Hi Steve,
Here's another fix patch for you. It fixes a leak of the pages extracted
for direct I/O. This was causing the page refcount to overflow and become
negative and then try_grab_page() would always fail on it.
I've pushed the patch here also:
https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/log/?h=iov-cifs
David
David Howells (1):
cifs: Fix memory leak in direct I/O
fs/cifs/file.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
When __cifs_readv() and __cifs_writev() extract pages from a user-backed
iterator into a BVEC-type iterator, they set ->bv_need_unpin to note
whether they need to unpin the pages later. However, in both cases they
examine the BVEC-type iterator and not the source iterator - and so
bv_need_unpin doesn't get set and the pages are leaked.
I think this may be responsible for the generic/208 xfstest failing
occasionally with:
WARNING: CPU: 0 PID: 3064 at mm/gup.c:218 try_grab_page+0x65/0x100
RIP: 0010:try_grab_page+0x65/0x100
follow_page_pte+0x1a7/0x570
__get_user_pages+0x1a2/0x650
__gup_longterm_locked+0xdc/0xb50
internal_get_user_pages_fast+0x17f/0x310
pin_user_pages_fast+0x46/0x60
iov_iter_extract_pages+0xc9/0x510
? __kmalloc_large_node+0xb1/0x120
? __kmalloc_node+0xbe/0x130
netfs_extract_user_iter+0xbf/0x200 [netfs]
__cifs_writev+0x150/0x330 [cifs]
vfs_write+0x2a8/0x3c0
ksys_pwrite64+0x65/0xa0
with the page refcount going negative. This is less unlikely than it seems
because the page is being pinned, not simply got, and so the refcount
increased by 1024 each time, and so only needs to be called around ~2097152
for the refcount to go negative.
Further, the test program (aio-dio-invalidate-failure) uses a 32MiB static
buffer and all the PTEs covering it refer to the same page because it's
never written to.
The warning in try_grab_page():
if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
return -ENOMEM;
then trips and prevents us ever using the page again for DIO at least.
Fixes: d08089f649a0 ("cifs: Change the I/O paths to use an iterator rather than a page list")
Reported-by: Murphy Zhou <[email protected]>
Link: https://lore.kernel.org/r/CAH2r5mvaTsJ---n=265a4zqRA7pP+o4MJ36WCQUS6oPrOij8cw@mail.gmail.com
Signed-off-by: David Howells <[email protected]>
cc: Steve French <[email protected]>
cc: Shyam Prasad N <[email protected]>
cc: Rohith Surabattula <[email protected]>
cc: Paulo Alcantara <[email protected]>
cc: Jeff Layton <[email protected]>
cc: [email protected]
---
fs/cifs/file.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/fs/cifs/file.c b/fs/cifs/file.c
index ec0694a65c7b..4d4a2d82636d 100644
--- a/fs/cifs/file.c
+++ b/fs/cifs/file.c
@@ -3612,7 +3612,7 @@ static ssize_t __cifs_writev(
ctx->nr_pinned_pages = rc;
ctx->bv = (void *)ctx->iter.bvec;
- ctx->bv_need_unpin = iov_iter_extract_will_pin(&ctx->iter);
+ ctx->bv_need_unpin = iov_iter_extract_will_pin(from);
} else if ((iov_iter_is_bvec(from) || iov_iter_is_kvec(from)) &&
!is_sync_kiocb(iocb)) {
/*
@@ -4148,7 +4148,7 @@ static ssize_t __cifs_readv(
ctx->nr_pinned_pages = rc;
ctx->bv = (void *)ctx->iter.bvec;
- ctx->bv_need_unpin = iov_iter_extract_will_pin(&ctx->iter);
+ ctx->bv_need_unpin = iov_iter_extract_will_pin(to);
ctx->should_dirty = true;
} else if ((iov_iter_is_bvec(to) || iov_iter_is_kvec(to)) &&
!is_sync_kiocb(iocb)) {
merged into cifs-2.6.git for-next pending more testing
On Tue, Feb 28, 2023 at 4:38 PM David Howells <[email protected]> wrote:
>
> When __cifs_readv() and __cifs_writev() extract pages from a user-backed
> iterator into a BVEC-type iterator, they set ->bv_need_unpin to note
> whether they need to unpin the pages later. However, in both cases they
> examine the BVEC-type iterator and not the source iterator - and so
> bv_need_unpin doesn't get set and the pages are leaked.
>
> I think this may be responsible for the generic/208 xfstest failing
> occasionally with:
>
> WARNING: CPU: 0 PID: 3064 at mm/gup.c:218 try_grab_page+0x65/0x100
> RIP: 0010:try_grab_page+0x65/0x100
> follow_page_pte+0x1a7/0x570
> __get_user_pages+0x1a2/0x650
> __gup_longterm_locked+0xdc/0xb50
> internal_get_user_pages_fast+0x17f/0x310
> pin_user_pages_fast+0x46/0x60
> iov_iter_extract_pages+0xc9/0x510
> ? __kmalloc_large_node+0xb1/0x120
> ? __kmalloc_node+0xbe/0x130
> netfs_extract_user_iter+0xbf/0x200 [netfs]
> __cifs_writev+0x150/0x330 [cifs]
> vfs_write+0x2a8/0x3c0
> ksys_pwrite64+0x65/0xa0
>
> with the page refcount going negative. This is less unlikely than it seems
> because the page is being pinned, not simply got, and so the refcount
> increased by 1024 each time, and so only needs to be called around ~2097152
> for the refcount to go negative.
>
> Further, the test program (aio-dio-invalidate-failure) uses a 32MiB static
> buffer and all the PTEs covering it refer to the same page because it's
> never written to.
>
> The warning in try_grab_page():
>
> if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
> return -ENOMEM;
>
> then trips and prevents us ever using the page again for DIO at least.
>
> Fixes: d08089f649a0 ("cifs: Change the I/O paths to use an iterator rather than a page list")
> Reported-by: Murphy Zhou <[email protected]>
> Link: https://lore.kernel.org/r/CAH2r5mvaTsJ---n=265a4zqRA7pP+o4MJ36WCQUS6oPrOij8cw@mail.gmail.com
> Signed-off-by: David Howells <[email protected]>
> cc: Steve French <[email protected]>
> cc: Shyam Prasad N <[email protected]>
> cc: Rohith Surabattula <[email protected]>
> cc: Paulo Alcantara <[email protected]>
> cc: Jeff Layton <[email protected]>
> cc: [email protected]
> ---
> fs/cifs/file.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/fs/cifs/file.c b/fs/cifs/file.c
> index ec0694a65c7b..4d4a2d82636d 100644
> --- a/fs/cifs/file.c
> +++ b/fs/cifs/file.c
> @@ -3612,7 +3612,7 @@ static ssize_t __cifs_writev(
>
> ctx->nr_pinned_pages = rc;
> ctx->bv = (void *)ctx->iter.bvec;
> - ctx->bv_need_unpin = iov_iter_extract_will_pin(&ctx->iter);
> + ctx->bv_need_unpin = iov_iter_extract_will_pin(from);
> } else if ((iov_iter_is_bvec(from) || iov_iter_is_kvec(from)) &&
> !is_sync_kiocb(iocb)) {
> /*
> @@ -4148,7 +4148,7 @@ static ssize_t __cifs_readv(
>
> ctx->nr_pinned_pages = rc;
> ctx->bv = (void *)ctx->iter.bvec;
> - ctx->bv_need_unpin = iov_iter_extract_will_pin(&ctx->iter);
> + ctx->bv_need_unpin = iov_iter_extract_will_pin(to);
> ctx->should_dirty = true;
> } else if ((iov_iter_is_bvec(to) || iov_iter_is_kvec(to)) &&
> !is_sync_kiocb(iocb)) {
>
--
Thanks,
Steve
David Howells <[email protected]> writes:
> When __cifs_readv() and __cifs_writev() extract pages from a user-backed
> iterator into a BVEC-type iterator, they set ->bv_need_unpin to note
> whether they need to unpin the pages later. However, in both cases they
> examine the BVEC-type iterator and not the source iterator - and so
> bv_need_unpin doesn't get set and the pages are leaked.
>
> I think this may be responsible for the generic/208 xfstest failing
> occasionally with:
>
> WARNING: CPU: 0 PID: 3064 at mm/gup.c:218 try_grab_page+0x65/0x100
> RIP: 0010:try_grab_page+0x65/0x100
> follow_page_pte+0x1a7/0x570
> __get_user_pages+0x1a2/0x650
> __gup_longterm_locked+0xdc/0xb50
> internal_get_user_pages_fast+0x17f/0x310
> pin_user_pages_fast+0x46/0x60
> iov_iter_extract_pages+0xc9/0x510
> ? __kmalloc_large_node+0xb1/0x120
> ? __kmalloc_node+0xbe/0x130
> netfs_extract_user_iter+0xbf/0x200 [netfs]
> __cifs_writev+0x150/0x330 [cifs]
> vfs_write+0x2a8/0x3c0
> ksys_pwrite64+0x65/0xa0
>
> with the page refcount going negative. This is less unlikely than it seems
> because the page is being pinned, not simply got, and so the refcount
> increased by 1024 each time, and so only needs to be called around ~2097152
> for the refcount to go negative.
>
> Further, the test program (aio-dio-invalidate-failure) uses a 32MiB static
> buffer and all the PTEs covering it refer to the same page because it's
> never written to.
>
> The warning in try_grab_page():
>
> if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
> return -ENOMEM;
>
> then trips and prevents us ever using the page again for DIO at least.
>
> Fixes: d08089f649a0 ("cifs: Change the I/O paths to use an iterator rather than a page list")
> Reported-by: Murphy Zhou <[email protected]>
> Link: https://lore.kernel.org/r/CAH2r5mvaTsJ---n=265a4zqRA7pP+o4MJ36WCQUS6oPrOij8cw@mail.gmail.com
> Signed-off-by: David Howells <[email protected]>
> cc: Steve French <[email protected]>
> cc: Shyam Prasad N <[email protected]>
> cc: Rohith Surabattula <[email protected]>
> cc: Paulo Alcantara <[email protected]>
> cc: Jeff Layton <[email protected]>
> cc: [email protected]
> ---
> fs/cifs/file.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
Reviewed-by: Paulo Alcantara (SUSE) <[email protected]>
I also verified that this fixes the problem that Murphy pointed out - thx
On Wed, Mar 1, 2023 at 9:11 AM Paulo Alcantara <[email protected]> wrote:
>
> David Howells <[email protected]> writes:
>
> > When __cifs_readv() and __cifs_writev() extract pages from a user-backed
> > iterator into a BVEC-type iterator, they set ->bv_need_unpin to note
> > whether they need to unpin the pages later. However, in both cases they
> > examine the BVEC-type iterator and not the source iterator - and so
> > bv_need_unpin doesn't get set and the pages are leaked.
> >
> > I think this may be responsible for the generic/208 xfstest failing
> > occasionally with:
> >
> > WARNING: CPU: 0 PID: 3064 at mm/gup.c:218 try_grab_page+0x65/0x100
> > RIP: 0010:try_grab_page+0x65/0x100
> > follow_page_pte+0x1a7/0x570
> > __get_user_pages+0x1a2/0x650
> > __gup_longterm_locked+0xdc/0xb50
> > internal_get_user_pages_fast+0x17f/0x310
> > pin_user_pages_fast+0x46/0x60
> > iov_iter_extract_pages+0xc9/0x510
> > ? __kmalloc_large_node+0xb1/0x120
> > ? __kmalloc_node+0xbe/0x130
> > netfs_extract_user_iter+0xbf/0x200 [netfs]
> > __cifs_writev+0x150/0x330 [cifs]
> > vfs_write+0x2a8/0x3c0
> > ksys_pwrite64+0x65/0xa0
> >
> > with the page refcount going negative. This is less unlikely than it seems
> > because the page is being pinned, not simply got, and so the refcount
> > increased by 1024 each time, and so only needs to be called around ~2097152
> > for the refcount to go negative.
> >
> > Further, the test program (aio-dio-invalidate-failure) uses a 32MiB static
> > buffer and all the PTEs covering it refer to the same page because it's
> > never written to.
> >
> > The warning in try_grab_page():
> >
> > if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
> > return -ENOMEM;
> >
> > then trips and prevents us ever using the page again for DIO at least.
> >
> > Fixes: d08089f649a0 ("cifs: Change the I/O paths to use an iterator rather than a page list")
> > Reported-by: Murphy Zhou <[email protected]>
> > Link: https://lore.kernel.org/r/CAH2r5mvaTsJ---n=265a4zqRA7pP+o4MJ36WCQUS6oPrOij8cw@mail.gmail.com
> > Signed-off-by: David Howells <[email protected]>
> > cc: Steve French <[email protected]>
> > cc: Shyam Prasad N <[email protected]>
> > cc: Rohith Surabattula <[email protected]>
> > cc: Paulo Alcantara <[email protected]>
> > cc: Jeff Layton <[email protected]>
> > cc: [email protected]
> > ---
> > fs/cifs/file.c | 4 ++--
> > 1 file changed, 2 insertions(+), 2 deletions(-)
>
> Reviewed-by: Paulo Alcantara (SUSE) <[email protected]>
--
Thanks,
Steve