* [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path
@ 2020-06-04 7:05 Yafang Shao
2020-06-09 13:53 ` Yafang Shao
2020-06-09 14:03 ` Christoph Hellwig
0 siblings, 2 replies; 6+ messages in thread
From: Yafang Shao @ 2020-06-04 7:05 UTC (permalink / raw)
To: darrick.wong, david, hch; +Cc: linux-xfs, linux-fsdevel, Yafang Shao
Recently there is a XFS deadlock on our server with an old kernel.
This deadlock is caused by allocating memory in xfs_map_blocks() while
doing writeback on behalf of memroy reclaim. Although this deadlock happens
on an old kernel, I think it could happen on the upstream as well. This
issue only happens once and can't be reproduced, so I haven't tried to
reproduce it on upsteam kernel.
Bellow is the call trace of this deadlock.
[480594.790087] INFO: task redis-server:16212 blocked for more than 120 seconds.
[480594.790087] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[480594.790088] redis-server D ffffffff8168bd60 0 16212 14347 0x00000004
[480594.790090] ffff880da128f070 0000000000000082 ffff880f94a2eeb0 ffff880da128ffd8
[480594.790092] ffff880da128ffd8 ffff880da128ffd8 ffff880f94a2eeb0 ffff88103f9d6c40
[480594.790094] 0000000000000000 7fffffffffffffff ffff88207ffc0ee8 ffffffff8168bd60
[480594.790096] Call Trace:
[480594.790101] [<ffffffff8168dce9>] schedule+0x29/0x70
[480594.790103] [<ffffffff8168b749>] schedule_timeout+0x239/0x2c0
[480594.790111] [<ffffffff8168d28e>] io_schedule_timeout+0xae/0x130
[480594.790114] [<ffffffff8168d328>] io_schedule+0x18/0x20
[480594.790116] [<ffffffff8168bd71>] bit_wait_io+0x11/0x50
[480594.790118] [<ffffffff8168b895>] __wait_on_bit+0x65/0x90
[480594.790121] [<ffffffff811814e1>] wait_on_page_bit+0x81/0xa0
[480594.790125] [<ffffffff81196ad2>] shrink_page_list+0x6d2/0xaf0
[480594.790130] [<ffffffff811975a3>] shrink_inactive_list+0x223/0x710
[480594.790135] [<ffffffff81198225>] shrink_lruvec+0x3b5/0x810
[480594.790139] [<ffffffff8119873a>] shrink_zone+0xba/0x1e0
[480594.790141] [<ffffffff81198c20>] do_try_to_free_pages+0x100/0x510
[480594.790143] [<ffffffff8119928d>] try_to_free_mem_cgroup_pages+0xdd/0x170
[480594.790145] [<ffffffff811f32de>] mem_cgroup_reclaim+0x4e/0x120
[480594.790147] [<ffffffff811f37cc>] __mem_cgroup_try_charge+0x41c/0x670
[480594.790153] [<ffffffff811f5cb6>] __memcg_kmem_newpage_charge+0xf6/0x180
[480594.790157] [<ffffffff8118c72d>] __alloc_pages_nodemask+0x22d/0x420
[480594.790162] [<ffffffff811d0c7a>] alloc_pages_current+0xaa/0x170
[480594.790165] [<ffffffff811db8fc>] new_slab+0x30c/0x320
[480594.790168] [<ffffffff811dd17c>] ___slab_alloc+0x3ac/0x4f0
[480594.790204] [<ffffffff81685656>] __slab_alloc+0x40/0x5c
[480594.790206] [<ffffffff811dfc43>] kmem_cache_alloc+0x193/0x1e0
[480594.790233] [<ffffffffa04fab67>] kmem_zone_alloc+0x97/0x130 [xfs]
[480594.790247] [<ffffffffa04f90ba>] _xfs_trans_alloc+0x3a/0xa0 [xfs]
[480594.790261] [<ffffffffa04f915c>] xfs_trans_alloc+0x3c/0x50 [xfs]
[480594.790276] [<ffffffffa04e958b>] xfs_iomap_write_allocate+0x1cb/0x390 [xfs]
[480594.790299] [<ffffffffa04d3616>] xfs_map_blocks+0x1a6/0x210 [xfs]
[480594.790312] [<ffffffffa04d416b>] xfs_do_writepage+0x17b/0x550 [xfs]
[480594.790314] [<ffffffff8118d881>] write_cache_pages+0x251/0x4d0 [xfs]
[480594.790338] [<ffffffffa04d3e05>] xfs_vm_writepages+0xc5/0xe0 [xfs]
[480594.790341] [<ffffffff8118ebfe>] do_writepages+0x1e/0x40
[480594.790343] [<ffffffff811837b5>] __filemap_fdatawrite_range+0x65/0x80
[480594.790346] [<ffffffff81183901>] filemap_write_and_wait_range+0x41/0x90
[480594.790360] [<ffffffffa04df2c6>] xfs_file_fsync+0x66/0x1e0 [xfs]
[480594.790363] [<ffffffff81231cf5>] do_fsync+0x65/0xa0
[480594.790365] [<ffffffff81231fe3>] SyS_fdatasync+0x13/0x20
[480594.790367] [<ffffffff81698d09>] system_call_fastpath+0x16/0x1b
Note that xfs_iomap_write_allocate() is replaced by xfs_convert_blocks() in
commit 4ad765edb02a ("xfs: move xfs_iomap_write_allocate to xfs_aops.c")
and write_cache_pages() is replaced by iomap_writepages() in
commit 598ecfbaa742 ("iomap: lift the xfs writeback code to iomap").
So for upsteam, the call trace should be,
xfs_vm_writepages
-> iomap_writepages
-> write_cache_pages
-> iomap_do_writepage
-> xfs_map_blocks
-> xfs_convert_blocks
-> xfs_bmapi_convert_delalloc
-> xfs_trans_alloc //It should alloc page with GFP_NOFS
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
---
v1 - >v2:
- retile the subject from "xfs: avoid deadlock when tigger memory reclam in xfs_map_blocks()"
- set GFP_NOFS in iomap_do_writepage(), per Dave.
---
fs/iomap/buffered-io.c | 14 +++++++++++++-
1 file changed, 13 insertions(+), 1 deletion(-)
diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index a1ed762..f5176e3 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -17,6 +17,7 @@
#include <linux/bio.h>
#include <linux/sched/signal.h>
#include <linux/migrate.h>
+#include <linux/sched/mm.h>
#include "trace.h"
#include "../internal.h"
@@ -1478,9 +1479,11 @@ static void iomap_writepage_end_bio(struct bio *bio)
{
struct iomap_writepage_ctx *wpc = data;
struct inode *inode = page->mapping->host;
+ unsigned int nofs_flag;
pgoff_t end_index;
u64 end_offset;
loff_t offset;
+ int ret;
trace_iomap_writepage(inode, page_offset(page), PAGE_SIZE);
@@ -1571,7 +1574,16 @@ static void iomap_writepage_end_bio(struct bio *bio)
end_offset = offset;
}
- return iomap_writepage_map(wpc, wbc, inode, page, end_offset);
+ /*
+ * We can allocate memory here while doing writeback on behalf of
+ * memory reclaim. To avoid memory allocation deadlocks set the
+ * task-wide nofs context for the following operations.
+ */
+ nofs_flag = memalloc_nofs_save();
+ ret = iomap_writepage_map(wpc, wbc, inode, page, end_offset);
+ memalloc_nofs_restore(nofs_flag);
+
+ return ret;
redirty:
redirty_page_for_writepage(wbc, page);
--
1.8.3.1
^ permalink raw reply related [flat|nested] 6+ messages in thread
* Re: [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path
2020-06-04 7:05 [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path Yafang Shao
@ 2020-06-09 13:53 ` Yafang Shao
2020-06-09 14:03 ` Christoph Hellwig
1 sibling, 0 replies; 6+ messages in thread
From: Yafang Shao @ 2020-06-09 13:53 UTC (permalink / raw)
To: Darrick J. Wong, Dave Chinner, hch; +Cc: linux-xfs, linux-fsdevel
On Thu, Jun 4, 2020 at 3:06 PM Yafang Shao <laoar.shao@gmail.com> wrote:
>
> Recently there is a XFS deadlock on our server with an old kernel.
> This deadlock is caused by allocating memory in xfs_map_blocks() while
> doing writeback on behalf of memroy reclaim. Although this deadlock happens
> on an old kernel, I think it could happen on the upstream as well. This
> issue only happens once and can't be reproduced, so I haven't tried to
> reproduce it on upsteam kernel.
>
> Bellow is the call trace of this deadlock.
> [480594.790087] INFO: task redis-server:16212 blocked for more than 120 seconds.
> [480594.790087] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> [480594.790088] redis-server D ffffffff8168bd60 0 16212 14347 0x00000004
> [480594.790090] ffff880da128f070 0000000000000082 ffff880f94a2eeb0 ffff880da128ffd8
> [480594.790092] ffff880da128ffd8 ffff880da128ffd8 ffff880f94a2eeb0 ffff88103f9d6c40
> [480594.790094] 0000000000000000 7fffffffffffffff ffff88207ffc0ee8 ffffffff8168bd60
> [480594.790096] Call Trace:
> [480594.790101] [<ffffffff8168dce9>] schedule+0x29/0x70
> [480594.790103] [<ffffffff8168b749>] schedule_timeout+0x239/0x2c0
> [480594.790111] [<ffffffff8168d28e>] io_schedule_timeout+0xae/0x130
> [480594.790114] [<ffffffff8168d328>] io_schedule+0x18/0x20
> [480594.790116] [<ffffffff8168bd71>] bit_wait_io+0x11/0x50
> [480594.790118] [<ffffffff8168b895>] __wait_on_bit+0x65/0x90
> [480594.790121] [<ffffffff811814e1>] wait_on_page_bit+0x81/0xa0
> [480594.790125] [<ffffffff81196ad2>] shrink_page_list+0x6d2/0xaf0
> [480594.790130] [<ffffffff811975a3>] shrink_inactive_list+0x223/0x710
> [480594.790135] [<ffffffff81198225>] shrink_lruvec+0x3b5/0x810
> [480594.790139] [<ffffffff8119873a>] shrink_zone+0xba/0x1e0
> [480594.790141] [<ffffffff81198c20>] do_try_to_free_pages+0x100/0x510
> [480594.790143] [<ffffffff8119928d>] try_to_free_mem_cgroup_pages+0xdd/0x170
> [480594.790145] [<ffffffff811f32de>] mem_cgroup_reclaim+0x4e/0x120
> [480594.790147] [<ffffffff811f37cc>] __mem_cgroup_try_charge+0x41c/0x670
> [480594.790153] [<ffffffff811f5cb6>] __memcg_kmem_newpage_charge+0xf6/0x180
> [480594.790157] [<ffffffff8118c72d>] __alloc_pages_nodemask+0x22d/0x420
> [480594.790162] [<ffffffff811d0c7a>] alloc_pages_current+0xaa/0x170
> [480594.790165] [<ffffffff811db8fc>] new_slab+0x30c/0x320
> [480594.790168] [<ffffffff811dd17c>] ___slab_alloc+0x3ac/0x4f0
> [480594.790204] [<ffffffff81685656>] __slab_alloc+0x40/0x5c
> [480594.790206] [<ffffffff811dfc43>] kmem_cache_alloc+0x193/0x1e0
> [480594.790233] [<ffffffffa04fab67>] kmem_zone_alloc+0x97/0x130 [xfs]
> [480594.790247] [<ffffffffa04f90ba>] _xfs_trans_alloc+0x3a/0xa0 [xfs]
> [480594.790261] [<ffffffffa04f915c>] xfs_trans_alloc+0x3c/0x50 [xfs]
> [480594.790276] [<ffffffffa04e958b>] xfs_iomap_write_allocate+0x1cb/0x390 [xfs]
> [480594.790299] [<ffffffffa04d3616>] xfs_map_blocks+0x1a6/0x210 [xfs]
> [480594.790312] [<ffffffffa04d416b>] xfs_do_writepage+0x17b/0x550 [xfs]
> [480594.790314] [<ffffffff8118d881>] write_cache_pages+0x251/0x4d0 [xfs]
> [480594.790338] [<ffffffffa04d3e05>] xfs_vm_writepages+0xc5/0xe0 [xfs]
> [480594.790341] [<ffffffff8118ebfe>] do_writepages+0x1e/0x40
> [480594.790343] [<ffffffff811837b5>] __filemap_fdatawrite_range+0x65/0x80
> [480594.790346] [<ffffffff81183901>] filemap_write_and_wait_range+0x41/0x90
> [480594.790360] [<ffffffffa04df2c6>] xfs_file_fsync+0x66/0x1e0 [xfs]
> [480594.790363] [<ffffffff81231cf5>] do_fsync+0x65/0xa0
> [480594.790365] [<ffffffff81231fe3>] SyS_fdatasync+0x13/0x20
> [480594.790367] [<ffffffff81698d09>] system_call_fastpath+0x16/0x1b
>
> Note that xfs_iomap_write_allocate() is replaced by xfs_convert_blocks() in
> commit 4ad765edb02a ("xfs: move xfs_iomap_write_allocate to xfs_aops.c")
> and write_cache_pages() is replaced by iomap_writepages() in
> commit 598ecfbaa742 ("iomap: lift the xfs writeback code to iomap").
> So for upsteam, the call trace should be,
> xfs_vm_writepages
> -> iomap_writepages
> -> write_cache_pages
> -> iomap_do_writepage
> -> xfs_map_blocks
> -> xfs_convert_blocks
> -> xfs_bmapi_convert_delalloc
> -> xfs_trans_alloc //It should alloc page with GFP_NOFS
>
> Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
>
> ---
> v1 - >v2:
> - retile the subject from "xfs: avoid deadlock when tigger memory reclam in xfs_map_blocks()"
> - set GFP_NOFS in iomap_do_writepage(), per Dave.
> ---
> fs/iomap/buffered-io.c | 14 +++++++++++++-
> 1 file changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
> index a1ed762..f5176e3 100644
> --- a/fs/iomap/buffered-io.c
> +++ b/fs/iomap/buffered-io.c
> @@ -17,6 +17,7 @@
> #include <linux/bio.h>
> #include <linux/sched/signal.h>
> #include <linux/migrate.h>
> +#include <linux/sched/mm.h>
> #include "trace.h"
>
> #include "../internal.h"
> @@ -1478,9 +1479,11 @@ static void iomap_writepage_end_bio(struct bio *bio)
> {
> struct iomap_writepage_ctx *wpc = data;
> struct inode *inode = page->mapping->host;
> + unsigned int nofs_flag;
> pgoff_t end_index;
> u64 end_offset;
> loff_t offset;
> + int ret;
>
> trace_iomap_writepage(inode, page_offset(page), PAGE_SIZE);
>
> @@ -1571,7 +1574,16 @@ static void iomap_writepage_end_bio(struct bio *bio)
> end_offset = offset;
> }
>
> - return iomap_writepage_map(wpc, wbc, inode, page, end_offset);
> + /*
> + * We can allocate memory here while doing writeback on behalf of
> + * memory reclaim. To avoid memory allocation deadlocks set the
> + * task-wide nofs context for the following operations.
> + */
> + nofs_flag = memalloc_nofs_save();
> + ret = iomap_writepage_map(wpc, wbc, inode, page, end_offset);
> + memalloc_nofs_restore(nofs_flag);
> +
> + return ret;
>
> redirty:
> redirty_page_for_writepage(wbc, page);
> --
> 1.8.3.1
>
Dave, Darrick,
Any comments on this version ?
Thanks
Yafang
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path
2020-06-04 7:05 [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path Yafang Shao
2020-06-09 13:53 ` Yafang Shao
@ 2020-06-09 14:03 ` Christoph Hellwig
2020-06-09 14:28 ` Yafang Shao
1 sibling, 1 reply; 6+ messages in thread
From: Christoph Hellwig @ 2020-06-09 14:03 UTC (permalink / raw)
To: Yafang Shao; +Cc: darrick.wong, david, hch, linux-xfs, linux-fsdevel
On Thu, Jun 04, 2020 at 03:05:47AM -0400, Yafang Shao wrote:
> Recently there is a XFS deadlock on our server with an old kernel.
> This deadlock is caused by allocating memory in xfs_map_blocks() while
> doing writeback on behalf of memroy reclaim. Although this deadlock happens
> on an old kernel, I think it could happen on the upstream as well. This
> issue only happens once and can't be reproduced, so I haven't tried to
> reproduce it on upsteam kernel.
The report looks sensible, but I don't think the iomap code is the
right place for this. Until/unless the VM people agree that
->writepages(s) generally should not recurse into the fs I think the
low-level file system allocating is the right place, so xfs_map_blocks
would seem like the correct place.
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path
2020-06-09 14:03 ` Christoph Hellwig
@ 2020-06-09 14:28 ` Yafang Shao
2020-06-09 14:32 ` Christoph Hellwig
0 siblings, 1 reply; 6+ messages in thread
From: Yafang Shao @ 2020-06-09 14:28 UTC (permalink / raw)
To: Christoph Hellwig; +Cc: Darrick J. Wong, Dave Chinner, linux-xfs, linux-fsdevel
On Tue, Jun 9, 2020 at 10:03 PM Christoph Hellwig <hch@infradead.org> wrote:
>
> On Thu, Jun 04, 2020 at 03:05:47AM -0400, Yafang Shao wrote:
> > Recently there is a XFS deadlock on our server with an old kernel.
> > This deadlock is caused by allocating memory in xfs_map_blocks() while
> > doing writeback on behalf of memroy reclaim. Although this deadlock happens
> > on an old kernel, I think it could happen on the upstream as well. This
> > issue only happens once and can't be reproduced, so I haven't tried to
> > reproduce it on upsteam kernel.
>
> The report looks sensible, but I don't think the iomap code is the
> right place for this. Until/unless the VM people agree that
> ->writepages(s) generally should not recurse into the fs I think the
> low-level file system allocating is the right place, so xfs_map_blocks
> would seem like the correct place.
Thanks for your comment.
That is what I did in the previous version [1].
So should I resend the v1 ?
[1]. https://lore.kernel.org/linux-xfs/1591179035-9270-1-git-send-email-laoar.shao@gmail.com/
--
Thanks
Yafang
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path
2020-06-09 14:28 ` Yafang Shao
@ 2020-06-09 14:32 ` Christoph Hellwig
2020-06-09 15:36 ` Yafang Shao
0 siblings, 1 reply; 6+ messages in thread
From: Christoph Hellwig @ 2020-06-09 14:32 UTC (permalink / raw)
To: Yafang Shao
Cc: Christoph Hellwig, Darrick J. Wong, Dave Chinner, linux-xfs,
linux-fsdevel
On Tue, Jun 09, 2020 at 10:28:06PM +0800, Yafang Shao wrote:
> On Tue, Jun 9, 2020 at 10:03 PM Christoph Hellwig <hch@infradead.org> wrote:
> >
> > On Thu, Jun 04, 2020 at 03:05:47AM -0400, Yafang Shao wrote:
> > > Recently there is a XFS deadlock on our server with an old kernel.
> > > This deadlock is caused by allocating memory in xfs_map_blocks() while
> > > doing writeback on behalf of memroy reclaim. Although this deadlock happens
> > > on an old kernel, I think it could happen on the upstream as well. This
> > > issue only happens once and can't be reproduced, so I haven't tried to
> > > reproduce it on upsteam kernel.
> >
> > The report looks sensible, but I don't think the iomap code is the
> > right place for this. Until/unless the VM people agree that
> > ->writepages(s) generally should not recurse into the fs I think the
> > low-level file system allocating is the right place, so xfs_map_blocks
> > would seem like the correct place.
>
> Thanks for your comment.
> That is what I did in the previous version [1].
> So should I resend the v1 ?
Well, v1 won't apply. But I do prefer the approach there.
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path
2020-06-09 14:32 ` Christoph Hellwig
@ 2020-06-09 15:36 ` Yafang Shao
0 siblings, 0 replies; 6+ messages in thread
From: Yafang Shao @ 2020-06-09 15:36 UTC (permalink / raw)
To: Christoph Hellwig, Michal Hocko, Andrew Morton
Cc: Darrick J. Wong, Dave Chinner, linux-xfs, linux-fsdevel, Linux MM
On Tue, Jun 9, 2020 at 10:32 PM Christoph Hellwig <hch@infradead.org> wrote:
>
> On Tue, Jun 09, 2020 at 10:28:06PM +0800, Yafang Shao wrote:
> > On Tue, Jun 9, 2020 at 10:03 PM Christoph Hellwig <hch@infradead.org> wrote:
> > >
> > > On Thu, Jun 04, 2020 at 03:05:47AM -0400, Yafang Shao wrote:
> > > > Recently there is a XFS deadlock on our server with an old kernel.
> > > > This deadlock is caused by allocating memory in xfs_map_blocks() while
> > > > doing writeback on behalf of memroy reclaim. Although this deadlock happens
> > > > on an old kernel, I think it could happen on the upstream as well. This
> > > > issue only happens once and can't be reproduced, so I haven't tried to
> > > > reproduce it on upsteam kernel.
> > >
> > > The report looks sensible, but I don't think the iomap code is the
> > > right place for this. Until/unless the VM people agree that
> > > ->writepages(s) generally should not recurse into the fs I think the
> > > low-level file system allocating is the right place, so xfs_map_blocks
> > > would seem like the correct place.
> >
> > Thanks for your comment.
> > That is what I did in the previous version [1].
> > So should I resend the v1 ?
>
> Well, v1 won't apply. But I do prefer the approach there.
All right. Let's include MM maintainers and see the opinion from them.
Hi Michal, Andrew,
What's your opinion on this XFS deadlock ?
Should ->writepages(s) generally not recurse into the fs ?
--
Thanks
Yafang
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2020-06-09 15:37 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-06-04 7:05 [PATCH v2] iomap: avoid deadlock if memory reclaim is triggered in writepage path Yafang Shao
2020-06-09 13:53 ` Yafang Shao
2020-06-09 14:03 ` Christoph Hellwig
2020-06-09 14:28 ` Yafang Shao
2020-06-09 14:32 ` Christoph Hellwig
2020-06-09 15:36 ` Yafang Shao
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).