| Message ID | 1424914057.41161.8.camel@primarydata.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Feb 25, 2015, at 8:27 PM, Trond Myklebust <trond.myklebust@primarydata.com> wrote: > On Wed, 2015-02-25 at 19:43 -0500, Trond Myklebust wrote: >> On Wed, 2015-02-25 at 19:37 -0500, Trond Myklebust wrote: >>> On Wed, 2015-02-25 at 17:32 -0500, Chuck Lever wrote: >>>> FWIW it’s easy to reproduce a similar race with fsx, and I encounter >>>> it frequently while running xfstests on fast NFS servers. >>>> >>>> fsx invokes ftruncate following a set of asynchronous reads >>>> (generated possibly due to readahead). The reads are started first, >>>> then the SETATTR, but they complete out of order. >>>> >>>> The SETATTR changes the test file’s size, and the completion >>>> updates the file size in the client’s inode. Then the read requests >>>> complete on the client and set the file’s size back to its old value. >>>> >>>> All it takes is one late read completion, and the cached file size >>>> is corrupted. fsx detects the file size mismatch and terminates the >>>> test. The file size is corrected by a subsequent GETATTR (say, an >>>> “ls -l” to check it after fsx has terminated). >>>> >>>> While SETATTR blocks concurrent writes, there’s no serialization >>>> on either the client or server to help guarantee the ordering of >>>> SETATTR with read operations. >>>> >>>> I’ve found a successful workaround by forcing the client to ignore >>>> post-op attrs in read replies. A stronger solution might simply set >>>> the “file attributes need update” flag in the inode if any file >>>> attribute mutation is noticed during a read completion. >>> >>> That's different. We definitely should aim to fix this kind of issue >>> since you are talking about a single client being the only thing >>> accessing the file on the server. >>> How about the following patch? >> >> Let me retry without the typos. The following will actually >> compile... :-/ > > Third version: this contains a minor optimisation so that we don't force > the re-read of the last page in the file. Also a little more detail in > the changelog. Thanks, I’ll give this a shot, hopefully today. > 8<-------------------------------------------------------------------- > From d2c822d64a68542665e4887b4d7bccd6e8e3a741 Mon Sep 17 00:00:00 2001 > From: Trond Myklebust <trond.myklebust@primarydata.com> > Date: Wed, 25 Feb 2015 19:26:28 -0500 > Subject: [PATCH] NFS: Quiesce reads before updating the file size after > truncating > > Chuck Lever reports seeing readaheads racing with truncate operations > and causing the file size to be reverted. Fix is to quiesce reads > after truncating the file on the server, but before updating the > size on the client. > > Note: > > 1) We believe that waiting for those reads that lie outside the new > file boundaries is sufficient, since this means any attempt to read > the data in those areas will trigger a new RPC call, which will be > ordered w.r.t. the size change on the server. > > 2) We do not move the call to truncate_pagecache(), since we want > to ensure that any partial page zeroing is ordered w.r.t. the > update of inode->i_size. > > Reported-by: Chuck Lever <chuck.lever@oracle.com> > Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com> > --- > fs/nfs/inode.c | 5 +++++ > 1 file changed, 5 insertions(+) > > diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c > index 83107be3dd01..daf3b1e183fe 100644 > --- a/fs/nfs/inode.c > +++ b/fs/nfs/inode.c > @@ -565,6 +565,11 @@ static int nfs_vmtruncate(struct inode * inode, loff_t offset) > if (err) > goto out; > > + /* Quiesce reads that lie outside the new file size */ > + invalidate_inode_pages2_range(inode->i_mapping, > + (offset + PAGE_CACHE_SHIFT - 1) >> PAGE_CACHE_SHIFT, > + -1); > + > spin_lock(&inode->i_lock); > i_size_write(inode, offset); > /* Optimisation */ > -- > 2.1.0 > > > -- Chuck Lever chuck[dot]lever[at]oracle[dot]com -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
I’m breaking this into a separate thread, since you correctly pointed out that this is not the same problem that Chris sees. Begin forwarded message: > From: Trond Myklebust <trond.myklebust@primarydata.com> > Subject: Re: File Read Returns Non-existent Null Bytes > Date: February 25, 2015 at 8:27:37 PM EST > To: Chuck Lever <chuck.lever@oracle.com> > Cc: Chris Perl <cperl@janestreet.com>, Linux NFS Mailing List <linux-nfs@vger.kernel.org>, Chris Perl <chris.perl@gmail.com> > > On Wed, 2015-02-25 at 19:43 -0500, Trond Myklebust wrote: >> On Wed, 2015-02-25 at 19:37 -0500, Trond Myklebust wrote: >>> On Wed, 2015-02-25 at 17:32 -0500, Chuck Lever wrote: >>>> FWIW it’s easy to reproduce a similar race with fsx, and I encounter >>>> it frequently while running xfstests on fast NFS servers. >>>> >>>> fsx invokes ftruncate following a set of asynchronous reads >>>> (generated possibly due to readahead). The reads are started first, >>>> then the SETATTR, but they complete out of order. >>>> >>>> The SETATTR changes the test file’s size, and the completion >>>> updates the file size in the client’s inode. Then the read requests >>>> complete on the client and set the file’s size back to its old value. >>>> >>>> All it takes is one late read completion, and the cached file size >>>> is corrupted. fsx detects the file size mismatch and terminates the >>>> test. The file size is corrected by a subsequent GETATTR (say, an >>>> “ls -l” to check it after fsx has terminated). >>>> >>>> While SETATTR blocks concurrent writes, there’s no serialization >>>> on either the client or server to help guarantee the ordering of >>>> SETATTR with read operations. >>>> >>>> I’ve found a successful workaround by forcing the client to ignore >>>> post-op attrs in read replies. A stronger solution might simply set >>>> the “file attributes need update” flag in the inode if any file >>>> attribute mutation is noticed during a read completion. >>> >>> That's different. We definitely should aim to fix this kind of issue >>> since you are talking about a single client being the only thing >>> accessing the file on the server. >>> How about the following patch? >> >> Let me retry without the typos. The following will actually >> compile... :-/ > > Third version: this contains a minor optimisation so that we don't force > the re-read of the last page in the file. Also a little more detail in > the changelog. Test results: generic/127 74s ... - output mismatch (see /home/cel/src/xfstests/results//generic/127.out.bad) --- tests/generic/127.out 2014-02-13 15:40:45.202103271 -0500 +++ /home/cel/src/xfstests/results//generic/127.out.bad 2015-02-26 10:46:24.284735977 -0500 @@ -4,7 +4,10011 @@ === FSX Light Mode, Memory Mapping === All operations completed A-OK! === FSX Standard Mode, No Memory Mapping === -All operations completed A-OK! +ltp/fsx -q -l 262144 -o 65536 -S 191110531 -N 100000 -R -W fsx_std_nommap +Size error: expected 0x36f6f stat 0x3e67a seek 0x3e67a +LOG DUMP (83542 total operations): ... (Run 'diff -u tests/generic/127.out /home/cel/src/xfstests/results//generic/127.out.bad' to see the entire diff) This happens when the server exports a tmpfs, for example, which mimics the performance characteristics of an NVM-based filesystem. Otherwise, SETATTR will be slow and the race is typically avoided. Well, that’s my theory, anyway. > 8<-------------------------------------------------------------------- > From d2c822d64a68542665e4887b4d7bccd6e8e3a741 Mon Sep 17 00:00:00 2001 > From: Trond Myklebust <trond.myklebust@primarydata.com> > Date: Wed, 25 Feb 2015 19:26:28 -0500 > Subject: [PATCH] NFS: Quiesce reads before updating the file size after > truncating > > Chuck Lever reports seeing readaheads racing with truncate operations > and causing the file size to be reverted. Fix is to quiesce reads > after truncating the file on the server, but before updating the > size on the client. > > Note: > > 1) We believe that waiting for those reads that lie outside the new > file boundaries is sufficient, since this means any attempt to read > the data in those areas will trigger a new RPC call, which will be > ordered w.r.t. the size change on the server. Certainly reads outside the new file size have _data_ that can be discarded. But in this case: The server processes the operations in the order the client sent them. The READs first, then the SETATTR. That’s why the READ replies still have the old file size. But queuing on the server causes the replies to be returned to the client out of order. The client applies the post-op attributes in that order, and the old file size replaces the new size. Therefore I think _any_ READ that is performed on the server before the SETATTR, but completes on the client after the SETATTR, will poison the file’s attribute cache on the client. > 2) We do not move the call to truncate_pagecache(), since we want > to ensure that any partial page zeroing is ordered w.r.t. the > update of inode->i_size. > > Reported-by: Chuck Lever <chuck.lever@oracle.com> > Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com> > --- > fs/nfs/inode.c | 5 +++++ > 1 file changed, 5 insertions(+) > > diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c > index 83107be3dd01..daf3b1e183fe 100644 > --- a/fs/nfs/inode.c > +++ b/fs/nfs/inode.c > @@ -565,6 +565,11 @@ static int nfs_vmtruncate(struct inode * inode, loff_t offset) > if (err) > goto out; > > + /* Quiesce reads that lie outside the new file size */ > + invalidate_inode_pages2_range(inode->i_mapping, > + (offset + PAGE_CACHE_SHIFT - 1) >> PAGE_CACHE_SHIFT, > + -1); > + > spin_lock(&inode->i_lock); > i_size_write(inode, offset); > /* Optimisation */ > -- > 2.1.0 > > > -- Chuck Lever chuck[dot]lever[at]oracle[dot]com -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Feb 26, 2015 at 11:26 AM, Chuck Lever <chuck.lever@oracle.com> wrote: > I’m breaking this into a separate thread, since you correctly > pointed out that this is not the same problem that Chris sees. > > Begin forwarded message: > >> From: Trond Myklebust <trond.myklebust@primarydata.com> >> Subject: Re: File Read Returns Non-existent Null Bytes >> Date: February 25, 2015 at 8:27:37 PM EST >> To: Chuck Lever <chuck.lever@oracle.com> >> Cc: Chris Perl <cperl@janestreet.com>, Linux NFS Mailing List <linux-nfs@vger.kernel.org>, Chris Perl <chris.perl@gmail.com> >> >> On Wed, 2015-02-25 at 19:43 -0500, Trond Myklebust wrote: >>> On Wed, 2015-02-25 at 19:37 -0500, Trond Myklebust wrote: >>>> On Wed, 2015-02-25 at 17:32 -0500, Chuck Lever wrote: >>>>> FWIW it’s easy to reproduce a similar race with fsx, and I encounter >>>>> it frequently while running xfstests on fast NFS servers. >>>>> >>>>> fsx invokes ftruncate following a set of asynchronous reads >>>>> (generated possibly due to readahead). The reads are started first, >>>>> then the SETATTR, but they complete out of order. >>>>> >>>>> The SETATTR changes the test file’s size, and the completion >>>>> updates the file size in the client’s inode. Then the read requests >>>>> complete on the client and set the file’s size back to its old value. >>>>> >>>>> All it takes is one late read completion, and the cached file size >>>>> is corrupted. fsx detects the file size mismatch and terminates the >>>>> test. The file size is corrected by a subsequent GETATTR (say, an >>>>> “ls -l” to check it after fsx has terminated). >>>>> >>>>> While SETATTR blocks concurrent writes, there’s no serialization >>>>> on either the client or server to help guarantee the ordering of >>>>> SETATTR with read operations. >>>>> >>>>> I’ve found a successful workaround by forcing the client to ignore >>>>> post-op attrs in read replies. A stronger solution might simply set >>>>> the “file attributes need update” flag in the inode if any file >>>>> attribute mutation is noticed during a read completion. >>>> >>>> That's different. We definitely should aim to fix this kind of issue >>>> since you are talking about a single client being the only thing >>>> accessing the file on the server. >>>> How about the following patch? >>> >>> Let me retry without the typos. The following will actually >>> compile... :-/ >> >> Third version: this contains a minor optimisation so that we don't force >> the re-read of the last page in the file. Also a little more detail in >> the changelog. > > Test results: > > generic/127 74s ... - output mismatch (see /home/cel/src/xfstests/results//generic/127.out.bad) > --- tests/generic/127.out 2014-02-13 15:40:45.202103271 -0500 > +++ /home/cel/src/xfstests/results//generic/127.out.bad 2015-02-26 10:46:24.284735977 -0500 > @@ -4,7 +4,10011 @@ > === FSX Light Mode, Memory Mapping === > All operations completed A-OK! > === FSX Standard Mode, No Memory Mapping === > -All operations completed A-OK! > +ltp/fsx -q -l 262144 -o 65536 -S 191110531 -N 100000 -R -W fsx_std_nommap > +Size error: expected 0x36f6f stat 0x3e67a seek 0x3e67a > +LOG DUMP (83542 total operations): > ... > (Run 'diff -u tests/generic/127.out /home/cel/src/xfstests/results//generic/127.out.bad' to see the entire diff) > > This happens when the server exports a tmpfs, for example, which > mimics the performance characteristics of an NVM-based filesystem. > Otherwise, SETATTR will be slow and the race is typically avoided. > Well, that’s my theory, anyway. > > >> 8<-------------------------------------------------------------------- >> From d2c822d64a68542665e4887b4d7bccd6e8e3a741 Mon Sep 17 00:00:00 2001 >> From: Trond Myklebust <trond.myklebust@primarydata.com> >> Date: Wed, 25 Feb 2015 19:26:28 -0500 >> Subject: [PATCH] NFS: Quiesce reads before updating the file size after >> truncating >> >> Chuck Lever reports seeing readaheads racing with truncate operations >> and causing the file size to be reverted. Fix is to quiesce reads >> after truncating the file on the server, but before updating the >> size on the client. >> >> Note: >> >> 1) We believe that waiting for those reads that lie outside the new >> file boundaries is sufficient, since this means any attempt to read >> the data in those areas will trigger a new RPC call, which will be >> ordered w.r.t. the size change on the server. > > Certainly reads outside the new file size have _data_ that can > be discarded. > > But in this case: The server processes the operations in the order > the client sent them. The READs first, then the SETATTR. That’s why > the READ replies still have the old file size. > > But queuing on the server causes the replies to be returned to the > client out of order. The client applies the post-op attributes in > that order, and the old file size replaces the new size. > > Therefore I think _any_ READ that is performed on the server before > the SETATTR, but completes on the client after the SETATTR, will > poison the file’s attribute cache on the client. > Why doesn't our usual trick of using the ctime to discriminate between older and newer file sizes work here? Is the problem that the resolution is insufficient? The alternative would be to add a "barrier" operation to allow the setattr code to wait for all RPC calls before ours to complete (it would have to wait not just for READ, but also GETATTR and all other operations to the file). We could do that too.
On Feb 26, 2015, at 12:27 PM, Trond Myklebust <trond.myklebust@primarydata.com> wrote: > On Thu, Feb 26, 2015 at 11:26 AM, Chuck Lever <chuck.lever@oracle.com> wrote: >> I’m breaking this into a separate thread, since you correctly >> pointed out that this is not the same problem that Chris sees. >> >> Begin forwarded message: >> >>> From: Trond Myklebust <trond.myklebust@primarydata.com> >>> Subject: Re: File Read Returns Non-existent Null Bytes >>> Date: February 25, 2015 at 8:27:37 PM EST >>> To: Chuck Lever <chuck.lever@oracle.com> >>> Cc: Chris Perl <cperl@janestreet.com>, Linux NFS Mailing List <linux-nfs@vger.kernel.org>, Chris Perl <chris.perl@gmail.com> >>> >>> On Wed, 2015-02-25 at 19:43 -0500, Trond Myklebust wrote: >>>> On Wed, 2015-02-25 at 19:37 -0500, Trond Myklebust wrote: >>>>> On Wed, 2015-02-25 at 17:32 -0500, Chuck Lever wrote: >>>>>> FWIW it’s easy to reproduce a similar race with fsx, and I encounter >>>>>> it frequently while running xfstests on fast NFS servers. >>>>>> >>>>>> fsx invokes ftruncate following a set of asynchronous reads >>>>>> (generated possibly due to readahead). The reads are started first, >>>>>> then the SETATTR, but they complete out of order. >>>>>> >>>>>> The SETATTR changes the test file’s size, and the completion >>>>>> updates the file size in the client’s inode. Then the read requests >>>>>> complete on the client and set the file’s size back to its old value. >>>>>> >>>>>> All it takes is one late read completion, and the cached file size >>>>>> is corrupted. fsx detects the file size mismatch and terminates the >>>>>> test. The file size is corrected by a subsequent GETATTR (say, an >>>>>> “ls -l” to check it after fsx has terminated). >>>>>> >>>>>> While SETATTR blocks concurrent writes, there’s no serialization >>>>>> on either the client or server to help guarantee the ordering of >>>>>> SETATTR with read operations. >>>>>> >>>>>> I’ve found a successful workaround by forcing the client to ignore >>>>>> post-op attrs in read replies. A stronger solution might simply set >>>>>> the “file attributes need update” flag in the inode if any file >>>>>> attribute mutation is noticed during a read completion. >>>>> >>>>> That's different. We definitely should aim to fix this kind of issue >>>>> since you are talking about a single client being the only thing >>>>> accessing the file on the server. >>>>> How about the following patch? >>>> >>>> Let me retry without the typos. The following will actually >>>> compile... :-/ >>> >>> Third version: this contains a minor optimisation so that we don't force >>> the re-read of the last page in the file. Also a little more detail in >>> the changelog. >> >> Test results: >> >> generic/127 74s ... - output mismatch (see /home/cel/src/xfstests/results//generic/127.out.bad) >> --- tests/generic/127.out 2014-02-13 15:40:45.202103271 -0500 >> +++ /home/cel/src/xfstests/results//generic/127.out.bad 2015-02-26 10:46:24.284735977 -0500 >> @@ -4,7 +4,10011 @@ >> === FSX Light Mode, Memory Mapping === >> All operations completed A-OK! >> === FSX Standard Mode, No Memory Mapping === >> -All operations completed A-OK! >> +ltp/fsx -q -l 262144 -o 65536 -S 191110531 -N 100000 -R -W fsx_std_nommap >> +Size error: expected 0x36f6f stat 0x3e67a seek 0x3e67a >> +LOG DUMP (83542 total operations): >> ... >> (Run 'diff -u tests/generic/127.out /home/cel/src/xfstests/results//generic/127.out.bad' to see the entire diff) >> >> This happens when the server exports a tmpfs, for example, which >> mimics the performance characteristics of an NVM-based filesystem. >> Otherwise, SETATTR will be slow and the race is typically avoided. >> Well, that’s my theory, anyway. >> >> >>> 8<-------------------------------------------------------------------- >>> From d2c822d64a68542665e4887b4d7bccd6e8e3a741 Mon Sep 17 00:00:00 2001 >>> From: Trond Myklebust <trond.myklebust@primarydata.com> >>> Date: Wed, 25 Feb 2015 19:26:28 -0500 >>> Subject: [PATCH] NFS: Quiesce reads before updating the file size after >>> truncating >>> >>> Chuck Lever reports seeing readaheads racing with truncate operations >>> and causing the file size to be reverted. Fix is to quiesce reads >>> after truncating the file on the server, but before updating the >>> size on the client. >>> >>> Note: >>> >>> 1) We believe that waiting for those reads that lie outside the new >>> file boundaries is sufficient, since this means any attempt to read >>> the data in those areas will trigger a new RPC call, which will be >>> ordered w.r.t. the size change on the server. >> >> Certainly reads outside the new file size have _data_ that can >> be discarded. >> >> But in this case: The server processes the operations in the order >> the client sent them. The READs first, then the SETATTR. That’s why >> the READ replies still have the old file size. >> >> But queuing on the server causes the replies to be returned to the >> client out of order. The client applies the post-op attributes in >> that order, and the old file size replaces the new size. >> >> Therefore I think _any_ READ that is performed on the server before >> the SETATTR, but completes on the client after the SETATTR, will >> poison the file’s attribute cache on the client. >> > > Why doesn't our usual trick of using the ctime to discriminate between > older and newer file sizes work here? Is the problem that the > resolution is insufficient? I’m not sure. These operations can occur during the same millisecond, so it is plausible there’s a timestamp resolution issue. Or the server doesn’t return pre-op attrs. Timestamp is more likely, though: I have seen this issue occur on NFSv4 as well. > The alternative would be to add a "barrier" operation to allow the > setattr code to wait for all RPC calls before ours to complete (it > would have to wait not just for READ, but also GETATTR and all other > operations to the file). We could do that too. Yes. A barrier is usually heavyweight, though. And I thought there was some exclusion between GETATTR and SETATTR, so forcing an inode revalidation should work too? Or would that result in an unneeded data cache invalidation? -- Chuck Lever chuck[dot]lever[at]oracle[dot]com -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Feb 26, 2015 at 2:00 PM, Chuck Lever <chuck.lever@oracle.com> wrote: >> The alternative would be to add a "barrier" operation to allow the >> setattr code to wait for all RPC calls before ours to complete (it >> would have to wait not just for READ, but also GETATTR and all other >> operations to the file). We could do that too. > > Yes. A barrier is usually heavyweight, though. And I thought there > was some exclusion between GETATTR and SETATTR, so forcing an > inode revalidation should work too? Or would that result in an > unneeded data cache invalidation? > I found a better solution, I think. Please see the 5 patches I just sent out.
diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c index 83107be3dd01..daf3b1e183fe 100644 --- a/fs/nfs/inode.c +++ b/fs/nfs/inode.c @@ -565,6 +565,11 @@ static int nfs_vmtruncate(struct inode * inode, loff_t offset) if (err) goto out; + /* Quiesce reads that lie outside the new file size */ + invalidate_inode_pages2_range(inode->i_mapping, + (offset + PAGE_CACHE_SHIFT - 1) >> PAGE_CACHE_SHIFT, + -1); + spin_lock(&inode->i_lock); i_size_write(inode, offset); /* Optimisation */