On Tue, Sep 12 2017, Jeff Layton wrote: > On Tue, 2017-09-12 at 07:52 +1000, NeilBrown wrote: >> > On Mon, 2017-09-11 at 13:24 +1000, NeilBrown wrote: >> > > On Thu, Sep 07 2017, Trond Myklebust wrote: >> > > >> > > > On Thu, 2017-09-07 at 07:35 -0400, Jeff Layton wrote: >> > > > > On Thu, 2017-09-07 at 13:37 +1000, NeilBrown wrote: >> > > > > > On Tue, Aug 29 2017, Jeff Layton wrote: >> > > > > > >> > > > > > > On Tue, 2017-08-29 at 11:23 +1000, NeilBrown wrote: >> > > > > > > > On Mon, Aug 28 2017, Jeff Layton wrote: >> > > > > > > > >> > > > > > > > > On Mon, 2017-08-28 at 09:24 +1000, NeilBrown wrote: >> > > > > > > > > > On Fri, Aug 25 2017, Jeff Layton wrote: >> > > > > > > > > > >> > > > > > > > > > > On Thu, 2017-07-20 at 15:42 -0400, Jeff Layton wrote: >> > > > > > > > > > > > From: Jeff Layton <jlayton@xxxxxxxxxx> >> > > > > > > > > > > > >> > > > > > > > > > > > There is some ambiguity in nfs about how writeback >> > > > > > > > > > > > errors are >> > > > > > > > > > > > tracked. >> > > > > > > > > > > > >> > > > > > > > > > > > For instance, nfs_pageio_add_request calls >> > > > > > > > > > > > mapping_set_error when >> > > > > > > > > > > > the >> > > > > > > > > > > > add fails, but we track errors that occur after adding >> > > > > > > > > > > > the >> > > > > > > > > > > > request >> > > > > > > > > > > > with a dedicated int error in the open context. >> > > > > > > > > > > > >> > > > > > > > > > > > Now that we have better infrastructure for the vfs >> > > > > > > > > > > > layer, this >> > > > > > > > > > > > latter int is now unnecessary. Just have >> > > > > > > > > > > > nfs_context_set_write_error set >> > > > > > > > > > > > the error in the mapping when one occurs. >> > > > > > > > > > > > >> > > > > > > > > > > > Have NFS use file_write_and_wait_range to initiate and >> > > > > > > > > > > > wait on >> > > > > > > > > > > > writeback >> > > > > > > > > > > > of the data, and then check again after issuing the >> > > > > > > > > > > > commit(s). >> > > > > > > > > > > > >> > > > > > > > > > > > With this, we also don't need to pay attention to the >> > > > > > > > > > > > ERROR_WRITE >> > > > > > > > > > > > flag for reporting, and just clear it to indicate to >> > > > > > > > > > > > subsequent >> > > > > > > > > > > > writers that they should try to go asynchronous again. >> > > > > > > > > > > > >> > > > > > > > > > > > In nfs_page_async_flush, sample the error before >> > > > > > > > > > > > locking and >> > > > > > > > > > > > joining >> > > > > > > > > > > > the requests, and check for errors since that point. >> > > > > > > > > > > > >> > > > > > > > > > > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx> >> > > > > > > > > > > > --- >> > > > > > > > > > > > fs/nfs/file.c | 24 +++++++++++------------- >> > > > > > > > > > > > fs/nfs/inode.c | 3 +-- >> > > > > > > > > > > > fs/nfs/write.c | 8 ++++++-- >> > > > > > > > > > > > include/linux/nfs_fs.h | 1 - >> > > > > > > > > > > > 4 files changed, 18 insertions(+), 18 deletions(-) >> > > > > > > > > > > > >> > > > > > > > > > > > I have a baling wire and duct tape solution for testing >> > > > > > > > > > > > this with >> > > > > > > > > > > > xfstests (using iptables REJECT targets and soft >> > > > > > > > > > > > mounts). This >> > > > > > > > > > > > seems to >> > > > > > > > > > > > make nfs do the right thing. >> > > > > > > > > > > > >> > > > > > > > > > > > diff --git a/fs/nfs/file.c b/fs/nfs/file.c >> > > > > > > > > > > > index 5713eb32a45e..15d3c6faafd3 100644 >> > > > > > > > > > > > --- a/fs/nfs/file.c >> > > > > > > > > > > > +++ b/fs/nfs/file.c >> > > > > > > > > > > > @@ -212,25 +212,23 @@ nfs_file_fsync_commit(struct file >> > > > > > > > > > > > *file, >> > > > > > > > > > > > loff_t start, loff_t end, int datasync) >> > > > > > > > > > > > { >> > > > > > > > > > > > struct nfs_open_context *ctx = >> > > > > > > > > > > > nfs_file_open_context(file); >> > > > > > > > > > > > struct inode *inode = file_inode(file); >> > > > > > > > > > > > - int have_error, do_resend, status; >> > > > > > > > > > > > - int ret = 0; >> > > > > > > > > > > > + int do_resend, status; >> > > > > > > > > > > > + int ret; >> > > > > > > > > > > > >> > > > > > > > > > > > dprintk("NFS: fsync file(%pD2) datasync %d\n", >> > > > > > > > > > > > file, >> > > > > > > > > > > > datasync); >> > > > > > > > > > > > >> > > > > > > > > > > > nfs_inc_stats(inode, NFSIOS_VFSFSYNC); >> > > > > > > > > > > > do_resend = >> > > > > > > > > > > > test_and_clear_bit(NFS_CONTEXT_RESEND_WRITES, &ctx- >> > > > > > > > > > > > > flags); >> > > > > > > > > > > > >> > > > > > > > > > > > - have_error = >> > > > > > > > > > > > test_and_clear_bit(NFS_CONTEXT_ERROR_WRITE, >> > > > > > > > > > > > &ctx->flags); >> > > > > > > > > > > > - status = nfs_commit_inode(inode, FLUSH_SYNC); >> > > > > > > > > > > > - have_error |= >> > > > > > > > > > > > test_bit(NFS_CONTEXT_ERROR_WRITE, &ctx- >> > > > > > > > > > > > > flags); >> > > > > > > > > > > > >> > > > > > > > > > > > - if (have_error) { >> > > > > > > > > > > > - ret = xchg(&ctx->error, 0); >> > > > > > > > > > > > - if (ret) >> > > > > > > > > > > > - goto out; >> > > > > > > > > > > > - } >> > > > > > > > > > > > - if (status < 0) { >> > > > > > > > > > > > + clear_bit(NFS_CONTEXT_ERROR_WRITE, &ctx- >> > > > > > > > > > > > > flags); >> > > > > > > > > > > > >> > > > > > > > > > > > + ret = nfs_commit_inode(inode, FLUSH_SYNC); >> > > > > > > > > > > > + >> > > > > > > > > > > > + /* Recheck and advance after the commit */ >> > > > > > > > > > > > + status = file_check_and_advance_wb_err(file); >> > > > > > > > > > >> > > > > > > > > > This change makes the code inconsistent with the comment >> > > > > > > > > > above the >> > > > > > > > > > function, which still references ctx->error. The intent of >> > > > > > > > > > the >> > > > > > > > > > comment >> > > > > > > > > > is still correct, but the details have changed. >> > > > > > > > > > >> > > > > > > > > >> > > > > > > > > Good catch. I'll fix that up in a respin. >> > > > > > > > > >> > > > > > > > > > Also, there is a call to mapping_set_error() in >> > > > > > > > > > nfs_pageio_add_request(). >> > > > > > > > > > I wonder if that should be changed to >> > > > > > > > > > nfs_context_set_write_error(req->wb_context, desc- >> > > > > > > > > > > pg_error) >> > > > > > > > > > >> > > > > > > > > > ?? >> > > > > > > > > > >> > > > > > > > > >> > > > > > > > > Trickier question... >> > > > > > > > > >> > > > > > > > > I'm not quite sure what semantics we're looking for with >> > > > > > > > > NFS_CONTEXT_ERROR_WRITE. I know that it forces writes to be >> > > > > > > > > synchronous, but I'm not quite sure why it gets cleared the >> > > > > > > > > way it >> > > > > > > > > does. It's set on any error but cleared before issuing a >> > > > > > > > > commit. >> > > > > > > > > >> > > > > > > > > I added a similar flag to Ceph inodes recently, but only >> > > > > > > > > clear it when >> > > > > > > > > a write succeeds. Wouldn't that make more sense here as well? >> > > > > > > > >> > > > > > > > It is a bit hard to wrap one's mind around. >> > > > > > > > >> > > > > > > > In the original code (commit 7b159fc18d417980) it looks like: >> > > > > > > > - test-and-clear bit >> > > > > > > > - write and sync >> > > > > > > > - test-bit >> > > > > > > > >> > > > > > > > This does, I think, seem safer than "clear on successful write" >> > > > > > > > as the >> > > > > > > > writes could complete out-of-order and I wouldn't be surprised >> > > > > > > > if the >> > > > > > > > unsuccessful ones completed with an error before the successful >> > > > > > > > one - >> > > > > > > > particularly with an error like EDQUOT. >> > > > > > > > >> > > > > > > > However the current code does the writes before the test-and- >> > > > > > > > clear, and >> > > > > > > > only does the commit afterwards. That makes it less clear why >> > > > > > > > the >> > > > > > > > current sequence is a good idea. >> > > > > > > > >> > > > > > > > However ... nfs_file_fsync_commit() is only called if >> > > > > > > > filemap_write_and_wait_range() returned with success, so we >> > > > > > > > only clear >> > > > > > > > the flag after successful writes(?). >> > > > > > > > >> > > > > > > > Oh.... >> > > > > > > > This patch from me: >> > > > > > > > >> > > > > > > > Commit: 2edb6bc3852c ("NFS - fix recent breakage to NFS error >> > > > > > > > handling.") >> > > > > > > > >> > > > > > > > seems to have been reverted by >> > > > > > > > >> > > > > > > > Commit: 7b281ee02655 ("NFS: fsync() must exit with an error if >> > > > > > > > page writeback failed") >> > > > > > > > >> > > > > > > > which probably isn't good. It appears that this code is very >> > > > > > > > fragile >> > > > > > > > and easily broken. >> > > > > > >> > > > > > On further investigation, I think the problem that I fixed and then >> > > > > > we >> > > > > > reintroduced will be fixed again - more permanently - by your >> > > > > > patch. >> > > > > > The root problem is that nfs keeps error codes in a different way >> > > > > > to the >> > > > > > MM core. By unifying those, the problem goes. >> > > > > > (The specific problem is that writes which hit EDQUOT on the server >> > > > > > can >> > > > > > report EIO on the client). >> > > > > > >> > > > > > >> > > > > > > > Maybe we need to work out exactly what is required, and >> > > > > > > > document it - so >> > > > > > > > we can stop breaking it. >> > > > > > > > Or maybe we need some unit tests..... >> > > > > > > > >> > > > > > > >> > > > > > > Yes, laying out what's necessary for this would be very helpful. >> > > > > > > We >> > > > > > > clearly want to set the flag when an error occurs. Under what >> > > > > > > circumstances should we be clearing it? >> > > > > > >> > > > > > Well.... looking back at 7b159fc18d417980f57ae which introduced >> > > > > > the >> > > > > > flag, prior to that write errors (ctx->error) were only reported by >> > > > > > nfs_file_flush and nfs_fsync, so only one close() and fsync(). >> > > > > > >> > > > > > After that commit, setting the flag would mean that errors could be >> > > > > > returned by 'write'. So clearing as part of returning the error >> > > > > > makes >> > > > > > perfect sense. >> > > > > > >> > > > > > As long as the error gets recorded, and gets returned when it is >> > > > > > recorded, it doesn't much matter when the flag is cleared. With >> > > > > > your >> > > > > > patches we don't need to flag any more to get errors reliably >> > > > > > reported. >> > > > > > >> > > > > > Leaving the flag set means that writes go more slowly - we don't >> > > > > > get >> > > > > > large queue of background rights building up but destined for >> > > > > > failure. >> > > > > > This is the main point made in the comment message when the flag >> > > > > > was >> > > > > > introduced. >> > > > > > Of course, by the time we first get an error there could already >> > > > > > by a large queue, so we probably want that to drain completely >> > > > > > before >> > > > > > allowing async writes again. >> > > > >> > > > We already have this functionality implemented in the existing code. >> > > > >> > > > > > >> > > > > > It might make sense to have 2 flags. One which says "writes should >> > > > > > be >> > > > > > synchronous", another that says "There was an error recently". >> > > > > > We clear the error flag before calling nfs_fsync, and if it is >> > > > > > still >> > > > > > clear afterwards, we clear the sync-writes flag. Maybe that is >> > > > > > more >> > > > > > complex than needed though. >> > > > > > >> > > > >> > > > We also need to preserve the NFS_CONTEXT_RESEND_WRITES flag. I don't >> > > > see any global mechanism that will replace that. >> > > > >> > > > > > I'm leaning towards your suggestion that it doesn't matter very >> > > > > > much >> > > > > > when it gets cleared, and clearing it on any successful write is >> > > > > > simplest. >> > > > > > >> > > > > > So I'm still in favor of using nfs_context_set_write_error() in >> > > > > > nfs_pageio_add_request(), primarily because it is most consistent - >> > > > > > we >> > > > > > don't need exceptions. >> > > > > >> > > > > Thanks for taking a closer look. I can easily make the change above, >> > > > > and >> > > > > I do think that keeping this mechanism as simple as possible will >> > > > > make >> > > > > it easier to prevent bitrot. >> > > > > >> > > > > That said... NFS_CONTEXT_ERROR_WRITE is a per ctx flag, and the ctx >> > > > > is a >> > > > > per open file description object. >> > > > > >> > > > > Is that the correct way to track this? All of the ctx's will share >> > > > > the >> > > > > same inode. If we're getting writeback errors for one context, it's >> > > > > quite likely that we'll be seeing them via others. >> > > > > >> > > > > I suppose the counterargument is when we have things like expiring >> > > > > krb5 >> > > > > tickets. Write failures via an expiring set of creds may have no >> > > > > effect >> > > > > on writeback via other creds. >> > > > > >> > > > > Still, I think a per-inode flag might make more sense here. >> > > > > >> > > > > Thoughts? >> > > > >> > > > As far as I'm concerned, that would be a regression. The most common >> > > > problem when flushing writeback data to the server aside from ENOSPC >> > > > (and possibly ESTALE) is EACCES, which is particular to the file >> > > > descriptor that opened the file. >> > > > >> > > > File contexts, and NFS_CONTEXT_ERROR_WRITE solve that problem by being >> > > > private to the file descriptor. >> > > >> > > Thanks for the reminder that errors are per-context and this patch drops >> > > this. The per-context nature of errors in NFS was the reason that I >> > > nagged Jeff to make errseq_t a stand-alone type rather than just a part >> > > of address_space. I had envisaged that it would be embedded in the >> > > open_context as well. >> > > We still could do that, but as there is precisely one open-file for each >> > > open_context, the gains are not great. >> > > >> > > However, while looking over the code to make sure I really understood it >> > > and all the possible consequences of changing to errseq_t I found a few >> > > anomalies. The patch below addresses them all. >> > > >> > > Would you see if they may sense to you? >> > > >> > > Thanks, >> > > NeilBrown >> > > >> > > >> > > From: NeilBrown <neilb@xxxxxxxx> >> > > Date: Mon, 11 Sep 2017 13:15:50 +1000 >> > > Subject: [PATCH] NFS: various changes relating to reporting IO errors. >> > > >> > > 1/ remove 'start' and 'end' args from nfs_file_fsync_commit(). >> > > They aren't used. >> > > >> > > 2/ Make nfs_context_set_write_error() a "static inline" in internal.h >> > > so we can... >> > > >> > > 3/ Use nfs_context_set_write_error() instead of mapping_set_error() >> > > if nfs_pageio_add_request() fails before sending any request. >> > > NFS generally keeps errors in the open_context, not the mapping, >> > > so this is more consistent. >> > > >> > > 4/ If filemap_write_and_write_range() reports any error, still >> > > check ctx->error. The value in ctx->error is likely to be >> > > more useful. As part of this, NFS_CONTEXT_ERROR_WRITE is >> > > cleared slightly earlier, before nfs_file_fsync_commit() is called, >> > > rather than at the start of that function. >> > > >> > > Signed-off-by: NeilBrown <neilb@xxxxxxxx> >> > > --- >> > > fs/nfs/file.c | 16 ++++++++++------ >> > > fs/nfs/internal.h | 7 +++++++ >> > > fs/nfs/pagelist.c | 4 ++-- >> > > fs/nfs/write.c | 7 ------- >> > > 4 files changed, 19 insertions(+), 15 deletions(-) >> > > >> > > diff --git a/fs/nfs/file.c b/fs/nfs/file.c >> > > index af330c31f627..ab324f14081f 100644 >> > > --- a/fs/nfs/file.c >> > > +++ b/fs/nfs/file.c >> > > @@ -208,21 +208,19 @@ EXPORT_SYMBOL_GPL(nfs_file_mmap); >> > > * fall back to doing a synchronous write. >> > > */ >> > > static int >> > > -nfs_file_fsync_commit(struct file *file, loff_t start, loff_t end, int datasync) >> > > +nfs_file_fsync_commit(struct file *file, int datasync) >> > > { >> > > struct nfs_open_context *ctx = nfs_file_open_context(file); >> > > struct inode *inode = file_inode(file); >> > > - int have_error, do_resend, status; >> > > + int do_resend, status; >> > > int ret = 0; >> > > >> > > dprintk("NFS: fsync file(%pD2) datasync %d\n", file, datasync); >> > > >> > > nfs_inc_stats(inode, NFSIOS_VFSFSYNC); >> > > do_resend = test_and_clear_bit(NFS_CONTEXT_RESEND_WRITES, &ctx->flags); >> > > - have_error = test_and_clear_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags); >> > > status = nfs_commit_inode(inode, FLUSH_SYNC); >> > > - have_error |= test_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags); >> > > - if (have_error) { >> > > + if (test_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags)) { >> > > ret = xchg(&ctx->error, 0); >> > > if (ret) >> > > goto out; >> > > @@ -247,10 +245,16 @@ nfs_file_fsync(struct file *file, loff_t start, loff_t end, int datasync) >> > > trace_nfs_fsync_enter(inode); >> > > >> > > do { >> > > + struct nfs_open_context *ctx = nfs_file_open_context(file); >> > > ret = filemap_write_and_wait_range(inode->i_mapping, start, end); >> > > + if (test_and_clear_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags)) { >> > > + int ret2 = xchg(&ctx->error, 0); >> > > + if (ret2) >> > > + ret = ret2; >> > > + } >> > > if (ret != 0) >> > > break; >> > > - ret = nfs_file_fsync_commit(file, start, end, datasync); >> > > + ret = nfs_file_fsync_commit(file, datasync); >> > > if (!ret) >> > > ret = pnfs_sync_inode(inode, !!datasync); >> > > /* >> > > diff --git a/fs/nfs/internal.h b/fs/nfs/internal.h >> > > index dc456416d2be..44c8962fec91 100644 >> > > --- a/fs/nfs/internal.h >> > > +++ b/fs/nfs/internal.h >> > > @@ -769,3 +769,10 @@ static inline bool nfs_error_is_fatal(int err) >> > > return false; >> > > } >> > > } >> > > + >> > > +static inline void nfs_context_set_write_error(struct nfs_open_context *ctx, int error) >> > > +{ >> > > + ctx->error = error; >> > > + smp_wmb(); >> > > + set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags); >> > > +} >> > > diff --git a/fs/nfs/pagelist.c b/fs/nfs/pagelist.c >> > > index de9066a92c0d..0ebd26b9a6bd 100644 >> > > --- a/fs/nfs/pagelist.c >> > > +++ b/fs/nfs/pagelist.c >> > > @@ -1198,8 +1198,8 @@ out_failed: >> > > >> > > /* remember fatal errors */ >> > > if (nfs_error_is_fatal(desc->pg_error)) >> > > - mapping_set_error(desc->pg_inode->i_mapping, >> > > - desc->pg_error); >> > > + nfs_context_set_write_error(req->wb_context, >> > > + desc->pg_error); >> > > >> > > func = desc->pg_completion_ops->error_cleanup; >> > > for (midx = 0; midx < desc->pg_mirror_count; midx++) { >> > > diff --git a/fs/nfs/write.c b/fs/nfs/write.c >> > > index b1af5dee5e0a..f702bf2def79 100644 >> > > --- a/fs/nfs/write.c >> > > +++ b/fs/nfs/write.c >> > > @@ -147,13 +147,6 @@ static void nfs_io_completion_put(struct nfs_io_completion *ioc) >> > > kref_put(&ioc->refcount, nfs_io_completion_release); >> > > } >> > > >> > > -static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error) >> > > -{ >> > > - ctx->error = error; >> > > - smp_wmb(); >> > > - set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags); >> > > -} >> > > - >> > > /* >> > > * nfs_page_find_head_request_locked - find head request associated with @page >> > > * >> > >> > This should probably be broken out into at least a 2-3 different >> > patches. >> > >> > Ok, so to make sure I understand: >> > >> > All writeback is done under the aegis of an open context, and writes >> > from different open contexts are not mergeable. We also flush to the >> > server in the case that a dirty page is written via an incompatible open >> > context. So with that we can always tie >> >> Not quite. Writes from different open contexts are sometimes mergeable, >> providing the credential is the same and there are no locks that might >> get in the way. (nfs_flush_incompatible() gets rid of conflicts writes >> to the same page as part of nfs_write_begin(). >> When writes are merged, all contexts remain reachable from the request >> through an 'nfs_page'. nfs_write_completion() iterates over all the >> nfs_pages attached to the nfs_pgio_header, and sets the context >> write_error from the hdr->error. >> > > Ok, by this account, NFS should already have "correct" error reporting > semantics on fsync. i.e. when the file is written via multiple fds, you > should get back an error on all fds if those writebacks failed. > > I have a test for nfs for the new-style error reporting: > > https://git.kernel.org/pub/scm/linux/kernel/git/jlayton/xfstests-dev.git/log/?h=wberr > > The nfs test is still pretty rickety, using soft mounts and iptables to > cause requests to fail. With the patch I originally proposed, this test > would pass. When I run this test on normal mainline kernels, it fails: > > -------------------------------8<------------------------------ > FSTYP -- nfs > PLATFORM -- Linux/x86_64 wberr 4.12.11-300.fc26.x86_64 > MKFS_OPTIONS -- knfsdsrv:/export/scratch > MOUNT_OPTIONS -- -o context=system_u:object_r:root_t:s0 knfsdsrv:/export/scratch /mnt/scratch > > nfs/002 - output mismatch (see /home/jlayton/git/xfstests/results//nfs/002.out.bad) > --- tests/nfs/002.out 2017-07-19 13:12:59.354561869 -0400 > +++ /home/jlayton/git/xfstests/results//nfs/002.out.bad 2017-09-12 11:17:17.335943539 -0400 > @@ -1,3 +1,3 @@ > QA output created by 002 > Format and mount > -Test passed! > +Success on second fsync on fd[1]! Interesting. I haven't reviewed the kernel code yet, but having looked at the test code I see that the one file is opened 10 time and that the *same* block in the file (65k?) is written via each fd. If I write to a file, and then someone else over-writes all the bytes that I wrote, then the page is written to the server and gets an error, then you could argue that as none of the bytes I wrote were involved in the error, I don't need to see the error status - someone else has taken on responsibility for that range. Maybe the test should/could write a few bytes with a different offset for each fd ?? NeilBrown > ... > (Run 'diff -u tests/nfs/002.out /home/jlayton/git/xfstests/results//nfs/002.out.bad' to see the entire diff) > Ran: nfs/002 > Failures: nfs/002 > Failed 1 of 1 tests > -------------------------------8<------------------------------ > > I'm not sure that errors are really propagated to all struct files like > you suggest above. I'll plan to look a little more closely at what's > happening here, when I get some time. > >> > >> > In that case, yes...mixing in errseq_t doesn't really buy us much here, >> > and I agree with most of the changes above. >> > >> > That said...I'm still not thrilled with how NFS_CONTEXT_ERROR_WRITE is >> > handled in this code. That flag is set when a write fails, but is only >> > cleared on fsync. >> > >> > That seems wrong to me. Why wait for an fsync to start doing async >> > writes again once they start working? What if the application never does >> > an fsync? Clearing that flag on a successful WRITE seems like it'd make >> > more sense. >> >> We don't really 'wait' for an fsync. Having NFS_CONTEXT_ERROR_WRITE >> means that the very next write will force an fsync >> (nfs_need_check_write()). So we really just wait for the next write. >> The current code doesn't seem "obviously right" to me, but it isn't >> "obviously wrong" either, and I can only make it obviously right to me >> by making it more complex, and I don't think I can justify that. > > Thanks for pointing this out. I missed the bit about it forcing the > fsync when this fails. I agree that that should be fine. > > -- > Jeff Layton <jlayton@xxxxxxxxxx> > -- > To unsubscribe from this list: send the line "unsubscribe linux-nfs" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html
Attachment:
signature.asc
Description: PGP signature