1
1
openmpi/oshmem/mca/sshmem/sysv/sshmem_sysv_module.c
Alex Mikheev 692021f637
oshmem: sshmem sysv: auto huge page alloc can fallback to regular pages.
Fallback to the regular pages if huge page allocation is set to auto
and it was not possible to allocate requested amount of memory with
the hugepages.

Signed-off-by: Alex Mikheev <alexm@mellanox.com>
2017-08-06 13:33:04 +03:00

330 строки
9.5 KiB
C

/*
* Copyright (c) 2014 Mellanox Technologies, Inc.
* All rights reserved.
* Copyright (c) 2014 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "oshmem_config.h"
#include <errno.h>
#ifdef HAVE_FCNTL_H
#include <fcntl.h>
#endif /* HAVE_FCNTL_H */
#ifdef HAVE_SYS_MMAN_H
#include <sys/mman.h>
#endif /* HAVE_SYS_MMAN_H */
#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif /* HAVE_UNISTD_H */
#ifdef HAVE_SYS_TYPES_H
#include <sys/types.h>
#endif /* HAVE_SYS_TYPES_H */
#ifdef HAVE_SYS_IPC_H
#include <sys/ipc.h>
#endif /* HAVE_SYS_IPC_H */
#if HAVE_SYS_SHM_H
#include <sys/shm.h>
#endif /* HAVE_SYS_SHM_H */
#if HAVE_SYS_STAT_H
#include <sys/stat.h>
#endif /* HAVE_SYS_STAT_H */
#include <string.h>
#ifdef HAVE_NETDB_H
#include <netdb.h>
#endif /* HAVE_NETDB_H */
#include "opal/constants.h"
#include "opal_stdint.h"
#include "opal/util/output.h"
#include "opal/util/path.h"
#include "opal/util/show_help.h"
#include "orte/util/show_help.h"
#include "oshmem/proc/proc.h"
#include "oshmem/mca/sshmem/sshmem.h"
#include "oshmem/mca/sshmem/base/base.h"
#include "sshmem_sysv.h"
/* ////////////////////////////////////////////////////////////////////////// */
/* local functions */
static int
module_init(void);
static int
segment_create(map_segment_t *ds_buf,
const char *file_name,
size_t size);
static void *
segment_attach(map_segment_t *ds_buf, sshmem_mkey_t *mkey);
static int
segment_detach(map_segment_t *ds_buf, sshmem_mkey_t *mkey);
static int
segment_unlink(map_segment_t *ds_buf);
static int
module_finalize(void);
/* sysv shmem module */
mca_sshmem_sysv_module_t mca_sshmem_sysv_module = {
/* super */
{
module_init,
segment_create,
segment_attach,
segment_detach,
segment_unlink,
module_finalize
}
};
/* ////////////////////////////////////////////////////////////////////////// */
static int
module_init(void)
{
/* nothing to do */
return OSHMEM_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////// */
static int
module_finalize(void)
{
/* nothing to do */
return OSHMEM_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////// */
static int
segment_create(map_segment_t *ds_buf,
const char *file_name,
size_t size)
{
int rc = OSHMEM_SUCCESS;
void *addr = NULL;
int shmid = MAP_SEGMENT_SHM_INVALID;
int flags;
int try_hp;
assert(ds_buf);
/* init the contents of map_segment_t */
shmem_ds_reset(ds_buf);
/* for sysv shared memory we don't have to worry about the backing store
* being located on a network file system... so no check is needed here.
*/
/* create a new shared memory segment and save the shmid. note the use of
* real_size here
*/
flags = IPC_CREAT | IPC_EXCL | S_IRUSR | S_IWUSR;
try_hp = mca_sshmem_sysv_component.use_hp;
#if defined (SHM_HUGETLB)
flags |= ((0 != try_hp) ? SHM_HUGETLB : 0);
size = ((size + sshmem_sysv_gethugepagesize() - 1) / sshmem_sysv_gethugepagesize()) * sshmem_sysv_gethugepagesize();
#endif
/* Create a new shared memory segment and save the shmid. */
retry_alloc:
shmid = shmget(IPC_PRIVATE, size, flags);
if (shmid == MAP_SEGMENT_SHM_INVALID) {
/* hugepage alloc was set to auto. Hopefully it failed because there are no
* enough hugepages on the system. Turn it off and retry.
*/
if (-1 == try_hp) {
OPAL_OUTPUT_VERBOSE(
(10, oshmem_sshmem_base_framework.framework_output,
"failed to allocate %llu bytes with huge pages. "
"Using regular pages", (unsigned long long)size));
flags = IPC_CREAT | IPC_EXCL | S_IRUSR | S_IWUSR;
try_hp = 0;
goto retry_alloc;
}
opal_show_help("help-oshmem-sshmem.txt",
"create segment failure",
true,
"sysv",
orte_process_info.nodename, (unsigned long long) size,
strerror(errno), errno);
opal_show_help("help-oshmem-sshmem-sysv.txt",
"sysv:create segment failure",
true);
return OSHMEM_ERROR;
}
/* Attach to the segment */
addr = shmat(shmid, (void *) mca_sshmem_base_start_address, 0);
if (addr == (void *) -1L) {
opal_show_help("help-oshmem-sshmem.txt",
"create segment failure",
true,
"sysv",
orte_process_info.nodename, (unsigned long long) size,
strerror(errno), errno);
opal_show_help("help-oshmem-sshmem-sysv.txt",
"sysv:create segment failure",
true);
shmctl(shmid, IPC_RMID, NULL);
return OSHMEM_ERR_OUT_OF_RESOURCE;
}
shmctl(shmid, IPC_RMID, NULL );
ds_buf->type = MAP_SEGMENT_ALLOC_SHM;
ds_buf->seg_id = shmid;
ds_buf->super.va_base = addr;
ds_buf->seg_size = size;
ds_buf->super.va_end = (void*)((uintptr_t)ds_buf->super.va_base + ds_buf->seg_size);
OPAL_OUTPUT_VERBOSE(
(70, oshmem_sshmem_base_framework.framework_output,
"%s: %s: create %s "
"(id: %d, addr: %p size: %lu)\n",
mca_sshmem_sysv_component.super.base_version.mca_type_name,
mca_sshmem_sysv_component.super.base_version.mca_component_name,
(rc ? "failure" : "successful"),
ds_buf->seg_id, ds_buf->super.va_base, (unsigned long)ds_buf->seg_size)
);
return rc;
}
/* ////////////////////////////////////////////////////////////////////////// */
/**
* segment_attach can only be called after a successful call to segment_create
*/
static void *
segment_attach(map_segment_t *ds_buf, sshmem_mkey_t *mkey)
{
assert(ds_buf);
assert(mkey->va_base == 0);
if (MAP_SEGMENT_SHM_INVALID == (int)(mkey->u.key)) {
return (mkey->va_base);
}
mkey->va_base = shmat((int)(mkey->u.key), 0, 0);
OPAL_OUTPUT_VERBOSE(
(70, oshmem_sshmem_base_framework.framework_output,
"%s: %s: attach successful "
"(id: %d, addr: %p size: %lu | va_base: 0x%p len: %d key %llx)\n",
mca_sshmem_sysv_component.super.base_version.mca_type_name,
mca_sshmem_sysv_component.super.base_version.mca_component_name,
ds_buf->seg_id, ds_buf->super.va_base, (unsigned long)ds_buf->seg_size,
mkey->va_base, mkey->len, (unsigned long long)mkey->u.key)
);
/* update returned base pointer with an offset that hides our stuff */
return (mkey->va_base);
}
/* ////////////////////////////////////////////////////////////////////////// */
static int
segment_detach(map_segment_t *ds_buf, sshmem_mkey_t *mkey)
{
int rc = OSHMEM_SUCCESS;
assert(ds_buf);
OPAL_OUTPUT_VERBOSE(
(70, oshmem_sshmem_base_framework.framework_output,
"%s: %s: detaching "
"(id: %d, addr: %p size: %lu)\n",
mca_sshmem_sysv_component.super.base_version.mca_type_name,
mca_sshmem_sysv_component.super.base_version.mca_component_name,
ds_buf->seg_id, ds_buf->super.va_base, (unsigned long)ds_buf->seg_size)
);
if (ds_buf->seg_id != MAP_SEGMENT_SHM_INVALID) {
shmctl(ds_buf->seg_id, IPC_RMID, NULL );
}
if (mca_sshmem_sysv_component.use_hp != 0) {
/**
* Workaround kernel panic when detaching huge pages from user space simultanously from several processes
* dont detach here instead let kernel do it during process cleanup
*/
/* shmdt((void *)ds_buf->seg_base_addr); */
}
/* reset the contents of the map_segment_t associated with this
* shared memory segment.
*/
shmem_ds_reset(ds_buf);
return rc;
}
/* ////////////////////////////////////////////////////////////////////////// */
static int
segment_unlink(map_segment_t *ds_buf)
{
/* not much unlink work needed for sysv */
OPAL_OUTPUT_VERBOSE(
(70, oshmem_sshmem_base_framework.framework_output,
"%s: %s: unlinking "
"(id: %d, size: %lu)\n",
mca_sshmem_sysv_component.super.base_version.mca_type_name,
mca_sshmem_sysv_component.super.base_version.mca_component_name,
ds_buf->seg_id, (unsigned long)ds_buf->seg_size)
);
/* don't completely reset. in particular, only reset
* the id and flip the invalid bit. size and name values will remain valid
* across unlinks. other information stored in flags will remain untouched.
*/
ds_buf->seg_id = MAP_SEGMENT_SHM_INVALID;
/* note: this is only changing the valid bit to 0. */
MAP_SEGMENT_INVALIDATE(ds_buf);
return OSHMEM_SUCCESS;
}
/*
* Get current huge page size
*
*/
size_t sshmem_sysv_gethugepagesize(void)
{
static size_t huge_page_size = 0;
char buf[256];
int size_kb;
FILE *f;
/* Cache the huge page size value */
if (huge_page_size == 0) {
f = fopen("/proc/meminfo", "r");
if (f != NULL) {
while (fgets(buf, sizeof(buf), f)) {
if (sscanf(buf, "Hugepagesize: %d kB", &size_kb) == 1) {
huge_page_size = size_kb * 1024L;
break;
}
}
fclose(f);
}
if (huge_page_size == 0) {
huge_page_size = 2 * 1024L *1024L;
}
}
return huge_page_size;
}