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

1  2 
arch/arm/mach-iop13xx/setup.c
arch/arm/plat-iop/adma.c
drivers/dma/dmaengine.c
drivers/dma/dmatest.c
drivers/md/raid5.c
include/linux/dmaengine.h

@@@ -307,7 -307,7 +307,7 @@@ static struct resource iop13xx_adma_2_r
        }
  };
  
 -static u64 iop13xx_adma_dmamask = DMA_64BIT_MASK;
 +static u64 iop13xx_adma_dmamask = DMA_BIT_MASK(64);
  static struct iop_adma_platform_data iop13xx_adma_0_data = {
        .hw_id = 0,
        .pool_size = PAGE_SIZE,
@@@ -331,7 -331,7 +331,7 @@@ static struct platform_device iop13xx_a
        .resource = iop13xx_adma_0_resources,
        .dev = {
                .dma_mask = &iop13xx_adma_dmamask,
 -              .coherent_dma_mask = DMA_64BIT_MASK,
 +              .coherent_dma_mask = DMA_BIT_MASK(64),
                .platform_data = (void *) &iop13xx_adma_0_data,
        },
  };
@@@ -343,7 -343,7 +343,7 @@@ static struct platform_device iop13xx_a
        .resource = iop13xx_adma_1_resources,
        .dev = {
                .dma_mask = &iop13xx_adma_dmamask,
 -              .coherent_dma_mask = DMA_64BIT_MASK,
 +              .coherent_dma_mask = DMA_BIT_MASK(64),
                .platform_data = (void *) &iop13xx_adma_1_data,
        },
  };
@@@ -355,7 -355,7 +355,7 @@@ static struct platform_device iop13xx_a
        .resource = iop13xx_adma_2_resources,
        .dev = {
                .dma_mask = &iop13xx_adma_dmamask,
 -              .coherent_dma_mask = DMA_64BIT_MASK,
 +              .coherent_dma_mask = DMA_BIT_MASK(64),
                .platform_data = (void *) &iop13xx_adma_2_data,
        },
  };
@@@ -478,7 -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);
                        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_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;
                }
        }
diff --combined arch/arm/plat-iop/adma.c
@@@ -119,7 -119,7 +119,7 @@@ static struct resource iop3xx_aau_resou
        }
  };
  
 -static u64 iop3xx_adma_dmamask = DMA_32BIT_MASK;
 +static u64 iop3xx_adma_dmamask = DMA_BIT_MASK(32);
  
  static struct iop_adma_platform_data iop3xx_dma_0_data = {
        .hw_id = DMA0_ID,
@@@ -143,7 -143,7 +143,7 @@@ struct platform_device iop3xx_dma_0_cha
        .resource = iop3xx_dma_0_resources,
        .dev = {
                .dma_mask = &iop3xx_adma_dmamask,
 -              .coherent_dma_mask = DMA_64BIT_MASK,
 +              .coherent_dma_mask = DMA_BIT_MASK(64),
                .platform_data = (void *) &iop3xx_dma_0_data,
        },
  };
@@@ -155,7 -155,7 +155,7 @@@ struct platform_device iop3xx_dma_1_cha
        .resource = iop3xx_dma_1_resources,
        .dev = {
                .dma_mask = &iop3xx_adma_dmamask,
 -              .coherent_dma_mask = DMA_64BIT_MASK,
 +              .coherent_dma_mask = DMA_BIT_MASK(64),
                .platform_data = (void *) &iop3xx_dma_1_data,
        },
  };
@@@ -167,7 -167,7 +167,7 @@@ struct platform_device iop3xx_aau_chann
        .resource = iop3xx_aau_resources,
        .dev = {
                .dma_mask = &iop3xx_adma_dmamask,
 -              .coherent_dma_mask = DMA_64BIT_MASK,
 +              .coherent_dma_mask = DMA_BIT_MASK(64),
                .platform_data = (void *) &iop3xx_aau_data,
        },
  };
@@@ -198,7 -198,7 +198,7 @@@ static int __init iop3xx_adma_cap_init(
        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
diff --combined drivers/dma/dmaengine.c
@@@ -644,8 -644,12 +644,12 @@@ int dma_async_device_register(struct dm
                !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) &&
@@@ -804,14 -808,11 +808,14 @@@ dma_async_memcpy_buf_to_buf(struct dma_
        dma_addr_t dma_dest, dma_src;
        dma_cookie_t cookie;
        int cpu;
 +      unsigned long flags;
  
        dma_src = dma_map_single(dev->dev, src, len, DMA_TO_DEVICE);
        dma_dest = dma_map_single(dev->dev, dest, len, DMA_FROM_DEVICE);
 -      tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
 -                                       DMA_CTRL_ACK);
 +      flags = DMA_CTRL_ACK |
 +              DMA_COMPL_SRC_UNMAP_SINGLE |
 +              DMA_COMPL_DEST_UNMAP_SINGLE;
 +      tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len, flags);
  
        if (!tx) {
                dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
@@@ -853,12 -854,11 +857,12 @@@ dma_async_memcpy_buf_to_pg(struct dma_c
        dma_addr_t dma_dest, dma_src;
        dma_cookie_t cookie;
        int cpu;
 +      unsigned long flags;
  
        dma_src = dma_map_single(dev->dev, kdata, len, DMA_TO_DEVICE);
        dma_dest = dma_map_page(dev->dev, page, offset, len, DMA_FROM_DEVICE);
 -      tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
 -                                       DMA_CTRL_ACK);
 +      flags = DMA_CTRL_ACK | DMA_COMPL_SRC_UNMAP_SINGLE;
 +      tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len, flags);
  
        if (!tx) {
                dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
@@@ -902,13 -902,12 +906,13 @@@ dma_async_memcpy_pg_to_pg(struct dma_ch
        dma_addr_t dma_dest, dma_src;
        dma_cookie_t cookie;
        int cpu;
 +      unsigned long flags;
  
        dma_src = dma_map_page(dev->dev, src_pg, src_off, len, DMA_TO_DEVICE);
        dma_dest = dma_map_page(dev->dev, dest_pg, dest_off, len,
                                DMA_FROM_DEVICE);
 -      tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
 -                                       DMA_CTRL_ACK);
 +      flags = DMA_CTRL_ACK;
 +      tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len, flags);
  
        if (!tx) {
                dma_unmap_page(dev->dev, dma_src, len, DMA_TO_DEVICE);
@@@ -939,49 -938,24 +943,24 @@@ EXPORT_SYMBOL(dma_async_tx_descriptor_i
  
  /* 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);
  
diff --combined drivers/dma/dmatest.c
@@@ -43,6 -43,11 +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 +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;
        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;
  
                                                      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 +466,8 @@@ static int dmatest_add_threads(struct d
                op = "copy";
        else if (type == DMA_XOR)
                op = "xor";
+       else if (type == DMA_PQ)
+               op = "pq";
        else
                return -EINVAL;
  
@@@ -501,6 -523,10 +523,10 @@@ static int dmatest_add_channel(struct d
                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));
@@@ -531,7 -557,9 +557,7 @@@ static int __init dmatest_init(void
                chan = dma_request_channel(mask, filter, NULL);
                if (chan) {
                        err = dmatest_add_channel(chan);
 -                      if (err == 0)
 -                              continue;
 -                      else {
 +                      if (err) {
                                dma_release_channel(chan);
                                break; /* add_channel failed, punt */
                        }
diff --combined drivers/md/raid5.c
@@@ -47,7 -47,9 +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"
@@@ -362,7 -364,7 +364,7 @@@ static void raid5_unplug_device(struct 
  
  static struct stripe_head *
  get_active_stripe(raid5_conf_t *conf, sector_t sector,
 -                int previous, int noblock)
 +                int previous, int noblock, int noquiesce)
  {
        struct stripe_head *sh;
  
  
        do {
                wait_event_lock_irq(conf->wait_for_stripe,
 -                                  conf->quiesce == 0,
 +                                  conf->quiesce == 0 || noquiesce,
                                    conf->device_lock, /* nothing */);
                sh = __find_stripe(conf, sector, conf->generation - previous);
                if (!sh) {
@@@ -499,11 -501,14 +501,14 @@@ async_copy_data(int frombio, struct bi
        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;
                        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 +596,7 @@@ static void ops_run_biofill(struct stri
  {
        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__,
        }
  
        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;
        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",
  
        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;
  }
  
  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;
                        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 +958,21 @@@ ops_run_biodrain(struct stripe_head *sh
        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);
        }
  
  }
  
  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;
         * 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)
        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--; ) {
                        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 +1239,28 @@@ static int grow_stripes(raid5_conf_t *c
        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.
        struct stripe_head *osh, *nsh;
        LIST_HEAD(newstripes);
        struct disk_info *ndisks;
+       unsigned long cpu;
        int err;
        struct kmem_cache *sc;
        int i;
        /* 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) {
        } 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 +1928,13 @@@ static sector_t compute_blocknr(struct 
  }
  
  
- /*
-  * 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
                } 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];
                                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];
                }
        }
  
-       /* 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 +2085,6 @@@ static int add_stripe_bio(struct stripe
  
  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 +2225,10 @@@ static int fetch_block5(struct stripe_h
                        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 +2267,104 @@@ static void handle_stripe_fill5(struct 
        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 +2498,61 @@@ static void handle_stripe_dirtying5(rai
         */
        /* 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);
                        }
                }
        }
        /* 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 +2611,7 @@@ static void handle_parity_checks5(raid5
                 * 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)
                         */
                                set_bit(R5_Wantcompute,
                                        &sh->dev[sh->pd_idx].flags);
                                sh->ops.target = sh->pd_idx;
+                               sh->ops.target2 = -1;
                                s->uptodate++;
                        }
                }
  
  
  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++;
                        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);
                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();
        }
  }
  
@@@ -2667,11 -2821,12 +2821,12 @@@ static void handle_stripe_expansion(rai
                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,
                                                          &dd_idx, NULL);
 -                      sh2 = get_active_stripe(conf, s, 0, 1);
 +                      sh2 = get_active_stripe(conf, s, 0, 1, 1);
                        if (sh2 == NULL)
                                /* so far only the early blocks of this stripe
                                 * have been requested.  When later blocks
                        }
  
                        /* 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);
@@@ -2944,7 -3100,7 +3100,7 @@@ static bool handle_stripe5(struct strip
        /* Finish reconstruct operations initiated by the expansion process */
        if (sh->reconstruct_state == reconstruct_state_result) {
                struct stripe_head *sh2
 -                      = get_active_stripe(conf, sh->sector, 1, 1);
 +                      = get_active_stripe(conf, sh->sector, 1, 1, 1);
                if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
                        /* sh cannot be written until sh2 has been read.
                         * so arrange for sh to be delayed a little
                /* 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);
                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);
  
        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;
        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);
  
                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++;
                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,
         * 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);
                                        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);
 +                      = get_active_stripe(conf, sh->sector, 1, 1, 1);
                if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
                        /* sh cannot be written until sh2 has been read.
                         * so arrange for sh to be delayed a little
                /* 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);
        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);
  }
  
  /* 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) {
@@@ -3288,7 -3491,7 +3491,7 @@@ static void unplug_slaves(mddev_t *mdde
        int i;
  
        rcu_read_lock();
 -      for (i=0; i<mddev->raid_disks; i++) {
 +      for (i = 0; i < conf->raid_disks; i++) {
                mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
                if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
                        struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
@@@ -3675,7 -3878,7 +3878,7 @@@ static int make_request(struct request_
                        (unsigned long long)logical_sector);
  
                sh = get_active_stripe(conf, new_sector, previous,
 -                                     (bi->bi_rw&RWA_MASK));
 +                                     (bi->bi_rw&RWA_MASK), 0);
                if (sh) {
                        if (unlikely(previous)) {
                                /* expansion might have moved on while waiting for a
@@@ -3811,13 -4014,13 +4014,13 @@@ static sector_t reshape_request(mddev_
        safepos = conf->reshape_safe;
        sector_div(safepos, data_disks);
        if (mddev->delta_disks < 0) {
 -              writepos -= reshape_sectors;
 +              writepos -= min_t(sector_t, reshape_sectors, writepos);
                readpos += reshape_sectors;
                safepos += reshape_sectors;
        } else {
                writepos += reshape_sectors;
 -              readpos -= reshape_sectors;
 -              safepos -= reshape_sectors;
 +              readpos -= min_t(sector_t, reshape_sectors, readpos);
 +              safepos -= min_t(sector_t, reshape_sectors, safepos);
        }
  
        /* 'writepos' is the most advanced device address we might write.
                wait_event(conf->wait_for_overlap,
                           atomic_read(&conf->reshape_stripes)==0);
                mddev->reshape_position = conf->reshape_progress;
 +              mddev->curr_resync_completed = mddev->curr_resync;
                conf->reshape_checkpoint = jiffies;
                set_bit(MD_CHANGE_DEVS, &mddev->flags);
                md_wakeup_thread(mddev->thread);
                conf->reshape_safe = mddev->reshape_position;
                spin_unlock_irq(&conf->device_lock);
                wake_up(&conf->wait_for_overlap);
 +              sysfs_notify(&mddev->kobj, NULL, "sync_completed");
        }
  
        if (mddev->delta_disks < 0) {
        for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
                int j;
                int skipped = 0;
 -              sh = get_active_stripe(conf, stripe_addr+i, 0, 0);
 +              sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
                set_bit(STRIPE_EXPANDING, &sh->state);
                atomic_inc(&conf->reshape_stripes);
                /* If any of this stripe is beyond the end of the old
                raid5_compute_sector(conf, stripe_addr*(new_data_disks),
                                     1, &dd_idx, NULL);
        last_sector =
 -              raid5_compute_sector(conf, ((stripe_addr+conf->chunk_size/512)
 +              raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
                                            *(new_data_disks) - 1),
                                     1, &dd_idx, NULL);
        if (last_sector >= mddev->dev_sectors)
                last_sector = mddev->dev_sectors - 1;
        while (first_sector <= last_sector) {
 -              sh = get_active_stripe(conf, first_sector, 1, 0);
 +              sh = get_active_stripe(conf, first_sector, 1, 0, 1);
                set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
                set_bit(STRIPE_HANDLE, &sh->state);
                release_stripe(sh);
         * then we need to write out the superblock.
         */
        sector_nr += reshape_sectors;
 -      if (sector_nr >= mddev->resync_max) {
 +      if ((sector_nr - mddev->curr_resync_completed) * 2
 +          >= mddev->resync_max - mddev->curr_resync_completed) {
                /* Cannot proceed until we've updated the superblock... */
                wait_event(conf->wait_for_overlap,
                           atomic_read(&conf->reshape_stripes) == 0);
                mddev->reshape_position = conf->reshape_progress;
 +              mddev->curr_resync_completed = mddev->curr_resync;
                conf->reshape_checkpoint = jiffies;
                set_bit(MD_CHANGE_DEVS, &mddev->flags);
                md_wakeup_thread(mddev->thread);
                conf->reshape_safe = mddev->reshape_position;
                spin_unlock_irq(&conf->device_lock);
                wake_up(&conf->wait_for_overlap);
 +              sysfs_notify(&mddev->kobj, NULL, "sync_completed");
        }
        return reshape_sectors;
  }
@@@ -4022,9 -4220,9 +4225,9 @@@ static inline sector_t sync_request(mdd
  
        bitmap_cond_end_sync(mddev->bitmap, sector_nr);
  
 -      sh = get_active_stripe(conf, sector_nr, 0, 1);
 +      sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
        if (sh == NULL) {
 -              sh = get_active_stripe(conf, sector_nr, 0, 0);
 +              sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
                /* make sure we don't swamp the stripe cache if someone else
                 * is trying to get access
                 */
         * We don't need to check the 'failed' flag as when that gets set,
         * recovery aborts.
         */
 -      for (i=0; i<mddev->raid_disks; i++)
 +      for (i = 0; i < conf->raid_disks; i++)
                if (conf->disks[i].rdev == NULL)
                        still_degraded = 1;
  
        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);
  
@@@ -4086,7 -4284,7 +4289,7 @@@ static int  retry_aligned_read(raid5_co
                        /* already done this stripe */
                        continue;
  
 -              sh = get_active_stripe(conf, sector, 0, 1);
 +              sh = get_active_stripe(conf, sector, 0, 1, 0);
  
                if (!sh) {
                        /* failed to get a stripe - must wait */
                        return handled;
                }
  
-               handle_stripe(sh, NULL);
+               handle_stripe(sh);
                release_stripe(sh);
                handled++;
        }
        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 -4359,7 +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");
  
                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);
        }
  
        spin_unlock_irq(&conf->device_lock);
  
+       synchronize_stripe_processing(&raid_domain);
        async_tx_issue_pending_all();
        unplug_slaves(mddev);
  
@@@ -4308,6 -4537,118 +4542,118 @@@ raid5_size(mddev_t *mddev, sector_t sec
        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;
                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
        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);
        }
  
        conf->chunk_size = mddev->new_chunk;
-       conf->level = mddev->new_level;
        if (conf->level == 6)
                conf->max_degraded = 2;
        else
  
   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 -4943,8 +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 -4959,10 +4964,10 @@@ static int stop(mddev_t *mddev
  
        md_unregister_thread(mddev->thread);
        mddev->thread = NULL;
        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;
  }
@@@ -52,11 -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,
  
  /**
   * 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),
        DMA_CTRL_ACK = (1 << 1),
        DMA_COMPL_SKIP_SRC_UNMAP = (1 << 2),
        DMA_COMPL_SKIP_DEST_UNMAP = (1 << 3),
 -      DMA_PREP_PQ_DISABLE_P = (1 << 4),
 -      DMA_PREP_PQ_DISABLE_Q = (1 << 5),
 -      DMA_PREP_CONTINUE = (1 << 6),
 +      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 -234,7 +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
   * @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 -259,9 +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;
        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);
        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