| Message ID | 1502369997-15665-1-git-send-email-lczerner@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu 10-08-17 14:59:57, Lukas Czerner wrote: > Currently when mixing buffered reads and asynchronous direct writes it > is possible to end up with the situation where we have stale data in the > page cache while the new data is already written to disk. This is > permanent until the affected pages are flushed away. Despite the fact > that mixing buffered and direct IO is ill-advised it does pose a thread > for a data integrity, is unexpected and should be fixed. > > Fix this by deferring completion of asynchronous direct writes to a > process context in the case that there are mapped pages to be found in > the inode. Later before the completion in dio_complete() invalidate > the pages in question. This ensures that after the completion the pages > in the written area are either unmapped, or populated with up-to-date > data. Also do the same for the iomap case which uses > iomap_dio_complete() instead. > > This has a side effect of deferring the completion to a process context > for every AIO DIO that happens on inode that has pages mapped. However > since the consensus is that this is ill-advised practice the performance > implication should not be a problem. > > This was based on proposal from Jeff Moyer, thanks! It seems the invalidation can be also removed from generic_file_direct_write(), can't it? It is duplicit there the same way as it was in the iomap code... Honza > > Signed-off-by: Lukas Czerner <lczerner@redhat.com> > Cc: Jeff Moyer <jmoyer@redhat.com> > --- > v2: Remove leftover ret variable from invalidate call in iomap_dio_complete > v3: Do not invalidate in case of error. Add some coments > v4: Remove unnecessary variable, remove unnecessary inner braces > v5: Style changes > v6: Remove redundant invalidatepage, add warning and comment > > fs/direct-io.c | 49 +++++++++++++++++++++++++++++++++++++++++++------ > fs/iomap.c | 29 ++++++++++++++++------------- > 2 files changed, 59 insertions(+), 19 deletions(-) > > diff --git a/fs/direct-io.c b/fs/direct-io.c > index 08cf278..ffb9e19 100644 > --- a/fs/direct-io.c > +++ b/fs/direct-io.c > @@ -229,6 +229,7 @@ static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async) > { > loff_t offset = dio->iocb->ki_pos; > ssize_t transferred = 0; > + int err; > > /* > * AIO submission can race with bio completion to get here while > @@ -258,8 +259,22 @@ static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async) > if (ret == 0) > ret = transferred; > > + /* > + * Try again to invalidate clean pages which might have been cached by > + * non-direct readahead, or faulted in by get_user_pages() if the source > + * of the write was an mmap'ed region of the file we're writing. Either > + * one is a pretty crazy thing to do, so we don't support it 100%. If > + * this invalidation fails, tough, the write still worked... > + */ > + if (ret > 0 && dio->op == REQ_OP_WRITE && > + dio->inode->i_mapping->nrpages) { > + err = invalidate_inode_pages2_range(dio->inode->i_mapping, > + offset >> PAGE_SHIFT, > + (offset + ret - 1) >> PAGE_SHIFT); > + WARN_ON_ONCE(err); > + } > + > if (dio->end_io) { > - int err; > > // XXX: ki_pos?? > err = dio->end_io(dio->iocb, offset, ret, dio->private); > @@ -304,6 +319,7 @@ static void dio_bio_end_aio(struct bio *bio) > struct dio *dio = bio->bi_private; > unsigned long remaining; > unsigned long flags; > + bool defer_completion = false; > > /* cleanup the bio */ > dio_bio_complete(dio, bio); > @@ -315,7 +331,19 @@ static void dio_bio_end_aio(struct bio *bio) > spin_unlock_irqrestore(&dio->bio_lock, flags); > > if (remaining == 0) { > - if (dio->result && dio->defer_completion) { > + /* > + * Defer completion when defer_completion is set or > + * when the inode has pages mapped and this is AIO write. > + * We need to invalidate those pages because there is a > + * chance they contain stale data in the case buffered IO > + * went in between AIO submission and completion into the > + * same region. > + */ > + if (dio->result) > + defer_completion = dio->defer_completion || > + (dio->op == REQ_OP_WRITE && > + dio->inode->i_mapping->nrpages); > + if (defer_completion) { > INIT_WORK(&dio->complete_work, dio_aio_complete_work); > queue_work(dio->inode->i_sb->s_dio_done_wq, > &dio->complete_work); > @@ -1210,10 +1238,19 @@ do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, > * For AIO O_(D)SYNC writes we need to defer completions to a workqueue > * so that we can call ->fsync. > */ > - if (dio->is_async && iov_iter_rw(iter) == WRITE && > - ((iocb->ki_filp->f_flags & O_DSYNC) || > - IS_SYNC(iocb->ki_filp->f_mapping->host))) { > - retval = dio_set_defer_completion(dio); > + if (dio->is_async && iov_iter_rw(iter) == WRITE) { > + retval = 0; > + if ((iocb->ki_filp->f_flags & O_DSYNC) || > + IS_SYNC(iocb->ki_filp->f_mapping->host)) > + retval = dio_set_defer_completion(dio); > + else if (!dio->inode->i_sb->s_dio_done_wq) { > + /* > + * In case of AIO write racing with buffered read we > + * need to defer completion. We can't decide this now, > + * however the workqueue needs to be initialized here. > + */ > + retval = sb_init_dio_done_wq(dio->inode->i_sb); > + } > if (retval) { > /* > * We grab i_mutex only for reads so we don't have > diff --git a/fs/iomap.c b/fs/iomap.c > index 0392661..c3e299a 100644 > --- a/fs/iomap.c > +++ b/fs/iomap.c > @@ -713,8 +713,24 @@ struct iomap_dio { > static ssize_t iomap_dio_complete(struct iomap_dio *dio) > { > struct kiocb *iocb = dio->iocb; > + struct inode *inode = file_inode(iocb->ki_filp); > ssize_t ret; > > + /* > + * Try again to invalidate clean pages which might have been cached by > + * non-direct readahead, or faulted in by get_user_pages() if the source > + * of the write was an mmap'ed region of the file we're writing. Either > + * one is a pretty crazy thing to do, so we don't support it 100%. If > + * this invalidation fails, tough, the write still worked... > + */ > + if (!dio->error && > + (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) { > + ret = invalidate_inode_pages2_range(inode->i_mapping, > + iocb->ki_pos >> PAGE_SHIFT, > + (iocb->ki_pos + dio->size - 1) >> PAGE_SHIFT); > + WARN_ON_ONCE(ret); > + } > + > if (dio->end_io) { > ret = dio->end_io(iocb, > dio->error ? dio->error : dio->size, > @@ -1042,19 +1058,6 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter, > > ret = iomap_dio_complete(dio); > > - /* > - * Try again to invalidate clean pages which might have been cached by > - * non-direct readahead, or faulted in by get_user_pages() if the source > - * of the write was an mmap'ed region of the file we're writing. Either > - * one is a pretty crazy thing to do, so we don't support it 100%. If > - * this invalidation fails, tough, the write still worked... > - */ > - if (iov_iter_rw(iter) == WRITE) { > - int err = invalidate_inode_pages2_range(mapping, > - start >> PAGE_SHIFT, end >> PAGE_SHIFT); > - WARN_ON_ONCE(err); > - } > - > return ret; > > out_free_dio: > -- > 2.7.5 >
Jan Kara <jack@suse.cz> writes: > On Thu 10-08-17 14:59:57, Lukas Czerner wrote: >> Currently when mixing buffered reads and asynchronous direct writes it >> is possible to end up with the situation where we have stale data in the >> page cache while the new data is already written to disk. This is >> permanent until the affected pages are flushed away. Despite the fact >> that mixing buffered and direct IO is ill-advised it does pose a thread >> for a data integrity, is unexpected and should be fixed. >> >> Fix this by deferring completion of asynchronous direct writes to a >> process context in the case that there are mapped pages to be found in >> the inode. Later before the completion in dio_complete() invalidate >> the pages in question. This ensures that after the completion the pages >> in the written area are either unmapped, or populated with up-to-date >> data. Also do the same for the iomap case which uses >> iomap_dio_complete() instead. >> >> This has a side effect of deferring the completion to a process context >> for every AIO DIO that happens on inode that has pages mapped. However >> since the consensus is that this is ill-advised practice the performance >> implication should not be a problem. >> >> This was based on proposal from Jeff Moyer, thanks! > > It seems the invalidation can be also removed from > generic_file_direct_write(), can't it? It is duplicit there the same way as > it was in the iomap code... Yep, sure looks that way. -Jeff
On Thu, Aug 10, 2017 at 10:22:47AM -0400, Jeff Moyer wrote: > Jan Kara <jack@suse.cz> writes: > > > On Thu 10-08-17 14:59:57, Lukas Czerner wrote: > >> Currently when mixing buffered reads and asynchronous direct writes it > >> is possible to end up with the situation where we have stale data in the > >> page cache while the new data is already written to disk. This is > >> permanent until the affected pages are flushed away. Despite the fact > >> that mixing buffered and direct IO is ill-advised it does pose a thread > >> for a data integrity, is unexpected and should be fixed. > >> > >> Fix this by deferring completion of asynchronous direct writes to a > >> process context in the case that there are mapped pages to be found in > >> the inode. Later before the completion in dio_complete() invalidate > >> the pages in question. This ensures that after the completion the pages > >> in the written area are either unmapped, or populated with up-to-date > >> data. Also do the same for the iomap case which uses > >> iomap_dio_complete() instead. > >> > >> This has a side effect of deferring the completion to a process context > >> for every AIO DIO that happens on inode that has pages mapped. However > >> since the consensus is that this is ill-advised practice the performance > >> implication should not be a problem. > >> > >> This was based on proposal from Jeff Moyer, thanks! > > > > It seems the invalidation can be also removed from > > generic_file_direct_write(), can't it? It is duplicit there the same way as > > it was in the iomap code... > > Yep, sure looks that way. Hrm, ok. Technically speaking generic_file_direct_write does not have to eventually end up with dio_complete() being called. This will change the behaviour for those that implement dio differently. Looking at the users now, vast majority will end up with complete_dio() so maybe this is not a problem. This is in contrast with iomap_dio_rw() which will end up calling iomap_dio_complete() so the situation is different there. Maybe adding mapping->nrpages check would be better than outright removing it ? -Lukas > > -Jeff
Hello, On Fri 11-08-17 11:03:01, Lukas Czerner wrote: > On Thu, Aug 10, 2017 at 10:22:47AM -0400, Jeff Moyer wrote: > > Jan Kara <jack@suse.cz> writes: > > > > > On Thu 10-08-17 14:59:57, Lukas Czerner wrote: > > >> Currently when mixing buffered reads and asynchronous direct writes it > > >> is possible to end up with the situation where we have stale data in the > > >> page cache while the new data is already written to disk. This is > > >> permanent until the affected pages are flushed away. Despite the fact > > >> that mixing buffered and direct IO is ill-advised it does pose a thread > > >> for a data integrity, is unexpected and should be fixed. > > >> > > >> Fix this by deferring completion of asynchronous direct writes to a > > >> process context in the case that there are mapped pages to be found in > > >> the inode. Later before the completion in dio_complete() invalidate > > >> the pages in question. This ensures that after the completion the pages > > >> in the written area are either unmapped, or populated with up-to-date > > >> data. Also do the same for the iomap case which uses > > >> iomap_dio_complete() instead. > > >> > > >> This has a side effect of deferring the completion to a process context > > >> for every AIO DIO that happens on inode that has pages mapped. However > > >> since the consensus is that this is ill-advised practice the performance > > >> implication should not be a problem. > > >> > > >> This was based on proposal from Jeff Moyer, thanks! > > > > > > It seems the invalidation can be also removed from > > > generic_file_direct_write(), can't it? It is duplicit there the same way as > > > it was in the iomap code... > > > > Yep, sure looks that way. > > Hrm, ok. Technically speaking generic_file_direct_write does not have to > eventually end up with dio_complete() being called. This will change the > behaviour for those that implement dio differently. Looking at the users > now, vast majority will end up with complete_dio() so maybe this is not > a problem. OK, so this seems to be the problem with 9p, fuse, nfs, lustre. > This is in contrast with iomap_dio_rw() which will end up calling > iomap_dio_complete() so the situation is different there. > > Maybe adding mapping->nrpages check would be better than outright > removing it ? OK, I agree we cannot just remove the invalidation. But shouldn't we rather fix the above mentioned filesystems? Otherwise they will keep having issues you are trying to fix? But for now I could live with keeping the invalidation behind nrpages check and adding a comment why we kept it there... Honza
On Mon, Aug 14, 2017 at 11:43:31AM +0200, Jan Kara wrote: > Hello, > > On Fri 11-08-17 11:03:01, Lukas Czerner wrote: > > On Thu, Aug 10, 2017 at 10:22:47AM -0400, Jeff Moyer wrote: > > > Jan Kara <jack@suse.cz> writes: > > > > > > > On Thu 10-08-17 14:59:57, Lukas Czerner wrote: > > > >> Currently when mixing buffered reads and asynchronous direct writes it > > > >> is possible to end up with the situation where we have stale data in the > > > >> page cache while the new data is already written to disk. This is > > > >> permanent until the affected pages are flushed away. Despite the fact > > > >> that mixing buffered and direct IO is ill-advised it does pose a thread > > > >> for a data integrity, is unexpected and should be fixed. > > > >> > > > >> Fix this by deferring completion of asynchronous direct writes to a > > > >> process context in the case that there are mapped pages to be found in > > > >> the inode. Later before the completion in dio_complete() invalidate > > > >> the pages in question. This ensures that after the completion the pages > > > >> in the written area are either unmapped, or populated with up-to-date > > > >> data. Also do the same for the iomap case which uses > > > >> iomap_dio_complete() instead. > > > >> > > > >> This has a side effect of deferring the completion to a process context > > > >> for every AIO DIO that happens on inode that has pages mapped. However > > > >> since the consensus is that this is ill-advised practice the performance > > > >> implication should not be a problem. > > > >> > > > >> This was based on proposal from Jeff Moyer, thanks! > > > > > > > > It seems the invalidation can be also removed from > > > > generic_file_direct_write(), can't it? It is duplicit there the same way as > > > > it was in the iomap code... > > > > > > Yep, sure looks that way. > > > > Hrm, ok. Technically speaking generic_file_direct_write does not have to > > eventually end up with dio_complete() being called. This will change the > > behaviour for those that implement dio differently. Looking at the users > > now, vast majority will end up with complete_dio() so maybe this is not > > a problem. > > OK, so this seems to be the problem with 9p, fuse, nfs, lustre. > > > This is in contrast with iomap_dio_rw() which will end up calling > > iomap_dio_complete() so the situation is different there. > > > > Maybe adding mapping->nrpages check would be better than outright > > removing it ? > > OK, I agree we cannot just remove the invalidation. But shouldn't we rather > fix the above mentioned filesystems? Otherwise they will keep having issues > you are trying to fix? But for now I could live with keeping the > invalidation behind nrpages check and adding a comment why we kept it > there... Right, I'd rather have closere on this neverending patch. The rest of the fs can be fixed later. -Lukas > > Honza > -- > Jan Kara <jack@suse.com> > SUSE Labs, CR
diff --git a/fs/direct-io.c b/fs/direct-io.c index 08cf278..ffb9e19 100644 --- a/fs/direct-io.c +++ b/fs/direct-io.c @@ -229,6 +229,7 @@ static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async) { loff_t offset = dio->iocb->ki_pos; ssize_t transferred = 0; + int err; /* * AIO submission can race with bio completion to get here while @@ -258,8 +259,22 @@ static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async) if (ret == 0) ret = transferred; + /* + * Try again to invalidate clean pages which might have been cached by + * non-direct readahead, or faulted in by get_user_pages() if the source + * of the write was an mmap'ed region of the file we're writing. Either + * one is a pretty crazy thing to do, so we don't support it 100%. If + * this invalidation fails, tough, the write still worked... + */ + if (ret > 0 && dio->op == REQ_OP_WRITE && + dio->inode->i_mapping->nrpages) { + err = invalidate_inode_pages2_range(dio->inode->i_mapping, + offset >> PAGE_SHIFT, + (offset + ret - 1) >> PAGE_SHIFT); + WARN_ON_ONCE(err); + } + if (dio->end_io) { - int err; // XXX: ki_pos?? err = dio->end_io(dio->iocb, offset, ret, dio->private); @@ -304,6 +319,7 @@ static void dio_bio_end_aio(struct bio *bio) struct dio *dio = bio->bi_private; unsigned long remaining; unsigned long flags; + bool defer_completion = false; /* cleanup the bio */ dio_bio_complete(dio, bio); @@ -315,7 +331,19 @@ static void dio_bio_end_aio(struct bio *bio) spin_unlock_irqrestore(&dio->bio_lock, flags); if (remaining == 0) { - if (dio->result && dio->defer_completion) { + /* + * Defer completion when defer_completion is set or + * when the inode has pages mapped and this is AIO write. + * We need to invalidate those pages because there is a + * chance they contain stale data in the case buffered IO + * went in between AIO submission and completion into the + * same region. + */ + if (dio->result) + defer_completion = dio->defer_completion || + (dio->op == REQ_OP_WRITE && + dio->inode->i_mapping->nrpages); + if (defer_completion) { INIT_WORK(&dio->complete_work, dio_aio_complete_work); queue_work(dio->inode->i_sb->s_dio_done_wq, &dio->complete_work); @@ -1210,10 +1238,19 @@ do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, * For AIO O_(D)SYNC writes we need to defer completions to a workqueue * so that we can call ->fsync. */ - if (dio->is_async && iov_iter_rw(iter) == WRITE && - ((iocb->ki_filp->f_flags & O_DSYNC) || - IS_SYNC(iocb->ki_filp->f_mapping->host))) { - retval = dio_set_defer_completion(dio); + if (dio->is_async && iov_iter_rw(iter) == WRITE) { + retval = 0; + if ((iocb->ki_filp->f_flags & O_DSYNC) || + IS_SYNC(iocb->ki_filp->f_mapping->host)) + retval = dio_set_defer_completion(dio); + else if (!dio->inode->i_sb->s_dio_done_wq) { + /* + * In case of AIO write racing with buffered read we + * need to defer completion. We can't decide this now, + * however the workqueue needs to be initialized here. + */ + retval = sb_init_dio_done_wq(dio->inode->i_sb); + } if (retval) { /* * We grab i_mutex only for reads so we don't have diff --git a/fs/iomap.c b/fs/iomap.c index 0392661..c3e299a 100644 --- a/fs/iomap.c +++ b/fs/iomap.c @@ -713,8 +713,24 @@ struct iomap_dio { static ssize_t iomap_dio_complete(struct iomap_dio *dio) { struct kiocb *iocb = dio->iocb; + struct inode *inode = file_inode(iocb->ki_filp); ssize_t ret; + /* + * Try again to invalidate clean pages which might have been cached by + * non-direct readahead, or faulted in by get_user_pages() if the source + * of the write was an mmap'ed region of the file we're writing. Either + * one is a pretty crazy thing to do, so we don't support it 100%. If + * this invalidation fails, tough, the write still worked... + */ + if (!dio->error && + (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) { + ret = invalidate_inode_pages2_range(inode->i_mapping, + iocb->ki_pos >> PAGE_SHIFT, + (iocb->ki_pos + dio->size - 1) >> PAGE_SHIFT); + WARN_ON_ONCE(ret); + } + if (dio->end_io) { ret = dio->end_io(iocb, dio->error ? dio->error : dio->size, @@ -1042,19 +1058,6 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter, ret = iomap_dio_complete(dio); - /* - * Try again to invalidate clean pages which might have been cached by - * non-direct readahead, or faulted in by get_user_pages() if the source - * of the write was an mmap'ed region of the file we're writing. Either - * one is a pretty crazy thing to do, so we don't support it 100%. If - * this invalidation fails, tough, the write still worked... - */ - if (iov_iter_rw(iter) == WRITE) { - int err = invalidate_inode_pages2_range(mapping, - start >> PAGE_SHIFT, end >> PAGE_SHIFT); - WARN_ON_ONCE(err); - } - return ret; out_free_dio:
Currently when mixing buffered reads and asynchronous direct writes it is possible to end up with the situation where we have stale data in the page cache while the new data is already written to disk. This is permanent until the affected pages are flushed away. Despite the fact that mixing buffered and direct IO is ill-advised it does pose a thread for a data integrity, is unexpected and should be fixed. Fix this by deferring completion of asynchronous direct writes to a process context in the case that there are mapped pages to be found in the inode. Later before the completion in dio_complete() invalidate the pages in question. This ensures that after the completion the pages in the written area are either unmapped, or populated with up-to-date data. Also do the same for the iomap case which uses iomap_dio_complete() instead. This has a side effect of deferring the completion to a process context for every AIO DIO that happens on inode that has pages mapped. However since the consensus is that this is ill-advised practice the performance implication should not be a problem. This was based on proposal from Jeff Moyer, thanks! Signed-off-by: Lukas Czerner <lczerner@redhat.com> Cc: Jeff Moyer <jmoyer@redhat.com> --- v2: Remove leftover ret variable from invalidate call in iomap_dio_complete v3: Do not invalidate in case of error. Add some coments v4: Remove unnecessary variable, remove unnecessary inner braces v5: Style changes v6: Remove redundant invalidatepage, add warning and comment fs/direct-io.c | 49 +++++++++++++++++++++++++++++++++++++++++++------ fs/iomap.c | 29 ++++++++++++++++------------- 2 files changed, 59 insertions(+), 19 deletions(-)