1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
|
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2009 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006-2007 Mellanox Technologies. All rights reserved.
* Copyright (c) 2010-2013 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2011 NVIDIA Corporation. All rights reserved.
* Copyright (c) 2012-2018 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "opal_config.h"
#include "opal/align.h"
#include "opal/class/opal_free_list.h"
#include "opal/mca/mpool/base/base.h"
#include "opal/mca/mpool/mpool.h"
#include "opal/mca/rcache/rcache.h"
#include "opal/util/output.h"
#include "opal/util/sys_limits.h"
typedef struct opal_free_list_item_t opal_free_list_memory_t;
OBJ_CLASS_INSTANCE(opal_free_list_item_t, opal_list_item_t, NULL, NULL);
static void opal_free_list_construct(opal_free_list_t *fl)
{
OBJ_CONSTRUCT(&fl->fl_lock, opal_mutex_t);
OBJ_CONSTRUCT(&fl->fl_condition, opal_condition_t);
fl->fl_max_to_alloc = 0;
fl->fl_num_allocated = 0;
fl->fl_num_per_alloc = 0;
fl->fl_num_waiting = 0;
fl->fl_frag_size = sizeof(opal_free_list_item_t);
fl->fl_frag_alignment = 0;
fl->fl_payload_buffer_size = 0;
fl->fl_payload_buffer_alignment = 0;
fl->fl_frag_class = OBJ_CLASS(opal_free_list_item_t);
fl->fl_mpool = NULL;
fl->fl_rcache = NULL;
/* default flags */
fl->fl_rcache_reg_flags = MCA_RCACHE_FLAGS_CACHE_BYPASS | MCA_RCACHE_FLAGS_ACCELERATOR_REGISTER_MEM;
fl->ctx = NULL;
OBJ_CONSTRUCT(&(fl->fl_allocations), opal_list_t);
}
static void opal_free_list_allocation_release(opal_free_list_t *fl, opal_free_list_memory_t *fl_mem)
{
if (NULL != fl->fl_rcache) {
fl->fl_rcache->rcache_deregister(fl->fl_rcache, fl_mem->registration);
}
if (NULL != fl->fl_mpool) {
fl->fl_mpool->mpool_free(fl->fl_mpool, fl_mem->ptr);
} else if (fl_mem->ptr) {
free(fl_mem->ptr);
}
/* destruct the item (we constructed it), then free the memory chunk */
OBJ_DESTRUCT(fl_mem);
free(fl_mem);
}
static void opal_free_list_destruct(opal_free_list_t *fl)
{
opal_list_item_t *item;
opal_free_list_item_t *fl_item;
#if 0 && OPAL_ENABLE_DEBUG
if(opal_list_get_size(&fl->super) != fl->fl_num_allocated) {
opal_output(0, "opal_free_list: %d allocated %d returned: %s:%d\n",
fl->fl_num_allocated, opal_list_get_size(&fl->super),
fl->super.super.cls_init_file_name, fl->super.super.cls_init_lineno);
}
#endif
while (NULL != (item = opal_lifo_pop(&(fl->super)))) {
fl_item = (opal_free_list_item_t *) item;
/* destruct the item (we constructed it), the underlying memory will be
* reclaimed when we free the slab (opal_free_list_memory_t ptr)
* containing it */
OBJ_DESTRUCT(fl_item);
}
while (NULL != (item = opal_list_remove_first(&fl->fl_allocations))) {
opal_free_list_allocation_release(fl, (opal_free_list_memory_t *) item);
}
OBJ_DESTRUCT(&fl->fl_allocations);
OBJ_DESTRUCT(&fl->fl_condition);
OBJ_DESTRUCT(&fl->fl_lock);
}
OBJ_CLASS_INSTANCE(opal_free_list_t, opal_lifo_t, opal_free_list_construct,
opal_free_list_destruct);
int opal_free_list_init(opal_free_list_t *flist, size_t frag_size, size_t frag_alignment,
opal_class_t *frag_class, size_t payload_buffer_size,
size_t payload_buffer_alignment, int num_elements_to_alloc,
int max_elements_to_alloc, int num_elements_per_alloc,
mca_mpool_base_module_t *mpool, int rcache_reg_flags,
mca_rcache_base_module_t *rcache, opal_free_list_item_init_fn_t item_init,
void *ctx)
{
/* alignment must be more than zero and power of two */
if (frag_alignment <= 1 || (frag_alignment & (frag_alignment - 1))) {
return OPAL_ERROR;
}
if (0 < payload_buffer_size) {
if (payload_buffer_alignment <= 1
|| (payload_buffer_alignment & (payload_buffer_alignment - 1))) {
return OPAL_ERROR;
}
}
if (frag_class && frag_size < frag_class->cls_sizeof) {
frag_size = frag_class->cls_sizeof;
}
if (frag_size > flist->fl_frag_size) {
flist->fl_frag_size = frag_size;
}
if (frag_class) {
flist->fl_frag_class = frag_class;
}
flist->fl_payload_buffer_size = payload_buffer_size;
flist->fl_max_to_alloc = max_elements_to_alloc;
flist->fl_num_allocated = 0;
flist->fl_num_per_alloc = num_elements_per_alloc;
flist->fl_mpool = mpool ? mpool : mca_mpool_base_default_module;
flist->fl_rcache = rcache;
flist->fl_frag_alignment = frag_alignment;
flist->fl_payload_buffer_alignment = payload_buffer_alignment;
flist->item_init = item_init;
flist->fl_rcache_reg_flags |= rcache_reg_flags;
flist->ctx = ctx;
if (num_elements_to_alloc) {
return opal_free_list_grow_st(flist, num_elements_to_alloc, NULL);
}
return OPAL_SUCCESS;
}
int opal_free_list_grow_st(opal_free_list_t *flist, size_t num_elements,
opal_free_list_item_t **item_out)
{
unsigned char *ptr, *payload_ptr = NULL;
opal_free_list_memory_t *alloc_ptr;
size_t alloc_size, head_size, elem_size = 0, buffer_size = 0, align = 0;
mca_rcache_base_registration_t *reg = NULL;
int rc = OPAL_SUCCESS;
if (flist->fl_max_to_alloc
&& (flist->fl_num_allocated + num_elements) > flist->fl_max_to_alloc) {
num_elements = flist->fl_max_to_alloc - flist->fl_num_allocated;
}
if (num_elements == 0) {
return OPAL_ERR_TEMP_OUT_OF_RESOURCE;
}
head_size = OPAL_ALIGN(flist->fl_frag_size, flist->fl_frag_alignment, size_t);
/* NTH: calculate allocation alignment first as it might change the number of elements */
if (0 != flist->fl_payload_buffer_size) {
elem_size = OPAL_ALIGN(flist->fl_payload_buffer_size, flist->fl_payload_buffer_alignment,
size_t);
/* elem_size should not be 0 here */
assert(elem_size > 0);
buffer_size = num_elements * elem_size;
align = flist->fl_payload_buffer_alignment;
if (MCA_RCACHE_FLAGS_ACCELERATOR_REGISTER_MEM & flist->fl_rcache_reg_flags) {
size_t pagesize = opal_getpagesize();
/* CUDA cannot handle registering overlapping regions, so make
* sure each region is page sized and page aligned. */
align = OPAL_ALIGN(align, pagesize, size_t);
buffer_size = OPAL_ALIGN(buffer_size, pagesize, size_t);
/* avoid wasting space in the buffer */
num_elements = buffer_size / elem_size;
}
}
/* calculate head allocation size */
alloc_size = num_elements * head_size + sizeof(opal_free_list_memory_t)
+ flist->fl_frag_alignment;
alloc_ptr = (opal_free_list_memory_t *) malloc(alloc_size);
if (OPAL_UNLIKELY(NULL == alloc_ptr)) {
return OPAL_ERR_TEMP_OUT_OF_RESOURCE;
}
if (0 != flist->fl_payload_buffer_size) {
/* allocate the rest from the mpool (or use memalign/malloc) */
payload_ptr = (unsigned char *) flist->fl_mpool->mpool_alloc(flist->fl_mpool, buffer_size,
align, 0);
if (NULL == payload_ptr) {
free(alloc_ptr);
return OPAL_ERR_TEMP_OUT_OF_RESOURCE;
}
if (flist->fl_rcache) {
rc = flist->fl_rcache->rcache_register(flist->fl_rcache, payload_ptr,
num_elements * elem_size,
flist->fl_rcache_reg_flags,
MCA_RCACHE_ACCESS_ANY, ®);
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
free(alloc_ptr);
flist->fl_mpool->mpool_free(flist->fl_mpool, payload_ptr);
return rc;
}
}
}
/* make the alloc_ptr a list item, save the chunk in the allocations list,
* and have ptr point to memory right after the list item structure */
OBJ_CONSTRUCT(alloc_ptr, opal_free_list_item_t);
opal_list_append(&(flist->fl_allocations), (opal_list_item_t *) alloc_ptr);
alloc_ptr->registration = reg;
alloc_ptr->ptr = payload_ptr;
ptr = (unsigned char *) alloc_ptr + sizeof(opal_free_list_memory_t);
ptr = OPAL_ALIGN_PTR(ptr, flist->fl_frag_alignment, unsigned char *);
for (size_t i = 0; i < num_elements; ++i) {
opal_free_list_item_t *item = (opal_free_list_item_t *) ptr;
item->registration = reg;
item->ptr = payload_ptr;
OBJ_CONSTRUCT_INTERNAL(item, flist->fl_frag_class);
item->super.item_free = 0;
/* run the initialize function if present */
if (flist->item_init) {
if (OPAL_SUCCESS != (rc = flist->item_init(item, flist->ctx))) {
num_elements = i;
OBJ_DESTRUCT(item);
break;
}
}
/* NTH: in case the free list may be accessed from multiple threads
* use the atomic lifo push. The overhead is small compared to the
* overall overhead of opal_free_list_grow(). */
if (item_out && 0 == i) {
/* ensure the thread that is growing the free list always gets an item
* if one is available */
*item_out = item;
} else {
opal_lifo_push_atomic(&flist->super, &item->super);
}
ptr += head_size;
if (NULL != payload_ptr) {
payload_ptr += elem_size;
}
}
if (OPAL_SUCCESS != rc && 0 == num_elements) {
/* couldn't initialize any items */
opal_list_remove_item(&flist->fl_allocations, (opal_list_item_t *) alloc_ptr);
opal_free_list_allocation_release(flist, alloc_ptr);
return OPAL_ERR_OUT_OF_RESOURCE;
}
flist->fl_num_allocated += num_elements;
return OPAL_SUCCESS;
}
/**
* This function resize the free_list to contain at least the specified
* number of elements. We do not create all of them in the same memory
* segment. Instead we will several time the fl_num_per_alloc elements
* until we reach the required number of the maximum allowed by the
* initialization.
*/
int opal_free_list_resize_mt(opal_free_list_t *flist, size_t size)
{
ssize_t inc_num;
int ret = OPAL_SUCCESS;
if (flist->fl_num_allocated > size) {
return OPAL_SUCCESS;
}
opal_mutex_lock(&flist->fl_lock);
do {
ret = opal_free_list_grow_st(flist, flist->fl_num_per_alloc, NULL);
if (OPAL_SUCCESS != ret) {
break;
}
inc_num = (ssize_t) size - (ssize_t) flist->fl_num_allocated;
} while (inc_num > 0);
opal_mutex_unlock(&flist->fl_lock);
return ret;
}
|