1/*
2 * Vhost User library
3 *
4 * Copyright IBM, Corp. 2007
5 * Copyright (c) 2016 Red Hat, Inc.
6 *
7 * Authors:
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Marc-André Lureau <mlureau@redhat.com>
10 * Victor Kaplansky <victork@redhat.com>
11 *
12 * This work is licensed under the terms of the GNU GPL, version 2 or
13 * later. See the COPYING file in the top-level directory.
14 */
15
16/* this code avoids GLib dependency */
17#include <stdlib.h>
18#include <stdio.h>
19#include <unistd.h>
20#include <stdarg.h>
21#include <errno.h>
22#include <string.h>
23#include <assert.h>
24#include <inttypes.h>
25#include <sys/types.h>
26#include <sys/socket.h>
27#include <sys/eventfd.h>
28#include <sys/mman.h>
29#include "qemu/compiler.h"
30
31#if defined(__linux__)
32#include <sys/syscall.h>
33#include <fcntl.h>
34#include <sys/ioctl.h>
35#include <linux/vhost.h>
36
37#ifdef __NR_userfaultfd
38#include <linux/userfaultfd.h>
39#endif
40
41#endif
42
43#include "qemu/atomic.h"
44#include "qemu/osdep.h"
45#include "qemu/memfd.h"
46
47#include "libvhost-user.h"
48
49/* usually provided by GLib */
50#ifndef MIN
51#define MIN(x, y) ({ \
52 typeof(x) _min1 = (x); \
53 typeof(y) _min2 = (y); \
54 (void) (&_min1 == &_min2); \
55 _min1 < _min2 ? _min1 : _min2; })
56#endif
57
58/* Round number down to multiple */
59#define ALIGN_DOWN(n, m) ((n) / (m) * (m))
60
61/* Round number up to multiple */
62#define ALIGN_UP(n, m) ALIGN_DOWN((n) + (m) - 1, (m))
63
64/* Align each region to cache line size in inflight buffer */
65#define INFLIGHT_ALIGNMENT 64
66
67/* The version of inflight buffer */
68#define INFLIGHT_VERSION 1
69
70#define VHOST_USER_HDR_SIZE offsetof(VhostUserMsg, payload.u64)
71
72/* The version of the protocol we support */
73#define VHOST_USER_VERSION 1
74#define LIBVHOST_USER_DEBUG 0
75
76#define DPRINT(...) \
77 do { \
78 if (LIBVHOST_USER_DEBUG) { \
79 fprintf(stderr, __VA_ARGS__); \
80 } \
81 } while (0)
82
83static inline
84bool has_feature(uint64_t features, unsigned int fbit)
85{
86 assert(fbit < 64);
87 return !!(features & (1ULL << fbit));
88}
89
90static inline
91bool vu_has_feature(VuDev *dev,
92 unsigned int fbit)
93{
94 return has_feature(dev->features, fbit);
95}
96
97static inline bool vu_has_protocol_feature(VuDev *dev, unsigned int fbit)
98{
99 return has_feature(dev->protocol_features, fbit);
100}
101
102static const char *
103vu_request_to_string(unsigned int req)
104{
105#define REQ(req) [req] = #req
106 static const char *vu_request_str[] = {
107 REQ(VHOST_USER_NONE),
108 REQ(VHOST_USER_GET_FEATURES),
109 REQ(VHOST_USER_SET_FEATURES),
110 REQ(VHOST_USER_SET_OWNER),
111 REQ(VHOST_USER_RESET_OWNER),
112 REQ(VHOST_USER_SET_MEM_TABLE),
113 REQ(VHOST_USER_SET_LOG_BASE),
114 REQ(VHOST_USER_SET_LOG_FD),
115 REQ(VHOST_USER_SET_VRING_NUM),
116 REQ(VHOST_USER_SET_VRING_ADDR),
117 REQ(VHOST_USER_SET_VRING_BASE),
118 REQ(VHOST_USER_GET_VRING_BASE),
119 REQ(VHOST_USER_SET_VRING_KICK),
120 REQ(VHOST_USER_SET_VRING_CALL),
121 REQ(VHOST_USER_SET_VRING_ERR),
122 REQ(VHOST_USER_GET_PROTOCOL_FEATURES),
123 REQ(VHOST_USER_SET_PROTOCOL_FEATURES),
124 REQ(VHOST_USER_GET_QUEUE_NUM),
125 REQ(VHOST_USER_SET_VRING_ENABLE),
126 REQ(VHOST_USER_SEND_RARP),
127 REQ(VHOST_USER_NET_SET_MTU),
128 REQ(VHOST_USER_SET_SLAVE_REQ_FD),
129 REQ(VHOST_USER_IOTLB_MSG),
130 REQ(VHOST_USER_SET_VRING_ENDIAN),
131 REQ(VHOST_USER_GET_CONFIG),
132 REQ(VHOST_USER_SET_CONFIG),
133 REQ(VHOST_USER_POSTCOPY_ADVISE),
134 REQ(VHOST_USER_POSTCOPY_LISTEN),
135 REQ(VHOST_USER_POSTCOPY_END),
136 REQ(VHOST_USER_GET_INFLIGHT_FD),
137 REQ(VHOST_USER_SET_INFLIGHT_FD),
138 REQ(VHOST_USER_GPU_SET_SOCKET),
139 REQ(VHOST_USER_MAX),
140 };
141#undef REQ
142
143 if (req < VHOST_USER_MAX) {
144 return vu_request_str[req];
145 } else {
146 return "unknown";
147 }
148}
149
150static void
151vu_panic(VuDev *dev, const char *msg, ...)
152{
153 char *buf = NULL;
154 va_list ap;
155
156 va_start(ap, msg);
157 if (vasprintf(&buf, msg, ap) < 0) {
158 buf = NULL;
159 }
160 va_end(ap);
161
162 dev->broken = true;
163 dev->panic(dev, buf);
164 free(buf);
165
166 /* FIXME: find a way to call virtio_error? */
167}
168
169/* Translate guest physical address to our virtual address. */
170void *
171vu_gpa_to_va(VuDev *dev, uint64_t *plen, uint64_t guest_addr)
172{
173 int i;
174
175 if (*plen == 0) {
176 return NULL;
177 }
178
179 /* Find matching memory region. */
180 for (i = 0; i < dev->nregions; i++) {
181 VuDevRegion *r = &dev->regions[i];
182
183 if ((guest_addr >= r->gpa) && (guest_addr < (r->gpa + r->size))) {
184 if ((guest_addr + *plen) > (r->gpa + r->size)) {
185 *plen = r->gpa + r->size - guest_addr;
186 }
187 return (void *)(uintptr_t)
188 guest_addr - r->gpa + r->mmap_addr + r->mmap_offset;
189 }
190 }
191
192 return NULL;
193}
194
195/* Translate qemu virtual address to our virtual address. */
196static void *
197qva_to_va(VuDev *dev, uint64_t qemu_addr)
198{
199 int i;
200
201 /* Find matching memory region. */
202 for (i = 0; i < dev->nregions; i++) {
203 VuDevRegion *r = &dev->regions[i];
204
205 if ((qemu_addr >= r->qva) && (qemu_addr < (r->qva + r->size))) {
206 return (void *)(uintptr_t)
207 qemu_addr - r->qva + r->mmap_addr + r->mmap_offset;
208 }
209 }
210
211 return NULL;
212}
213
214static void
215vmsg_close_fds(VhostUserMsg *vmsg)
216{
217 int i;
218
219 for (i = 0; i < vmsg->fd_num; i++) {
220 close(vmsg->fds[i]);
221 }
222}
223
224/* Set reply payload.u64 and clear request flags and fd_num */
225static void vmsg_set_reply_u64(VhostUserMsg *vmsg, uint64_t val)
226{
227 vmsg->flags = 0; /* defaults will be set by vu_send_reply() */
228 vmsg->size = sizeof(vmsg->payload.u64);
229 vmsg->payload.u64 = val;
230 vmsg->fd_num = 0;
231}
232
233/* A test to see if we have userfault available */
234static bool
235have_userfault(void)
236{
237#if defined(__linux__) && defined(__NR_userfaultfd) &&\
238 defined(UFFD_FEATURE_MISSING_SHMEM) &&\
239 defined(UFFD_FEATURE_MISSING_HUGETLBFS)
240 /* Now test the kernel we're running on really has the features */
241 int ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
242 struct uffdio_api api_struct;
243 if (ufd < 0) {
244 return false;
245 }
246
247 api_struct.api = UFFD_API;
248 api_struct.features = UFFD_FEATURE_MISSING_SHMEM |
249 UFFD_FEATURE_MISSING_HUGETLBFS;
250 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
251 close(ufd);
252 return false;
253 }
254 close(ufd);
255 return true;
256
257#else
258 return false;
259#endif
260}
261
262static bool
263vu_message_read(VuDev *dev, int conn_fd, VhostUserMsg *vmsg)
264{
265 char control[CMSG_SPACE(VHOST_MEMORY_MAX_NREGIONS * sizeof(int))] = { };
266 struct iovec iov = {
267 .iov_base = (char *)vmsg,
268 .iov_len = VHOST_USER_HDR_SIZE,
269 };
270 struct msghdr msg = {
271 .msg_iov = &iov,
272 .msg_iovlen = 1,
273 .msg_control = control,
274 .msg_controllen = sizeof(control),
275 };
276 size_t fd_size;
277 struct cmsghdr *cmsg;
278 int rc;
279
280 do {
281 rc = recvmsg(conn_fd, &msg, 0);
282 } while (rc < 0 && (errno == EINTR || errno == EAGAIN));
283
284 if (rc < 0) {
285 vu_panic(dev, "Error while recvmsg: %s", strerror(errno));
286 return false;
287 }
288
289 vmsg->fd_num = 0;
290 for (cmsg = CMSG_FIRSTHDR(&msg);
291 cmsg != NULL;
292 cmsg = CMSG_NXTHDR(&msg, cmsg))
293 {
294 if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS) {
295 fd_size = cmsg->cmsg_len - CMSG_LEN(0);
296 vmsg->fd_num = fd_size / sizeof(int);
297 memcpy(vmsg->fds, CMSG_DATA(cmsg), fd_size);
298 break;
299 }
300 }
301
302 if (vmsg->size > sizeof(vmsg->payload)) {
303 vu_panic(dev,
304 "Error: too big message request: %d, size: vmsg->size: %u, "
305 "while sizeof(vmsg->payload) = %zu\n",
306 vmsg->request, vmsg->size, sizeof(vmsg->payload));
307 goto fail;
308 }
309
310 if (vmsg->size) {
311 do {
312 rc = read(conn_fd, &vmsg->payload, vmsg->size);
313 } while (rc < 0 && (errno == EINTR || errno == EAGAIN));
314
315 if (rc <= 0) {
316 vu_panic(dev, "Error while reading: %s", strerror(errno));
317 goto fail;
318 }
319
320 assert(rc == vmsg->size);
321 }
322
323 return true;
324
325fail:
326 vmsg_close_fds(vmsg);
327
328 return false;
329}
330
331static bool
332vu_message_write(VuDev *dev, int conn_fd, VhostUserMsg *vmsg)
333{
334 int rc;
335 uint8_t *p = (uint8_t *)vmsg;
336 char control[CMSG_SPACE(VHOST_MEMORY_MAX_NREGIONS * sizeof(int))] = { };
337 struct iovec iov = {
338 .iov_base = (char *)vmsg,
339 .iov_len = VHOST_USER_HDR_SIZE,
340 };
341 struct msghdr msg = {
342 .msg_iov = &iov,
343 .msg_iovlen = 1,
344 .msg_control = control,
345 };
346 struct cmsghdr *cmsg;
347
348 memset(control, 0, sizeof(control));
349 assert(vmsg->fd_num <= VHOST_MEMORY_MAX_NREGIONS);
350 if (vmsg->fd_num > 0) {
351 size_t fdsize = vmsg->fd_num * sizeof(int);
352 msg.msg_controllen = CMSG_SPACE(fdsize);
353 cmsg = CMSG_FIRSTHDR(&msg);
354 cmsg->cmsg_len = CMSG_LEN(fdsize);
355 cmsg->cmsg_level = SOL_SOCKET;
356 cmsg->cmsg_type = SCM_RIGHTS;
357 memcpy(CMSG_DATA(cmsg), vmsg->fds, fdsize);
358 } else {
359 msg.msg_controllen = 0;
360 }
361
362 do {
363 rc = sendmsg(conn_fd, &msg, 0);
364 } while (rc < 0 && (errno == EINTR || errno == EAGAIN));
365
366 if (vmsg->size) {
367 do {
368 if (vmsg->data) {
369 rc = write(conn_fd, vmsg->data, vmsg->size);
370 } else {
371 rc = write(conn_fd, p + VHOST_USER_HDR_SIZE, vmsg->size);
372 }
373 } while (rc < 0 && (errno == EINTR || errno == EAGAIN));
374 }
375
376 if (rc <= 0) {
377 vu_panic(dev, "Error while writing: %s", strerror(errno));
378 return false;
379 }
380
381 return true;
382}
383
384static bool
385vu_send_reply(VuDev *dev, int conn_fd, VhostUserMsg *vmsg)
386{
387 /* Set the version in the flags when sending the reply */
388 vmsg->flags &= ~VHOST_USER_VERSION_MASK;
389 vmsg->flags |= VHOST_USER_VERSION;
390 vmsg->flags |= VHOST_USER_REPLY_MASK;
391
392 return vu_message_write(dev, conn_fd, vmsg);
393}
394
395static bool
396vu_process_message_reply(VuDev *dev, const VhostUserMsg *vmsg)
397{
398 VhostUserMsg msg_reply;
399
400 if ((vmsg->flags & VHOST_USER_NEED_REPLY_MASK) == 0) {
401 return true;
402 }
403
404 if (!vu_message_read(dev, dev->slave_fd, &msg_reply)) {
405 return false;
406 }
407
408 if (msg_reply.request != vmsg->request) {
409 DPRINT("Received unexpected msg type. Expected %d received %d",
410 vmsg->request, msg_reply.request);
411 return false;
412 }
413
414 return msg_reply.payload.u64 == 0;
415}
416
417/* Kick the log_call_fd if required. */
418static void
419vu_log_kick(VuDev *dev)
420{
421 if (dev->log_call_fd != -1) {
422 DPRINT("Kicking the QEMU's log...\n");
423 if (eventfd_write(dev->log_call_fd, 1) < 0) {
424 vu_panic(dev, "Error writing eventfd: %s", strerror(errno));
425 }
426 }
427}
428
429static void
430vu_log_page(uint8_t *log_table, uint64_t page)
431{
432 DPRINT("Logged dirty guest page: %"PRId64"\n", page);
433 atomic_or(&log_table[page / 8], 1 << (page % 8));
434}
435
436static void
437vu_log_write(VuDev *dev, uint64_t address, uint64_t length)
438{
439 uint64_t page;
440
441 if (!(dev->features & (1ULL << VHOST_F_LOG_ALL)) ||
442 !dev->log_table || !length) {
443 return;
444 }
445
446 assert(dev->log_size > ((address + length - 1) / VHOST_LOG_PAGE / 8));
447
448 page = address / VHOST_LOG_PAGE;
449 while (page * VHOST_LOG_PAGE < address + length) {
450 vu_log_page(dev->log_table, page);
451 page += 1;
452 }
453
454 vu_log_kick(dev);
455}
456
457static void
458vu_kick_cb(VuDev *dev, int condition, void *data)
459{
460 int index = (intptr_t)data;
461 VuVirtq *vq = &dev->vq[index];
462 int sock = vq->kick_fd;
463 eventfd_t kick_data;
464 ssize_t rc;
465
466 rc = eventfd_read(sock, &kick_data);
467 if (rc == -1) {
468 vu_panic(dev, "kick eventfd_read(): %s", strerror(errno));
469 dev->remove_watch(dev, dev->vq[index].kick_fd);
470 } else {
471 DPRINT("Got kick_data: %016"PRIx64" handler:%p idx:%d\n",
472 kick_data, vq->handler, index);
473 if (vq->handler) {
474 vq->handler(dev, index);
475 }
476 }
477}
478
479static bool
480vu_get_features_exec(VuDev *dev, VhostUserMsg *vmsg)
481{
482 vmsg->payload.u64 =
483 1ULL << VHOST_F_LOG_ALL |
484 1ULL << VHOST_USER_F_PROTOCOL_FEATURES;
485
486 if (dev->iface->get_features) {
487 vmsg->payload.u64 |= dev->iface->get_features(dev);
488 }
489
490 vmsg->size = sizeof(vmsg->payload.u64);
491 vmsg->fd_num = 0;
492
493 DPRINT("Sending back to guest u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
494
495 return true;
496}
497
498static void
499vu_set_enable_all_rings(VuDev *dev, bool enabled)
500{
501 uint16_t i;
502
503 for (i = 0; i < dev->max_queues; i++) {
504 dev->vq[i].enable = enabled;
505 }
506}
507
508static bool
509vu_set_features_exec(VuDev *dev, VhostUserMsg *vmsg)
510{
511 DPRINT("u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
512
513 dev->features = vmsg->payload.u64;
514
515 if (!(dev->features & VHOST_USER_F_PROTOCOL_FEATURES)) {
516 vu_set_enable_all_rings(dev, true);
517 }
518
519 if (dev->iface->set_features) {
520 dev->iface->set_features(dev, dev->features);
521 }
522
523 return false;
524}
525
526static bool
527vu_set_owner_exec(VuDev *dev, VhostUserMsg *vmsg)
528{
529 return false;
530}
531
532static void
533vu_close_log(VuDev *dev)
534{
535 if (dev->log_table) {
536 if (munmap(dev->log_table, dev->log_size) != 0) {
537 perror("close log munmap() error");
538 }
539
540 dev->log_table = NULL;
541 }
542 if (dev->log_call_fd != -1) {
543 close(dev->log_call_fd);
544 dev->log_call_fd = -1;
545 }
546}
547
548static bool
549vu_reset_device_exec(VuDev *dev, VhostUserMsg *vmsg)
550{
551 vu_set_enable_all_rings(dev, false);
552
553 return false;
554}
555
556static bool
557vu_set_mem_table_exec_postcopy(VuDev *dev, VhostUserMsg *vmsg)
558{
559 int i;
560 VhostUserMemory m = vmsg->payload.memory, *memory = &m;
561 dev->nregions = memory->nregions;
562
563 DPRINT("Nregions: %d\n", memory->nregions);
564 for (i = 0; i < dev->nregions; i++) {
565 void *mmap_addr;
566 VhostUserMemoryRegion *msg_region = &memory->regions[i];
567 VuDevRegion *dev_region = &dev->regions[i];
568
569 DPRINT("Region %d\n", i);
570 DPRINT(" guest_phys_addr: 0x%016"PRIx64"\n",
571 msg_region->guest_phys_addr);
572 DPRINT(" memory_size: 0x%016"PRIx64"\n",
573 msg_region->memory_size);
574 DPRINT(" userspace_addr 0x%016"PRIx64"\n",
575 msg_region->userspace_addr);
576 DPRINT(" mmap_offset 0x%016"PRIx64"\n",
577 msg_region->mmap_offset);
578
579 dev_region->gpa = msg_region->guest_phys_addr;
580 dev_region->size = msg_region->memory_size;
581 dev_region->qva = msg_region->userspace_addr;
582 dev_region->mmap_offset = msg_region->mmap_offset;
583
584 /* We don't use offset argument of mmap() since the
585 * mapped address has to be page aligned, and we use huge
586 * pages.
587 * In postcopy we're using PROT_NONE here to catch anyone
588 * accessing it before we userfault
589 */
590 mmap_addr = mmap(0, dev_region->size + dev_region->mmap_offset,
591 PROT_NONE, MAP_SHARED,
592 vmsg->fds[i], 0);
593
594 if (mmap_addr == MAP_FAILED) {
595 vu_panic(dev, "region mmap error: %s", strerror(errno));
596 } else {
597 dev_region->mmap_addr = (uint64_t)(uintptr_t)mmap_addr;
598 DPRINT(" mmap_addr: 0x%016"PRIx64"\n",
599 dev_region->mmap_addr);
600 }
601
602 /* Return the address to QEMU so that it can translate the ufd
603 * fault addresses back.
604 */
605 msg_region->userspace_addr = (uintptr_t)(mmap_addr +
606 dev_region->mmap_offset);
607 close(vmsg->fds[i]);
608 }
609
610 /* Send the message back to qemu with the addresses filled in */
611 vmsg->fd_num = 0;
612 if (!vu_send_reply(dev, dev->sock, vmsg)) {
613 vu_panic(dev, "failed to respond to set-mem-table for postcopy");
614 return false;
615 }
616
617 /* Wait for QEMU to confirm that it's registered the handler for the
618 * faults.
619 */
620 if (!vu_message_read(dev, dev->sock, vmsg) ||
621 vmsg->size != sizeof(vmsg->payload.u64) ||
622 vmsg->payload.u64 != 0) {
623 vu_panic(dev, "failed to receive valid ack for postcopy set-mem-table");
624 return false;
625 }
626
627 /* OK, now we can go and register the memory and generate faults */
628 for (i = 0; i < dev->nregions; i++) {
629 VuDevRegion *dev_region = &dev->regions[i];
630 int ret;
631#ifdef UFFDIO_REGISTER
632 /* We should already have an open ufd. Mark each memory
633 * range as ufd.
634 * Discard any mapping we have here; note I can't use MADV_REMOVE
635 * or fallocate to make the hole since I don't want to lose
636 * data that's already arrived in the shared process.
637 * TODO: How to do hugepage
638 */
639 ret = madvise((void *)(uintptr_t)dev_region->mmap_addr,
640 dev_region->size + dev_region->mmap_offset,
641 MADV_DONTNEED);
642 if (ret) {
643 fprintf(stderr,
644 "%s: Failed to madvise(DONTNEED) region %d: %s\n",
645 __func__, i, strerror(errno));
646 }
647 /* Turn off transparent hugepages so we dont get lose wakeups
648 * in neighbouring pages.
649 * TODO: Turn this backon later.
650 */
651 ret = madvise((void *)(uintptr_t)dev_region->mmap_addr,
652 dev_region->size + dev_region->mmap_offset,
653 MADV_NOHUGEPAGE);
654 if (ret) {
655 /* Note: This can happen legally on kernels that are configured
656 * without madvise'able hugepages
657 */
658 fprintf(stderr,
659 "%s: Failed to madvise(NOHUGEPAGE) region %d: %s\n",
660 __func__, i, strerror(errno));
661 }
662 struct uffdio_register reg_struct;
663 reg_struct.range.start = (uintptr_t)dev_region->mmap_addr;
664 reg_struct.range.len = dev_region->size + dev_region->mmap_offset;
665 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
666
667 if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER, &reg_struct)) {
668 vu_panic(dev, "%s: Failed to userfault region %d "
669 "@%p + size:%zx offset: %zx: (ufd=%d)%s\n",
670 __func__, i,
671 dev_region->mmap_addr,
672 dev_region->size, dev_region->mmap_offset,
673 dev->postcopy_ufd, strerror(errno));
674 return false;
675 }
676 if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) {
677 vu_panic(dev, "%s Region (%d) doesn't support COPY",
678 __func__, i);
679 return false;
680 }
681 DPRINT("%s: region %d: Registered userfault for %"
682 PRIx64 " + %" PRIx64 "\n", __func__, i,
683 (uint64_t)reg_struct.range.start,
684 (uint64_t)reg_struct.range.len);
685 /* Now it's registered we can let the client at it */
686 if (mprotect((void *)(uintptr_t)dev_region->mmap_addr,
687 dev_region->size + dev_region->mmap_offset,
688 PROT_READ | PROT_WRITE)) {
689 vu_panic(dev, "failed to mprotect region %d for postcopy (%s)",
690 i, strerror(errno));
691 return false;
692 }
693 /* TODO: Stash 'zero' support flags somewhere */
694#endif
695 }
696
697 return false;
698}
699
700static bool
701vu_set_mem_table_exec(VuDev *dev, VhostUserMsg *vmsg)
702{
703 int i;
704 VhostUserMemory m = vmsg->payload.memory, *memory = &m;
705
706 for (i = 0; i < dev->nregions; i++) {
707 VuDevRegion *r = &dev->regions[i];
708 void *m = (void *) (uintptr_t) r->mmap_addr;
709
710 if (m) {
711 munmap(m, r->size + r->mmap_offset);
712 }
713 }
714 dev->nregions = memory->nregions;
715
716 if (dev->postcopy_listening) {
717 return vu_set_mem_table_exec_postcopy(dev, vmsg);
718 }
719
720 DPRINT("Nregions: %d\n", memory->nregions);
721 for (i = 0; i < dev->nregions; i++) {
722 void *mmap_addr;
723 VhostUserMemoryRegion *msg_region = &memory->regions[i];
724 VuDevRegion *dev_region = &dev->regions[i];
725
726 DPRINT("Region %d\n", i);
727 DPRINT(" guest_phys_addr: 0x%016"PRIx64"\n",
728 msg_region->guest_phys_addr);
729 DPRINT(" memory_size: 0x%016"PRIx64"\n",
730 msg_region->memory_size);
731 DPRINT(" userspace_addr 0x%016"PRIx64"\n",
732 msg_region->userspace_addr);
733 DPRINT(" mmap_offset 0x%016"PRIx64"\n",
734 msg_region->mmap_offset);
735
736 dev_region->gpa = msg_region->guest_phys_addr;
737 dev_region->size = msg_region->memory_size;
738 dev_region->qva = msg_region->userspace_addr;
739 dev_region->mmap_offset = msg_region->mmap_offset;
740
741 /* We don't use offset argument of mmap() since the
742 * mapped address has to be page aligned, and we use huge
743 * pages. */
744 mmap_addr = mmap(0, dev_region->size + dev_region->mmap_offset,
745 PROT_READ | PROT_WRITE, MAP_SHARED,
746 vmsg->fds[i], 0);
747
748 if (mmap_addr == MAP_FAILED) {
749 vu_panic(dev, "region mmap error: %s", strerror(errno));
750 } else {
751 dev_region->mmap_addr = (uint64_t)(uintptr_t)mmap_addr;
752 DPRINT(" mmap_addr: 0x%016"PRIx64"\n",
753 dev_region->mmap_addr);
754 }
755
756 close(vmsg->fds[i]);
757 }
758
759 return false;
760}
761
762static bool
763vu_set_log_base_exec(VuDev *dev, VhostUserMsg *vmsg)
764{
765 int fd;
766 uint64_t log_mmap_size, log_mmap_offset;
767 void *rc;
768
769 if (vmsg->fd_num != 1 ||
770 vmsg->size != sizeof(vmsg->payload.log)) {
771 vu_panic(dev, "Invalid log_base message");
772 return true;
773 }
774
775 fd = vmsg->fds[0];
776 log_mmap_offset = vmsg->payload.log.mmap_offset;
777 log_mmap_size = vmsg->payload.log.mmap_size;
778 DPRINT("Log mmap_offset: %"PRId64"\n", log_mmap_offset);
779 DPRINT("Log mmap_size: %"PRId64"\n", log_mmap_size);
780
781 rc = mmap(0, log_mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd,
782 log_mmap_offset);
783 close(fd);
784 if (rc == MAP_FAILED) {
785 perror("log mmap error");
786 }
787
788 if (dev->log_table) {
789 munmap(dev->log_table, dev->log_size);
790 }
791 dev->log_table = rc;
792 dev->log_size = log_mmap_size;
793
794 vmsg->size = sizeof(vmsg->payload.u64);
795 vmsg->fd_num = 0;
796
797 return true;
798}
799
800static bool
801vu_set_log_fd_exec(VuDev *dev, VhostUserMsg *vmsg)
802{
803 if (vmsg->fd_num != 1) {
804 vu_panic(dev, "Invalid log_fd message");
805 return false;
806 }
807
808 if (dev->log_call_fd != -1) {
809 close(dev->log_call_fd);
810 }
811 dev->log_call_fd = vmsg->fds[0];
812 DPRINT("Got log_call_fd: %d\n", vmsg->fds[0]);
813
814 return false;
815}
816
817static bool
818vu_set_vring_num_exec(VuDev *dev, VhostUserMsg *vmsg)
819{
820 unsigned int index = vmsg->payload.state.index;
821 unsigned int num = vmsg->payload.state.num;
822
823 DPRINT("State.index: %d\n", index);
824 DPRINT("State.num: %d\n", num);
825 dev->vq[index].vring.num = num;
826
827 return false;
828}
829
830static bool
831vu_set_vring_addr_exec(VuDev *dev, VhostUserMsg *vmsg)
832{
833 struct vhost_vring_addr addr = vmsg->payload.addr, *vra = &addr;
834 unsigned int index = vra->index;
835 VuVirtq *vq = &dev->vq[index];
836
837 DPRINT("vhost_vring_addr:\n");
838 DPRINT(" index: %d\n", vra->index);
839 DPRINT(" flags: %d\n", vra->flags);
840 DPRINT(" desc_user_addr: 0x%016" PRIx64 "\n", vra->desc_user_addr);
841 DPRINT(" used_user_addr: 0x%016" PRIx64 "\n", vra->used_user_addr);
842 DPRINT(" avail_user_addr: 0x%016" PRIx64 "\n", vra->avail_user_addr);
843 DPRINT(" log_guest_addr: 0x%016" PRIx64 "\n", vra->log_guest_addr);
844
845 vq->vring.flags = vra->flags;
846 vq->vring.desc = qva_to_va(dev, vra->desc_user_addr);
847 vq->vring.used = qva_to_va(dev, vra->used_user_addr);
848 vq->vring.avail = qva_to_va(dev, vra->avail_user_addr);
849 vq->vring.log_guest_addr = vra->log_guest_addr;
850
851 DPRINT("Setting virtq addresses:\n");
852 DPRINT(" vring_desc at %p\n", vq->vring.desc);
853 DPRINT(" vring_used at %p\n", vq->vring.used);
854 DPRINT(" vring_avail at %p\n", vq->vring.avail);
855
856 if (!(vq->vring.desc && vq->vring.used && vq->vring.avail)) {
857 vu_panic(dev, "Invalid vring_addr message");
858 return false;
859 }
860
861 vq->used_idx = vq->vring.used->idx;
862
863 if (vq->last_avail_idx != vq->used_idx) {
864 bool resume = dev->iface->queue_is_processed_in_order &&
865 dev->iface->queue_is_processed_in_order(dev, index);
866
867 DPRINT("Last avail index != used index: %u != %u%s\n",
868 vq->last_avail_idx, vq->used_idx,
869 resume ? ", resuming" : "");
870
871 if (resume) {
872 vq->shadow_avail_idx = vq->last_avail_idx = vq->used_idx;
873 }
874 }
875
876 return false;
877}
878
879static bool
880vu_set_vring_base_exec(VuDev *dev, VhostUserMsg *vmsg)
881{
882 unsigned int index = vmsg->payload.state.index;
883 unsigned int num = vmsg->payload.state.num;
884
885 DPRINT("State.index: %d\n", index);
886 DPRINT("State.num: %d\n", num);
887 dev->vq[index].shadow_avail_idx = dev->vq[index].last_avail_idx = num;
888
889 return false;
890}
891
892static bool
893vu_get_vring_base_exec(VuDev *dev, VhostUserMsg *vmsg)
894{
895 unsigned int index = vmsg->payload.state.index;
896
897 DPRINT("State.index: %d\n", index);
898 vmsg->payload.state.num = dev->vq[index].last_avail_idx;
899 vmsg->size = sizeof(vmsg->payload.state);
900
901 dev->vq[index].started = false;
902 if (dev->iface->queue_set_started) {
903 dev->iface->queue_set_started(dev, index, false);
904 }
905
906 if (dev->vq[index].call_fd != -1) {
907 close(dev->vq[index].call_fd);
908 dev->vq[index].call_fd = -1;
909 }
910 if (dev->vq[index].kick_fd != -1) {
911 dev->remove_watch(dev, dev->vq[index].kick_fd);
912 close(dev->vq[index].kick_fd);
913 dev->vq[index].kick_fd = -1;
914 }
915
916 return true;
917}
918
919static bool
920vu_check_queue_msg_file(VuDev *dev, VhostUserMsg *vmsg)
921{
922 int index = vmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
923
924 if (index >= dev->max_queues) {
925 vmsg_close_fds(vmsg);
926 vu_panic(dev, "Invalid queue index: %u", index);
927 return false;
928 }
929
930 if (vmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK ||
931 vmsg->fd_num != 1) {
932 vmsg_close_fds(vmsg);
933 vu_panic(dev, "Invalid fds in request: %d", vmsg->request);
934 return false;
935 }
936
937 return true;
938}
939
940static int
941inflight_desc_compare(const void *a, const void *b)
942{
943 VuVirtqInflightDesc *desc0 = (VuVirtqInflightDesc *)a,
944 *desc1 = (VuVirtqInflightDesc *)b;
945
946 if (desc1->counter > desc0->counter &&
947 (desc1->counter - desc0->counter) < VIRTQUEUE_MAX_SIZE * 2) {
948 return 1;
949 }
950
951 return -1;
952}
953
954static int
955vu_check_queue_inflights(VuDev *dev, VuVirtq *vq)
956{
957 int i = 0;
958
959 if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)) {
960 return 0;
961 }
962
963 if (unlikely(!vq->inflight)) {
964 return -1;
965 }
966
967 if (unlikely(!vq->inflight->version)) {
968 /* initialize the buffer */
969 vq->inflight->version = INFLIGHT_VERSION;
970 return 0;
971 }
972
973 vq->used_idx = vq->vring.used->idx;
974 vq->resubmit_num = 0;
975 vq->resubmit_list = NULL;
976 vq->counter = 0;
977
978 if (unlikely(vq->inflight->used_idx != vq->used_idx)) {
979 vq->inflight->desc[vq->inflight->last_batch_head].inflight = 0;
980
981 barrier();
982
983 vq->inflight->used_idx = vq->used_idx;
984 }
985
986 for (i = 0; i < vq->inflight->desc_num; i++) {
987 if (vq->inflight->desc[i].inflight == 1) {
988 vq->inuse++;
989 }
990 }
991
992 vq->shadow_avail_idx = vq->last_avail_idx = vq->inuse + vq->used_idx;
993
994 if (vq->inuse) {
995 vq->resubmit_list = malloc(sizeof(VuVirtqInflightDesc) * vq->inuse);
996 if (!vq->resubmit_list) {
997 return -1;
998 }
999
1000 for (i = 0; i < vq->inflight->desc_num; i++) {
1001 if (vq->inflight->desc[i].inflight) {
1002 vq->resubmit_list[vq->resubmit_num].index = i;
1003 vq->resubmit_list[vq->resubmit_num].counter =
1004 vq->inflight->desc[i].counter;
1005 vq->resubmit_num++;
1006 }
1007 }
1008
1009 if (vq->resubmit_num > 1) {
1010 qsort(vq->resubmit_list, vq->resubmit_num,
1011 sizeof(VuVirtqInflightDesc), inflight_desc_compare);
1012 }
1013 vq->counter = vq->resubmit_list[0].counter + 1;
1014 }
1015
1016 /* in case of I/O hang after reconnecting */
1017 if (eventfd_write(vq->kick_fd, 1)) {
1018 return -1;
1019 }
1020
1021 return 0;
1022}
1023
1024static bool
1025vu_set_vring_kick_exec(VuDev *dev, VhostUserMsg *vmsg)
1026{
1027 int index = vmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1028
1029 DPRINT("u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
1030
1031 if (!vu_check_queue_msg_file(dev, vmsg)) {
1032 return false;
1033 }
1034
1035 if (dev->vq[index].kick_fd != -1) {
1036 dev->remove_watch(dev, dev->vq[index].kick_fd);
1037 close(dev->vq[index].kick_fd);
1038 dev->vq[index].kick_fd = -1;
1039 }
1040
1041 dev->vq[index].kick_fd = vmsg->fds[0];
1042 DPRINT("Got kick_fd: %d for vq: %d\n", vmsg->fds[0], index);
1043
1044 dev->vq[index].started = true;
1045 if (dev->iface->queue_set_started) {
1046 dev->iface->queue_set_started(dev, index, true);
1047 }
1048
1049 if (dev->vq[index].kick_fd != -1 && dev->vq[index].handler) {
1050 dev->set_watch(dev, dev->vq[index].kick_fd, VU_WATCH_IN,
1051 vu_kick_cb, (void *)(long)index);
1052
1053 DPRINT("Waiting for kicks on fd: %d for vq: %d\n",
1054 dev->vq[index].kick_fd, index);
1055 }
1056
1057 if (vu_check_queue_inflights(dev, &dev->vq[index])) {
1058 vu_panic(dev, "Failed to check inflights for vq: %d\n", index);
1059 }
1060
1061 return false;
1062}
1063
1064void vu_set_queue_handler(VuDev *dev, VuVirtq *vq,
1065 vu_queue_handler_cb handler)
1066{
1067 int qidx = vq - dev->vq;
1068
1069 vq->handler = handler;
1070 if (vq->kick_fd >= 0) {
1071 if (handler) {
1072 dev->set_watch(dev, vq->kick_fd, VU_WATCH_IN,
1073 vu_kick_cb, (void *)(long)qidx);
1074 } else {
1075 dev->remove_watch(dev, vq->kick_fd);
1076 }
1077 }
1078}
1079
1080bool vu_set_queue_host_notifier(VuDev *dev, VuVirtq *vq, int fd,
1081 int size, int offset)
1082{
1083 int qidx = vq - dev->vq;
1084 int fd_num = 0;
1085 VhostUserMsg vmsg = {
1086 .request = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
1087 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY_MASK,
1088 .size = sizeof(vmsg.payload.area),
1089 .payload.area = {
1090 .u64 = qidx & VHOST_USER_VRING_IDX_MASK,
1091 .size = size,
1092 .offset = offset,
1093 },
1094 };
1095
1096 if (fd == -1) {
1097 vmsg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
1098 } else {
1099 vmsg.fds[fd_num++] = fd;
1100 }
1101
1102 vmsg.fd_num = fd_num;
1103
1104 if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) {
1105 return false;
1106 }
1107
1108 if (!vu_message_write(dev, dev->slave_fd, &vmsg)) {
1109 return false;
1110 }
1111
1112 return vu_process_message_reply(dev, &vmsg);
1113}
1114
1115static bool
1116vu_set_vring_call_exec(VuDev *dev, VhostUserMsg *vmsg)
1117{
1118 int index = vmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1119
1120 DPRINT("u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
1121
1122 if (!vu_check_queue_msg_file(dev, vmsg)) {
1123 return false;
1124 }
1125
1126 if (dev->vq[index].call_fd != -1) {
1127 close(dev->vq[index].call_fd);
1128 dev->vq[index].call_fd = -1;
1129 }
1130
1131 dev->vq[index].call_fd = vmsg->fds[0];
1132
1133 /* in case of I/O hang after reconnecting */
1134 if (eventfd_write(vmsg->fds[0], 1)) {
1135 return -1;
1136 }
1137
1138 DPRINT("Got call_fd: %d for vq: %d\n", vmsg->fds[0], index);
1139
1140 return false;
1141}
1142
1143static bool
1144vu_set_vring_err_exec(VuDev *dev, VhostUserMsg *vmsg)
1145{
1146 int index = vmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1147
1148 DPRINT("u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
1149
1150 if (!vu_check_queue_msg_file(dev, vmsg)) {
1151 return false;
1152 }
1153
1154 if (dev->vq[index].err_fd != -1) {
1155 close(dev->vq[index].err_fd);
1156 dev->vq[index].err_fd = -1;
1157 }
1158
1159 dev->vq[index].err_fd = vmsg->fds[0];
1160
1161 return false;
1162}
1163
1164static bool
1165vu_get_protocol_features_exec(VuDev *dev, VhostUserMsg *vmsg)
1166{
1167 uint64_t features = 1ULL << VHOST_USER_PROTOCOL_F_MQ |
1168 1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD |
1169 1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ |
1170 1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER |
1171 1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD;
1172
1173 if (have_userfault()) {
1174 features |= 1ULL << VHOST_USER_PROTOCOL_F_PAGEFAULT;
1175 }
1176
1177 if (dev->iface->get_config && dev->iface->set_config) {
1178 features |= 1ULL << VHOST_USER_PROTOCOL_F_CONFIG;
1179 }
1180
1181 if (dev->iface->get_protocol_features) {
1182 features |= dev->iface->get_protocol_features(dev);
1183 }
1184
1185 vmsg_set_reply_u64(vmsg, features);
1186 return true;
1187}
1188
1189static bool
1190vu_set_protocol_features_exec(VuDev *dev, VhostUserMsg *vmsg)
1191{
1192 uint64_t features = vmsg->payload.u64;
1193
1194 DPRINT("u64: 0x%016"PRIx64"\n", features);
1195
1196 dev->protocol_features = vmsg->payload.u64;
1197
1198 if (dev->iface->set_protocol_features) {
1199 dev->iface->set_protocol_features(dev, features);
1200 }
1201
1202 return false;
1203}
1204
1205static bool
1206vu_get_queue_num_exec(VuDev *dev, VhostUserMsg *vmsg)
1207{
1208 vmsg_set_reply_u64(vmsg, dev->max_queues);
1209 return true;
1210}
1211
1212static bool
1213vu_set_vring_enable_exec(VuDev *dev, VhostUserMsg *vmsg)
1214{
1215 unsigned int index = vmsg->payload.state.index;
1216 unsigned int enable = vmsg->payload.state.num;
1217
1218 DPRINT("State.index: %d\n", index);
1219 DPRINT("State.enable: %d\n", enable);
1220
1221 if (index >= dev->max_queues) {
1222 vu_panic(dev, "Invalid vring_enable index: %u", index);
1223 return false;
1224 }
1225
1226 dev->vq[index].enable = enable;
1227 return false;
1228}
1229
1230static bool
1231vu_set_slave_req_fd(VuDev *dev, VhostUserMsg *vmsg)
1232{
1233 if (vmsg->fd_num != 1) {
1234 vu_panic(dev, "Invalid slave_req_fd message (%d fd's)", vmsg->fd_num);
1235 return false;
1236 }
1237
1238 if (dev->slave_fd != -1) {
1239 close(dev->slave_fd);
1240 }
1241 dev->slave_fd = vmsg->fds[0];
1242 DPRINT("Got slave_fd: %d\n", vmsg->fds[0]);
1243
1244 return false;
1245}
1246
1247static bool
1248vu_get_config(VuDev *dev, VhostUserMsg *vmsg)
1249{
1250 int ret = -1;
1251
1252 if (dev->iface->get_config) {
1253 ret = dev->iface->get_config(dev, vmsg->payload.config.region,
1254 vmsg->payload.config.size);
1255 }
1256
1257 if (ret) {
1258 /* resize to zero to indicate an error to master */
1259 vmsg->size = 0;
1260 }
1261
1262 return true;
1263}
1264
1265static bool
1266vu_set_config(VuDev *dev, VhostUserMsg *vmsg)
1267{
1268 int ret = -1;
1269
1270 if (dev->iface->set_config) {
1271 ret = dev->iface->set_config(dev, vmsg->payload.config.region,
1272 vmsg->payload.config.offset,
1273 vmsg->payload.config.size,
1274 vmsg->payload.config.flags);
1275 if (ret) {
1276 vu_panic(dev, "Set virtio configuration space failed");
1277 }
1278 }
1279
1280 return false;
1281}
1282
1283static bool
1284vu_set_postcopy_advise(VuDev *dev, VhostUserMsg *vmsg)
1285{
1286 dev->postcopy_ufd = -1;
1287#ifdef UFFDIO_API
1288 struct uffdio_api api_struct;
1289
1290 dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1291 vmsg->size = 0;
1292#endif
1293
1294 if (dev->postcopy_ufd == -1) {
1295 vu_panic(dev, "Userfaultfd not available: %s", strerror(errno));
1296 goto out;
1297 }
1298
1299#ifdef UFFDIO_API
1300 api_struct.api = UFFD_API;
1301 api_struct.features = 0;
1302 if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1303 vu_panic(dev, "Failed UFFDIO_API: %s", strerror(errno));
1304 close(dev->postcopy_ufd);
1305 dev->postcopy_ufd = -1;
1306 goto out;
1307 }
1308 /* TODO: Stash feature flags somewhere */
1309#endif
1310
1311out:
1312 /* Return a ufd to the QEMU */
1313 vmsg->fd_num = 1;
1314 vmsg->fds[0] = dev->postcopy_ufd;
1315 return true; /* = send a reply */
1316}
1317
1318static bool
1319vu_set_postcopy_listen(VuDev *dev, VhostUserMsg *vmsg)
1320{
1321 if (dev->nregions) {
1322 vu_panic(dev, "Regions already registered at postcopy-listen");
1323 vmsg_set_reply_u64(vmsg, -1);
1324 return true;
1325 }
1326 dev->postcopy_listening = true;
1327
1328 vmsg_set_reply_u64(vmsg, 0);
1329 return true;
1330}
1331
1332static bool
1333vu_set_postcopy_end(VuDev *dev, VhostUserMsg *vmsg)
1334{
1335 DPRINT("%s: Entry\n", __func__);
1336 dev->postcopy_listening = false;
1337 if (dev->postcopy_ufd > 0) {
1338 close(dev->postcopy_ufd);
1339 dev->postcopy_ufd = -1;
1340 DPRINT("%s: Done close\n", __func__);
1341 }
1342
1343 vmsg_set_reply_u64(vmsg, 0);
1344 DPRINT("%s: exit\n", __func__);
1345 return true;
1346}
1347
1348static inline uint64_t
1349vu_inflight_queue_size(uint16_t queue_size)
1350{
1351 return ALIGN_UP(sizeof(VuDescStateSplit) * queue_size +
1352 sizeof(uint16_t), INFLIGHT_ALIGNMENT);
1353}
1354
1355static bool
1356vu_get_inflight_fd(VuDev *dev, VhostUserMsg *vmsg)
1357{
1358 int fd;
1359 void *addr;
1360 uint64_t mmap_size;
1361 uint16_t num_queues, queue_size;
1362
1363 if (vmsg->size != sizeof(vmsg->payload.inflight)) {
1364 vu_panic(dev, "Invalid get_inflight_fd message:%d", vmsg->size);
1365 vmsg->payload.inflight.mmap_size = 0;
1366 return true;
1367 }
1368
1369 num_queues = vmsg->payload.inflight.num_queues;
1370 queue_size = vmsg->payload.inflight.queue_size;
1371
1372 DPRINT("set_inflight_fd num_queues: %"PRId16"\n", num_queues);
1373 DPRINT("set_inflight_fd queue_size: %"PRId16"\n", queue_size);
1374
1375 mmap_size = vu_inflight_queue_size(queue_size) * num_queues;
1376
1377 addr = qemu_memfd_alloc("vhost-inflight", mmap_size,
1378 F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1379 &fd, NULL);
1380
1381 if (!addr) {
1382 vu_panic(dev, "Failed to alloc vhost inflight area");
1383 vmsg->payload.inflight.mmap_size = 0;
1384 return true;
1385 }
1386
1387 memset(addr, 0, mmap_size);
1388
1389 dev->inflight_info.addr = addr;
1390 dev->inflight_info.size = vmsg->payload.inflight.mmap_size = mmap_size;
1391 dev->inflight_info.fd = vmsg->fds[0] = fd;
1392 vmsg->fd_num = 1;
1393 vmsg->payload.inflight.mmap_offset = 0;
1394
1395 DPRINT("send inflight mmap_size: %"PRId64"\n",
1396 vmsg->payload.inflight.mmap_size);
1397 DPRINT("send inflight mmap offset: %"PRId64"\n",
1398 vmsg->payload.inflight.mmap_offset);
1399
1400 return true;
1401}
1402
1403static bool
1404vu_set_inflight_fd(VuDev *dev, VhostUserMsg *vmsg)
1405{
1406 int fd, i;
1407 uint64_t mmap_size, mmap_offset;
1408 uint16_t num_queues, queue_size;
1409 void *rc;
1410
1411 if (vmsg->fd_num != 1 ||
1412 vmsg->size != sizeof(vmsg->payload.inflight)) {
1413 vu_panic(dev, "Invalid set_inflight_fd message size:%d fds:%d",
1414 vmsg->size, vmsg->fd_num);
1415 return false;
1416 }
1417
1418 fd = vmsg->fds[0];
1419 mmap_size = vmsg->payload.inflight.mmap_size;
1420 mmap_offset = vmsg->payload.inflight.mmap_offset;
1421 num_queues = vmsg->payload.inflight.num_queues;
1422 queue_size = vmsg->payload.inflight.queue_size;
1423
1424 DPRINT("set_inflight_fd mmap_size: %"PRId64"\n", mmap_size);
1425 DPRINT("set_inflight_fd mmap_offset: %"PRId64"\n", mmap_offset);
1426 DPRINT("set_inflight_fd num_queues: %"PRId16"\n", num_queues);
1427 DPRINT("set_inflight_fd queue_size: %"PRId16"\n", queue_size);
1428
1429 rc = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1430 fd, mmap_offset);
1431
1432 if (rc == MAP_FAILED) {
1433 vu_panic(dev, "set_inflight_fd mmap error: %s", strerror(errno));
1434 return false;
1435 }
1436
1437 if (dev->inflight_info.fd) {
1438 close(dev->inflight_info.fd);
1439 }
1440
1441 if (dev->inflight_info.addr) {
1442 munmap(dev->inflight_info.addr, dev->inflight_info.size);
1443 }
1444
1445 dev->inflight_info.fd = fd;
1446 dev->inflight_info.addr = rc;
1447 dev->inflight_info.size = mmap_size;
1448
1449 for (i = 0; i < num_queues; i++) {
1450 dev->vq[i].inflight = (VuVirtqInflight *)rc;
1451 dev->vq[i].inflight->desc_num = queue_size;
1452 rc = (void *)((char *)rc + vu_inflight_queue_size(queue_size));
1453 }
1454
1455 return false;
1456}
1457
1458static bool
1459vu_process_message(VuDev *dev, VhostUserMsg *vmsg)
1460{
1461 int do_reply = 0;
1462
1463 /* Print out generic part of the request. */
1464 DPRINT("================ Vhost user message ================\n");
1465 DPRINT("Request: %s (%d)\n", vu_request_to_string(vmsg->request),
1466 vmsg->request);
1467 DPRINT("Flags: 0x%x\n", vmsg->flags);
1468 DPRINT("Size: %d\n", vmsg->size);
1469
1470 if (vmsg->fd_num) {
1471 int i;
1472 DPRINT("Fds:");
1473 for (i = 0; i < vmsg->fd_num; i++) {
1474 DPRINT(" %d", vmsg->fds[i]);
1475 }
1476 DPRINT("\n");
1477 }
1478
1479 if (dev->iface->process_msg &&
1480 dev->iface->process_msg(dev, vmsg, &do_reply)) {
1481 return do_reply;
1482 }
1483
1484 switch (vmsg->request) {
1485 case VHOST_USER_GET_FEATURES:
1486 return vu_get_features_exec(dev, vmsg);
1487 case VHOST_USER_SET_FEATURES:
1488 return vu_set_features_exec(dev, vmsg);
1489 case VHOST_USER_GET_PROTOCOL_FEATURES:
1490 return vu_get_protocol_features_exec(dev, vmsg);
1491 case VHOST_USER_SET_PROTOCOL_FEATURES:
1492 return vu_set_protocol_features_exec(dev, vmsg);
1493 case VHOST_USER_SET_OWNER:
1494 return vu_set_owner_exec(dev, vmsg);
1495 case VHOST_USER_RESET_OWNER:
1496 return vu_reset_device_exec(dev, vmsg);
1497 case VHOST_USER_SET_MEM_TABLE:
1498 return vu_set_mem_table_exec(dev, vmsg);
1499 case VHOST_USER_SET_LOG_BASE:
1500 return vu_set_log_base_exec(dev, vmsg);
1501 case VHOST_USER_SET_LOG_FD:
1502 return vu_set_log_fd_exec(dev, vmsg);
1503 case VHOST_USER_SET_VRING_NUM:
1504 return vu_set_vring_num_exec(dev, vmsg);
1505 case VHOST_USER_SET_VRING_ADDR:
1506 return vu_set_vring_addr_exec(dev, vmsg);
1507 case VHOST_USER_SET_VRING_BASE:
1508 return vu_set_vring_base_exec(dev, vmsg);
1509 case VHOST_USER_GET_VRING_BASE:
1510 return vu_get_vring_base_exec(dev, vmsg);
1511 case VHOST_USER_SET_VRING_KICK:
1512 return vu_set_vring_kick_exec(dev, vmsg);
1513 case VHOST_USER_SET_VRING_CALL:
1514 return vu_set_vring_call_exec(dev, vmsg);
1515 case VHOST_USER_SET_VRING_ERR:
1516 return vu_set_vring_err_exec(dev, vmsg);
1517 case VHOST_USER_GET_QUEUE_NUM:
1518 return vu_get_queue_num_exec(dev, vmsg);
1519 case VHOST_USER_SET_VRING_ENABLE:
1520 return vu_set_vring_enable_exec(dev, vmsg);
1521 case VHOST_USER_SET_SLAVE_REQ_FD:
1522 return vu_set_slave_req_fd(dev, vmsg);
1523 case VHOST_USER_GET_CONFIG:
1524 return vu_get_config(dev, vmsg);
1525 case VHOST_USER_SET_CONFIG:
1526 return vu_set_config(dev, vmsg);
1527 case VHOST_USER_NONE:
1528 /* if you need processing before exit, override iface->process_msg */
1529 exit(0);
1530 case VHOST_USER_POSTCOPY_ADVISE:
1531 return vu_set_postcopy_advise(dev, vmsg);
1532 case VHOST_USER_POSTCOPY_LISTEN:
1533 return vu_set_postcopy_listen(dev, vmsg);
1534 case VHOST_USER_POSTCOPY_END:
1535 return vu_set_postcopy_end(dev, vmsg);
1536 case VHOST_USER_GET_INFLIGHT_FD:
1537 return vu_get_inflight_fd(dev, vmsg);
1538 case VHOST_USER_SET_INFLIGHT_FD:
1539 return vu_set_inflight_fd(dev, vmsg);
1540 default:
1541 vmsg_close_fds(vmsg);
1542 vu_panic(dev, "Unhandled request: %d", vmsg->request);
1543 }
1544
1545 return false;
1546}
1547
1548bool
1549vu_dispatch(VuDev *dev)
1550{
1551 VhostUserMsg vmsg = { 0, };
1552 int reply_requested;
1553 bool success = false;
1554
1555 if (!vu_message_read(dev, dev->sock, &vmsg)) {
1556 goto end;
1557 }
1558
1559 reply_requested = vu_process_message(dev, &vmsg);
1560 if (!reply_requested) {
1561 success = true;
1562 goto end;
1563 }
1564
1565 if (!vu_send_reply(dev, dev->sock, &vmsg)) {
1566 goto end;
1567 }
1568
1569 success = true;
1570
1571end:
1572 free(vmsg.data);
1573 return success;
1574}
1575
1576void
1577vu_deinit(VuDev *dev)
1578{
1579 int i;
1580
1581 for (i = 0; i < dev->nregions; i++) {
1582 VuDevRegion *r = &dev->regions[i];
1583 void *m = (void *) (uintptr_t) r->mmap_addr;
1584 if (m != MAP_FAILED) {
1585 munmap(m, r->size + r->mmap_offset);
1586 }
1587 }
1588 dev->nregions = 0;
1589
1590 for (i = 0; i < dev->max_queues; i++) {
1591 VuVirtq *vq = &dev->vq[i];
1592
1593 if (vq->call_fd != -1) {
1594 close(vq->call_fd);
1595 vq->call_fd = -1;
1596 }
1597
1598 if (vq->kick_fd != -1) {
1599 close(vq->kick_fd);
1600 vq->kick_fd = -1;
1601 }
1602
1603 if (vq->err_fd != -1) {
1604 close(vq->err_fd);
1605 vq->err_fd = -1;
1606 }
1607
1608 if (vq->resubmit_list) {
1609 free(vq->resubmit_list);
1610 vq->resubmit_list = NULL;
1611 }
1612
1613 vq->inflight = NULL;
1614 }
1615
1616 if (dev->inflight_info.addr) {
1617 munmap(dev->inflight_info.addr, dev->inflight_info.size);
1618 dev->inflight_info.addr = NULL;
1619 }
1620
1621 if (dev->inflight_info.fd > 0) {
1622 close(dev->inflight_info.fd);
1623 dev->inflight_info.fd = -1;
1624 }
1625
1626 vu_close_log(dev);
1627 if (dev->slave_fd != -1) {
1628 close(dev->slave_fd);
1629 dev->slave_fd = -1;
1630 }
1631
1632 if (dev->sock != -1) {
1633 close(dev->sock);
1634 }
1635
1636 free(dev->vq);
1637 dev->vq = NULL;
1638}
1639
1640bool
1641vu_init(VuDev *dev,
1642 uint16_t max_queues,
1643 int socket,
1644 vu_panic_cb panic,
1645 vu_set_watch_cb set_watch,
1646 vu_remove_watch_cb remove_watch,
1647 const VuDevIface *iface)
1648{
1649 uint16_t i;
1650
1651 assert(max_queues > 0);
1652 assert(socket >= 0);
1653 assert(set_watch);
1654 assert(remove_watch);
1655 assert(iface);
1656 assert(panic);
1657
1658 memset(dev, 0, sizeof(*dev));
1659
1660 dev->sock = socket;
1661 dev->panic = panic;
1662 dev->set_watch = set_watch;
1663 dev->remove_watch = remove_watch;
1664 dev->iface = iface;
1665 dev->log_call_fd = -1;
1666 dev->slave_fd = -1;
1667 dev->max_queues = max_queues;
1668
1669 dev->vq = malloc(max_queues * sizeof(dev->vq[0]));
1670 if (!dev->vq) {
1671 DPRINT("%s: failed to malloc virtqueues\n", __func__);
1672 return false;
1673 }
1674
1675 for (i = 0; i < max_queues; i++) {
1676 dev->vq[i] = (VuVirtq) {
1677 .call_fd = -1, .kick_fd = -1, .err_fd = -1,
1678 .notification = true,
1679 };
1680 }
1681
1682 return true;
1683}
1684
1685VuVirtq *
1686vu_get_queue(VuDev *dev, int qidx)
1687{
1688 assert(qidx < dev->max_queues);
1689 return &dev->vq[qidx];
1690}
1691
1692bool
1693vu_queue_enabled(VuDev *dev, VuVirtq *vq)
1694{
1695 return vq->enable;
1696}
1697
1698bool
1699vu_queue_started(const VuDev *dev, const VuVirtq *vq)
1700{
1701 return vq->started;
1702}
1703
1704static inline uint16_t
1705vring_avail_flags(VuVirtq *vq)
1706{
1707 return vq->vring.avail->flags;
1708}
1709
1710static inline uint16_t
1711vring_avail_idx(VuVirtq *vq)
1712{
1713 vq->shadow_avail_idx = vq->vring.avail->idx;
1714
1715 return vq->shadow_avail_idx;
1716}
1717
1718static inline uint16_t
1719vring_avail_ring(VuVirtq *vq, int i)
1720{
1721 return vq->vring.avail->ring[i];
1722}
1723
1724static inline uint16_t
1725vring_get_used_event(VuVirtq *vq)
1726{
1727 return vring_avail_ring(vq, vq->vring.num);
1728}
1729
1730static int
1731virtqueue_num_heads(VuDev *dev, VuVirtq *vq, unsigned int idx)
1732{
1733 uint16_t num_heads = vring_avail_idx(vq) - idx;
1734
1735 /* Check it isn't doing very strange things with descriptor numbers. */
1736 if (num_heads > vq->vring.num) {
1737 vu_panic(dev, "Guest moved used index from %u to %u",
1738 idx, vq->shadow_avail_idx);
1739 return -1;
1740 }
1741 if (num_heads) {
1742 /* On success, callers read a descriptor at vq->last_avail_idx.
1743 * Make sure descriptor read does not bypass avail index read. */
1744 smp_rmb();
1745 }
1746
1747 return num_heads;
1748}
1749
1750static bool
1751virtqueue_get_head(VuDev *dev, VuVirtq *vq,
1752 unsigned int idx, unsigned int *head)
1753{
1754 /* Grab the next descriptor number they're advertising, and increment
1755 * the index we've seen. */
1756 *head = vring_avail_ring(vq, idx % vq->vring.num);
1757
1758 /* If their number is silly, that's a fatal mistake. */
1759 if (*head >= vq->vring.num) {
1760 vu_panic(dev, "Guest says index %u is available", head);
1761 return false;
1762 }
1763
1764 return true;
1765}
1766
1767static int
1768virtqueue_read_indirect_desc(VuDev *dev, struct vring_desc *desc,
1769 uint64_t addr, size_t len)
1770{
1771 struct vring_desc *ori_desc;
1772 uint64_t read_len;
1773
1774 if (len > (VIRTQUEUE_MAX_SIZE * sizeof(struct vring_desc))) {
1775 return -1;
1776 }
1777
1778 if (len == 0) {
1779 return -1;
1780 }
1781
1782 while (len) {
1783 read_len = len;
1784 ori_desc = vu_gpa_to_va(dev, &read_len, addr);
1785 if (!ori_desc) {
1786 return -1;
1787 }
1788
1789 memcpy(desc, ori_desc, read_len);
1790 len -= read_len;
1791 addr += read_len;
1792 desc += read_len;
1793 }
1794
1795 return 0;
1796}
1797
1798enum {
1799 VIRTQUEUE_READ_DESC_ERROR = -1,
1800 VIRTQUEUE_READ_DESC_DONE = 0, /* end of chain */
1801 VIRTQUEUE_READ_DESC_MORE = 1, /* more buffers in chain */
1802};
1803
1804static int
1805virtqueue_read_next_desc(VuDev *dev, struct vring_desc *desc,
1806 int i, unsigned int max, unsigned int *next)
1807{
1808 /* If this descriptor says it doesn't chain, we're done. */
1809 if (!(desc[i].flags & VRING_DESC_F_NEXT)) {
1810 return VIRTQUEUE_READ_DESC_DONE;
1811 }
1812
1813 /* Check they're not leading us off end of descriptors. */
1814 *next = desc[i].next;
1815 /* Make sure compiler knows to grab that: we don't want it changing! */
1816 smp_wmb();
1817
1818 if (*next >= max) {
1819 vu_panic(dev, "Desc next is %u", next);
1820 return VIRTQUEUE_READ_DESC_ERROR;
1821 }
1822
1823 return VIRTQUEUE_READ_DESC_MORE;
1824}
1825
1826void
1827vu_queue_get_avail_bytes(VuDev *dev, VuVirtq *vq, unsigned int *in_bytes,
1828 unsigned int *out_bytes,
1829 unsigned max_in_bytes, unsigned max_out_bytes)
1830{
1831 unsigned int idx;
1832 unsigned int total_bufs, in_total, out_total;
1833 int rc;
1834
1835 idx = vq->last_avail_idx;
1836
1837 total_bufs = in_total = out_total = 0;
1838 if (unlikely(dev->broken) ||
1839 unlikely(!vq->vring.avail)) {
1840 goto done;
1841 }
1842
1843 while ((rc = virtqueue_num_heads(dev, vq, idx)) > 0) {
1844 unsigned int max, desc_len, num_bufs, indirect = 0;
1845 uint64_t desc_addr, read_len;
1846 struct vring_desc *desc;
1847 struct vring_desc desc_buf[VIRTQUEUE_MAX_SIZE];
1848 unsigned int i;
1849
1850 max = vq->vring.num;
1851 num_bufs = total_bufs;
1852 if (!virtqueue_get_head(dev, vq, idx++, &i)) {
1853 goto err;
1854 }
1855 desc = vq->vring.desc;
1856
1857 if (desc[i].flags & VRING_DESC_F_INDIRECT) {
1858 if (desc[i].len % sizeof(struct vring_desc)) {
1859 vu_panic(dev, "Invalid size for indirect buffer table");
1860 goto err;
1861 }
1862
1863 /* If we've got too many, that implies a descriptor loop. */
1864 if (num_bufs >= max) {
1865 vu_panic(dev, "Looped descriptor");
1866 goto err;
1867 }
1868
1869 /* loop over the indirect descriptor table */
1870 indirect = 1;
1871 desc_addr = desc[i].addr;
1872 desc_len = desc[i].len;
1873 max = desc_len / sizeof(struct vring_desc);
1874 read_len = desc_len;
1875 desc = vu_gpa_to_va(dev, &read_len, desc_addr);
1876 if (unlikely(desc && read_len != desc_len)) {
1877 /* Failed to use zero copy */
1878 desc = NULL;
1879 if (!virtqueue_read_indirect_desc(dev, desc_buf,
1880 desc_addr,
1881 desc_len)) {
1882 desc = desc_buf;
1883 }
1884 }
1885 if (!desc) {
1886 vu_panic(dev, "Invalid indirect buffer table");
1887 goto err;
1888 }
1889 num_bufs = i = 0;
1890 }
1891
1892 do {
1893 /* If we've got too many, that implies a descriptor loop. */
1894 if (++num_bufs > max) {
1895 vu_panic(dev, "Looped descriptor");
1896 goto err;
1897 }
1898
1899 if (desc[i].flags & VRING_DESC_F_WRITE) {
1900 in_total += desc[i].len;
1901 } else {
1902 out_total += desc[i].len;
1903 }
1904 if (in_total >= max_in_bytes && out_total >= max_out_bytes) {
1905 goto done;
1906 }
1907 rc = virtqueue_read_next_desc(dev, desc, i, max, &i);
1908 } while (rc == VIRTQUEUE_READ_DESC_MORE);
1909
1910 if (rc == VIRTQUEUE_READ_DESC_ERROR) {
1911 goto err;
1912 }
1913
1914 if (!indirect) {
1915 total_bufs = num_bufs;
1916 } else {
1917 total_bufs++;
1918 }
1919 }
1920 if (rc < 0) {
1921 goto err;
1922 }
1923done:
1924 if (in_bytes) {
1925 *in_bytes = in_total;
1926 }
1927 if (out_bytes) {
1928 *out_bytes = out_total;
1929 }
1930 return;
1931
1932err:
1933 in_total = out_total = 0;
1934 goto done;
1935}
1936
1937bool
1938vu_queue_avail_bytes(VuDev *dev, VuVirtq *vq, unsigned int in_bytes,
1939 unsigned int out_bytes)
1940{
1941 unsigned int in_total, out_total;
1942
1943 vu_queue_get_avail_bytes(dev, vq, &in_total, &out_total,
1944 in_bytes, out_bytes);
1945
1946 return in_bytes <= in_total && out_bytes <= out_total;
1947}
1948
1949/* Fetch avail_idx from VQ memory only when we really need to know if
1950 * guest has added some buffers. */
1951bool
1952vu_queue_empty(VuDev *dev, VuVirtq *vq)
1953{
1954 if (unlikely(dev->broken) ||
1955 unlikely(!vq->vring.avail)) {
1956 return true;
1957 }
1958
1959 if (vq->shadow_avail_idx != vq->last_avail_idx) {
1960 return false;
1961 }
1962
1963 return vring_avail_idx(vq) == vq->last_avail_idx;
1964}
1965
1966static bool
1967vring_notify(VuDev *dev, VuVirtq *vq)
1968{
1969 uint16_t old, new;
1970 bool v;
1971
1972 /* We need to expose used array entries before checking used event. */
1973 smp_mb();
1974
1975 /* Always notify when queue is empty (when feature acknowledge) */
1976 if (vu_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1977 !vq->inuse && vu_queue_empty(dev, vq)) {
1978 return true;
1979 }
1980
1981 if (!vu_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1982 return !(vring_avail_flags(vq) & VRING_AVAIL_F_NO_INTERRUPT);
1983 }
1984
1985 v = vq->signalled_used_valid;
1986 vq->signalled_used_valid = true;
1987 old = vq->signalled_used;
1988 new = vq->signalled_used = vq->used_idx;
1989 return !v || vring_need_event(vring_get_used_event(vq), new, old);
1990}
1991
1992void
1993vu_queue_notify(VuDev *dev, VuVirtq *vq)
1994{
1995 if (unlikely(dev->broken) ||
1996 unlikely(!vq->vring.avail)) {
1997 return;
1998 }
1999
2000 if (!vring_notify(dev, vq)) {
2001 DPRINT("skipped notify...\n");
2002 return;
2003 }
2004
2005 if (eventfd_write(vq->call_fd, 1) < 0) {
2006 vu_panic(dev, "Error writing eventfd: %s", strerror(errno));
2007 }
2008}
2009
2010static inline void
2011vring_used_flags_set_bit(VuVirtq *vq, int mask)
2012{
2013 uint16_t *flags;
2014
2015 flags = (uint16_t *)((char*)vq->vring.used +
2016 offsetof(struct vring_used, flags));
2017 *flags |= mask;
2018}
2019
2020static inline void
2021vring_used_flags_unset_bit(VuVirtq *vq, int mask)
2022{
2023 uint16_t *flags;
2024
2025 flags = (uint16_t *)((char*)vq->vring.used +
2026 offsetof(struct vring_used, flags));
2027 *flags &= ~mask;
2028}
2029
2030static inline void
2031vring_set_avail_event(VuVirtq *vq, uint16_t val)
2032{
2033 if (!vq->notification) {
2034 return;
2035 }
2036
2037 *((uint16_t *) &vq->vring.used->ring[vq->vring.num]) = val;
2038}
2039
2040void
2041vu_queue_set_notification(VuDev *dev, VuVirtq *vq, int enable)
2042{
2043 vq->notification = enable;
2044 if (vu_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
2045 vring_set_avail_event(vq, vring_avail_idx(vq));
2046 } else if (enable) {
2047 vring_used_flags_unset_bit(vq, VRING_USED_F_NO_NOTIFY);
2048 } else {
2049 vring_used_flags_set_bit(vq, VRING_USED_F_NO_NOTIFY);
2050 }
2051 if (enable) {
2052 /* Expose avail event/used flags before caller checks the avail idx. */
2053 smp_mb();
2054 }
2055}
2056
2057static void
2058virtqueue_map_desc(VuDev *dev,
2059 unsigned int *p_num_sg, struct iovec *iov,
2060 unsigned int max_num_sg, bool is_write,
2061 uint64_t pa, size_t sz)
2062{
2063 unsigned num_sg = *p_num_sg;
2064
2065 assert(num_sg <= max_num_sg);
2066
2067 if (!sz) {
2068 vu_panic(dev, "virtio: zero sized buffers are not allowed");
2069 return;
2070 }
2071
2072 while (sz) {
2073 uint64_t len = sz;
2074
2075 if (num_sg == max_num_sg) {
2076 vu_panic(dev, "virtio: too many descriptors in indirect table");
2077 return;
2078 }
2079
2080 iov[num_sg].iov_base = vu_gpa_to_va(dev, &len, pa);
2081 if (iov[num_sg].iov_base == NULL) {
2082 vu_panic(dev, "virtio: invalid address for buffers");
2083 return;
2084 }
2085 iov[num_sg].iov_len = len;
2086 num_sg++;
2087 sz -= len;
2088 pa += len;
2089 }
2090
2091 *p_num_sg = num_sg;
2092}
2093
2094static void *
2095virtqueue_alloc_element(size_t sz,
2096 unsigned out_num, unsigned in_num)
2097{
2098 VuVirtqElement *elem;
2099 size_t in_sg_ofs = ALIGN_UP(sz, __alignof__(elem->in_sg[0]));
2100 size_t out_sg_ofs = in_sg_ofs + in_num * sizeof(elem->in_sg[0]);
2101 size_t out_sg_end = out_sg_ofs + out_num * sizeof(elem->out_sg[0]);
2102
2103 assert(sz >= sizeof(VuVirtqElement));
2104 elem = malloc(out_sg_end);
2105 elem->out_num = out_num;
2106 elem->in_num = in_num;
2107 elem->in_sg = (void *)elem + in_sg_ofs;
2108 elem->out_sg = (void *)elem + out_sg_ofs;
2109 return elem;
2110}
2111
2112static void *
2113vu_queue_map_desc(VuDev *dev, VuVirtq *vq, unsigned int idx, size_t sz)
2114{
2115 struct vring_desc *desc = vq->vring.desc;
2116 uint64_t desc_addr, read_len;
2117 unsigned int desc_len;
2118 unsigned int max = vq->vring.num;
2119 unsigned int i = idx;
2120 VuVirtqElement *elem;
2121 unsigned int out_num = 0, in_num = 0;
2122 struct iovec iov[VIRTQUEUE_MAX_SIZE];
2123 struct vring_desc desc_buf[VIRTQUEUE_MAX_SIZE];
2124 int rc;
2125
2126 if (desc[i].flags & VRING_DESC_F_INDIRECT) {
2127 if (desc[i].len % sizeof(struct vring_desc)) {
2128 vu_panic(dev, "Invalid size for indirect buffer table");
2129 }
2130
2131 /* loop over the indirect descriptor table */
2132 desc_addr = desc[i].addr;
2133 desc_len = desc[i].len;
2134 max = desc_len / sizeof(struct vring_desc);
2135 read_len = desc_len;
2136 desc = vu_gpa_to_va(dev, &read_len, desc_addr);
2137 if (unlikely(desc && read_len != desc_len)) {
2138 /* Failed to use zero copy */
2139 desc = NULL;
2140 if (!virtqueue_read_indirect_desc(dev, desc_buf,
2141 desc_addr,
2142 desc_len)) {
2143 desc = desc_buf;
2144 }
2145 }
2146 if (!desc) {
2147 vu_panic(dev, "Invalid indirect buffer table");
2148 return NULL;
2149 }
2150 i = 0;
2151 }
2152
2153 /* Collect all the descriptors */
2154 do {
2155 if (desc[i].flags & VRING_DESC_F_WRITE) {
2156 virtqueue_map_desc(dev, &in_num, iov + out_num,
2157 VIRTQUEUE_MAX_SIZE - out_num, true,
2158 desc[i].addr, desc[i].len);
2159 } else {
2160 if (in_num) {
2161 vu_panic(dev, "Incorrect order for descriptors");
2162 return NULL;
2163 }
2164 virtqueue_map_desc(dev, &out_num, iov,
2165 VIRTQUEUE_MAX_SIZE, false,
2166 desc[i].addr, desc[i].len);
2167 }
2168
2169 /* If we've got too many, that implies a descriptor loop. */
2170 if ((in_num + out_num) > max) {
2171 vu_panic(dev, "Looped descriptor");
2172 }
2173 rc = virtqueue_read_next_desc(dev, desc, i, max, &i);
2174 } while (rc == VIRTQUEUE_READ_DESC_MORE);
2175
2176 if (rc == VIRTQUEUE_READ_DESC_ERROR) {
2177 vu_panic(dev, "read descriptor error");
2178 return NULL;
2179 }
2180
2181 /* Now copy what we have collected and mapped */
2182 elem = virtqueue_alloc_element(sz, out_num, in_num);
2183 elem->index = idx;
2184 for (i = 0; i < out_num; i++) {
2185 elem->out_sg[i] = iov[i];
2186 }
2187 for (i = 0; i < in_num; i++) {
2188 elem->in_sg[i] = iov[out_num + i];
2189 }
2190
2191 return elem;
2192}
2193
2194static int
2195vu_queue_inflight_get(VuDev *dev, VuVirtq *vq, int desc_idx)
2196{
2197 if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)) {
2198 return 0;
2199 }
2200
2201 if (unlikely(!vq->inflight)) {
2202 return -1;
2203 }
2204
2205 vq->inflight->desc[desc_idx].counter = vq->counter++;
2206 vq->inflight->desc[desc_idx].inflight = 1;
2207
2208 return 0;
2209}
2210
2211static int
2212vu_queue_inflight_pre_put(VuDev *dev, VuVirtq *vq, int desc_idx)
2213{
2214 if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)) {
2215 return 0;
2216 }
2217
2218 if (unlikely(!vq->inflight)) {
2219 return -1;
2220 }
2221
2222 vq->inflight->last_batch_head = desc_idx;
2223
2224 return 0;
2225}
2226
2227static int
2228vu_queue_inflight_post_put(VuDev *dev, VuVirtq *vq, int desc_idx)
2229{
2230 if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)) {
2231 return 0;
2232 }
2233
2234 if (unlikely(!vq->inflight)) {
2235 return -1;
2236 }
2237
2238 barrier();
2239
2240 vq->inflight->desc[desc_idx].inflight = 0;
2241
2242 barrier();
2243
2244 vq->inflight->used_idx = vq->used_idx;
2245
2246 return 0;
2247}
2248
2249void *
2250vu_queue_pop(VuDev *dev, VuVirtq *vq, size_t sz)
2251{
2252 int i;
2253 unsigned int head;
2254 VuVirtqElement *elem;
2255
2256 if (unlikely(dev->broken) ||
2257 unlikely(!vq->vring.avail)) {
2258 return NULL;
2259 }
2260
2261 if (unlikely(vq->resubmit_list && vq->resubmit_num > 0)) {
2262 i = (--vq->resubmit_num);
2263 elem = vu_queue_map_desc(dev, vq, vq->resubmit_list[i].index, sz);
2264
2265 if (!vq->resubmit_num) {
2266 free(vq->resubmit_list);
2267 vq->resubmit_list = NULL;
2268 }
2269
2270 return elem;
2271 }
2272
2273 if (vu_queue_empty(dev, vq)) {
2274 return NULL;
2275 }
2276 /*
2277 * Needed after virtio_queue_empty(), see comment in
2278 * virtqueue_num_heads().
2279 */
2280 smp_rmb();
2281
2282 if (vq->inuse >= vq->vring.num) {
2283 vu_panic(dev, "Virtqueue size exceeded");
2284 return NULL;
2285 }
2286
2287 if (!virtqueue_get_head(dev, vq, vq->last_avail_idx++, &head)) {
2288 return NULL;
2289 }
2290
2291 if (vu_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
2292 vring_set_avail_event(vq, vq->last_avail_idx);
2293 }
2294
2295 elem = vu_queue_map_desc(dev, vq, head, sz);
2296
2297 if (!elem) {
2298 return NULL;
2299 }
2300
2301 vq->inuse++;
2302
2303 vu_queue_inflight_get(dev, vq, head);
2304
2305 return elem;
2306}
2307
2308static void
2309vu_queue_detach_element(VuDev *dev, VuVirtq *vq, VuVirtqElement *elem,
2310 size_t len)
2311{
2312 vq->inuse--;
2313 /* unmap, when DMA support is added */
2314}
2315
2316void
2317vu_queue_unpop(VuDev *dev, VuVirtq *vq, VuVirtqElement *elem,
2318 size_t len)
2319{
2320 vq->last_avail_idx--;
2321 vu_queue_detach_element(dev, vq, elem, len);
2322}
2323
2324bool
2325vu_queue_rewind(VuDev *dev, VuVirtq *vq, unsigned int num)
2326{
2327 if (num > vq->inuse) {
2328 return false;
2329 }
2330 vq->last_avail_idx -= num;
2331 vq->inuse -= num;
2332 return true;
2333}
2334
2335static inline
2336void vring_used_write(VuDev *dev, VuVirtq *vq,
2337 struct vring_used_elem *uelem, int i)
2338{
2339 struct vring_used *used = vq->vring.used;
2340
2341 used->ring[i] = *uelem;
2342 vu_log_write(dev, vq->vring.log_guest_addr +
2343 offsetof(struct vring_used, ring[i]),
2344 sizeof(used->ring[i]));
2345}
2346
2347
2348static void
2349vu_log_queue_fill(VuDev *dev, VuVirtq *vq,
2350 const VuVirtqElement *elem,
2351 unsigned int len)
2352{
2353 struct vring_desc *desc = vq->vring.desc;
2354 unsigned int i, max, min, desc_len;
2355 uint64_t desc_addr, read_len;
2356 struct vring_desc desc_buf[VIRTQUEUE_MAX_SIZE];
2357 unsigned num_bufs = 0;
2358
2359 max = vq->vring.num;
2360 i = elem->index;
2361
2362 if (desc[i].flags & VRING_DESC_F_INDIRECT) {
2363 if (desc[i].len % sizeof(struct vring_desc)) {
2364 vu_panic(dev, "Invalid size for indirect buffer table");
2365 }
2366
2367 /* loop over the indirect descriptor table */
2368 desc_addr = desc[i].addr;
2369 desc_len = desc[i].len;
2370 max = desc_len / sizeof(struct vring_desc);
2371 read_len = desc_len;
2372 desc = vu_gpa_to_va(dev, &read_len, desc_addr);
2373 if (unlikely(desc && read_len != desc_len)) {
2374 /* Failed to use zero copy */
2375 desc = NULL;
2376 if (!virtqueue_read_indirect_desc(dev, desc_buf,
2377 desc_addr,
2378 desc_len)) {
2379 desc = desc_buf;
2380 }
2381 }
2382 if (!desc) {
2383 vu_panic(dev, "Invalid indirect buffer table");
2384 return;
2385 }
2386 i = 0;
2387 }
2388
2389 do {
2390 if (++num_bufs > max) {
2391 vu_panic(dev, "Looped descriptor");
2392 return;
2393 }
2394
2395 if (desc[i].flags & VRING_DESC_F_WRITE) {
2396 min = MIN(desc[i].len, len);
2397 vu_log_write(dev, desc[i].addr, min);
2398 len -= min;
2399 }
2400
2401 } while (len > 0 &&
2402 (virtqueue_read_next_desc(dev, desc, i, max, &i)
2403 == VIRTQUEUE_READ_DESC_MORE));
2404}
2405
2406void
2407vu_queue_fill(VuDev *dev, VuVirtq *vq,
2408 const VuVirtqElement *elem,
2409 unsigned int len, unsigned int idx)
2410{
2411 struct vring_used_elem uelem;
2412
2413 if (unlikely(dev->broken) ||
2414 unlikely(!vq->vring.avail)) {
2415 return;
2416 }
2417
2418 vu_log_queue_fill(dev, vq, elem, len);
2419
2420 idx = (idx + vq->used_idx) % vq->vring.num;
2421
2422 uelem.id = elem->index;
2423 uelem.len = len;
2424 vring_used_write(dev, vq, &uelem, idx);
2425}
2426
2427static inline
2428void vring_used_idx_set(VuDev *dev, VuVirtq *vq, uint16_t val)
2429{
2430 vq->vring.used->idx = val;
2431 vu_log_write(dev,
2432 vq->vring.log_guest_addr + offsetof(struct vring_used, idx),
2433 sizeof(vq->vring.used->idx));
2434
2435 vq->used_idx = val;
2436}
2437
2438void
2439vu_queue_flush(VuDev *dev, VuVirtq *vq, unsigned int count)
2440{
2441 uint16_t old, new;
2442
2443 if (unlikely(dev->broken) ||
2444 unlikely(!vq->vring.avail)) {
2445 return;
2446 }
2447
2448 /* Make sure buffer is written before we update index. */
2449 smp_wmb();
2450
2451 old = vq->used_idx;
2452 new = old + count;
2453 vring_used_idx_set(dev, vq, new);
2454 vq->inuse -= count;
2455 if (unlikely((int16_t)(new - vq->signalled_used) < (uint16_t)(new - old))) {
2456 vq->signalled_used_valid = false;
2457 }
2458}
2459
2460void
2461vu_queue_push(VuDev *dev, VuVirtq *vq,
2462 const VuVirtqElement *elem, unsigned int len)
2463{
2464 vu_queue_fill(dev, vq, elem, len, 0);
2465 vu_queue_inflight_pre_put(dev, vq, elem->index);
2466 vu_queue_flush(dev, vq, 1);
2467 vu_queue_inflight_post_put(dev, vq, elem->index);
2468}
2469