Skip to contentSkip to conversation

cifs: Avoid calling unlock_page() twice in cifs_readpage() when using fscache

PATCH In review Sachin Prabhu 5 participants 6 messages Sep 12, 2013 – Sep 13, 2013 4 messages not in the archive

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

PATCH 1/1 cifs: Avoid calling unlock_page() twice in cifs_readpage() when using fscache

When reading a single page with cifs_readpage(), we make a call to fscache_read_or_alloc_page() which once done, asynchronously calls the completion function cifs_readpage_from_fscache_complete(). This completion function unlocks the page once it has been populated from cache. The module then attempts to unlock the page a second time in cifs_readpage() which leads to warning messages.

In case of a successful call to fscache_read_or_alloc_page() we should skip the second unlock_page() since this will be called by the cifs_readpage_from_fscache_complete() once the page has been populated by fscache.

With the modifications to cifs_readpage_worker(), we will need to re-grab the page lock in cifs_write_begin().

---
 fs/cifs/file.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
fs/cifs/file.c +7 -3
index 69e8431..98e5222 100644
@@ -3423,6 +3423,7 @@ static int cifs_readpage_worker(struct file *file, struct page *page,
 io_error:
 	kunmap(page);
 	page_cache_release(page);
+	unlock_page(page);
 
 read_complete:
 	return rc;
@@ -3447,8 +3448,6 @@ static int cifs_readpage(struct file *file, struct page *page)
 
 	rc = cifs_readpage_worker(file, page, &offset);
 
-	unlock_page(page);
-
 	free_xid(xid);
 	return rc;
 }
@@ -3502,6 +3501,7 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 			loff_t pos, unsigned len, unsigned flags,
 			struct page **pagep, void **fsdata)
 {
+	int oncethru = 0;
 	pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 	loff_t offset = pos & (PAGE_CACHE_SIZE - 1);
 	loff_t page_start = pos & PAGE_MASK;
@@ -3511,6 +3511,7 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 
 	cifs_dbg(FYI, "write_begin from %lld len %d\n", (long long)pos, len);
 
+start:
 	page = grab_cache_page_write_begin(mapping, index, flags);
 	if (!page) {
 		rc = -ENOMEM;
@@ -3552,13 +3553,16 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 		}
 	}
 
-	if ((file->f_flags & O_ACCMODE) != O_WRONLY) {
+	if ((file->f_flags & O_ACCMODE) != O_WRONLY && !oncethru) {
 		/*
 		 * might as well read a page, it is fast enough. If we get
 		 * an error, we don't need to return it. cifs_write_end will
 		 * do a sync write instead since PG_uptodate isn't set.
 		 */
 		cifs_readpage_worker(file, page, &page_start);
+		page_cache_release(page);
+		oncethru = 1;
+		goto start;
 	} else {
 		/* we could try using another file handle if there is one -
 		   but how would we lock it to prevent close of that handle
Signature · 1 line
1.8.3.1

A message that is not in the archive (sent off-list or not yet received): 1378997931-19954-1-git-send-email-sprabhu-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org

Reviewed-by
#2 lore raw
Quoted 65 lines · On Thu, 12 Sep 2013 15:58:51 +0100 Sachin Prabhu…

When reading a single page with cifs_readpage(), we make a call to fscache_read_or_alloc_page() which once done, asynchronously calls the completion function cifs_readpage_from_fscache_complete(). This completion function unlocks the page once it has been populated from cache. The module then attempts to unlock the page a second time in cifs_readpage() which leads to warning messages.

In case of a successful call to fscache_read_or_alloc_page() we should skip the second unlock_page() since this will be called by the cifs_readpage_from_fscache_complete() once the page has been populated by fscache.

With the modifications to cifs_readpage_worker(), we will need to re-grab the page lock in cifs_write_begin().

---
 fs/cifs/file.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
fs/cifs/file.c +4 -3
index 69e8431..98e5222 100644
@@ -3423,6 +3423,7 @@ static int cifs_readpage_worker(struct file *file, struct page *page,
 io_error:
 	kunmap(page);
 	page_cache_release(page);
+	unlock_page(page);
 
 read_complete:
 	return rc;
@@ -3447,8 +3448,6 @@ static int cifs_readpage(struct file *file, struct page *page)
 
 	rc = cifs_readpage_worker(file, page, &offset);
 
-	unlock_page(page);
-
 	free_xid(xid);
 	return rc;
 }
@@ -3502,6 +3501,7 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 			loff_t pos, unsigned len, unsigned flags,
 			struct page **pagep, void **fsdata)
 {
+	int oncethru = 0;
 	pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 	loff_t offset = pos & (PAGE_CACHE_SIZE - 1);
 	loff_t page_start = pos & PAGE_MASK;
@@ -3511,6 +3511,7 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 
 	cifs_dbg(FYI, "write_begin from %lld len %d\n", (long long)pos, len);
 
+start:
 	page = grab_cache_page_write_begin(mapping, index, flags);
 	if (!page) {
 		rc = -ENOMEM;
@@ -3552,13 +3553,16 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 		}
 	}
 
-	if ((file->f_flags & O_ACCMODE) != O_WRONLY) {
+	if ((file->f_flags & O_ACCMODE) != O_WRONLY && !oncethru) {
 		/*
 		 * might as well read a page, it is fast enough. If we get

fs/cifs/file.c static int cifs_write_begin(struct file *file, struct address_space *mapping, · line 3568

Quoted 10 lines
fs/cifs/file.c +3 -0
 		 * an error, we don't need to return it. cifs_write_end will
 		 * do a sync write instead since PG_uptodate isn't set.
 		 */
 		cifs_readpage_worker(file, page, &page_start);
+		page_cache_release(page);
+		oncethru = 1;
+		goto start;
 	} else {
 		/* we could try using another file handle if there is one -
 		   but how would we lock it to prevent close of that handle

Looks correct. Nice catch!

A message that is not in the archive (sent off-list or not yet received): 20130912113527.2eb9b5cd-9yPaYZwiELC+kQycOl6kW4xkIHaj4LzF@public.gmane.org

#4 lore raw
Quoted 23 lines · On Thu, 12 Sep 2013 11:35:27 -0400 Jeff Layton…
Quoted 20 lines · On Thu, 12 Sep 2013 15:58:51 +0100 Sachin Prabhu…

When reading a single page with cifs_readpage(), we make a call to fscache_read_or_alloc_page() which once done, asynchronously calls the completion function cifs_readpage_from_fscache_complete(). This completion function unlocks the page once it has been populated from cache. The module then attempts to unlock the page a second time in cifs_readpage() which leads to warning messages.

In case of a successful call to fscache_read_or_alloc_page() we should skip the second unlock_page() since this will be called by the cifs_readpage_from_fscache_complete() once the page has been populated by fscache.

With the modifications to cifs_readpage_worker(), we will need to re-grab the page lock in cifs_write_begin().

---
 fs/cifs/file.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
Quoted 10 lines

fs/cifs/file.c static int cifs_readpage_worker(struct file *file, struct page *page, · line 3427

Quoted 10 lines
fs/cifs/file.c +1 -0
index 69e8431..98e5222 100644
@@ -3423,6 +3423,7 @@ static int cifs_readpage_worker(struct file *file, struct page *page,
 io_error:
 	kunmap(page);
 	page_cache_release(page);
+	unlock_page(page);
 

Actually...one preexisting bug that you should probably fix while you're in there. It's a bad idea to unlock the page *after* you release the reference to it. You probably want to move that unlock_page call before the page_cache_release.

OTOH...it's not clear to me why we're bumping the refcount on the page at all in cifs_readpage_worker. Clearly we must have a reference to it already or it won't be ok to just pass in the pointer to it. Maybe it'd be better to just make it clear that cifs_readpage_worker must be called with the page pinned and get rid of the extra refcounting in that function altogether.

Sound reasonable?

Quoted 53 lines
Quoted 35 lines
fs/cifs/file.c +3 -3
 read_complete:
 	return rc;
@@ -3447,8 +3448,6 @@ static int cifs_readpage(struct file *file, struct page *page)
 
 	rc = cifs_readpage_worker(file, page, &offset);
 
-	unlock_page(page);
-
 	free_xid(xid);
 	return rc;
 }
@@ -3502,6 +3501,7 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 			loff_t pos, unsigned len, unsigned flags,
 			struct page **pagep, void **fsdata)
 {
+	int oncethru = 0;
 	pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 	loff_t offset = pos & (PAGE_CACHE_SIZE - 1);
 	loff_t page_start = pos & PAGE_MASK;
@@ -3511,6 +3511,7 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 
 	cifs_dbg(FYI, "write_begin from %lld len %d\n", (long long)pos, len);
 
+start:
 	page = grab_cache_page_write_begin(mapping, index, flags);
 	if (!page) {
 		rc = -ENOMEM;
@@ -3552,13 +3553,16 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 		}
 	}
 
-	if ((file->f_flags & O_ACCMODE) != O_WRONLY) {
+	if ((file->f_flags & O_ACCMODE) != O_WRONLY && !oncethru) {
 		/*
 		 * might as well read a page, it is fast enough. If we get

fs/cifs/file.c static int cifs_write_begin(struct file *file, struct address_space *mapping, · line 3568

Quoted 10 lines
fs/cifs/file.c +3 -0
 		 * an error, we don't need to return it. cifs_write_end will
 		 * do a sync write instead since PG_uptodate isn't set.
 		 */
 		cifs_readpage_worker(file, page, &page_start);
+		page_cache_release(page);
+		oncethru = 1;
+		goto start;
 	} else {
 		/* we could try using another file handle if there is one -
 		   but how would we lock it to prevent close of that handle

Looks correct. Nice catch!

--
To unsubscribe from this list: send the line "unsubscribe linux-cifs" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

Signature · 1 line
Jeff Layton <[email protected]>

A message that is not in the archive (sent off-list or not yet received): 20130912114209.20319df5-9yPaYZwiELC+kQycOl6kW4xkIHaj4LzF@public.gmane.org

#6 lore raw

Also would be helpful to have the link to any external bug report (if any)

Quoted 108 lines · Jeff Layton
Quoted 23 lines · On Thu, 12 Sep 2013 11:35:27 -0400 Jeff Layton…
Quoted 20 lines · On Thu, 12 Sep 2013 15:58:51 +0100 Sachin Prabhu…

When reading a single page with cifs_readpage(), we make a call to fscache_read_or_alloc_page() which once done, asynchronously calls the completion function cifs_readpage_from_fscache_complete(). This completion function unlocks the page once it has been populated from cache. The module then attempts to unlock the page a second time in cifs_readpage() which leads to warning messages.

In case of a successful call to fscache_read_or_alloc_page() we should skip the second unlock_page() since this will be called by the cifs_readpage_from_fscache_complete() once the page has been populated by fscache.

With the modifications to cifs_readpage_worker(), we will need to re-grab the page lock in cifs_write_begin().

---
 fs/cifs/file.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
Quoted 10 lines

fs/cifs/file.c static int cifs_readpage_worker(struct file *file, struct page *page, · line 3427

Quoted 10 lines
fs/cifs/file.c +1 -0
index 69e8431..98e5222 100644
@@ -3423,6 +3423,7 @@ static int cifs_readpage_worker(struct file *file, struct page *page,
 io_error:
    kunmap(page);
    page_cache_release(page);
+   unlock_page(page);
 

Actually...one preexisting bug that you should probably fix while you're in there. It's a bad idea to unlock the page *after* you release the reference to it. You probably want to move that unlock_page call before the page_cache_release.

OTOH...it's not clear to me why we're bumping the refcount on the page at all in cifs_readpage_worker. Clearly we must have a reference to it already or it won't be ok to just pass in the pointer to it. Maybe it'd be better to just make it clear that cifs_readpage_worker must be called with the page pinned and get rid of the extra refcounting in that function altogether.

Sound reasonable?

Quoted 48 lines
Quoted 35 lines
fs/cifs/file.c +3 -3
 read_complete:
    return rc;
@@ -3447,8 +3448,6 @@ static int cifs_readpage(struct file *file, struct page *page)
 
    rc = cifs_readpage_worker(file, page, &offset);
 
-   unlock_page(page);
-
    free_xid(xid);
    return rc;
 }
@@ -3502,6 +3501,7 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
                    loff_t pos, unsigned len, unsigned flags,
                    struct page **pagep, void **fsdata)
 {
+   int oncethru = 0;
    pgoff_t index = pos >> PAGE_CACHE_SHIFT;
    loff_t offset = pos & (PAGE_CACHE_SIZE - 1);
    loff_t page_start = pos & PAGE_MASK;
@@ -3511,6 +3511,7 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 
    cifs_dbg(FYI, "write_begin from %lld len %d\n", (long long)pos, len);
 
+start:
    page = grab_cache_page_write_begin(mapping, index, flags);
    if (!page) {
            rc = -ENOMEM;
@@ -3552,13 +3553,16 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
            }
    }
 
-   if ((file->f_flags & O_ACCMODE) != O_WRONLY) {
+   if ((file->f_flags & O_ACCMODE) != O_WRONLY && !oncethru) {
            /*
             * might as well read a page, it is fast enough. If we get

fs/cifs/file.c static int cifs_write_begin(struct file *file, struct address_space *mapping, · line 3568

Quoted 10 lines
fs/cifs/file.c +3 -0
             * an error, we don't need to return it. cifs_write_end will
             * do a sync write instead since PG_uptodate isn't set.
             */
            cifs_readpage_worker(file, page, &page_start);
+           page_cache_release(page);
+           oncethru = 1;
+           goto start;
    } else {
            /* we could try using another file handle if there is one -
               but how would we lock it to prevent close of that handle

Looks correct. Nice catch!

Quoted 5 lines

--
To unsubscribe from this list: send the line "unsubscribe linux-cifs" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

--
Jeff Layton <[email protected]>

Signature · 3 lines
Thanks,

Steve
Acked-by
#7 lore raw
In reply to a message not in the archive

Looking at this:

	static int cifs_readpage_worker(struct file *file, struct page *page,
		loff_t *poffset)
	{
		char *read_data;
		int rc;

		/* Is the page cached? */
		rc = cifs_readpage_from_fscache(file_inode(file), page);
		if (rc == 0)
			goto read_complete;

Should you return here if rc == -ENOMEM or -ERESTARTSYS? That might break cifs_write_begin() though - which perhaps ought to check the return value.

Your patch, however, looks okay otherwise, so feel free to add:

David

A message that is not in the archive (sent off-list or not yet received): 31932.1379079033-S6HVgzuS8uM4Awkfq6JHfwNdhmdF6hFW@public.gmane.org

#9 lore raw
Quoted 13 lines · David Howells

Looking at this:

	static int cifs_readpage_worker(struct file *file, struct page *page,
		loff_t *poffset)
	{
		char *read_data;
		int rc;

		/* Is the page cached? */
		rc = cifs_readpage_from_fscache(file_inode(file), page);
		if (rc == 0)
			goto read_complete;
Quoted 8 lines

Should you return here if rc == -ENOMEM or -ERESTARTSYS? That might break cifs_write_begin() though - which perhaps ought to check the return value.

Your patch, however, looks okay otherwise, so feel free to add:

David

According to the comment already above the call to cifs_readpage_worker() in cifs_write_begin():

/*
* might as well read a page, it is fast enough. If we get
* an error, we don't need to return it. cifs_write_end will
* do a sync write instead since PG_uptodate isn't set. */

With the patch, we ignore the error and set oncethru to 1 so that we don't attempt to call cifs_readpage_worker() again and let cifs_write_end do a sync_write.

Sachin Prabhu