[I/OAT]: Utility functions for offloading sk_buff to iovec copies
Chris Leech [Wed, 24 May 2006 00:50:37 +0000 (17:50 -0700)]
Provides for pinning user space pages in memory, copying to iovecs,
and copying from sk_buffs including fragmented and chained sk_buffs.

Signed-off-by: Chris Leech <christopher.leech@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/dma/Makefile
drivers/dma/iovlock.c [new file with mode: 0644]
include/linux/dmaengine.h
include/net/netdma.h
net/core/Makefile
net/core/user_dma.c [new file with mode: 0644]

index c8a5f56..bdcfdbd 100644 (file)
@@ -1,2 +1,3 @@
-obj-y += dmaengine.o
+obj-$(CONFIG_DMA_ENGINE) += dmaengine.o
+obj-$(CONFIG_NET_DMA) += iovlock.o
 obj-$(CONFIG_INTEL_IOATDMA) += ioatdma.o
diff --git a/drivers/dma/iovlock.c b/drivers/dma/iovlock.c
new file mode 100644 (file)
index 0000000..5ed327e
--- /dev/null
@@ -0,0 +1,301 @@
+/*
+ * Copyright(c) 2004 - 2006 Intel Corporation. All rights reserved.
+ * Portions based on net/core/datagram.c and copyrighted by their authors.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License as published by the Free
+ * Software Foundation; either version 2 of the License, or (at your option)
+ * any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 59
+ * Temple Place - Suite 330, Boston, MA  02111-1307, USA.
+ *
+ * The full GNU General Public License is included in this distribution in the
+ * file called COPYING.
+ */
+
+/*
+ * This code allows the net stack to make use of a DMA engine for
+ * skb to iovec copies.
+ */
+
+#include <linux/dmaengine.h>
+#include <linux/pagemap.h>
+#include <net/tcp.h> /* for memcpy_toiovec */
+#include <asm/io.h>
+#include <asm/uaccess.h>
+
+int num_pages_spanned(struct iovec *iov)
+{
+       return
+       ((PAGE_ALIGN((unsigned long)iov->iov_base + iov->iov_len) -
+       ((unsigned long)iov->iov_base & PAGE_MASK)) >> PAGE_SHIFT);
+}
+
+/*
+ * Pin down all the iovec pages needed for len bytes.
+ * Return a struct dma_pinned_list to keep track of pages pinned down.
+ *
+ * We are allocating a single chunk of memory, and then carving it up into
+ * 3 sections, the latter 2 whose size depends on the number of iovecs and the
+ * total number of pages, respectively.
+ */
+struct dma_pinned_list *dma_pin_iovec_pages(struct iovec *iov, size_t len)
+{
+       struct dma_pinned_list *local_list;
+       struct page **pages;
+       int i;
+       int ret;
+       int nr_iovecs = 0;
+       int iovec_len_used = 0;
+       int iovec_pages_used = 0;
+       long err;
+
+       /* don't pin down non-user-based iovecs */
+       if (segment_eq(get_fs(), KERNEL_DS))
+               return NULL;
+
+       /* determine how many iovecs/pages there are, up front */
+       do {
+               iovec_len_used += iov[nr_iovecs].iov_len;
+               iovec_pages_used += num_pages_spanned(&iov[nr_iovecs]);
+               nr_iovecs++;
+       } while (iovec_len_used < len);
+
+       /* single kmalloc for pinned list, page_list[], and the page arrays */
+       local_list = kmalloc(sizeof(*local_list)
+               + (nr_iovecs * sizeof (struct dma_page_list))
+               + (iovec_pages_used * sizeof (struct page*)), GFP_KERNEL);
+       if (!local_list) {
+               err = -ENOMEM;
+               goto out;
+       }
+
+       /* list of pages starts right after the page list array */
+       pages = (struct page **) &local_list->page_list[nr_iovecs];
+
+       for (i = 0; i < nr_iovecs; i++) {
+               struct dma_page_list *page_list = &local_list->page_list[i];
+
+               len -= iov[i].iov_len;
+
+               if (!access_ok(VERIFY_WRITE, iov[i].iov_base, iov[i].iov_len)) {
+                       err = -EFAULT;
+                       goto unpin;
+               }
+
+               page_list->nr_pages = num_pages_spanned(&iov[i]);
+               page_list->base_address = iov[i].iov_base;
+
+               page_list->pages = pages;
+               pages += page_list->nr_pages;
+
+               /* pin pages down */
+               down_read(&current->mm->mmap_sem);
+               ret = get_user_pages(
+                       current,
+                       current->mm,
+                       (unsigned long) iov[i].iov_base,
+                       page_list->nr_pages,
+                       1,      /* write */
+                       0,      /* force */
+                       page_list->pages,
+                       NULL);
+               up_read(&current->mm->mmap_sem);
+
+               if (ret != page_list->nr_pages) {
+                       err = -ENOMEM;
+                       goto unpin;
+               }
+
+               local_list->nr_iovecs = i + 1;
+       }
+
+       return local_list;
+
+unpin:
+       dma_unpin_iovec_pages(local_list);
+out:
+       return ERR_PTR(err);
+}
+
+void dma_unpin_iovec_pages(struct dma_pinned_list *pinned_list)
+{
+       int i, j;
+
+       if (!pinned_list)
+               return;
+
+       for (i = 0; i < pinned_list->nr_iovecs; i++) {
+               struct dma_page_list *page_list = &pinned_list->page_list[i];
+               for (j = 0; j < page_list->nr_pages; j++) {
+                       set_page_dirty_lock(page_list->pages[j]);
+                       page_cache_release(page_list->pages[j]);
+               }
+       }
+
+       kfree(pinned_list);
+}
+
+static dma_cookie_t dma_memcpy_to_kernel_iovec(struct dma_chan *chan, struct
+       iovec *iov, unsigned char *kdata, size_t len)
+{
+       dma_cookie_t dma_cookie = 0;
+
+       while (len > 0) {
+               if (iov->iov_len) {
+                       int copy = min_t(unsigned int, iov->iov_len, len);
+                       dma_cookie = dma_async_memcpy_buf_to_buf(
+                                       chan,
+                                       iov->iov_base,
+                                       kdata,
+                                       copy);
+                       kdata += copy;
+                       len -= copy;
+                       iov->iov_len -= copy;
+                       iov->iov_base += copy;
+               }
+               iov++;
+       }
+
+       return dma_cookie;
+}
+
+/*
+ * We have already pinned down the pages we will be using in the iovecs.
+ * Each entry in iov array has corresponding entry in pinned_list->page_list.
+ * Using array indexing to keep iov[] and page_list[] in sync.
+ * Initial elements in iov array's iov->iov_len will be 0 if already copied into
+ *   by another call.
+ * iov array length remaining guaranteed to be bigger than len.
+ */
+dma_cookie_t dma_memcpy_to_iovec(struct dma_chan *chan, struct iovec *iov,
+       struct dma_pinned_list *pinned_list, unsigned char *kdata, size_t len)
+{
+       int iov_byte_offset;
+       int copy;
+       dma_cookie_t dma_cookie = 0;
+       int iovec_idx;
+       int page_idx;
+
+       if (!chan)
+               return memcpy_toiovec(iov, kdata, len);
+
+       /* -> kernel copies (e.g. smbfs) */
+       if (!pinned_list)
+               return dma_memcpy_to_kernel_iovec(chan, iov, kdata, len);
+
+       iovec_idx = 0;
+       while (iovec_idx < pinned_list->nr_iovecs) {
+               struct dma_page_list *page_list;
+
+               /* skip already used-up iovecs */
+               while (!iov[iovec_idx].iov_len)
+                       iovec_idx++;
+
+               page_list = &pinned_list->page_list[iovec_idx];
+
+               iov_byte_offset = ((unsigned long)iov[iovec_idx].iov_base & ~PAGE_MASK);
+               page_idx = (((unsigned long)iov[iovec_idx].iov_base & PAGE_MASK)
+                        - ((unsigned long)page_list->base_address & PAGE_MASK)) >> PAGE_SHIFT;
+
+               /* break up copies to not cross page boundary */
+               while (iov[iovec_idx].iov_len) {
+                       copy = min_t(int, PAGE_SIZE - iov_byte_offset, len);
+                       copy = min_t(int, copy, iov[iovec_idx].iov_len);
+
+                       dma_cookie = dma_async_memcpy_buf_to_pg(chan,
+                                       page_list->pages[page_idx],
+                                       iov_byte_offset,
+                                       kdata,
+                                       copy);
+
+                       len -= copy;
+                       iov[iovec_idx].iov_len -= copy;
+                       iov[iovec_idx].iov_base += copy;
+
+                       if (!len)
+                               return dma_cookie;
+
+                       kdata += copy;
+                       iov_byte_offset = 0;
+                       page_idx++;
+               }
+               iovec_idx++;
+       }
+
+       /* really bad if we ever run out of iovecs */
+       BUG();
+       return -EFAULT;
+}
+
+dma_cookie_t dma_memcpy_pg_to_iovec(struct dma_chan *chan, struct iovec *iov,
+       struct dma_pinned_list *pinned_list, struct page *page,
+       unsigned int offset, size_t len)
+{
+       int iov_byte_offset;
+       int copy;
+       dma_cookie_t dma_cookie = 0;
+       int iovec_idx;
+       int page_idx;
+       int err;
+
+       /* this needs as-yet-unimplemented buf-to-buff, so punt. */
+       /* TODO: use dma for this */
+       if (!chan || !pinned_list) {
+               u8 *vaddr = kmap(page);
+               err = memcpy_toiovec(iov, vaddr + offset, len);
+               kunmap(page);
+               return err;
+       }
+
+       iovec_idx = 0;
+       while (iovec_idx < pinned_list->nr_iovecs) {
+               struct dma_page_list *page_list;
+
+               /* skip already used-up iovecs */
+               while (!iov[iovec_idx].iov_len)
+                       iovec_idx++;
+
+               page_list = &pinned_list->page_list[iovec_idx];
+
+               iov_byte_offset = ((unsigned long)iov[iovec_idx].iov_base & ~PAGE_MASK);
+               page_idx = (((unsigned long)iov[iovec_idx].iov_base & PAGE_MASK)
+                        - ((unsigned long)page_list->base_address & PAGE_MASK)) >> PAGE_SHIFT;
+
+               /* break up copies to not cross page boundary */
+               while (iov[iovec_idx].iov_len) {
+                       copy = min_t(int, PAGE_SIZE - iov_byte_offset, len);
+                       copy = min_t(int, copy, iov[iovec_idx].iov_len);
+
+                       dma_cookie = dma_async_memcpy_pg_to_pg(chan,
+                                       page_list->pages[page_idx],
+                                       iov_byte_offset,
+                                       page,
+                                       offset,
+                                       copy);
+
+                       len -= copy;
+                       iov[iovec_idx].iov_len -= copy;
+                       iov[iovec_idx].iov_base += copy;
+
+                       if (!len)
+                               return dma_cookie;
+
+                       offset += copy;
+                       iov_byte_offset = 0;
+                       page_idx++;
+               }
+               iovec_idx++;
+       }
+
+       /* really bad if we ever run out of iovecs */
+       BUG();
+       return -EFAULT;
+}
index 3078154..78b236c 100644 (file)
@@ -333,5 +333,27 @@ static inline enum dma_status dma_async_is_complete(dma_cookie_t cookie,
 int dma_async_device_register(struct dma_device *device);
 void dma_async_device_unregister(struct dma_device *device);
 
+/* --- Helper iov-locking functions --- */
+
+struct dma_page_list {
+       char *base_address;
+       int nr_pages;
+       struct page **pages;
+};
+
+struct dma_pinned_list {
+       int nr_iovecs;
+       struct dma_page_list page_list[0];
+};
+
+struct dma_pinned_list *dma_pin_iovec_pages(struct iovec *iov, size_t len);
+void dma_unpin_iovec_pages(struct dma_pinned_list* pinned_list);
+
+dma_cookie_t dma_memcpy_to_iovec(struct dma_chan *chan, struct iovec *iov,
+       struct dma_pinned_list *pinned_list, unsigned char *kdata, size_t len);
+dma_cookie_t dma_memcpy_pg_to_iovec(struct dma_chan *chan, struct iovec *iov,
+       struct dma_pinned_list *pinned_list, struct page *page,
+       unsigned int offset, size_t len);
+
 #endif /* CONFIG_DMA_ENGINE */
 #endif /* DMAENGINE_H */
index cbfe89d..19760eb 100644 (file)
@@ -23,6 +23,7 @@
 #include <linux/config.h>
 #ifdef CONFIG_NET_DMA
 #include <linux/dmaengine.h>
+#include <linux/skbuff.h>
 
 static inline struct dma_chan *get_softnet_dma(void)
 {
@@ -34,5 +35,10 @@ static inline struct dma_chan *get_softnet_dma(void)
        rcu_read_unlock();
        return chan;
 }
+
+int dma_skb_copy_datagram_iovec(struct dma_chan* chan,
+               const struct sk_buff *skb, int offset, struct iovec *to,
+               size_t len, struct dma_pinned_list *pinned_list);
+
 #endif /* CONFIG_NET_DMA */
 #endif /* NETDMA_H */
index 79fe12c..e9bd246 100644 (file)
@@ -16,3 +16,4 @@ obj-$(CONFIG_NET_DIVERT) += dv.o
 obj-$(CONFIG_NET_PKTGEN) += pktgen.o
 obj-$(CONFIG_WIRELESS_EXT) += wireless.o
 obj-$(CONFIG_NETPOLL) += netpoll.o
+obj-$(CONFIG_NET_DMA) += user_dma.o
diff --git a/net/core/user_dma.c b/net/core/user_dma.c
new file mode 100644 (file)
index 0000000..9eee91b
--- /dev/null
@@ -0,0 +1,127 @@
+/*
+ * Copyright(c) 2004 - 2006 Intel Corporation. All rights reserved.
+ * Portions based on net/core/datagram.c and copyrighted by their authors.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License as published by the Free
+ * Software Foundation; either version 2 of the License, or (at your option)
+ * any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 59
+ * Temple Place - Suite 330, Boston, MA  02111-1307, USA.
+ *
+ * The full GNU General Public License is included in this distribution in the
+ * file called COPYING.
+ */
+
+/*
+ * This code allows the net stack to make use of a DMA engine for
+ * skb to iovec copies.
+ */
+
+#include <linux/dmaengine.h>
+#include <linux/socket.h>
+#include <linux/rtnetlink.h> /* for BUG_TRAP */
+#include <net/tcp.h>
+
+/**
+ *     dma_skb_copy_datagram_iovec - Copy a datagram to an iovec.
+ *     @skb - buffer to copy
+ *     @offset - offset in the buffer to start copying from
+ *     @iovec - io vector to copy to
+ *     @len - amount of data to copy from buffer to iovec
+ *     @pinned_list - locked iovec buffer data
+ *
+ *     Note: the iovec is modified during the copy.
+ */
+int dma_skb_copy_datagram_iovec(struct dma_chan *chan,
+                       struct sk_buff *skb, int offset, struct iovec *to,
+                       size_t len, struct dma_pinned_list *pinned_list)
+{
+       int start = skb_headlen(skb);
+       int i, copy = start - offset;
+       dma_cookie_t cookie = 0;
+
+       /* Copy header. */
+       if (copy > 0) {
+               if (copy > len)
+                       copy = len;
+               cookie = dma_memcpy_to_iovec(chan, to, pinned_list,
+                                           skb->data + offset, copy);
+               if (cookie < 0)
+                       goto fault;
+               len -= copy;
+               if (len == 0)
+                       goto end;
+               offset += copy;
+       }
+
+       /* Copy paged appendix. Hmm... why does this look so complicated? */
+       for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
+               int end;
+
+               BUG_TRAP(start <= offset + len);
+
+               end = start + skb_shinfo(skb)->frags[i].size;
+               copy = end - offset;
+               if ((copy = end - offset) > 0) {
+                       skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
+                       struct page *page = frag->page;
+
+                       if (copy > len)
+                               copy = len;
+
+                       cookie = dma_memcpy_pg_to_iovec(chan, to, pinned_list, page,
+                                       frag->page_offset + offset - start, copy);
+                       if (cookie < 0)
+                               goto fault;
+                       len -= copy;
+                       if (len == 0)
+                               goto end;
+                       offset += copy;
+               }
+               start = end;
+       }
+
+       if (skb_shinfo(skb)->frag_list) {
+               struct sk_buff *list = skb_shinfo(skb)->frag_list;
+
+               for (; list; list = list->next) {
+                       int end;
+
+                       BUG_TRAP(start <= offset + len);
+
+                       end = start + list->len;
+                       copy = end - offset;
+                       if (copy > 0) {
+                               if (copy > len)
+                                       copy = len;
+                               cookie = dma_skb_copy_datagram_iovec(chan, list,
+                                               offset - start, to, copy,
+                                               pinned_list);
+                               if (cookie < 0)
+                                       goto fault;
+                               len -= copy;
+                               if (len == 0)
+                                       goto end;
+                               offset += copy;
+                       }
+                       start = end;
+               }
+       }
+
+end:
+       if (!len) {
+               skb->dma_cookie = cookie;
+               return cookie;
+       }
+
+fault:
+       return -EFAULT;
+}