Skip to contentSkip to conversation

CIFS: Implement cifs_strict_write (try #2)

PATCH In review Pavel Shilovsky 2 participants 3 messages Dec 17, 2010 – Dec 22, 2010 2 messages not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 5/6 CIFS: Implement cifs_strict_write (try #2)

If we don't have Exclusive oplock we write a data to the server through CISSMBWrite2. Also set invalidate_mapping flag on the inode if we wrote something to the server.

---
 fs/cifs/cifsfs.c    |  228 ++++++++++++++++++++++++++++++++++++++++++++++++++-
 fs/cifs/cifsproto.h |    2 +
 fs/cifs/file.c      |    2 +-
 3 files changed, 227 insertions(+), 5 deletions(-)
fs/cifs/cifsfs.c +224 -4
index bcb4d2c..fb50883 100644
@@ -606,15 +606,235 @@ static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov,
 	return read;
 }
 
+/* Assume that dest points to memory area that is enough for source data */
+static int cifs_copy_iovs(struct kvec *dst, const struct iovec **source,
+			  unsigned long *lastlen, unsigned long len)
+{
+	const struct iovec *src = *source;
+	unsigned long dst_lastlen, src_lastlen;
+	unsigned long cur_len;
+
+	dst_lastlen = dst->iov_len;
+	src_lastlen = *lastlen;
+
+	while (len > 0) {
+		cur_len = min_t(unsigned long, dst_lastlen, src_lastlen);
+		cur_len = min_t(unsigned long, cur_len, len);
+
+		if (copy_from_user(dst->iov_base + dst->iov_len - dst_lastlen,
+				   src->iov_base + src->iov_len - src_lastlen,
+				   cur_len))
+			return -EFAULT;
+
+		dst_lastlen -= cur_len;
+		if (!dst_lastlen) {
+			dst++;
+			dst_lastlen = dst->iov_len;
+		}
+
+		src_lastlen -= cur_len;
+		if (!src_lastlen) {
+			src++;
+			src_lastlen = src->iov_len;
+		}
+
+		len -= cur_len;
+	}
+
+	/* update length of the last element in dest */
+	if (dst_lastlen)
+		dst->iov_len -= dst_lastlen;
+
+	/* save offset and length of the last element in source */
+	*lastlen = src_lastlen;
+	*source = src;
+
+	return 0;
+}
+
+static int
+cifs_strict_write_helper(struct page ***ppages, struct kvec **pto_send,
+			 unsigned long num_pages)
+{
+	int rc = 0;
+	unsigned long i;
+	struct page **pages;
+	struct kvec *to_send;
+
+	*ppages = kmalloc(sizeof(struct pages *)*num_pages, GFP_KERNEL);
+	if (!ppages)
+		return -ENOMEM;
+
+	*pto_send = kmalloc(sizeof(struct kvec)*(num_pages + 1),
+			  GFP_KERNEL);
+	if (!pto_send) {
+		kfree(ppages);
+		return -ENOMEM;
+	}
+
+	pages = *ppages;
+	to_send = *pto_send;
+
+	for (i = 0; i < num_pages; i++) {
+		pages[i] = alloc_page(__GFP_HIGHMEM);
+		if (!pages[i]) {
+			/*
+			 * save number of pages we have already allocated and
+			 * return with ENOMEM error
+			 */
+			num_pages = i;
+			rc = -ENOMEM;
+			goto error;
+		}
+		to_send[i+1].iov_base = kmap(pages[i]);
+		to_send[i+1].iov_len = PAGE_CACHE_SIZE;
+	}
+
+	return rc;
+
+error:
+	for (i = 0; i < num_pages; i++) {
+		kunmap(pages[i]);
+		put_page(pages[i]);
+	}
+	kfree(to_send);
+	kfree(pages);
+	return rc;
+}
+
+static ssize_t cifs_strict_write(struct kiocb *iocb, const struct iovec *iov,
+				 unsigned long nr_segs, loff_t pos)
+{
+	struct inode *inode;
+	struct cifs_sb_info *cifs_sb;
+	ssize_t written = 0, total_written = 0;
+	unsigned long i, num_pages, cur_numpages;
+	unsigned long len, cur_len, last_len;
+	struct kvec *to_send;
+	const struct iovec *offset;
+	struct page **pages;
+	struct cifsFileInfo *open_file;
+	struct cifsTconInfo *pTcon;
+	int xid, rc;
+
+	inode = iocb->ki_filp->f_path.dentry->d_inode;
+
+	if (CIFS_I(inode)->clientCanCacheAll)
+		return generic_file_aio_write(iocb, iov, nr_segs, pos);
+
+	cifs_sb = CIFS_SB(iocb->ki_filp->f_path.dentry->d_sb);
+
+	/*
+	 * In strict cache mode we need to write the data to the server exactly
+	 * from the pos to pos+len-1 rather than flush all affected pages
+	 * because it may cause a error with mandatory locks on these pages but
+	 * not on the region from pos to ppos+len-1.
+	 */
+
+	len = 0;
+	for (i = 0; i < nr_segs; i++)
+		len += iov[i].iov_len;
+
+	if (!len)
+		return 0;
+
+	/*
+	 * BB - optimize the way when signing is disabled. We can drop this
+	 * extra memory-to-memory copying and use iovec buffers for constructing
+	 * write request.
+	 */
+
+	xid = GetXid();
+	open_file = iocb->ki_filp->private_data;
+	pTcon = tlink_tcon(open_file->tlink);
+
+	/* calculate number of pages we need for process this write */
+	cur_len = min_t(unsigned long, len, cifs_sb->wsize);
+	num_pages = cur_len / PAGE_CACHE_SIZE;
+	if (cur_len % PAGE_CACHE_SIZE)
+		num_pages++;
+
+	rc = cifs_strict_write_helper(&pages, &to_send, num_pages);
+	if (rc) {
+		total_written = rc;
+		goto error;
+	}
+
+	last_len = iov->iov_len;
+	offset = iov;
+	cur_numpages = num_pages;
+
+	do {
+		rc = cifs_copy_iovs(to_send+1, &offset, &last_len, cur_len);
+		if (rc) {
+			total_written = -EFAULT;
+			goto exit;
+		}
+
+		do {
+			rc = CIFSSMBWrite2(xid, pTcon, open_file->netfid,
+					   cur_len, pos, &written, to_send,
+					   cur_numpages, 0);
+		} while (rc == -EAGAIN);
+
+		if (written) {
+			len -= written;
+			total_written += written;
+			cifs_update_eof(CIFS_I(inode), pos, written);
+			pos += written;
+		} else if (rc < 0) {
+			if (!total_written)
+				total_written = rc;
+			goto exit;
+		}
+
+		/* get length and number of kvecs of the next write */
+		cur_len = min_t(unsigned long, len, cifs_sb->wsize);
+		cur_numpages = cur_len / PAGE_CACHE_SIZE;
+		if (cur_len % PAGE_CACHE_SIZE)
+			cur_numpages++;
+	} while (len > 0);
+
+exit:
+	if (total_written > 0) {
+		CIFS_I(inode)->invalid_mapping = true;
+		iocb->ki_pos += total_written;
+		spin_lock(&inode->i_lock);
+		if (pos > inode->i_size)
+			i_size_write(inode, pos);
+		spin_unlock(&inode->i_lock);
+	}
+
+	cifs_stats_bytes_written(pTcon, total_written);
+	mark_inode_dirty_sync(inode);
+
+	for (i = 0; i < num_pages; i++) {
+		kunmap(pages[i]);
+		put_page(pages[i]);
+	}
+	kfree(to_send);
+	kfree(pages);
+error:
+	FreeXid(xid);
+	return total_written;
+}
+
 static ssize_t cifs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 				   unsigned long nr_segs, loff_t pos)
 {
 	struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 	ssize_t written;
+	int rc;
 
 	written = generic_file_aio_write(iocb, iov, nr_segs, pos);
-	if (!CIFS_I(inode)->clientCanCacheAll)
-		filemap_fdatawrite(inode->i_mapping);
+
+	if (CIFS_I(inode)->clientCanCacheAll)
+		return written;
+
+	rc = filemap_fdatawrite(inode->i_mapping);
+	if (rc)
+		cFYI(1, "cifs_file_aio_write: %d rc on %p inode", rc, inode);
+
 	return written;
 }
 
@@ -748,7 +968,7 @@ const struct file_operations cifs_file_strict_ops = {
 	.read = do_sync_read,
 	.write = do_sync_write,
 	.aio_read = cifs_strict_read,
-	.aio_write = cifs_file_aio_write,
+	.aio_write = cifs_strict_write,
 	.open = cifs_open,
 	.release = cifs_close,
 	.lock = cifs_lock,
@@ -804,7 +1024,7 @@ const struct file_operations cifs_file_strict_nobrl_ops = {
 	.read = do_sync_read,
 	.write = do_sync_write,
 	.aio_read = cifs_strict_read,
-	.aio_write = cifs_file_aio_write,
+	.aio_write = cifs_strict_write,
 	.open = cifs_open,
 	.release = cifs_close,
 	.fsync = cifs_strict_fsync,
fs/cifs/cifsproto.h +2 -0
index e6d1481..e7d6c24 100644
@@ -79,6 +79,8 @@ extern int checkSMB(struct smb_hdr *smb, __u16 mid, unsigned int length);
 extern bool is_valid_oplock_break(struct smb_hdr *smb,
 				  struct TCP_Server_Info *);
 extern bool is_size_safe_to_change(struct cifsInodeInfo *, __u64 eof);
+extern void cifs_update_eof(struct cifsInodeInfo *cifsi, loff_t offset,
+			    unsigned int bytes_written);
 extern struct cifsFileInfo *find_writable_file(struct cifsInodeInfo *, bool);
 extern struct cifsFileInfo *find_readable_file(struct cifsInodeInfo *, bool);
 extern unsigned int smbCalcSize(struct smb_hdr *ptr);
fs/cifs/file.c +1 -1
index bc53e9e..8260f53 100644
@@ -924,7 +924,7 @@ cifs_write_timeout(struct cifsInodeInfo *cifsi, loff_t offset)
 }
 
 /* update the file size (if needed) after a write */
-static void
+void
 cifs_update_eof(struct cifsInodeInfo *cifsi, loff_t offset,
 		      unsigned int bytes_written)
 {
Signature · 1 line
1.7.3.2

A message that is not in the archive (sent off-list or not yet received): 1292590154-4768-1-git-send-email-piastryyy-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#2 lore raw
Quoted 57 lines · On Fri, 17 Dec 2010 15:49:14 +0300 Pavel Shilovsky…

If we don't have Exclusive oplock we write a data to the server through CISSMBWrite2. Also set invalidate_mapping flag on the inode if we wrote something to the server.

---
 fs/cifs/cifsfs.c    |  228 ++++++++++++++++++++++++++++++++++++++++++++++++++-
 fs/cifs/cifsproto.h |    2 +
 fs/cifs/file.c      |    2 +-
 3 files changed, 227 insertions(+), 5 deletions(-)
fs/cifs/cifsfs.c +38 -0
index bcb4d2c..fb50883 100644
@@ -606,15 +606,235 @@ static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov,
 	return read;
 }
 
+/* Assume that dest points to memory area that is enough for source data */
+static int cifs_copy_iovs(struct kvec *dst, const struct iovec **source,
+			  unsigned long *lastlen, unsigned long len)
+{
+	const struct iovec *src = *source;
+	unsigned long dst_lastlen, src_lastlen;
+	unsigned long cur_len;
+
+	dst_lastlen = dst->iov_len;
+	src_lastlen = *lastlen;
+
+	while (len > 0) {
+		cur_len = min_t(unsigned long, dst_lastlen, src_lastlen);
+		cur_len = min_t(unsigned long, cur_len, len);
+
+		if (copy_from_user(dst->iov_base + dst->iov_len - dst_lastlen,
+				   src->iov_base + src->iov_len - src_lastlen,
+				   cur_len))
+			return -EFAULT;
+
+		dst_lastlen -= cur_len;
+		if (!dst_lastlen) {
+			dst++;
+			dst_lastlen = dst->iov_len;
+		}
+
+		src_lastlen -= cur_len;
+		if (!src_lastlen) {
+			src++;
+			src_lastlen = src->iov_len;
+		}
+
+		len -= cur_len;
+	}
+
+	/* update length of the last element in dest */
+	if (dst_lastlen)
+		dst->iov_len -= dst_lastlen;

fs/cifs/cifsfs.c static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov, · line 656

Quoted 10 lines
fs/cifs/cifsfs.c +10 -0
+
+	/* save offset and length of the last element in source */
+	*lastlen = src_lastlen;
+	*source = src;
+
+	return 0;
+}
+
+static int
+cifs_strict_write_helper(struct page ***ppages, struct kvec **pto_send,
					^^^^
				Wow, a triple pointer? Ugly...
Quoted 24 lines
fs/cifs/cifsfs.c +24 -0
+			 unsigned long num_pages)
+{
+	int rc = 0;
+	unsigned long i;
+	struct page **pages;
+	struct kvec *to_send;
+
+	*ppages = kmalloc(sizeof(struct pages *)*num_pages, GFP_KERNEL);
+	if (!ppages)
+		return -ENOMEM;
+
+	*pto_send = kmalloc(sizeof(struct kvec)*(num_pages + 1),
+			  GFP_KERNEL);
+	if (!pto_send) {
+		kfree(ppages);
+		return -ENOMEM;
+	}
+
+	pages = *ppages;
+	to_send = *pto_send;
+
+	for (i = 0; i < num_pages; i++) {
+		pages[i] = alloc_page(__GFP_HIGHMEM);
+		if (!pages[i]) {

fs/cifs/cifsfs.c static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov, · line 690

Quoted 10 lines
fs/cifs/cifsfs.c +10 -0
+			/*
+			 * save number of pages we have already allocated and
+			 * return with ENOMEM error
+			 */
+			num_pages = i;
+			rc = -ENOMEM;
+			goto error;
+		}
+		to_send[i+1].iov_base = kmap(pages[i]);
+		to_send[i+1].iov_len = PAGE_CACHE_SIZE;

^^^^^^^^^^^
kmap space is limited on 32-bit architectures. You really want to limit the amount of pages that you have simultaneously mapped.

What's going to happen if someone passes down a massive write this way? I expect that it would fail badly.

I think that you really only want to allocate the memory needed to perform just the write that you're getting ready to perform, and then free it when that's done.

Quoted 4 lines
fs/cifs/cifsfs.c +4 -0
+	}
+
+	return rc;
+

fs/cifs/cifsfs.c static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov, · line 704

Quoted 10 lines
fs/cifs/cifsfs.c +10 -0
+error:
+	for (i = 0; i < num_pages; i++) {
+		kunmap(pages[i]);
+		put_page(pages[i]);
+	}
+	kfree(to_send);
+	kfree(pages);
+	return rc;
+}
+

Instead of rolling your own functions for this, might it be better to try and use the iov_iter infrastructure? Look at what generic_perform_write does, IOW. This situation is a little different, but there are similarities.

Quoted 182 lines
fs/cifs/cifsfs.c +128 -4
+static ssize_t cifs_strict_write(struct kiocb *iocb, const struct iovec *iov,
+				 unsigned long nr_segs, loff_t pos)
+{
+	struct inode *inode;
+	struct cifs_sb_info *cifs_sb;
+	ssize_t written = 0, total_written = 0;
+	unsigned long i, num_pages, cur_numpages;
+	unsigned long len, cur_len, last_len;
+	struct kvec *to_send;
+	const struct iovec *offset;
+	struct page **pages;
+	struct cifsFileInfo *open_file;
+	struct cifsTconInfo *pTcon;
+	int xid, rc;
+
+	inode = iocb->ki_filp->f_path.dentry->d_inode;
+
+	if (CIFS_I(inode)->clientCanCacheAll)
+		return generic_file_aio_write(iocb, iov, nr_segs, pos);
+
+	cifs_sb = CIFS_SB(iocb->ki_filp->f_path.dentry->d_sb);
+
+	/*
+	 * In strict cache mode we need to write the data to the server exactly
+	 * from the pos to pos+len-1 rather than flush all affected pages
+	 * because it may cause a error with mandatory locks on these pages but
+	 * not on the region from pos to ppos+len-1.
+	 */
+
+	len = 0;
+	for (i = 0; i < nr_segs; i++)
+		len += iov[i].iov_len;
+
+	if (!len)
+		return 0;
+
+	/*
+	 * BB - optimize the way when signing is disabled. We can drop this
+	 * extra memory-to-memory copying and use iovec buffers for constructing
+	 * write request.
+	 */
+
+	xid = GetXid();
+	open_file = iocb->ki_filp->private_data;
+	pTcon = tlink_tcon(open_file->tlink);
+
+	/* calculate number of pages we need for process this write */
+	cur_len = min_t(unsigned long, len, cifs_sb->wsize);
+	num_pages = cur_len / PAGE_CACHE_SIZE;
+	if (cur_len % PAGE_CACHE_SIZE)
+		num_pages++;
+
+	rc = cifs_strict_write_helper(&pages, &to_send, num_pages);
+	if (rc) {
+		total_written = rc;
+		goto error;
+	}
+
+	last_len = iov->iov_len;
+	offset = iov;
+	cur_numpages = num_pages;
+
+	do {
+		rc = cifs_copy_iovs(to_send+1, &offset, &last_len, cur_len);
+		if (rc) {
+			total_written = -EFAULT;
+			goto exit;
+		}
+
+		do {
+			rc = CIFSSMBWrite2(xid, pTcon, open_file->netfid,
+					   cur_len, pos, &written, to_send,
+					   cur_numpages, 0);
+		} while (rc == -EAGAIN);
+
+		if (written) {
+			len -= written;
+			total_written += written;
+			cifs_update_eof(CIFS_I(inode), pos, written);
+			pos += written;
+		} else if (rc < 0) {
+			if (!total_written)
+				total_written = rc;
+			goto exit;
+		}
+
+		/* get length and number of kvecs of the next write */
+		cur_len = min_t(unsigned long, len, cifs_sb->wsize);
+		cur_numpages = cur_len / PAGE_CACHE_SIZE;
+		if (cur_len % PAGE_CACHE_SIZE)
+			cur_numpages++;
+	} while (len > 0);
+
+exit:
+	if (total_written > 0) {
+		CIFS_I(inode)->invalid_mapping = true;
+		iocb->ki_pos += total_written;
+		spin_lock(&inode->i_lock);
+		if (pos > inode->i_size)
+			i_size_write(inode, pos);
+		spin_unlock(&inode->i_lock);
+	}
+
+	cifs_stats_bytes_written(pTcon, total_written);
+	mark_inode_dirty_sync(inode);
+
+	for (i = 0; i < num_pages; i++) {
+		kunmap(pages[i]);
+		put_page(pages[i]);
+	}
+	kfree(to_send);
+	kfree(pages);
+error:
+	FreeXid(xid);
+	return total_written;
+}
+
 static ssize_t cifs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 				   unsigned long nr_segs, loff_t pos)
 {
 	struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 	ssize_t written;
+	int rc;
 
 	written = generic_file_aio_write(iocb, iov, nr_segs, pos);
-	if (!CIFS_I(inode)->clientCanCacheAll)
-		filemap_fdatawrite(inode->i_mapping);
+
+	if (CIFS_I(inode)->clientCanCacheAll)
+		return written;
+
+	rc = filemap_fdatawrite(inode->i_mapping);
+	if (rc)
+		cFYI(1, "cifs_file_aio_write: %d rc on %p inode", rc, inode);
+
 	return written;
 }
 
@@ -748,7 +968,7 @@ const struct file_operations cifs_file_strict_ops = {
 	.read = do_sync_read,
 	.write = do_sync_write,
 	.aio_read = cifs_strict_read,
-	.aio_write = cifs_file_aio_write,
+	.aio_write = cifs_strict_write,
 	.open = cifs_open,
 	.release = cifs_close,
 	.lock = cifs_lock,
@@ -804,7 +1024,7 @@ const struct file_operations cifs_file_strict_nobrl_ops = {
 	.read = do_sync_read,
 	.write = do_sync_write,
 	.aio_read = cifs_strict_read,
-	.aio_write = cifs_file_aio_write,
+	.aio_write = cifs_strict_write,
 	.open = cifs_open,
 	.release = cifs_close,
 	.fsync = cifs_strict_fsync,
fs/cifs/cifsproto.h +2 -0
index e6d1481..e7d6c24 100644
@@ -79,6 +79,8 @@ extern int checkSMB(struct smb_hdr *smb, __u16 mid, unsigned int length);
 extern bool is_valid_oplock_break(struct smb_hdr *smb,
 				  struct TCP_Server_Info *);
 extern bool is_size_safe_to_change(struct cifsInodeInfo *, __u64 eof);
+extern void cifs_update_eof(struct cifsInodeInfo *cifsi, loff_t offset,
+			    unsigned int bytes_written);
 extern struct cifsFileInfo *find_writable_file(struct cifsInodeInfo *, bool);
 extern struct cifsFileInfo *find_readable_file(struct cifsInodeInfo *, bool);
 extern unsigned int smbCalcSize(struct smb_hdr *ptr);
fs/cifs/file.c +1 -1
index bc53e9e..8260f53 100644
@@ -924,7 +924,7 @@ cifs_write_timeout(struct cifsInodeInfo *cifsi, loff_t offset)
 }
 
 /* update the file size (if needed) after a write */
-static void
+void
 cifs_update_eof(struct cifsInodeInfo *cifsi, loff_t offset,
 		      unsigned int bytes_written)
 {
Signature · 1 line
Jeff Layton <[email protected]>

A message that is not in the archive (sent off-list or not yet received): 20101221140941.17fd3ece-9yPaYZwiELC+kQycOl6kW4xkIHaj4LzF@public.gmane.org

#4 lore raw

2010/12/21 Jeff Layton <[email protected]>:

Quoted 318 lines
Quoted 57 lines · On Fri, 17 Dec 2010 15:49:14 +0300 Pavel Shilovsky…

If we don't have Exclusive oplock we write a data to the server through CISSMBWrite2. Also set invalidate_mapping flag on the inode if we wrote something to the server.

---
 fs/cifs/cifsfs.c    |  228 ++++++++++++++++++++++++++++++++++++++++++++++++++- fs/cifs/cifsproto.h |    2 +
 fs/cifs/file.c      |    2 +-

 3 files changed, 227 insertions(+), 5 deletions(-)
fs/cifs/cifsfs.c +38 -0
index bcb4d2c..fb50883 100644
@@ -606,15 +606,235 @@ static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov,
       return read;
  }
 
+/* Assume that dest points to memory area that is enough for source data */
+static int cifs_copy_iovs(struct kvec *dst, const struct iovec **source,
+                       unsigned long *lastlen, unsigned long len)
+{
+     const struct iovec *src = *source;
+     unsigned long dst_lastlen, src_lastlen;
+     unsigned long cur_len;
+
+     dst_lastlen = dst->iov_len;
+     src_lastlen = *lastlen;
+
+     while (len > 0) {
+             cur_len = min_t(unsigned long, dst_lastlen, src_lastlen);
+             cur_len = min_t(unsigned long, cur_len, len);
+
+             if (copy_from_user(dst->iov_base + dst->iov_len - dst_lastlen,
+                                src->iov_base + src->iov_len - src_lastlen,
+                                cur_len))
+                     return -EFAULT;
+
+             dst_lastlen -= cur_len;
+             if (!dst_lastlen) {
+                     dst++;
+                     dst_lastlen = dst->iov_len;
+             }
+
+             src_lastlen -= cur_len;
+             if (!src_lastlen) {
+                     src++;
+                     src_lastlen = src->iov_len;
+             }
+
+             len -= cur_len;
+     }
+
+     /* update length of the last element in dest */
+     if (dst_lastlen)
+             dst->iov_len -= dst_lastlen;

fs/cifs/cifsfs.c static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov, · line 656

Quoted 10 lines
fs/cifs/cifsfs.c +10 -0
+
+     /* save offset and length of the last element in source */
+     *lastlen = src_lastlen;
+     *source = src;
+
+     return 0;
+}
+
+static int
+cifs_strict_write_helper(struct page ***ppages, struct kvec **pto_send,

                                       ^^^^
                               Wow, a triple pointer? Ugly...

Quoted 24 lines
fs/cifs/cifsfs.c +24 -0
+                      unsigned long num_pages)
+{
+     int rc = 0;
+     unsigned long i;
+     struct page **pages;
+     struct kvec *to_send;
+
+     *ppages = kmalloc(sizeof(struct pages *)*num_pages, GFP_KERNEL);
+     if (!ppages)
+             return -ENOMEM;
+
+     *pto_send = kmalloc(sizeof(struct kvec)*(num_pages + 1),
+                       GFP_KERNEL);
+     if (!pto_send) {
+             kfree(ppages);
+             return -ENOMEM;
+     }
+
+     pages = *ppages;
+     to_send = *pto_send;
+
+     for (i = 0; i < num_pages; i++) {
+             pages[i] = alloc_page(__GFP_HIGHMEM);
+             if (!pages[i]) {

fs/cifs/cifsfs.c static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov, · line 690

Quoted 10 lines
fs/cifs/cifsfs.c +10 -0
+                     /*
+                      * save number of pages we have already allocated and
+                      * return with ENOMEM error
+                      */
+                     num_pages = i;
+                     rc = -ENOMEM;
+                     goto error;
+             }
+             to_send[i+1].iov_base = kmap(pages[i]);
+             to_send[i+1].iov_len = PAGE_CACHE_SIZE;

                               ^^^^^^^^^^^
                       kmap space is limited on 32-bit architectures. You really want to limit the amount of pages that you have simultaneously mapped.

                       What's going to happen if someone passes down a massive write this way? I expect that it would fail badly.

                       I think that you really only want to allocate the memory needed to perform just the write that you're getting ready to perform, and then free it when that's done.

Quoted 4 lines
fs/cifs/cifsfs.c +4 -0
+     }
+
+     return rc;
+

fs/cifs/cifsfs.c static ssize_t cifs_strict_read(struct kiocb *iocb, const struct iovec *iov, · line 704

Quoted 10 lines
fs/cifs/cifsfs.c +10 -0
+error:
+     for (i = 0; i < num_pages; i++) {
+             kunmap(pages[i]);
+             put_page(pages[i]);
+     }
+     kfree(to_send);
+     kfree(pages);
+     return rc;
+}
+

Instead of rolling your own functions for this, might it be better to try and use the iov_iter infrastructure? Look at what generic_perform_write does, IOW. This situation is a little different, but there are similarities.

Quoted 176 lines
fs/cifs/cifsfs.c +128 -4
+static ssize_t cifs_strict_write(struct kiocb *iocb, const struct iovec *iov,
+                              unsigned long nr_segs, loff_t pos)
+{
+     struct inode *inode;
+     struct cifs_sb_info *cifs_sb;
+     ssize_t written = 0, total_written = 0;
+     unsigned long i, num_pages, cur_numpages;
+     unsigned long len, cur_len, last_len;
+     struct kvec *to_send;
+     const struct iovec *offset;
+     struct page **pages;
+     struct cifsFileInfo *open_file;
+     struct cifsTconInfo *pTcon;
+     int xid, rc;
+
+     inode = iocb->ki_filp->f_path.dentry->d_inode;
+
+     if (CIFS_I(inode)->clientCanCacheAll)
+             return generic_file_aio_write(iocb, iov, nr_segs, pos);
+
+     cifs_sb = CIFS_SB(iocb->ki_filp->f_path.dentry->d_sb);
+
+     /*
+      * In strict cache mode we need to write the data to the server exactly
+      * from the pos to pos+len-1 rather than flush all affected pages
+      * because it may cause a error with mandatory locks on these pages but
+      * not on the region from pos to ppos+len-1.
+      */
+
+     len = 0;
+     for (i = 0; i < nr_segs; i++)
+             len += iov[i].iov_len;
+
+     if (!len)
+             return 0;
+
+     /*
+      * BB - optimize the way when signing is disabled. We can drop this
+      * extra memory-to-memory copying and use iovec buffers for constructing
+      * write request.
+      */
+
+     xid = GetXid();
+     open_file = iocb->ki_filp->private_data;
+     pTcon = tlink_tcon(open_file->tlink);
+
+     /* calculate number of pages we need for process this write */
+     cur_len = min_t(unsigned long, len, cifs_sb->wsize);
+     num_pages = cur_len / PAGE_CACHE_SIZE;
+     if (cur_len % PAGE_CACHE_SIZE)
+             num_pages++;
+
+     rc = cifs_strict_write_helper(&pages, &to_send, num_pages);
+     if (rc) {
+             total_written = rc;
+             goto error;
+     }
+
+     last_len = iov->iov_len;
+     offset = iov;
+     cur_numpages = num_pages;
+
+     do {
+             rc = cifs_copy_iovs(to_send+1, &offset, &last_len, cur_len);
+             if (rc) {
+                     total_written = -EFAULT;
+                     goto exit;
+             }
+
+             do {
+                     rc = CIFSSMBWrite2(xid, pTcon, open_file->netfid,
+                                        cur_len, pos, &written, to_send,
+                                        cur_numpages, 0);
+             } while (rc == -EAGAIN);
+
+             if (written) {
+                     len -= written;
+                     total_written += written;
+                     cifs_update_eof(CIFS_I(inode), pos, written);
+                     pos += written;
+             } else if (rc < 0) {
+                     if (!total_written)
+                             total_written = rc;
+                     goto exit;
+             }
+
+             /* get length and number of kvecs of the next write */
+             cur_len = min_t(unsigned long, len, cifs_sb->wsize);
+             cur_numpages = cur_len / PAGE_CACHE_SIZE;
+             if (cur_len % PAGE_CACHE_SIZE)
+                     cur_numpages++;
+     } while (len > 0);
+
+exit:
+     if (total_written > 0) {
+             CIFS_I(inode)->invalid_mapping = true;
+             iocb->ki_pos += total_written;
+             spin_lock(&inode->i_lock);
+             if (pos > inode->i_size)
+                     i_size_write(inode, pos);
+             spin_unlock(&inode->i_lock);
+     }
+
+     cifs_stats_bytes_written(pTcon, total_written);
+     mark_inode_dirty_sync(inode);
+
+     for (i = 0; i < num_pages; i++) {
+             kunmap(pages[i]);
+             put_page(pages[i]);
+     }
+     kfree(to_send);
+     kfree(pages);
+error:
+     FreeXid(xid);
+     return total_written;
+}
+
  static ssize_t cifs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
                                  unsigned long nr_segs, loff_t pos)
  {
       struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
       ssize_t written;
+     int rc;
 
       written = generic_file_aio_write(iocb, iov, nr_segs, pos);
-     if (!CIFS_I(inode)->clientCanCacheAll)
-             filemap_fdatawrite(inode->i_mapping);
+
+     if (CIFS_I(inode)->clientCanCacheAll)
+             return written;
+
+     rc = filemap_fdatawrite(inode->i_mapping);
+     if (rc)
+             cFYI(1, "cifs_file_aio_write: %d rc on %p inode", rc, inode);
+
       return written;
  }
 
@@ -748,7 +968,7 @@ const struct file_operations cifs_file_strict_ops = {
       .read = do_sync_read,
       .write = do_sync_write,
       .aio_read = cifs_strict_read,
-     .aio_write = cifs_file_aio_write,
+     .aio_write = cifs_strict_write,
       .open = cifs_open,
       .release = cifs_close,
       .lock = cifs_lock,
@@ -804,7 +1024,7 @@ const struct file_operations cifs_file_strict_nobrl_ops = {
       .read = do_sync_read,
       .write = do_sync_write,
       .aio_read = cifs_strict_read,
-     .aio_write = cifs_file_aio_write,
+     .aio_write = cifs_strict_write,
       .open = cifs_open,
       .release = cifs_close,
       .fsync = cifs_strict_fsync,
fs/cifs/cifsproto.h +2 -0
index e6d1481..e7d6c24 100644
@@ -79,6 +79,8 @@ extern int checkSMB(struct smb_hdr *smb, __u16 mid, unsigned int length);
  extern bool is_valid_oplock_break(struct smb_hdr *smb,
                                 struct TCP_Server_Info *);
  extern bool is_size_safe_to_change(struct cifsInodeInfo *, __u64 eof);
+extern void cifs_update_eof(struct cifsInodeInfo *cifsi, loff_t offset,
+                         unsigned int bytes_written);
  extern struct cifsFileInfo *find_writable_file(struct cifsInodeInfo *, bool);
  extern struct cifsFileInfo *find_readable_file(struct cifsInodeInfo *, bool);
  extern unsigned int smbCalcSize(struct smb_hdr *ptr);
fs/cifs/file.c +0 -0
index bc53e9e..8260f53 100644
@@ -924,7 +924,7 @@ cifs_write_timeout(struct cifsInodeInfo *cifsi, loff_t offset)
  }
 
Quoted 11 lines

fs/cifs/file.c cifs_write_timeout(struct cifsInodeInfo *cifsi, loff_t offset) · line 930

Quoted 6 lines
fs/cifs/file.c +1 -1
  /* update the file size (if needed) after a write */
-static void
+void
  cifs_update_eof(struct cifsInodeInfo *cifsi, loff_t offset,
                     unsigned int bytes_written)
  {

--
Jeff Layton <[email protected]>

Ok, I understood your idea and agree with it. iov_iter is what I should use for this. So, I am going to rewrite this patch also with kmap changes you wrote about above.

Signature · 2 lines
Best regards,
Pavel Shilovsky.