-
Notifications
You must be signed in to change notification settings - Fork 380
/
hmem_cuda_gdrcopy.c
428 lines (359 loc) · 10.7 KB
/
hmem_cuda_gdrcopy.c
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
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
/*
* Copyright (c) 2020 Amazon.com, Inc. or its affiliates.
* All rights reserved.
*
* This software is available to you under a choice of one of two
* licenses. You may choose to be licensed under the terms of the GNU
* General Public License (GPL) Version 2, available from the file
* COPYING in the main directory of this source tree, or the
* BSD license below:
*
* Redistribution and use in source and binary forms, with or
* without modification, are permitted provided that the following
* conditions are met:
*
* - Redistributions of source code must retain the above
* copyright notice, this list of conditions and the following
* disclaimer.
*
* - Redistributions in binary form must reproduce the above
* copyright notice, this list of conditions and the following
* disclaimer in the documentation and/or other materials
* provided with the distribution.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#if HAVE_CONFIG_H
#include <config.h>
#endif
#include "ofi_hmem.h"
#include "ofi.h"
#include "ofi_iov.h"
#if HAVE_GDRCOPY
#include <pthread.h>
#include <gdrapi.h>
struct gdrcopy_handle {
gdr_mh_t mh; /* memory handler */
void *cuda_ptr; /* page aligned gpu pointer */
void *user_ptr; /* user space ptr mapped to GPU memory */
size_t length; /* page aligned length */
};
struct gdrcopy_ops {
gdr_t (*gdr_open)();
int (*gdr_close)(gdr_t g);
int (*gdr_pin_buffer)(gdr_t g, unsigned long addr, size_t size,
uint64_t p2p_token, uint32_t va_space,
gdr_mh_t *handle);
int (*gdr_unpin_buffer)(gdr_t g, gdr_mh_t handle);
int (*gdr_map)(gdr_t g, gdr_mh_t handle, void **va, size_t size);
int (*gdr_unmap)(gdr_t g, gdr_mh_t handle, void *va, size_t size);
int (*gdr_copy_to_mapping)(gdr_mh_t handle, void *map_d_ptr,
const void *h_ptr, size_t size);
int (*gdr_copy_from_mapping)(gdr_mh_t handle, void *map_d_ptr,
const void *h_ptr, size_t size);
};
enum gdrcopy_dir {
GDRCOPY_TO_DEVICE,
GDRCOPY_FROM_DEVICE,
};
static gdr_t global_gdr;
static pthread_spinlock_t global_gdr_lock;
#if ENABLE_GDRCOPY_DLOPEN
#include <dlfcn.h>
static void *gdrapi_handle;
static struct gdrcopy_ops global_gdrcopy_ops;
static int cuda_gdrcopy_dl_hmem_init(void)
{
gdrapi_handle = dlopen("libgdrapi.so", RTLD_NOW);
if (!gdrapi_handle) {
FI_INFO(&core_prov, FI_LOG_CORE,
"Failed to dlopen libgdrapi.so\n");
return -FI_ENOSYS;
}
global_gdrcopy_ops.gdr_open = dlsym(gdrapi_handle, "gdr_open");
if (!global_gdrcopy_ops.gdr_open) {
FI_WARN(&core_prov, FI_LOG_CORE, "Failed to find gdr_open\n");
goto err_dlclose_gdrapi;
}
global_gdrcopy_ops.gdr_close = dlsym(gdrapi_handle, "gdr_close");
if (!global_gdrcopy_ops.gdr_close) {
FI_WARN(&core_prov, FI_LOG_CORE, "Failed to find gdr_close\n");
goto err_dlclose_gdrapi;
}
global_gdrcopy_ops.gdr_pin_buffer = dlsym(gdrapi_handle, "gdr_pin_buffer");
if (!global_gdrcopy_ops.gdr_pin_buffer) {
FI_WARN(&core_prov, FI_LOG_CORE,
"Failed to find gdr_pin_buffer\n");
goto err_dlclose_gdrapi;
}
global_gdrcopy_ops.gdr_unpin_buffer = dlsym(gdrapi_handle, "gdr_unpin_buffer");
if (!global_gdrcopy_ops.gdr_unpin_buffer) {
FI_WARN(&core_prov, FI_LOG_CORE,
"Failed to find gdr_unpin_buffer\n");
goto err_dlclose_gdrapi;
}
global_gdrcopy_ops.gdr_map = dlsym(gdrapi_handle, "gdr_map");
if (!global_gdrcopy_ops.gdr_map) {
FI_WARN(&core_prov, FI_LOG_CORE,
"Failed to find gdr_map\n");
goto err_dlclose_gdrapi;
}
global_gdrcopy_ops.gdr_unmap = dlsym(gdrapi_handle, "gdr_unmap");
if (!global_gdrcopy_ops.gdr_unmap) {
FI_WARN(&core_prov, FI_LOG_CORE,
"Failed to find gdr_unmap\n");
goto err_dlclose_gdrapi;
}
global_gdrcopy_ops.gdr_copy_to_mapping = dlsym(gdrapi_handle, "gdr_copy_to_mapping");
if (!global_gdrcopy_ops.gdr_copy_to_mapping) {
FI_WARN(&core_prov, FI_LOG_CORE,
"Failed to find gdr_copy_to_mapping\n");
goto err_dlclose_gdrapi;
}
global_gdrcopy_ops.gdr_copy_from_mapping = dlsym(gdrapi_handle, "gdr_copy_from_mapping");
if (!global_gdrcopy_ops.gdr_copy_from_mapping) {
FI_WARN(&core_prov, FI_LOG_CORE,
"Failed to find gdr_copy_from_mapping\n");
goto err_dlclose_gdrapi;
}
return FI_SUCCESS;
err_dlclose_gdrapi:
memset(&global_gdrcopy_ops, 0, sizeof(global_gdrcopy_ops));
dlclose(gdrapi_handle);
return -FI_ENODATA;
}
static int cuda_gdrcopy_dl_hmem_cleanup(void)
{
dlclose(gdrapi_handle);
return FI_SUCCESS;
}
#else
static struct gdrcopy_ops global_gdrcopy_ops = {
.gdr_open = gdr_open,
.gdr_close = gdr_close,
.gdr_pin_buffer = gdr_pin_buffer,
.gdr_unpin_buffer = gdr_unpin_buffer,
.gdr_map = gdr_map,
.gdr_unmap = gdr_unmap,
.gdr_copy_to_mapping = gdr_copy_to_mapping,
.gdr_copy_from_mapping = gdr_copy_from_mapping
};
static int cuda_gdrcopy_dl_hmem_init(void)
{
return FI_SUCCESS;
}
static int cuda_gdrcopy_dl_hmem_cleanup(void)
{
return FI_SUCCESS;
}
#endif /* ENABLE_CUDA_DLOPEN */
int cuda_gdrcopy_hmem_init(void)
{
int err;
err = cuda_gdrcopy_dl_hmem_init();
if (err) {
FI_WARN(&core_prov, FI_LOG_CORE,
"gdrcopy_dl_hmem_init failed!\n");
return -FI_ENOSYS;
}
assert(global_gdrcopy_ops.gdr_open);
global_gdr = global_gdrcopy_ops.gdr_open();
if (!global_gdr) {
FI_WARN(&core_prov, FI_LOG_CORE,
"gdr_open failed!\n");
err = -FI_ENOMEM;
goto error_exit;
}
err = pthread_spin_init(&global_gdr_lock, 0);
if (err) {
assert(global_gdrcopy_ops.gdr_close);
global_gdrcopy_ops.gdr_close(global_gdr);
err = -err;
goto error_exit;
}
return 0;
error_exit:
cuda_gdrcopy_dl_hmem_cleanup();
return err;
}
int cuda_gdrcopy_hmem_cleanup(void)
{
int err, ret = 0;
err = pthread_spin_destroy(&global_gdr_lock);
if (err) {
FI_WARN(&core_prov, FI_LOG_CORE,
"destroy global_gdr_lock failed! err: %s\n",
strerror(err));
ret = err;
}
assert(global_gdrcopy_ops.gdr_close);
err = global_gdrcopy_ops.gdr_close(global_gdr);
if (err) {
FI_WARN(&core_prov, FI_LOG_CORE,
"close global_gdr failed! err: %s\n",
strerror(err));
ret = err;
}
err = cuda_gdrcopy_dl_hmem_cleanup();
if (err) {
FI_WARN(&core_prov, FI_LOG_CORE,
"cuda_gdrcopy_dl_hmem_cleaup() failed! err: %s\n",
strerror(err));
ret = err;
}
return ret;
}
void cuda_gdrcopy_impl(uint64_t handle, void *devptr,
void *hostptr, size_t len,
enum gdrcopy_dir dir)
{
ssize_t off;
struct gdrcopy_handle *gdrcopy;
void *gdrcopy_user_ptr;
assert(global_gdrcopy_ops.gdr_copy_to_mapping);
assert(handle);
gdrcopy = (struct gdrcopy_handle *)handle;
off = (char *)devptr - (char *)gdrcopy->cuda_ptr;
assert(off >= 0 && off + len <= gdrcopy->length);
gdrcopy_user_ptr = (char *)gdrcopy->user_ptr + off;
if (dir == GDRCOPY_TO_DEVICE) {
global_gdrcopy_ops.gdr_copy_to_mapping(gdrcopy->mh,
gdrcopy_user_ptr,
hostptr, len);
} else {
assert(dir == GDRCOPY_FROM_DEVICE);
global_gdrcopy_ops.gdr_copy_from_mapping(gdrcopy->mh,
hostptr,
gdrcopy_user_ptr,
len);
}
}
void cuda_gdrcopy_to_dev(uint64_t handle, void *devptr,
const void *hostptr, size_t len)
{
cuda_gdrcopy_impl(handle, devptr, (void *)hostptr, len,
GDRCOPY_TO_DEVICE);
}
void cuda_gdrcopy_from_dev(uint64_t handle, void *hostptr,
const void *devptr, size_t len)
{
cuda_gdrcopy_impl(handle, (void *)devptr, hostptr, len,
GDRCOPY_FROM_DEVICE);
}
int cuda_gdrcopy_dev_register(const void *buf, size_t len, uint64_t *handle)
{
int err;
uintptr_t regbgn, regend;
size_t reglen;
struct gdrcopy_handle *gdrcopy;
assert(global_gdr);
assert(global_gdrcopy_ops.gdr_pin_buffer);
assert(global_gdrcopy_ops.gdr_map);
regbgn = (uintptr_t)ofi_get_page_start(buf, GPU_PAGE_SIZE);
regend = (uintptr_t)buf + len;
reglen = ofi_get_aligned_size(regend - regbgn, GPU_PAGE_SIZE);
gdrcopy = malloc(sizeof(struct gdrcopy_handle));
if (!gdrcopy)
return -FI_ENOMEM;
assert(global_gdr);
pthread_spin_lock(&global_gdr_lock);
err = global_gdrcopy_ops.gdr_pin_buffer(global_gdr, regbgn,
reglen, 0, 0, &gdrcopy->mh);
if (err) {
FI_WARN(&core_prov, FI_LOG_CORE,
"gdr_pin_buffer failed! error: %s ptr: %p len: %ld\n",
strerror(err), buf, len);
free(gdrcopy);
goto exit;
}
gdrcopy->cuda_ptr = (void *)regbgn;
gdrcopy->length = reglen;
err = global_gdrcopy_ops.gdr_map(global_gdr, gdrcopy->mh,
&gdrcopy->user_ptr, gdrcopy->length);
if (err) {
FI_WARN(&core_prov, FI_LOG_CORE, "gdr_map failed! error: %s\n",
strerror(err));
global_gdrcopy_ops.gdr_unpin_buffer(global_gdr, gdrcopy->mh);
free(gdrcopy);
goto exit;
}
*handle = (uint64_t)gdrcopy;
exit:
pthread_spin_unlock(&global_gdr_lock);
return err;
}
int cuda_gdrcopy_dev_unregister(uint64_t handle)
{
int err;
struct gdrcopy_handle *gdrcopy;
assert(global_gdr);
assert(global_gdrcopy_ops.gdr_unmap);
assert(global_gdrcopy_ops.gdr_unpin_buffer);
gdrcopy = (struct gdrcopy_handle *)handle;
assert(gdrcopy);
pthread_spin_lock(&global_gdr_lock);
err = global_gdrcopy_ops.gdr_unmap(global_gdr, gdrcopy->mh,
gdrcopy->user_ptr, gdrcopy->length);
if (err) {
FI_WARN(&core_prov, FI_LOG_CORE,
"gdr_unmap failed! error: %s\n",
strerror(err));
goto exit;
}
err = global_gdrcopy_ops.gdr_unpin_buffer(global_gdr, gdrcopy->mh);
if (err) {
FI_WARN(&core_prov, FI_LOG_MR,
"gdr_unpin_buffer failed! error: %s\n",
strerror(err));
goto exit;
}
exit:
pthread_spin_unlock(&global_gdr_lock);
free(gdrcopy);
return err;
}
#else
int cuda_gdrcopy_hmem_init(void)
{
return -FI_ENOSYS;
}
int cuda_gdrcopy_hmem_cleanup(void)
{
return FI_SUCCESS;
}
void cuda_gdrcopy_to_dev(uint64_t devhandle, void *devptr,
const void *hostptr, size_t len)
{
}
void cuda_gdrcopy_from_dev(uint64_t devhandle, void *hostptr,
const void *devptr, size_t len)
{
}
int cuda_gdrcopy_dev_register(const void *buf, size_t len, uint64_t *handle)
{
return FI_SUCCESS;
}
int cuda_gdrcopy_dev_unregister(uint64_t handle)
{
return FI_SUCCESS;
}
ssize_t ofi_gdrcopy_to_cuda_iov(uint64_t handle, const struct iovec *iov,
size_t iov_count, uint64_t iov_offset,
const void *host, size_t len)
{
return -FI_ENOSYS;
}
ssize_t ofi_gdrcopy_from_cuda_iov(uint64_t handle, void *host,
const struct iovec *iov, size_t iov_count,
uint64_t iov_offset, size_t len)
{
return -FI_ENOSYS;
}
#endif /* HAVE_GDRCOPY */