5_conf_t *conf,
         */
        /* since handle_stripe can be called at any time we need to handle the
         * case where a compute block operation has been submitted and then a
-        * subsequent call wants to start a write request.  raid5_run_ops only
-        * handles the case where compute block and postxor are requested
+        * subsequent call wants to start a write request.  raid_run_ops only
+        * handles the case where compute block and reconstruct are requested
         * simultaneously.  If this is not the case then new writes need to be
         * held off until the compute completes.
         */
        if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
            (s->locked == 0 && (rcw == 0 || rmw == 0) &&
            !test_bit(STRIPE_BIT_DELAY, &sh->state)))
-               schedule_reconstruction5(sh, s, rcw == 0, 0);
+               schedule_reconstruction(sh, s, rcw == 0, 0);
 }
 
 static void handle_stripe_dirtying6(raid5_conf_t *conf,
                struct stripe_head *sh, struct stripe_head_state *s,
                struct r6_state *r6s, int disks)
 {
-       int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
+       int rcw = 0, pd_idx = sh->pd_idx, i;
        int qd_idx = sh->qd_idx;
+
+       set_bit(STRIPE_HANDLE, &sh->state);
        for (i = disks; i--; ) {
                struct r5dev *dev = &sh->dev[i];
-               /* Would I have to read this buffer for reconstruct_write */
-               if (!test_bit(R5_OVERWRITE, &dev->flags)
-                   && i != pd_idx && i != qd_idx
-                   && (!test_bit(R5_LOCKED, &dev->flags)
-                           ) &&
-                   !test_bit(R5_UPTODATE, &dev->flags)) {
-                       if (test_bit(R5_Insync, &dev->flags)) rcw++;
-                       else {
-                               pr_debug("raid6: must_compute: "
-                                       "disk %d flags=%#lx\n", i, dev->flags);
-                               must_compute++;
+               /* check if we haven't enough data */
+               if (!test_bit(R5_OVERWRITE, &dev->flags) &&
+                   i != pd_idx && i != qd_idx &&
+                   !test_bit(R5_LOCKED, &dev->flags) &&
+                   !(test_bit(R5_UPTODATE, &dev->flags) ||
+                     test_bit(R5_Wantcompute, &dev->flags))) {
+                       rcw++;
+                       if (!test_bit(R5_Insync, &dev->flags))
+                               continue; /* it's a failed drive */
+
+                       if (
+                         test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
+                               pr_debug("Read_old stripe %llu "
+                                       "block %d for Reconstruct\n",
+                                    (unsigned long long)sh->sector, i);
+                               set_bit(R5_LOCKED, &dev->flags);
+                               set_bit(R5_Wantread, &dev->flags);
+                               s->locked++;
+                       } else {
+                               pr_debug("Request delayed stripe %llu "
+                                       "block %d for Reconstruct\n",
+                                    (unsigned long long)sh->sector, i);
+                               set_bit(STRIPE_DELAYED, &sh->state);
+                               set_bit(STRIPE_HANDLE, &sh->state);
                        }
                }
        }
-       pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
-              (unsigned long long)sh->sector, rcw, must_compute);
-       set_bit(STRIPE_HANDLE, &sh->state);
-
-       if (rcw > 0)
-               /* want reconstruct write, but need to get some data */
-               for (i = disks; i--; ) {
-                       struct r5dev *dev = &sh->dev[i];
-                       if (!test_bit(R5_OVERWRITE, &dev->flags)
-                           && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
-                           && !test_bit(R5_LOCKED, &dev->flags) &&
-                           !test_bit(R5_UPTODATE, &dev->flags) &&
-                           test_bit(R5_Insync, &dev->flags)) {
-                               if (
-                                 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
-                                       pr_debug("Read_old stripe %llu "
-                                               "block %d for Reconstruct\n",
-                                            (unsigned long long)sh->sector, i);
-                                       set_bit(R5_LOCKED, &dev->flags);
-                                       set_bit(R5_Wantread, &dev->flags);
-                                       s->locked++;
-                               } else {
-                                       pr_debug("Request delayed stripe %llu "
-                                               "block %d for Reconstruct\n",
-                                            (unsigned long long)sh->sector, i);
-                                       set_bit(STRIPE_DELAYED, &sh->state);
-                                       set_bit(STRIPE_HANDLE, &sh->state);
-                               }
-                       }
-               }
        /* now if nothing is locked, and if we have enough data, we can start a
         * write request
         */
-       if (s->locked == 0 && rcw == 0 &&
+       if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
+           s->locked == 0 && rcw == 0 &&
            !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
-               if (must_compute > 0) {
-                       /* We have failed blocks and need to compute them */
-                       switch (s->failed) {
-                       case 0:
-                               BUG();
-                       case 1:
-                               compute_block_1(sh, r6s->failed_num[0], 0);
-                               break;
-                       case 2:
-                               compute_block_2(sh, r6s->failed_num[0],
-                                               r6s->failed_num[1]);
-                               break;
-                       default: /* This request should have been failed? */
-                               BUG();
-                       }
-               }
-
-               pr_debug("Computing parity for stripe %llu\n",
-                       (unsigned long long)sh->sector);
-               compute_parity6(sh, RECONSTRUCT_WRITE);
-               /* now every locked buffer is ready to be written */
-               for (i = disks; i--; )
-                       if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
-                               pr_debug("Writing stripe %llu block %d\n",
-                                      (unsigned long long)sh->sector, i);
-                               s->locked++;
-                               set_bit(R5_Wantwrite, &sh->dev[i].flags);
-                       }
-               if (s->locked == disks)
-                       if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
-                               atomic_inc(&conf->pending_full_writes);
-               /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
-               set_bit(STRIPE_INSYNC, &sh->state);
-
-               if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
-                       atomic_dec(&conf->preread_active_stripes);
-                       if (atomic_read(&conf->preread_active_stripes) <
-                           IO_THRESHOLD)
-                               md_wakeup_thread(conf->mddev->thread);
-               }
+               schedule_reconstruction(sh, s, 1, 0);
        }
 }
 
@@ -2528,7 +2611,7 @@ static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
                 * we are done.  Otherwise update the mismatch count and repair
                 * parity if !MD_RECOVERY_CHECK
                 */
-               if (sh->ops.zero_sum_result == 0)
+               if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
                        /* parity is correct (on disc,
                         * not in buffer any more)
                         */
@@ -2545,6 +2628,7 @@ static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
                                set_bit(R5_Wantcompute,
                                        &sh->dev[sh->pd_idx].flags);
                                sh->ops.target = sh->pd_idx;
+                               sh->ops.target2 = -1;
                                s->uptodate++;
                        }
                }
@@ -2561,67 +2645,74 @@ static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
 
 
 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
-                               struct stripe_head_state *s,
-                               struct r6_state *r6s, struct page *tmp_page,
-                               int disks)
+                                 struct stripe_head_state *s,
+                                 struct r6_state *r6s, int disks)
 {
-       int update_p = 0, update_q = 0;
-       struct r5dev *dev;
        int pd_idx = sh->pd_idx;
        int qd_idx = sh->qd_idx;
+       struct r5dev *dev;
 
        set_bit(STRIPE_HANDLE, &sh->state);
 
        BUG_ON(s->failed > 2);
-       BUG_ON(s->uptodate < disks);
+
        /* Want to check and possibly repair P and Q.
         * However there could be one 'failed' device, in which
         * case we can only check one of them, possibly using the
         * other to generate missing data
         */
 
-       /* If !tmp_page, we cannot do the calculations,
-        * but as we have set STRIPE_HANDLE, we will soon be called
-        * by stripe_handle with a tmp_page - just wait until then.
-        */
-       if (tmp_page) {
+       switch (sh->check_state) {
+       case check_state_idle:
+               /* start a new check operation if there are < 2 failures */
                if (s->failed == r6s->q_failed) {
-                       /* The only possible failed device holds 'Q', so it
+                       /* The only possible failed device holds Q, so it
                         * makes sense to check P (If anything else were failed,
                         * we would have used P to recreate it).
                         */
-                       compute_block_1(sh, pd_idx, 1);
-                       if (!page_is_zero(sh->dev[pd_idx].page)) {
-                               compute_block_1(sh, pd_idx, 0);
-                               update_p = 1;
-                       }
+                       sh->check_state = check_state_run;
                }
                if (!r6s->q_failed && s->failed < 2) {
-                       /* q is not failed, and we didn't use it to generate
+                       /* Q is not failed, and we didn't use it to generate
                         * anything, so it makes sense to check it
                         */
-                       memcpy(page_address(tmp_page),
-                              page_address(sh->dev[qd_idx].page),
-                              STRIPE_SIZE);
-                       compute_parity6(sh, UPDATE_PARITY);
-                       if (memcmp(page_address(tmp_page),
-                                  page_address(sh->dev[qd_idx].page),
-                                  STRIPE_SIZE) != 0) {
-                               clear_bit(STRIPE_INSYNC, &sh->state);
-                               update_q = 1;
-                       }
+                       if (sh->check_state == check_state_run)
+                               sh->check_state = check_state_run_pq;
+                       else
+                               sh->check_state = check_state_run_q;
                }
-               if (update_p || update_q) {
-                       conf->mddev->resync_mismatches += STRIPE_SECTORS;
-                       if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
-                               /* don't try to repair!! */
-                               update_p = update_q = 0;
+
+               /* discard potentially stale zero_sum_result */
+               sh->ops.zero_sum_result = 0;
+
+               if (sh->check_state == check_state_run) {
+                       /* async_xor_zero_sum destroys the contents of P */
+                       clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
+                       s->uptodate--;
                }
+               if (sh->check_state >= check_state_run &&
+                   sh->check_state <= check_state_run_pq) {
+                       /* async_syndrome_zero_sum preserves P and Q, so
+                        * no need to mark them !uptodate here
+                        */
+                       set_bit(STRIPE_OP_CHECK, &s->ops_request);
+                       break;
+               }
+
+               /* we have 2-disk failure */
+               BUG_ON(s->failed != 2);
+               /* fall through */
+       case check_state_compute_result:
+               sh->check_state = check_state_idle;
+
+               /* check that a write has not made the stripe insync */
+               if (test_bit(STRIPE_INSYNC, &sh->state))
+                       break;
 
                /* now write out any block on a failed drive,
-                * or P or Q if they need it
+                * or P or Q if they were recomputed
                 */
-
+               BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
                if (s->failed == 2) {
                        dev = &sh->dev[r6s->failed_num[1]];
                        s->locked++;
@@ -2634,14 +2725,13 @@ static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
                        set_bit(R5_LOCKED, &dev->flags);
                        set_bit(R5_Wantwrite, &dev->flags);
                }
-
-               if (update_p) {
+               if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
                        dev = &sh->dev[pd_idx];
                        s->locked++;
                        set_bit(R5_LOCKED, &dev->flags);
                        set_bit(R5_Wantwrite, &dev->flags);
                }
-               if (update_q) {
+               if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
                        dev = &sh->dev[qd_idx];
                        s->locked++;
                        set_bit(R5_LOCKED, &dev->flags);
@@ -2650,6 +2740,70 @@ static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
                clear_bit(STRIPE_DEGRADED, &sh->state);
 
                set_bit(STRIPE_INSYNC, &sh->state);
+               break;
+       case check_state_run:
+       case check_state_run_q:
+       case check_state_run_pq:
+               break; /* we will be called again upon completion */
+       case check_state_check_result:
+               sh->check_state = check_state_idle;
+
+               /* handle a successful check operation, if parity is correct
+                * we are done.  Otherwise update the mismatch count and repair
+                * parity if !MD_RECOVERY_CHECK
+                */
+               if (sh->ops.zero_sum_result == 0) {
+                       /* both parities are correct */
+                       if (!s->failed)
+                               set_bit(STRIPE_INSYNC, &sh->state);
+                       else {
+                               /* in contrast to the raid5 case we can validate
+                                * parity, but still have a failure to write
+                                * back
+                                */
+                               sh->check_state = check_state_compute_result;
+                               /* Returning at this point means that we may go
+                                * off and bring p and/or q uptodate again so
+                                * we make sure to check zero_sum_result again
+                                * to verify if p or q need writeback
+                                */
+                       }
+               } else {
+                       conf->mddev->resync_mismatches += STRIPE_SECTORS;
+                       if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
+                               /* don't try to repair!! */
+                               set_bit(STRIPE_INSYNC, &sh->state);
+                       else {
+                               int *target = &sh->ops.target;
+
+                               sh->ops.target = -1;
+                               sh->ops.target2 = -1;
+                               sh->check_state = check_state_compute_run;
+                               set_bit(STRIPE_COMPUTE_RUN, &sh->state);
+                               set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
+                               if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
+                                       set_bit(R5_Wantcompute,
+                                               &sh->dev[pd_idx].flags);
+                                       *target = pd_idx;
+                                       target = &sh->ops.target2;
+                                       s->uptodate++;
+                               }
+                               if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
+                                       set_bit(R5_Wantcompute,
+                                               &sh->dev[qd_idx].flags);
+                                       *target = qd_idx;
+                                       s->uptodate++;
+                               }
+                       }
+               }
+               break;
+       case check_state_compute_run:
+               break;
+       default:
+               printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
+                      __func__, sh->check_state,
+                      (unsigned long long) sh->sector);
+               BUG();
        }
 }
 
@@ -2667,6 +2821,7 @@ static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
                if (i != sh->pd_idx && i != sh->qd_idx) {
                        int dd_idx, j;
                        struct stripe_head *sh2;
+                       struct async_submit_ctl submit;
 
                        sector_t bn = compute_blocknr(sh, i, 1);
                        sector_t s = raid5_compute_sector(conf, bn, 0,
@@ -2686,9 +2841,10 @@ static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
                        }
 
                        /* place all the copies on one channel */
+                       init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
                        tx = async_memcpy(sh2->dev[dd_idx].page,
-                               sh->dev[i].page, 0, 0, STRIPE_SIZE,
-                               ASYNC_TX_DEP_ACK, tx, NULL, NULL);
+                                         sh->dev[i].page, 0, 0, STRIPE_SIZE,
+                                         &submit);
 
                        set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
                        set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
@@ -2974,7 +3130,7 @@ static bool handle_stripe5(struct stripe_head *sh)
                /* Need to write out all blocks after computing parity */
                sh->disks = conf->raid_disks;
                stripe_set_idx(sh->sector, conf, 0, sh);
-               schedule_reconstruction5(sh, &s, 1, 1);
+               schedule_reconstruction(sh, &s, 1, 1);
        } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
                clear_bit(STRIPE_EXPAND_READY, &sh->state);
                atomic_dec(&conf->reshape_stripes);
@@ -2994,7 +3150,7 @@ static bool handle_stripe5(struct stripe_head *sh)
                md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
 
        if (s.ops_request)
-               raid5_run_ops(sh, s.ops_request);
+               raid_run_ops(sh, s.ops_request);
 
        ops_run_io(sh, &s);
 
@@ -3003,7 +3159,7 @@ static bool handle_stripe5(struct stripe_head *sh)
        return blocked_rdev == NULL;
 }
 
-static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
+static bool handle_stripe6(struct stripe_head *sh)
 {
        raid5_conf_t *conf = sh->raid_conf;
        int disks = sh->disks;
@@ -3015,9 +3171,10 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
        mdk_rdev_t *blocked_rdev = NULL;
 
        pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
-               "pd_idx=%d, qd_idx=%d\n",
+               "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
               (unsigned long long)sh->sector, sh->state,
-              atomic_read(&sh->count), pd_idx, qd_idx);
+              atomic_read(&sh->count), pd_idx, qd_idx,
+              sh->check_state, sh->reconstruct_state);
        memset(&s, 0, sizeof(s));
 
        spin_lock(&sh->lock);
@@ -3037,35 +3194,24 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
 
                pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
                        i, dev->flags, dev->toread, dev->towrite, dev->written);
-               /* maybe we can reply to a read */
-               if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
-                       struct bio *rbi, *rbi2;
-                       pr_debug("Return read for disc %d\n", i);
-                       spin_lock_irq(&conf->device_lock);
-                       rbi = dev->toread;
-                       dev->toread = NULL;
-                       if (test_and_clear_bit(R5_Overlap, &dev->flags))
-                               wake_up(&conf->wait_for_overlap);
-                       spin_unlock_irq(&conf->device_lock);
-                       while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
-                               copy_data(0, rbi, dev->page, dev->sector);
-                               rbi2 = r5_next_bio(rbi, dev->sector);
-                               spin_lock_irq(&conf->device_lock);
-                               if (!raid5_dec_bi_phys_segments(rbi)) {
-                                       rbi->bi_next = return_bi;
-                                       return_bi = rbi;
-                               }
-                               spin_unlock_irq(&conf->device_lock);
-                               rbi = rbi2;
-                       }
-               }
+               /* maybe we can reply to a read
+                *
+                * new wantfill requests are only permitted while
+                * ops_complete_biofill is guaranteed to be inactive
+                */
+               if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
+                   !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
+                       set_bit(R5_Wantfill, &dev->flags);
 
                /* now count some things */
                if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
                if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
+               if (test_bit(R5_Wantcompute, &dev->flags))
+                       BUG_ON(++s.compute > 2);
 
-
-               if (dev->toread)
+               if (test_bit(R5_Wantfill, &dev->flags)) {
+                       s.to_fill++;
+               } else if (dev->toread)
                        s.to_read++;
                if (dev->towrite) {
                        s.to_write++;
@@ -3106,6 +3252,11 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
                blocked_rdev = NULL;
        }
 
+       if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
+               set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
+               set_bit(STRIPE_BIOFILL_RUN, &sh->state);
+       }
+
        pr_debug("locked=%d uptodate=%d to_read=%d"
               " to_write=%d failed=%d failed_num=%d,%d\n",
               s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
@@ -3146,19 +3297,62 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
         * or to load a block that is being partially written.
         */
        if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
-           (s.syncing && (s.uptodate < disks)) || s.expanding)
+           (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
                handle_stripe_fill6(sh, &s, &r6s, disks);
 
-       /* now to consider writing and what else, if anything should be read */
-       if (s.to_write)
+       /* Now we check to see if any write operations have recently
+        * completed
+        */
+       if (sh->reconstruct_state == reconstruct_state_drain_result) {
+               int qd_idx = sh->qd_idx;
+
+               sh->reconstruct_state = reconstruct_state_idle;
+               /* All the 'written' buffers and the parity blocks are ready to
+                * be written back to disk
+                */
+               BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
+               BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
+               for (i = disks; i--; ) {
+                       dev = &sh->dev[i];
+                       if (test_bit(R5_LOCKED, &dev->flags) &&
+                           (i == sh->pd_idx || i == qd_idx ||
+                            dev->written)) {
+                               pr_debug("Writing block %d\n", i);
+                               BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
+                               set_bit(R5_Wantwrite, &dev->flags);
+                               if (!test_bit(R5_Insync, &dev->flags) ||
+                                   ((i == sh->pd_idx || i == qd_idx) &&
+                                     s.failed == 0))
+                                       set_bit(STRIPE_INSYNC, &sh->state);
+                       }
+               }
+               if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
+                       atomic_dec(&conf->preread_active_stripes);
+                       if (atomic_read(&conf->preread_active_stripes) <
+                               IO_THRESHOLD)
+                               md_wakeup_thread(conf->mddev->thread);
+               }
+       }
+
+       /* Now to consider new write requests and what else, if anything
+        * should be read.  We do not handle new writes when:
+        * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
+        * 2/ A 'check' operation is in flight, as it may clobber the parity
+        *    block.
+        */
+       if (s.to_write && !sh->reconstruct_state && !sh->check_state)
                handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
 
        /* maybe we need to check and possibly fix the parity for this stripe
         * Any reads will already have been scheduled, so we just see if enough
-        * data is available
+        * data is available.  The parity check is held off while parity
+        * dependent operations are in flight.
         */
-       if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
-               handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
+       if (sh->check_state ||
+           (s.syncing && s.locked == 0 &&
+            !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
+            !test_bit(STRIPE_INSYNC, &sh->state)))
+               handle_parity_checks6(conf, sh, &s, &r6s, disks);
 
        if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
                md_done_sync(conf->mddev, STRIPE_SECTORS,1);
@@ -3179,15 +3373,29 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
                                        set_bit(R5_Wantwrite, &dev->flags);
                                        set_bit(R5_ReWrite, &dev->flags);
                                        set_bit(R5_LOCKED, &dev->flags);
+                                       s.locked++;
                                } else {
                                        /* let's read it back */
                                        set_bit(R5_Wantread, &dev->flags);
                                        set_bit(R5_LOCKED, &dev->flags);
+                                       s.locked++;
                                }
                        }
                }
 
-       if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
+       /* Finish reconstruct operations initiated by the expansion process */
+       if (sh->reconstruct_state == reconstruct_state_result) {
+               sh->reconstruct_state = reconstruct_state_idle;
+               clear_bit(STRIPE_EXPANDING, &sh->state);
+               for (i = conf->raid_disks; i--; ) {
+                       set_bit(R5_Wantwrite, &sh->dev[i].flags);
+                       set_bit(R5_LOCKED, &sh->dev[i].flags);
+                       s.locked++;
+               }
+       }
+
+       if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
+           !sh->reconstruct_state) {
                struct stripe_head *sh2
                        = get_active_stripe(conf, sh->sector, 1, 1, 1);
                if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
@@ -3208,14 +3416,8 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
                /* Need to write out all blocks after computing P&Q */
                sh->disks = conf->raid_disks;
                stripe_set_idx(sh->sector, conf, 0, sh);
-               compute_parity6(sh, RECONSTRUCT_WRITE);
-               for (i = conf->raid_disks ; i-- ;  ) {
-                       set_bit(R5_LOCKED, &sh->dev[i].flags);
-                       s.locked++;
-                       set_bit(R5_Wantwrite, &sh->dev[i].flags);
-               }
-               clear_bit(STRIPE_EXPANDING, &sh->state);
-       } else if (s.expanded) {
+               schedule_reconstruction(sh, &s, 1, 1);
+       } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
                clear_bit(STRIPE_EXPAND_READY, &sh->state);
                atomic_dec(&conf->reshape_stripes);
                wake_up(&conf->wait_for_overlap);
@@ -3233,6 +3435,9 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
        if (unlikely(blocked_rdev))
                md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
 
+       if (s.ops_request)
+               raid_run_ops(sh, s.ops_request);
+
        ops_run_io(sh, &s);
 
        return_io(return_bi);
@@ -3241,16 +3446,14 @@ static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
 }
 
 /* returns true if the stripe was handled */
-static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
+static bool handle_stripe(struct stripe_head *sh)
 {
        if (sh->raid_conf->level == 6)
-               return handle_stripe6(sh, tmp_page);
+               return handle_stripe6(sh);
        else
                return handle_stripe5(sh);
 }
 
-
-
 static void raid5_activate_delayed(raid5_conf_t *conf)
 {
        if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
@@ -4046,7 +4249,7 @@ static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *ski
        spin_unlock(&sh->lock);
 
        /* wait for any blocked device to be handled */
-       while(unlikely(!handle_stripe(sh, NULL)))
+       while (unlikely(!handle_stripe(sh)))
                ;
        release_stripe(sh);
 
@@ -4103,7 +4306,7 @@ static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
                        return handled;
                }
 
-               handle_stripe(sh, NULL);
+               handle_stripe(sh);
                release_stripe(sh);
                handled++;
        }
@@ -4117,6 +4320,36 @@ static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
        return handled;
 }
 
+#ifdef CONFIG_MULTICORE_RAID456
+static void __process_stripe(void *param, async_cookie_t cookie)
+{
+       struct stripe_head *sh = param;
+
+       handle_stripe(sh);
+       release_stripe(sh);
+}
+
+static void process_stripe(struct stripe_head *sh, struct list_head *domain)
+{
+       async_schedule_domain(__process_stripe, sh, domain);
+}
+
+static void synchronize_stripe_processing(struct list_head *domain)
+{
+       async_synchronize_full_domain(domain);
+}
+#else
+static void process_stripe(struct stripe_head *sh, struct list_head *domain)
+{
+       handle_stripe(sh);
+       release_stripe(sh);
+       cond_resched();
+}
+
+static void synchronize_stripe_processing(struct list_head *domain)
+{
+}
+#endif
 
 
 /*
@@ -4131,6 +4364,7 @@ static void raid5d(mddev_t *mddev)
        struct stripe_head *sh;
        raid5_conf_t *conf = mddev_to_conf(mddev);
        int handled;
+       LIST_HEAD(raid_domain);
 
        pr_debug("+++ raid5d active\n");
 
@@ -4167,8 +4401,7 @@ static void raid5d(mddev_t *mddev)
                spin_unlock_irq(&conf->device_lock);
                
                handled++;
-               handle_stripe(sh, conf->spare_page);
-               release_stripe(sh);
+               process_stripe(sh, &raid_domain);
 
                spin_lock_irq(&conf->device_lock);
        }
@@ -4176,6 +4409,7 @@ static void raid5d(mddev_t *mddev)
 
        spin_unlock_irq(&conf->device_lock);
 
+       synchronize_stripe_processing(&raid_domain);
        async_tx_issue_pending_all();
        unplug_slaves(mddev);
 
@@ -4308,6 +4542,118 @@ raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
        return sectors * (raid_disks - conf->max_degraded);
 }
 
+static void raid5_free_percpu(raid5_conf_t *conf)
+{
+       struct raid5_percpu *percpu;
+       unsigned long cpu;
+
+       if (!conf->percpu)
+               return;
+
+       get_online_cpus();
+       for_each_possible_cpu(cpu) {
+               percpu = per_cpu_ptr(conf->percpu, cpu);
+               safe_put_page(percpu->spare_page);
+               kfree(percpu->scribble);
+       }
+#ifdef CONFIG_HOTPLUG_CPU
+       unregister_cpu_notifier(&conf->cpu_notify);
+#endif
+       put_online_cpus();
+
+       free_percpu(conf->percpu);
+}
+
+static void free_conf(raid5_conf_t *conf)
+{
+       shrink_stripes(conf);
+       raid5_free_percpu(conf);
+       kfree(conf->disks);
+       kfree(conf->stripe_hashtbl);
+       kfree(conf);
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
+                             void *hcpu)
+{
+       raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
+       long cpu = (long)hcpu;
+       struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
+
+       switch (action) {
+       case CPU_UP_PREPARE:
+       case CPU_UP_PREPARE_FROZEN:
+               if (conf->level == 6 && !percpu->spare_page)
+                       percpu->spare_page = alloc_page(GFP_KERNEL);
+               if (!percpu->scribble)
+                       percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
+
+               if (!percpu->scribble ||
+                   (conf->level == 6 && !percpu->spare_page)) {
+                       safe_put_page(percpu->spare_page);
+                       kfree(percpu->scribble);
+                       pr_err("%s: failed memory allocation for cpu%ld\n",
+                              __func__, cpu);
+                       return NOTIFY_BAD;
+               }
+               break;
+       case CPU_DEAD:
+       case CPU_DEAD_FROZEN:
+               safe_put_page(percpu->spare_page);
+               kfree(percpu->scribble);
+               percpu->spare_page = NULL;
+               percpu->scribble = NULL;
+               break;
+       default:
+               break;
+       }
+       return NOTIFY_OK;
+}
+#endif
+
+static int raid5_alloc_percpu(raid5_conf_t *conf)
+{
+       unsigned long cpu;
+       struct page *spare_page;
+       struct raid5_percpu *allcpus;
+       void *scribble;
+       int err;
+
+       allcpus = alloc_percpu(struct raid5_percpu);
+       if (!allcpus)
+               return -ENOMEM;
+       conf->percpu = allcpus;
+
+       get_online_cpus();
+       err = 0;
+       for_each_present_cpu(cpu) {
+               if (conf->level == 6) {
+                       spare_page = alloc_page(GFP_KERNEL);
+                       if (!spare_page) {
+                               err = -ENOMEM;
+                               break;
+                       }
+                       per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
+               }
+               scribble = kmalloc(scribble_len(conf->raid_disks), GFP_KERNEL);
+               if (!scribble) {
+                       err = -ENOMEM;
+                       break;
+               }
+               per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
+       }
+#ifdef CONFIG_HOTPLUG_CPU
+       conf->cpu_notify.notifier_call = raid456_cpu_notify;
+       conf->cpu_notify.priority = 0;
+       if (err == 0)
+               err = register_cpu_notifier(&conf->cpu_notify);
+#endif
+       put_online_cpus();
+
+       return err;
+}
+
 static raid5_conf_t *setup_conf(mddev_t *mddev)
 {
        raid5_conf_t *conf;
@@ -4347,6 +4693,7 @@ static raid5_conf_t *setup_conf(mddev_t *mddev)
                goto abort;
 
        conf->raid_disks = mddev->raid_disks;
+       conf->scribble_len = scribble_len(conf->raid_disks);
        if (mddev->reshape_position == MaxSector)
                conf->previous_raid_disks = mddev->raid_disks;
        else
@@ -4362,11 +4709,10 @@ static raid5_conf_t *setup_conf(mddev_t *mddev)
        if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
                goto abort;
 
-       if (mddev->new_level == 6) {
-               conf->spare_page = alloc_page(GFP_KERNEL);
-               if (!conf->spare_page)
-                       goto abort;
-       }
+       conf->level = mddev->new_level;
+       if (raid5_alloc_percpu(conf) != 0)
+               goto abort;
+
        spin_lock_init(&conf->device_lock);
        init_waitqueue_head(&conf->wait_for_stripe);
        init_waitqueue_head(&conf->wait_for_overlap);
@@ -4402,7 +4748,6 @@ static raid5_conf_t *setup_conf(mddev_t *mddev)
        }
 
        conf->chunk_size = mddev->new_chunk;
-       conf->level = mddev->new_level;
        if (conf->level == 6)
                conf->max_degraded = 2;
        else
@@ -4437,11 +4782,7 @@ static raid5_conf_t *setup_conf(mddev_t *mddev)
 
  abort:
        if (conf) {
-               shrink_stripes(conf);
-               safe_put_page(conf->spare_page);
-               kfree(conf->disks);
-               kfree(conf->stripe_hashtbl);
-               kfree(conf);
+               free_conf(conf);
                return ERR_PTR(-EIO);
        } else
                return ERR_PTR(-ENOMEM);
@@ -4607,12 +4948,8 @@ abort:
        md_unregister_thread(mddev->thread);
        mddev->thread = NULL;
        if (conf) {
-               shrink_stripes(conf);
                print_raid5_conf(conf);
-               safe_put_page(conf->spare_page);
-               kfree(conf->disks);
-               kfree(conf->stripe_hashtbl);
-               kfree(conf);
+               free_conf(conf);
        }
        mddev->private = NULL;
        printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
@@ -4627,13 +4964,10 @@ static int stop(mddev_t *mddev)
 
        md_unregister_thread(mddev->thread);
        mddev->thread = NULL;
-       shrink_stripes(conf);
-       kfree(conf->stripe_hashtbl);
        mddev->queue->backing_dev_info.congested_fn = NULL;
        blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
        sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
-       kfree(conf->disks);
-       kfree(conf);
+       free_conf(conf);
        mddev->private = NULL;
        return 0;
 }
index 52ba999..116d0b4 100644 (file)
@@ -2,6 +2,7 @@
 #define _RAID5_H
 
 #include <linux/raid/xor.h>
+#include <linux/dmaengine.h>
 
 /*
  *
  */
 enum check_states {
        check_state_idle = 0,
-       check_state_run, /* parity check */
+       check_state_run, /* xor parity check */
+       check_state_run_q, /* q-parity check */
+       check_state_run_pq, /* pq dual parity check */
        check_state_check_result,
        check_state_compute_run, /* parity repair */
        check_state_compute_result,
@@ -215,8 +218,8 @@ struct stripe_head {
         * @target - STRIPE_OP_COMPUTE_BLK target
         */
        struct stripe_operations {
-               int                target;
-               u32                zero_sum_result;
+               int                  target, target2;
+               enum sum_check_flags zero_sum_result;
        } ops;
        struct r5dev {
                struct bio      req;
@@ -298,7 +301,7 @@ struct r6_state {
 #define STRIPE_OP_COMPUTE_BLK  1
 #define STRIPE_OP_PREXOR       2
 #define STRIPE_OP_BIODRAIN     3
-#define STRIPE_OP_POSTXOR      4
+#define STRIPE_OP_RECONSTRUCT  4
 #define STRIPE_OP_CHECK        5
 
 /*
@@ -383,8 +386,21 @@ struct raid5_private_data {
                                            * (fresh device added).
                                            * Cleared when a sync completes.
                                            */
-
-       struct page             *spare_page; /* Used when checking P/Q in raid6 */
+       /* per cpu variables */
+       struct raid5_percpu {
+               struct page     *spare_page; /* Used when checking P/Q in raid6 */
+               void            *scribble;   /* space for constructing buffer
+                                             * lists and performing address
+                                             * conversions
+                                             */
+       } *percpu;
+       size_t                  scribble_len; /* size of scribble region must be
+                                              * associated with conf to handle
+                                              * cpu hotplug while reshaping
+                                              */
+#ifdef CONFIG_HOTPLUG_CPU
+       struct notifier_block   cpu_notify;
+#endif
 
        /*
         * Free stripes pool
index 5fc2ef8..866e61c 100644 (file)
@@ -58,25 +58,57 @@ struct dma_chan_ref {
  * array.
  * @ASYNC_TX_ACK: immediately ack the descriptor, precludes setting up a
  * dependency chain
- * @ASYNC_TX_DEP_ACK: ack the dependency descriptor.  Useful for chaining.
  */
 enum async_tx_flags {
        ASYNC_TX_XOR_ZERO_DST    = (1 << 0),
        ASYNC_TX_XOR_DROP_DST    = (1 << 1),
-       ASYNC_TX_ACK             = (1 << 3),
-       ASYNC_TX_DEP_ACK         = (1 << 4),
+       ASYNC_TX_ACK             = (1 << 2),
+};
+
+/**
+ * struct async_submit_ctl - async_tx submission/completion modifiers
+ * @flags: submission modifiers
+ * @depend_tx: parent dependency of the current operation being submitted
+ * @cb_fn: callback routine to run at operation completion
+ * @cb_param: parameter for the callback routine
+ * @scribble: caller provided space for dma/page address conversions
+ */
+struct async_submit_ctl {
+       enum async_tx_flags flags;
+       struct dma_async_tx_descriptor *depend_tx;
+       dma_async_tx_callback cb_fn;
+       void *cb_param;
+       void *scribble;
 };
 
 #ifdef CONFIG_DMA_ENGINE
 #define async_tx_issue_pending_all dma_issue_pending_all
+
+/**
+ * async_tx_issue_pending - send pending descriptor to the hardware channel
+ * @tx: descriptor handle to retrieve hardware context
+ *
+ * Note: any dependent operations will have already been issued by
+ * async_tx_channel_switch, or (in the case of no channel switch) will
+ * be already pending on this channel.
+ */
+static inline void async_tx_issue_pending(struct dma_async_tx_descriptor *tx)
+{
+       if (likely(tx)) {
+               struct dma_chan *chan = tx->chan;
+               struct dma_device *dma = chan->device;
+
+               dma->device_issue_pending(chan);
+       }
+}
 #ifdef CONFIG_ARCH_HAS_ASYNC_TX_FIND_CHANNEL
 #include <asm/async_tx.h>
 #else
 #define async_tx_find_channel(dep, type, dst, dst_count, src, src_count, len) \
         __async_tx_find_channel(dep, type)
 struct dma_chan *
-__async_tx_find_channel(struct dma_async_tx_descriptor *depend_tx,
-       enum dma_transaction_type tx_type);
+__async_tx_find_channel(struct async_submit_ctl *submit,
+                       enum dma_transaction_type tx_type);
 #endif /* CONFIG_ARCH_HAS_ASYNC_TX_FIND_CHANNEL */
 #else
 static inline void async_tx_issue_pending_all(void)
@@ -84,10 +116,16 @@ static inline void async_tx_issue_pending_all(void)
        do { } while (0);
 }
 
+static inline void async_tx_issue_pending(struct dma_async_tx_descriptor *tx)
+{
+       do { } while (0);
+}
+
 static inline struct dma_chan *
-async_tx_find_channel(struct dma_async_tx_descriptor *depend_tx,
-       enum dma_transaction_type tx_type, struct page **dst, int dst_count,
-       struct page **src, int src_count, size_t len)
+async_tx_find_channel(struct async_submit_ctl *submit,
+                     enum dma_transaction_type tx_type, struct page **dst,
+                     int dst_count, struct page **src, int src_count,
+                     size_t len)
 {
        return NULL;
 }
@@ -99,46 +137,70 @@ async_tx_find_channel(struct dma_async_tx_descriptor *depend_tx,
  * @cb_fn_param: parameter to pass to the callback routine
  */
 static inline void
-async_tx_sync_epilog(dma_async_tx_callback cb_fn, void *cb_fn_param)
+async_tx_sync_epilog(struct async_submit_ctl *submit)
 {
-       if (cb_fn)
-               cb_fn(cb_fn_param);
+       if (submit->cb_fn)
+               submit->cb_fn(submit->cb_param);
 }
 
-void
-async_tx_submit(struct dma_chan *chan, struct dma_async_tx_descriptor *tx,
-       enum async_tx_flags flags, struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_fn_param);
+typedef union {
+       unsigned long addr;
+       struct page *page;
+       dma_addr_t dma;
+} addr_conv_t;
+
+static inline void
+init_async_submit(struct async_submit_ctl *args, enum async_tx_flags flags,
+                 struct dma_async_tx_descriptor *tx,
+                 dma_async_tx_callback cb_fn, void *cb_param,
+                 addr_conv_t *scribble)
+{
+       args->flags = flags;
+       args->depend_tx = tx;
+       args->cb_fn = cb_fn;
+       args->cb_param = cb_param;
+       args->scribble = scribble;
+}
+
+void async_tx_submit(struct dma_chan *chan, struct dma_async_tx_descriptor *tx,
+                    struct async_submit_ctl *submit);
 
 struct dma_async_tx_descriptor *
 async_xor(struct page *dest, struct page **src_list, unsigned int offset,
-       int src_cnt, size_t len, enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_fn_param);
+         int src_cnt, size_t len, struct async_submit_ctl *submit);
 
 struct dma_async_tx_descriptor *
-async_xor_zero_sum(struct page *dest, struct page **src_list,
-       unsigned int offset, int src_cnt, size_t len,
-       u32 *result, enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_fn_param);
+async_xor_val(struct page *dest, struct page **src_list, unsigned int offset,
+             int src_cnt, size_t len, enum sum_check_flags *result,
+             struct async_submit_ctl *submit);
 
 struct dma_async_tx_descriptor *
 async_memcpy(struct page *dest, struct page *src, unsigned int dest_offset,
-       unsigned int src_offset, size_t len, enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_fn_param);
+            unsigned int src_offset, size_t len,
+            struct async_submit_ctl *submit);
 
 struct dma_async_tx_descriptor *
 async_memset(struct page *dest, int val, unsigned int offset,
-       size_t len, enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_fn_param);
+            size_t len, struct async_submit_ctl *submit);
+
+struct dma_async_tx_descriptor *async_trigger_callback(struct async_submit_ctl *submit);
+
+struct dma_async_tx_descriptor *
+async_gen_syndrome(struct page **blocks, unsigned int offset, int src_cnt,
+                  size_t len, struct async_submit_ctl *submit);
+
+struct dma_async_tx_descriptor *
+async_syndrome_val(struct page **blocks, unsigned int offset, int src_cnt,
+                  size_t len, enum sum_check_flags *pqres, struct page *spare,
+                  struct async_submit_ctl *submit);
+
+struct dma_async_tx_descriptor *
+async_raid6_2data_recov(int src_num, size_t bytes, int faila, int failb,
+                       struct page **ptrs, struct async_submit_ctl *submit);
 
 struct dma_async_tx_descriptor *
-async_trigger_callback(enum async_tx_flags flags,
-       struct dma_async_tx_descriptor *depend_tx,
-       dma_async_tx_callback cb_fn, void *cb_fn_param);
+async_raid6_datap_recov(int src_num, size_t bytes, int faila,
+                       struct page **ptrs, struct async_submit_ctl *submit);
 
 void async_tx_quiesce(struct dma_async_tx_descriptor **tx);
 #endif /* _ASYNC_TX_H_ */
index ffefba8..1012f1a 100644 (file)
@@ -52,11 +52,11 @@ enum dma_status {
 enum dma_transaction_type {
        DMA_MEMCPY,
        DMA_XOR,
-       DMA_PQ_XOR,
+       DMA_PQ,
        DMA_DUAL_XOR,
        DMA_PQ_UPDATE,
-       DMA_ZERO_SUM,
-       DMA_PQ_ZERO_SUM,
+       DMA_XOR_VAL,
+       DMA_PQ_VAL,
        DMA_MEMSET,
        DMA_MEMCPY_CRC32C,
        DMA_INTERRUPT,
@@ -70,18 +70,23 @@ enum dma_transaction_type {
 
 /**
  * enum dma_ctrl_flags - DMA flags to augment operation preparation,
- *     control completion, and communicate status.
+ *  control completion, and communicate status.
  * @DMA_PREP_INTERRUPT - trigger an interrupt (callback) upon completion of
- *     this transaction
+ *  this transaction
  * @DMA_CTRL_ACK - the descriptor cannot be reused until the client
- *     acknowledges receipt, i.e. has has a chance to establish any
- *     dependency chains
+ *  acknowledges receipt, i.e. has has a chance to establish any dependency
+ *  chains
  * @DMA_COMPL_SKIP_SRC_UNMAP - set to disable dma-unmapping the source buffer(s)
  * @DMA_COMPL_SKIP_DEST_UNMAP - set to disable dma-unmapping the destination(s)
  * @DMA_COMPL_SRC_UNMAP_SINGLE - set to do the source dma-unmapping as single
  *     (if not set, do the source dma-unmapping as page)
  * @DMA_COMPL_DEST_UNMAP_SINGLE - set to do the destination dma-unmapping as single
  *     (if not set, do the destination dma-unmapping as page)
+ * @DMA_PREP_PQ_DISABLE_P - prevent generation of P while generating Q
+ * @DMA_PREP_PQ_DISABLE_Q - prevent generation of Q while generating P
+ * @DMA_PREP_CONTINUE - indicate to a driver that it is reusing buffers as
+ *  sources that were the result of a previous operation, in the case of a PQ
+ *  operation it continues the calculation with new sources
  */
 enum dma_ctrl_flags {
        DMA_PREP_INTERRUPT = (1 << 0),
@@ -90,9 +95,31 @@ enum dma_ctrl_flags {
        DMA_COMPL_SKIP_DEST_UNMAP = (1 << 3),
        DMA_COMPL_SRC_UNMAP_SINGLE = (1 << 4),
        DMA_COMPL_DEST_UNMAP_SINGLE = (1 << 5),
+       DMA_PREP_PQ_DISABLE_P = (1 << 6),
+       DMA_PREP_PQ_DISABLE_Q = (1 << 7),
+       DMA_PREP_CONTINUE = (1 << 8),
 };
 
 /**
+ * enum sum_check_bits - bit position of pq_check_flags
+ */
+enum sum_check_bits {
+       SUM_CHECK_P = 0,
+       SUM_CHECK_Q = 1,
+};
+
+/**
+ * enum pq_check_flags - result of async_{xor,pq}_zero_sum operations
+ * @SUM_CHECK_P_RESULT - 1 if xor zero sum error, 0 otherwise
+ * @SUM_CHECK_Q_RESULT - 1 if reed-solomon zero sum error, 0 otherwise
+ */
+enum sum_check_flags {
+       SUM_CHECK_P_RESULT = (1 << SUM_CHECK_P),
+       SUM_CHECK_Q_RESULT = (1 << SUM_CHECK_Q),
+};
+
+
+/**
  * dma_cap_mask_t - capabilities bitmap modeled after cpumask_t.
  * See linux/cpumask.h
  */
@@ -213,6 +240,7 @@ struct dma_async_tx_descriptor {
  * @global_node: list_head for global dma_device_list
  * @cap_mask: one or more dma_capability flags
  * @max_xor: maximum number of xor sources, 0 if no capability
+ * @max_pq: maximum number of PQ sources and PQ-continue capability
  * @dev_id: unique device ID
  * @dev: struct device reference for dma mapping api
  * @device_alloc_chan_resources: allocate resources and return the
@@ -220,7 +248,9 @@ struct dma_async_tx_descriptor {
  * @device_free_chan_resources: release DMA channel's resources
  * @device_prep_dma_memcpy: prepares a memcpy operation
  * @device_prep_dma_xor: prepares a xor operation
- * @device_prep_dma_zero_sum: prepares a zero_sum operation
+ * @device_prep_dma_xor_val: prepares a xor validation operation
+ * @device_prep_dma_pq: prepares a pq operation
+ * @device_prep_dma_pq_val: prepares a pqzero_sum operation
  * @device_prep_dma_memset: prepares a memset operation
  * @device_prep_dma_interrupt: prepares an end of chain interrupt operation
  * @device_prep_slave_sg: prepares a slave dma operation
@@ -235,7 +265,9 @@ struct dma_device {
        struct list_head channels;
        struct list_head global_node;
        dma_cap_mask_t  cap_mask;
-       int max_xor;
+       unsigned short max_xor;
+       unsigned short max_pq;
+       #define DMA_HAS_PQ_CONTINUE (1 << 15)
 
        int dev_id;
        struct device *dev;
@@ -249,9 +281,17 @@ struct dma_device {
        struct dma_async_tx_descriptor *(*device_prep_dma_xor)(
                struct dma_chan *chan, dma_addr_t dest, dma_addr_t *src,
                unsigned int src_cnt, size_t len, unsigned long flags);
-       struct dma_async_tx_descriptor *(*device_prep_dma_zero_sum)(
+       struct dma_async_tx_descriptor *(*device_prep_dma_xor_val)(
                struct dma_chan *chan, dma_addr_t *src, unsigned int src_cnt,
-               size_t len, u32 *result, unsigned long flags);
+               size_t len, enum sum_check_flags *result, unsigned long flags);
+       struct dma_async_tx_descriptor *(*device_prep_dma_pq)(
+               struct dma_chan *chan, dma_addr_t *dst, dma_addr_t *src,
+               unsigned int src_cnt, const unsigned char *scf,
+               size_t len, unsigned long flags);
+       struct dma_async_tx_descriptor *(*device_prep_dma_pq_val)(
+               struct dma_chan *chan, dma_addr_t *pq, dma_addr_t *src,
+               unsigned int src_cnt, const unsigned char *scf, size_t len,
+               enum sum_check_flags *pqres, unsigned long flags);
        struct dma_async_tx_descriptor *(*device_prep_dma_memset)(
                struct dma_chan *chan, dma_addr_t dest, int value, size_t len,
                unsigned long flags);
@@ -270,6 +310,60 @@ struct dma_device {
        void (*device_issue_pending)(struct dma_chan *chan);
 };
 
+static inline void
+dma_set_maxpq(struct dma_device *dma, int maxpq, int has_pq_continue)
+{
+       dma->max_pq = maxpq;
+       if (has_pq_continue)
+               dma->max_pq |= DMA_HAS_PQ_CONTINUE;
+}
+
+static inline bool dmaf_continue(enum dma_ctrl_flags flags)
+{
+       return (flags & DMA_PREP_CONTINUE) == DMA_PREP_CONTINUE;
+}
+
+static inline bool dmaf_p_disabled_continue(enum dma_ctrl_flags flags)
+{
+       enum dma_ctrl_flags mask = DMA_PREP_CONTINUE | DMA_PREP_PQ_DISABLE_P;
+
+       return (flags & mask) == mask;
+}
+
+static inline bool dma_dev_has_pq_continue(struct dma_device *dma)
+{
+       return (dma->max_pq & DMA_HAS_PQ_CONTINUE) == DMA_HAS_PQ_CONTINUE;
+}
+
+static unsigned short dma_dev_to_maxpq(struct dma_device *dma)
+{
+       return dma->max_pq & ~DMA_HAS_PQ_CONTINUE;
+}
+
+/* dma_maxpq - reduce maxpq in the face of continued operations
+ * @dma - dma device with PQ capability
+ * @flags - to check if DMA_PREP_CONTINUE and DMA_PREP_PQ_DISABLE_P are set
+ *
+ * When an engine does not support native continuation we need 3 extra
+ * source slots to reuse P and Q with the following coefficients:
+ * 1/ {00} * P : remove P from Q', but use it as a source for P'
+ * 2/ {01} * Q : use Q to continue Q' calculation
+ * 3/ {00} * Q : subtract Q from P' to cancel (2)
+ *
+ * In the case where P is disabled we only need 1 extra source:
+ * 1/ {01} * Q : use Q to continue Q' calculation
+ */
+static inline int dma_maxpq(struct dma_device *dma, enum dma_ctrl_flags flags)
+{
+       if (dma_dev_has_pq_continue(dma) || !dmaf_continue(flags))
+               return dma_dev_to_maxpq(dma);
+       else if (dmaf_p_disabled_continue(flags))
+               return dma_dev_to_maxpq(dma) - 1;
+       else if (dmaf_continue(flags))
+               return dma_dev_to_maxpq(dma) - 3;
+       BUG();
+}
+
 /* --- public DMA engine API --- */
 
 #ifdef CONFIG_DMA_ENGINE