/*
* register memory
*/
static int mca_rcache_grdma_register (mca_rcache_base_module_t *rcache, void *addr,
size_t size, uint32_t flags, int32_t access_flags,
mca_rcache_base_registration_t **reg)
{
mca_rcache_grdma_module_t *rcache_grdma = (mca_rcache_grdma_module_t*)rcache;
const bool bypass_cache = !!(flags & MCA_RCACHE_FLAGS_CACHE_BYPASS);
const bool persist = !!(flags & MCA_RCACHE_FLAGS_PERSIST);
mca_rcache_base_registration_t *grdma_reg;
opal_free_list_item_t *item;
unsigned char *base, *bound;
unsigned int page_size = opal_getpagesize ();
int rc;
*reg = NULL;
/* if cache bypass is requested don't use the cache */
base = OPAL_DOWN_ALIGN_PTR(addr, page_size, unsigned char *);
bound = OPAL_ALIGN_PTR((intptr_t) addr + size, page_size, unsigned char *) - 1;
#if OPAL_CUDA_GDR_SUPPORT
if (flags & MCA_RCACHE_FLAGS_CUDA_GPU_MEM) {
size_t psize;
mca_common_cuda_get_address_range(&base, &psize, addr);
bound = base + psize - 1;
/* Check to see if this memory is in the cache and if it has been freed. If so,
* this call will boot it out of the cache. */
check_for_cuda_freed_memory(rcache, base, psize);
}
#endif /* OPAL_CUDA_GDR_SUPPORT */
do_unregistration_gc (rcache);
/* look through existing regs if not persistent registration requested.
* Persistent registration are always registered and placed in the cache */
if (!(bypass_cache || persist)) {
mca_rcache_base_find_args_t find_args = {.reg = NULL, .rcache_grdma = rcache_grdma,
.base = base, .bound = bound,
.access_flags = access_flags};
/* check to see if memory is registered */
rc = mca_rcache_base_vma_iterate (rcache_grdma->cache->vma_module, base, size,
mca_rcache_grdma_check_cached, (void *) &find_args);
if (1 == rc) {
*reg = find_args.reg;
return OPAL_SUCCESS;
}
/* get updated access flags */
access_flags = find_args.access_flags;
OPAL_THREAD_ADD32((volatile int32_t *) &rcache_grdma->stat_cache_miss, 1);
}
void mca_mpool_grdma_finalize(struct mca_mpool_base_module_t *mpool)
{
mca_mpool_grdma_module_t *mpool_grdma = (mca_mpool_grdma_module_t*)mpool;
mca_mpool_base_registration_t *regs[GRDMA_MPOOL_NREGS];
int reg_cnt, i;
/* Statistic */
if (true == mca_mpool_grdma_component.print_stats) {
opal_output(0, "%s grdma: stats "
"(hit/miss/found/not found/evicted): %d/%d/%d/%d/%d\n",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
mpool_grdma->stat_cache_hit, mpool_grdma->stat_cache_miss,
mpool_grdma->stat_cache_found, mpool_grdma->stat_cache_notfound,
mpool_grdma->stat_evicted);
}
OPAL_THREAD_LOCK(&mpool->rcache->lock);
do_unregistration_gc(mpool);
do {
reg_cnt = mpool->rcache->rcache_find_all(mpool->rcache, 0, (size_t)-1,
regs, GRDMA_MPOOL_NREGS);
for (i = 0 ; i < reg_cnt ; ++i) {
if (regs[i]->ref_count) {
regs[i]->ref_count = 0; /* otherwise dereg will fail on assert */
} else if (mca_mpool_grdma_component.leave_pinned) {
opal_list_remove_item(&mpool_grdma->pool->lru_list,
(opal_list_item_t *) regs[i]);
}
(void) dereg_mem(regs[i]);
}
} while (reg_cnt == GRDMA_MPOOL_NREGS);
OBJ_RELEASE(mpool_grdma->pool);
OBJ_DESTRUCT(&mpool_grdma->reg_list);
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
/* Cleanup any vmas that we have deferred deletion on */
mpool->rcache->rcache_clean(mpool->rcache);
/* this mpool was allocated by grdma_init in mpool_grdma_component.c */
free(mpool);
}
/*
* register memory
*/
int mca_mpool_grdma_register(mca_mpool_base_module_t *mpool, void *addr,
size_t size, uint32_t flags,
mca_mpool_base_registration_t **reg)
{
mca_mpool_grdma_module_t *mpool_grdma = (mca_mpool_grdma_module_t*)mpool;
mca_mpool_base_registration_t *grdma_reg;
ompi_free_list_item_t *item;
unsigned char *base, *bound;
bool bypass_cache = !!(flags & MCA_MPOOL_FLAGS_CACHE_BYPASS);
int rc;
OPAL_THREAD_LOCK(&mpool->rcache->lock);
/* if cache bypass is requested don't use the cache */
base = (unsigned char *) down_align_addr(addr, mca_mpool_base_page_size_log);
bound = (unsigned char *) up_align_addr((void*)((char*) addr + size - 1),
mca_mpool_base_page_size_log);
if (!opal_list_is_empty (&mpool_grdma->pool->gc_list))
do_unregistration_gc(mpool);
/* look through existing regs if not persistent registration requested.
* Persistent registration are always registered and placed in the cache */
if(!(flags & MCA_MPOOL_FLAGS_PERSIST) && !bypass_cache) {
/* check to see if memory is registered */
mpool->rcache->rcache_find(mpool->rcache, addr, size, reg);
if(*reg != NULL &&
(mca_mpool_grdma_component.leave_pinned ||
((*reg)->flags & MCA_MPOOL_FLAGS_PERSIST) ||
((*reg)->base == base && (*reg)->bound == bound))) {
if(0 == (*reg)->ref_count &&
mca_mpool_grdma_component.leave_pinned) {
opal_list_remove_item(&mpool_grdma->pool->lru_list,
(opal_list_item_t*)(*reg));
}
mpool_grdma->stat_cache_hit++;
(*reg)->ref_count++;
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
return OMPI_SUCCESS;
}
mpool_grdma->stat_cache_miss++;
*reg = NULL; /* in case previous find found something */
/* If no suitable registration is in cache and leave_pinned isn't
* set don't use the cache.
* This is optimisation in case limit is not set. If limit is set we
* have to put registration into the cache to determine when we hit
* memory registration limit.
* NONE: cache is still used for persistent registrations so previous
* find can find something */
if(!mca_mpool_grdma_component.leave_pinned) {
bypass_cache = true;
}
}
OMPI_FREE_LIST_GET(&mpool_grdma->reg_list, item, rc);
if(OMPI_SUCCESS != rc) {
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
return rc;
}
grdma_reg = (mca_mpool_base_registration_t*)item;
grdma_reg->mpool = mpool;
grdma_reg->base = base;
grdma_reg->bound = bound;
grdma_reg->flags = flags;
if (false == bypass_cache) {
rc = mpool->rcache->rcache_insert(mpool->rcache, grdma_reg, 0);
if (OPAL_UNLIKELY(rc != OMPI_SUCCESS)) {
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
OMPI_FREE_LIST_RETURN(&mpool_grdma->reg_list, item);
return rc;
}
}
while (OMPI_ERR_OUT_OF_RESOURCE ==
(rc = mpool_grdma->resources.register_mem(mpool_grdma->resources.reg_data,
base, bound - base + 1, grdma_reg))) {
/* try to remove one unused reg and retry */
if (!mca_mpool_grdma_evict (mpool)) {
break;
}
}
if (OPAL_UNLIKELY(rc != OMPI_SUCCESS)) {
if (false == bypass_cache) {
mpool->rcache->rcache_delete(mpool->rcache, grdma_reg);
}
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
OMPI_FREE_LIST_RETURN(&mpool_grdma->reg_list, item);
return rc;
}
*reg = grdma_reg;
(*reg)->ref_count++;
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
/* Cleanup any vmas that we have deferred deletion on */
mpool->rcache->rcache_clean(mpool->rcache);
return OMPI_SUCCESS;
}
void mca_mpool_rdma_finalize(struct mca_mpool_base_module_t *mpool)
{
mca_mpool_rdma_module_t *mpool_rdma = (mca_mpool_rdma_module_t*)mpool;
mca_mpool_base_registration_t *reg;
mca_mpool_base_registration_t *regs[RDMA_MPOOL_NREGS];
int reg_cnt, i;
int rc;
/* Statistic */
if(true == mca_mpool_rdma_component.print_stats) {
opal_output(0, "%s rdma: stats "
"(hit/miss/found/not found/evicted): %d/%d/%d/%d/%d\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
mpool_rdma->stat_cache_hit, mpool_rdma->stat_cache_miss,
mpool_rdma->stat_cache_found, mpool_rdma->stat_cache_notfound,
mpool_rdma->stat_evicted);
}
OPAL_THREAD_LOCK(&mpool->rcache->lock);
if(!opal_list_is_empty(&mpool_rdma->gc_list))
do_unregistration_gc(mpool);
do {
reg_cnt = mpool->rcache->rcache_find_all(mpool->rcache, 0, (size_t)-1,
regs, RDMA_MPOOL_NREGS);
for(i = 0; i < reg_cnt; i++) {
reg = regs[i];
if(reg->ref_count) {
reg->ref_count = 0; /* otherway dereg will fail on assert */
} else if (mca_mpool_rdma_component.leave_pinned) {
opal_list_remove_item(&mpool_rdma->mru_list,
(opal_list_item_t*)reg);
}
/* Remove from rcache first */
mpool->rcache->rcache_delete(mpool->rcache, reg);
/* Drop lock before deregistering memory */
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
rc = dereg_mem(mpool, reg);
OPAL_THREAD_LOCK(&mpool->rcache->lock);
if(rc != OMPI_SUCCESS) {
/* Potentially lose track of registrations
do we have to put it back? */
continue;
}
OMPI_FREE_LIST_RETURN(&mpool_rdma->reg_list,
(ompi_free_list_item_t*)reg);
}
} while(reg_cnt == RDMA_MPOOL_NREGS);
OBJ_DESTRUCT(&mpool_rdma->mru_list);
OBJ_DESTRUCT(&mpool_rdma->gc_list);
OBJ_DESTRUCT(&mpool_rdma->reg_list);
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
/* Cleanup any vmas that we have deferred deletion on */
mpool->rcache->rcache_clean(mpool->rcache);
}
/*
* register memory
*/
int mca_mpool_rdma_register(mca_mpool_base_module_t *mpool, void *addr,
size_t size, uint32_t flags,
mca_mpool_base_registration_t **reg)
{
mca_mpool_rdma_module_t *mpool_rdma = (mca_mpool_rdma_module_t*)mpool;
mca_mpool_base_registration_t *rdma_reg;
ompi_free_list_item_t *item;
unsigned char *base, *bound;
int rc;
/* if cache bypass is requested don't use the cache */
if(flags & MCA_MPOOL_FLAGS_CACHE_BYPASS) {
return register_cache_bypass(mpool, addr, size, flags, reg);
}
base = (unsigned char *) down_align_addr(addr, mca_mpool_base_page_size_log);
bound = (unsigned char *) up_align_addr((void*)((char*) addr + size - 1),
mca_mpool_base_page_size_log);
OPAL_THREAD_LOCK(&mpool->rcache->lock);
if(!opal_list_is_empty(&mpool_rdma->gc_list))
do_unregistration_gc(mpool);
/* look through existing regs if not persistent registration requested.
* Persistent registration are always registered and placed in the cache */
if(!(flags & MCA_MPOOL_FLAGS_PERSIST)) {
/* check to see if memory is registered */
mpool->rcache->rcache_find(mpool->rcache, addr, size, reg);
if(*reg != NULL &&
(mca_mpool_rdma_component.leave_pinned ||
((*reg)->flags & MCA_MPOOL_FLAGS_PERSIST) ||
((*reg)->base == base && (*reg)->bound == bound))) {
if(0 == (*reg)->ref_count &&
mca_mpool_rdma_component.leave_pinned) {
opal_list_remove_item(&mpool_rdma->mru_list,
(opal_list_item_t*)(*reg));
}
mpool_rdma->stat_cache_hit++;
(*reg)->ref_count++;
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
return OMPI_SUCCESS;
}
mpool_rdma->stat_cache_miss++;
*reg = NULL; /* in case previous find found something */
/* If no suitable registration is in cache and leave_pinned isn't
* set and size of registration cache is unlimited don't use the cache.
* This is optimisation in case limit is not set. If limit is set we
* have to put registration into the cache to determine when we hit
* memory registration limit.
* NONE: cache is still used for persistent registrations so previous
* find can find something */
if(!mca_mpool_rdma_component.leave_pinned &&
mca_mpool_rdma_component.rcache_size_limit == 0) {
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
return register_cache_bypass(mpool, addr, size, flags, reg);
}
}
OMPI_FREE_LIST_GET(&mpool_rdma->reg_list, item, rc);
if(OMPI_SUCCESS != rc) {
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
return rc;
}
rdma_reg = (mca_mpool_base_registration_t*)item;
rdma_reg->mpool = mpool;
rdma_reg->base = base;
rdma_reg->bound = bound;
rdma_reg->flags = flags;
while((rc = mpool->rcache->rcache_insert(mpool->rcache, rdma_reg,
mca_mpool_rdma_component.rcache_size_limit)) ==
OMPI_ERR_TEMP_OUT_OF_RESOURCE) {
mca_mpool_base_registration_t *old_reg;
/* try to remove one unused reg and retry */
old_reg = (mca_mpool_base_registration_t*)
opal_list_get_last(&mpool_rdma->mru_list);
if(opal_list_get_end(&mpool_rdma->mru_list) !=
(opal_list_item_t*)old_reg) {
/* Remove the registration from the cache and list before
deregistering the memory */
mpool->rcache->rcache_delete(mpool->rcache, old_reg);
opal_list_remove_item(&mpool_rdma->mru_list,
(opal_list_item_t*)old_reg);
/* Drop the rcache lock while we deregister the memory */
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
rc = dereg_mem(mpool, old_reg);
OPAL_THREAD_LOCK(&mpool->rcache->lock);
/* This introduces a potential leak of registrations if
the deregistration fails to occur as we no longer have
a reference to it. Is this possible? */
if(OMPI_SUCCESS == rc) {
OMPI_FREE_LIST_RETURN(&mpool_rdma->reg_list,
(ompi_free_list_item_t*)old_reg);
mpool_rdma->stat_evicted++;
} else
break;
} else
break;
}
if(rc != OMPI_SUCCESS) {
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
OMPI_FREE_LIST_RETURN(&mpool_rdma->reg_list, item);
return rc;
}
rc = mpool_rdma->resources.register_mem(mpool_rdma->resources.reg_data,
base, bound - base + 1, rdma_reg);
if(rc != OMPI_SUCCESS) {
mpool->rcache->rcache_delete(mpool->rcache, rdma_reg);
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
OMPI_FREE_LIST_RETURN(&mpool_rdma->reg_list, item);
return rc;
}
*reg = rdma_reg;
(*reg)->ref_count++;
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
/* Cleanup any vmas that we have deferred deletion on */
mpool->rcache->rcache_clean(mpool->rcache);
return OMPI_SUCCESS;
}
/*
* register memory
*/
int mca_mpool_grdma_register(mca_mpool_base_module_t *mpool, void *addr,
size_t size, uint32_t flags,
mca_mpool_base_registration_t **reg)
{
mca_mpool_grdma_module_t *mpool_grdma = (mca_mpool_grdma_module_t*)mpool;
const bool bypass_cache = !!(flags & MCA_MPOOL_FLAGS_CACHE_BYPASS);
const bool persist = !!(flags & MCA_MPOOL_FLAGS_PERSIST);
mca_mpool_base_registration_t *grdma_reg;
opal_free_list_item_t *item;
unsigned char *base, *bound;
int rc;
OPAL_THREAD_LOCK(&mpool->rcache->lock);
/* if cache bypass is requested don't use the cache */
base = (unsigned char *) down_align_addr(addr, mca_mpool_base_page_size_log);
bound = (unsigned char *) up_align_addr((void*)((char*) addr + size - 1),
mca_mpool_base_page_size_log);
if (!opal_list_is_empty (&mpool_grdma->pool->gc_list))
do_unregistration_gc(mpool);
#if OPAL_CUDA_GDR_SUPPORT
if (flags & MCA_MPOOL_FLAGS_CUDA_GPU_MEM) {
size_t psize;
mca_common_cuda_get_address_range(&base, &psize, addr);
bound = base + psize - 1;
/* Check to see if this memory is in the cache and if it has been freed. If so,
* this call will boot it out of the cache. */
check_for_cuda_freed_memory(mpool, base, psize);
}
#endif /* OPAL_CUDA_GDR_SUPPORT */
/* look through existing regs if not persistent registration requested.
* Persistent registration are always registered and placed in the cache */
if(!(bypass_cache || persist)) {
/* check to see if memory is registered */
mpool->rcache->rcache_find(mpool->rcache, base, bound - base + 1, reg);
if (*reg && !(flags & MCA_MPOOL_FLAGS_INVALID)) {
if (0 == (*reg)->ref_count) {
/* Leave pinned must be set for this to still be in the rcache. */
opal_list_remove_item(&mpool_grdma->pool->lru_list,
(opal_list_item_t *)(*reg));
}
/* This segment fits fully within an existing segment. */
mpool_grdma->stat_cache_hit++;
(*reg)->ref_count++;
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
return OPAL_SUCCESS;
}
mpool_grdma->stat_cache_miss++;
*reg = NULL; /* in case previous find found something */
/* Unless explicitly requested by the caller always store the
* registration in the rcache. This will speed up the case where
* no leave pinned protocol is in use but the same segment is in
* use in multiple simultaneous transactions. We used to set bypass_cache
* here is !mca_mpool_grdma_component.leave_pinned. */
}
item = opal_free_list_get (&mpool_grdma->reg_list);
if(NULL == item) {
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
return OPAL_ERR_OUT_OF_RESOURCE;
}
grdma_reg = (mca_mpool_base_registration_t*)item;
grdma_reg->mpool = mpool;
grdma_reg->base = base;
grdma_reg->bound = bound;
grdma_reg->flags = flags;
#if OPAL_CUDA_GDR_SUPPORT
if (flags & MCA_MPOOL_FLAGS_CUDA_GPU_MEM) {
mca_common_cuda_get_buffer_id(grdma_reg);
}
#endif /* OPAL_CUDA_GDR_SUPPORT */
if (false == bypass_cache) {
rc = mpool->rcache->rcache_insert(mpool->rcache, grdma_reg, 0);
if (OPAL_UNLIKELY(rc != OPAL_SUCCESS)) {
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
opal_free_list_return (&mpool_grdma->reg_list, item);
return rc;
}
}
while (OPAL_ERR_OUT_OF_RESOURCE ==
(rc = mpool_grdma->resources.register_mem(mpool_grdma->resources.reg_data,
base, bound - base + 1, grdma_reg))) {
/* try to remove one unused reg and retry */
if (!mca_mpool_grdma_evict (mpool)) {
break;
}
}
if (OPAL_UNLIKELY(rc != OPAL_SUCCESS)) {
if (false == bypass_cache) {
mpool->rcache->rcache_delete(mpool->rcache, grdma_reg);
}
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
opal_free_list_return (&mpool_grdma->reg_list, item);
return rc;
}
*reg = grdma_reg;
(*reg)->ref_count++;
OPAL_THREAD_UNLOCK(&mpool->rcache->lock);
/* Cleanup any vmas that we have deferred deletion on */
mpool->rcache->rcache_clean(mpool->rcache);
return OPAL_SUCCESS;
}
