On Thu, Mar 28, 2019 at 08:17:44AM -0700, Darrick J. Wong wrote: > On Thu, Mar 28, 2019 at 10:10:10AM -0400, Brian Foster wrote: > > On Tue, Mar 26, 2019 at 08:06:34PM -0700, Darrick J. Wong wrote: > > > From: Darrick J. Wong <darrick.wong@xxxxxxxxxx> > > > > > > When we're processing an ioend on the list of io completions, check to > > > see if the next items on the list are both adjacent and of the same > > > type. If so, we can merge the completions to reduce transaction > > > overhead. > > > > > > Signed-off-by: Darrick J. Wong <darrick.wong@xxxxxxxxxx> > > > --- > > > > I'm curious of the value of this one... what situations allow for > > batching on the ioend completion side that we haven't already accounted > > for in the ioend construction side? > > I was skeptical too, but Dave (I think?) pointed out that writeback can > split into 1GB chunks so it actually is possible to end up with adjacent > ioends. So I wrote this patch and added a tracepoint, and lo it > actually did trigger when there's a lot of data to flush out, and we > succeed at allocating a single extent for the entire delalloc reservation. > That doesn't seem like a huge overhead to me, but I am curious where that splitting logic is. Somewhere in the writeback code..? (I assume Dave can chime in on some of this stuff if he's more familiar with it..). > > The latter already batches until we > > cross a change in fork type, extent state, or a break in logical or > > physical contiguity. The former looks like it follows similar logic for > > merging with the exceptions of allowing for merges of physically > > discontiguous extents and disallowing merges of those with different > > append status. That seems like a smallish window of opportunity to me.. > > am I missing something? > > Yep, it's a smallish window; small discontiguous writes don't benefit > here at all. > Ok. The whole append thing is somewhat dynamic/non-deterministic as well, along with the fact that we're subject to completion/wq timing to allow for this kind of batching to occur. E.g., if we're completing a series of 1GB ioends, what are the odds the next ioend completes before the current one starts in the wq? I'd guess not great, but tbh I have no idea.. > > If that is the gist but there is enough benefit for the more lenient > > merging, I also wonder whether it would be more efficient to try and > > also accomplish that on the construction side rather than via completion > > post-processing. For example, could we abstract a single ioend to cover > > an arbitrary list of bio/page -> sector mappings with the same higher > > level semantics? We already have a bio chaining mechanism, it's just > > only used for when a bio is full. Could we reuse that for dealing with > > physical discontiguity? > > I suppose we could, though the bigger the ioend the longer it'll take to > process responses. Also, I think it's the case that if any of the bios > fail then we treat all of the chained ones as failed? (Meh, it's > writeback, it's not like you get to know /which/ writes failed unless > you do a stupid write()/fsync() dance...) > This already seems to be the case to a large degree with our current batching. That 1GB ioend above is already a fairly large bio chain, as I think that a bio can't have more than 256 (BIO_MAX_PAGES) pages. Eh, I guess my take is that this doesn't necessarily seem like an unreasonable change and it's not like it's a huge amount of code, but it does seem like potentially more standalone code than it's worth for minimal benefit. It seems the amount of code and processing could be reduced and benefit slightly increased by allocating fewer ioends in the first place if this were pre-processed vs. post-processed. I'd prefer to see more analysis of the potential benefits either way... > The other thing is that directio completions look very similar to > writeback completions, including the potential for having the thundering > herds pounding on the ILOCK. I was thinking about refactoring those to > use the per-inode queue as a next step, though the directio completion > paths are murky. > Indeed, though we don't have the degree of submission batching taking place for direct I/O. Perhaps it just depends more on the workload since aio can presumably drive a deeper queue. A couple things to note.. I vaguely recall that the dio completion code has a bit of a messy history of being switched back and forth between using ioends or bitmasks or whatever else to avoid ioend allocations. This doesn't rule out potential new benefits of using ioends that might be achieved in light of new features like reflink, but FWIW I'd be more skeptical of refactors along those lines that don't come along with a measurable benefit or solve a tangible problem. Another thing to be aware of here is that the iomap code looks like it already can invoke our callback in a wq and it may complete the aio as soon as we return... Brian > > Brian > > > > > fs/xfs/xfs_aops.c | 86 +++++++++++++++++++++++++++++++++++++++++++++++++++++ > > > 1 file changed, 86 insertions(+) > > > > > > diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c > > > index f7a9bb661826..53afa2e6e3e7 100644 > > > --- a/fs/xfs/xfs_aops.c > > > +++ b/fs/xfs/xfs_aops.c > > > @@ -237,6 +237,7 @@ STATIC void > > > xfs_end_ioend( > > > struct xfs_ioend *ioend) > > > { > > > + struct list_head ioend_list; > > > struct xfs_inode *ip = XFS_I(ioend->io_inode); > > > xfs_off_t offset = ioend->io_offset; > > > size_t size = ioend->io_size; > > > @@ -273,7 +274,89 @@ xfs_end_ioend( > > > done: > > > if (ioend->io_append_trans) > > > error = xfs_setfilesize_ioend(ioend, error); > > > + list_replace_init(&ioend->io_list, &ioend_list); > > > xfs_destroy_ioend(ioend, error); > > > + > > > + while (!list_empty(&ioend_list)) { > > > + ioend = list_first_entry(&ioend_list, struct xfs_ioend, > > > + io_list); > > > + list_del_init(&ioend->io_list); > > > + xfs_destroy_ioend(ioend, error); > > > + } > > > +} > > > + > > > +/* > > > + * We can merge two adjacent ioends if they have the same set of work to do. > > > + */ > > > +static bool > > > +xfs_ioend_can_merge( > > > + struct xfs_ioend *ioend, > > > + int ioend_error, > > > + struct xfs_ioend *next) > > > +{ > > > + int next_error; > > > + > > > + next_error = blk_status_to_errno(next->io_bio->bi_status); > > > + if (ioend_error != next_error) > > > + return false; > > > + if ((ioend->io_fork == XFS_COW_FORK) ^ (next->io_fork == XFS_COW_FORK)) > > > + return false; > > > + if ((ioend->io_state == XFS_EXT_UNWRITTEN) ^ > > > + (next->io_state == XFS_EXT_UNWRITTEN)) > > > + return false; > > > + if (ioend->io_offset + ioend->io_size != next->io_offset) > > > + return false; > > > + if (xfs_ioend_is_append(ioend) != xfs_ioend_is_append(next)) > > > + return false; > > > + return true; > > > +} > > > + > > > +/* Try to merge adjacent completions. */ > > > +STATIC void > > > +xfs_ioend_try_merge( > > > + struct xfs_ioend *ioend, > > > + struct list_head *more_ioends) > > > +{ > > > + struct xfs_ioend *next_ioend; > > > + int ioend_error; > > > + int error; > > > + > > > + if (list_empty(more_ioends)) > > > + return; > > > + > > > + ioend_error = blk_status_to_errno(ioend->io_bio->bi_status); > > > + > > > + while (!list_empty(more_ioends)) { > > > + next_ioend = list_first_entry(more_ioends, struct xfs_ioend, > > > + io_list); > > > + if (!xfs_ioend_can_merge(ioend, ioend_error, next_ioend)) > > > + break; > > > + list_move_tail(&next_ioend->io_list, &ioend->io_list); > > > + ioend->io_size += next_ioend->io_size; > > > + if (ioend->io_append_trans) { > > > + error = xfs_setfilesize_ioend(next_ioend, 1); > > > + ASSERT(error == 1); > > > + } > > > + } > > > +} > > > + > > > +/* list_sort compare function for ioends */ > > > +static int > > > +xfs_ioend_compare( > > > + void *priv, > > > + struct list_head *a, > > > + struct list_head *b) > > > +{ > > > + struct xfs_ioend *ia; > > > + struct xfs_ioend *ib; > > > + > > > + ia = container_of(a, struct xfs_ioend, io_list); > > > + ib = container_of(b, struct xfs_ioend, io_list); > > > + if (ia->io_offset < ib->io_offset) > > > + return -1; > > > + else if (ia->io_offset > ib->io_offset) > > > + return 1; > > > + return 0; > > > } > > > > > > /* Finish all pending io completions. */ > > > @@ -292,10 +375,13 @@ xfs_end_io( > > > list_replace_init(&ip->i_iodone_list, &completion_list); > > > spin_unlock_irqrestore(&ip->i_iodone_lock, flags); > > > > > > + list_sort(NULL, &completion_list, xfs_ioend_compare); > > > + > > > while (!list_empty(&completion_list)) { > > > ioend = list_first_entry(&completion_list, struct xfs_ioend, > > > io_list); > > > list_del_init(&ioend->io_list); > > > + xfs_ioend_try_merge(ioend, &completion_list); > > > xfs_end_ioend(ioend); > > > } > > > }