summaryrefslogtreecommitdiff
path: root/fs/pipe.c
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2021-08-05 20:04:43 +0300
committerLinus Torvalds <torvalds@linux-foundation.org>2021-08-18 21:39:46 +0300
commit3b844826b6c6affa80755254da322b017358a2f4 (patch)
tree65c3e589debdb07b539012bf4d5c2e567b85f8cf /fs/pipe.c
parent614cb2751d3150850d459bee596c397f344a7936 (diff)
downloadlinux-3b844826b6c6affa80755254da322b017358a2f4.tar.xz
pipe: avoid unnecessary EPOLLET wakeups under normal loads
I had forgotten just how sensitive hackbench is to extra pipe wakeups, and commit 3a34b13a88ca ("pipe: make pipe writes always wake up readers") ended up causing a quite noticeable regression on larger machines. Now, hackbench isn't necessarily a hugely meaningful benchmark, and it's not clear that this matters in real life all that much, but as Mel points out, it's used often enough when comparing kernels and so the performance regression shows up like a sore thumb. It's easy enough to fix at least for the common cases where pipes are used purely for data transfer, and you never have any exciting poll usage at all. So set a special 'poll_usage' flag when there is polling activity, and make the ugly "EPOLLET has crazy legacy expectations" semantics explicit to only that case. I would love to limit it to just the broken EPOLLET case, but the pipe code can't see the difference between epoll and regular select/poll, so any non-read/write waiting will trigger the extra wakeup behavior. That is sufficient for at least the hackbench case. Apart from making the odd extra wakeup cases more explicitly about EPOLLET, this also makes the extra wakeup be at the _end_ of the pipe write, not at the first write chunk. That is actually much saner semantics (as much as you can call any of the legacy edge-triggered expectations for EPOLLET "sane") since it means that you know the wakeup will happen once the write is done, rather than possibly in the middle of one. [ For stable people: I'm putting a "Fixes" tag on this, but I leave it up to you to decide whether you actually want to backport it or not. It likely has no impact outside of synthetic benchmarks - Linus ] Link: https://lore.kernel.org/lkml/20210802024945.GA8372@xsang-OptiPlex-9020/ Fixes: 3a34b13a88ca ("pipe: make pipe writes always wake up readers") Reported-by: kernel test robot <oliver.sang@intel.com> Tested-by: Sandeep Patil <sspatil@android.com> Tested-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'fs/pipe.c')
-rw-r--r--fs/pipe.c15
1 files changed, 9 insertions, 6 deletions
diff --git a/fs/pipe.c b/fs/pipe.c
index 8e6ef62aeb1c..678dee2a8228 100644
--- a/fs/pipe.c
+++ b/fs/pipe.c
@@ -444,9 +444,6 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
#endif
/*
- * Epoll nonsensically wants a wakeup whether the pipe
- * was already empty or not.
- *
* If it wasn't empty we try to merge new data into
* the last buffer.
*
@@ -455,9 +452,9 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
* spanning multiple pages.
*/
head = pipe->head;
- was_empty = true;
+ was_empty = pipe_empty(head, pipe->tail);
chars = total_len & (PAGE_SIZE-1);
- if (chars && !pipe_empty(head, pipe->tail)) {
+ if (chars && !was_empty) {
unsigned int mask = pipe->ring_size - 1;
struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
int offset = buf->offset + buf->len;
@@ -590,8 +587,11 @@ out:
* This is particularly important for small writes, because of
* how (for example) the GNU make jobserver uses small writes to
* wake up pending jobs
+ *
+ * Epoll nonsensically wants a wakeup whether the pipe
+ * was already empty or not.
*/
- if (was_empty) {
+ if (was_empty || pipe->poll_usage) {
wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
}
@@ -654,6 +654,9 @@ pipe_poll(struct file *filp, poll_table *wait)
struct pipe_inode_info *pipe = filp->private_data;
unsigned int head, tail;
+ /* Epoll has some historical nasty semantics, this enables them */
+ pipe->poll_usage = 1;
+
/*
* Reading pipe state only -- no need for acquiring the semaphore.
*