Line data Source code
1 : /* SPDX-License-Identifier: BSD-3-Clause
2 : * Copyright (C) 2017 Intel Corporation.
3 : * All rights reserved.
4 : */
5 :
6 : #include "spdk/stdinc.h"
7 :
8 : #include "env_internal.h"
9 : #include "pci_dpdk.h"
10 :
11 : #include <rte_config.h>
12 : #include <rte_memory.h>
13 : #include <rte_eal_memconfig.h>
14 : #include <rte_dev.h>
15 : #include <rte_pci.h>
16 :
17 : #include "spdk_internal/assert.h"
18 :
19 : #include "spdk/assert.h"
20 : #include "spdk/likely.h"
21 : #include "spdk/queue.h"
22 : #include "spdk/util.h"
23 : #include "spdk/memory.h"
24 : #include "spdk/env_dpdk.h"
25 : #include "spdk/log.h"
26 :
27 : #ifdef __linux__
28 : #include <linux/version.h>
29 : #if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 6, 0)
30 : #include <linux/vfio.h>
31 : #include <rte_vfio.h>
32 :
33 : struct spdk_vfio_dma_map {
34 : struct vfio_iommu_type1_dma_map map;
35 : TAILQ_ENTRY(spdk_vfio_dma_map) tailq;
36 : };
37 :
38 : struct vfio_cfg {
39 : int fd;
40 : bool enabled;
41 : bool noiommu_enabled;
42 : unsigned device_ref;
43 : TAILQ_HEAD(, spdk_vfio_dma_map) maps;
44 : pthread_mutex_t mutex;
45 : };
46 :
47 : static struct vfio_cfg g_vfio = {
48 : .fd = -1,
49 : .enabled = false,
50 : .noiommu_enabled = false,
51 : .device_ref = 0,
52 : .maps = TAILQ_HEAD_INITIALIZER(g_vfio.maps),
53 : .mutex = PTHREAD_MUTEX_INITIALIZER
54 : };
55 : #endif
56 : #endif
57 :
58 : #if DEBUG
59 : #define DEBUG_PRINT(...) SPDK_ERRLOG(__VA_ARGS__)
60 : #else
61 : #define DEBUG_PRINT(...)
62 : #endif
63 :
64 : #define FN_2MB_TO_4KB(fn) (fn << (SHIFT_2MB - SHIFT_4KB))
65 : #define FN_4KB_TO_2MB(fn) (fn >> (SHIFT_2MB - SHIFT_4KB))
66 :
67 : #define MAP_256TB_IDX(vfn_2mb) ((vfn_2mb) >> (SHIFT_1GB - SHIFT_2MB))
68 : #define MAP_1GB_IDX(vfn_2mb) ((vfn_2mb) & ((1ULL << (SHIFT_1GB - SHIFT_2MB)) - 1))
69 :
70 : /* Page is registered */
71 : #define REG_MAP_REGISTERED (1ULL << 62)
72 :
73 : /* A notification region barrier. The 2MB translation entry that's marked
74 : * with this flag must be unregistered separately. This allows contiguous
75 : * regions to be unregistered in the same chunks they were registered.
76 : */
77 : #define REG_MAP_NOTIFY_START (1ULL << 63)
78 :
79 : /* Translation of a single 2MB page. */
80 : struct map_2mb {
81 : uint64_t translation_2mb;
82 : };
83 :
84 : /* Second-level map table indexed by bits [21..29] of the virtual address.
85 : * Each entry contains the address translation or error for entries that haven't
86 : * been retrieved yet.
87 : */
88 : struct map_1gb {
89 : struct map_2mb map[1ULL << (SHIFT_1GB - SHIFT_2MB)];
90 : };
91 :
92 : /* Top-level map table indexed by bits [30..47] of the virtual address.
93 : * Each entry points to a second-level map table or NULL.
94 : */
95 : struct map_256tb {
96 : struct map_1gb *map[1ULL << (SHIFT_256TB - SHIFT_1GB)];
97 : };
98 :
99 : /* Page-granularity memory address translation */
100 : struct spdk_mem_map {
101 : struct map_256tb map_256tb;
102 : pthread_mutex_t mutex;
103 : uint64_t default_translation;
104 : struct spdk_mem_map_ops ops;
105 : void *cb_ctx;
106 : TAILQ_ENTRY(spdk_mem_map) tailq;
107 : };
108 :
109 : /* Registrations map. The 64 bit translations are bit fields with the
110 : * following layout (starting with the low bits):
111 : * 0 - 61 : reserved
112 : * 62 - 63 : flags
113 : */
114 : static struct spdk_mem_map *g_mem_reg_map;
115 : static TAILQ_HEAD(spdk_mem_map_head, spdk_mem_map) g_spdk_mem_maps =
116 : TAILQ_HEAD_INITIALIZER(g_spdk_mem_maps);
117 : static pthread_mutex_t g_spdk_mem_map_mutex = PTHREAD_MUTEX_INITIALIZER;
118 :
119 : static bool g_legacy_mem;
120 : static bool g_huge_pages = true;
121 :
122 : /*
123 : * Walk the currently registered memory via the main memory registration map
124 : * and call the new map's notify callback for each virtually contiguous region.
125 : */
126 : static int
127 0 : mem_map_notify_walk(struct spdk_mem_map *map, enum spdk_mem_map_notify_action action)
128 : {
129 : size_t idx_256tb;
130 : uint64_t idx_1gb;
131 0 : uint64_t contig_start = UINT64_MAX;
132 0 : uint64_t contig_end = UINT64_MAX;
133 : struct map_1gb *map_1gb;
134 : int rc;
135 :
136 0 : if (!g_mem_reg_map) {
137 0 : return -EINVAL;
138 : }
139 :
140 : /* Hold the memory registration map mutex so no new registrations can be added while we are looping. */
141 0 : pthread_mutex_lock(&g_mem_reg_map->mutex);
142 :
143 0 : for (idx_256tb = 0;
144 0 : idx_256tb < sizeof(g_mem_reg_map->map_256tb.map) / sizeof(g_mem_reg_map->map_256tb.map[0]);
145 0 : idx_256tb++) {
146 0 : map_1gb = g_mem_reg_map->map_256tb.map[idx_256tb];
147 :
148 0 : if (!map_1gb) {
149 0 : if (contig_start != UINT64_MAX) {
150 : /* End of of a virtually contiguous range */
151 0 : rc = map->ops.notify_cb(map->cb_ctx, map, action,
152 : (void *)contig_start,
153 0 : contig_end - contig_start + VALUE_2MB);
154 : /* Don't bother handling unregister failures. It can't be any worse */
155 0 : if (rc != 0 && action == SPDK_MEM_MAP_NOTIFY_REGISTER) {
156 0 : goto err_unregister;
157 : }
158 : }
159 0 : contig_start = UINT64_MAX;
160 0 : continue;
161 : }
162 :
163 0 : for (idx_1gb = 0; idx_1gb < sizeof(map_1gb->map) / sizeof(map_1gb->map[0]); idx_1gb++) {
164 0 : if ((map_1gb->map[idx_1gb].translation_2mb & REG_MAP_REGISTERED) &&
165 0 : (contig_start == UINT64_MAX ||
166 0 : (map_1gb->map[idx_1gb].translation_2mb & REG_MAP_NOTIFY_START) == 0)) {
167 : /* Rebuild the virtual address from the indexes */
168 0 : uint64_t vaddr = (idx_256tb << SHIFT_1GB) | (idx_1gb << SHIFT_2MB);
169 :
170 0 : if (contig_start == UINT64_MAX) {
171 0 : contig_start = vaddr;
172 : }
173 :
174 0 : contig_end = vaddr;
175 : } else {
176 0 : if (contig_start != UINT64_MAX) {
177 : /* End of of a virtually contiguous range */
178 0 : rc = map->ops.notify_cb(map->cb_ctx, map, action,
179 : (void *)contig_start,
180 0 : contig_end - contig_start + VALUE_2MB);
181 : /* Don't bother handling unregister failures. It can't be any worse */
182 0 : if (rc != 0 && action == SPDK_MEM_MAP_NOTIFY_REGISTER) {
183 0 : goto err_unregister;
184 : }
185 :
186 : /* This page might be a part of a neighbour region, so process
187 : * it again. The idx_1gb will be incremented immediately.
188 : */
189 0 : idx_1gb--;
190 : }
191 0 : contig_start = UINT64_MAX;
192 : }
193 : }
194 : }
195 :
196 0 : pthread_mutex_unlock(&g_mem_reg_map->mutex);
197 0 : return 0;
198 :
199 0 : err_unregister:
200 : /* Unwind to the first empty translation so we don't unregister
201 : * a region that just failed to register.
202 : */
203 0 : idx_256tb = MAP_256TB_IDX((contig_start >> SHIFT_2MB) - 1);
204 0 : idx_1gb = MAP_1GB_IDX((contig_start >> SHIFT_2MB) - 1);
205 0 : contig_start = UINT64_MAX;
206 0 : contig_end = UINT64_MAX;
207 :
208 : /* Unregister any memory we managed to register before the failure */
209 0 : for (; idx_256tb < SIZE_MAX; idx_256tb--) {
210 0 : map_1gb = g_mem_reg_map->map_256tb.map[idx_256tb];
211 :
212 0 : if (!map_1gb) {
213 0 : if (contig_end != UINT64_MAX) {
214 : /* End of of a virtually contiguous range */
215 0 : map->ops.notify_cb(map->cb_ctx, map,
216 : SPDK_MEM_MAP_NOTIFY_UNREGISTER,
217 : (void *)contig_start,
218 0 : contig_end - contig_start + VALUE_2MB);
219 : }
220 0 : contig_end = UINT64_MAX;
221 0 : continue;
222 : }
223 :
224 0 : for (; idx_1gb < UINT64_MAX; idx_1gb--) {
225 : /* Rebuild the virtual address from the indexes */
226 0 : uint64_t vaddr = (idx_256tb << SHIFT_1GB) | (idx_1gb << SHIFT_2MB);
227 0 : if ((map_1gb->map[idx_1gb].translation_2mb & REG_MAP_REGISTERED) &&
228 0 : (contig_end == UINT64_MAX || (map_1gb->map[idx_1gb].translation_2mb & REG_MAP_NOTIFY_START) == 0)) {
229 :
230 0 : if (contig_end == UINT64_MAX) {
231 0 : contig_end = vaddr;
232 : }
233 0 : contig_start = vaddr;
234 : } else {
235 0 : if (contig_end != UINT64_MAX) {
236 0 : if (map_1gb->map[idx_1gb].translation_2mb & REG_MAP_NOTIFY_START) {
237 0 : contig_start = vaddr;
238 : }
239 : /* End of of a virtually contiguous range */
240 0 : map->ops.notify_cb(map->cb_ctx, map,
241 : SPDK_MEM_MAP_NOTIFY_UNREGISTER,
242 : (void *)contig_start,
243 0 : contig_end - contig_start + VALUE_2MB);
244 : }
245 0 : contig_end = UINT64_MAX;
246 : }
247 : }
248 0 : idx_1gb = sizeof(map_1gb->map) / sizeof(map_1gb->map[0]) - 1;
249 : }
250 :
251 0 : pthread_mutex_unlock(&g_mem_reg_map->mutex);
252 0 : return rc;
253 : }
254 :
255 : struct spdk_mem_map *
256 0 : spdk_mem_map_alloc(uint64_t default_translation, const struct spdk_mem_map_ops *ops, void *cb_ctx)
257 : {
258 : struct spdk_mem_map *map;
259 : int rc;
260 : size_t i;
261 :
262 0 : map = calloc(1, sizeof(*map));
263 0 : if (map == NULL) {
264 0 : return NULL;
265 : }
266 :
267 0 : if (pthread_mutex_init(&map->mutex, NULL)) {
268 0 : free(map);
269 0 : return NULL;
270 : }
271 :
272 0 : map->default_translation = default_translation;
273 0 : map->cb_ctx = cb_ctx;
274 0 : if (ops) {
275 0 : map->ops = *ops;
276 : }
277 :
278 0 : if (ops && ops->notify_cb) {
279 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
280 0 : rc = mem_map_notify_walk(map, SPDK_MEM_MAP_NOTIFY_REGISTER);
281 0 : if (rc != 0) {
282 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
283 0 : DEBUG_PRINT("Initial mem_map notify failed\n");
284 0 : pthread_mutex_destroy(&map->mutex);
285 0 : for (i = 0; i < sizeof(map->map_256tb.map) / sizeof(map->map_256tb.map[0]); i++) {
286 0 : free(map->map_256tb.map[i]);
287 : }
288 0 : free(map);
289 0 : return NULL;
290 : }
291 0 : TAILQ_INSERT_TAIL(&g_spdk_mem_maps, map, tailq);
292 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
293 : }
294 :
295 0 : return map;
296 : }
297 :
298 : void
299 0 : spdk_mem_map_free(struct spdk_mem_map **pmap)
300 : {
301 : struct spdk_mem_map *map;
302 : size_t i;
303 :
304 0 : if (!pmap) {
305 0 : return;
306 : }
307 :
308 0 : map = *pmap;
309 :
310 0 : if (!map) {
311 0 : return;
312 : }
313 :
314 0 : if (map->ops.notify_cb) {
315 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
316 0 : mem_map_notify_walk(map, SPDK_MEM_MAP_NOTIFY_UNREGISTER);
317 0 : TAILQ_REMOVE(&g_spdk_mem_maps, map, tailq);
318 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
319 : }
320 :
321 0 : for (i = 0; i < sizeof(map->map_256tb.map) / sizeof(map->map_256tb.map[0]); i++) {
322 0 : free(map->map_256tb.map[i]);
323 : }
324 :
325 0 : pthread_mutex_destroy(&map->mutex);
326 :
327 0 : free(map);
328 0 : *pmap = NULL;
329 : }
330 :
331 : int
332 0 : spdk_mem_register(void *_vaddr, size_t len)
333 : {
334 : struct spdk_mem_map *map;
335 : int rc;
336 0 : uint64_t vaddr = (uintptr_t)_vaddr;
337 : uint64_t seg_vaddr;
338 : size_t seg_len;
339 : uint64_t reg;
340 :
341 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
342 0 : DEBUG_PRINT("invalid usermode virtual address %jx\n", vaddr);
343 0 : return -EINVAL;
344 : }
345 :
346 0 : if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
347 0 : DEBUG_PRINT("invalid %s parameters, vaddr=%jx len=%ju\n",
348 : __func__, vaddr, len);
349 0 : return -EINVAL;
350 : }
351 :
352 0 : if (len == 0) {
353 0 : return 0;
354 : }
355 :
356 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
357 :
358 0 : seg_vaddr = vaddr;
359 0 : seg_len = len;
360 0 : while (seg_len > 0) {
361 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
362 0 : if (reg & REG_MAP_REGISTERED) {
363 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
364 0 : return -EBUSY;
365 : }
366 0 : seg_vaddr += VALUE_2MB;
367 0 : seg_len -= VALUE_2MB;
368 : }
369 :
370 0 : seg_vaddr = vaddr;
371 0 : seg_len = 0;
372 0 : while (len > 0) {
373 0 : spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, VALUE_2MB,
374 : seg_len == 0 ? REG_MAP_REGISTERED | REG_MAP_NOTIFY_START : REG_MAP_REGISTERED);
375 0 : seg_len += VALUE_2MB;
376 0 : vaddr += VALUE_2MB;
377 0 : len -= VALUE_2MB;
378 : }
379 :
380 0 : TAILQ_FOREACH(map, &g_spdk_mem_maps, tailq) {
381 0 : rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_REGISTER,
382 : (void *)seg_vaddr, seg_len);
383 0 : if (rc != 0) {
384 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
385 0 : return rc;
386 : }
387 : }
388 :
389 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
390 0 : return 0;
391 : }
392 :
393 : int
394 0 : spdk_mem_unregister(void *_vaddr, size_t len)
395 : {
396 : struct spdk_mem_map *map;
397 : int rc;
398 0 : uint64_t vaddr = (uintptr_t)_vaddr;
399 : uint64_t seg_vaddr;
400 : size_t seg_len;
401 : uint64_t reg, newreg;
402 :
403 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
404 0 : DEBUG_PRINT("invalid usermode virtual address %jx\n", vaddr);
405 0 : return -EINVAL;
406 : }
407 :
408 0 : if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
409 0 : DEBUG_PRINT("invalid %s parameters, vaddr=%jx len=%ju\n",
410 : __func__, vaddr, len);
411 0 : return -EINVAL;
412 : }
413 :
414 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
415 :
416 : /* The first page must be a start of a region. Also check if it's
417 : * registered to make sure we don't return -ERANGE for non-registered
418 : * regions.
419 : */
420 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)vaddr, NULL);
421 0 : if ((reg & REG_MAP_REGISTERED) && (reg & REG_MAP_NOTIFY_START) == 0) {
422 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
423 0 : return -ERANGE;
424 : }
425 :
426 0 : seg_vaddr = vaddr;
427 0 : seg_len = len;
428 0 : while (seg_len > 0) {
429 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
430 0 : if ((reg & REG_MAP_REGISTERED) == 0) {
431 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
432 0 : return -EINVAL;
433 : }
434 0 : seg_vaddr += VALUE_2MB;
435 0 : seg_len -= VALUE_2MB;
436 : }
437 :
438 0 : newreg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
439 : /* If the next page is registered, it must be a start of a region as well,
440 : * otherwise we'd be unregistering only a part of a region.
441 : */
442 0 : if ((newreg & REG_MAP_NOTIFY_START) == 0 && (newreg & REG_MAP_REGISTERED)) {
443 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
444 0 : return -ERANGE;
445 : }
446 0 : seg_vaddr = vaddr;
447 0 : seg_len = 0;
448 :
449 0 : while (len > 0) {
450 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)vaddr, NULL);
451 0 : spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, VALUE_2MB, 0);
452 :
453 0 : if (seg_len > 0 && (reg & REG_MAP_NOTIFY_START)) {
454 0 : TAILQ_FOREACH_REVERSE(map, &g_spdk_mem_maps, spdk_mem_map_head, tailq) {
455 0 : rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_UNREGISTER,
456 : (void *)seg_vaddr, seg_len);
457 0 : if (rc != 0) {
458 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
459 0 : return rc;
460 : }
461 : }
462 :
463 0 : seg_vaddr = vaddr;
464 0 : seg_len = VALUE_2MB;
465 : } else {
466 0 : seg_len += VALUE_2MB;
467 : }
468 :
469 0 : vaddr += VALUE_2MB;
470 0 : len -= VALUE_2MB;
471 : }
472 :
473 0 : if (seg_len > 0) {
474 0 : TAILQ_FOREACH_REVERSE(map, &g_spdk_mem_maps, spdk_mem_map_head, tailq) {
475 0 : rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_UNREGISTER,
476 : (void *)seg_vaddr, seg_len);
477 0 : if (rc != 0) {
478 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
479 0 : return rc;
480 : }
481 : }
482 : }
483 :
484 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
485 0 : return 0;
486 : }
487 :
488 : int
489 0 : spdk_mem_reserve(void *vaddr, size_t len)
490 : {
491 : struct spdk_mem_map *map;
492 : void *seg_vaddr;
493 : size_t seg_len;
494 : uint64_t reg;
495 :
496 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
497 0 : DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
498 0 : return -EINVAL;
499 : }
500 :
501 0 : if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
502 0 : DEBUG_PRINT("invalid %s parameters, vaddr=%p len=%ju\n",
503 : __func__, vaddr, len);
504 0 : return -EINVAL;
505 : }
506 :
507 0 : if (len == 0) {
508 0 : return 0;
509 : }
510 :
511 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
512 :
513 : /* Check if any part of this range is already registered */
514 0 : seg_vaddr = vaddr;
515 0 : seg_len = len;
516 0 : while (seg_len > 0) {
517 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
518 0 : if (reg & REG_MAP_REGISTERED) {
519 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
520 0 : return -EBUSY;
521 : }
522 0 : seg_vaddr += VALUE_2MB;
523 0 : seg_len -= VALUE_2MB;
524 : }
525 :
526 : /* Simply set the translation to the memory map's default. This allocates the space in the
527 : * map but does not provide a valid translation. */
528 0 : spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, len,
529 0 : g_mem_reg_map->default_translation);
530 :
531 0 : TAILQ_FOREACH(map, &g_spdk_mem_maps, tailq) {
532 0 : spdk_mem_map_set_translation(map, (uint64_t)vaddr, len, map->default_translation);
533 : }
534 :
535 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
536 0 : return 0;
537 : }
538 :
539 : static struct map_1gb *
540 0 : mem_map_get_map_1gb(struct spdk_mem_map *map, uint64_t vfn_2mb)
541 : {
542 : struct map_1gb *map_1gb;
543 0 : uint64_t idx_256tb = MAP_256TB_IDX(vfn_2mb);
544 : size_t i;
545 :
546 0 : if (spdk_unlikely(idx_256tb >= SPDK_COUNTOF(map->map_256tb.map))) {
547 0 : return NULL;
548 : }
549 :
550 0 : map_1gb = map->map_256tb.map[idx_256tb];
551 :
552 0 : if (!map_1gb) {
553 0 : pthread_mutex_lock(&map->mutex);
554 :
555 : /* Recheck to make sure nobody else got the mutex first. */
556 0 : map_1gb = map->map_256tb.map[idx_256tb];
557 0 : if (!map_1gb) {
558 0 : map_1gb = malloc(sizeof(struct map_1gb));
559 0 : if (map_1gb) {
560 : /* initialize all entries to default translation */
561 0 : for (i = 0; i < SPDK_COUNTOF(map_1gb->map); i++) {
562 0 : map_1gb->map[i].translation_2mb = map->default_translation;
563 : }
564 0 : map->map_256tb.map[idx_256tb] = map_1gb;
565 : }
566 : }
567 :
568 0 : pthread_mutex_unlock(&map->mutex);
569 :
570 0 : if (!map_1gb) {
571 0 : DEBUG_PRINT("allocation failed\n");
572 0 : return NULL;
573 : }
574 : }
575 :
576 0 : return map_1gb;
577 : }
578 :
579 : int
580 0 : spdk_mem_map_set_translation(struct spdk_mem_map *map, uint64_t vaddr, uint64_t size,
581 : uint64_t translation)
582 : {
583 : uint64_t vfn_2mb;
584 : struct map_1gb *map_1gb;
585 : uint64_t idx_1gb;
586 : struct map_2mb *map_2mb;
587 :
588 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
589 0 : DEBUG_PRINT("invalid usermode virtual address %" PRIu64 "\n", vaddr);
590 0 : return -EINVAL;
591 : }
592 :
593 : /* For now, only 2 MB-aligned registrations are supported */
594 0 : if (((uintptr_t)vaddr & MASK_2MB) || (size & MASK_2MB)) {
595 0 : DEBUG_PRINT("invalid %s parameters, vaddr=%" PRIu64 " len=%" PRIu64 "\n",
596 : __func__, vaddr, size);
597 0 : return -EINVAL;
598 : }
599 :
600 0 : vfn_2mb = vaddr >> SHIFT_2MB;
601 :
602 0 : while (size) {
603 0 : map_1gb = mem_map_get_map_1gb(map, vfn_2mb);
604 0 : if (!map_1gb) {
605 0 : DEBUG_PRINT("could not get %p map\n", (void *)vaddr);
606 0 : return -ENOMEM;
607 : }
608 :
609 0 : idx_1gb = MAP_1GB_IDX(vfn_2mb);
610 0 : map_2mb = &map_1gb->map[idx_1gb];
611 0 : map_2mb->translation_2mb = translation;
612 :
613 0 : size -= VALUE_2MB;
614 0 : vfn_2mb++;
615 : }
616 :
617 0 : return 0;
618 : }
619 :
620 : int
621 0 : spdk_mem_map_clear_translation(struct spdk_mem_map *map, uint64_t vaddr, uint64_t size)
622 : {
623 0 : return spdk_mem_map_set_translation(map, vaddr, size, map->default_translation);
624 : }
625 :
626 : inline uint64_t
627 0 : spdk_mem_map_translate(const struct spdk_mem_map *map, uint64_t vaddr, uint64_t *size)
628 : {
629 : const struct map_1gb *map_1gb;
630 : const struct map_2mb *map_2mb;
631 : uint64_t idx_256tb;
632 : uint64_t idx_1gb;
633 : uint64_t vfn_2mb;
634 : uint64_t cur_size;
635 : uint64_t prev_translation;
636 : uint64_t orig_translation;
637 :
638 0 : if (spdk_unlikely(vaddr & ~MASK_256TB)) {
639 0 : DEBUG_PRINT("invalid usermode virtual address %p\n", (void *)vaddr);
640 0 : return map->default_translation;
641 : }
642 :
643 0 : vfn_2mb = vaddr >> SHIFT_2MB;
644 0 : idx_256tb = MAP_256TB_IDX(vfn_2mb);
645 0 : idx_1gb = MAP_1GB_IDX(vfn_2mb);
646 :
647 0 : map_1gb = map->map_256tb.map[idx_256tb];
648 0 : if (spdk_unlikely(!map_1gb)) {
649 0 : return map->default_translation;
650 : }
651 :
652 0 : cur_size = VALUE_2MB - _2MB_OFFSET(vaddr);
653 0 : map_2mb = &map_1gb->map[idx_1gb];
654 0 : if (size == NULL || map->ops.are_contiguous == NULL ||
655 0 : map_2mb->translation_2mb == map->default_translation) {
656 0 : if (size != NULL) {
657 0 : *size = spdk_min(*size, cur_size);
658 : }
659 0 : return map_2mb->translation_2mb;
660 : }
661 :
662 0 : orig_translation = map_2mb->translation_2mb;
663 0 : prev_translation = orig_translation;
664 0 : while (cur_size < *size) {
665 0 : vfn_2mb++;
666 0 : idx_256tb = MAP_256TB_IDX(vfn_2mb);
667 0 : idx_1gb = MAP_1GB_IDX(vfn_2mb);
668 :
669 0 : map_1gb = map->map_256tb.map[idx_256tb];
670 0 : if (spdk_unlikely(!map_1gb)) {
671 0 : break;
672 : }
673 :
674 0 : map_2mb = &map_1gb->map[idx_1gb];
675 0 : if (!map->ops.are_contiguous(prev_translation, map_2mb->translation_2mb)) {
676 0 : break;
677 : }
678 :
679 0 : cur_size += VALUE_2MB;
680 0 : prev_translation = map_2mb->translation_2mb;
681 : }
682 :
683 0 : *size = spdk_min(*size, cur_size);
684 0 : return orig_translation;
685 : }
686 :
687 : static void
688 0 : memory_hotplug_cb(enum rte_mem_event event_type,
689 : const void *addr, size_t len, void *arg)
690 : {
691 0 : if (event_type == RTE_MEM_EVENT_ALLOC) {
692 0 : spdk_mem_register((void *)addr, len);
693 :
694 0 : if (!spdk_env_dpdk_external_init()) {
695 0 : return;
696 : }
697 :
698 : /* When the user initialized DPDK separately, we can't
699 : * be sure that --match-allocations RTE flag was specified.
700 : * Without this flag, DPDK can free memory in different units
701 : * than it was allocated. It doesn't work with things like RDMA MRs.
702 : *
703 : * For such cases, we mark segments so they aren't freed.
704 : */
705 0 : while (len > 0) {
706 : struct rte_memseg *seg;
707 :
708 0 : seg = rte_mem_virt2memseg(addr, NULL);
709 0 : assert(seg != NULL);
710 0 : seg->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
711 0 : addr = (void *)((uintptr_t)addr + seg->hugepage_sz);
712 0 : len -= seg->hugepage_sz;
713 : }
714 0 : } else if (event_type == RTE_MEM_EVENT_FREE) {
715 0 : spdk_mem_unregister((void *)addr, len);
716 : }
717 : }
718 :
719 : static int
720 0 : memory_iter_cb(const struct rte_memseg_list *msl,
721 : const struct rte_memseg *ms, size_t len, void *arg)
722 : {
723 0 : return spdk_mem_register(ms->addr, len);
724 : }
725 :
726 : int
727 0 : mem_map_init(bool legacy_mem)
728 : {
729 0 : g_legacy_mem = legacy_mem;
730 :
731 0 : g_mem_reg_map = spdk_mem_map_alloc(0, NULL, NULL);
732 0 : if (g_mem_reg_map == NULL) {
733 0 : DEBUG_PRINT("memory registration map allocation failed\n");
734 0 : return -ENOMEM;
735 : }
736 :
737 : /*
738 : * Walk all DPDK memory segments and register them
739 : * with the main memory map
740 : */
741 0 : if (g_huge_pages) {
742 0 : rte_mem_event_callback_register("spdk", memory_hotplug_cb, NULL);
743 0 : rte_memseg_contig_walk(memory_iter_cb, NULL);
744 : }
745 0 : return 0;
746 : }
747 :
748 : bool
749 0 : spdk_iommu_is_enabled(void)
750 : {
751 : #if VFIO_ENABLED
752 0 : return g_vfio.enabled && !g_vfio.noiommu_enabled;
753 : #else
754 : return false;
755 : #endif
756 : }
757 :
758 : struct spdk_vtophys_pci_device {
759 : struct rte_pci_device *pci_device;
760 : TAILQ_ENTRY(spdk_vtophys_pci_device) tailq;
761 : };
762 :
763 : static pthread_mutex_t g_vtophys_pci_devices_mutex = PTHREAD_MUTEX_INITIALIZER;
764 : static TAILQ_HEAD(, spdk_vtophys_pci_device) g_vtophys_pci_devices =
765 : TAILQ_HEAD_INITIALIZER(g_vtophys_pci_devices);
766 :
767 : static struct spdk_mem_map *g_vtophys_map;
768 : static struct spdk_mem_map *g_phys_ref_map;
769 :
770 : #if VFIO_ENABLED
771 : static int
772 0 : _vfio_iommu_map_dma(uint64_t vaddr, uint64_t iova, uint64_t size)
773 : {
774 : struct spdk_vfio_dma_map *dma_map;
775 : int ret;
776 :
777 0 : dma_map = calloc(1, sizeof(*dma_map));
778 0 : if (dma_map == NULL) {
779 0 : return -ENOMEM;
780 : }
781 :
782 0 : dma_map->map.argsz = sizeof(dma_map->map);
783 0 : dma_map->map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
784 0 : dma_map->map.vaddr = vaddr;
785 0 : dma_map->map.iova = iova;
786 0 : dma_map->map.size = size;
787 :
788 0 : if (g_vfio.device_ref == 0) {
789 : /* VFIO requires at least one device (IOMMU group) to be added to
790 : * a VFIO container before it is possible to perform any IOMMU
791 : * operations on that container. This memory will be mapped once
792 : * the first device (IOMMU group) is hotplugged.
793 : *
794 : * Since the vfio container is managed internally by DPDK, it is
795 : * also possible that some device is already in that container, but
796 : * it's not managed by SPDK - e.g. an NIC attached internally
797 : * inside DPDK. We could map the memory straight away in such
798 : * scenario, but there's no need to do it. DPDK devices clearly
799 : * don't need our mappings and hence we defer the mapping
800 : * unconditionally until the first SPDK-managed device is
801 : * hotplugged.
802 : */
803 0 : goto out_insert;
804 : }
805 :
806 0 : ret = ioctl(g_vfio.fd, VFIO_IOMMU_MAP_DMA, &dma_map->map);
807 0 : if (ret) {
808 : /* There are cases the vfio container doesn't have IOMMU group, it's safe for this case */
809 0 : SPDK_NOTICELOG("Cannot set up DMA mapping, error %d, ignored\n", errno);
810 : }
811 :
812 0 : out_insert:
813 0 : TAILQ_INSERT_TAIL(&g_vfio.maps, dma_map, tailq);
814 0 : return 0;
815 : }
816 :
817 :
818 : static int
819 0 : vtophys_iommu_map_dma(uint64_t vaddr, uint64_t iova, uint64_t size)
820 : {
821 : uint64_t refcount;
822 : int ret;
823 :
824 0 : refcount = spdk_mem_map_translate(g_phys_ref_map, iova, NULL);
825 0 : assert(refcount < UINT64_MAX);
826 0 : if (refcount > 0) {
827 0 : spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount + 1);
828 0 : return 0;
829 : }
830 :
831 0 : pthread_mutex_lock(&g_vfio.mutex);
832 0 : ret = _vfio_iommu_map_dma(vaddr, iova, size);
833 0 : pthread_mutex_unlock(&g_vfio.mutex);
834 0 : if (ret) {
835 0 : return ret;
836 : }
837 :
838 0 : spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount + 1);
839 0 : return 0;
840 : }
841 :
842 : int
843 0 : vtophys_iommu_map_dma_bar(uint64_t vaddr, uint64_t iova, uint64_t size)
844 : {
845 : int ret;
846 :
847 0 : pthread_mutex_lock(&g_vfio.mutex);
848 0 : ret = _vfio_iommu_map_dma(vaddr, iova, size);
849 0 : pthread_mutex_unlock(&g_vfio.mutex);
850 :
851 0 : return ret;
852 : }
853 :
854 : static int
855 0 : _vfio_iommu_unmap_dma(struct spdk_vfio_dma_map *dma_map)
856 : {
857 0 : struct vfio_iommu_type1_dma_unmap unmap = {};
858 : int ret;
859 :
860 0 : if (g_vfio.device_ref == 0) {
861 : /* Memory is not mapped anymore, just remove it's references */
862 0 : goto out_remove;
863 : }
864 :
865 0 : unmap.argsz = sizeof(unmap);
866 0 : unmap.flags = 0;
867 0 : unmap.iova = dma_map->map.iova;
868 0 : unmap.size = dma_map->map.size;
869 0 : ret = ioctl(g_vfio.fd, VFIO_IOMMU_UNMAP_DMA, &unmap);
870 0 : if (ret) {
871 0 : SPDK_NOTICELOG("Cannot clear DMA mapping, error %d, ignored\n", errno);
872 : }
873 :
874 0 : out_remove:
875 0 : TAILQ_REMOVE(&g_vfio.maps, dma_map, tailq);
876 0 : free(dma_map);
877 0 : return 0;
878 : }
879 :
880 : static int
881 0 : vtophys_iommu_unmap_dma(uint64_t iova, uint64_t size)
882 : {
883 : struct spdk_vfio_dma_map *dma_map;
884 : uint64_t refcount;
885 : int ret;
886 :
887 0 : pthread_mutex_lock(&g_vfio.mutex);
888 0 : TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
889 0 : if (dma_map->map.iova == iova) {
890 0 : break;
891 : }
892 : }
893 :
894 0 : if (dma_map == NULL) {
895 0 : DEBUG_PRINT("Cannot clear DMA mapping for IOVA %"PRIx64" - it's not mapped\n", iova);
896 0 : pthread_mutex_unlock(&g_vfio.mutex);
897 0 : return -ENXIO;
898 : }
899 :
900 0 : refcount = spdk_mem_map_translate(g_phys_ref_map, iova, NULL);
901 0 : assert(refcount < UINT64_MAX);
902 0 : if (refcount > 0) {
903 0 : spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount - 1);
904 : }
905 :
906 : /* We still have outstanding references, don't clear it. */
907 0 : if (refcount > 1) {
908 0 : pthread_mutex_unlock(&g_vfio.mutex);
909 0 : return 0;
910 : }
911 :
912 : /** don't support partial or multiple-page unmap for now */
913 0 : assert(dma_map->map.size == size);
914 :
915 0 : ret = _vfio_iommu_unmap_dma(dma_map);
916 0 : pthread_mutex_unlock(&g_vfio.mutex);
917 :
918 0 : return ret;
919 : }
920 :
921 : int
922 0 : vtophys_iommu_unmap_dma_bar(uint64_t vaddr)
923 : {
924 : struct spdk_vfio_dma_map *dma_map;
925 : int ret;
926 :
927 0 : pthread_mutex_lock(&g_vfio.mutex);
928 0 : TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
929 0 : if (dma_map->map.vaddr == vaddr) {
930 0 : break;
931 : }
932 : }
933 :
934 0 : if (dma_map == NULL) {
935 0 : DEBUG_PRINT("Cannot clear DMA mapping for address %"PRIx64" - it's not mapped\n", vaddr);
936 0 : pthread_mutex_unlock(&g_vfio.mutex);
937 0 : return -ENXIO;
938 : }
939 :
940 0 : ret = _vfio_iommu_unmap_dma(dma_map);
941 0 : pthread_mutex_unlock(&g_vfio.mutex);
942 0 : return ret;
943 : }
944 : #endif
945 :
946 : static uint64_t
947 0 : vtophys_get_paddr_memseg(uint64_t vaddr)
948 : {
949 : uintptr_t paddr;
950 : struct rte_memseg *seg;
951 :
952 0 : seg = rte_mem_virt2memseg((void *)(uintptr_t)vaddr, NULL);
953 0 : if (seg != NULL) {
954 0 : paddr = seg->iova;
955 0 : if (paddr == RTE_BAD_IOVA) {
956 0 : return SPDK_VTOPHYS_ERROR;
957 : }
958 0 : paddr += (vaddr - (uintptr_t)seg->addr);
959 0 : return paddr;
960 : }
961 :
962 0 : return SPDK_VTOPHYS_ERROR;
963 : }
964 :
965 : /* Try to get the paddr from /proc/self/pagemap */
966 : static uint64_t
967 0 : vtophys_get_paddr_pagemap(uint64_t vaddr)
968 : {
969 : uintptr_t paddr;
970 :
971 : /* Silence static analyzers */
972 0 : assert(vaddr != 0);
973 0 : paddr = rte_mem_virt2iova((void *)vaddr);
974 0 : if (paddr == RTE_BAD_IOVA) {
975 : /*
976 : * The vaddr may be valid but doesn't have a backing page
977 : * assigned yet. Touch the page to ensure a backing page
978 : * gets assigned, then try to translate again.
979 : */
980 0 : rte_atomic64_read((rte_atomic64_t *)vaddr);
981 0 : paddr = rte_mem_virt2iova((void *)vaddr);
982 : }
983 0 : if (paddr == RTE_BAD_IOVA) {
984 : /* Unable to get to the physical address. */
985 0 : return SPDK_VTOPHYS_ERROR;
986 : }
987 :
988 0 : return paddr;
989 : }
990 :
991 : static uint64_t
992 0 : pci_device_vtophys(struct rte_pci_device *dev, uint64_t vaddr, size_t len)
993 : {
994 : struct rte_mem_resource *res;
995 : uint64_t paddr;
996 : unsigned r;
997 :
998 0 : for (r = 0; r < PCI_MAX_RESOURCE; r++) {
999 0 : res = dpdk_pci_device_get_mem_resource(dev, r);
1000 :
1001 0 : if (res->phys_addr == 0 || vaddr < (uint64_t)res->addr ||
1002 0 : (vaddr + len) >= (uint64_t)res->addr + res->len) {
1003 0 : continue;
1004 : }
1005 :
1006 : #if VFIO_ENABLED
1007 0 : if (spdk_iommu_is_enabled() && rte_eal_iova_mode() == RTE_IOVA_VA) {
1008 : /*
1009 : * The IOMMU is on and we're using IOVA == VA. The BAR was
1010 : * automatically registered when it was mapped, so just return
1011 : * the virtual address here.
1012 : */
1013 0 : return vaddr;
1014 : }
1015 : #endif
1016 0 : paddr = res->phys_addr + (vaddr - (uint64_t)res->addr);
1017 0 : return paddr;
1018 : }
1019 :
1020 0 : return SPDK_VTOPHYS_ERROR;
1021 : }
1022 :
1023 : /* Try to get the paddr from pci devices */
1024 : static uint64_t
1025 0 : vtophys_get_paddr_pci(uint64_t vaddr, size_t len)
1026 : {
1027 : struct spdk_vtophys_pci_device *vtophys_dev;
1028 : uintptr_t paddr;
1029 : struct rte_pci_device *dev;
1030 :
1031 0 : pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
1032 0 : TAILQ_FOREACH(vtophys_dev, &g_vtophys_pci_devices, tailq) {
1033 0 : dev = vtophys_dev->pci_device;
1034 0 : paddr = pci_device_vtophys(dev, vaddr, len);
1035 0 : if (paddr != SPDK_VTOPHYS_ERROR) {
1036 0 : pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
1037 0 : return paddr;
1038 : }
1039 : }
1040 0 : pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
1041 :
1042 0 : return SPDK_VTOPHYS_ERROR;
1043 : }
1044 :
1045 : static int
1046 0 : vtophys_notify(void *cb_ctx, struct spdk_mem_map *map,
1047 : enum spdk_mem_map_notify_action action,
1048 : void *vaddr, size_t len)
1049 : {
1050 0 : int rc = 0;
1051 : uint64_t paddr;
1052 :
1053 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
1054 0 : DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
1055 0 : return -EINVAL;
1056 : }
1057 :
1058 0 : if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
1059 0 : DEBUG_PRINT("invalid parameters, vaddr=%p len=%ju\n",
1060 : vaddr, len);
1061 0 : return -EINVAL;
1062 : }
1063 :
1064 : /* Get the physical address from the DPDK memsegs */
1065 0 : paddr = vtophys_get_paddr_memseg((uint64_t)vaddr);
1066 :
1067 0 : switch (action) {
1068 0 : case SPDK_MEM_MAP_NOTIFY_REGISTER:
1069 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1070 : /* This is not an address that DPDK is managing. */
1071 :
1072 : /* Check if this is a PCI BAR. They need special handling */
1073 0 : paddr = vtophys_get_paddr_pci((uint64_t)vaddr, len);
1074 0 : if (paddr != SPDK_VTOPHYS_ERROR) {
1075 : /* Get paddr for each 2MB chunk in this address range */
1076 0 : while (len > 0) {
1077 0 : paddr = vtophys_get_paddr_pci((uint64_t)vaddr, VALUE_2MB);
1078 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1079 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1080 0 : return -EFAULT;
1081 : }
1082 :
1083 0 : rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
1084 0 : if (rc != 0) {
1085 0 : return rc;
1086 : }
1087 :
1088 0 : vaddr += VALUE_2MB;
1089 0 : len -= VALUE_2MB;
1090 : }
1091 :
1092 0 : return 0;
1093 : }
1094 :
1095 : #if VFIO_ENABLED
1096 : enum rte_iova_mode iova_mode;
1097 :
1098 0 : iova_mode = rte_eal_iova_mode();
1099 :
1100 0 : if (spdk_iommu_is_enabled() && iova_mode == RTE_IOVA_VA) {
1101 : /* We'll use the virtual address as the iova to match DPDK. */
1102 0 : paddr = (uint64_t)vaddr;
1103 0 : rc = vtophys_iommu_map_dma((uint64_t)vaddr, paddr, len);
1104 0 : if (rc) {
1105 0 : return -EFAULT;
1106 : }
1107 0 : while (len > 0) {
1108 0 : rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
1109 0 : if (rc != 0) {
1110 0 : return rc;
1111 : }
1112 0 : vaddr += VALUE_2MB;
1113 0 : paddr += VALUE_2MB;
1114 0 : len -= VALUE_2MB;
1115 : }
1116 : } else
1117 : #endif
1118 : {
1119 : /* Get the physical address from /proc/self/pagemap. */
1120 0 : paddr = vtophys_get_paddr_pagemap((uint64_t)vaddr);
1121 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1122 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1123 0 : return -EFAULT;
1124 : }
1125 :
1126 : /* Get paddr for each 2MB chunk in this address range */
1127 0 : while (len > 0) {
1128 : /* Get the physical address from /proc/self/pagemap. */
1129 0 : paddr = vtophys_get_paddr_pagemap((uint64_t)vaddr);
1130 :
1131 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1132 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1133 0 : return -EFAULT;
1134 : }
1135 :
1136 0 : if (paddr & MASK_2MB) {
1137 0 : DEBUG_PRINT("invalid paddr 0x%" PRIx64 " - must be 2MB aligned\n", paddr);
1138 0 : return -EINVAL;
1139 : }
1140 : #if VFIO_ENABLED
1141 : /* If the IOMMU is on, but DPDK is using iova-mode=pa, we want to register this memory
1142 : * with the IOMMU using the physical address to match. */
1143 0 : if (spdk_iommu_is_enabled()) {
1144 0 : rc = vtophys_iommu_map_dma((uint64_t)vaddr, paddr, VALUE_2MB);
1145 0 : if (rc) {
1146 0 : DEBUG_PRINT("Unable to assign vaddr %p to paddr 0x%" PRIx64 "\n", vaddr, paddr);
1147 0 : return -EFAULT;
1148 : }
1149 : }
1150 : #endif
1151 :
1152 0 : rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
1153 0 : if (rc != 0) {
1154 0 : return rc;
1155 : }
1156 :
1157 0 : vaddr += VALUE_2MB;
1158 0 : len -= VALUE_2MB;
1159 : }
1160 : }
1161 : } else {
1162 : /* This is an address managed by DPDK. Just setup the translations. */
1163 0 : while (len > 0) {
1164 0 : paddr = vtophys_get_paddr_memseg((uint64_t)vaddr);
1165 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1166 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1167 0 : return -EFAULT;
1168 : }
1169 :
1170 0 : rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
1171 0 : if (rc != 0) {
1172 0 : return rc;
1173 : }
1174 :
1175 0 : vaddr += VALUE_2MB;
1176 0 : len -= VALUE_2MB;
1177 : }
1178 : }
1179 :
1180 0 : break;
1181 0 : case SPDK_MEM_MAP_NOTIFY_UNREGISTER:
1182 : #if VFIO_ENABLED
1183 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1184 : /*
1185 : * This is not an address that DPDK is managing.
1186 : */
1187 :
1188 : /* Check if this is a PCI BAR. They need special handling */
1189 0 : paddr = vtophys_get_paddr_pci((uint64_t)vaddr, len);
1190 0 : if (paddr != SPDK_VTOPHYS_ERROR) {
1191 : /* Get paddr for each 2MB chunk in this address range */
1192 0 : while (len > 0) {
1193 0 : paddr = vtophys_get_paddr_pci((uint64_t)vaddr, VALUE_2MB);
1194 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1195 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1196 0 : return -EFAULT;
1197 : }
1198 :
1199 0 : rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, VALUE_2MB);
1200 0 : if (rc != 0) {
1201 0 : return rc;
1202 : }
1203 :
1204 0 : vaddr += VALUE_2MB;
1205 0 : len -= VALUE_2MB;
1206 : }
1207 :
1208 0 : return 0;
1209 : }
1210 :
1211 : /* If vfio is enabled,
1212 : * we need to unmap the range from the IOMMU
1213 : */
1214 0 : if (spdk_iommu_is_enabled()) {
1215 0 : uint64_t buffer_len = len;
1216 0 : uint8_t *va = vaddr;
1217 : enum rte_iova_mode iova_mode;
1218 :
1219 0 : iova_mode = rte_eal_iova_mode();
1220 : /*
1221 : * In virtual address mode, the region is contiguous and can be done in
1222 : * one unmap.
1223 : */
1224 0 : if (iova_mode == RTE_IOVA_VA) {
1225 0 : paddr = spdk_mem_map_translate(map, (uint64_t)va, &buffer_len);
1226 0 : if (buffer_len != len || paddr != (uintptr_t)va) {
1227 0 : DEBUG_PRINT("Unmapping %p with length %lu failed because "
1228 : "translation had address 0x%" PRIx64 " and length %lu\n",
1229 : va, len, paddr, buffer_len);
1230 0 : return -EINVAL;
1231 : }
1232 0 : rc = vtophys_iommu_unmap_dma(paddr, len);
1233 0 : if (rc) {
1234 0 : DEBUG_PRINT("Failed to iommu unmap paddr 0x%" PRIx64 "\n", paddr);
1235 0 : return -EFAULT;
1236 : }
1237 0 : } else if (iova_mode == RTE_IOVA_PA) {
1238 : /* Get paddr for each 2MB chunk in this address range */
1239 0 : while (buffer_len > 0) {
1240 0 : paddr = spdk_mem_map_translate(map, (uint64_t)va, NULL);
1241 :
1242 0 : if (paddr == SPDK_VTOPHYS_ERROR || buffer_len < VALUE_2MB) {
1243 0 : DEBUG_PRINT("could not get phys addr for %p\n", va);
1244 0 : return -EFAULT;
1245 : }
1246 :
1247 0 : rc = vtophys_iommu_unmap_dma(paddr, VALUE_2MB);
1248 0 : if (rc) {
1249 0 : DEBUG_PRINT("Failed to iommu unmap paddr 0x%" PRIx64 "\n", paddr);
1250 0 : return -EFAULT;
1251 : }
1252 :
1253 0 : va += VALUE_2MB;
1254 0 : buffer_len -= VALUE_2MB;
1255 : }
1256 : }
1257 : }
1258 : }
1259 : #endif
1260 0 : while (len > 0) {
1261 0 : rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, VALUE_2MB);
1262 0 : if (rc != 0) {
1263 0 : return rc;
1264 : }
1265 :
1266 0 : vaddr += VALUE_2MB;
1267 0 : len -= VALUE_2MB;
1268 : }
1269 :
1270 0 : break;
1271 0 : default:
1272 0 : SPDK_UNREACHABLE();
1273 : }
1274 :
1275 0 : return rc;
1276 : }
1277 :
1278 : static int
1279 0 : vtophys_check_contiguous_entries(uint64_t paddr1, uint64_t paddr2)
1280 : {
1281 : /* This function is always called with paddrs for two subsequent
1282 : * 2MB chunks in virtual address space, so those chunks will be only
1283 : * physically contiguous if the physical addresses are 2MB apart
1284 : * from each other as well.
1285 : */
1286 0 : return (paddr2 - paddr1 == VALUE_2MB);
1287 : }
1288 :
1289 : #if VFIO_ENABLED
1290 :
1291 : static bool
1292 0 : vfio_enabled(void)
1293 : {
1294 0 : return rte_vfio_is_enabled("vfio_pci");
1295 : }
1296 :
1297 : /* Check if IOMMU is enabled on the system */
1298 : static bool
1299 0 : has_iommu_groups(void)
1300 : {
1301 0 : int count = 0;
1302 0 : DIR *dir = opendir("/sys/kernel/iommu_groups");
1303 :
1304 0 : if (dir == NULL) {
1305 0 : return false;
1306 : }
1307 :
1308 0 : while (count < 3 && readdir(dir) != NULL) {
1309 0 : count++;
1310 : }
1311 :
1312 0 : closedir(dir);
1313 : /* there will always be ./ and ../ entries */
1314 0 : return count > 2;
1315 : }
1316 :
1317 : static bool
1318 0 : vfio_noiommu_enabled(void)
1319 : {
1320 0 : return rte_vfio_noiommu_is_enabled();
1321 : }
1322 :
1323 : static void
1324 0 : vtophys_iommu_init(void)
1325 : {
1326 0 : char proc_fd_path[PATH_MAX + 1];
1327 0 : char link_path[PATH_MAX + 1];
1328 0 : const char vfio_path[] = "/dev/vfio/vfio";
1329 : DIR *dir;
1330 : struct dirent *d;
1331 :
1332 0 : if (!vfio_enabled()) {
1333 0 : return;
1334 : }
1335 :
1336 0 : if (vfio_noiommu_enabled()) {
1337 0 : g_vfio.noiommu_enabled = true;
1338 0 : } else if (!has_iommu_groups()) {
1339 0 : return;
1340 : }
1341 :
1342 0 : dir = opendir("/proc/self/fd");
1343 0 : if (!dir) {
1344 0 : DEBUG_PRINT("Failed to open /proc/self/fd (%d)\n", errno);
1345 0 : return;
1346 : }
1347 :
1348 0 : while ((d = readdir(dir)) != NULL) {
1349 0 : if (d->d_type != DT_LNK) {
1350 0 : continue;
1351 : }
1352 :
1353 0 : snprintf(proc_fd_path, sizeof(proc_fd_path), "/proc/self/fd/%s", d->d_name);
1354 0 : if (readlink(proc_fd_path, link_path, sizeof(link_path)) != (sizeof(vfio_path) - 1)) {
1355 0 : continue;
1356 : }
1357 :
1358 0 : if (memcmp(link_path, vfio_path, sizeof(vfio_path) - 1) == 0) {
1359 0 : sscanf(d->d_name, "%d", &g_vfio.fd);
1360 0 : break;
1361 : }
1362 : }
1363 :
1364 0 : closedir(dir);
1365 :
1366 0 : if (g_vfio.fd < 0) {
1367 0 : DEBUG_PRINT("Failed to discover DPDK VFIO container fd.\n");
1368 0 : return;
1369 : }
1370 :
1371 0 : g_vfio.enabled = true;
1372 :
1373 0 : return;
1374 : }
1375 :
1376 : #endif
1377 :
1378 : void
1379 0 : vtophys_pci_device_added(struct rte_pci_device *pci_device)
1380 : {
1381 : struct spdk_vtophys_pci_device *vtophys_dev;
1382 :
1383 0 : pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
1384 :
1385 0 : vtophys_dev = calloc(1, sizeof(*vtophys_dev));
1386 0 : if (vtophys_dev) {
1387 0 : vtophys_dev->pci_device = pci_device;
1388 0 : TAILQ_INSERT_TAIL(&g_vtophys_pci_devices, vtophys_dev, tailq);
1389 : } else {
1390 0 : DEBUG_PRINT("Memory allocation error\n");
1391 : }
1392 0 : pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
1393 :
1394 : #if VFIO_ENABLED
1395 : struct spdk_vfio_dma_map *dma_map;
1396 : int ret;
1397 :
1398 0 : if (!g_vfio.enabled) {
1399 0 : return;
1400 : }
1401 :
1402 0 : pthread_mutex_lock(&g_vfio.mutex);
1403 0 : g_vfio.device_ref++;
1404 0 : if (g_vfio.device_ref > 1) {
1405 0 : pthread_mutex_unlock(&g_vfio.mutex);
1406 0 : return;
1407 : }
1408 :
1409 : /* This is the first SPDK device using DPDK vfio. This means that the first
1410 : * IOMMU group might have been just been added to the DPDK vfio container.
1411 : * From this point it is certain that the memory can be mapped now.
1412 : */
1413 0 : TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
1414 0 : ret = ioctl(g_vfio.fd, VFIO_IOMMU_MAP_DMA, &dma_map->map);
1415 0 : if (ret) {
1416 0 : DEBUG_PRINT("Cannot update DMA mapping, error %d\n", errno);
1417 0 : break;
1418 : }
1419 : }
1420 0 : pthread_mutex_unlock(&g_vfio.mutex);
1421 : #endif
1422 : }
1423 :
1424 : void
1425 0 : vtophys_pci_device_removed(struct rte_pci_device *pci_device)
1426 : {
1427 : struct spdk_vtophys_pci_device *vtophys_dev;
1428 :
1429 0 : pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
1430 0 : TAILQ_FOREACH(vtophys_dev, &g_vtophys_pci_devices, tailq) {
1431 0 : if (vtophys_dev->pci_device == pci_device) {
1432 0 : TAILQ_REMOVE(&g_vtophys_pci_devices, vtophys_dev, tailq);
1433 0 : free(vtophys_dev);
1434 0 : break;
1435 : }
1436 : }
1437 0 : pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
1438 :
1439 : #if VFIO_ENABLED
1440 : struct spdk_vfio_dma_map *dma_map;
1441 : int ret;
1442 :
1443 0 : if (!g_vfio.enabled) {
1444 0 : return;
1445 : }
1446 :
1447 0 : pthread_mutex_lock(&g_vfio.mutex);
1448 0 : assert(g_vfio.device_ref > 0);
1449 0 : g_vfio.device_ref--;
1450 0 : if (g_vfio.device_ref > 0) {
1451 0 : pthread_mutex_unlock(&g_vfio.mutex);
1452 0 : return;
1453 : }
1454 :
1455 : /* This is the last SPDK device using DPDK vfio. If DPDK doesn't have
1456 : * any additional devices using it's vfio container, all the mappings
1457 : * will be automatically removed by the Linux vfio driver. We unmap
1458 : * the memory manually to be able to easily re-map it later regardless
1459 : * of other, external factors.
1460 : */
1461 0 : TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
1462 0 : struct vfio_iommu_type1_dma_unmap unmap = {};
1463 0 : unmap.argsz = sizeof(unmap);
1464 0 : unmap.flags = 0;
1465 0 : unmap.iova = dma_map->map.iova;
1466 0 : unmap.size = dma_map->map.size;
1467 0 : ret = ioctl(g_vfio.fd, VFIO_IOMMU_UNMAP_DMA, &unmap);
1468 0 : if (ret) {
1469 0 : DEBUG_PRINT("Cannot unmap DMA memory, error %d\n", errno);
1470 0 : break;
1471 : }
1472 : }
1473 0 : pthread_mutex_unlock(&g_vfio.mutex);
1474 : #endif
1475 : }
1476 :
1477 : int
1478 0 : vtophys_init(void)
1479 : {
1480 0 : const struct spdk_mem_map_ops vtophys_map_ops = {
1481 : .notify_cb = vtophys_notify,
1482 : .are_contiguous = vtophys_check_contiguous_entries,
1483 : };
1484 :
1485 0 : const struct spdk_mem_map_ops phys_ref_map_ops = {
1486 : .notify_cb = NULL,
1487 : .are_contiguous = NULL,
1488 : };
1489 :
1490 : #if VFIO_ENABLED
1491 0 : vtophys_iommu_init();
1492 : #endif
1493 :
1494 0 : g_phys_ref_map = spdk_mem_map_alloc(0, &phys_ref_map_ops, NULL);
1495 0 : if (g_phys_ref_map == NULL) {
1496 0 : DEBUG_PRINT("phys_ref map allocation failed.\n");
1497 0 : return -ENOMEM;
1498 : }
1499 :
1500 0 : if (g_huge_pages) {
1501 0 : g_vtophys_map = spdk_mem_map_alloc(SPDK_VTOPHYS_ERROR, &vtophys_map_ops, NULL);
1502 0 : if (g_vtophys_map == NULL) {
1503 0 : DEBUG_PRINT("vtophys map allocation failed\n");
1504 0 : spdk_mem_map_free(&g_phys_ref_map);
1505 0 : return -ENOMEM;
1506 : }
1507 : }
1508 0 : return 0;
1509 : }
1510 :
1511 : uint64_t
1512 0 : spdk_vtophys(const void *buf, uint64_t *size)
1513 : {
1514 : uint64_t vaddr, paddr_2mb;
1515 :
1516 0 : if (!g_huge_pages) {
1517 0 : return SPDK_VTOPHYS_ERROR;
1518 : }
1519 :
1520 0 : vaddr = (uint64_t)buf;
1521 0 : paddr_2mb = spdk_mem_map_translate(g_vtophys_map, vaddr, size);
1522 :
1523 : /*
1524 : * SPDK_VTOPHYS_ERROR has all bits set, so if the lookup returned SPDK_VTOPHYS_ERROR,
1525 : * we will still bitwise-or it with the buf offset below, but the result will still be
1526 : * SPDK_VTOPHYS_ERROR. However now that we do + rather than | (due to PCI vtophys being
1527 : * unaligned) we must now check the return value before addition.
1528 : */
1529 : SPDK_STATIC_ASSERT(SPDK_VTOPHYS_ERROR == UINT64_C(-1), "SPDK_VTOPHYS_ERROR should be all 1s");
1530 0 : if (paddr_2mb == SPDK_VTOPHYS_ERROR) {
1531 0 : return SPDK_VTOPHYS_ERROR;
1532 : } else {
1533 0 : return paddr_2mb + (vaddr & MASK_2MB);
1534 : }
1535 : }
1536 :
1537 : int
1538 0 : spdk_mem_get_fd_and_offset(void *vaddr, uint64_t *offset)
1539 : {
1540 : struct rte_memseg *seg;
1541 : int ret, fd;
1542 :
1543 0 : seg = rte_mem_virt2memseg(vaddr, NULL);
1544 0 : if (!seg) {
1545 0 : SPDK_ERRLOG("memory %p doesn't exist\n", vaddr);
1546 0 : return -ENOENT;
1547 : }
1548 :
1549 0 : fd = rte_memseg_get_fd_thread_unsafe(seg);
1550 0 : if (fd < 0) {
1551 0 : return fd;
1552 : }
1553 :
1554 0 : ret = rte_memseg_get_fd_offset_thread_unsafe(seg, offset);
1555 0 : if (ret < 0) {
1556 0 : return ret;
1557 : }
1558 :
1559 0 : return fd;
1560 : }
1561 :
1562 : void
1563 0 : mem_disable_huge_pages(void)
1564 : {
1565 0 : g_huge_pages = false;
1566 0 : }
|