| Message ID | 20241112160055.1829361-1-maharmstone@fb.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | btrfs: fix lockdep warnings on io_uring encoded reads | expand |
On 12.11.24 17:01, Mark Harmstone wrote: > Lockdep doesn't like the fact that btrfs_uring_read_extent() returns to > userspace still holding the inode lock, even though we release it once > the I/O finishes. Add calls to rwsem_release() and rwsem_acquire_read() to > work round this. > > Signed-off-by: Mark Harmstone <maharmstone@fb.com> > Reported-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> > --- > fs/btrfs/ioctl.c | 14 ++++++++++++++ > 1 file changed, 14 insertions(+) > > diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c > index 1fdeb216bf6c..6ea01e4f940e 100644 > --- a/fs/btrfs/ioctl.c > +++ b/fs/btrfs/ioctl.c > @@ -4752,6 +4752,11 @@ static void btrfs_uring_read_finished(struct io_uring_cmd *cmd, unsigned int iss > size_t page_offset; > ssize_t ret; > > +#ifdef CONFIG_DEBUG_LOCK_ALLOC > + /* The inode lock has already been acquired in btrfs_uring_read_extent. */ > + rwsem_acquire_read(&inode->vfs_inode.i_rwsem.dep_map, 0, 0, _THIS_IP_); > +#endif > + > if (priv->err) { > ret = priv->err; > goto out; > @@ -4860,6 +4865,15 @@ static int btrfs_uring_read_extent(struct kiocb *iocb, struct iov_iter *iter, > * and inode and freeing the allocations. > */ > > +#ifdef CONFIG_DEBUG_LOCK_ALLOC > + /* > + * We're returning to userspace with the inode lock held, and that's > + * okay - it'll get unlocked in a kthread. Call rwsem_release to > + * avoid confusing lockdep. > + */ > + rwsem_release(&inode->vfs_inode.i_rwsem.dep_map, _THIS_IP_); > +#endif > + > return -EIOCBQUEUED; > > out_fail: Can't say anything about the correctness (as I have no clue), but we have wrappers around rwsem_release (btrfs_lockdep_release()) and rwsem_acquire_read (btrfs_lockdep_acquire()) that I think serve the documentation purpose.
On 13/11/24 08:21, Johannes Thumshirn wrote: > > > On 12.11.24 17:01, Mark Harmstone wrote: >> Lockdep doesn't like the fact that btrfs_uring_read_extent() returns to >> userspace still holding the inode lock, even though we release it once >> the I/O finishes. Add calls to rwsem_release() and rwsem_acquire_read() to >> work round this. >> >> Signed-off-by: Mark Harmstone <maharmstone@fb.com> >> Reported-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> >> --- >> fs/btrfs/ioctl.c | 14 ++++++++++++++ >> 1 file changed, 14 insertions(+) >> >> diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c >> index 1fdeb216bf6c..6ea01e4f940e 100644 >> --- a/fs/btrfs/ioctl.c >> +++ b/fs/btrfs/ioctl.c >> @@ -4752,6 +4752,11 @@ static void btrfs_uring_read_finished(struct io_uring_cmd *cmd, unsigned int iss >> size_t page_offset; >> ssize_t ret; >> >> +#ifdef CONFIG_DEBUG_LOCK_ALLOC >> + /* The inode lock has already been acquired in btrfs_uring_read_extent. */ >> + rwsem_acquire_read(&inode->vfs_inode.i_rwsem.dep_map, 0, 0, _THIS_IP_); >> +#endif >> + >> if (priv->err) { >> ret = priv->err; >> goto out; >> @@ -4860,6 +4865,15 @@ static int btrfs_uring_read_extent(struct kiocb *iocb, struct iov_iter *iter, >> * and inode and freeing the allocations. >> */ >> >> +#ifdef CONFIG_DEBUG_LOCK_ALLOC >> + /* >> + * We're returning to userspace with the inode lock held, and that's >> + * okay - it'll get unlocked in a kthread. Call rwsem_release to >> + * avoid confusing lockdep. >> + */ >> + rwsem_release(&inode->vfs_inode.i_rwsem.dep_map, _THIS_IP_); >> +#endif >> + >> return -EIOCBQUEUED; >> >> out_fail: > > Can't say anything about the correctness (as I have no clue), but we > have wrappers around rwsem_release (btrfs_lockdep_release()) and > rwsem_acquire_read (btrfs_lockdep_acquire()) that I think serve the > documentation purpose. btrfs_lockdep_acquire(a,b) calls rwsem_acquire_read on &a->b_map, so can't be used for something located at &inode->vfs_inode.i_rwsem.dep_map. Mark
On 13.11.24 11:24, Mark Harmstone wrote: > On 13/11/24 08:21, Johannes Thumshirn wrote: >>> >> On 12.11.24 17:01, Mark Harmstone wrote: >>> Lockdep doesn't like the fact that btrfs_uring_read_extent() returns to >>> userspace still holding the inode lock, even though we release it once >>> the I/O finishes. Add calls to rwsem_release() and rwsem_acquire_read() to >>> work round this. >>> >>> Signed-off-by: Mark Harmstone <maharmstone@fb.com> >>> Reported-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> >>> --- >>> fs/btrfs/ioctl.c | 14 ++++++++++++++ >>> 1 file changed, 14 insertions(+) >>> >>> diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c >>> index 1fdeb216bf6c..6ea01e4f940e 100644 >>> --- a/fs/btrfs/ioctl.c >>> +++ b/fs/btrfs/ioctl.c >>> @@ -4752,6 +4752,11 @@ static void btrfs_uring_read_finished(struct io_uring_cmd *cmd, unsigned int iss >>> size_t page_offset; >>> ssize_t ret; >>> >>> +#ifdef CONFIG_DEBUG_LOCK_ALLOC >>> + /* The inode lock has already been acquired in btrfs_uring_read_extent. */ >>> + rwsem_acquire_read(&inode->vfs_inode.i_rwsem.dep_map, 0, 0, _THIS_IP_); >>> +#endif >>> + >>> if (priv->err) { >>> ret = priv->err; >>> goto out; >>> @@ -4860,6 +4865,15 @@ static int btrfs_uring_read_extent(struct kiocb *iocb, struct iov_iter *iter, >>> * and inode and freeing the allocations. >>> */ >>> >>> +#ifdef CONFIG_DEBUG_LOCK_ALLOC >>> + /* >>> + * We're returning to userspace with the inode lock held, and that's >>> + * okay - it'll get unlocked in a kthread. Call rwsem_release to >>> + * avoid confusing lockdep. >>> + */ >>> + rwsem_release(&inode->vfs_inode.i_rwsem.dep_map, _THIS_IP_); >>> +#endif >>> + >>> return -EIOCBQUEUED; >>> >>> out_fail: >> >> Can't say anything about the correctness (as I have no clue), but we >> have wrappers around rwsem_release (btrfs_lockdep_release()) and >> rwsem_acquire_read (btrfs_lockdep_acquire()) that I think serve the >> documentation purpose. > > btrfs_lockdep_acquire(a,b) calls rwsem_acquire_read on &a->b_map, so > can't be used for something located at &inode->vfs_inode.i_rwsem.dep_map. Ah my bad, sorry.
On Tue, Nov 12, 2024 at 04:00:49PM +0000, Mark Harmstone wrote: > Lockdep doesn't like the fact that btrfs_uring_read_extent() returns to > userspace still holding the inode lock, even though we release it once > the I/O finishes. Add calls to rwsem_release() and rwsem_acquire_read() to > work round this. > > Signed-off-by: Mark Harmstone <maharmstone@fb.com> > Reported-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> > --- > fs/btrfs/ioctl.c | 14 ++++++++++++++ > 1 file changed, 14 insertions(+) > > diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c > index 1fdeb216bf6c..6ea01e4f940e 100644 > --- a/fs/btrfs/ioctl.c > +++ b/fs/btrfs/ioctl.c > @@ -4752,6 +4752,11 @@ static void btrfs_uring_read_finished(struct io_uring_cmd *cmd, unsigned int iss > size_t page_offset; > ssize_t ret; > > +#ifdef CONFIG_DEBUG_LOCK_ALLOC > + /* The inode lock has already been acquired in btrfs_uring_read_extent. */ > + rwsem_acquire_read(&inode->vfs_inode.i_rwsem.dep_map, 0, 0, _THIS_IP_); > +#endif Please put that to a wrapper, we want to avoid #ifdefs in the middle of functions (there are exceptions), especially when the macro is for lock debugging. The wrapper name can follow naming and syntax of the other btrfs_lockde_* wrappers, so like btrfs_lockdep_inode_acquire(owner, lock). There is only one rwsem in inode, but for clarity and consistency I think it makes sense. btrfs_lockdep_inode_acquire(inode, i_rwsem);
diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c index 1fdeb216bf6c..6ea01e4f940e 100644 --- a/fs/btrfs/ioctl.c +++ b/fs/btrfs/ioctl.c @@ -4752,6 +4752,11 @@ static void btrfs_uring_read_finished(struct io_uring_cmd *cmd, unsigned int iss size_t page_offset; ssize_t ret; +#ifdef CONFIG_DEBUG_LOCK_ALLOC + /* The inode lock has already been acquired in btrfs_uring_read_extent. */ + rwsem_acquire_read(&inode->vfs_inode.i_rwsem.dep_map, 0, 0, _THIS_IP_); +#endif + if (priv->err) { ret = priv->err; goto out; @@ -4860,6 +4865,15 @@ static int btrfs_uring_read_extent(struct kiocb *iocb, struct iov_iter *iter, * and inode and freeing the allocations. */ +#ifdef CONFIG_DEBUG_LOCK_ALLOC + /* + * We're returning to userspace with the inode lock held, and that's + * okay - it'll get unlocked in a kthread. Call rwsem_release to + * avoid confusing lockdep. + */ + rwsem_release(&inode->vfs_inode.i_rwsem.dep_map, _THIS_IP_); +#endif + return -EIOCBQUEUED; out_fail:
Lockdep doesn't like the fact that btrfs_uring_read_extent() returns to userspace still holding the inode lock, even though we release it once the I/O finishes. Add calls to rwsem_release() and rwsem_acquire_read() to work round this. Signed-off-by: Mark Harmstone <maharmstone@fb.com> Reported-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> --- fs/btrfs/ioctl.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)