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 : struct map_page_cfg {
65 : uint64_t shift;
66 : uint64_t size;
67 : uint64_t mask;
68 : uint64_t num_pages_per_gb;
69 : };
70 :
71 : /**
72 : * g_map_page_cfg can not be static because it is used in the inline function
73 : * spdk_mem_map_translate.
74 : */
75 : struct map_page_cfg g_map_page_cfg = {
76 : .shift = SHIFT_2MB,
77 : .size = VALUE_2MB,
78 : .mask = MASK_2MB,
79 : .num_pages_per_gb = 1UL << (SHIFT_1GB - SHIFT_2MB),
80 : };
81 :
82 : #define MAP_PAGE_SHIFT (g_map_page_cfg.shift)
83 : #define MAP_PAGE_SIZE (g_map_page_cfg.size)
84 : #define MAP_PAGE_MASK (g_map_page_cfg.mask)
85 : #define MAP_NUM_PAGES_PER_GB (g_map_page_cfg.num_pages_per_gb)
86 :
87 : #define MAP_256TB_IDX(vfn_page) ((vfn_page) >> (SHIFT_1GB - MAP_PAGE_SHIFT))
88 : #define MAP_1GB_IDX(vfn_page) ((vfn_page) & ((1ULL << (SHIFT_1GB - MAP_PAGE_SHIFT)) - 1))
89 : #define MAP_PAGE_OFFSET(ptr) (((uintptr_t)(ptr)) & MAP_PAGE_MASK)
90 :
91 : /* Page is registered */
92 : #define REG_MAP_REGISTERED (1ULL << 62)
93 :
94 : /* A notification region barrier. The page translation entry that's marked
95 : * with this flag must be unregistered separately. This allows contiguous
96 : * regions to be unregistered in the same chunks they were registered.
97 : */
98 : #define REG_MAP_NOTIFY_START (1ULL << 63)
99 :
100 : /* Translation of a single page. */
101 : struct map_page {
102 : uint64_t translation;
103 : };
104 :
105 : /* Second-level map table indexed by bits [page_shift..29] of the virtual address.
106 : * Each entry contains the address translation or error for entries that haven't
107 : * been retrieved yet.
108 : */
109 : struct map_1gb {
110 : struct map_page map[0];
111 : /**
112 : * Page table space.
113 : * Do not put any fields after this!
114 : */
115 : };
116 :
117 : #define MAP_SIZE_OF_MAP_1GB (sizeof(struct map_1gb) + MAP_NUM_PAGES_PER_GB * sizeof(struct map_page))
118 :
119 : /* Top-level map table indexed by bits [30..47] of the virtual address.
120 : * Each entry points to a second-level map table or NULL.
121 : */
122 : struct map_256tb {
123 : struct map_1gb *map[1ULL << (SHIFT_256TB - SHIFT_1GB)];
124 : };
125 :
126 : /* Page-granularity memory address translation */
127 : struct spdk_mem_map {
128 : struct map_256tb map_256tb;
129 : pthread_mutex_t mutex;
130 : uint64_t default_translation;
131 : struct spdk_mem_map_ops ops;
132 : void *cb_ctx;
133 : TAILQ_ENTRY(spdk_mem_map) tailq;
134 : };
135 :
136 : /* Registrations map. The 64 bit translations are bit fields with the
137 : * following layout (starting with the low bits):
138 : * 0 - 61 : reserved
139 : * 62 - 63 : flags
140 : */
141 : static struct spdk_mem_map *g_mem_reg_map;
142 : static TAILQ_HEAD(spdk_mem_map_head, spdk_mem_map) g_spdk_mem_maps =
143 : TAILQ_HEAD_INITIALIZER(g_spdk_mem_maps);
144 : static pthread_mutex_t g_spdk_mem_map_mutex = PTHREAD_MUTEX_INITIALIZER;
145 :
146 : static bool g_legacy_mem;
147 : static bool g_huge_pages = true;
148 : static bool g_mem_event_cb_registered = false;
149 :
150 : /*
151 : * Walk the currently registered memory via the main memory registration map
152 : * and call the new map's notify callback for each virtually contiguous region.
153 : */
154 : static int
155 0 : mem_map_notify_walk(struct spdk_mem_map *map, enum spdk_mem_map_notify_action action)
156 : {
157 : size_t idx_256tb;
158 : uint64_t idx_1gb;
159 0 : uint64_t contig_start = UINT64_MAX;
160 0 : uint64_t contig_end = UINT64_MAX;
161 : struct map_1gb *map_1gb;
162 : int rc;
163 :
164 0 : if (!g_mem_reg_map) {
165 0 : return -EINVAL;
166 : }
167 :
168 : /* Hold the memory registration map mutex so no new registrations can be added while we are looping. */
169 0 : pthread_mutex_lock(&g_mem_reg_map->mutex);
170 :
171 0 : for (idx_256tb = 0;
172 0 : idx_256tb < sizeof(g_mem_reg_map->map_256tb.map) / sizeof(g_mem_reg_map->map_256tb.map[0]);
173 0 : idx_256tb++) {
174 0 : map_1gb = g_mem_reg_map->map_256tb.map[idx_256tb];
175 :
176 0 : if (!map_1gb) {
177 0 : if (contig_start != UINT64_MAX) {
178 : /* End of of a virtually contiguous range */
179 0 : rc = map->ops.notify_cb(map->cb_ctx, map, action,
180 0 : (void *)contig_start,
181 0 : contig_end - contig_start + MAP_PAGE_SIZE);
182 : /* Don't bother handling unregister failures. It can't be any worse */
183 0 : if (rc != 0 && action == SPDK_MEM_MAP_NOTIFY_REGISTER) {
184 0 : goto err_unregister;
185 : }
186 0 : }
187 0 : contig_start = UINT64_MAX;
188 0 : continue;
189 : }
190 :
191 0 : for (idx_1gb = 0; idx_1gb < MAP_NUM_PAGES_PER_GB; idx_1gb++) {
192 0 : if ((map_1gb->map[idx_1gb].translation & REG_MAP_REGISTERED) &&
193 0 : (contig_start == UINT64_MAX ||
194 0 : (map_1gb->map[idx_1gb].translation & REG_MAP_NOTIFY_START) == 0)) {
195 : /* Rebuild the virtual address from the indexes */
196 0 : uint64_t vaddr = (idx_256tb << SHIFT_1GB) | (idx_1gb << MAP_PAGE_SHIFT);
197 :
198 0 : if (contig_start == UINT64_MAX) {
199 0 : contig_start = vaddr;
200 0 : }
201 :
202 0 : contig_end = vaddr;
203 0 : } else {
204 0 : if (contig_start != UINT64_MAX) {
205 : /* End of of a virtually contiguous range */
206 0 : rc = map->ops.notify_cb(map->cb_ctx, map, action,
207 0 : (void *)contig_start,
208 0 : contig_end - contig_start + MAP_PAGE_SIZE);
209 : /* Don't bother handling unregister failures. It can't be any worse */
210 0 : if (rc != 0 && action == SPDK_MEM_MAP_NOTIFY_REGISTER) {
211 0 : goto err_unregister;
212 : }
213 :
214 : /* This page might be a part of a neighbour region, so process
215 : * it again. The idx_1gb will be incremented immediately.
216 : */
217 0 : idx_1gb--;
218 0 : }
219 0 : contig_start = UINT64_MAX;
220 : }
221 0 : }
222 0 : }
223 :
224 0 : pthread_mutex_unlock(&g_mem_reg_map->mutex);
225 0 : return 0;
226 :
227 : err_unregister:
228 : /* Unwind to the first empty translation so we don't unregister
229 : * a region that just failed to register.
230 : */
231 0 : idx_256tb = MAP_256TB_IDX((contig_start >> MAP_PAGE_SHIFT) - 1);
232 0 : idx_1gb = MAP_1GB_IDX((contig_start >> MAP_PAGE_SHIFT) - 1);
233 0 : contig_start = UINT64_MAX;
234 0 : contig_end = UINT64_MAX;
235 :
236 : /* Unregister any memory we managed to register before the failure */
237 0 : for (; idx_256tb < SIZE_MAX; idx_256tb--) {
238 0 : map_1gb = g_mem_reg_map->map_256tb.map[idx_256tb];
239 :
240 0 : if (!map_1gb) {
241 0 : if (contig_end != UINT64_MAX) {
242 : /* End of of a virtually contiguous range */
243 0 : map->ops.notify_cb(map->cb_ctx, map,
244 : SPDK_MEM_MAP_NOTIFY_UNREGISTER,
245 0 : (void *)contig_start,
246 0 : contig_end - contig_start + MAP_PAGE_SIZE);
247 0 : }
248 0 : contig_end = UINT64_MAX;
249 0 : continue;
250 : }
251 :
252 0 : for (; idx_1gb < UINT64_MAX; idx_1gb--) {
253 : /* Rebuild the virtual address from the indexes */
254 0 : uint64_t vaddr = (idx_256tb << SHIFT_1GB) | (idx_1gb << MAP_PAGE_SHIFT);
255 0 : if ((map_1gb->map[idx_1gb].translation & REG_MAP_REGISTERED) &&
256 0 : (contig_end == UINT64_MAX || (map_1gb->map[idx_1gb].translation & REG_MAP_NOTIFY_START) == 0)) {
257 :
258 0 : if (contig_end == UINT64_MAX) {
259 0 : contig_end = vaddr;
260 0 : }
261 0 : contig_start = vaddr;
262 0 : } else {
263 0 : if (contig_end != UINT64_MAX) {
264 0 : if (map_1gb->map[idx_1gb].translation & REG_MAP_NOTIFY_START) {
265 0 : contig_start = vaddr;
266 0 : }
267 : /* End of of a virtually contiguous range */
268 0 : map->ops.notify_cb(map->cb_ctx, map,
269 : SPDK_MEM_MAP_NOTIFY_UNREGISTER,
270 0 : (void *)contig_start,
271 0 : contig_end - contig_start + MAP_PAGE_SIZE);
272 0 : }
273 0 : contig_end = UINT64_MAX;
274 : }
275 0 : }
276 0 : idx_1gb = MAP_NUM_PAGES_PER_GB - 1;
277 0 : }
278 :
279 0 : pthread_mutex_unlock(&g_mem_reg_map->mutex);
280 0 : return rc;
281 0 : }
282 :
283 : struct spdk_mem_map *
284 0 : spdk_mem_map_alloc(uint64_t default_translation, const struct spdk_mem_map_ops *ops, void *cb_ctx)
285 : {
286 : struct spdk_mem_map *map;
287 : int rc;
288 : size_t i;
289 :
290 0 : map = calloc(1, sizeof(*map));
291 0 : if (map == NULL) {
292 0 : return NULL;
293 : }
294 :
295 0 : if (pthread_mutex_init(&map->mutex, NULL)) {
296 0 : free(map);
297 0 : return NULL;
298 : }
299 :
300 0 : map->default_translation = default_translation;
301 0 : map->cb_ctx = cb_ctx;
302 0 : if (ops) {
303 0 : map->ops = *ops;
304 0 : }
305 :
306 0 : if (ops && ops->notify_cb) {
307 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
308 0 : rc = mem_map_notify_walk(map, SPDK_MEM_MAP_NOTIFY_REGISTER);
309 0 : if (rc != 0) {
310 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
311 0 : DEBUG_PRINT("Initial mem_map notify failed\n");
312 0 : pthread_mutex_destroy(&map->mutex);
313 0 : for (i = 0; i < sizeof(map->map_256tb.map) / sizeof(map->map_256tb.map[0]); i++) {
314 0 : free(map->map_256tb.map[i]);
315 0 : }
316 0 : free(map);
317 0 : return NULL;
318 : }
319 0 : TAILQ_INSERT_TAIL(&g_spdk_mem_maps, map, tailq);
320 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
321 0 : }
322 :
323 0 : return map;
324 0 : }
325 :
326 : void
327 0 : spdk_mem_map_free(struct spdk_mem_map **pmap)
328 : {
329 : struct spdk_mem_map *map;
330 : size_t i;
331 :
332 0 : if (!pmap) {
333 0 : return;
334 : }
335 :
336 0 : map = *pmap;
337 :
338 0 : if (!map) {
339 0 : return;
340 : }
341 :
342 0 : if (map->ops.notify_cb) {
343 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
344 0 : mem_map_notify_walk(map, SPDK_MEM_MAP_NOTIFY_UNREGISTER);
345 0 : TAILQ_REMOVE(&g_spdk_mem_maps, map, tailq);
346 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
347 0 : }
348 :
349 0 : for (i = 0; i < sizeof(map->map_256tb.map) / sizeof(map->map_256tb.map[0]); i++) {
350 0 : free(map->map_256tb.map[i]);
351 0 : }
352 :
353 0 : pthread_mutex_destroy(&map->mutex);
354 :
355 0 : free(map);
356 0 : *pmap = NULL;
357 0 : }
358 :
359 : uint64_t
360 0 : spdk_mem_map_get_page_size(void)
361 : {
362 0 : return g_map_page_cfg.size;
363 : }
364 :
365 : int
366 0 : spdk_mem_register(void *_vaddr, size_t len)
367 : {
368 : struct spdk_mem_map *map;
369 : int rc;
370 0 : uint64_t vaddr = (uintptr_t)_vaddr;
371 : uint64_t seg_vaddr;
372 : size_t seg_len;
373 : uint64_t reg;
374 :
375 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
376 0 : DEBUG_PRINT("invalid usermode virtual address %jx\n", vaddr);
377 0 : return -EINVAL;
378 : }
379 :
380 0 : if (((uintptr_t)vaddr & MAP_PAGE_MASK) || (len & MAP_PAGE_MASK)) {
381 0 : DEBUG_PRINT("invalid %s parameters, vaddr=%jx len=%ju\n",
382 : __func__, vaddr, len);
383 0 : return -EINVAL;
384 : }
385 :
386 0 : if (len == 0) {
387 0 : return 0;
388 : }
389 :
390 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
391 :
392 0 : seg_vaddr = vaddr;
393 0 : seg_len = len;
394 0 : while (seg_len > 0) {
395 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
396 0 : if (reg & REG_MAP_REGISTERED) {
397 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
398 0 : return -EBUSY;
399 : }
400 0 : seg_vaddr += MAP_PAGE_SIZE;
401 0 : seg_len -= MAP_PAGE_SIZE;
402 : }
403 :
404 0 : seg_vaddr = vaddr;
405 0 : seg_len = 0;
406 0 : while (len > 0) {
407 0 : spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, MAP_PAGE_SIZE,
408 0 : seg_len == 0 ? REG_MAP_REGISTERED | REG_MAP_NOTIFY_START : REG_MAP_REGISTERED);
409 0 : seg_len += MAP_PAGE_SIZE;
410 0 : vaddr += MAP_PAGE_SIZE;
411 0 : len -= MAP_PAGE_SIZE;
412 : }
413 :
414 0 : TAILQ_FOREACH(map, &g_spdk_mem_maps, tailq) {
415 0 : rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_REGISTER,
416 0 : (void *)seg_vaddr, seg_len);
417 0 : if (rc != 0) {
418 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
419 0 : return rc;
420 : }
421 0 : }
422 :
423 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
424 0 : return 0;
425 0 : }
426 :
427 : int
428 0 : spdk_mem_unregister(void *_vaddr, size_t len)
429 : {
430 : struct spdk_mem_map *map;
431 : int rc;
432 0 : uint64_t vaddr = (uintptr_t)_vaddr;
433 : uint64_t seg_vaddr;
434 : size_t seg_len;
435 : uint64_t reg, newreg;
436 :
437 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
438 0 : DEBUG_PRINT("invalid usermode virtual address %jx\n", vaddr);
439 0 : return -EINVAL;
440 : }
441 :
442 0 : if (((uintptr_t)vaddr & MAP_PAGE_MASK) || (len & MAP_PAGE_MASK)) {
443 0 : DEBUG_PRINT("invalid %s parameters, vaddr=%jx len=%ju\n",
444 : __func__, vaddr, len);
445 0 : return -EINVAL;
446 : }
447 :
448 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
449 :
450 : /* The first page must be a start of a region. Also check if it's
451 : * registered to make sure we don't return -ERANGE for non-registered
452 : * regions.
453 : */
454 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)vaddr, NULL);
455 0 : if ((reg & REG_MAP_REGISTERED) && (reg & REG_MAP_NOTIFY_START) == 0) {
456 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
457 0 : return -ERANGE;
458 : }
459 :
460 0 : seg_vaddr = vaddr;
461 0 : seg_len = len;
462 0 : while (seg_len > 0) {
463 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
464 0 : if ((reg & REG_MAP_REGISTERED) == 0) {
465 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
466 0 : return -EINVAL;
467 : }
468 0 : seg_vaddr += MAP_PAGE_SIZE;
469 0 : seg_len -= MAP_PAGE_SIZE;
470 : }
471 :
472 0 : newreg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
473 : /* If the next page is registered, it must be a start of a region as well,
474 : * otherwise we'd be unregistering only a part of a region.
475 : */
476 0 : if ((newreg & REG_MAP_NOTIFY_START) == 0 && (newreg & REG_MAP_REGISTERED)) {
477 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
478 0 : return -ERANGE;
479 : }
480 0 : seg_vaddr = vaddr;
481 0 : seg_len = 0;
482 :
483 0 : while (len > 0) {
484 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)vaddr, NULL);
485 0 : spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, MAP_PAGE_SIZE, 0);
486 :
487 0 : if (seg_len > 0 && (reg & REG_MAP_NOTIFY_START)) {
488 0 : TAILQ_FOREACH_REVERSE(map, &g_spdk_mem_maps, spdk_mem_map_head, tailq) {
489 0 : rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_UNREGISTER,
490 0 : (void *)seg_vaddr, seg_len);
491 0 : if (rc != 0) {
492 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
493 0 : return rc;
494 : }
495 0 : }
496 :
497 0 : seg_vaddr = vaddr;
498 0 : seg_len = MAP_PAGE_SIZE;
499 0 : } else {
500 0 : seg_len += MAP_PAGE_SIZE;
501 : }
502 :
503 0 : vaddr += MAP_PAGE_SIZE;
504 0 : len -= MAP_PAGE_SIZE;
505 : }
506 :
507 0 : if (seg_len > 0) {
508 0 : TAILQ_FOREACH_REVERSE(map, &g_spdk_mem_maps, spdk_mem_map_head, tailq) {
509 0 : rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_UNREGISTER,
510 0 : (void *)seg_vaddr, seg_len);
511 0 : if (rc != 0) {
512 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
513 0 : return rc;
514 : }
515 0 : }
516 0 : }
517 :
518 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
519 0 : return 0;
520 0 : }
521 :
522 : int
523 0 : spdk_mem_reserve(void *vaddr, size_t len)
524 : {
525 : struct spdk_mem_map *map;
526 : void *seg_vaddr;
527 : size_t seg_len;
528 : uint64_t reg;
529 :
530 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
531 0 : DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
532 0 : return -EINVAL;
533 : }
534 :
535 0 : if (((uintptr_t)vaddr & MAP_PAGE_MASK) || (len & MAP_PAGE_MASK)) {
536 0 : DEBUG_PRINT("invalid %s parameters, vaddr=%p len=%ju\n",
537 : __func__, vaddr, len);
538 0 : return -EINVAL;
539 : }
540 :
541 0 : if (len == 0) {
542 0 : return 0;
543 : }
544 :
545 0 : pthread_mutex_lock(&g_spdk_mem_map_mutex);
546 :
547 : /* Check if any part of this range is already registered */
548 0 : seg_vaddr = vaddr;
549 0 : seg_len = len;
550 0 : while (seg_len > 0) {
551 0 : reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
552 0 : if (reg & REG_MAP_REGISTERED) {
553 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
554 0 : return -EBUSY;
555 : }
556 0 : seg_vaddr += MAP_PAGE_SIZE;
557 0 : seg_len -= MAP_PAGE_SIZE;
558 : }
559 :
560 : /* Simply set the translation to the memory map's default. This allocates the space in the
561 : * map but does not provide a valid translation. */
562 0 : spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, len,
563 0 : g_mem_reg_map->default_translation);
564 :
565 0 : TAILQ_FOREACH(map, &g_spdk_mem_maps, tailq) {
566 0 : spdk_mem_map_set_translation(map, (uint64_t)vaddr, len, map->default_translation);
567 0 : }
568 :
569 0 : pthread_mutex_unlock(&g_spdk_mem_map_mutex);
570 0 : return 0;
571 0 : }
572 :
573 : static struct map_1gb *
574 0 : mem_map_get_map_1gb(struct spdk_mem_map *map, uint64_t vfn_page)
575 : {
576 : struct map_1gb *map_1gb;
577 0 : uint64_t idx_256tb = MAP_256TB_IDX(vfn_page);
578 : size_t i;
579 :
580 0 : if (spdk_unlikely(idx_256tb >= SPDK_COUNTOF(map->map_256tb.map))) {
581 0 : return NULL;
582 : }
583 :
584 0 : map_1gb = map->map_256tb.map[idx_256tb];
585 :
586 0 : if (!map_1gb) {
587 0 : pthread_mutex_lock(&map->mutex);
588 :
589 : /* Recheck to make sure nobody else got the mutex first. */
590 0 : map_1gb = map->map_256tb.map[idx_256tb];
591 0 : if (!map_1gb) {
592 0 : map_1gb = malloc(MAP_SIZE_OF_MAP_1GB);
593 0 : if (map_1gb) {
594 : /* initialize all entries to default translation */
595 0 : for (i = 0; i < MAP_NUM_PAGES_PER_GB; i++) {
596 0 : map_1gb->map[i].translation = map->default_translation;
597 0 : }
598 0 : map->map_256tb.map[idx_256tb] = map_1gb;
599 0 : }
600 0 : }
601 :
602 0 : pthread_mutex_unlock(&map->mutex);
603 :
604 0 : if (!map_1gb) {
605 0 : DEBUG_PRINT("allocation failed\n");
606 0 : return NULL;
607 : }
608 0 : }
609 :
610 0 : return map_1gb;
611 0 : }
612 :
613 : int
614 0 : spdk_mem_map_set_translation(struct spdk_mem_map *map, uint64_t vaddr, uint64_t size,
615 : uint64_t translation)
616 : {
617 : uint64_t vfn_page;
618 : struct map_1gb *map_1gb;
619 : uint64_t idx_1gb;
620 : struct map_page *map_page;
621 :
622 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
623 0 : DEBUG_PRINT("invalid usermode virtual address %" PRIu64 "\n", vaddr);
624 0 : return -EINVAL;
625 : }
626 :
627 : /* Only page-aligned registrations are supported */
628 0 : if (((uintptr_t)vaddr & MAP_PAGE_MASK) || (size & MAP_PAGE_MASK)) {
629 0 : DEBUG_PRINT("invalid %s parameters, vaddr=%" PRIu64 " len=%" PRIu64 "\n",
630 : __func__, vaddr, size);
631 0 : return -EINVAL;
632 : }
633 :
634 0 : vfn_page = vaddr >> MAP_PAGE_SHIFT;
635 :
636 0 : while (size) {
637 0 : map_1gb = mem_map_get_map_1gb(map, vfn_page);
638 0 : if (!map_1gb) {
639 0 : DEBUG_PRINT("could not get %p map\n", (void *)vaddr);
640 0 : return -ENOMEM;
641 : }
642 :
643 0 : idx_1gb = MAP_1GB_IDX(vfn_page);
644 0 : map_page = &map_1gb->map[idx_1gb];
645 0 : map_page->translation = translation;
646 :
647 0 : size -= MAP_PAGE_SIZE;
648 0 : vfn_page++;
649 : }
650 :
651 0 : return 0;
652 0 : }
653 :
654 : int
655 0 : spdk_mem_map_clear_translation(struct spdk_mem_map *map, uint64_t vaddr, uint64_t size)
656 : {
657 0 : return spdk_mem_map_set_translation(map, vaddr, size, map->default_translation);
658 : }
659 :
660 : inline uint64_t
661 0 : spdk_mem_map_translate(const struct spdk_mem_map *map, uint64_t vaddr, uint64_t *size)
662 : {
663 : const struct map_1gb *map_1gb;
664 : const struct map_page *map_page;
665 : uint64_t idx_256tb;
666 : uint64_t idx_1gb;
667 : uint64_t vfn_page;
668 : uint64_t cur_size;
669 : uint64_t prev_translation;
670 : uint64_t orig_translation;
671 :
672 0 : if (spdk_unlikely(vaddr & ~MASK_256TB)) {
673 0 : DEBUG_PRINT("invalid usermode virtual address %p\n", (void *)vaddr);
674 0 : return map->default_translation;
675 : }
676 :
677 0 : vfn_page = vaddr >> MAP_PAGE_SHIFT;
678 0 : idx_256tb = MAP_256TB_IDX(vfn_page);
679 0 : idx_1gb = MAP_1GB_IDX(vfn_page);
680 :
681 0 : map_1gb = map->map_256tb.map[idx_256tb];
682 0 : if (spdk_unlikely(!map_1gb)) {
683 0 : return map->default_translation;
684 : }
685 :
686 0 : cur_size = MAP_PAGE_SIZE - MAP_PAGE_OFFSET(vaddr);
687 0 : map_page = &map_1gb->map[idx_1gb];
688 0 : if (size == NULL || map->ops.are_contiguous == NULL ||
689 0 : map_page->translation == map->default_translation) {
690 0 : if (size != NULL) {
691 0 : *size = spdk_min(*size, cur_size);
692 0 : }
693 0 : return map_page->translation;
694 : }
695 :
696 0 : orig_translation = map_page->translation;
697 0 : prev_translation = orig_translation;
698 0 : while (cur_size < *size) {
699 0 : vfn_page++;
700 0 : idx_256tb = MAP_256TB_IDX(vfn_page);
701 0 : idx_1gb = MAP_1GB_IDX(vfn_page);
702 :
703 0 : map_1gb = map->map_256tb.map[idx_256tb];
704 0 : if (spdk_unlikely(!map_1gb)) {
705 0 : break;
706 : }
707 :
708 0 : map_page = &map_1gb->map[idx_1gb];
709 0 : if (!map->ops.are_contiguous(prev_translation, map_page->translation)) {
710 0 : break;
711 : }
712 :
713 0 : cur_size += MAP_PAGE_SIZE;
714 0 : prev_translation = map_page->translation;
715 : }
716 :
717 0 : *size = spdk_min(*size, cur_size);
718 0 : return orig_translation;
719 0 : }
720 :
721 : static void
722 0 : memory_hotplug_cb(enum rte_mem_event event_type,
723 : const void *addr, size_t len, void *arg)
724 : {
725 0 : if (event_type == RTE_MEM_EVENT_ALLOC) {
726 0 : spdk_mem_register((void *)addr, len);
727 :
728 0 : if (!spdk_env_dpdk_external_init()) {
729 0 : return;
730 : }
731 :
732 : /* When the user initialized DPDK separately, we can't
733 : * be sure that --match-allocations RTE flag was specified.
734 : * Without this flag, DPDK can free memory in different units
735 : * than it was allocated. It doesn't work with things like RDMA MRs.
736 : *
737 : * For such cases, we mark segments so they aren't freed.
738 : */
739 0 : while (len > 0) {
740 : struct rte_memseg *seg;
741 :
742 0 : seg = rte_mem_virt2memseg(addr, NULL);
743 0 : assert(seg != NULL);
744 0 : seg->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
745 0 : addr = (void *)((uintptr_t)addr + seg->hugepage_sz);
746 0 : len -= seg->hugepage_sz;
747 : }
748 0 : } else if (event_type == RTE_MEM_EVENT_FREE) {
749 0 : spdk_mem_unregister((void *)addr, len);
750 0 : }
751 0 : }
752 :
753 : static int
754 0 : memory_iter_cb(const struct rte_memseg_list *msl,
755 : const struct rte_memseg *ms, size_t len, void *arg)
756 : {
757 0 : return spdk_mem_register(ms->addr, len);
758 : }
759 :
760 : int
761 0 : mem_map_init(bool legacy_mem)
762 : {
763 : int rc;
764 :
765 0 : g_legacy_mem = legacy_mem;
766 :
767 0 : g_mem_reg_map = spdk_mem_map_alloc(0, NULL, NULL);
768 0 : if (g_mem_reg_map == NULL) {
769 0 : DEBUG_PRINT("memory registration map allocation failed\n");
770 0 : return -ENOMEM;
771 : }
772 :
773 0 : if (g_huge_pages) {
774 0 : rc = rte_mem_event_callback_register("spdk", memory_hotplug_cb, NULL);
775 0 : if (rc != 0) {
776 0 : DEBUG_PRINT("memory event callback registration failed, rc = %d\n", rc);
777 0 : goto err_free_reg_map;
778 : }
779 0 : g_mem_event_cb_registered = true;
780 0 : }
781 :
782 : /*
783 : * Walk all DPDK memory segments and register them
784 : * with the main memory map
785 : */
786 0 : rc = rte_memseg_contig_walk(memory_iter_cb, NULL);
787 0 : if (rc != 0) {
788 0 : DEBUG_PRINT("memory segments walking failed, rc = %d\n", rc);
789 0 : goto err_unregister_mem_cb;
790 : }
791 :
792 0 : return 0;
793 :
794 : err_unregister_mem_cb:
795 0 : if (g_mem_event_cb_registered) {
796 0 : g_mem_event_cb_registered = false;
797 0 : rte_mem_event_callback_unregister("spdk", NULL);
798 0 : }
799 : err_free_reg_map:
800 0 : spdk_mem_map_free(&g_mem_reg_map);
801 0 : return rc;
802 0 : }
803 :
804 : void
805 0 : mem_map_fini(void)
806 : {
807 0 : if (g_mem_event_cb_registered) {
808 0 : g_mem_event_cb_registered = false;
809 0 : rte_mem_event_callback_unregister("spdk", NULL);
810 0 : }
811 0 : spdk_mem_map_free(&g_mem_reg_map);
812 0 : }
813 :
814 : bool
815 0 : spdk_iommu_is_enabled(void)
816 : {
817 : #if VFIO_ENABLED
818 : return g_vfio.enabled && !g_vfio.noiommu_enabled;
819 : #else
820 0 : return false;
821 : #endif
822 : }
823 :
824 : struct spdk_vtophys_pci_device {
825 : struct rte_pci_device *pci_device;
826 : TAILQ_ENTRY(spdk_vtophys_pci_device) tailq;
827 : };
828 :
829 : static pthread_mutex_t g_vtophys_pci_devices_mutex = PTHREAD_MUTEX_INITIALIZER;
830 : static TAILQ_HEAD(, spdk_vtophys_pci_device) g_vtophys_pci_devices =
831 : TAILQ_HEAD_INITIALIZER(g_vtophys_pci_devices);
832 :
833 : static struct spdk_mem_map *g_vtophys_map = NULL;
834 : static struct spdk_mem_map *g_phys_ref_map = NULL;
835 : static struct spdk_mem_map *g_numa_map = NULL;
836 :
837 : #if VFIO_ENABLED
838 : static int
839 : _vfio_iommu_map_dma(uint64_t vaddr, uint64_t iova, uint64_t size)
840 : {
841 : struct spdk_vfio_dma_map *dma_map;
842 : int ret;
843 :
844 : dma_map = calloc(1, sizeof(*dma_map));
845 : if (dma_map == NULL) {
846 : return -ENOMEM;
847 : }
848 :
849 : dma_map->map.argsz = sizeof(dma_map->map);
850 : dma_map->map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
851 : dma_map->map.vaddr = vaddr;
852 : dma_map->map.iova = iova;
853 : dma_map->map.size = size;
854 :
855 : if (g_vfio.device_ref == 0) {
856 : /* VFIO requires at least one device (IOMMU group) to be added to
857 : * a VFIO container before it is possible to perform any IOMMU
858 : * operations on that container. This memory will be mapped once
859 : * the first device (IOMMU group) is hotplugged.
860 : *
861 : * Since the vfio container is managed internally by DPDK, it is
862 : * also possible that some device is already in that container, but
863 : * it's not managed by SPDK - e.g. an NIC attached internally
864 : * inside DPDK. We could map the memory straight away in such
865 : * scenario, but there's no need to do it. DPDK devices clearly
866 : * don't need our mappings and hence we defer the mapping
867 : * unconditionally until the first SPDK-managed device is
868 : * hotplugged.
869 : */
870 : goto out_insert;
871 : }
872 :
873 : ret = ioctl(g_vfio.fd, VFIO_IOMMU_MAP_DMA, &dma_map->map);
874 : if (ret) {
875 : /* There are cases the vfio container doesn't have IOMMU group, it's safe for this case */
876 : SPDK_NOTICELOG("Cannot set up DMA mapping, error %d, ignored\n", errno);
877 : }
878 :
879 : out_insert:
880 : TAILQ_INSERT_TAIL(&g_vfio.maps, dma_map, tailq);
881 : return 0;
882 : }
883 :
884 :
885 : static int
886 : vtophys_iommu_map_dma(uint64_t vaddr, uint64_t iova, uint64_t size)
887 : {
888 : uint64_t refcount;
889 : int ret;
890 :
891 : refcount = spdk_mem_map_translate(g_phys_ref_map, iova, NULL);
892 : assert(refcount < UINT64_MAX);
893 : if (refcount > 0) {
894 : spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount + 1);
895 : return 0;
896 : }
897 :
898 : pthread_mutex_lock(&g_vfio.mutex);
899 : ret = _vfio_iommu_map_dma(vaddr, iova, size);
900 : pthread_mutex_unlock(&g_vfio.mutex);
901 : if (ret) {
902 : return ret;
903 : }
904 :
905 : spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount + 1);
906 : return 0;
907 : }
908 :
909 : int
910 : vtophys_iommu_map_dma_bar(uint64_t vaddr, uint64_t iova, uint64_t size)
911 : {
912 : int ret;
913 :
914 : pthread_mutex_lock(&g_vfio.mutex);
915 : ret = _vfio_iommu_map_dma(vaddr, iova, size);
916 : pthread_mutex_unlock(&g_vfio.mutex);
917 :
918 : return ret;
919 : }
920 :
921 : static int
922 : _vfio_iommu_unmap_dma(struct spdk_vfio_dma_map *dma_map)
923 : {
924 : struct vfio_iommu_type1_dma_unmap unmap = {};
925 : int ret;
926 :
927 : if (g_vfio.device_ref == 0) {
928 : /* Memory is not mapped anymore, just remove it's references */
929 : goto out_remove;
930 : }
931 :
932 : unmap.argsz = sizeof(unmap);
933 : unmap.flags = 0;
934 : unmap.iova = dma_map->map.iova;
935 : unmap.size = dma_map->map.size;
936 : ret = ioctl(g_vfio.fd, VFIO_IOMMU_UNMAP_DMA, &unmap);
937 : if (ret) {
938 : SPDK_NOTICELOG("Cannot clear DMA mapping, error %d, ignored\n", errno);
939 : }
940 :
941 : out_remove:
942 : TAILQ_REMOVE(&g_vfio.maps, dma_map, tailq);
943 : free(dma_map);
944 : return 0;
945 : }
946 :
947 : static int
948 : vtophys_iommu_unmap_dma(uint64_t iova, uint64_t size)
949 : {
950 : struct spdk_vfio_dma_map *dma_map;
951 : uint64_t refcount;
952 : int ret;
953 :
954 : pthread_mutex_lock(&g_vfio.mutex);
955 : TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
956 : if (dma_map->map.iova == iova) {
957 : break;
958 : }
959 : }
960 :
961 : if (dma_map == NULL) {
962 : DEBUG_PRINT("Cannot clear DMA mapping for IOVA %"PRIx64" - it's not mapped\n", iova);
963 : pthread_mutex_unlock(&g_vfio.mutex);
964 : return -ENXIO;
965 : }
966 :
967 : refcount = spdk_mem_map_translate(g_phys_ref_map, iova, NULL);
968 : assert(refcount < UINT64_MAX);
969 : if (refcount > 0) {
970 : spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount - 1);
971 : }
972 :
973 : /* We still have outstanding references, don't clear it. */
974 : if (refcount > 1) {
975 : pthread_mutex_unlock(&g_vfio.mutex);
976 : return 0;
977 : }
978 :
979 : /** don't support partial or multiple-page unmap for now */
980 : assert(dma_map->map.size == size);
981 :
982 : ret = _vfio_iommu_unmap_dma(dma_map);
983 : pthread_mutex_unlock(&g_vfio.mutex);
984 :
985 : return ret;
986 : }
987 :
988 : int
989 : vtophys_iommu_unmap_dma_bar(uint64_t vaddr)
990 : {
991 : struct spdk_vfio_dma_map *dma_map;
992 : int ret;
993 :
994 : pthread_mutex_lock(&g_vfio.mutex);
995 : TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
996 : if (dma_map->map.vaddr == vaddr) {
997 : break;
998 : }
999 : }
1000 :
1001 : if (dma_map == NULL) {
1002 : DEBUG_PRINT("Cannot clear DMA mapping for address %"PRIx64" - it's not mapped\n", vaddr);
1003 : pthread_mutex_unlock(&g_vfio.mutex);
1004 : return -ENXIO;
1005 : }
1006 :
1007 : ret = _vfio_iommu_unmap_dma(dma_map);
1008 : pthread_mutex_unlock(&g_vfio.mutex);
1009 : return ret;
1010 : }
1011 : #endif
1012 :
1013 : static uint64_t
1014 0 : vtophys_get_paddr_memseg(uint64_t vaddr)
1015 : {
1016 : uintptr_t paddr;
1017 : struct rte_memseg *seg;
1018 :
1019 0 : seg = rte_mem_virt2memseg((void *)(uintptr_t)vaddr, NULL);
1020 0 : if (seg != NULL) {
1021 0 : paddr = seg->iova;
1022 0 : if (paddr == RTE_BAD_IOVA) {
1023 0 : return SPDK_VTOPHYS_ERROR;
1024 : }
1025 0 : paddr += (vaddr - (uintptr_t)seg->addr);
1026 0 : return paddr;
1027 : }
1028 :
1029 0 : return SPDK_VTOPHYS_ERROR;
1030 0 : }
1031 :
1032 : /* Try to get the paddr from /proc/self/pagemap */
1033 : static uint64_t
1034 0 : vtophys_get_paddr_pagemap(uint64_t vaddr)
1035 : {
1036 : uintptr_t paddr;
1037 :
1038 : /* Silence static analyzers */
1039 0 : assert(vaddr != 0);
1040 0 : paddr = rte_mem_virt2iova((void *)vaddr);
1041 0 : if (paddr == RTE_BAD_IOVA) {
1042 : /*
1043 : * The vaddr may be valid but doesn't have a backing page
1044 : * assigned yet. Touch the page to ensure a backing page
1045 : * gets assigned, then try to translate again.
1046 : */
1047 0 : rte_atomic64_read((rte_atomic64_t *)vaddr);
1048 0 : paddr = rte_mem_virt2iova((void *)vaddr);
1049 0 : }
1050 0 : if (paddr == RTE_BAD_IOVA) {
1051 : /* Unable to get to the physical address. */
1052 0 : return SPDK_VTOPHYS_ERROR;
1053 : }
1054 :
1055 0 : return paddr;
1056 0 : }
1057 :
1058 : static uint64_t
1059 0 : pci_device_vtophys(struct rte_pci_device *dev, uint64_t vaddr, size_t len)
1060 : {
1061 : struct rte_mem_resource *res;
1062 : uint64_t paddr;
1063 : unsigned r;
1064 :
1065 0 : for (r = 0; r < PCI_MAX_RESOURCE; r++) {
1066 0 : res = dpdk_pci_device_get_mem_resource(dev, r);
1067 :
1068 0 : if (res->phys_addr == 0 || vaddr < (uint64_t)res->addr ||
1069 0 : (vaddr + len) >= (uint64_t)res->addr + res->len) {
1070 0 : continue;
1071 : }
1072 :
1073 : #if VFIO_ENABLED
1074 : if (spdk_iommu_is_enabled() && rte_eal_iova_mode() == RTE_IOVA_VA) {
1075 : /*
1076 : * The IOMMU is on and we're using IOVA == VA. The BAR was
1077 : * automatically registered when it was mapped, so just return
1078 : * the virtual address here.
1079 : */
1080 : return vaddr;
1081 : }
1082 : #endif
1083 0 : paddr = res->phys_addr + (vaddr - (uint64_t)res->addr);
1084 0 : return paddr;
1085 : }
1086 :
1087 0 : return SPDK_VTOPHYS_ERROR;
1088 0 : }
1089 :
1090 : /* Try to get the paddr from pci devices */
1091 : static uint64_t
1092 0 : vtophys_get_paddr_pci(uint64_t vaddr, size_t len)
1093 : {
1094 : struct spdk_vtophys_pci_device *vtophys_dev;
1095 : uintptr_t paddr;
1096 : struct rte_pci_device *dev;
1097 :
1098 0 : pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
1099 0 : TAILQ_FOREACH(vtophys_dev, &g_vtophys_pci_devices, tailq) {
1100 0 : dev = vtophys_dev->pci_device;
1101 0 : paddr = pci_device_vtophys(dev, vaddr, len);
1102 0 : if (paddr != SPDK_VTOPHYS_ERROR) {
1103 0 : pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
1104 0 : return paddr;
1105 : }
1106 0 : }
1107 0 : pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
1108 :
1109 0 : return SPDK_VTOPHYS_ERROR;
1110 0 : }
1111 :
1112 : static int
1113 0 : vtophys_notify(void *cb_ctx, struct spdk_mem_map *map,
1114 : enum spdk_mem_map_notify_action action,
1115 : void *vaddr, size_t len)
1116 : {
1117 0 : int rc = 0;
1118 : uint64_t paddr;
1119 :
1120 0 : if ((uintptr_t)vaddr & ~MASK_256TB) {
1121 0 : DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
1122 0 : return -EINVAL;
1123 : }
1124 :
1125 0 : if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
1126 0 : DEBUG_PRINT("invalid parameters, vaddr=%p len=%ju\n",
1127 : vaddr, len);
1128 0 : return -EINVAL;
1129 : }
1130 :
1131 : /* Get the physical address from the DPDK memsegs */
1132 0 : paddr = vtophys_get_paddr_memseg((uint64_t)vaddr);
1133 :
1134 0 : switch (action) {
1135 : case SPDK_MEM_MAP_NOTIFY_REGISTER:
1136 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1137 : /* This is not an address that DPDK is managing. */
1138 :
1139 : /* Check if this is a PCI BAR. They need special handling */
1140 0 : paddr = vtophys_get_paddr_pci((uint64_t)vaddr, len);
1141 0 : if (paddr != SPDK_VTOPHYS_ERROR) {
1142 : /* Get paddr for each 2MB chunk in this address range */
1143 0 : while (len > 0) {
1144 0 : paddr = vtophys_get_paddr_pci((uint64_t)vaddr, VALUE_2MB);
1145 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1146 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1147 0 : return -EFAULT;
1148 : }
1149 :
1150 0 : rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
1151 0 : if (rc != 0) {
1152 0 : return rc;
1153 : }
1154 :
1155 0 : vaddr += VALUE_2MB;
1156 0 : len -= VALUE_2MB;
1157 : }
1158 :
1159 0 : return 0;
1160 : }
1161 :
1162 : #if VFIO_ENABLED
1163 : enum rte_iova_mode iova_mode;
1164 :
1165 : iova_mode = rte_eal_iova_mode();
1166 :
1167 : if (spdk_iommu_is_enabled() && iova_mode == RTE_IOVA_VA) {
1168 : /* We'll use the virtual address as the iova to match DPDK. */
1169 : paddr = (uint64_t)vaddr;
1170 : rc = vtophys_iommu_map_dma((uint64_t)vaddr, paddr, len);
1171 : if (rc) {
1172 : return -EFAULT;
1173 : }
1174 : while (len > 0) {
1175 : rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
1176 : if (rc != 0) {
1177 : return rc;
1178 : }
1179 : vaddr += VALUE_2MB;
1180 : paddr += VALUE_2MB;
1181 : len -= VALUE_2MB;
1182 : }
1183 : } else
1184 : #endif
1185 : {
1186 : /* Get the physical address from /proc/self/pagemap. */
1187 0 : paddr = vtophys_get_paddr_pagemap((uint64_t)vaddr);
1188 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1189 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1190 0 : return -EFAULT;
1191 : }
1192 :
1193 : /* Get paddr for each 2MB chunk in this address range */
1194 0 : while (len > 0) {
1195 : /* Get the physical address from /proc/self/pagemap. */
1196 0 : paddr = vtophys_get_paddr_pagemap((uint64_t)vaddr);
1197 :
1198 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1199 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1200 0 : return -EFAULT;
1201 : }
1202 :
1203 0 : if (paddr & MASK_2MB) {
1204 0 : DEBUG_PRINT("invalid paddr 0x%" PRIx64 " - must be 2MB aligned\n", paddr);
1205 0 : return -EINVAL;
1206 : }
1207 : #if VFIO_ENABLED
1208 : /* If the IOMMU is on, but DPDK is using iova-mode=pa, we want to register this memory
1209 : * with the IOMMU using the physical address to match. */
1210 : if (spdk_iommu_is_enabled()) {
1211 : rc = vtophys_iommu_map_dma((uint64_t)vaddr, paddr, VALUE_2MB);
1212 : if (rc) {
1213 : DEBUG_PRINT("Unable to assign vaddr %p to paddr 0x%" PRIx64 "\n", vaddr, paddr);
1214 : return -EFAULT;
1215 : }
1216 : }
1217 : #endif
1218 :
1219 0 : rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
1220 0 : if (rc != 0) {
1221 0 : return rc;
1222 : }
1223 :
1224 0 : vaddr += VALUE_2MB;
1225 0 : len -= VALUE_2MB;
1226 : }
1227 : }
1228 0 : } else {
1229 : /* This is an address managed by DPDK. Just setup the translations. */
1230 0 : while (len > 0) {
1231 0 : paddr = vtophys_get_paddr_memseg((uint64_t)vaddr);
1232 0 : if (paddr == SPDK_VTOPHYS_ERROR) {
1233 0 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1234 0 : return -EFAULT;
1235 : }
1236 :
1237 0 : rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
1238 0 : if (rc != 0) {
1239 0 : return rc;
1240 : }
1241 :
1242 0 : vaddr += VALUE_2MB;
1243 0 : len -= VALUE_2MB;
1244 : }
1245 : }
1246 :
1247 0 : break;
1248 : case SPDK_MEM_MAP_NOTIFY_UNREGISTER:
1249 : #if VFIO_ENABLED
1250 : if (paddr == SPDK_VTOPHYS_ERROR) {
1251 : /*
1252 : * This is not an address that DPDK is managing.
1253 : */
1254 :
1255 : /* Check if this is a PCI BAR. They need special handling */
1256 : paddr = vtophys_get_paddr_pci((uint64_t)vaddr, len);
1257 : if (paddr != SPDK_VTOPHYS_ERROR) {
1258 : /* Get paddr for each 2MB chunk in this address range */
1259 : while (len > 0) {
1260 : paddr = vtophys_get_paddr_pci((uint64_t)vaddr, VALUE_2MB);
1261 : if (paddr == SPDK_VTOPHYS_ERROR) {
1262 : DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
1263 : return -EFAULT;
1264 : }
1265 :
1266 : rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, VALUE_2MB);
1267 : if (rc != 0) {
1268 : return rc;
1269 : }
1270 :
1271 : vaddr += VALUE_2MB;
1272 : len -= VALUE_2MB;
1273 : }
1274 :
1275 : return 0;
1276 : }
1277 :
1278 : /* If vfio is enabled,
1279 : * we need to unmap the range from the IOMMU
1280 : */
1281 : if (spdk_iommu_is_enabled()) {
1282 : uint64_t buffer_len = len;
1283 : uint8_t *va = vaddr;
1284 : enum rte_iova_mode iova_mode;
1285 :
1286 : iova_mode = rte_eal_iova_mode();
1287 : /*
1288 : * In virtual address mode, the region is contiguous and can be done in
1289 : * one unmap.
1290 : */
1291 : if (iova_mode == RTE_IOVA_VA) {
1292 : paddr = spdk_mem_map_translate(map, (uint64_t)va, &buffer_len);
1293 : if (buffer_len != len || paddr != (uintptr_t)va) {
1294 : DEBUG_PRINT("Unmapping %p with length %lu failed because "
1295 : "translation had address 0x%" PRIx64 " and length %lu\n",
1296 : va, len, paddr, buffer_len);
1297 : return -EINVAL;
1298 : }
1299 : rc = vtophys_iommu_unmap_dma(paddr, len);
1300 : if (rc) {
1301 : DEBUG_PRINT("Failed to iommu unmap paddr 0x%" PRIx64 "\n", paddr);
1302 : return -EFAULT;
1303 : }
1304 : } else if (iova_mode == RTE_IOVA_PA) {
1305 : /* Get paddr for each 2MB chunk in this address range */
1306 : while (buffer_len > 0) {
1307 : paddr = spdk_mem_map_translate(map, (uint64_t)va, NULL);
1308 :
1309 : if (paddr == SPDK_VTOPHYS_ERROR || buffer_len < VALUE_2MB) {
1310 : DEBUG_PRINT("could not get phys addr for %p\n", va);
1311 : return -EFAULT;
1312 : }
1313 :
1314 : rc = vtophys_iommu_unmap_dma(paddr, VALUE_2MB);
1315 : if (rc) {
1316 : DEBUG_PRINT("Failed to iommu unmap paddr 0x%" PRIx64 "\n", paddr);
1317 : return -EFAULT;
1318 : }
1319 :
1320 : va += VALUE_2MB;
1321 : buffer_len -= VALUE_2MB;
1322 : }
1323 : }
1324 : }
1325 : }
1326 : #endif
1327 0 : while (len > 0) {
1328 0 : rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, VALUE_2MB);
1329 0 : if (rc != 0) {
1330 0 : return rc;
1331 : }
1332 :
1333 0 : vaddr += VALUE_2MB;
1334 0 : len -= VALUE_2MB;
1335 : }
1336 :
1337 0 : break;
1338 : default:
1339 0 : SPDK_UNREACHABLE();
1340 : }
1341 :
1342 0 : return rc;
1343 0 : }
1344 :
1345 : static int
1346 0 : numa_notify(void *cb_ctx, struct spdk_mem_map *map,
1347 : enum spdk_mem_map_notify_action action,
1348 : void *vaddr, size_t len)
1349 : {
1350 : struct rte_memseg *seg;
1351 :
1352 : /* We always return 0 from here, even if we aren't able to get a
1353 : * memseg for the address. This can happen in non-DPDK memory
1354 : * registration paths, for example vhost or vfio-user. That is OK,
1355 : * spdk_mem_get_numa_id() just returns SPDK_ENV_NUMA_ID_ANY for
1356 : * that kind of memory. If we return an error here, the
1357 : * spdk_mem_register() from vhost or vfio-user would fail which is
1358 : * not what we want.
1359 : */
1360 0 : seg = rte_mem_virt2memseg(vaddr, NULL);
1361 0 : if (seg == NULL) {
1362 0 : return 0;
1363 : }
1364 :
1365 0 : switch (action) {
1366 : case SPDK_MEM_MAP_NOTIFY_REGISTER:
1367 0 : spdk_mem_map_set_translation(map, (uint64_t)vaddr, len, seg->socket_id);
1368 0 : break;
1369 : case SPDK_MEM_MAP_NOTIFY_UNREGISTER:
1370 0 : spdk_mem_map_clear_translation(map, (uint64_t)vaddr, len);
1371 0 : break;
1372 : default:
1373 0 : break;
1374 : }
1375 :
1376 0 : return 0;
1377 0 : }
1378 :
1379 : static int
1380 0 : vtophys_check_contiguous_entries(uint64_t paddr1, uint64_t paddr2)
1381 : {
1382 : /* This function is always called with paddrs for two subsequent
1383 : * 2MB chunks in virtual address space, so those chunks will be only
1384 : * physically contiguous if the physical addresses are 2MB apart
1385 : * from each other as well.
1386 : */
1387 0 : return (paddr2 - paddr1 == VALUE_2MB);
1388 : }
1389 :
1390 : #if VFIO_ENABLED
1391 :
1392 : static bool
1393 : vfio_enabled(void)
1394 : {
1395 : return rte_vfio_is_enabled("vfio_pci");
1396 : }
1397 :
1398 : /* Check if IOMMU is enabled on the system */
1399 : static bool
1400 : has_iommu_groups(void)
1401 : {
1402 : int count = 0;
1403 : DIR *dir = opendir("/sys/kernel/iommu_groups");
1404 :
1405 : if (dir == NULL) {
1406 : return false;
1407 : }
1408 :
1409 : while (count < 3 && readdir(dir) != NULL) {
1410 : count++;
1411 : }
1412 :
1413 : closedir(dir);
1414 : /* there will always be ./ and ../ entries */
1415 : return count > 2;
1416 : }
1417 :
1418 : static bool
1419 : vfio_noiommu_enabled(void)
1420 : {
1421 : return rte_vfio_noiommu_is_enabled();
1422 : }
1423 :
1424 : static void
1425 : vtophys_iommu_init(void)
1426 : {
1427 : char proc_fd_path[PATH_MAX + 1];
1428 : char link_path[PATH_MAX + 1];
1429 : const char vfio_path[] = "/dev/vfio/vfio";
1430 : DIR *dir;
1431 : struct dirent *d;
1432 :
1433 : if (!vfio_enabled()) {
1434 : return;
1435 : }
1436 :
1437 : if (vfio_noiommu_enabled()) {
1438 : g_vfio.noiommu_enabled = true;
1439 : } else if (!has_iommu_groups()) {
1440 : return;
1441 : }
1442 :
1443 : dir = opendir("/proc/self/fd");
1444 : if (!dir) {
1445 : DEBUG_PRINT("Failed to open /proc/self/fd (%d)\n", errno);
1446 : return;
1447 : }
1448 :
1449 : while ((d = readdir(dir)) != NULL) {
1450 : if (d->d_type != DT_LNK) {
1451 : continue;
1452 : }
1453 :
1454 : snprintf(proc_fd_path, sizeof(proc_fd_path), "/proc/self/fd/%s", d->d_name);
1455 : if (readlink(proc_fd_path, link_path, sizeof(link_path)) != (sizeof(vfio_path) - 1)) {
1456 : continue;
1457 : }
1458 :
1459 : if (memcmp(link_path, vfio_path, sizeof(vfio_path) - 1) == 0) {
1460 : sscanf(d->d_name, "%d", &g_vfio.fd);
1461 : break;
1462 : }
1463 : }
1464 :
1465 : closedir(dir);
1466 :
1467 : if (g_vfio.fd < 0) {
1468 : DEBUG_PRINT("Failed to discover DPDK VFIO container fd.\n");
1469 : return;
1470 : }
1471 :
1472 : g_vfio.enabled = true;
1473 :
1474 : return;
1475 : }
1476 :
1477 : #endif
1478 :
1479 : void
1480 0 : vtophys_pci_device_added(struct rte_pci_device *pci_device)
1481 : {
1482 : struct spdk_vtophys_pci_device *vtophys_dev;
1483 :
1484 0 : pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
1485 :
1486 0 : vtophys_dev = calloc(1, sizeof(*vtophys_dev));
1487 0 : if (vtophys_dev) {
1488 0 : vtophys_dev->pci_device = pci_device;
1489 0 : TAILQ_INSERT_TAIL(&g_vtophys_pci_devices, vtophys_dev, tailq);
1490 0 : } else {
1491 0 : DEBUG_PRINT("Memory allocation error\n");
1492 : }
1493 0 : pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
1494 :
1495 : #if VFIO_ENABLED
1496 : struct spdk_vfio_dma_map *dma_map;
1497 : int ret;
1498 :
1499 : if (!g_vfio.enabled) {
1500 : return;
1501 : }
1502 :
1503 : pthread_mutex_lock(&g_vfio.mutex);
1504 : g_vfio.device_ref++;
1505 : if (g_vfio.device_ref > 1) {
1506 : pthread_mutex_unlock(&g_vfio.mutex);
1507 : return;
1508 : }
1509 :
1510 : /* This is the first SPDK device using DPDK vfio. This means that the first
1511 : * IOMMU group might have been just been added to the DPDK vfio container.
1512 : * From this point it is certain that the memory can be mapped now.
1513 : */
1514 : TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
1515 : ret = ioctl(g_vfio.fd, VFIO_IOMMU_MAP_DMA, &dma_map->map);
1516 : if (ret) {
1517 : DEBUG_PRINT("Cannot update DMA mapping, error %d\n", errno);
1518 : break;
1519 : }
1520 : }
1521 : pthread_mutex_unlock(&g_vfio.mutex);
1522 : #endif
1523 0 : }
1524 :
1525 : void
1526 0 : vtophys_pci_device_removed(struct rte_pci_device *pci_device)
1527 : {
1528 : struct spdk_vtophys_pci_device *vtophys_dev;
1529 :
1530 0 : pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
1531 0 : TAILQ_FOREACH(vtophys_dev, &g_vtophys_pci_devices, tailq) {
1532 0 : if (vtophys_dev->pci_device == pci_device) {
1533 0 : TAILQ_REMOVE(&g_vtophys_pci_devices, vtophys_dev, tailq);
1534 0 : free(vtophys_dev);
1535 0 : break;
1536 : }
1537 0 : }
1538 0 : pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
1539 :
1540 : #if VFIO_ENABLED
1541 : struct spdk_vfio_dma_map *dma_map;
1542 : int ret;
1543 :
1544 : if (!g_vfio.enabled) {
1545 : return;
1546 : }
1547 :
1548 : pthread_mutex_lock(&g_vfio.mutex);
1549 : assert(g_vfio.device_ref > 0);
1550 : g_vfio.device_ref--;
1551 : if (g_vfio.device_ref > 0) {
1552 : pthread_mutex_unlock(&g_vfio.mutex);
1553 : return;
1554 : }
1555 :
1556 : /* This is the last SPDK device using DPDK vfio. If DPDK doesn't have
1557 : * any additional devices using it's vfio container, all the mappings
1558 : * will be automatically removed by the Linux vfio driver. We unmap
1559 : * the memory manually to be able to easily re-map it later regardless
1560 : * of other, external factors.
1561 : */
1562 : TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
1563 : struct vfio_iommu_type1_dma_unmap unmap = {};
1564 : unmap.argsz = sizeof(unmap);
1565 : unmap.flags = 0;
1566 : unmap.iova = dma_map->map.iova;
1567 : unmap.size = dma_map->map.size;
1568 : ret = ioctl(g_vfio.fd, VFIO_IOMMU_UNMAP_DMA, &unmap);
1569 : if (ret) {
1570 : DEBUG_PRINT("Cannot unmap DMA memory, error %d\n", errno);
1571 : break;
1572 : }
1573 : }
1574 : pthread_mutex_unlock(&g_vfio.mutex);
1575 : #endif
1576 0 : }
1577 :
1578 : int
1579 0 : vtophys_init(void)
1580 : {
1581 0 : const struct spdk_mem_map_ops vtophys_map_ops = {
1582 : .notify_cb = vtophys_notify,
1583 : .are_contiguous = vtophys_check_contiguous_entries,
1584 : };
1585 :
1586 0 : const struct spdk_mem_map_ops phys_ref_map_ops = {
1587 : .notify_cb = NULL,
1588 : .are_contiguous = NULL,
1589 : };
1590 :
1591 0 : const struct spdk_mem_map_ops numa_map_ops = {
1592 : .notify_cb = numa_notify,
1593 : .are_contiguous = NULL,
1594 : };
1595 :
1596 : #if VFIO_ENABLED
1597 : vtophys_iommu_init();
1598 : #endif
1599 :
1600 0 : g_phys_ref_map = spdk_mem_map_alloc(0, &phys_ref_map_ops, NULL);
1601 0 : if (g_phys_ref_map == NULL) {
1602 0 : DEBUG_PRINT("phys_ref map allocation failed.\n");
1603 0 : return -ENOMEM;
1604 : }
1605 :
1606 0 : g_numa_map = spdk_mem_map_alloc(SPDK_ENV_NUMA_ID_ANY, &numa_map_ops, NULL);
1607 0 : if (g_numa_map == NULL) {
1608 0 : DEBUG_PRINT("numa map allocation failed.\n");
1609 0 : spdk_mem_map_free(&g_phys_ref_map);
1610 0 : return -ENOMEM;
1611 : }
1612 :
1613 0 : if (g_huge_pages) {
1614 0 : g_vtophys_map = spdk_mem_map_alloc(SPDK_VTOPHYS_ERROR, &vtophys_map_ops, NULL);
1615 0 : if (g_vtophys_map == NULL) {
1616 0 : DEBUG_PRINT("vtophys map allocation failed\n");
1617 0 : spdk_mem_map_free(&g_numa_map);
1618 0 : spdk_mem_map_free(&g_phys_ref_map);
1619 0 : return -ENOMEM;
1620 : }
1621 0 : }
1622 0 : return 0;
1623 0 : }
1624 :
1625 : void
1626 0 : vtophys_fini(void)
1627 : {
1628 0 : spdk_mem_map_free(&g_vtophys_map);
1629 0 : spdk_mem_map_free(&g_numa_map);
1630 0 : spdk_mem_map_free(&g_phys_ref_map);
1631 0 : }
1632 :
1633 : uint64_t
1634 0 : spdk_vtophys(const void *buf, uint64_t *size)
1635 : {
1636 : uint64_t vaddr, paddr_2mb;
1637 :
1638 0 : if (!g_huge_pages) {
1639 0 : return SPDK_VTOPHYS_ERROR;
1640 : }
1641 :
1642 0 : vaddr = (uint64_t)buf;
1643 0 : paddr_2mb = spdk_mem_map_translate(g_vtophys_map, vaddr, size);
1644 :
1645 : /*
1646 : * SPDK_VTOPHYS_ERROR has all bits set, so if the lookup returned SPDK_VTOPHYS_ERROR,
1647 : * we will still bitwise-or it with the buf offset below, but the result will still be
1648 : * SPDK_VTOPHYS_ERROR. However now that we do + rather than | (due to PCI vtophys being
1649 : * unaligned) we must now check the return value before addition.
1650 : */
1651 : SPDK_STATIC_ASSERT(SPDK_VTOPHYS_ERROR == UINT64_C(-1), "SPDK_VTOPHYS_ERROR should be all 1s");
1652 0 : if (paddr_2mb == SPDK_VTOPHYS_ERROR) {
1653 0 : return SPDK_VTOPHYS_ERROR;
1654 : } else {
1655 0 : return paddr_2mb + (vaddr & MASK_2MB);
1656 : }
1657 0 : }
1658 :
1659 : int32_t
1660 0 : spdk_mem_get_numa_id(const void *buf, uint64_t *size)
1661 : {
1662 0 : return spdk_mem_map_translate(g_numa_map, (uint64_t)buf, size);
1663 : }
1664 :
1665 : int
1666 0 : spdk_mem_get_fd_and_offset(void *vaddr, uint64_t *offset)
1667 : {
1668 : struct rte_memseg *seg;
1669 : int ret, fd;
1670 :
1671 0 : seg = rte_mem_virt2memseg(vaddr, NULL);
1672 0 : if (!seg) {
1673 0 : SPDK_ERRLOG("memory %p doesn't exist\n", vaddr);
1674 0 : return -ENOENT;
1675 : }
1676 :
1677 0 : fd = rte_memseg_get_fd_thread_unsafe(seg);
1678 0 : if (fd < 0) {
1679 0 : return fd;
1680 : }
1681 :
1682 0 : ret = rte_memseg_get_fd_offset_thread_unsafe(seg, offset);
1683 0 : if (ret < 0) {
1684 0 : return ret;
1685 : }
1686 :
1687 0 : return fd;
1688 0 : }
1689 :
1690 : void
1691 0 : mem_disable_huge_pages(void)
1692 : {
1693 0 : g_huge_pages = false;
1694 0 : mem_map_use_page_shift(SHIFT_4KB);
1695 0 : }
1696 :
1697 : void
1698 0 : mem_map_use_page_shift(uint32_t page_shift)
1699 : {
1700 0 : g_map_page_cfg.shift = page_shift;
1701 0 : g_map_page_cfg.size = 1UL << page_shift;
1702 0 : g_map_page_cfg.mask = g_map_page_cfg.size - 1;
1703 0 : g_map_page_cfg.num_pages_per_gb = 1UL << (SHIFT_1GB - page_shift);
1704 0 : }
|