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. -- Dave Chinner david@xxxxxxxxxxxxx _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
