talexander
/
freebsd_amp_hwpstate
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

geom_io: Shift to pause_sbt to eliminate bogus min and update comment. 

Update to eliminate bogus min to ensure 0 was never passed to
pause. Instead, requrest 1ms with an 'infinite' precision, which
defaults to whatever the underlying time counter can do. This should
ensure we run fairly quickly to start processing done events, while
still giving a small pause for the system to catch its breath. This rate
limiter still is less than ideal, and this commit doesn't change
that. It should really have no functional change: it just uses a better
interface to express the desired sleep.

Sponsored by:		Netflix
Reviewed by:		kib
Differential Revision:	https://reviews.freebsd.org/D45316
This commit is contained in:
Warner Losh 2024-05-24 08:31:55 -06:00
parent 32f40fc983
commit 6d83b38186
1 changed files with 12 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
sys/geom/geom_io.c
View File

@ -809,20 +809,17 @@ g_io_schedule_down(struct thread *tp __unused)
biotrack(bp, __func__);
if (pace != 0) {
/*
* There has been at least one memory allocation
* failure since the last I/O completed. Pause 1ms to
* give the system a chance to free up memory. We only
* do this once because a large number of allocations
* can fail in the direct dispatch case and there's no
* relationship between the number of these failures and
* the length of the outage. If there's still an outage,
* we'll pause again and again until it's
* resolved. Older versions paused longer and once per
* allocation failure. This was OK for a single threaded
* g_down, but with direct dispatch would lead to max of
* 10 IOPs for minutes at a time when transient memory
* issues prevented allocation for a batch of requests
* from the upper layers.
* There has been at least one memory allocation failure
* since the last I/O completed. Pause 1ms to give the
* system a chance to free up memory. Pause time is not
* scaled to the number of I/O failures since they tend
* to cluster and the number is not predictive of how
* long a pause is needed.
*
* Older versions had a longer pause, which limited the
* IOPS to 10, which prolonged memory shortages that could
* be alleviated by I/O completing since it eliminated
* direct dispatch as well.
*
* XXX This pacing is really lame. It needs to be solved
* by other methods. This is OK only because the worst
@ -833,7 +830,7 @@ g_io_schedule_down(struct thread *tp __unused)
*/
CTR0(KTR_GEOM, "g_down pacing self");
pause_count++; /* g_down has only one thread */
pause("g_down", min(hz/1000, 1));
pause_sbt("g_down", SBT_1MS, 0, 0);
pace = 0;
}
CTR2(KTR_GEOM, "g_down processing bp %p provider %s", bp,