| Message ID | 1467291229-13548-1-git-send-email-bfoster@redhat.com (mailing list archive) |
|---|---|
| State | Superseded, archived |
| Headers | show |
On Thu, Jun 30, 2016 at 08:53:49AM -0400, Brian Foster wrote: > Newly allocated XFS metadata buffers are added to the LRU once the hold > count is released, which typically occurs after I/O completion. There is > no other mechanism at current that tracks the existence or I/O state of > a new buffer. Further, readahead I/O tends to be submitted > asynchronously by nature, which means the I/O can remain in flight and > actually complete long after the calling context is gone. This means > that file descriptors or any other holds on the filesystem can be > released, allowing the filesystem to be unmounted while I/O is still in > flight. When I/O completion occurs, core data structures may have been > freed, causing completion to run into invalid memory accesses and likely > to panic. > > This problem is reproduced on XFS via directory readahead. A filesystem > is mounted, a directory is opened/closed and the filesystem immediately > unmounted. The open/close cycle triggers a directory readahead that if > delayed long enough, runs buffer I/O completion after the unmount has > completed. > > To work around this problem, add readahead buffers to the LRU earlier > than other buffers (when the buffer is allocated, specifically). The > buffer hold count will ultimately remain until I/O completion, which > means any shrinker activity will skip the buffer until then. This makes > the buffer visible to xfs_wait_buftarg(), however, which ensures that an > unmount or quiesce waits for I/O completion appropriately. > > Signed-off-by: Brian Foster <bfoster@redhat.com> > --- > > This addresses the problem reproduced by the recently posted xfstests > test: > > http://thread.gmane.org/gmane.comp.file-systems.fstests/2740 > > This could probably be made more involved, i.e., to create another list > of buffers in flight or some such. This seems more simple/sane to me, > however, and survives my testing so far... > > Brian > > fs/xfs/xfs_buf.c | 12 ++++++++++++ > 1 file changed, 12 insertions(+) > > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c > index 4665ff6..3f03df9 100644 > --- a/fs/xfs/xfs_buf.c > +++ b/fs/xfs/xfs_buf.c > @@ -590,8 +590,20 @@ xfs_buf_get_map( > return NULL; > } > > + /* > + * If the buffer found doesn't match the one allocated above, somebody > + * else beat us to insertion and we can toss the new one. > + * > + * If we did add the buffer and it happens to be readahead, add to the > + * LRU now rather than waiting until the hold is released. Otherwise, > + * the buffer is not visible to xfs_wait_buftarg() while in flight and > + * nothing else prevents an unmount before I/O completion. > + */ > if (bp != new_bp) > xfs_buf_free(new_bp); > + else if (flags & XBF_READ_AHEAD && > + list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) > + atomic_inc(&bp->b_hold); This doesn't sit right with me. The LRU is for "unused" objects, and readahead objects are not unused until IO completes and nobody is waiting on them. As it is, this points out another problem with readahead buffers - they aren't actually being cached properly because b_lru_ref == 0, which means they are immediately reclaimed on IO completion rather than being added to the LRU.... I also think that it's not sufficient to cover the generic case of async IO that has no waiter. i.e. we could do get_buf, submit async write, drop submitter reference, and now we have the same problem but on a write. i.e. this problem is and async IO issue, not a readahead issue. I think that it might be better to fix it by doing this: 1. ensure async IO submission always has b_lru_ref set, and if it isn't, set it to 1. This ensures the buffer will be added to the LRU on completion if it isn't already there. 2. keep a count of async buffer IO in progress. A per-cpu counter in the buftarg will be fine for this. Increment in xfs_buf_submit(), decrement in the xfs_buf_rele() call from xfs_buf_iodone() once we've determined if the buffer needs adding to the LRU or not. 3. make xfs_wait_buftarg() wait until the async IO count goes to zero before it gives up trying to release buffers on the LRU. That will ensure readahead buffers are cached, and we capture both async read and async write buffers in xfs_wait_buftarg(). Cheers, Dave.
On Fri, Jul 01, 2016 at 08:44:57AM +1000, Dave Chinner wrote: > On Thu, Jun 30, 2016 at 08:53:49AM -0400, Brian Foster wrote: > > Newly allocated XFS metadata buffers are added to the LRU once the hold > > count is released, which typically occurs after I/O completion. There is > > no other mechanism at current that tracks the existence or I/O state of > > a new buffer. Further, readahead I/O tends to be submitted > > asynchronously by nature, which means the I/O can remain in flight and > > actually complete long after the calling context is gone. This means > > that file descriptors or any other holds on the filesystem can be > > released, allowing the filesystem to be unmounted while I/O is still in > > flight. When I/O completion occurs, core data structures may have been > > freed, causing completion to run into invalid memory accesses and likely > > to panic. > > > > This problem is reproduced on XFS via directory readahead. A filesystem > > is mounted, a directory is opened/closed and the filesystem immediately > > unmounted. The open/close cycle triggers a directory readahead that if > > delayed long enough, runs buffer I/O completion after the unmount has > > completed. > > > > To work around this problem, add readahead buffers to the LRU earlier > > than other buffers (when the buffer is allocated, specifically). The > > buffer hold count will ultimately remain until I/O completion, which > > means any shrinker activity will skip the buffer until then. This makes > > the buffer visible to xfs_wait_buftarg(), however, which ensures that an > > unmount or quiesce waits for I/O completion appropriately. > > > > Signed-off-by: Brian Foster <bfoster@redhat.com> > > --- > > > > This addresses the problem reproduced by the recently posted xfstests > > test: > > > > http://thread.gmane.org/gmane.comp.file-systems.fstests/2740 > > > > This could probably be made more involved, i.e., to create another list > > of buffers in flight or some such. This seems more simple/sane to me, > > however, and survives my testing so far... > > > > Brian > > > > fs/xfs/xfs_buf.c | 12 ++++++++++++ > > 1 file changed, 12 insertions(+) > > > > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c > > index 4665ff6..3f03df9 100644 > > --- a/fs/xfs/xfs_buf.c > > +++ b/fs/xfs/xfs_buf.c > > @@ -590,8 +590,20 @@ xfs_buf_get_map( > > return NULL; > > } > > > > + /* > > + * If the buffer found doesn't match the one allocated above, somebody > > + * else beat us to insertion and we can toss the new one. > > + * > > + * If we did add the buffer and it happens to be readahead, add to the > > + * LRU now rather than waiting until the hold is released. Otherwise, > > + * the buffer is not visible to xfs_wait_buftarg() while in flight and > > + * nothing else prevents an unmount before I/O completion. > > + */ > > if (bp != new_bp) > > xfs_buf_free(new_bp); > > + else if (flags & XBF_READ_AHEAD && > > + list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) > > + atomic_inc(&bp->b_hold); > > This doesn't sit right with me. The LRU is for "unused" objects, and > readahead objects are not unused until IO completes and nobody is > waiting on them. > Sure, but don't buffers remain on the LRU once they are reused? My impression is that this patch really doesn't change behavior in that regard. IOWs, the buffer is going to end up on the LRU either way and either end up disposed if it is never actually accessed or "used" otherwise and the hold count is bumped. Further, isn't the hold count mechanism precisely to handle "used" buffers on the LRU? > As it is, this points out another problem with readahead buffers - > they aren't actually being cached properly because b_lru_ref == 0, > which means they are immediately reclaimed on IO completion rather > than being added to the LRU.... > Not following... _xfs_buf_alloc() sets b_lru_ref to 1 and I don't see it set/decremented unless we stale or dispose it. What am I missing? > I also think that it's not sufficient to cover the generic case of > async IO that has no waiter. i.e. we could do get_buf, submit async > write, drop submitter reference, and now we have the same problem > but on a write. i.e. this problem is and async IO issue, not a > readahead issue. > Indeed. I thought about making the current patch check for ASYNC, but opted to try and make it more isolated (I suppose ASYNC|READAHEAD would have been moreso). The other thought I had was to change where buffers are added to the LRU altogether, but didn't want to jump right to that. Is there any issue with populating the LRU with initially held buffers as such, or any particular reason LRU addition was deferred to I/O completion in the first place? > I think that it might be better to fix it by doing this: > > 1. ensure async IO submission always has b_lru_ref set, and > if it isn't, set it to 1. This ensures the buffer will be > added to the LRU on completion if it isn't already there. > See above. I'm not following why this is necessary (if it is, it seems indicative of a bug). > 2. keep a count of async buffer IO in progress. A per-cpu > counter in the buftarg will be fine for this. Increment in > xfs_buf_submit(), decrement in the xfs_buf_rele() call from > xfs_buf_iodone() once we've determined if the buffer needs > adding to the LRU or not. > > 3. make xfs_wait_buftarg() wait until the async IO count > goes to zero before it gives up trying to release buffers on > the LRU. > This is along the lines of what I was thinking wrt to a list (explicit I/O tracking one way or another). The count is more simple and does cover the arbitrary read/write case, but is still more code and seems slightly duplicative to me because in most cases the LRU wait already handles this. I'm also not totally sure we need to handle the write case. Are we susceptible to this issue anywhere that isn't already protected by the sb write protection mechanism, for example? OTOH, explicit tracking might provide a more generic way to deal with uncached buffers as opposed to lock cycling. Brian > That will ensure readahead buffers are cached, and we capture both > async read and async write buffers in xfs_wait_buftarg(). > > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com
On Thu, Jun 30, 2016 at 07:56:21PM -0400, Brian Foster wrote: > On Fri, Jul 01, 2016 at 08:44:57AM +1000, Dave Chinner wrote: > > On Thu, Jun 30, 2016 at 08:53:49AM -0400, Brian Foster wrote: > > > Newly allocated XFS metadata buffers are added to the LRU once the hold > > > count is released, which typically occurs after I/O completion. There is > > > no other mechanism at current that tracks the existence or I/O state of > > > a new buffer. Further, readahead I/O tends to be submitted > > > asynchronously by nature, which means the I/O can remain in flight and > > > actually complete long after the calling context is gone. This means > > > that file descriptors or any other holds on the filesystem can be > > > released, allowing the filesystem to be unmounted while I/O is still in > > > flight. When I/O completion occurs, core data structures may have been > > > freed, causing completion to run into invalid memory accesses and likely > > > to panic. > > > > > > This problem is reproduced on XFS via directory readahead. A filesystem > > > is mounted, a directory is opened/closed and the filesystem immediately > > > unmounted. The open/close cycle triggers a directory readahead that if > > > delayed long enough, runs buffer I/O completion after the unmount has > > > completed. > > > > > > To work around this problem, add readahead buffers to the LRU earlier > > > than other buffers (when the buffer is allocated, specifically). The > > > buffer hold count will ultimately remain until I/O completion, which > > > means any shrinker activity will skip the buffer until then. This makes > > > the buffer visible to xfs_wait_buftarg(), however, which ensures that an > > > unmount or quiesce waits for I/O completion appropriately. > > > > > > Signed-off-by: Brian Foster <bfoster@redhat.com> > > > --- > > > > > > This addresses the problem reproduced by the recently posted xfstests > > > test: > > > > > > http://thread.gmane.org/gmane.comp.file-systems.fstests/2740 > > > > > > This could probably be made more involved, i.e., to create another list > > > of buffers in flight or some such. This seems more simple/sane to me, > > > however, and survives my testing so far... > > > > > > Brian > > > > > > fs/xfs/xfs_buf.c | 12 ++++++++++++ > > > 1 file changed, 12 insertions(+) > > > > > > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c > > > index 4665ff6..3f03df9 100644 > > > --- a/fs/xfs/xfs_buf.c > > > +++ b/fs/xfs/xfs_buf.c > > > @@ -590,8 +590,20 @@ xfs_buf_get_map( > > > return NULL; > > > } > > > > > > + /* > > > + * If the buffer found doesn't match the one allocated above, somebody > > > + * else beat us to insertion and we can toss the new one. > > > + * > > > + * If we did add the buffer and it happens to be readahead, add to the > > > + * LRU now rather than waiting until the hold is released. Otherwise, > > > + * the buffer is not visible to xfs_wait_buftarg() while in flight and > > > + * nothing else prevents an unmount before I/O completion. > > > + */ > > > if (bp != new_bp) > > > xfs_buf_free(new_bp); > > > + else if (flags & XBF_READ_AHEAD && > > > + list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) > > > + atomic_inc(&bp->b_hold); > > > > This doesn't sit right with me. The LRU is for "unused" objects, and > > readahead objects are not unused until IO completes and nobody is > > waiting on them. > > > > Sure, but don't buffers remain on the LRU once they are reused? Yes, but that's once the contents have been used. > My > impression is that this patch really doesn't change behavior in that > regard. Not really - but it makes the readahead buffers completely different to the rest of the buffers on the LRU in that they don't contain valid data. For the same reason we also don't add buffers at creation time because they don't hold valid data until after the code that has created/read in the data has released it..... > IOWs, the buffer is going to end up on the LRU either way and > either end up disposed if it is never actually accessed or "used" > otherwise and the hold count is bumped. Further, isn't the hold count > mechanism precisely to handle "used" buffers on the LRU? ... hence we only add the buffer to the LRU once the initial hold count reaches zero and the owner has indicated that the buffer should be moved to the LRU by setting b_lru_ref. Once we've added the buffer to the LRU, we add a reference count that persists untile b_lru_ref is decremented to zero. At that point, we remove the euffer from the LRU and release it. That (b_lru_ref == 0) then triggers the buffer to be freed.... Note also that buffers that contain stale data (i.e. XBF_STALE) are not added to the LRU and xfs_buf_stale() removes buffers from the LRU, so shortcutting the above reclaim lifecycle in the case where a user invalidates the contents of the buffer. The end result is the same - there are no buffers with stale contents on the LRU. readahead is a special case - there is no accessor to say "cache this buffer for N references", but we have to keep it around for some time so that readahead is effective. We don't want to add it before we've done IO on it, and realistically we should only add it to the LRU if there was no IO error. We've had to work around bugs introduced by caching failed readahead buffers on the LRU in the past.... > > As it is, this points out another problem with readahead buffers - > > they aren't actually being cached properly because b_lru_ref == 0, > > which means they are immediately reclaimed on IO completion rather > > than being added to the LRU.... > > > > Not following... _xfs_buf_alloc() sets b_lru_ref to 1 and I don't see it > set/decremented unless we stale or dispose it. What am I missing? Sorry, I misread the cscope output I was looking at to quickly remind me of how it all worked - xfs_buf_stale() sets it to zero, not _xfs_buf_alloc(). > > I also think that it's not sufficient to cover the generic case of > > async IO that has no waiter. i.e. we could do get_buf, submit async > > write, drop submitter reference, and now we have the same problem > > but on a write. i.e. this problem is and async IO issue, not a > > readahead issue. > > > > Indeed. I thought about making the current patch check for ASYNC, but > opted to try and make it more isolated (I suppose ASYNC|READAHEAD would > have been moreso). > > The other thought I had was to change where buffers are added to the LRU > altogether, but didn't want to jump right to that. Is there any issue > with populating the LRU with initially held buffers as such, or any > particular reason LRU addition was deferred to I/O completion in the > first place? Yes, because then we have to deal with buffers that fail memory allocation, read IO or are allocated just to invalidate a range of disk blocks during a transaction (e.g. removing a remote attribute). There are probably other cases where we don't want to put buffers we allocate onto the LRU, but I can't think of any more right now. > > I think that it might be better to fix it by doing this: > > > > 1. ensure async IO submission always has b_lru_ref set, and > > if it isn't, set it to 1. This ensures the buffer will be > > added to the LRU on completion if it isn't already there. > > > > See above. I'm not following why this is necessary (if it is, it seems > indicative of a bug). > > > 2. keep a count of async buffer IO in progress. A per-cpu > > counter in the buftarg will be fine for this. Increment in > > xfs_buf_submit(), decrement in the xfs_buf_rele() call from > > xfs_buf_iodone() once we've determined if the buffer needs > > adding to the LRU or not. > > > > 3. make xfs_wait_buftarg() wait until the async IO count > > goes to zero before it gives up trying to release buffers on > > the LRU. > > > > This is along the lines of what I was thinking wrt to a list (explicit > I/O tracking one way or another). The count is more simple and does > cover the arbitrary read/write case, but is still more code and seems > slightly duplicative to me because in most cases the LRU wait already > handles this. Sure, but adding a list means you need global locks and when you have lots of concurrent readahead going on that will be a problem. readahead is supposed to have minimal overhead, so anything that adds submission serialisation is not the best idea... The counter has almost no overhead for the async IO case, read or write, and that's all we need to allow the IO to complete safely during unmount. > I'm also not totally sure we need to handle the write > case. Are we susceptible to this issue anywhere that isn't already > protected by the sb write protection mechanism, for example? We can do metadata buffer writes outside the superblock write protection context. A good example of that is log recovery on read-only filesystems.... > OTOH, explicit tracking might provide a more generic way to deal with > uncached buffers as opposed to lock cycling. Well, only if those uncached buffers are using async IO.... Cheers, Dave.
On Fri, Jul 01, 2016 at 02:33:31PM +1000, Dave Chinner wrote: > On Thu, Jun 30, 2016 at 07:56:21PM -0400, Brian Foster wrote: > > On Fri, Jul 01, 2016 at 08:44:57AM +1000, Dave Chinner wrote: > > > On Thu, Jun 30, 2016 at 08:53:49AM -0400, Brian Foster wrote: > > > > Newly allocated XFS metadata buffers are added to the LRU once the hold > > > > count is released, which typically occurs after I/O completion. There is > > > > no other mechanism at current that tracks the existence or I/O state of > > > > a new buffer. Further, readahead I/O tends to be submitted > > > > asynchronously by nature, which means the I/O can remain in flight and > > > > actually complete long after the calling context is gone. This means > > > > that file descriptors or any other holds on the filesystem can be > > > > released, allowing the filesystem to be unmounted while I/O is still in > > > > flight. When I/O completion occurs, core data structures may have been > > > > freed, causing completion to run into invalid memory accesses and likely > > > > to panic. > > > > > > > > This problem is reproduced on XFS via directory readahead. A filesystem > > > > is mounted, a directory is opened/closed and the filesystem immediately > > > > unmounted. The open/close cycle triggers a directory readahead that if > > > > delayed long enough, runs buffer I/O completion after the unmount has > > > > completed. > > > > > > > > To work around this problem, add readahead buffers to the LRU earlier > > > > than other buffers (when the buffer is allocated, specifically). The > > > > buffer hold count will ultimately remain until I/O completion, which > > > > means any shrinker activity will skip the buffer until then. This makes > > > > the buffer visible to xfs_wait_buftarg(), however, which ensures that an > > > > unmount or quiesce waits for I/O completion appropriately. > > > > > > > > Signed-off-by: Brian Foster <bfoster@redhat.com> > > > > --- > > > > > > > > This addresses the problem reproduced by the recently posted xfstests > > > > test: > > > > > > > > http://thread.gmane.org/gmane.comp.file-systems.fstests/2740 > > > > > > > > This could probably be made more involved, i.e., to create another list > > > > of buffers in flight or some such. This seems more simple/sane to me, > > > > however, and survives my testing so far... > > > > > > > > Brian > > > > > > > > fs/xfs/xfs_buf.c | 12 ++++++++++++ > > > > 1 file changed, 12 insertions(+) > > > > > > > > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c > > > > index 4665ff6..3f03df9 100644 > > > > --- a/fs/xfs/xfs_buf.c > > > > +++ b/fs/xfs/xfs_buf.c > > > > @@ -590,8 +590,20 @@ xfs_buf_get_map( > > > > return NULL; > > > > } > > > > > > > > + /* > > > > + * If the buffer found doesn't match the one allocated above, somebody > > > > + * else beat us to insertion and we can toss the new one. > > > > + * > > > > + * If we did add the buffer and it happens to be readahead, add to the > > > > + * LRU now rather than waiting until the hold is released. Otherwise, > > > > + * the buffer is not visible to xfs_wait_buftarg() while in flight and > > > > + * nothing else prevents an unmount before I/O completion. > > > > + */ > > > > if (bp != new_bp) > > > > xfs_buf_free(new_bp); > > > > + else if (flags & XBF_READ_AHEAD && > > > > + list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) > > > > + atomic_inc(&bp->b_hold); > > > > > > This doesn't sit right with me. The LRU is for "unused" objects, and > > > readahead objects are not unused until IO completes and nobody is > > > waiting on them. > > > > > > > Sure, but don't buffers remain on the LRU once they are reused? > > Yes, but that's once the contents have been used. > > > My > > impression is that this patch really doesn't change behavior in that > > regard. > > Not really - but it makes the readahead buffers completely different > to the rest of the buffers on the LRU in that they don't contain > valid data. > Yes, this does specially handle readahead buffers in terms of implementation, which is unfortunate... > For the same reason we also don't add buffers at creation time > because they don't hold valid data until after the code that has > created/read in the data has released it..... > ... but buffers are added to the LRU at I/O completion regardless of whether they have valid data or not (i.e., even on I/O error). > > IOWs, the buffer is going to end up on the LRU either way and > > either end up disposed if it is never actually accessed or "used" > > otherwise and the hold count is bumped. Further, isn't the hold count > > mechanism precisely to handle "used" buffers on the LRU? > > ... hence we only add the buffer to the LRU once the initial hold > count reaches zero and the owner has indicated that the buffer > should be moved to the LRU by setting b_lru_ref. Once we've added > the buffer to the LRU, we add a reference count that persists untile > b_lru_ref is decremented to zero. At that point, we remove the > euffer from the LRU and release it. That (b_lru_ref == 0) then > triggers the buffer to be freed.... > > Note also that buffers that contain stale data (i.e. XBF_STALE) are > not added to the LRU and xfs_buf_stale() removes buffers from the > LRU, so shortcutting the above reclaim lifecycle in the case where a > user invalidates the contents of the buffer. The end result is the > same - there are no buffers with stale contents on the LRU. > Yes, but this seems arbitrary. The buffer may or may not be on the LRU when it is marked stale. > readahead is a special case - there is no accessor to say "cache > this buffer for N references", but we have to keep it around for > some time so that readahead is effective. We don't want to add it > before we've done IO on it, and realistically we should only add it > to the LRU if there was no IO error. We've had to work around bugs > introduced by caching failed readahead buffers on the LRU in the > past.... > Ok, that makes sense. At least, I get the sense that you have a view of the LRU that is slightly different from what the code actually does (e.g., buffers w/ I/O errors on the LRU when they shouldn't be), which is why I'm trying to reconcile some of the reasoning here. In other words, if the LRU was an unused only buffer list where used or stale buffers were pulled off and later reinserted/reprioritized (in terms of reclaim), then this would all make complete sense to me. In actuality, we have an LRU list that has buffers with I/O errors, buffers in use, buffers that might or might not be stale, etc. > > > As it is, this points out another problem with readahead buffers - > > > they aren't actually being cached properly because b_lru_ref == 0, > > > which means they are immediately reclaimed on IO completion rather > > > than being added to the LRU.... > > > > > > > Not following... _xfs_buf_alloc() sets b_lru_ref to 1 and I don't see it > > set/decremented unless we stale or dispose it. What am I missing? > > Sorry, I misread the cscope output I was looking at to quickly > remind me of how it all worked - xfs_buf_stale() sets it to zero, > not _xfs_buf_alloc(). > > > > I also think that it's not sufficient to cover the generic case of > > > async IO that has no waiter. i.e. we could do get_buf, submit async > > > write, drop submitter reference, and now we have the same problem > > > but on a write. i.e. this problem is and async IO issue, not a > > > readahead issue. > > > > > > > Indeed. I thought about making the current patch check for ASYNC, but > > opted to try and make it more isolated (I suppose ASYNC|READAHEAD would > > have been moreso). > > > > The other thought I had was to change where buffers are added to the LRU > > altogether, but didn't want to jump right to that. Is there any issue > > with populating the LRU with initially held buffers as such, or any > > particular reason LRU addition was deferred to I/O completion in the > > first place? > > Yes, because then we have to deal with buffers that fail memory > allocation, read IO or are allocated just to invalidate a range of > disk blocks during a transaction (e.g. removing a remote attribute). > There are probably other cases where we don't want to put buffers we > allocate onto the LRU, but I can't think of any more right now. > I was thinking more along the lines of insertion on I/O submission rather than allocation. I.e., similar the proposed #1 below, but to actually insert to the list such that we never lose track of the buffer. My understanding is that for an async buffer, LRU insertion is imminent at this point as it is (unless perhaps it is looked up and marked stale by somebody else before I/O completion, but I doubt that's common). > > > I think that it might be better to fix it by doing this: > > > > > > 1. ensure async IO submission always has b_lru_ref set, and > > > if it isn't, set it to 1. This ensures the buffer will be > > > added to the LRU on completion if it isn't already there. > > > > > > > See above. I'm not following why this is necessary (if it is, it seems > > indicative of a bug). > > > > > 2. keep a count of async buffer IO in progress. A per-cpu > > > counter in the buftarg will be fine for this. Increment in > > > xfs_buf_submit(), decrement in the xfs_buf_rele() call from > > > xfs_buf_iodone() once we've determined if the buffer needs > > > adding to the LRU or not. > > > > > > 3. make xfs_wait_buftarg() wait until the async IO count > > > goes to zero before it gives up trying to release buffers on > > > the LRU. > > > > > > > This is along the lines of what I was thinking wrt to a list (explicit > > I/O tracking one way or another). The count is more simple and does > > cover the arbitrary read/write case, but is still more code and seems > > slightly duplicative to me because in most cases the LRU wait already > > handles this. > > Sure, but adding a list means you need global locks and when you > have lots of concurrent readahead going on that will be a problem. > readahead is supposed to have minimal overhead, so anything that > adds submission serialisation is not the best idea... The counter > has almost no overhead for the async IO case, read or write, and > that's all we need to allow the IO to complete safely during > unmount. > Indeed, I'm not advocating a list over counters. Just pointing out that either seem potentially unnecessary or more complexity than really necessary. I guess my initial approach was to make the fix as isolated as possible because this code is hairy and it seems we have a lot of little hacks around to deal with other corner cases such as this. Surely this adds yet another one, but afaics, so does a counter mechanism. At the end of the day, the only real issue we know it fixes is this readahead corner case. I think that if we're going to go as far as adding a list/counter mechanism, we should try to fix this in a way that is more broadly useful. By that I mean fix up the existing mechanism or add something that allows us to start unwinding some of the preexisting hacks such as cycling buffer locks to ensure I/O completion (perhaps alleviating the need to hold locks across I/O in the first place), etc. Perhaps the counter approach opens the door for that, I have to think about it some more.. > > I'm also not totally sure we need to handle the write > > case. Are we susceptible to this issue anywhere that isn't already > > protected by the sb write protection mechanism, for example? > > We can do metadata buffer writes outside the superblock write > protection context. A good example of that is log recovery on > read-only filesystems.... > This waits for I/O to complete so is unaffected by this problem. Now that I take a look at that, I'm not sure we'd be able to replace that with an I/O counter wait since technically that would wait for all in-flight I/O to complete. This context (xfs_buf_delwri_submit()) is only looking to wait for the locally submitted I/O to complete. Ugh. Brian > > OTOH, explicit tracking might provide a more generic way to deal with > > uncached buffers as opposed to lock cycling. > > Well, only if those uncached buffers are using async IO.... > > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com > > _______________________________________________ > xfs mailing list > xfs@oss.sgi.com > http://oss.sgi.com/mailman/listinfo/xfs
On Fri, Jul 01, 2016 at 08:44:57AM +1000, Dave Chinner wrote: > On Thu, Jun 30, 2016 at 08:53:49AM -0400, Brian Foster wrote: > > Newly allocated XFS metadata buffers are added to the LRU once the hold > > count is released, which typically occurs after I/O completion. There is > > no other mechanism at current that tracks the existence or I/O state of > > a new buffer. Further, readahead I/O tends to be submitted > > asynchronously by nature, which means the I/O can remain in flight and > > actually complete long after the calling context is gone. This means > > that file descriptors or any other holds on the filesystem can be > > released, allowing the filesystem to be unmounted while I/O is still in > > flight. When I/O completion occurs, core data structures may have been > > freed, causing completion to run into invalid memory accesses and likely > > to panic. > > > > This problem is reproduced on XFS via directory readahead. A filesystem > > is mounted, a directory is opened/closed and the filesystem immediately > > unmounted. The open/close cycle triggers a directory readahead that if > > delayed long enough, runs buffer I/O completion after the unmount has > > completed. > > > > To work around this problem, add readahead buffers to the LRU earlier > > than other buffers (when the buffer is allocated, specifically). The > > buffer hold count will ultimately remain until I/O completion, which > > means any shrinker activity will skip the buffer until then. This makes > > the buffer visible to xfs_wait_buftarg(), however, which ensures that an > > unmount or quiesce waits for I/O completion appropriately. > > > > Signed-off-by: Brian Foster <bfoster@redhat.com> > > --- > > > > This addresses the problem reproduced by the recently posted xfstests > > test: > > > > http://thread.gmane.org/gmane.comp.file-systems.fstests/2740 > > > > This could probably be made more involved, i.e., to create another list > > of buffers in flight or some such. This seems more simple/sane to me, > > however, and survives my testing so far... > > > > Brian > > > > fs/xfs/xfs_buf.c | 12 ++++++++++++ > > 1 file changed, 12 insertions(+) > > > > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c > > index 4665ff6..3f03df9 100644 > > --- a/fs/xfs/xfs_buf.c > > +++ b/fs/xfs/xfs_buf.c > > @@ -590,8 +590,20 @@ xfs_buf_get_map( > > return NULL; > > } > > > > + /* > > + * If the buffer found doesn't match the one allocated above, somebody > > + * else beat us to insertion and we can toss the new one. > > + * > > + * If we did add the buffer and it happens to be readahead, add to the > > + * LRU now rather than waiting until the hold is released. Otherwise, > > + * the buffer is not visible to xfs_wait_buftarg() while in flight and > > + * nothing else prevents an unmount before I/O completion. > > + */ > > if (bp != new_bp) > > xfs_buf_free(new_bp); > > + else if (flags & XBF_READ_AHEAD && > > + list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) > > + atomic_inc(&bp->b_hold); > > This doesn't sit right with me. The LRU is for "unused" objects, and > readahead objects are not unused until IO completes and nobody is > waiting on them. > > As it is, this points out another problem with readahead buffers - > they aren't actually being cached properly because b_lru_ref == 0, > which means they are immediately reclaimed on IO completion rather > than being added to the LRU.... > > I also think that it's not sufficient to cover the generic case of > async IO that has no waiter. i.e. we could do get_buf, submit async > write, drop submitter reference, and now we have the same problem > but on a write. i.e. this problem is and async IO issue, not a > readahead issue. > > I think that it might be better to fix it by doing this: > > 1. ensure async IO submission always has b_lru_ref set, and > if it isn't, set it to 1. This ensures the buffer will be > added to the LRU on completion if it isn't already there. > > 2. keep a count of async buffer IO in progress. A per-cpu > counter in the buftarg will be fine for this. Increment in > xfs_buf_submit(), decrement in the xfs_buf_rele() call from > xfs_buf_iodone() once we've determined if the buffer needs > adding to the LRU or not. > > 3. make xfs_wait_buftarg() wait until the async IO count > goes to zero before it gives up trying to release buffers on > the LRU. > After playing with this a bit this afternoon, I don't think it is so straightforward to maintain consistency between xfs_buf_submit() and xfs_buf_rele(). Some buffers are actually never released (superblock, log buffers). Other buffers can actually be submitted for I/O multiple times before they are ultimately released (e.g., log recovery buffer read -> delwri submission). I have a semi-functional patch that holds more of a pure I/O count, which means the count is decremented immediately in xfs_buf_ioend() rather than deferred to release. One downside is that while this technically still resolves the original problem, it's racy in that the count is dropped before the buffer is added to the LRU. This still works for the original problem because we also drain the ioend workqueue in xfs_wait_buftarg(), but it's not correct because we allow for non-deferred completion in the event of I/O errors (i.e., xfs_buf_ioend() called directly from xfs_buf_submit()). Brian > That will ensure readahead buffers are cached, and we capture both > async read and async write buffers in xfs_wait_buftarg(). > > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com > > _______________________________________________ > xfs mailing list > xfs@oss.sgi.com > http://oss.sgi.com/mailman/listinfo/xfs
On Fri, Jul 01, 2016 at 08:53:03AM -0400, Brian Foster wrote: > On Fri, Jul 01, 2016 at 02:33:31PM +1000, Dave Chinner wrote: > > On Thu, Jun 30, 2016 at 07:56:21PM -0400, Brian Foster wrote: > > Note also that buffers that contain stale data (i.e. XBF_STALE) are > > not added to the LRU and xfs_buf_stale() removes buffers from the > > LRU, so shortcutting the above reclaim lifecycle in the case where a > > user invalidates the contents of the buffer. The end result is the > > same - there are no buffers with stale contents on the LRU. > > > > Yes, but this seems arbitrary. The buffer may or may not be on the LRU > when it is marked stale. Yes, because (as I've already explained) we can mark buffers stale without needing to fill them with data. But whether or not the buffer is in the LRU when we mark it stale is irrelevant, because we only have a requirement that stale buffers are removed from the LRU so that they are torn down immediately on release. This is actually very important because this is how we prevent duplicate buffers from entering the cache when we free metadata. Once we free a metadata buffer and finish committing the transaction, the buffer must not be kept in the cache because the extent it covers could be reallocated at any time, and in a different shape. It could even be user data. In all cases, the buffer needs to be removed from the cache immediately, and that's why xfs_buf_stale() behaves like this. > > readahead is a special case - there is no accessor to say "cache > > this buffer for N references", but we have to keep it around for > > some time so that readahead is effective. We don't want to add it > > before we've done IO on it, and realistically we should only add it > > to the LRU if there was no IO error. We've had to work around bugs > > introduced by caching failed readahead buffers on the LRU in the > > past.... > > > > Ok, that makes sense. At least, I get the sense that you have a view of > the LRU that is slightly different from what the code actually does > (e.g., buffers w/ I/O errors on the LRU when they shouldn't be), which > is why I'm trying to reconcile some of the reasoning here. > > In other words, if the LRU was an unused only buffer list where used or > stale buffers were pulled off and later reinserted/reprioritized (in > terms of reclaim), then this would all make complete sense to me. In > actuality, we have an LRU list that has buffers with I/O errors, buffers > in use, buffers that might or might not be stale, etc. Nothing is as simple as it seems - clean design never remains that way once reality intrudes. However, the general rule is that buffers are not added to the LRU until the initial user has finished and release them. THis is done for many reasons. Firstly, We do not add them a lookup time because we do not know if the caller wants them to be added to the LRU. Secondly, we also don't add at lookup time because buffer lookup is the single hottest path in the XFS codebase. i.e. xfs_buf_find() and the struct xfs_buf are one of the few places where we've actually optimised the code to minimise cache footprint and CPU overhead. Currently, all of th einformation that a buffer lookup accesses is in the first cacheline of the struct xfs_buf, ensuring that we only take one cache miss penalty per buffer we walk through the rb-tree. If we now add a check to see if the buffer is on the LRU into the lookup path, we are - at minimum - adding an extra cacheline miss to the hot path. If we have to add the buffer to the LRU, then we're adding another couple of cacheline misses and another lock, all of which will have an impact on performance. THird, we add/remove buffers to/from the LRU lazily in paths that are not performance critical as a result. This is exactly the same algorithms the VFS uses for the inode and dentry cache LRUs, and we also use this for the XFS dquot LRU, too. The b_lru_ref count replaces the single referenced flag that the inode/dentry caches use so we can prioritise reclaim to match typical btree and directory access patterns (i.e. reclaim data before leaf before node before root) but otherwise the buffer LRU is no different to the VFS cache algorithms.... Finally, and perhaps the most important, is behaviour under low memory conditions. If we add the buffer to the LRU when we first look it up in the cache, the shrinker can now see it and decrement the b_lru_ref count while we are doing the first operation on that buffer. Say this transaction generates sufficient memory pressure that it causes the shrinker to completely decrement b_lru_ref - it will then remove the buffer from the LRU *before we've even finished with it*. This will cause the buffer to be freed immediately after the caller releases it (because b_lru_ref is zero), and so the next access requires the buffer to be read from disk again. What we end up with here is a situation where the buffer cache cannot maintian a working set of buffers between operations because the memory pressure within an operation causes the buffers to be removed from the LRU before the operation completes. So, yeah, there's lots of "assumed knowledge" in why the buffer cache LRU is structured the way it is. That, unfortunately, is pretty much typical of all the linux slab caches that are controlled by shrinkers.... FWIW, readahead is often not complete before the followup read comes in, which means there's a caller blocked on the buffer with an active reference by the time IO completion runs xfs_buf_relse() to unlock the buffer. In these cases, the readahead buffer does not go onto the LRU until after the blocked caller has finished with the buffer and called xfs_buf_rele() itself (i.e. after the transaction subsystem is done with it). > > > The other thought I had was to change where buffers are added > > > to the LRU altogether, but didn't want to jump right to that. > > > Is there any issue with populating the LRU with initially held > > > buffers as such, or any particular reason LRU addition was > > > deferred to I/O completion in the first place? > > > > Yes, because then we have to deal with buffers that fail memory > > allocation, read IO or are allocated just to invalidate a range > > of disk blocks during a transaction (e.g. removing a remote > > attribute). There are probably other cases where we don't want > > to put buffers we allocate onto the LRU, but I can't think of > > any more right now. > > I was thinking more along the lines of insertion on I/O submission > rather than allocation. I.e., similar the proposed #1 below, but > to actually insert to the list such that we never lose track of > the buffer. Why do we need a list? It's a similar situation to inode_dio_wait(): we don't track direct IO because of the per-IO overhead of doing so, especially as we only need to implement a "drain barrier". i.e. it only needs a counter to implement the barrier needed for truncate/hole punch operations. That's exactly the same case here xfs_wait_buftarg() is a drain barrier; we could simply count every single IO in flight and wait for that to return to zero before walking the LRU, and it would solve the problem for everything, regardless of the type of buffer or whether it is in or going to be put in the LRU.... Also, remember that xfs_wait_buftarg() is run with the assumption that all high level processing has been stopped, and we are only waiting for remaining users to drop their locks and go away (e.g. complete async IO from AIL pushing). It's not a general use case we are working on here... > I guess my initial approach was to make the fix as isolated as possible > because this code is hairy and it seems we have a lot of little hacks > around to deal with other corner cases such as this. Surely this adds > yet another one, but afaics, so does a counter mechanism. At the end of > the day, the only real issue we know it fixes is this readahead corner > case. > > I think that if we're going to go as far as adding a list/counter > mechanism, we should try to fix this in a way that is more broadly > useful. By that I mean fix up the existing mechanism or add something > that allows us to start unwinding some of the preexisting hacks such as > cycling buffer locks to ensure I/O completion (perhaps alleviating the > need to hold locks across I/O in the first place), etc. > > Perhaps the counter approach opens the door for that, I have to think > about it some more.. I don't think we can get rid of lock cycling - the counter can only exist within the IO context which is within the lock context. Hence to guarantee that all users of a buffer have finished, lock cycling is still necessary. i.e. lock cycling is a life cycle barrier, not an IO barrier... > > We can do metadata buffer writes outside the superblock write > > protection context. A good example of that is log recovery on > > read-only filesystems.... > > > > This waits for I/O to complete so is unaffected by this problem. Sure, but that's not what you were asking about. My point was that there is no specific high level superblock protection against async metadata IO being issued from non-VFS interfacing, internal filesystem subsystems. > Now > that I take a look at that, I'm not sure we'd be able to replace that > with an I/O counter wait since technically that would wait for all > in-flight I/O to complete. This context (xfs_buf_delwri_submit()) is > only looking to wait for the locally submitted I/O to complete. Ugh. I'm not advocating doing that at all. Cheers, Dave.
On Mon, Jul 04, 2016 at 10:08:38AM +1000, Dave Chinner wrote: > On Fri, Jul 01, 2016 at 08:53:03AM -0400, Brian Foster wrote: > > On Fri, Jul 01, 2016 at 02:33:31PM +1000, Dave Chinner wrote: > > > On Thu, Jun 30, 2016 at 07:56:21PM -0400, Brian Foster wrote: > > > Note also that buffers that contain stale data (i.e. XBF_STALE) are > > > not added to the LRU and xfs_buf_stale() removes buffers from the > > > LRU, so shortcutting the above reclaim lifecycle in the case where a > > > user invalidates the contents of the buffer. The end result is the > > > same - there are no buffers with stale contents on the LRU. > > > > > > > Yes, but this seems arbitrary. The buffer may or may not be on the LRU > > when it is marked stale. > > Yes, because (as I've already explained) we can mark buffers stale > without needing to fill them with data. But whether or not the > buffer is in the LRU when we mark it stale is irrelevant, because we > only have a requirement that stale buffers are removed from the LRU > so that they are torn down immediately on release. > > This is actually very important because this is how we prevent > duplicate buffers from entering the cache when we free metadata. > Once we free a metadata buffer and finish committing the transaction, > the buffer must not be kept in the cache because the extent it > covers could be reallocated at any time, and in a different shape. > It could even be user data. In all cases, the buffer needs to be > removed from the cache immediately, and that's why xfs_buf_stale() > behaves like this. > Yes, I understand. I think we're talking past eachother on this particular train of thought. Your original point was that the LRU is for "unused" buffers, and/or buffers aren't on the LRU until "their content is used." The example given was that stale buffers are immediately removed from the LRU and thus "buffers with stale data are never on the LRU." My response to that is that's an implementation detail of XBF_STALE buffers and not necessarily true in other cases. We do actually end up with buffers with "stale" (as opposed to XBF_STALE, perhaps "invalid" is a better term) data in the event of I/O error. If a callsite allocates a buffer, never issues I/O on it (i.e., it is !XBF_DONE) and releases it, it ends up on the LRU. (I don't know that we do that anywhere given this is a filesystem after all ;), but that's the natural lifecycle in that particular case afaict). So nothing about the LRU suggests to me that the contents must have valid data. AFAICT, it's primarily memory management mechanism to allow priority-ordered releasing of "unused" buffers, where an unused buffer is defined as a buffer without any active references, regardless of what the buffer contains. I'm sympathetic to the argument that in the ideal case the LRU might not contain buffers actively in use, but IMO that's beyond the scope of the original patch. IOW, if we want to change that, we should approach it separately. > > > readahead is a special case - there is no accessor to say "cache > > > this buffer for N references", but we have to keep it around for > > > some time so that readahead is effective. We don't want to add it > > > before we've done IO on it, and realistically we should only add it > > > to the LRU if there was no IO error. We've had to work around bugs > > > introduced by caching failed readahead buffers on the LRU in the > > > past.... > > > > > > > Ok, that makes sense. At least, I get the sense that you have a view of > > the LRU that is slightly different from what the code actually does > > (e.g., buffers w/ I/O errors on the LRU when they shouldn't be), which > > is why I'm trying to reconcile some of the reasoning here. > > > > In other words, if the LRU was an unused only buffer list where used or > > stale buffers were pulled off and later reinserted/reprioritized (in > > terms of reclaim), then this would all make complete sense to me. In > > actuality, we have an LRU list that has buffers with I/O errors, buffers > > in use, buffers that might or might not be stale, etc. > > Nothing is as simple as it seems - clean design never remains that > way once reality intrudes. However, the general rule is that buffers are not added to the LRU until the initial > user has finished and release them. THis is done for many reasons. > Agreed on the first point. :) I think that's the only reason for the original patch to exist. ;) (As noted previously, I agree that it is a hack). > Firstly, We do not add them a lookup time because we do not know if > the caller wants them to be added to the LRU. > We set b_lru_ref to 1 on buffer allocation. But it's true that we don't know if it's going to be marked stale before it is released, of course. > Secondly, we also don't add at lookup time because buffer lookup is > the single hottest path in the XFS codebase. i.e. xfs_buf_find() > and the struct xfs_buf are one of the few places where we've > actually optimised the code to minimise cache footprint and CPU > overhead. > Ok... > Currently, all of th einformation that a buffer lookup > accesses is in the first cacheline of the struct xfs_buf, ensuring > that we only take one cache miss penalty per buffer we walk through > the rb-tree. If we now add a check to see if the buffer is on the > LRU into the lookup path, we are - at minimum - adding an extra > cacheline miss to the hot path. If we have to add the buffer to the > LRU, then we're adding another couple of cacheline misses and > another lock, all of which will have an impact on performance. > Performance is a fair point against adding all buffers to the LRU immediately. I take that to mean we'd have to test such a change with that in mind, but it sounds like this path is already optimized without any more room in the first cacheline, thus probably not likely to be an option. I'm not sure this is as as critical applied solely to read ahead buffers, however... > THird, we add/remove buffers to/from the LRU lazily in paths that > are not performance critical as a result. This is exactly the same > algorithms the VFS uses for the inode and dentry cache LRUs, and we > also use this for the XFS dquot LRU, too. The b_lru_ref count > replaces the single referenced flag that the inode/dentry caches use > so we can prioritise reclaim to match typical btree and directory > access patterns (i.e. reclaim data before leaf before node before > root) but otherwise the buffer LRU is no different to the VFS cache > algorithms.... > > Finally, and perhaps the most important, is behaviour under low memory > conditions. If we add the buffer to the LRU when we first look it up > in the cache, the shrinker can now see it and decrement the > b_lru_ref count while we are doing the first operation on that > buffer. Say this transaction generates sufficient memory pressure > that it causes the shrinker to completely decrement b_lru_ref - it > will then remove the buffer from the LRU *before we've even finished > with it*. This will cause the buffer to be freed immediately after > the caller releases it (because b_lru_ref is zero), and so the next > access requires the buffer to be read from disk again. What we end > up with here is a situation where the buffer cache cannot maintian a > working set of buffers between operations because the memory > pressure within an operation causes the buffers to be removed from > the LRU before the operation completes. > This makes sense, we probably don't want to subject active buffers to deprioritization via the shrinker before their initial use. It sounds like this could technically be a flaw for buffers once they're actually reused, as well. I could see an argument for reused buffers remaining on the LRU such that their b_lru_ref remains consistent (in terms of priority) with associated buffers (e.g., btree node vs. leaf priority ordering) in the event of shrinker activity, but it seems like there are other ways for that to get out of sync just the same. It could also be the case that such buffers tend to be "held" together. Again, I'm not sure how important this is to readahead buffers since they are released on I/O completion. If we're waiting on the readahead before it completes then the buffer is not fully released, but that's a matter of timing more than anything. E.g., faster storage could disrupt that just as much as this particular change. Otherwise, this just means the readahead buffer could be freed before we try to use it, which is already the case today. > So, yeah, there's lots of "assumed knowledge" in why the buffer > cache LRU is structured the way it is. That, unfortunately, is > pretty much typical of all the linux slab caches that are controlled > by shrinkers.... > > FWIW, readahead is often not complete before the followup read > comes in, which means there's a caller blocked on the buffer with an > active reference by the time IO completion runs xfs_buf_relse() to > unlock the buffer. In these cases, the readahead buffer does not go > onto the LRU until after the blocked caller has finished with the > buffer and called xfs_buf_rele() itself (i.e. after the transaction > subsystem is done with it). > > > > > The other thought I had was to change where buffers are added > > > > to the LRU altogether, but didn't want to jump right to that. > > > > Is there any issue with populating the LRU with initially held > > > > buffers as such, or any particular reason LRU addition was > > > > deferred to I/O completion in the first place? > > > > > > Yes, because then we have to deal with buffers that fail memory > > > allocation, read IO or are allocated just to invalidate a range > > > of disk blocks during a transaction (e.g. removing a remote > > > attribute). There are probably other cases where we don't want > > > to put buffers we allocate onto the LRU, but I can't think of > > > any more right now. > > > > I was thinking more along the lines of insertion on I/O submission > > rather than allocation. I.e., similar the proposed #1 below, but > > to actually insert to the list such that we never lose track of > > the buffer. > > Why do we need a list? > I was referring to the LRU. > It's a similar situation to inode_dio_wait(): we don't track direct > IO because of the per-IO overhead of doing so, especially as we only > need to implement a "drain barrier". i.e. it only needs a counter to > implement the barrier needed for truncate/hole punch operations. > That's exactly the same case here xfs_wait_buftarg() is a drain > barrier; we could simply count every single IO in flight and wait > for that to return to zero before walking the LRU, and it would > solve the problem for everything, regardless of the type of buffer > or whether it is in or going to be put in the LRU.... > > Also, remember that xfs_wait_buftarg() is run with the assumption > that all high level processing has been stopped, and we are only > waiting for remaining users to drop their locks and go away (e.g. > complete async IO from AIL pushing). It's not a general use case we > are working on here... > > > I guess my initial approach was to make the fix as isolated as possible > > because this code is hairy and it seems we have a lot of little hacks > > around to deal with other corner cases such as this. Surely this adds > > yet another one, but afaics, so does a counter mechanism. At the end of > > the day, the only real issue we know it fixes is this readahead corner > > case. > > > > I think that if we're going to go as far as adding a list/counter > > mechanism, we should try to fix this in a way that is more broadly > > useful. By that I mean fix up the existing mechanism or add something > > that allows us to start unwinding some of the preexisting hacks such as > > cycling buffer locks to ensure I/O completion (perhaps alleviating the > > need to hold locks across I/O in the first place), etc. > > > > Perhaps the counter approach opens the door for that, I have to think > > about it some more.. > > I don't think we can get rid of lock cycling - the counter can only > exist within the IO context which is within the lock context. Hence > to guarantee that all users of a buffer have finished, lock cycling > is still necessary. i.e. lock cycling is a life cycle barrier, not > an IO barrier... > Ok, but we apparently use it as such in places like xfs_buf_delwri_submit(). > > > We can do metadata buffer writes outside the superblock write > > > protection context. A good example of that is log recovery on > > > read-only filesystems.... > > > > > > > This waits for I/O to complete so is unaffected by this problem. > > Sure, but that's not what you were asking about. My point was that > there is no specific high level superblock protection against > async metadata IO being issued from non-VFS interfacing, internal > filesystem subsystems. > Ok, I wasn't being clear... the sb protection was just an example. The fundamental question was whether we have to handle the write case due to not otherwise having some kind of synchronization or protection against tearing down infrastructure before I/Os complete. IOW, I don't care as much about the sb write protection specifically as whether we can identify any places where this is an actual, real problem that needs to be solved. (Again, I'm not against stepping back and taking a larger look at the current buffer mechanisms to try and resolve such problems in theory, but I would prefer to do that separately from this fix). > > Now > > that I take a look at that, I'm not sure we'd be able to replace that > > with an I/O counter wait since technically that would wait for all > > in-flight I/O to complete. This context (xfs_buf_delwri_submit()) is > > only looking to wait for the locally submitted I/O to complete. Ugh. > > I'm not advocating doing that at all. > I know, I was. I was hoping that if we go down the path of a more generic mechanism we might come up with something more useful beyond this particular readahead problem. Brian > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com > > _______________________________________________ > xfs mailing list > xfs@oss.sgi.com > http://oss.sgi.com/mailman/listinfo/xfs
diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c index 4665ff6..3f03df9 100644 --- a/fs/xfs/xfs_buf.c +++ b/fs/xfs/xfs_buf.c @@ -590,8 +590,20 @@ xfs_buf_get_map( return NULL; } + /* + * If the buffer found doesn't match the one allocated above, somebody + * else beat us to insertion and we can toss the new one. + * + * If we did add the buffer and it happens to be readahead, add to the + * LRU now rather than waiting until the hold is released. Otherwise, + * the buffer is not visible to xfs_wait_buftarg() while in flight and + * nothing else prevents an unmount before I/O completion. + */ if (bp != new_bp) xfs_buf_free(new_bp); + else if (flags & XBF_READ_AHEAD && + list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) + atomic_inc(&bp->b_hold); found: if (!bp->b_addr) {
Newly allocated XFS metadata buffers are added to the LRU once the hold count is released, which typically occurs after I/O completion. There is no other mechanism at current that tracks the existence or I/O state of a new buffer. Further, readahead I/O tends to be submitted asynchronously by nature, which means the I/O can remain in flight and actually complete long after the calling context is gone. This means that file descriptors or any other holds on the filesystem can be released, allowing the filesystem to be unmounted while I/O is still in flight. When I/O completion occurs, core data structures may have been freed, causing completion to run into invalid memory accesses and likely to panic. This problem is reproduced on XFS via directory readahead. A filesystem is mounted, a directory is opened/closed and the filesystem immediately unmounted. The open/close cycle triggers a directory readahead that if delayed long enough, runs buffer I/O completion after the unmount has completed. To work around this problem, add readahead buffers to the LRU earlier than other buffers (when the buffer is allocated, specifically). The buffer hold count will ultimately remain until I/O completion, which means any shrinker activity will skip the buffer until then. This makes the buffer visible to xfs_wait_buftarg(), however, which ensures that an unmount or quiesce waits for I/O completion appropriately. Signed-off-by: Brian Foster <bfoster@redhat.com> --- This addresses the problem reproduced by the recently posted xfstests test: http://thread.gmane.org/gmane.comp.file-systems.fstests/2740 This could probably be made more involved, i.e., to create another list of buffers in flight or some such. This seems more simple/sane to me, however, and survives my testing so far... Brian fs/xfs/xfs_buf.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)