mirror/postgresql
2
0
Fork 0
mirror of https://git.postgresql.org/git/postgresql.git synced 2025-01-24 18:55:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add infrastructure for making spins_per_delay variable depending on

Browse Source 
whether we seem to be running in a uniprocessor or multiprocessor.
The adjustment rules could probably still use further tweaking, but
I'm convinced this should be a win overall.
This commit is contained in:
Tom Lane 2005-10-11 20:41:32 +00:00
parent 9907b9775b
commit a72ee09090
4 changed files with 145 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
src/backend/storage/lmgr/proc.c
View File

@ -8,7 +8,7 @@
*
*
* IDENTIFICATION
* $PostgreSQL: pgsql/src/backend/storage/lmgr/proc.c,v 1.164 2005/09/19 17:21:47 momjian Exp $
* $PostgreSQL: pgsql/src/backend/storage/lmgr/proc.c,v 1.165 2005/10/11 20:41:32 tgl Exp $
*
*-------------------------------------------------------------------------
*/
@ -171,6 +171,8 @@ InitProcGlobal(void)
ProcGlobal->freeProcs = INVALID_OFFSET;
ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
/*
* Pre-create the PGPROC structures and create a semaphore for
* each.
@ -225,9 +227,14 @@ InitProcess(void)
/*
* Try to get a proc struct from the free list. If this fails, we
* must be out of PGPROC structures (not to mention semaphores).
*
* While we are holding the ProcStructLock, also copy the current
* shared estimate of spins_per_delay to local storage.
*/
SpinLockAcquire(ProcStructLock);
set_spins_per_delay(procglobal->spins_per_delay);
myOffset = procglobal->freeProcs;
if (myOffset != INVALID_OFFSET)
@ -319,21 +326,38 @@ InitDummyProcess(int proctype)
Assert(proctype >= 0 && proctype < NUM_DUMMY_PROCS);
/*
* Just for paranoia's sake, we use the ProcStructLock to protect
* assignment and releasing of DummyProcs entries.
*
* While we are holding the ProcStructLock, also copy the current
* shared estimate of spins_per_delay to local storage.
*/
SpinLockAcquire(ProcStructLock);
set_spins_per_delay(ProcGlobal->spins_per_delay);
dummyproc = &DummyProcs[proctype];
/*
* dummyproc should not presently be in use by anyone else
*/
if (dummyproc->pid != 0)
{
SpinLockRelease(ProcStructLock);
elog(FATAL, "DummyProc[%d] is in use by PID %d",
proctype, dummyproc->pid);
}
MyProc = dummyproc;
MyProc->pid = MyProcPid; /* marks dummy proc as in use by me */
SpinLockRelease(ProcStructLock);
/*
* Initialize all fields of MyProc, except MyProc->sem which was set
* up by InitProcGlobal.
*/
MyProc->pid = MyProcPid; /* marks dummy proc as in use by me */
SHMQueueElemInit(&(MyProc->links));
MyProc->waitStatus = STATUS_OK;
MyProc->xid = InvalidTransactionId;
@ -510,6 +534,9 @@ ProcKill(int code, Datum arg)
/* PGPROC struct isn't mine anymore */
MyProc = NULL;
/* Update shared estimate of spins_per_delay */
procglobal->spins_per_delay = update_spins_per_delay(procglobal->spins_per_delay);
SpinLockRelease(ProcStructLock);
}
@ -533,11 +560,18 @@ DummyProcKill(int code, Datum arg)
/* Release any LW locks I am holding (see notes above) */
LWLockReleaseAll();
SpinLockAcquire(ProcStructLock);
/* Mark dummy proc no longer in use */
MyProc->pid = 0;
/* PGPROC struct isn't mine anymore */
MyProc = NULL;
/* Update shared estimate of spins_per_delay */
ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
SpinLockRelease(ProcStructLock);
}

119
src/backend/storage/lmgr/s_lock.c
View File

@ -9,7 +9,7 @@
*
*
* IDENTIFICATION
* $PostgreSQL: pgsql/src/backend/storage/lmgr/s_lock.c,v 1.38 2005/08/26 14:47:35 tgl Exp $
* $PostgreSQL: pgsql/src/backend/storage/lmgr/s_lock.c,v 1.39 2005/10/11 20:41:32 tgl Exp $
*
*-------------------------------------------------------------------------
*/
@ -21,6 +21,10 @@
#include "storage/s_lock.h"
#include "miscadmin.h"
static int spins_per_delay = DEFAULT_SPINS_PER_DELAY;
/*
* s_lock_stuck() - complain about a stuck spinlock
*/
@ -49,54 +53,67 @@ s_lock(volatile slock_t *lock, const char *file, int line)
* We loop tightly for awhile, then delay using pg_usleep() and try
* again. Preferably, "awhile" should be a small multiple of the
* maximum time we expect a spinlock to be held. 100 iterations seems
* about right. In most multi-CPU scenarios, the spinlock is probably
* held by a process on another CPU and will be released before we
* finish 100 iterations. However, on a uniprocessor, the tight loop
* is just a waste of cycles, so don't iterate thousands of times.
* about right as an initial guess. However, on a uniprocessor the
* loop is a waste of cycles, while in a multi-CPU scenario it's usually
* better to spin a bit longer than to call the kernel, so we try to
* adapt the spin loop count depending on whether we seem to be in
* a uniprocessor or multiprocessor.
*
* Note: you might think MIN_SPINS_PER_DELAY should be just 1, but you'd
* be wrong; there are platforms where that can result in a "stuck
* spinlock" failure. This has been seen particularly on Alphas; it
* seems that the first TAS after returning from kernel space will always
* fail on that hardware.
*
* Once we do decide to block, we use randomly increasing pg_usleep()
* delays. The first delay is 10 msec, then the delay randomly
* increases to about one second, after which we reset to 10 msec and
* delays. The first delay is 1 msec, then the delay randomly
* increases to about one second, after which we reset to 1 msec and
* start again. The idea here is that in the presence of heavy
* contention we need to increase the delay, else the spinlock holder
* may never get to run and release the lock. (Consider situation
* where spinlock holder has been nice'd down in priority by the
* scheduler --- it will not get scheduled until all would-be
* acquirers are sleeping, so if we always use a 10-msec sleep, there
* acquirers are sleeping, so if we always use a 1-msec sleep, there
* is a real possibility of starvation.) But we can't just clamp the
* delay to an upper bound, else it would take a long time to make a
* reasonable number of tries.
*
* We time out and declare error after NUM_DELAYS delays (thus, exactly
* that many tries). With the given settings, this will usually take
* 3 or so minutes. It seems better to fix the total number of tries
* 2 or so minutes. It seems better to fix the total number of tries
* (and thus the probability of unintended failure) than to fix the
* total time spent.
*
* The pg_usleep() delays are measured in centiseconds (0.01 sec) because
* 10 msec is a common resolution limit at the OS level.
* The pg_usleep() delays are measured in milliseconds because 1 msec
* is a common resolution limit at the OS level for newer platforms.
* On older platforms the resolution limit is usually 10 msec, in
* which case the total delay before timeout will be a bit more.
*/
#define SPINS_PER_DELAY 100
#define MIN_SPINS_PER_DELAY 10
#define MAX_SPINS_PER_DELAY 1000
#define NUM_DELAYS 1000
#define MIN_DELAY_CSEC 1
#define MAX_DELAY_CSEC 100
#define MIN_DELAY_MSEC 1
#define MAX_DELAY_MSEC 1000
int spins = 0;
int delays = 0;
int cur_delay = MIN_DELAY_CSEC;
int cur_delay = 0;
while (TAS(lock))
{
/* CPU-specific delay each time through the loop */
SPIN_DELAY();
/* Block the process every SPINS_PER_DELAY tries */
if (++spins > SPINS_PER_DELAY)
/* Block the process every spins_per_delay tries */
if (++spins >= spins_per_delay)
{
if (++delays > NUM_DELAYS)
s_lock_stuck(lock, file, line);
pg_usleep(cur_delay * 10000L);
if (cur_delay == 0) /* first time to delay? */
cur_delay = MIN_DELAY_MSEC;
pg_usleep(cur_delay * 1000L);
#if defined(S_LOCK_TEST)
fprintf(stdout, "*");
@ -107,14 +124,76 @@ s_lock(volatile slock_t *lock, const char *file, int line)
cur_delay += (int) (cur_delay *
(((double) random()) / ((double) MAX_RANDOM_VALUE)) + 0.5);
/* wrap back to minimum delay when max is exceeded */
if (cur_delay > MAX_DELAY_CSEC)
cur_delay = MIN_DELAY_CSEC;
if (cur_delay > MAX_DELAY_MSEC)
cur_delay = MIN_DELAY_MSEC;
spins = 0;
}
}
/*
* If we were able to acquire the lock without delaying, it's a good
* indication we are in a multiprocessor. If we had to delay, it's
* a sign (but not a sure thing) that we are in a uniprocessor.
* Hence, we decrement spins_per_delay slowly when we had to delay,
* and increase it rapidly when we didn't. It's expected that
* spins_per_delay will converge to the minimum value on a uniprocessor
* and to the maximum value on a multiprocessor.
*
* Note: spins_per_delay is local within our current process.
* We want to average these observations across multiple backends,
* since it's relatively rare for this function to even get entered,
* and so a single backend might not live long enough to converge on
* a good value. That is handled by the two routines below.
*/
if (cur_delay == 0)
{
/* we never had to delay */
if (spins_per_delay < MAX_SPINS_PER_DELAY)
spins_per_delay = Min(spins_per_delay + 100, MAX_SPINS_PER_DELAY);
}
else
{
if (spins_per_delay > MIN_SPINS_PER_DELAY)
spins_per_delay = Max(spins_per_delay - 1, MIN_SPINS_PER_DELAY);
}
}
/*
* Set local copy of spins_per_delay during backend startup.
*
* NB: this has to be pretty fast as it is called while holding a spinlock
*/
void
set_spins_per_delay(int shared_spins_per_delay)
{
spins_per_delay = shared_spins_per_delay;
}
/*
* Update shared estimate of spins_per_delay during backend exit.
*
* NB: this has to be pretty fast as it is called while holding a spinlock
*/
int
update_spins_per_delay(int shared_spins_per_delay)
{
/*
* We use an exponential moving average with a relatively slow
* adaption rate, so that noise in any one backend's result won't
* affect the shared value too much. As long as both inputs are
* within the allowed range, the result must be too, so we need not
* worry about clamping the result.
*
* We deliberately truncate rather than rounding; this is so that
* single adjustments inside a backend can affect the shared estimate
* (see the asymmetric adjustment rules above).
*/
return (shared_spins_per_delay * 15 + spins_per_delay) / 16;
}
/*
* Various TAS implementations that cannot live in s_lock.h as no inline
* definition exists (yet).

4
src/include/storage/proc.h
View File

@ -7,7 +7,7 @@
* Portions Copyright (c) 1996-2005, PostgreSQL Global Development Group
* Portions Copyright (c) 1994, Regents of the University of California
*
* $PostgreSQL: pgsql/src/include/storage/proc.h,v 1.82 2005/09/19 17:21:48 momjian Exp $
* $PostgreSQL: pgsql/src/include/storage/proc.h,v 1.83 2005/10/11 20:41:32 tgl Exp $
*
*-------------------------------------------------------------------------
*/
@ -105,6 +105,8 @@ typedef struct PROC_HDR
{
/* Head of list of free PGPROC structures */
SHMEM_OFFSET freeProcs;
/* Current shared estimate of appropriate spins_per_delay value */
int spins_per_delay;
} PROC_HDR;

8
src/include/storage/s_lock.h
View File

@ -66,7 +66,7 @@
* Portions Copyright (c) 1996-2005, PostgreSQL Global Development Group
* Portions Copyright (c) 1994, Regents of the University of California
*
* $PostgreSQL: pgsql/src/include/storage/s_lock.h,v 1.141 2005/10/11 20:01:30 tgl Exp $
* $PostgreSQL: pgsql/src/include/storage/s_lock.h,v 1.142 2005/10/11 20:41:32 tgl Exp $
*
*-------------------------------------------------------------------------
*/
@ -872,4 +872,10 @@ extern int tas(volatile slock_t *lock); /* in port/.../tas.s, or
*/
extern void s_lock(volatile slock_t *lock, const char *file, int line);
/* Support for dynamic adjustment of spins_per_delay */
#define DEFAULT_SPINS_PER_DELAY 100
extern void set_spins_per_delay(int shared_spins_per_delay);
extern int update_spins_per_delay(int shared_spins_per_delay);
#endif /* S_LOCK_H */