From daa7527afc2274432094ebe7ceb03aa41f916607 Mon Sep 17 00:00:00 2001
From: Robert Haas <rhaas@postgresql.org>
Date: Wed, 8 Jan 2014 18:49:14 -0500
Subject: [PATCH] Reduce the number of semaphores used under
--disable-spinlocks.
Instead of allocating a semaphore from the operating system for every
spinlock, allocate a fixed number of semaphores (by default, 1024)
from the operating system and multiplex all the spinlocks that get
created onto them. This could self-deadlock if a process attempted
to acquire more than one spinlock at a time, but since processes
aren't supposed to execute anything other than short stretches of
straight-line code while holding a spinlock, that shouldn't happen.
One motivation for this change is that, with the introduction of
dynamic shared memory, it may be desirable to create spinlocks that
last for less than the lifetime of the server. Without this change,
attempting to use such facilities under --disable-spinlocks would
quickly exhaust any supply of available semaphores. Quite apart
from that, it's desirable to contain the quantity of semaphores
needed to run the server simply on convenience grounds, since using
too many may make it harder to get PostgreSQL running on a new
platform, which is mostly the point of --disable-spinlocks in the
first place.
Patch by me; review by Tom Lane.
---
src/backend/postmaster/postmaster.c | 9 ++++++
src/backend/storage/ipc/ipci.c | 1 +
src/backend/storage/ipc/shmem.c | 21 ++++++++++--
src/backend/storage/lmgr/spin.c | 50 ++++++++++++++++++-----------
src/include/pg_config_manual.h | 8 +++++
src/include/storage/s_lock.h | 2 +-
src/include/storage/spin.h | 6 ++++
7 files changed, 74 insertions(+), 23 deletions(-)
diff --git a/src/backend/postmaster/postmaster.c b/src/backend/postmaster/postmaster.c
index 377fa591a8b..a7e40cc4a59 100644
--- a/src/backend/postmaster/postmaster.c
+++ b/src/backend/postmaster/postmaster.c
@@ -471,6 +471,9 @@ typedef struct
slock_t *ShmemLock;
VariableCache ShmemVariableCache;
Backend *ShmemBackendArray;
+#ifndef HAVE_SPINLOCKS
+ PGSemaphore SpinlockSemaArray;
+#endif
LWLock *LWLockArray;
slock_t *ProcStructLock;
PROC_HDR *ProcGlobal;
@@ -5626,6 +5629,9 @@ save_backend_variables(BackendParameters *param, Port *port,
param->ShmemVariableCache = ShmemVariableCache;
param->ShmemBackendArray = ShmemBackendArray;
+#ifndef HAVE_SPINLOCKS
+ param->SpinlockSemaArray = SpinlockSemaArray;
+#endif
param->LWLockArray = LWLockArray;
param->ProcStructLock = ProcStructLock;
param->ProcGlobal = ProcGlobal;
@@ -5854,6 +5860,9 @@ restore_backend_variables(BackendParameters *param, Port *port)
ShmemVariableCache = param->ShmemVariableCache;
ShmemBackendArray = param->ShmemBackendArray;
+#ifndef HAVE_SPINLOCKS
+ SpinlockSemaArray = param->SpinlockSemaArray;
+#endif
LWLockArray = param->LWLockArray;
ProcStructLock = param->ProcStructLock;
ProcGlobal = param->ProcGlobal;
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index 1c8783f2c08..3c04fc31bae 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -105,6 +105,7 @@ CreateSharedMemoryAndSemaphores(bool makePrivate, int port)
* need to be so careful during the actual allocation phase.
*/
size = 100000;
+ size = add_size(size, SpinlockSemaSize());
size = add_size(size, hash_estimate_size(SHMEM_INDEX_SIZE,
sizeof(ShmemIndexEnt)));
size = add_size(size, BufferShmemSize());
diff --git a/src/backend/storage/ipc/shmem.c b/src/backend/storage/ipc/shmem.c
index e93b988ad74..70b02ca8384 100644
--- a/src/backend/storage/ipc/shmem.c
+++ b/src/backend/storage/ipc/shmem.c
@@ -116,9 +116,24 @@ InitShmemAllocation(void)
Assert(shmhdr != NULL);
/*
- * Initialize the spinlock used by ShmemAlloc. We have to do the space
- * allocation the hard way, since obviously ShmemAlloc can't be called
- * yet.
+ * If spinlocks are disabled, initialize emulation layer. We have to do
+ * the space allocation the hard way, since obviously ShmemAlloc can't be
+ * called yet.
+ */
+#ifndef HAVE_SPINLOCKS
+ {
+ PGSemaphore spinsemas;
+
+ spinsemas = (PGSemaphore) (((char *) shmhdr) + shmhdr->freeoffset);
+ shmhdr->freeoffset += MAXALIGN(SpinlockSemaSize());
+ SpinlockSemaInit(spinsemas);
+ Assert(shmhdr->freeoffset <= shmhdr->totalsize);
+ }
+#endif
+
+ /*
+ * Initialize the spinlock used by ShmemAlloc; we have to do this the hard
+ * way, too, for the same reasons as above.
*/
ShmemLock = (slock_t *) (((char *) shmhdr) + shmhdr->freeoffset);
shmhdr->freeoffset += MAXALIGN(sizeof(slock_t));
diff --git a/src/backend/storage/lmgr/spin.c b/src/backend/storage/lmgr/spin.c
index 08782178de9..3d116bc7a2e 100644
--- a/src/backend/storage/lmgr/spin.c
+++ b/src/backend/storage/lmgr/spin.c
@@ -29,6 +29,18 @@
#include "storage/spin.h"
+PGSemaphore SpinlockSemaArray;
+
+/*
+ * Report the amount of shared memory needed to store semaphores for spinlock
+ * support.
+ */
+Size
+SpinlockSemaSize(void)
+{
+ return SpinlockSemas() * sizeof(PGSemaphoreData);
+}
+
#ifdef HAVE_SPINLOCKS
/*
@@ -52,22 +64,20 @@ SpinlockSemas(void)
int
SpinlockSemas(void)
{
- int nsemas;
-
- /*
- * It would be cleaner to distribute this logic into the affected modules,
- * similar to the way shmem space estimation is handled.
- *
- * For now, though, there are few enough users of spinlocks that we just
- * keep the knowledge here.
- */
- nsemas = NumLWLocks(); /* one for each lwlock */
- nsemas += NBuffers; /* one for each buffer header */
- nsemas += max_wal_senders; /* one for each wal sender process */
- nsemas += num_xloginsert_slots; /* one for each WAL insertion slot */
- nsemas += 30; /* plus a bunch for other small-scale use */
-
- return nsemas;
+ return NUM_SPINLOCK_SEMAPHORES;
+}
+
+/*
+ * Initialize semaphores.
+ */
+extern void
+SpinlockSemaInit(PGSemaphore spinsemas)
+{
+ int i;
+
+ for (i = 0; i < NUM_SPINLOCK_SEMAPHORES; ++i)
+ PGSemaphoreCreate(&spinsemas[i]);
+ SpinlockSemaArray = spinsemas;
}
/*
@@ -77,13 +87,15 @@ SpinlockSemas(void)
void
s_init_lock_sema(volatile slock_t *lock)
{
- PGSemaphoreCreate((PGSemaphore) lock);
+ static int counter = 0;
+
+ *lock = (++counter) % NUM_SPINLOCK_SEMAPHORES;
}
void
s_unlock_sema(volatile slock_t *lock)
{
- PGSemaphoreUnlock((PGSemaphore) lock);
+ PGSemaphoreUnlock(&SpinlockSemaArray[*lock]);
}
bool
@@ -98,7 +110,7 @@ int
tas_sema(volatile slock_t *lock)
{
/* Note that TAS macros return 0 if *success* */
- return !PGSemaphoreTryLock((PGSemaphore) lock);
+ return !PGSemaphoreTryLock(&SpinlockSemaArray[*lock]);
}
#endif /* !HAVE_SPINLOCKS */
diff --git a/src/include/pg_config_manual.h b/src/include/pg_config_manual.h
index 2387a434aa4..20c5ff0e902 100644
--- a/src/include/pg_config_manual.h
+++ b/src/include/pg_config_manual.h
@@ -56,6 +56,14 @@
*/
#define NUM_USER_DEFINED_LWLOCKS 4
+/*
+ * When we don't have native spinlocks, we use semaphores to simulate them.
+ * Decreasing this value reduces consumption of OS resources; increasing it
+ * may improve performance, but supplying a real spinlock implementation is
+ * probably far better.
+ */
+#define NUM_SPINLOCK_SEMAPHORES 1024
+
/*
* Define this if you want to allow the lo_import and lo_export SQL
* functions to be executed by ordinary users. By default these
diff --git a/src/include/storage/s_lock.h b/src/include/storage/s_lock.h
index b52f0e7f852..2297f772805 100644
--- a/src/include/storage/s_lock.h
+++ b/src/include/storage/s_lock.h
@@ -915,7 +915,7 @@ spin_delay(void)
* to fall foul of kernel limits on number of semaphores, so don't use this
* unless you must! The subroutines appear in spin.c.
*/
-typedef PGSemaphoreData slock_t;
+typedef int slock_t;
extern bool s_lock_free_sema(volatile slock_t *lock);
extern void s_unlock_sema(volatile slock_t *lock);
diff --git a/src/include/storage/spin.h b/src/include/storage/spin.h
index e7201652499..2ac510db7a0 100644
--- a/src/include/storage/spin.h
+++ b/src/include/storage/spin.h
@@ -69,5 +69,11 @@
extern int SpinlockSemas(void);
+extern Size SpinlockSemaSize(void);
+
+#ifndef HAVE_SPINLOCKS
+extern void SpinlockSemaInit(PGSemaphore);
+extern PGSemaphore SpinlockSemaArray;
+#endif
#endif /* SPIN_H */
--
GitLab