Skip to contentSkip to conversation

Change page reference hanlding semantic of page cache

RFC In review Minchan Kim 6 participants 28 messages Dec 17, 2010 – Jan 10, 2011

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

cover Change page reference hanlding semantic of page cache

I copy description of 1/5.

Now we add page reference on add_to_page_cache but doesn't drop it in remove_from_page_cache. Such asymmetric makes confusing about page reference so that caller should notice it and comment why they release page reference. It's not good API.

Long time ago, Hugh tried it[1] but gave up of reason which reiser4's drop_page had to unlock the page between removing it from page cache and doing the page_cache_release. But now the situation is changed. I think at least things in current mainline doesn't have any obstacles. The problem is fs or somethings out of mainline. If it has done such thing like reiser4, this patch could be a problem.

Do anyone know the such things? Do we care about things out of mainline?

[1] http://lkml.org/lkml/2004/10/24/140

Minchan Kim (5):
drop page reference on remove_from_page_cache
fuse: Remove unnecessary page release
tlbfs: Remove unnecessary page release
swap: Remove unnecessary page release
truncate: Remove unnecessary page release

 fs/fuse/dev.c        |    1 -
 fs/hugetlbfs/inode.c |    1 -
 mm/filemap.c         |   12 ++++++++++++
 mm/shmem.c           |    1 -
 mm/truncate.c        |    1 -
 5 files changed, 12 insertions(+), 4 deletions(-)
#1 lore raw

RFC 1/5 drop page reference on remove_from_page_cache

Now we add page reference on add_to_page_cache but doesn't drop it in remove_from_page_cache. Such asymmetric makes confusing about page reference so that caller should notice it and comment why they release page reference. It's not good API.

Long time ago, Hugh tried it[1] but gave up of reason which reiser4's drop_page had to unlock the page between removing it from page cache and doing the page_cache_release. But now the situation is changed. I think at least things in current mainline doesn't have any obstacles. The problem is fs or somethings out of mainline. If it has done such thing like reiser4, this patch could be a problem.

Do anyone know the such things? Do we care about things out of mainline?

Note :
The comment of remove_from_page_cache make by copy & paste & s/swap/page/ from delete_from_swap_cache.

[1] http://lkml.org/lkml/2004/10/24/140

---
 mm/filemap.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)
mm/filemap.c +12 -0
index 095c393..fb9db36 100644
@@ -148,6 +148,16 @@ void __remove_from_page_cache(struct page *page)
 	}
 }
 
+/**
+ * remove_from_page_cache - remove page from page cache
+ *
+ * @page: the page which the kernel is trying to remove from page cache
+ *
+ * This must be called only on pages that have
+ * been verified to be in the page cache and locked.
+ * It will never put the page into the free list,
+ * the caller has a reference on the page.
+ */
 void remove_from_page_cache(struct page *page)
 {
 	struct address_space *mapping = page->mapping;
@@ -163,6 +173,8 @@ void remove_from_page_cache(struct page *page)
 
 	if (freepage)
 		freepage(page);
+
+	page_cache_release(page);
 }
 EXPORT_SYMBOL(remove_from_page_cache);
 
Signature · 8 lines
1.7.0.4

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#3 lore raw

RFC 2/5 fuse: Remove unnecessary page release

In reply to Minchan Kim

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 fs/fuse/dev.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)
fs/fuse/dev.c +0 -1
index cf8d28d..4adaf4b 100644
@@ -738,7 +738,6 @@ static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 		goto out_fallback_unlock;
 
 	remove_from_page_cache(oldpage);
-	page_cache_release(oldpage);
 
 	err = add_to_page_cache_locked(newpage, mapping, index, GFP_KERNEL);
 	if (err) {
Signature · 8 lines
1.7.0.4

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#5 lore raw

RFC 3/5 tlbfs: Remove unnecessary page release

In reply to Minchan Kim

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 fs/hugetlbfs/inode.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)
fs/hugetlbfs/inode.c +0 -1
index 9885082..4f32fb6 100644
@@ -333,7 +333,6 @@ static void truncate_huge_page(struct page *page)
 	cancel_dirty_page(page, /* No IO accounting for huge pages? */0);
 	ClearPageUptodate(page);
 	remove_from_page_cache(page);
-	put_page(page);
 }
 
 static void truncate_hugepages(struct inode *inode, loff_t lstart)
Signature · 8 lines
1.7.0.4

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#9 lore raw

RFC 4/5 swap: Remove unnecessary page release

In reply to Minchan Kim

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/shmem.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)
mm/shmem.c +0 -1
index fa9acc9..16800f2 100644
@@ -1091,7 +1091,6 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 		spin_unlock(&info->lock);
 		swap_shmem_alloc(swap);
 		BUG_ON(page_mapped(page));
-		page_cache_release(page);	/* pagecache ref */
 		swap_writepage(page, wbc);
 		if (inode) {
 			mutex_lock(&shmem_swaplist_mutex);
Signature · 8 lines
1.7.0.4

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#11 lore raw

RFC 5/5 truncate: Remove unnecessary page release

In reply to Minchan Kim

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
 	 * calls cleancache_put_page (and note page->mapping is now NULL)
 	 */
 	cleancache_flush_page(mapping, page);
-	page_cache_release(page);	/* pagecache ref */
 	return 0;
 }
 
Signature · 1 line
1.7.0.4
#2 lore raw

Re: [RFC 1/5] drop page reference on remove_from_page_cache

In reply to Minchan Kim
Quoted 18 lines · On Sat, 18 Dec 2010 02:13:36 +0900 Minchan Kim <[email protected]> wrote:

Now we add page reference on add_to_page_cache but doesn't drop it in remove_from_page_cache. Such asymmetric makes confusing about page reference so that caller should notice it and comment why they release page reference. It's not good API.

Long time ago, Hugh tried it[1] but gave up of reason which reiser4's drop_page had to unlock the page between removing it from page cache and doing the page_cache_release. But now the situation is changed. I think at least things in current mainline doesn't have any obstacles. The problem is fs or somethings out of mainline. If it has done such thing like reiser4, this patch could be a problem.

Do anyone know the such things? Do we care about things out of mainline?

Note :
The comment of remove_from_page_cache make by copy & paste & s/swap/page/ from delete_from_swap_cache.

Quoted 5 lines

[1] http://lkml.org/lkml/2004/10/24/140

I like this.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Reviewed-by on v1 2/5
#4 lore raw

Re: [RFC 2/5] fuse: Remove unnecessary page release

In reply to Minchan Kim
Quoted 7 lines · On Sat, 18 Dec 2010 02:13:37 +0900 Minchan Kim <[email protected]> wrote:

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Reviewed-by on v1 3/5
#6 lore raw

Re: [RFC 3/5] tlbfs: Remove unnecessary page release

In reply to Minchan Kim
Quoted 6 lines · On Sat, 18 Dec 2010 02:13:38 +0900 Minchan Kim <[email protected]> wrote:

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Reviewed-by on v1 4/5
#10 lore raw

Re: [RFC 4/5] swap: Remove unnecessary page release

In reply to Minchan Kim
Quoted 7 lines · On Sat, 18 Dec 2010 02:13:39 +0900 Minchan Kim <[email protected]> wrote:

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Reviewed-by on v1 5/5
#12 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

In reply to Minchan Kim
Quoted 8 lines · On Sat, 18 Dec 2010 02:13:40 +0900 Minchan Kim <[email protected]> wrote:

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#13 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

In reply to Minchan Kim
Quoted 17 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)
mm/truncate.c +0 -0
index 9ee5673..8decb93 100644

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 5 lines
mm/truncate.c +0 -1
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
 	 * calls cleancache_put_page (and note page->mapping is now NULL)
 	 */
 	cleancache_flush_page(mapping, page);
-	page_cache_release(page);	/* pagecache ref */
 	return 0;

Do we _always_ have stable page reference here? IOW, I can assume cleancache_flush_page() doesn't cause NULL deref?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#14 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

Quoted 12 lines · KOSAKI Motohiro
Quoted 12 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -

 1 files changed, 0 insertions(+), 1 deletions(-)
Quoted 12 lines

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 10 lines
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
        * calls cleancache_put_page (and note page->mapping is now NULL)
        */
       cleancache_flush_page(mapping, page);
-     page_cache_release(page);       /* pagecache ref */
       return 0;

Do we _always_ have stable page reference here? IOW, I can assume

I think so.
Because the page is locked so caller have to hold a ref to unlock it.

Quoted 5 lines

cleancache_flush_page() doesn't cause NULL deref?

Signature · 2 lines
Kind regards,
Minchan Kim
#15 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

Quoted 27 lines
Quoted 12 lines · KOSAKI Motohiro
Quoted 12 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -

 1 files changed, 0 insertions(+), 1 deletions(-)
Quoted 12 lines

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 10 lines
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
        * calls cleancache_put_page (and note page->mapping is now NULL)
        */
       cleancache_flush_page(mapping, page);
-     page_cache_release(page);       /* pagecache ref */
       return 0;

Do we _always_ have stable page reference here? IOW, I can assume

Quoted 2 lines

I think so.
Because the page is locked so caller have to hold a ref to unlock it.

Hmm...

Perhaps, I'm missing something. But I think __memory_failure() only lock compaund_head page. not all. example.

Quoted 2 lines
Quoted 2 lines

cleancache_flush_page() doesn't cause NULL deref?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#19 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

In reply to KOSAKI Motohiro
Quoted 12 lines · On Mon, 20 Dec 2010 11:21:52 +0900 (JST) KOSAKI Motohiro…
Quoted 12 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)
Quoted 14 lines

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 10 lines
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
 	 * calls cleancache_put_page (and note page->mapping is now NULL)
 	 */
 	cleancache_flush_page(mapping, page);
-	page_cache_release(page);	/* pagecache ref */
 	return 0;

Do we _always_ have stable page reference here? IOW, I can assume cleancache_flush_page() doesn't cause NULL deref?

Hmm, my review was bad.

I think cleancache_flush_page() here should eat (mapping, index) as argument rather than "page".

BTW, I can't understand
==
void __cleancache_flush_page(struct address_space *mapping, struct page *page)

{
        /* careful... page->mapping is NULL sometimes when this is called */
        int pool_id = mapping->host->i_sb->cleancache_poolid;
        struct cleancache_filekey key = { .u.key = { 0 } };

==

Why above is safe...
I think (mapping,index) should be passed instead of page.

-Kame

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#16 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

In reply to KOSAKI Motohiro
Quoted 27 lines · KOSAKI Motohiro
Quoted 27 lines
Quoted 24 lines · KOSAKI Motohiro
Quoted 17 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -

 1 files changed, 0 insertions(+), 1 deletions(-)
mm/truncate.c +0 -0
index 9ee5673..8decb93 100644

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 5 lines
mm/truncate.c +0 -1
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
        * calls cleancache_put_page (and note page->mapping is now NULL)
        */
       cleancache_flush_page(mapping, page);
-     page_cache_release(page);       /* pagecache ref */
       return 0;

Do we _always_ have stable page reference here? IOW, I can assume

Quoted 7 lines
Quoted 2 lines

I think so.
Because the page is locked so caller have to hold a ref to unlock it.

Hmm...

Perhaps, I'm missing something. But I think  __memory_failure() only lock compaund_head page. not all. example.

The page passed truncate_complete_page is only head page? Is it possible to pass the page which isn't head of compound in truncate_complete_page?

Quoted 7 lines
Quoted 3 lines
Quoted 3 lines

cleancache_flush_page() doesn't cause NULL deref?

Signature · 9 lines
Kind regards,
Minchan Kim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#20 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

In reply to KAMEZAWA Hiroyuki
Quoted 43 lines · KAMEZAWA Hiroyuki
Quoted 12 lines · On Mon, 20 Dec 2010 11:21:52 +0900 (JST) KOSAKI Motohiro…
Quoted 12 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -

 1 files changed, 0 insertions(+), 1 deletions(-)
Quoted 14 lines

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 10 lines
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
      * calls cleancache_put_page (and note page->mapping is now NULL)
      */
     cleancache_flush_page(mapping, page);
-   page_cache_release(page);       /* pagecache ref */
     return 0;

Do we _always_ have stable page reference here? IOW, I can assume cleancache_flush_page() doesn't cause NULL deref?

Hmm, my review was bad.

I think cleancache_flush_page() here should eat (mapping, index) as argument rather than "page".

BTW,  I can't understand
==
void __cleancache_flush_page(struct address_space *mapping, struct page *page) {
       /* careful... page->mapping is NULL sometimes when this is called */ int pool_id = mapping->host->i_sb->cleancache_poolid; struct cleancache_filekey key = { .u.key = { 0 } };
==

Quoted 2 lines

Why above is safe...
I think (mapping,index) should be passed instead of page.

I don't think current code isn't safe.

void __cleancache_flush_page(struct address_space *mapping, struct page *page)

{
        /* careful... page->mapping is NULL sometimes when this is called */
        int pool_id = mapping->host->i_sb->cleancache_poolid;
        struct cleancache_filekey key = { .u.key = { 0 } };

        if (pool_id >= 0) {
                VM_BUG_ON(!PageLocked(page));

it does check PageLocked. So caller should hold a page reference to prevent freeing ramined PG_locked
If the caller doesn't hold a ref of page, I think it's BUG of caller.

In our case, caller calls truncate_complete_page have to make sure it, I think.

Quoted 5 lines

-Kame

Signature · 9 lines
Kind regards,
Minchan Kim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#17 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

In reply to Minchan Kim
Quoted 32 lines
Quoted 30 lines · KOSAKI Motohiro
Quoted 29 lines
Quoted 12 lines · KOSAKI Motohiro
Quoted 12 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -

 1 files changed, 0 insertions(+), 1 deletions(-)
Quoted 12 lines

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 10 lines
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
        * calls cleancache_put_page (and note page->mapping is now NULL)
        */
       cleancache_flush_page(mapping, page);
-     page_cache_release(page);       /* pagecache ref */
       return 0;

Do we _always_ have stable page reference here? IOW, I can assume

I think so.
Because the page is locked so caller have to hold a ref to unlock it.

Quoted 8 lines
Quoted 4 lines

Hmm...

Perhaps, I'm missing something. But I think  __memory_failure() only lock compaund_head page. not all. example.

The page passed truncate_complete_page is only head page? Is it possible to pass the page which isn't head of compound in truncate_complete_page?

I dunno, really. My five miniture grep found following logic. therefore I asked you.

__memory_failure()

{
        p = pfn_to_page(pfn);
        hpage = compound_head(p);
(snip)
        res = -EBUSY;
        for (ps = error_states;; ps++) {
                if ((p->flags & ps->mask) == ps->res) {
                        res = page_action(ps, p, pfn);  // call truncate here
                        break;
                }
        }
out:
        unlock_page(hpage);
}

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#18 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

Quoted 45 lines · KOSAKI Motohiro
Quoted 32 lines
Quoted 30 lines · KOSAKI Motohiro
Quoted 29 lines
Quoted 12 lines · KOSAKI Motohiro
Quoted 12 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -

 1 files changed, 0 insertions(+), 1 deletions(-)
Quoted 12 lines

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 10 lines
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
        * calls cleancache_put_page (and note page->mapping is now NULL)
        */
       cleancache_flush_page(mapping, page);
-     page_cache_release(page);       /* pagecache ref */
       return 0;

Do we _always_ have stable page reference here? IOW, I can assume

I think so.
Because the page is locked so caller have to hold a ref to unlock it.

Quoted 8 lines
Quoted 4 lines

Hmm...

Perhaps, I'm missing something. But I think  __memory_failure() only lock compaund_head page. not all. example.

The page passed truncate_complete_page is only head page? Is it possible to pass the page which isn't head of compound in truncate_complete_page?

I dunno, really. My five miniture grep found following logic. therefore I asked you.

Quoted 17 lines

__memory_failure()
{
       p = pfn_to_page(pfn);
       hpage = compound_head(p);
(snip)
       res = -EBUSY;
       for (ps = error_states;; ps++) {
               if ((p->flags & ps->mask) == ps->res) {
                       res = page_action(ps, p, pfn);  // call truncate here break;
               }
       }
out:
       unlock_page(hpage);
}

AFAIK, We have to handle head page when we handle compound page. Internal page handling logic about tail pages is hidden by compound page internal.

So I think memory_failure also don't have a problem. For needing double check, Cced Andi.

Thanks for the review, KOSAKI.

Signature · 9 lines
Kind regards,
Minchan Kim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#21 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

In reply to Minchan Kim
Quoted 59 lines · On Mon, 20 Dec 2010 11:58:50 +0900 Minchan Kim <[email protected]> wrote:
Quoted 43 lines · KAMEZAWA Hiroyuki
Quoted 12 lines · On Mon, 20 Dec 2010 11:21:52 +0900 (JST) KOSAKI Motohiro…
Quoted 12 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -

 1 files changed, 0 insertions(+), 1 deletions(-)
Quoted 14 lines

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 10 lines
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
      * calls cleancache_put_page (and note page->mapping is now NULL)
      */
     cleancache_flush_page(mapping, page);
-   page_cache_release(page);       /* pagecache ref */
     return 0;

Do we _always_ have stable page reference here? IOW, I can assume cleancache_flush_page() doesn't cause NULL deref?

Hmm, my review was bad.

I think cleancache_flush_page() here should eat (mapping, index) as argument rather than "page".

BTW,  I can't understand
==
void __cleancache_flush_page(struct address_space *mapping, struct page *page) {
       /* careful... page->mapping is NULL sometimes when this is called */ int pool_id = mapping->host->i_sb->cleancache_poolid; struct cleancache_filekey key = { .u.key = { 0 } };
==

Quoted 2 lines

Why above is safe...
I think (mapping,index) should be passed instead of page.

I don't think current code isn't safe.

void __cleancache_flush_page(struct address_space *mapping, struct page *page)

{
        /* careful... page->mapping is NULL sometimes when this is called */
        int pool_id = mapping->host->i_sb->cleancache_poolid;
        struct cleancache_filekey key = { .u.key = { 0 } };

        if (pool_id >= 0) {
                VM_BUG_ON(!PageLocked(page));
Quoted 6 lines

it does check PageLocked. So caller should hold a page reference to prevent freeing ramined PG_locked
If the caller doesn't hold a ref of page, I think it's BUG of caller.

In our case, caller calls truncate_complete_page have to make sure it, I think.

Ah, my point is that this function trust page->index even if page->mapping is reset to NULL. And I'm not sure that there are any race that an other thread add a replacement page for (mapping, index) while a thread call this function.

Thanks,
-Kame

#22 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

Quoted 66 lines · KAMEZAWA Hiroyuki
Quoted 63 lines · On Mon, 20 Dec 2010 11:58:50 +0900 Minchan Kim <[email protected]> wrote:
Quoted 43 lines · KAMEZAWA Hiroyuki
Quoted 12 lines · On Mon, 20 Dec 2010 11:21:52 +0900 (JST) KOSAKI Motohiro…
Quoted 12 lines

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

---
 mm/truncate.c |    1 -

 1 files changed, 0 insertions(+), 1 deletions(-)
Quoted 14 lines

mm/truncate.c truncate_complete_page(struct address_space *mapping, struct page *page) · line 117

Quoted 10 lines
mm/truncate.c +0 -1
index 9ee5673..8decb93 100644
@@ -114,7 +114,6 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
      * calls cleancache_put_page (and note page->mapping is now NULL)
      */
     cleancache_flush_page(mapping, page);
-   page_cache_release(page);       /* pagecache ref */
     return 0;

Do we _always_ have stable page reference here? IOW, I can assume cleancache_flush_page() doesn't cause NULL deref?

Hmm, my review was bad.

I think cleancache_flush_page() here should eat (mapping, index) as argument rather than "page".

BTW,  I can't understand
==
void __cleancache_flush_page(struct address_space *mapping, struct page *page) {
       /* careful... page->mapping is NULL sometimes when this is called */ int pool_id = mapping->host->i_sb->cleancache_poolid; struct cleancache_filekey key = { .u.key = { 0 } };
==

Quoted 2 lines

Why above is safe...
I think (mapping,index) should be passed instead of page.

I don't think current code isn't safe.

void __cleancache_flush_page(struct address_space *mapping, struct page *page) {
        /* careful... page->mapping is NULL sometimes when this is called */ int pool_id = mapping->host->i_sb->cleancache_poolid; struct cleancache_filekey key = { .u.key = { 0 } };

        if (pool_id >= 0) {
                VM_BUG_ON(!PageLocked(page));

it does check PageLocked. So caller should hold a page reference to prevent freeing ramined PG_locked
If the caller doesn't hold a ref of page, I think it's BUG of caller.

Quoted 6 lines
Quoted 2 lines

In our case, caller calls truncate_complete_page have to make sure it, I think.

Ah, my point is that this function trust page->index even if page->mapping is reset to NULL. And I'm not sure that there are any race that an other thread add a replacement page for (mapping, index) while a thread call this function.

Because the page is locked and is detached from page cache, I guess it's no problem.
Anyway, I think It's off-topic.

Quoted 7 lines

Thanks,
-Kame

Signature · 9 lines
Kind regards,
Minchan Kim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#23 lore raw

Re: [RFC 5/5] truncate: Remove unnecessary page release

Quoted 4 lines · On Mon, 20 Dec 2010 17:09:11 +0900 Minchan Kim <[email protected]> wrote:

Because the page is locked and is detached from page cache, I guess it's no problem.
Anyway, I think It's off-topic.

yes.

Thanks,
-Kame

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#25 lore raw
Quoted 3 lines · Christoph Hellwig

You'll need to merge all patches into one, otherwise you create really nasty memory leaks when bisecting between them.

Okay. I will resend.

Thanks for the notice, Christoph.

Signature · 9 lines
Kind regards,
Minchan Kim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
maintainer
#26 lore raw
Quoted 8 lines · Minchan Kim
Quoted 3 lines · Christoph Hellwig

You'll need to merge all patches into one, otherwise you create really nasty memory leaks when bisecting between them.

Okay. I will resend.

Thanks for the notice, Christoph.

Good point from hch, but I feel even more strongly: if you're going to do this now, please rename remove_from_page_cache (delete_from_page_cache was what I chose back when I misdid it) - you're changing an EXPORTed function in a subtle (well, subtlish) confusing way, which could easily waste people's time down the line, whether in not-yet-in-tree filesystems or backports of fixes. I'd much rather you break someone's build, forcing them to look at what changed, than crash or leak at runtime.

If you do rename, you can keep your patch structure, introducing the new function as a wrapper to the old at the beginning, then removing the old function at the end.

(As you know, I do agree that it's right to decrement the reference count at the point of removing from page cache.)

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#27 lore raw
Quoted 18 lines · Hugh Dickins
Quoted 8 lines · Minchan Kim
Quoted 3 lines · Christoph Hellwig

You'll need to merge all patches into one, otherwise you create really nasty memory leaks when bisecting between them.

Okay. I will resend.

Thanks for the notice, Christoph.

Good point from hch, but I feel even more strongly: if you're going to do this now, please rename remove_from_page_cache (delete_from_page_cache was what I chose back when I misdid it) - you're changing an EXPORTed function in a subtle (well, subtlish) confusing way, which could easily waste people's time down the line, whether in not-yet-in-tree filesystems or backports of fixes.  I'd much rather you break someone's build, forcing them to look at what changed, than crash or leak at runtime.

Quoted 3 lines

If you do rename, you can keep your patch structure, introducing the new function as a wrapper to the old at the beginning, then removing the old function at the end.

It is very good idea!!
Thanks for good suggestion, Hugh.

Quoted 6 lines

(As you know, I do agree that it's right to decrement the reference count at the point of removing from page cache.)

Hugh

Signature · 9 lines
Kind regards,
Minchan Kim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#7 lore raw

Re: [RFC 3/5] tlbfs: Remove unnecessary page release

In reply to Minchan Kim
Quoted 6 lines · Minchan Kim

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

Other than the subject calling hugetlbfs tlbfs, I did not see any problem with this assuming the first patch of the series is also applied.

Signature · 3 lines
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
#8 lore raw

Re: [RFC 3/5] tlbfs: Remove unnecessary page release

Quoted 10 lines · Mel Gorman
Quoted 6 lines · Minchan Kim

This patch series changes remove_from_page_cache's page ref counting rule. page cache ref count is decreased in remove_from_page_cache. So we don't need call again in caller context.

Other than the subject calling hugetlbfs tlbfs, I did not see any problem with this assuming the first patch of the series is also applied.

Thanks, Mel.
Will fix.

Signature · 9 lines
Kind regards,
Minchan Kim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>