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 #ifndef _GNU_SOURCE
17 #define _GNU_SOURCE
18 #endif
19 
20 /* this code avoids GLib dependency */
21 #include <stdlib.h>
22 #include <stdio.h>
23 #include <unistd.h>
24 #include <stdarg.h>
25 #include <errno.h>
26 #include <string.h>
27 #include <assert.h>
28 #include <inttypes.h>
29 #include <sys/types.h>
30 #include <sys/socket.h>
31 #include <sys/eventfd.h>
32 #include <sys/mman.h>
33 #include <endian.h>
34 
35 /* Necessary to provide VIRTIO_F_VERSION_1 on system
36  * with older linux headers. Must appear before
37  * <linux/vhost.h> below.
38  */
39 #include "standard-headers/linux/virtio_config.h"
40 
41 #if defined(__linux__)
42 #include <sys/syscall.h>
43 #include <fcntl.h>
44 #include <sys/ioctl.h>
45 #include <linux/vhost.h>
46 #include <sys/vfs.h>
47 #include <linux/magic.h>
48 
49 #ifdef __NR_userfaultfd
50 #include <linux/userfaultfd.h>
51 #endif
52 
53 #endif
54 
55 #include "include/atomic.h"
56 
57 #include "libvhost-user.h"
58 
59 /* usually provided by GLib */
60 #if     __GNUC__ > 2 || (__GNUC__ == 2 && __GNUC_MINOR__ > 4)
61 #if !defined(__clang__) && (__GNUC__ == 4 && __GNUC_MINOR__ == 4)
62 #define G_GNUC_PRINTF(format_idx, arg_idx) \
63   __attribute__((__format__(gnu_printf, format_idx, arg_idx)))
64 #else
65 #define G_GNUC_PRINTF(format_idx, arg_idx) \
66   __attribute__((__format__(__printf__, format_idx, arg_idx)))
67 #endif
68 #else   /* !__GNUC__ */
69 #define G_GNUC_PRINTF(format_idx, arg_idx)
70 #endif  /* !__GNUC__ */
71 #ifndef MIN
72 #define MIN(x, y) ({                            \
73             __typeof__(x) _min1 = (x);          \
74             __typeof__(y) _min2 = (y);          \
75             (void) (&_min1 == &_min2);          \
76             _min1 < _min2 ? _min1 : _min2; })
77 #endif
78 
79 /* Round number down to multiple */
80 #define ALIGN_DOWN(n, m) ((n) / (m) * (m))
81 
82 /* Round number up to multiple */
83 #define ALIGN_UP(n, m) ALIGN_DOWN((n) + (m) - 1, (m))
84 
85 #ifndef unlikely
86 #define unlikely(x)   __builtin_expect(!!(x), 0)
87 #endif
88 
89 /* Align each region to cache line size in inflight buffer */
90 #define INFLIGHT_ALIGNMENT 64
91 
92 /* The version of inflight buffer */
93 #define INFLIGHT_VERSION 1
94 
95 /* The version of the protocol we support */
96 #define VHOST_USER_VERSION 1
97 #define LIBVHOST_USER_DEBUG 0
98 
99 #define DPRINT(...)                             \
100     do {                                        \
101         if (LIBVHOST_USER_DEBUG) {              \
102             fprintf(stderr, __VA_ARGS__);        \
103         }                                       \
104     } while (0)
105 
106 static inline
107 bool has_feature(uint64_t features, unsigned int fbit)
108 {
109     assert(fbit < 64);
110     return !!(features & (1ULL << fbit));
111 }
112 
113 static inline
114 bool vu_has_feature(VuDev *dev,
115                     unsigned int fbit)
116 {
117     return has_feature(dev->features, fbit);
118 }
119 
120 static inline bool vu_has_protocol_feature(VuDev *dev, unsigned int fbit)
121 {
122     return has_feature(dev->protocol_features, fbit);
123 }
124 
125 const char *
126 vu_request_to_string(unsigned int req)
127 {
128 #define REQ(req) [req] = #req
129     static const char *vu_request_str[] = {
130         REQ(VHOST_USER_NONE),
131         REQ(VHOST_USER_GET_FEATURES),
132         REQ(VHOST_USER_SET_FEATURES),
133         REQ(VHOST_USER_SET_OWNER),
134         REQ(VHOST_USER_RESET_OWNER),
135         REQ(VHOST_USER_SET_MEM_TABLE),
136         REQ(VHOST_USER_SET_LOG_BASE),
137         REQ(VHOST_USER_SET_LOG_FD),
138         REQ(VHOST_USER_SET_VRING_NUM),
139         REQ(VHOST_USER_SET_VRING_ADDR),
140         REQ(VHOST_USER_SET_VRING_BASE),
141         REQ(VHOST_USER_GET_VRING_BASE),
142         REQ(VHOST_USER_SET_VRING_KICK),
143         REQ(VHOST_USER_SET_VRING_CALL),
144         REQ(VHOST_USER_SET_VRING_ERR),
145         REQ(VHOST_USER_GET_PROTOCOL_FEATURES),
146         REQ(VHOST_USER_SET_PROTOCOL_FEATURES),
147         REQ(VHOST_USER_GET_QUEUE_NUM),
148         REQ(VHOST_USER_SET_VRING_ENABLE),
149         REQ(VHOST_USER_SEND_RARP),
150         REQ(VHOST_USER_NET_SET_MTU),
151         REQ(VHOST_USER_SET_BACKEND_REQ_FD),
152         REQ(VHOST_USER_IOTLB_MSG),
153         REQ(VHOST_USER_SET_VRING_ENDIAN),
154         REQ(VHOST_USER_GET_CONFIG),
155         REQ(VHOST_USER_SET_CONFIG),
156         REQ(VHOST_USER_POSTCOPY_ADVISE),
157         REQ(VHOST_USER_POSTCOPY_LISTEN),
158         REQ(VHOST_USER_POSTCOPY_END),
159         REQ(VHOST_USER_GET_INFLIGHT_FD),
160         REQ(VHOST_USER_SET_INFLIGHT_FD),
161         REQ(VHOST_USER_GPU_SET_SOCKET),
162         REQ(VHOST_USER_VRING_KICK),
163         REQ(VHOST_USER_GET_MAX_MEM_SLOTS),
164         REQ(VHOST_USER_ADD_MEM_REG),
165         REQ(VHOST_USER_REM_MEM_REG),
166         REQ(VHOST_USER_GET_SHARED_OBJECT),
167         REQ(VHOST_USER_MAX),
168     };
169 #undef REQ
170 
171     if (req < VHOST_USER_MAX) {
172         return vu_request_str[req];
173     } else {
174         return "unknown";
175     }
176 }
177 
178 static void G_GNUC_PRINTF(2, 3)
179 vu_panic(VuDev *dev, const char *msg, ...)
180 {
181     char *buf = NULL;
182     va_list ap;
183 
184     va_start(ap, msg);
185     if (vasprintf(&buf, msg, ap) < 0) {
186         buf = NULL;
187     }
188     va_end(ap);
189 
190     dev->broken = true;
191     dev->panic(dev, buf);
192     free(buf);
193 
194     /*
195      * FIXME:
196      * find a way to call virtio_error, or perhaps close the connection?
197      */
198 }
199 
200 /* Search for a memory region that covers this guest physical address. */
201 static VuDevRegion *
202 vu_gpa_to_mem_region(VuDev *dev, uint64_t guest_addr)
203 {
204     int low = 0;
205     int high = dev->nregions - 1;
206 
207     /*
208      * Memory regions cannot overlap in guest physical address space. Each
209      * GPA belongs to exactly one memory region, so there can only be one
210      * match.
211      *
212      * We store our memory regions ordered by GPA and can simply perform a
213      * binary search.
214      */
215     while (low <= high) {
216         unsigned int mid = low + (high - low) / 2;
217         VuDevRegion *cur = &dev->regions[mid];
218 
219         if (guest_addr >= cur->gpa && guest_addr < cur->gpa + cur->size) {
220             return cur;
221         }
222         if (guest_addr >= cur->gpa + cur->size) {
223             low = mid + 1;
224         }
225         if (guest_addr < cur->gpa) {
226             high = mid - 1;
227         }
228     }
229     return NULL;
230 }
231 
232 /* Translate guest physical address to our virtual address.  */
233 void *
234 vu_gpa_to_va(VuDev *dev, uint64_t *plen, uint64_t guest_addr)
235 {
236     VuDevRegion *r;
237 
238     if (*plen == 0) {
239         return NULL;
240     }
241 
242     r = vu_gpa_to_mem_region(dev, guest_addr);
243     if (!r) {
244         return NULL;
245     }
246 
247     if ((guest_addr + *plen) > (r->gpa + r->size)) {
248         *plen = r->gpa + r->size - guest_addr;
249     }
250     return (void *)(uintptr_t)guest_addr - r->gpa + r->mmap_addr +
251            r->mmap_offset;
252 }
253 
254 /* Translate qemu virtual address to our virtual address.  */
255 static void *
256 qva_to_va(VuDev *dev, uint64_t qemu_addr)
257 {
258     unsigned int i;
259 
260     /* Find matching memory region.  */
261     for (i = 0; i < dev->nregions; i++) {
262         VuDevRegion *r = &dev->regions[i];
263 
264         if ((qemu_addr >= r->qva) && (qemu_addr < (r->qva + r->size))) {
265             return (void *)(uintptr_t)
266                 qemu_addr - r->qva + r->mmap_addr + r->mmap_offset;
267         }
268     }
269 
270     return NULL;
271 }
272 
273 static void
274 vu_remove_all_mem_regs(VuDev *dev)
275 {
276     unsigned int i;
277 
278     for (i = 0; i < dev->nregions; i++) {
279         VuDevRegion *r = &dev->regions[i];
280 
281         munmap((void *)(uintptr_t)r->mmap_addr, r->size + r->mmap_offset);
282     }
283     dev->nregions = 0;
284 }
285 
286 static bool
287 map_ring(VuDev *dev, VuVirtq *vq)
288 {
289     vq->vring.desc = qva_to_va(dev, vq->vra.desc_user_addr);
290     vq->vring.used = qva_to_va(dev, vq->vra.used_user_addr);
291     vq->vring.avail = qva_to_va(dev, vq->vra.avail_user_addr);
292 
293     DPRINT("Setting virtq addresses:\n");
294     DPRINT("    vring_desc  at %p\n", vq->vring.desc);
295     DPRINT("    vring_used  at %p\n", vq->vring.used);
296     DPRINT("    vring_avail at %p\n", vq->vring.avail);
297 
298     return !(vq->vring.desc && vq->vring.used && vq->vring.avail);
299 }
300 
301 static bool
302 vu_is_vq_usable(VuDev *dev, VuVirtq *vq)
303 {
304     if (unlikely(dev->broken)) {
305         return false;
306     }
307 
308     if (likely(vq->vring.avail)) {
309         return true;
310     }
311 
312     /*
313      * In corner cases, we might temporarily remove a memory region that
314      * mapped a ring. When removing a memory region we make sure to
315      * unmap any rings that would be impacted. Let's try to remap if we
316      * already succeeded mapping this ring once.
317      */
318     if (!vq->vra.desc_user_addr || !vq->vra.used_user_addr ||
319         !vq->vra.avail_user_addr) {
320         return false;
321     }
322     if (map_ring(dev, vq)) {
323         vu_panic(dev, "remapping queue on access");
324         return false;
325     }
326     return true;
327 }
328 
329 static void
330 unmap_rings(VuDev *dev, VuDevRegion *r)
331 {
332     int i;
333 
334     for (i = 0; i < dev->max_queues; i++) {
335         VuVirtq *vq = &dev->vq[i];
336         const uintptr_t desc = (uintptr_t)vq->vring.desc;
337         const uintptr_t used = (uintptr_t)vq->vring.used;
338         const uintptr_t avail = (uintptr_t)vq->vring.avail;
339 
340         if (desc < r->mmap_addr || desc >= r->mmap_addr + r->size) {
341             continue;
342         }
343         if (used < r->mmap_addr || used >= r->mmap_addr + r->size) {
344             continue;
345         }
346         if (avail < r->mmap_addr || avail >= r->mmap_addr + r->size) {
347             continue;
348         }
349 
350         DPRINT("Unmapping rings of queue %d\n", i);
351         vq->vring.desc = NULL;
352         vq->vring.used = NULL;
353         vq->vring.avail = NULL;
354     }
355 }
356 
357 static size_t
358 get_fd_hugepagesize(int fd)
359 {
360 #if defined(__linux__)
361     struct statfs fs;
362     int ret;
363 
364     do {
365         ret = fstatfs(fd, &fs);
366     } while (ret != 0 && errno == EINTR);
367 
368     if (!ret && (unsigned int)fs.f_type == HUGETLBFS_MAGIC) {
369         return fs.f_bsize;
370     }
371 #endif
372     return 0;
373 }
374 
375 static void
376 _vu_add_mem_reg(VuDev *dev, VhostUserMemoryRegion *msg_region, int fd)
377 {
378     const uint64_t start_gpa = msg_region->guest_phys_addr;
379     const uint64_t end_gpa = start_gpa + msg_region->memory_size;
380     int prot = PROT_READ | PROT_WRITE;
381     uint64_t mmap_offset, fd_offset;
382     size_t hugepagesize;
383     VuDevRegion *r;
384     void *mmap_addr;
385     int low = 0;
386     int high = dev->nregions - 1;
387     unsigned int idx;
388 
389     DPRINT("Adding region %d\n", dev->nregions);
390     DPRINT("    guest_phys_addr: 0x%016"PRIx64"\n",
391            msg_region->guest_phys_addr);
392     DPRINT("    memory_size:     0x%016"PRIx64"\n",
393            msg_region->memory_size);
394     DPRINT("    userspace_addr:  0x%016"PRIx64"\n",
395            msg_region->userspace_addr);
396     DPRINT("    old mmap_offset: 0x%016"PRIx64"\n",
397            msg_region->mmap_offset);
398 
399     if (dev->postcopy_listening) {
400         /*
401          * In postcopy we're using PROT_NONE here to catch anyone
402          * accessing it before we userfault
403          */
404         prot = PROT_NONE;
405     }
406 
407     /*
408      * We will add memory regions into the array sorted by GPA. Perform a
409      * binary search to locate the insertion point: it will be at the low
410      * index.
411      */
412     while (low <= high) {
413         unsigned int mid = low + (high - low)  / 2;
414         VuDevRegion *cur = &dev->regions[mid];
415 
416         /* Overlap of GPA addresses. */
417         if (start_gpa < cur->gpa + cur->size && cur->gpa < end_gpa) {
418             vu_panic(dev, "regions with overlapping guest physical addresses");
419             return;
420         }
421         if (start_gpa >= cur->gpa + cur->size) {
422             low = mid + 1;
423         }
424         if (start_gpa < cur->gpa) {
425             high = mid - 1;
426         }
427     }
428     idx = low;
429 
430     /*
431      * Convert most of msg_region->mmap_offset to fd_offset. In almost all
432      * cases, this will leave us with mmap_offset == 0, mmap()'ing only
433      * what we really need. Only if a memory region would partially cover
434      * hugetlb pages, we'd get mmap_offset != 0, which usually doesn't happen
435      * anymore (i.e., modern QEMU).
436      *
437      * Note that mmap() with hugetlb would fail if the offset into the file
438      * is not aligned to the huge page size.
439      */
440     hugepagesize = get_fd_hugepagesize(fd);
441     if (hugepagesize) {
442         fd_offset = ALIGN_DOWN(msg_region->mmap_offset, hugepagesize);
443         mmap_offset = msg_region->mmap_offset - fd_offset;
444     } else {
445         fd_offset = msg_region->mmap_offset;
446         mmap_offset = 0;
447     }
448 
449     DPRINT("    fd_offset:       0x%016"PRIx64"\n",
450            fd_offset);
451     DPRINT("    new mmap_offset: 0x%016"PRIx64"\n",
452            mmap_offset);
453 
454     mmap_addr = mmap(0, msg_region->memory_size + mmap_offset,
455                      prot, MAP_SHARED | MAP_NORESERVE, fd, fd_offset);
456     if (mmap_addr == MAP_FAILED) {
457         vu_panic(dev, "region mmap error: %s", strerror(errno));
458         return;
459     }
460     DPRINT("    mmap_addr:       0x%016"PRIx64"\n",
461            (uint64_t)(uintptr_t)mmap_addr);
462 
463 #if defined(__linux__)
464     /* Don't include all guest memory in a coredump. */
465     madvise(mmap_addr, msg_region->memory_size + mmap_offset,
466             MADV_DONTDUMP);
467 #endif
468 
469     /* Shift all affected entries by 1 to open a hole at idx. */
470     r = &dev->regions[idx];
471     memmove(r + 1, r, sizeof(VuDevRegion) * (dev->nregions - idx));
472     r->gpa = msg_region->guest_phys_addr;
473     r->size = msg_region->memory_size;
474     r->qva = msg_region->userspace_addr;
475     r->mmap_addr = (uint64_t)(uintptr_t)mmap_addr;
476     r->mmap_offset = mmap_offset;
477     dev->nregions++;
478 
479     if (dev->postcopy_listening) {
480         /*
481          * Return the address to QEMU so that it can translate the ufd
482          * fault addresses back.
483          */
484         msg_region->userspace_addr = r->mmap_addr + r->mmap_offset;
485     }
486 }
487 
488 static void
489 vmsg_close_fds(VhostUserMsg *vmsg)
490 {
491     int i;
492 
493     for (i = 0; i < vmsg->fd_num; i++) {
494         close(vmsg->fds[i]);
495     }
496 }
497 
498 /* Set reply payload.u64 and clear request flags and fd_num */
499 static void vmsg_set_reply_u64(VhostUserMsg *vmsg, uint64_t val)
500 {
501     vmsg->flags = 0; /* defaults will be set by vu_send_reply() */
502     vmsg->size = sizeof(vmsg->payload.u64);
503     vmsg->payload.u64 = val;
504     vmsg->fd_num = 0;
505 }
506 
507 /* A test to see if we have userfault available */
508 static bool
509 have_userfault(void)
510 {
511 #if defined(__linux__) && defined(__NR_userfaultfd) &&\
512         defined(UFFD_FEATURE_MISSING_SHMEM) &&\
513         defined(UFFD_FEATURE_MISSING_HUGETLBFS)
514     /* Now test the kernel we're running on really has the features */
515     int ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
516     struct uffdio_api api_struct;
517     if (ufd < 0) {
518         return false;
519     }
520 
521     api_struct.api = UFFD_API;
522     api_struct.features = UFFD_FEATURE_MISSING_SHMEM |
523                           UFFD_FEATURE_MISSING_HUGETLBFS;
524     if (ioctl(ufd, UFFDIO_API, &api_struct)) {
525         close(ufd);
526         return false;
527     }
528     close(ufd);
529     return true;
530 
531 #else
532     return false;
533 #endif
534 }
535 
536 static bool
537 vu_message_read_default(VuDev *dev, int conn_fd, VhostUserMsg *vmsg)
538 {
539     char control[CMSG_SPACE(VHOST_MEMORY_BASELINE_NREGIONS * sizeof(int))] = {};
540     struct iovec iov = {
541         .iov_base = (char *)vmsg,
542         .iov_len = VHOST_USER_HDR_SIZE,
543     };
544     struct msghdr msg = {
545         .msg_iov = &iov,
546         .msg_iovlen = 1,
547         .msg_control = control,
548         .msg_controllen = sizeof(control),
549     };
550     size_t fd_size;
551     struct cmsghdr *cmsg;
552     int rc;
553 
554     do {
555         rc = recvmsg(conn_fd, &msg, 0);
556     } while (rc < 0 && (errno == EINTR || errno == EAGAIN));
557 
558     if (rc < 0) {
559         vu_panic(dev, "Error while recvmsg: %s", strerror(errno));
560         return false;
561     }
562 
563     vmsg->fd_num = 0;
564     for (cmsg = CMSG_FIRSTHDR(&msg);
565          cmsg != NULL;
566          cmsg = CMSG_NXTHDR(&msg, cmsg))
567     {
568         if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS) {
569             fd_size = cmsg->cmsg_len - CMSG_LEN(0);
570             vmsg->fd_num = fd_size / sizeof(int);
571             assert(vmsg->fd_num <= VHOST_MEMORY_BASELINE_NREGIONS);
572             memcpy(vmsg->fds, CMSG_DATA(cmsg), fd_size);
573             break;
574         }
575     }
576 
577     if (vmsg->size > sizeof(vmsg->payload)) {
578         vu_panic(dev,
579                  "Error: too big message request: %d, size: vmsg->size: %u, "
580                  "while sizeof(vmsg->payload) = %zu\n",
581                  vmsg->request, vmsg->size, sizeof(vmsg->payload));
582         goto fail;
583     }
584 
585     if (vmsg->size) {
586         do {
587             rc = read(conn_fd, &vmsg->payload, vmsg->size);
588         } while (rc < 0 && (errno == EINTR || errno == EAGAIN));
589 
590         if (rc <= 0) {
591             vu_panic(dev, "Error while reading: %s", strerror(errno));
592             goto fail;
593         }
594 
595         assert((uint32_t)rc == vmsg->size);
596     }
597 
598     return true;
599 
600 fail:
601     vmsg_close_fds(vmsg);
602 
603     return false;
604 }
605 
606 static bool
607 vu_message_write(VuDev *dev, int conn_fd, VhostUserMsg *vmsg)
608 {
609     int rc;
610     uint8_t *p = (uint8_t *)vmsg;
611     char control[CMSG_SPACE(VHOST_MEMORY_BASELINE_NREGIONS * sizeof(int))] = {};
612     struct iovec iov = {
613         .iov_base = (char *)vmsg,
614         .iov_len = VHOST_USER_HDR_SIZE,
615     };
616     struct msghdr msg = {
617         .msg_iov = &iov,
618         .msg_iovlen = 1,
619         .msg_control = control,
620     };
621     struct cmsghdr *cmsg;
622 
623     memset(control, 0, sizeof(control));
624     assert(vmsg->fd_num <= VHOST_MEMORY_BASELINE_NREGIONS);
625     if (vmsg->fd_num > 0) {
626         size_t fdsize = vmsg->fd_num * sizeof(int);
627         msg.msg_controllen = CMSG_SPACE(fdsize);
628         cmsg = CMSG_FIRSTHDR(&msg);
629         cmsg->cmsg_len = CMSG_LEN(fdsize);
630         cmsg->cmsg_level = SOL_SOCKET;
631         cmsg->cmsg_type = SCM_RIGHTS;
632         memcpy(CMSG_DATA(cmsg), vmsg->fds, fdsize);
633     } else {
634         msg.msg_controllen = 0;
635         msg.msg_control = NULL;
636     }
637 
638     do {
639         rc = sendmsg(conn_fd, &msg, 0);
640     } while (rc < 0 && (errno == EINTR || errno == EAGAIN));
641 
642     if (rc <= 0) {
643         vu_panic(dev, "Error while writing: %s", strerror(errno));
644         return false;
645     }
646 
647     if (vmsg->size) {
648         do {
649             if (vmsg->data) {
650                 rc = write(conn_fd, vmsg->data, vmsg->size);
651             } else {
652                 rc = write(conn_fd, p + VHOST_USER_HDR_SIZE, vmsg->size);
653             }
654         } while (rc < 0 && (errno == EINTR || errno == EAGAIN));
655     }
656 
657     if (rc <= 0) {
658         vu_panic(dev, "Error while writing: %s", strerror(errno));
659         return false;
660     }
661 
662     return true;
663 }
664 
665 static bool
666 vu_send_reply(VuDev *dev, int conn_fd, VhostUserMsg *vmsg)
667 {
668     /* Set the version in the flags when sending the reply */
669     vmsg->flags &= ~VHOST_USER_VERSION_MASK;
670     vmsg->flags |= VHOST_USER_VERSION;
671     vmsg->flags |= VHOST_USER_REPLY_MASK;
672 
673     return vu_message_write(dev, conn_fd, vmsg);
674 }
675 
676 /*
677  * Processes a reply on the backend channel.
678  * Entered with backend_mutex held and releases it before exit.
679  * Returns true on success.
680  */
681 static bool
682 vu_process_message_reply(VuDev *dev, const VhostUserMsg *vmsg)
683 {
684     VhostUserMsg msg_reply;
685     bool result = false;
686 
687     if ((vmsg->flags & VHOST_USER_NEED_REPLY_MASK) == 0) {
688         result = true;
689         goto out;
690     }
691 
692     if (!vu_message_read_default(dev, dev->backend_fd, &msg_reply)) {
693         goto out;
694     }
695 
696     if (msg_reply.request != vmsg->request) {
697         DPRINT("Received unexpected msg type. Expected %d received %d",
698                vmsg->request, msg_reply.request);
699         goto out;
700     }
701 
702     result = msg_reply.payload.u64 == 0;
703 
704 out:
705     pthread_mutex_unlock(&dev->backend_mutex);
706     return result;
707 }
708 
709 /* Kick the log_call_fd if required. */
710 static void
711 vu_log_kick(VuDev *dev)
712 {
713     if (dev->log_call_fd != -1) {
714         DPRINT("Kicking the QEMU's log...\n");
715         if (eventfd_write(dev->log_call_fd, 1) < 0) {
716             vu_panic(dev, "Error writing eventfd: %s", strerror(errno));
717         }
718     }
719 }
720 
721 static void
722 vu_log_page(uint8_t *log_table, uint64_t page)
723 {
724     DPRINT("Logged dirty guest page: %"PRId64"\n", page);
725     qatomic_or(&log_table[page / 8], 1 << (page % 8));
726 }
727 
728 static void
729 vu_log_write(VuDev *dev, uint64_t address, uint64_t length)
730 {
731     uint64_t page;
732 
733     if (!(dev->features & (1ULL << VHOST_F_LOG_ALL)) ||
734         !dev->log_table || !length) {
735         return;
736     }
737 
738     assert(dev->log_size > ((address + length - 1) / VHOST_LOG_PAGE / 8));
739 
740     page = address / VHOST_LOG_PAGE;
741     while (page * VHOST_LOG_PAGE < address + length) {
742         vu_log_page(dev->log_table, page);
743         page += 1;
744     }
745 
746     vu_log_kick(dev);
747 }
748 
749 static void
750 vu_kick_cb(VuDev *dev, int condition, void *data)
751 {
752     int index = (intptr_t)data;
753     VuVirtq *vq = &dev->vq[index];
754     int sock = vq->kick_fd;
755     eventfd_t kick_data;
756     ssize_t rc;
757 
758     rc = eventfd_read(sock, &kick_data);
759     if (rc == -1) {
760         vu_panic(dev, "kick eventfd_read(): %s", strerror(errno));
761         dev->remove_watch(dev, dev->vq[index].kick_fd);
762     } else {
763         DPRINT("Got kick_data: %016"PRIx64" handler:%p idx:%d\n",
764                kick_data, vq->handler, index);
765         if (vq->handler) {
766             vq->handler(dev, index);
767         }
768     }
769 }
770 
771 static bool
772 vu_get_features_exec(VuDev *dev, VhostUserMsg *vmsg)
773 {
774     vmsg->payload.u64 =
775         /*
776          * The following VIRTIO feature bits are supported by our virtqueue
777          * implementation:
778          */
779         1ULL << VIRTIO_F_NOTIFY_ON_EMPTY |
780         1ULL << VIRTIO_RING_F_INDIRECT_DESC |
781         1ULL << VIRTIO_RING_F_EVENT_IDX |
782         1ULL << VIRTIO_F_VERSION_1 |
783 
784         /* vhost-user feature bits */
785         1ULL << VHOST_F_LOG_ALL |
786         1ULL << VHOST_USER_F_PROTOCOL_FEATURES;
787 
788     if (dev->iface->get_features) {
789         vmsg->payload.u64 |= dev->iface->get_features(dev);
790     }
791 
792     vmsg->size = sizeof(vmsg->payload.u64);
793     vmsg->fd_num = 0;
794 
795     DPRINT("Sending back to guest u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
796 
797     return true;
798 }
799 
800 static void
801 vu_set_enable_all_rings(VuDev *dev, bool enabled)
802 {
803     uint16_t i;
804 
805     for (i = 0; i < dev->max_queues; i++) {
806         dev->vq[i].enable = enabled;
807     }
808 }
809 
810 static bool
811 vu_set_features_exec(VuDev *dev, VhostUserMsg *vmsg)
812 {
813     DPRINT("u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
814 
815     dev->features = vmsg->payload.u64;
816     if (!vu_has_feature(dev, VIRTIO_F_VERSION_1)) {
817         /*
818          * We only support devices conforming to VIRTIO 1.0 or
819          * later
820          */
821         vu_panic(dev, "virtio legacy devices aren't supported by libvhost-user");
822         return false;
823     }
824 
825     if (!(dev->features & VHOST_USER_F_PROTOCOL_FEATURES)) {
826         vu_set_enable_all_rings(dev, true);
827     }
828 
829     if (dev->iface->set_features) {
830         dev->iface->set_features(dev, dev->features);
831     }
832 
833     return false;
834 }
835 
836 static bool
837 vu_set_owner_exec(VuDev *dev, VhostUserMsg *vmsg)
838 {
839     return false;
840 }
841 
842 static void
843 vu_close_log(VuDev *dev)
844 {
845     if (dev->log_table) {
846         if (munmap(dev->log_table, dev->log_size) != 0) {
847             perror("close log munmap() error");
848         }
849 
850         dev->log_table = NULL;
851     }
852     if (dev->log_call_fd != -1) {
853         close(dev->log_call_fd);
854         dev->log_call_fd = -1;
855     }
856 }
857 
858 static bool
859 vu_reset_device_exec(VuDev *dev, VhostUserMsg *vmsg)
860 {
861     vu_set_enable_all_rings(dev, false);
862 
863     return false;
864 }
865 
866 static bool
867 generate_faults(VuDev *dev) {
868     unsigned int i;
869     for (i = 0; i < dev->nregions; i++) {
870 #ifdef UFFDIO_REGISTER
871         VuDevRegion *dev_region = &dev->regions[i];
872         int ret;
873         struct uffdio_register reg_struct;
874 
875         /*
876          * We should already have an open ufd. Mark each memory
877          * range as ufd.
878          * Discard any mapping we have here; note I can't use MADV_REMOVE
879          * or fallocate to make the hole since I don't want to lose
880          * data that's already arrived in the shared process.
881          * TODO: How to do hugepage
882          */
883         ret = madvise((void *)(uintptr_t)dev_region->mmap_addr,
884                       dev_region->size + dev_region->mmap_offset,
885                       MADV_DONTNEED);
886         if (ret) {
887             fprintf(stderr,
888                     "%s: Failed to madvise(DONTNEED) region %d: %s\n",
889                     __func__, i, strerror(errno));
890         }
891         /*
892          * Turn off transparent hugepages so we dont get lose wakeups
893          * in neighbouring pages.
894          * TODO: Turn this backon later.
895          */
896         ret = madvise((void *)(uintptr_t)dev_region->mmap_addr,
897                       dev_region->size + dev_region->mmap_offset,
898                       MADV_NOHUGEPAGE);
899         if (ret) {
900             /*
901              * Note: This can happen legally on kernels that are configured
902              * without madvise'able hugepages
903              */
904             fprintf(stderr,
905                     "%s: Failed to madvise(NOHUGEPAGE) region %d: %s\n",
906                     __func__, i, strerror(errno));
907         }
908 
909         reg_struct.range.start = (uintptr_t)dev_region->mmap_addr;
910         reg_struct.range.len = dev_region->size + dev_region->mmap_offset;
911         reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
912 
913         if (ioctl(dev->postcopy_ufd, UFFDIO_REGISTER, &reg_struct)) {
914             vu_panic(dev, "%s: Failed to userfault region %d "
915                           "@%" PRIx64 " + size:%" PRIx64 " offset: %" PRIx64
916                           ": (ufd=%d)%s\n",
917                      __func__, i,
918                      dev_region->mmap_addr,
919                      dev_region->size, dev_region->mmap_offset,
920                      dev->postcopy_ufd, strerror(errno));
921             return false;
922         }
923         if (!(reg_struct.ioctls & (1ULL << _UFFDIO_COPY))) {
924             vu_panic(dev, "%s Region (%d) doesn't support COPY",
925                      __func__, i);
926             return false;
927         }
928         DPRINT("%s: region %d: Registered userfault for %"
929                PRIx64 " + %" PRIx64 "\n", __func__, i,
930                (uint64_t)reg_struct.range.start,
931                (uint64_t)reg_struct.range.len);
932         /* Now it's registered we can let the client at it */
933         if (mprotect((void *)(uintptr_t)dev_region->mmap_addr,
934                      dev_region->size + dev_region->mmap_offset,
935                      PROT_READ | PROT_WRITE)) {
936             vu_panic(dev, "failed to mprotect region %d for postcopy (%s)",
937                      i, strerror(errno));
938             return false;
939         }
940         /* TODO: Stash 'zero' support flags somewhere */
941 #endif
942     }
943 
944     return true;
945 }
946 
947 static bool
948 vu_add_mem_reg(VuDev *dev, VhostUserMsg *vmsg) {
949     VhostUserMemoryRegion m = vmsg->payload.memreg.region, *msg_region = &m;
950 
951     if (vmsg->fd_num != 1) {
952         vmsg_close_fds(vmsg);
953         vu_panic(dev, "VHOST_USER_ADD_MEM_REG received %d fds - only 1 fd "
954                       "should be sent for this message type", vmsg->fd_num);
955         return false;
956     }
957 
958     if (vmsg->size < VHOST_USER_MEM_REG_SIZE) {
959         close(vmsg->fds[0]);
960         vu_panic(dev, "VHOST_USER_ADD_MEM_REG requires a message size of at "
961                       "least %zu bytes and only %d bytes were received",
962                       VHOST_USER_MEM_REG_SIZE, vmsg->size);
963         return false;
964     }
965 
966     if (dev->nregions == VHOST_USER_MAX_RAM_SLOTS) {
967         close(vmsg->fds[0]);
968         vu_panic(dev, "failing attempt to hot add memory via "
969                       "VHOST_USER_ADD_MEM_REG message because the backend has "
970                       "no free ram slots available");
971         return false;
972     }
973 
974     /*
975      * If we are in postcopy mode and we receive a u64 payload with a 0 value
976      * we know all the postcopy client bases have been received, and we
977      * should start generating faults.
978      */
979     if (dev->postcopy_listening &&
980         vmsg->size == sizeof(vmsg->payload.u64) &&
981         vmsg->payload.u64 == 0) {
982         (void)generate_faults(dev);
983         return false;
984     }
985 
986     _vu_add_mem_reg(dev, msg_region, vmsg->fds[0]);
987     close(vmsg->fds[0]);
988 
989     if (dev->postcopy_listening) {
990         /* Send the message back to qemu with the addresses filled in. */
991         vmsg->fd_num = 0;
992         DPRINT("Successfully added new region in postcopy\n");
993         return true;
994     }
995     DPRINT("Successfully added new region\n");
996     return false;
997 }
998 
999 static inline bool reg_equal(VuDevRegion *vudev_reg,
1000                              VhostUserMemoryRegion *msg_reg)
1001 {
1002     if (vudev_reg->gpa == msg_reg->guest_phys_addr &&
1003         vudev_reg->qva == msg_reg->userspace_addr &&
1004         vudev_reg->size == msg_reg->memory_size) {
1005         return true;
1006     }
1007 
1008     return false;
1009 }
1010 
1011 static bool
1012 vu_rem_mem_reg(VuDev *dev, VhostUserMsg *vmsg) {
1013     VhostUserMemoryRegion m = vmsg->payload.memreg.region, *msg_region = &m;
1014     unsigned int idx;
1015     VuDevRegion *r;
1016 
1017     if (vmsg->fd_num > 1) {
1018         vmsg_close_fds(vmsg);
1019         vu_panic(dev, "VHOST_USER_REM_MEM_REG received %d fds - at most 1 fd "
1020                       "should be sent for this message type", vmsg->fd_num);
1021         return false;
1022     }
1023 
1024     if (vmsg->size < VHOST_USER_MEM_REG_SIZE) {
1025         vmsg_close_fds(vmsg);
1026         vu_panic(dev, "VHOST_USER_REM_MEM_REG requires a message size of at "
1027                       "least %zu bytes and only %d bytes were received",
1028                       VHOST_USER_MEM_REG_SIZE, vmsg->size);
1029         return false;
1030     }
1031 
1032     DPRINT("Removing region:\n");
1033     DPRINT("    guest_phys_addr: 0x%016"PRIx64"\n",
1034            msg_region->guest_phys_addr);
1035     DPRINT("    memory_size:     0x%016"PRIx64"\n",
1036            msg_region->memory_size);
1037     DPRINT("    userspace_addr   0x%016"PRIx64"\n",
1038            msg_region->userspace_addr);
1039     DPRINT("    mmap_offset      0x%016"PRIx64"\n",
1040            msg_region->mmap_offset);
1041 
1042     r = vu_gpa_to_mem_region(dev, msg_region->guest_phys_addr);
1043     if (!r || !reg_equal(r, msg_region)) {
1044         vmsg_close_fds(vmsg);
1045         vu_panic(dev, "Specified region not found\n");
1046         return false;
1047     }
1048 
1049     /*
1050      * There might be valid cases where we temporarily remove memory regions
1051      * to readd them again, or remove memory regions and don't use the rings
1052      * anymore before we set the ring addresses and restart the device.
1053      *
1054      * Unmap all affected rings, remapping them on demand later. This should
1055      * be a corner case.
1056      */
1057     unmap_rings(dev, r);
1058 
1059     munmap((void *)(uintptr_t)r->mmap_addr, r->size + r->mmap_offset);
1060 
1061     idx = r - dev->regions;
1062     assert(idx < dev->nregions);
1063     /* Shift all affected entries by 1 to close the hole. */
1064     memmove(r, r + 1, sizeof(VuDevRegion) * (dev->nregions - idx - 1));
1065     DPRINT("Successfully removed a region\n");
1066     dev->nregions--;
1067 
1068     vmsg_close_fds(vmsg);
1069 
1070     return false;
1071 }
1072 
1073 static bool
1074 vu_get_shared_object(VuDev *dev, VhostUserMsg *vmsg)
1075 {
1076     int fd_num = 0;
1077     int dmabuf_fd = -1;
1078     if (dev->iface->get_shared_object) {
1079         dmabuf_fd = dev->iface->get_shared_object(
1080             dev, &vmsg->payload.object.uuid[0]);
1081     }
1082     if (dmabuf_fd != -1) {
1083         DPRINT("dmabuf_fd found for requested UUID\n");
1084         vmsg->fds[fd_num++] = dmabuf_fd;
1085     }
1086     vmsg->fd_num = fd_num;
1087 
1088     return true;
1089 }
1090 
1091 static bool
1092 vu_set_mem_table_exec(VuDev *dev, VhostUserMsg *vmsg)
1093 {
1094     VhostUserMemory m = vmsg->payload.memory, *memory = &m;
1095     unsigned int i;
1096 
1097     vu_remove_all_mem_regs(dev);
1098 
1099     DPRINT("Nregions: %u\n", memory->nregions);
1100     for (i = 0; i < memory->nregions; i++) {
1101         _vu_add_mem_reg(dev, &memory->regions[i], vmsg->fds[i]);
1102         close(vmsg->fds[i]);
1103     }
1104 
1105     if (dev->postcopy_listening) {
1106         /* Send the message back to qemu with the addresses filled in */
1107         vmsg->fd_num = 0;
1108         if (!vu_send_reply(dev, dev->sock, vmsg)) {
1109             vu_panic(dev, "failed to respond to set-mem-table for postcopy");
1110             return false;
1111         }
1112 
1113         /*
1114          * Wait for QEMU to confirm that it's registered the handler for the
1115          * faults.
1116          */
1117         if (!dev->read_msg(dev, dev->sock, vmsg) ||
1118             vmsg->size != sizeof(vmsg->payload.u64) ||
1119             vmsg->payload.u64 != 0) {
1120             vu_panic(dev, "failed to receive valid ack for postcopy set-mem-table");
1121             return false;
1122         }
1123 
1124         /* OK, now we can go and register the memory and generate faults */
1125         (void)generate_faults(dev);
1126         return false;
1127     }
1128 
1129     for (i = 0; i < dev->max_queues; i++) {
1130         if (dev->vq[i].vring.desc) {
1131             if (map_ring(dev, &dev->vq[i])) {
1132                 vu_panic(dev, "remapping queue %d during setmemtable", i);
1133             }
1134         }
1135     }
1136 
1137     return false;
1138 }
1139 
1140 static bool
1141 vu_set_log_base_exec(VuDev *dev, VhostUserMsg *vmsg)
1142 {
1143     int fd;
1144     uint64_t log_mmap_size, log_mmap_offset;
1145     void *rc;
1146 
1147     if (vmsg->fd_num != 1 ||
1148         vmsg->size != sizeof(vmsg->payload.log)) {
1149         vu_panic(dev, "Invalid log_base message");
1150         return true;
1151     }
1152 
1153     fd = vmsg->fds[0];
1154     log_mmap_offset = vmsg->payload.log.mmap_offset;
1155     log_mmap_size = vmsg->payload.log.mmap_size;
1156     DPRINT("Log mmap_offset: %"PRId64"\n", log_mmap_offset);
1157     DPRINT("Log mmap_size:   %"PRId64"\n", log_mmap_size);
1158 
1159     rc = mmap(0, log_mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd,
1160               log_mmap_offset);
1161     close(fd);
1162     if (rc == MAP_FAILED) {
1163         perror("log mmap error");
1164     }
1165 
1166     if (dev->log_table) {
1167         munmap(dev->log_table, dev->log_size);
1168     }
1169     dev->log_table = rc;
1170     dev->log_size = log_mmap_size;
1171 
1172     vmsg->size = sizeof(vmsg->payload.u64);
1173     vmsg->fd_num = 0;
1174 
1175     return true;
1176 }
1177 
1178 static bool
1179 vu_set_log_fd_exec(VuDev *dev, VhostUserMsg *vmsg)
1180 {
1181     if (vmsg->fd_num != 1) {
1182         vu_panic(dev, "Invalid log_fd message");
1183         return false;
1184     }
1185 
1186     if (dev->log_call_fd != -1) {
1187         close(dev->log_call_fd);
1188     }
1189     dev->log_call_fd = vmsg->fds[0];
1190     DPRINT("Got log_call_fd: %d\n", vmsg->fds[0]);
1191 
1192     return false;
1193 }
1194 
1195 static bool
1196 vu_set_vring_num_exec(VuDev *dev, VhostUserMsg *vmsg)
1197 {
1198     unsigned int index = vmsg->payload.state.index;
1199     unsigned int num = vmsg->payload.state.num;
1200 
1201     DPRINT("State.index: %u\n", index);
1202     DPRINT("State.num:   %u\n", num);
1203     dev->vq[index].vring.num = num;
1204 
1205     return false;
1206 }
1207 
1208 static bool
1209 vu_set_vring_addr_exec(VuDev *dev, VhostUserMsg *vmsg)
1210 {
1211     struct vhost_vring_addr addr = vmsg->payload.addr, *vra = &addr;
1212     unsigned int index = vra->index;
1213     VuVirtq *vq = &dev->vq[index];
1214 
1215     DPRINT("vhost_vring_addr:\n");
1216     DPRINT("    index:  %d\n", vra->index);
1217     DPRINT("    flags:  %d\n", vra->flags);
1218     DPRINT("    desc_user_addr:   0x%016" PRIx64 "\n", (uint64_t)vra->desc_user_addr);
1219     DPRINT("    used_user_addr:   0x%016" PRIx64 "\n", (uint64_t)vra->used_user_addr);
1220     DPRINT("    avail_user_addr:  0x%016" PRIx64 "\n", (uint64_t)vra->avail_user_addr);
1221     DPRINT("    log_guest_addr:   0x%016" PRIx64 "\n", (uint64_t)vra->log_guest_addr);
1222 
1223     vq->vra = *vra;
1224     vq->vring.flags = vra->flags;
1225     vq->vring.log_guest_addr = vra->log_guest_addr;
1226 
1227 
1228     if (map_ring(dev, vq)) {
1229         vu_panic(dev, "Invalid vring_addr message");
1230         return false;
1231     }
1232 
1233     vq->used_idx = le16toh(vq->vring.used->idx);
1234 
1235     if (vq->last_avail_idx != vq->used_idx) {
1236         bool resume = dev->iface->queue_is_processed_in_order &&
1237             dev->iface->queue_is_processed_in_order(dev, index);
1238 
1239         DPRINT("Last avail index != used index: %u != %u%s\n",
1240                vq->last_avail_idx, vq->used_idx,
1241                resume ? ", resuming" : "");
1242 
1243         if (resume) {
1244             vq->shadow_avail_idx = vq->last_avail_idx = vq->used_idx;
1245         }
1246     }
1247 
1248     return false;
1249 }
1250 
1251 static bool
1252 vu_set_vring_base_exec(VuDev *dev, VhostUserMsg *vmsg)
1253 {
1254     unsigned int index = vmsg->payload.state.index;
1255     unsigned int num = vmsg->payload.state.num;
1256 
1257     DPRINT("State.index: %u\n", index);
1258     DPRINT("State.num:   %u\n", num);
1259     dev->vq[index].shadow_avail_idx = dev->vq[index].last_avail_idx = num;
1260 
1261     return false;
1262 }
1263 
1264 static bool
1265 vu_get_vring_base_exec(VuDev *dev, VhostUserMsg *vmsg)
1266 {
1267     unsigned int index = vmsg->payload.state.index;
1268 
1269     DPRINT("State.index: %u\n", index);
1270     vmsg->payload.state.num = dev->vq[index].last_avail_idx;
1271     vmsg->size = sizeof(vmsg->payload.state);
1272 
1273     dev->vq[index].started = false;
1274     if (dev->iface->queue_set_started) {
1275         dev->iface->queue_set_started(dev, index, false);
1276     }
1277 
1278     if (dev->vq[index].call_fd != -1) {
1279         close(dev->vq[index].call_fd);
1280         dev->vq[index].call_fd = -1;
1281     }
1282     if (dev->vq[index].kick_fd != -1) {
1283         dev->remove_watch(dev, dev->vq[index].kick_fd);
1284         close(dev->vq[index].kick_fd);
1285         dev->vq[index].kick_fd = -1;
1286     }
1287 
1288     return true;
1289 }
1290 
1291 static bool
1292 vu_check_queue_msg_file(VuDev *dev, VhostUserMsg *vmsg)
1293 {
1294     int index = vmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1295     bool nofd = vmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK;
1296 
1297     if (index >= dev->max_queues) {
1298         vmsg_close_fds(vmsg);
1299         vu_panic(dev, "Invalid queue index: %u", index);
1300         return false;
1301     }
1302 
1303     if (nofd) {
1304         vmsg_close_fds(vmsg);
1305         return true;
1306     }
1307 
1308     if (vmsg->fd_num != 1) {
1309         vmsg_close_fds(vmsg);
1310         vu_panic(dev, "Invalid fds in request: %d", vmsg->request);
1311         return false;
1312     }
1313 
1314     return true;
1315 }
1316 
1317 static int
1318 inflight_desc_compare(const void *a, const void *b)
1319 {
1320     VuVirtqInflightDesc *desc0 = (VuVirtqInflightDesc *)a,
1321                         *desc1 = (VuVirtqInflightDesc *)b;
1322 
1323     if (desc1->counter > desc0->counter &&
1324         (desc1->counter - desc0->counter) < VIRTQUEUE_MAX_SIZE * 2) {
1325         return 1;
1326     }
1327 
1328     return -1;
1329 }
1330 
1331 static int
1332 vu_check_queue_inflights(VuDev *dev, VuVirtq *vq)
1333 {
1334     int i = 0;
1335 
1336     if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)) {
1337         return 0;
1338     }
1339 
1340     if (unlikely(!vq->inflight)) {
1341         return -1;
1342     }
1343 
1344     if (unlikely(!vq->inflight->version)) {
1345         /* initialize the buffer */
1346         vq->inflight->version = INFLIGHT_VERSION;
1347         return 0;
1348     }
1349 
1350     vq->used_idx = le16toh(vq->vring.used->idx);
1351     vq->resubmit_num = 0;
1352     vq->resubmit_list = NULL;
1353     vq->counter = 0;
1354 
1355     if (unlikely(vq->inflight->used_idx != vq->used_idx)) {
1356         vq->inflight->desc[vq->inflight->last_batch_head].inflight = 0;
1357 
1358         barrier();
1359 
1360         vq->inflight->used_idx = vq->used_idx;
1361     }
1362 
1363     for (i = 0; i < vq->inflight->desc_num; i++) {
1364         if (vq->inflight->desc[i].inflight == 1) {
1365             vq->inuse++;
1366         }
1367     }
1368 
1369     vq->shadow_avail_idx = vq->last_avail_idx = vq->inuse + vq->used_idx;
1370 
1371     if (vq->inuse) {
1372         vq->resubmit_list = calloc(vq->inuse, sizeof(VuVirtqInflightDesc));
1373         if (!vq->resubmit_list) {
1374             return -1;
1375         }
1376 
1377         for (i = 0; i < vq->inflight->desc_num; i++) {
1378             if (vq->inflight->desc[i].inflight) {
1379                 vq->resubmit_list[vq->resubmit_num].index = i;
1380                 vq->resubmit_list[vq->resubmit_num].counter =
1381                                         vq->inflight->desc[i].counter;
1382                 vq->resubmit_num++;
1383             }
1384         }
1385 
1386         if (vq->resubmit_num > 1) {
1387             qsort(vq->resubmit_list, vq->resubmit_num,
1388                   sizeof(VuVirtqInflightDesc), inflight_desc_compare);
1389         }
1390         vq->counter = vq->resubmit_list[0].counter + 1;
1391     }
1392 
1393     /* in case of I/O hang after reconnecting */
1394     if (eventfd_write(vq->kick_fd, 1)) {
1395         return -1;
1396     }
1397 
1398     return 0;
1399 }
1400 
1401 static bool
1402 vu_set_vring_kick_exec(VuDev *dev, VhostUserMsg *vmsg)
1403 {
1404     int index = vmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1405     bool nofd = vmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK;
1406 
1407     DPRINT("u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
1408 
1409     if (!vu_check_queue_msg_file(dev, vmsg)) {
1410         return false;
1411     }
1412 
1413     if (dev->vq[index].kick_fd != -1) {
1414         dev->remove_watch(dev, dev->vq[index].kick_fd);
1415         close(dev->vq[index].kick_fd);
1416         dev->vq[index].kick_fd = -1;
1417     }
1418 
1419     dev->vq[index].kick_fd = nofd ? -1 : vmsg->fds[0];
1420     DPRINT("Got kick_fd: %d for vq: %d\n", dev->vq[index].kick_fd, index);
1421 
1422     dev->vq[index].started = true;
1423     if (dev->iface->queue_set_started) {
1424         dev->iface->queue_set_started(dev, index, true);
1425     }
1426 
1427     if (dev->vq[index].kick_fd != -1 && dev->vq[index].handler) {
1428         dev->set_watch(dev, dev->vq[index].kick_fd, VU_WATCH_IN,
1429                        vu_kick_cb, (void *)(long)index);
1430 
1431         DPRINT("Waiting for kicks on fd: %d for vq: %d\n",
1432                dev->vq[index].kick_fd, index);
1433     }
1434 
1435     if (vu_check_queue_inflights(dev, &dev->vq[index])) {
1436         vu_panic(dev, "Failed to check inflights for vq: %d\n", index);
1437     }
1438 
1439     return false;
1440 }
1441 
1442 void vu_set_queue_handler(VuDev *dev, VuVirtq *vq,
1443                           vu_queue_handler_cb handler)
1444 {
1445     int qidx = vq - dev->vq;
1446 
1447     vq->handler = handler;
1448     if (vq->kick_fd >= 0) {
1449         if (handler) {
1450             dev->set_watch(dev, vq->kick_fd, VU_WATCH_IN,
1451                            vu_kick_cb, (void *)(long)qidx);
1452         } else {
1453             dev->remove_watch(dev, vq->kick_fd);
1454         }
1455     }
1456 }
1457 
1458 bool vu_set_queue_host_notifier(VuDev *dev, VuVirtq *vq, int fd,
1459                                 int size, int offset)
1460 {
1461     int qidx = vq - dev->vq;
1462     int fd_num = 0;
1463     VhostUserMsg vmsg = {
1464         .request = VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG,
1465         .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY_MASK,
1466         .size = sizeof(vmsg.payload.area),
1467         .payload.area = {
1468             .u64 = qidx & VHOST_USER_VRING_IDX_MASK,
1469             .size = size,
1470             .offset = offset,
1471         },
1472     };
1473 
1474     if (fd == -1) {
1475         vmsg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
1476     } else {
1477         vmsg.fds[fd_num++] = fd;
1478     }
1479 
1480     vmsg.fd_num = fd_num;
1481 
1482     if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_BACKEND_SEND_FD)) {
1483         return false;
1484     }
1485 
1486     pthread_mutex_lock(&dev->backend_mutex);
1487     if (!vu_message_write(dev, dev->backend_fd, &vmsg)) {
1488         pthread_mutex_unlock(&dev->backend_mutex);
1489         return false;
1490     }
1491 
1492     /* Also unlocks the backend_mutex */
1493     return vu_process_message_reply(dev, &vmsg);
1494 }
1495 
1496 bool
1497 vu_lookup_shared_object(VuDev *dev, unsigned char uuid[UUID_LEN],
1498                         int *dmabuf_fd)
1499 {
1500     bool result = false;
1501     VhostUserMsg msg_reply;
1502     VhostUserMsg msg = {
1503         .request = VHOST_USER_BACKEND_SHARED_OBJECT_LOOKUP,
1504         .size = sizeof(msg.payload.object),
1505         .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY_MASK,
1506     };
1507 
1508     memcpy(msg.payload.object.uuid, uuid, sizeof(uuid[0]) * UUID_LEN);
1509 
1510     if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_SHARED_OBJECT)) {
1511         return false;
1512     }
1513 
1514     pthread_mutex_lock(&dev->backend_mutex);
1515     if (!vu_message_write(dev, dev->backend_fd, &msg)) {
1516         goto out;
1517     }
1518 
1519     if (!vu_message_read_default(dev, dev->backend_fd, &msg_reply)) {
1520         goto out;
1521     }
1522 
1523     if (msg_reply.request != msg.request) {
1524         DPRINT("Received unexpected msg type. Expected %d, received %d",
1525                msg.request, msg_reply.request);
1526         goto out;
1527     }
1528 
1529     if (msg_reply.fd_num != 1) {
1530         DPRINT("Received unexpected number of fds. Expected 1, received %d",
1531                msg_reply.fd_num);
1532         goto out;
1533     }
1534 
1535     *dmabuf_fd = msg_reply.fds[0];
1536     result = *dmabuf_fd > 0 && msg_reply.payload.u64 == 0;
1537 out:
1538     pthread_mutex_unlock(&dev->backend_mutex);
1539 
1540     return result;
1541 }
1542 
1543 static bool
1544 vu_send_message(VuDev *dev, VhostUserMsg *vmsg)
1545 {
1546     bool result = false;
1547     pthread_mutex_lock(&dev->backend_mutex);
1548     if (!vu_message_write(dev, dev->backend_fd, vmsg)) {
1549         goto out;
1550     }
1551 
1552     result = true;
1553 out:
1554     pthread_mutex_unlock(&dev->backend_mutex);
1555 
1556     return result;
1557 }
1558 
1559 bool
1560 vu_add_shared_object(VuDev *dev, unsigned char uuid[UUID_LEN])
1561 {
1562     VhostUserMsg msg = {
1563         .request = VHOST_USER_BACKEND_SHARED_OBJECT_ADD,
1564         .size = sizeof(msg.payload.object),
1565         .flags = VHOST_USER_VERSION,
1566     };
1567 
1568     memcpy(msg.payload.object.uuid, uuid, sizeof(uuid[0]) * UUID_LEN);
1569 
1570     if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_SHARED_OBJECT)) {
1571         return false;
1572     }
1573 
1574     return vu_send_message(dev, &msg);
1575 }
1576 
1577 bool
1578 vu_rm_shared_object(VuDev *dev, unsigned char uuid[UUID_LEN])
1579 {
1580     VhostUserMsg msg = {
1581         .request = VHOST_USER_BACKEND_SHARED_OBJECT_REMOVE,
1582         .size = sizeof(msg.payload.object),
1583         .flags = VHOST_USER_VERSION,
1584     };
1585 
1586     memcpy(msg.payload.object.uuid, uuid, sizeof(uuid[0]) * UUID_LEN);
1587 
1588     if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_SHARED_OBJECT)) {
1589         return false;
1590     }
1591 
1592     return vu_send_message(dev, &msg);
1593 }
1594 
1595 static bool
1596 vu_set_vring_call_exec(VuDev *dev, VhostUserMsg *vmsg)
1597 {
1598     int index = vmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1599     bool nofd = vmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK;
1600 
1601     DPRINT("u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
1602 
1603     if (!vu_check_queue_msg_file(dev, vmsg)) {
1604         return false;
1605     }
1606 
1607     if (dev->vq[index].call_fd != -1) {
1608         close(dev->vq[index].call_fd);
1609         dev->vq[index].call_fd = -1;
1610     }
1611 
1612     dev->vq[index].call_fd = nofd ? -1 : vmsg->fds[0];
1613 
1614     /* in case of I/O hang after reconnecting */
1615     if (dev->vq[index].call_fd != -1 && eventfd_write(vmsg->fds[0], 1)) {
1616         return -1;
1617     }
1618 
1619     DPRINT("Got call_fd: %d for vq: %d\n", dev->vq[index].call_fd, index);
1620 
1621     return false;
1622 }
1623 
1624 static bool
1625 vu_set_vring_err_exec(VuDev *dev, VhostUserMsg *vmsg)
1626 {
1627     int index = vmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1628     bool nofd = vmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK;
1629 
1630     DPRINT("u64: 0x%016"PRIx64"\n", vmsg->payload.u64);
1631 
1632     if (!vu_check_queue_msg_file(dev, vmsg)) {
1633         return false;
1634     }
1635 
1636     if (dev->vq[index].err_fd != -1) {
1637         close(dev->vq[index].err_fd);
1638         dev->vq[index].err_fd = -1;
1639     }
1640 
1641     dev->vq[index].err_fd = nofd ? -1 : vmsg->fds[0];
1642 
1643     return false;
1644 }
1645 
1646 static bool
1647 vu_get_protocol_features_exec(VuDev *dev, VhostUserMsg *vmsg)
1648 {
1649     /*
1650      * Note that we support, but intentionally do not set,
1651      * VHOST_USER_PROTOCOL_F_INBAND_NOTIFICATIONS. This means that
1652      * a device implementation can return it in its callback
1653      * (get_protocol_features) if it wants to use this for
1654      * simulation, but it is otherwise not desirable (if even
1655      * implemented by the frontend.)
1656      */
1657     uint64_t features = 1ULL << VHOST_USER_PROTOCOL_F_MQ |
1658                         1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD |
1659                         1ULL << VHOST_USER_PROTOCOL_F_BACKEND_REQ |
1660                         1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER |
1661                         1ULL << VHOST_USER_PROTOCOL_F_BACKEND_SEND_FD |
1662                         1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK |
1663                         1ULL << VHOST_USER_PROTOCOL_F_CONFIGURE_MEM_SLOTS;
1664 
1665     if (have_userfault()) {
1666         features |= 1ULL << VHOST_USER_PROTOCOL_F_PAGEFAULT;
1667     }
1668 
1669     if (dev->iface->get_config && dev->iface->set_config) {
1670         features |= 1ULL << VHOST_USER_PROTOCOL_F_CONFIG;
1671     }
1672 
1673     if (dev->iface->get_protocol_features) {
1674         features |= dev->iface->get_protocol_features(dev);
1675     }
1676 
1677     vmsg_set_reply_u64(vmsg, features);
1678     return true;
1679 }
1680 
1681 static bool
1682 vu_set_protocol_features_exec(VuDev *dev, VhostUserMsg *vmsg)
1683 {
1684     uint64_t features = vmsg->payload.u64;
1685 
1686     DPRINT("u64: 0x%016"PRIx64"\n", features);
1687 
1688     dev->protocol_features = vmsg->payload.u64;
1689 
1690     if (vu_has_protocol_feature(dev,
1691                                 VHOST_USER_PROTOCOL_F_INBAND_NOTIFICATIONS) &&
1692         (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_BACKEND_REQ) ||
1693          !vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_REPLY_ACK))) {
1694         /*
1695          * The use case for using messages for kick/call is simulation, to make
1696          * the kick and call synchronous. To actually get that behaviour, both
1697          * of the other features are required.
1698          * Theoretically, one could use only kick messages, or do them without
1699          * having F_REPLY_ACK, but too many (possibly pending) messages on the
1700          * socket will eventually cause the frontend to hang, to avoid this in
1701          * scenarios where not desired enforce that the settings are in a way
1702          * that actually enables the simulation case.
1703          */
1704         vu_panic(dev,
1705                  "F_IN_BAND_NOTIFICATIONS requires F_BACKEND_REQ && F_REPLY_ACK");
1706         return false;
1707     }
1708 
1709     if (dev->iface->set_protocol_features) {
1710         dev->iface->set_protocol_features(dev, features);
1711     }
1712 
1713     return false;
1714 }
1715 
1716 static bool
1717 vu_get_queue_num_exec(VuDev *dev, VhostUserMsg *vmsg)
1718 {
1719     vmsg_set_reply_u64(vmsg, dev->max_queues);
1720     return true;
1721 }
1722 
1723 static bool
1724 vu_set_vring_enable_exec(VuDev *dev, VhostUserMsg *vmsg)
1725 {
1726     unsigned int index = vmsg->payload.state.index;
1727     unsigned int enable = vmsg->payload.state.num;
1728 
1729     DPRINT("State.index: %u\n", index);
1730     DPRINT("State.enable:   %u\n", enable);
1731 
1732     if (index >= dev->max_queues) {
1733         vu_panic(dev, "Invalid vring_enable index: %u", index);
1734         return false;
1735     }
1736 
1737     dev->vq[index].enable = enable;
1738     return false;
1739 }
1740 
1741 static bool
1742 vu_set_backend_req_fd(VuDev *dev, VhostUserMsg *vmsg)
1743 {
1744     if (vmsg->fd_num != 1) {
1745         vu_panic(dev, "Invalid backend_req_fd message (%d fd's)", vmsg->fd_num);
1746         return false;
1747     }
1748 
1749     if (dev->backend_fd != -1) {
1750         close(dev->backend_fd);
1751     }
1752     dev->backend_fd = vmsg->fds[0];
1753     DPRINT("Got backend_fd: %d\n", vmsg->fds[0]);
1754 
1755     return false;
1756 }
1757 
1758 static bool
1759 vu_get_config(VuDev *dev, VhostUserMsg *vmsg)
1760 {
1761     int ret = -1;
1762 
1763     if (dev->iface->get_config) {
1764         ret = dev->iface->get_config(dev, vmsg->payload.config.region,
1765                                      vmsg->payload.config.size);
1766     }
1767 
1768     if (ret) {
1769         /* resize to zero to indicate an error to frontend */
1770         vmsg->size = 0;
1771     }
1772 
1773     return true;
1774 }
1775 
1776 static bool
1777 vu_set_config(VuDev *dev, VhostUserMsg *vmsg)
1778 {
1779     int ret = -1;
1780 
1781     if (dev->iface->set_config) {
1782         ret = dev->iface->set_config(dev, vmsg->payload.config.region,
1783                                      vmsg->payload.config.offset,
1784                                      vmsg->payload.config.size,
1785                                      vmsg->payload.config.flags);
1786         if (ret) {
1787             vu_panic(dev, "Set virtio configuration space failed");
1788         }
1789     }
1790 
1791     return false;
1792 }
1793 
1794 static bool
1795 vu_set_postcopy_advise(VuDev *dev, VhostUserMsg *vmsg)
1796 {
1797 #ifdef UFFDIO_API
1798     struct uffdio_api api_struct;
1799 
1800     dev->postcopy_ufd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1801     vmsg->size = 0;
1802 #else
1803     dev->postcopy_ufd = -1;
1804 #endif
1805 
1806     if (dev->postcopy_ufd == -1) {
1807         vu_panic(dev, "Userfaultfd not available: %s", strerror(errno));
1808         goto out;
1809     }
1810 
1811 #ifdef UFFDIO_API
1812     api_struct.api = UFFD_API;
1813     api_struct.features = 0;
1814     if (ioctl(dev->postcopy_ufd, UFFDIO_API, &api_struct)) {
1815         vu_panic(dev, "Failed UFFDIO_API: %s", strerror(errno));
1816         close(dev->postcopy_ufd);
1817         dev->postcopy_ufd = -1;
1818         goto out;
1819     }
1820     /* TODO: Stash feature flags somewhere */
1821 #endif
1822 
1823 out:
1824     /* Return a ufd to the QEMU */
1825     vmsg->fd_num = 1;
1826     vmsg->fds[0] = dev->postcopy_ufd;
1827     return true; /* = send a reply */
1828 }
1829 
1830 static bool
1831 vu_set_postcopy_listen(VuDev *dev, VhostUserMsg *vmsg)
1832 {
1833     if (dev->nregions) {
1834         vu_panic(dev, "Regions already registered at postcopy-listen");
1835         vmsg_set_reply_u64(vmsg, -1);
1836         return true;
1837     }
1838     dev->postcopy_listening = true;
1839 
1840     vmsg_set_reply_u64(vmsg, 0);
1841     return true;
1842 }
1843 
1844 static bool
1845 vu_set_postcopy_end(VuDev *dev, VhostUserMsg *vmsg)
1846 {
1847     DPRINT("%s: Entry\n", __func__);
1848     dev->postcopy_listening = false;
1849     if (dev->postcopy_ufd > 0) {
1850         close(dev->postcopy_ufd);
1851         dev->postcopy_ufd = -1;
1852         DPRINT("%s: Done close\n", __func__);
1853     }
1854 
1855     vmsg_set_reply_u64(vmsg, 0);
1856     DPRINT("%s: exit\n", __func__);
1857     return true;
1858 }
1859 
1860 static inline uint64_t
1861 vu_inflight_queue_size(uint16_t queue_size)
1862 {
1863     return ALIGN_UP(sizeof(VuDescStateSplit) * queue_size +
1864            sizeof(uint16_t), INFLIGHT_ALIGNMENT);
1865 }
1866 
1867 #ifdef MFD_ALLOW_SEALING
1868 static void *
1869 memfd_alloc(const char *name, size_t size, unsigned int flags, int *fd)
1870 {
1871     void *ptr;
1872     int ret;
1873 
1874     *fd = memfd_create(name, MFD_ALLOW_SEALING);
1875     if (*fd < 0) {
1876         return NULL;
1877     }
1878 
1879     ret = ftruncate(*fd, size);
1880     if (ret < 0) {
1881         close(*fd);
1882         return NULL;
1883     }
1884 
1885     ret = fcntl(*fd, F_ADD_SEALS, flags);
1886     if (ret < 0) {
1887         close(*fd);
1888         return NULL;
1889     }
1890 
1891     ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, *fd, 0);
1892     if (ptr == MAP_FAILED) {
1893         close(*fd);
1894         return NULL;
1895     }
1896 
1897     return ptr;
1898 }
1899 #endif
1900 
1901 static bool
1902 vu_get_inflight_fd(VuDev *dev, VhostUserMsg *vmsg)
1903 {
1904     int fd = -1;
1905     void *addr = NULL;
1906     uint64_t mmap_size;
1907     uint16_t num_queues, queue_size;
1908 
1909     if (vmsg->size != sizeof(vmsg->payload.inflight)) {
1910         vu_panic(dev, "Invalid get_inflight_fd message:%d", vmsg->size);
1911         vmsg->payload.inflight.mmap_size = 0;
1912         return true;
1913     }
1914 
1915     num_queues = vmsg->payload.inflight.num_queues;
1916     queue_size = vmsg->payload.inflight.queue_size;
1917 
1918     DPRINT("set_inflight_fd num_queues: %"PRId16"\n", num_queues);
1919     DPRINT("set_inflight_fd queue_size: %"PRId16"\n", queue_size);
1920 
1921     mmap_size = vu_inflight_queue_size(queue_size) * num_queues;
1922 
1923 #ifdef MFD_ALLOW_SEALING
1924     addr = memfd_alloc("vhost-inflight", mmap_size,
1925                        F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1926                        &fd);
1927 #else
1928     vu_panic(dev, "Not implemented: memfd support is missing");
1929 #endif
1930 
1931     if (!addr) {
1932         vu_panic(dev, "Failed to alloc vhost inflight area");
1933         vmsg->payload.inflight.mmap_size = 0;
1934         return true;
1935     }
1936 
1937     memset(addr, 0, mmap_size);
1938 
1939     dev->inflight_info.addr = addr;
1940     dev->inflight_info.size = vmsg->payload.inflight.mmap_size = mmap_size;
1941     dev->inflight_info.fd = vmsg->fds[0] = fd;
1942     vmsg->fd_num = 1;
1943     vmsg->payload.inflight.mmap_offset = 0;
1944 
1945     DPRINT("send inflight mmap_size: %"PRId64"\n",
1946            vmsg->payload.inflight.mmap_size);
1947     DPRINT("send inflight mmap offset: %"PRId64"\n",
1948            vmsg->payload.inflight.mmap_offset);
1949 
1950     return true;
1951 }
1952 
1953 static bool
1954 vu_set_inflight_fd(VuDev *dev, VhostUserMsg *vmsg)
1955 {
1956     int fd, i;
1957     uint64_t mmap_size, mmap_offset;
1958     uint16_t num_queues, queue_size;
1959     void *rc;
1960 
1961     if (vmsg->fd_num != 1 ||
1962         vmsg->size != sizeof(vmsg->payload.inflight)) {
1963         vu_panic(dev, "Invalid set_inflight_fd message size:%d fds:%d",
1964                  vmsg->size, vmsg->fd_num);
1965         return false;
1966     }
1967 
1968     fd = vmsg->fds[0];
1969     mmap_size = vmsg->payload.inflight.mmap_size;
1970     mmap_offset = vmsg->payload.inflight.mmap_offset;
1971     num_queues = vmsg->payload.inflight.num_queues;
1972     queue_size = vmsg->payload.inflight.queue_size;
1973 
1974     DPRINT("set_inflight_fd mmap_size: %"PRId64"\n", mmap_size);
1975     DPRINT("set_inflight_fd mmap_offset: %"PRId64"\n", mmap_offset);
1976     DPRINT("set_inflight_fd num_queues: %"PRId16"\n", num_queues);
1977     DPRINT("set_inflight_fd queue_size: %"PRId16"\n", queue_size);
1978 
1979     rc = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
1980               fd, mmap_offset);
1981 
1982     if (rc == MAP_FAILED) {
1983         vu_panic(dev, "set_inflight_fd mmap error: %s", strerror(errno));
1984         return false;
1985     }
1986 
1987     if (dev->inflight_info.fd) {
1988         close(dev->inflight_info.fd);
1989     }
1990 
1991     if (dev->inflight_info.addr) {
1992         munmap(dev->inflight_info.addr, dev->inflight_info.size);
1993     }
1994 
1995     dev->inflight_info.fd = fd;
1996     dev->inflight_info.addr = rc;
1997     dev->inflight_info.size = mmap_size;
1998 
1999     for (i = 0; i < num_queues; i++) {
2000         dev->vq[i].inflight = (VuVirtqInflight *)rc;
2001         dev->vq[i].inflight->desc_num = queue_size;
2002         rc = (void *)((char *)rc + vu_inflight_queue_size(queue_size));
2003     }
2004 
2005     return false;
2006 }
2007 
2008 static bool
2009 vu_handle_vring_kick(VuDev *dev, VhostUserMsg *vmsg)
2010 {
2011     unsigned int index = vmsg->payload.state.index;
2012 
2013     if (index >= dev->max_queues) {
2014         vu_panic(dev, "Invalid queue index: %u", index);
2015         return false;
2016     }
2017 
2018     DPRINT("Got kick message: handler:%p idx:%u\n",
2019            dev->vq[index].handler, index);
2020 
2021     if (!dev->vq[index].started) {
2022         dev->vq[index].started = true;
2023 
2024         if (dev->iface->queue_set_started) {
2025             dev->iface->queue_set_started(dev, index, true);
2026         }
2027     }
2028 
2029     if (dev->vq[index].handler) {
2030         dev->vq[index].handler(dev, index);
2031     }
2032 
2033     return false;
2034 }
2035 
2036 static bool vu_handle_get_max_memslots(VuDev *dev, VhostUserMsg *vmsg)
2037 {
2038     vmsg_set_reply_u64(vmsg, VHOST_USER_MAX_RAM_SLOTS);
2039 
2040     DPRINT("u64: 0x%016"PRIx64"\n", (uint64_t) VHOST_USER_MAX_RAM_SLOTS);
2041 
2042     return true;
2043 }
2044 
2045 static bool
2046 vu_process_message(VuDev *dev, VhostUserMsg *vmsg)
2047 {
2048     int do_reply = 0;
2049 
2050     /* Print out generic part of the request. */
2051     DPRINT("================ Vhost user message ================\n");
2052     DPRINT("Request: %s (%d)\n", vu_request_to_string(vmsg->request),
2053            vmsg->request);
2054     DPRINT("Flags:   0x%x\n", vmsg->flags);
2055     DPRINT("Size:    %u\n", vmsg->size);
2056 
2057     if (vmsg->fd_num) {
2058         int i;
2059         DPRINT("Fds:");
2060         for (i = 0; i < vmsg->fd_num; i++) {
2061             DPRINT(" %d", vmsg->fds[i]);
2062         }
2063         DPRINT("\n");
2064     }
2065 
2066     if (dev->iface->process_msg &&
2067         dev->iface->process_msg(dev, vmsg, &do_reply)) {
2068         return do_reply;
2069     }
2070 
2071     switch (vmsg->request) {
2072     case VHOST_USER_GET_FEATURES:
2073         return vu_get_features_exec(dev, vmsg);
2074     case VHOST_USER_SET_FEATURES:
2075         return vu_set_features_exec(dev, vmsg);
2076     case VHOST_USER_GET_PROTOCOL_FEATURES:
2077         return vu_get_protocol_features_exec(dev, vmsg);
2078     case VHOST_USER_SET_PROTOCOL_FEATURES:
2079         return vu_set_protocol_features_exec(dev, vmsg);
2080     case VHOST_USER_SET_OWNER:
2081         return vu_set_owner_exec(dev, vmsg);
2082     case VHOST_USER_RESET_OWNER:
2083         return vu_reset_device_exec(dev, vmsg);
2084     case VHOST_USER_SET_MEM_TABLE:
2085         return vu_set_mem_table_exec(dev, vmsg);
2086     case VHOST_USER_SET_LOG_BASE:
2087         return vu_set_log_base_exec(dev, vmsg);
2088     case VHOST_USER_SET_LOG_FD:
2089         return vu_set_log_fd_exec(dev, vmsg);
2090     case VHOST_USER_SET_VRING_NUM:
2091         return vu_set_vring_num_exec(dev, vmsg);
2092     case VHOST_USER_SET_VRING_ADDR:
2093         return vu_set_vring_addr_exec(dev, vmsg);
2094     case VHOST_USER_SET_VRING_BASE:
2095         return vu_set_vring_base_exec(dev, vmsg);
2096     case VHOST_USER_GET_VRING_BASE:
2097         return vu_get_vring_base_exec(dev, vmsg);
2098     case VHOST_USER_SET_VRING_KICK:
2099         return vu_set_vring_kick_exec(dev, vmsg);
2100     case VHOST_USER_SET_VRING_CALL:
2101         return vu_set_vring_call_exec(dev, vmsg);
2102     case VHOST_USER_SET_VRING_ERR:
2103         return vu_set_vring_err_exec(dev, vmsg);
2104     case VHOST_USER_GET_QUEUE_NUM:
2105         return vu_get_queue_num_exec(dev, vmsg);
2106     case VHOST_USER_SET_VRING_ENABLE:
2107         return vu_set_vring_enable_exec(dev, vmsg);
2108     case VHOST_USER_SET_BACKEND_REQ_FD:
2109         return vu_set_backend_req_fd(dev, vmsg);
2110     case VHOST_USER_GET_CONFIG:
2111         return vu_get_config(dev, vmsg);
2112     case VHOST_USER_SET_CONFIG:
2113         return vu_set_config(dev, vmsg);
2114     case VHOST_USER_NONE:
2115         /* if you need processing before exit, override iface->process_msg */
2116         exit(0);
2117     case VHOST_USER_POSTCOPY_ADVISE:
2118         return vu_set_postcopy_advise(dev, vmsg);
2119     case VHOST_USER_POSTCOPY_LISTEN:
2120         return vu_set_postcopy_listen(dev, vmsg);
2121     case VHOST_USER_POSTCOPY_END:
2122         return vu_set_postcopy_end(dev, vmsg);
2123     case VHOST_USER_GET_INFLIGHT_FD:
2124         return vu_get_inflight_fd(dev, vmsg);
2125     case VHOST_USER_SET_INFLIGHT_FD:
2126         return vu_set_inflight_fd(dev, vmsg);
2127     case VHOST_USER_VRING_KICK:
2128         return vu_handle_vring_kick(dev, vmsg);
2129     case VHOST_USER_GET_MAX_MEM_SLOTS:
2130         return vu_handle_get_max_memslots(dev, vmsg);
2131     case VHOST_USER_ADD_MEM_REG:
2132         return vu_add_mem_reg(dev, vmsg);
2133     case VHOST_USER_REM_MEM_REG:
2134         return vu_rem_mem_reg(dev, vmsg);
2135     case VHOST_USER_GET_SHARED_OBJECT:
2136         return vu_get_shared_object(dev, vmsg);
2137     default:
2138         vmsg_close_fds(vmsg);
2139         vu_panic(dev, "Unhandled request: %d", vmsg->request);
2140     }
2141 
2142     return false;
2143 }
2144 
2145 bool
2146 vu_dispatch(VuDev *dev)
2147 {
2148     VhostUserMsg vmsg = { 0, };
2149     int reply_requested;
2150     bool need_reply, success = false;
2151 
2152     if (!dev->read_msg(dev, dev->sock, &vmsg)) {
2153         goto end;
2154     }
2155 
2156     need_reply = vmsg.flags & VHOST_USER_NEED_REPLY_MASK;
2157 
2158     reply_requested = vu_process_message(dev, &vmsg);
2159     if (!reply_requested && need_reply) {
2160         vmsg_set_reply_u64(&vmsg, 0);
2161         reply_requested = 1;
2162     }
2163 
2164     if (!reply_requested) {
2165         success = true;
2166         goto end;
2167     }
2168 
2169     if (!vu_send_reply(dev, dev->sock, &vmsg)) {
2170         goto end;
2171     }
2172 
2173     success = true;
2174 
2175 end:
2176     free(vmsg.data);
2177     return success;
2178 }
2179 
2180 void
2181 vu_deinit(VuDev *dev)
2182 {
2183     unsigned int i;
2184 
2185     vu_remove_all_mem_regs(dev);
2186 
2187     for (i = 0; i < dev->max_queues; i++) {
2188         VuVirtq *vq = &dev->vq[i];
2189 
2190         if (vq->call_fd != -1) {
2191             close(vq->call_fd);
2192             vq->call_fd = -1;
2193         }
2194 
2195         if (vq->kick_fd != -1) {
2196             dev->remove_watch(dev, vq->kick_fd);
2197             close(vq->kick_fd);
2198             vq->kick_fd = -1;
2199         }
2200 
2201         if (vq->err_fd != -1) {
2202             close(vq->err_fd);
2203             vq->err_fd = -1;
2204         }
2205 
2206         if (vq->resubmit_list) {
2207             free(vq->resubmit_list);
2208             vq->resubmit_list = NULL;
2209         }
2210 
2211         vq->inflight = NULL;
2212     }
2213 
2214     if (dev->inflight_info.addr) {
2215         munmap(dev->inflight_info.addr, dev->inflight_info.size);
2216         dev->inflight_info.addr = NULL;
2217     }
2218 
2219     if (dev->inflight_info.fd > 0) {
2220         close(dev->inflight_info.fd);
2221         dev->inflight_info.fd = -1;
2222     }
2223 
2224     vu_close_log(dev);
2225     if (dev->backend_fd != -1) {
2226         close(dev->backend_fd);
2227         dev->backend_fd = -1;
2228     }
2229     pthread_mutex_destroy(&dev->backend_mutex);
2230 
2231     if (dev->sock != -1) {
2232         close(dev->sock);
2233     }
2234 
2235     free(dev->vq);
2236     dev->vq = NULL;
2237     free(dev->regions);
2238     dev->regions = NULL;
2239 }
2240 
2241 bool
2242 vu_init(VuDev *dev,
2243         uint16_t max_queues,
2244         int socket,
2245         vu_panic_cb panic,
2246         vu_read_msg_cb read_msg,
2247         vu_set_watch_cb set_watch,
2248         vu_remove_watch_cb remove_watch,
2249         const VuDevIface *iface)
2250 {
2251     uint16_t i;
2252 
2253     assert(max_queues > 0);
2254     assert(socket >= 0);
2255     assert(set_watch);
2256     assert(remove_watch);
2257     assert(iface);
2258     assert(panic);
2259 
2260     memset(dev, 0, sizeof(*dev));
2261 
2262     dev->sock = socket;
2263     dev->panic = panic;
2264     dev->read_msg = read_msg ? read_msg : vu_message_read_default;
2265     dev->set_watch = set_watch;
2266     dev->remove_watch = remove_watch;
2267     dev->iface = iface;
2268     dev->log_call_fd = -1;
2269     pthread_mutex_init(&dev->backend_mutex, NULL);
2270     dev->backend_fd = -1;
2271     dev->max_queues = max_queues;
2272 
2273     dev->regions = malloc(VHOST_USER_MAX_RAM_SLOTS * sizeof(dev->regions[0]));
2274     if (!dev->regions) {
2275         DPRINT("%s: failed to malloc mem regions\n", __func__);
2276         return false;
2277     }
2278 
2279     dev->vq = malloc(max_queues * sizeof(dev->vq[0]));
2280     if (!dev->vq) {
2281         DPRINT("%s: failed to malloc virtqueues\n", __func__);
2282         free(dev->regions);
2283         dev->regions = NULL;
2284         return false;
2285     }
2286 
2287     for (i = 0; i < max_queues; i++) {
2288         dev->vq[i] = (VuVirtq) {
2289             .call_fd = -1, .kick_fd = -1, .err_fd = -1,
2290             .notification = true,
2291         };
2292     }
2293 
2294     return true;
2295 }
2296 
2297 VuVirtq *
2298 vu_get_queue(VuDev *dev, int qidx)
2299 {
2300     assert(qidx < dev->max_queues);
2301     return &dev->vq[qidx];
2302 }
2303 
2304 bool
2305 vu_queue_enabled(VuDev *dev, VuVirtq *vq)
2306 {
2307     return vq->enable;
2308 }
2309 
2310 bool
2311 vu_queue_started(const VuDev *dev, const VuVirtq *vq)
2312 {
2313     return vq->started;
2314 }
2315 
2316 static inline uint16_t
2317 vring_avail_flags(VuVirtq *vq)
2318 {
2319     return le16toh(vq->vring.avail->flags);
2320 }
2321 
2322 static inline uint16_t
2323 vring_avail_idx(VuVirtq *vq)
2324 {
2325     vq->shadow_avail_idx = le16toh(vq->vring.avail->idx);
2326 
2327     return vq->shadow_avail_idx;
2328 }
2329 
2330 static inline uint16_t
2331 vring_avail_ring(VuVirtq *vq, int i)
2332 {
2333     return le16toh(vq->vring.avail->ring[i]);
2334 }
2335 
2336 static inline uint16_t
2337 vring_get_used_event(VuVirtq *vq)
2338 {
2339     return vring_avail_ring(vq, vq->vring.num);
2340 }
2341 
2342 static int
2343 virtqueue_num_heads(VuDev *dev, VuVirtq *vq, unsigned int idx)
2344 {
2345     uint16_t num_heads = vring_avail_idx(vq) - idx;
2346 
2347     /* Check it isn't doing very strange things with descriptor numbers. */
2348     if (num_heads > vq->vring.num) {
2349         vu_panic(dev, "Guest moved used index from %u to %u",
2350                  idx, vq->shadow_avail_idx);
2351         return -1;
2352     }
2353     if (num_heads) {
2354         /* On success, callers read a descriptor at vq->last_avail_idx.
2355          * Make sure descriptor read does not bypass avail index read. */
2356         smp_rmb();
2357     }
2358 
2359     return num_heads;
2360 }
2361 
2362 static bool
2363 virtqueue_get_head(VuDev *dev, VuVirtq *vq,
2364                    unsigned int idx, unsigned int *head)
2365 {
2366     /* Grab the next descriptor number they're advertising, and increment
2367      * the index we've seen. */
2368     *head = vring_avail_ring(vq, idx % vq->vring.num);
2369 
2370     /* If their number is silly, that's a fatal mistake. */
2371     if (*head >= vq->vring.num) {
2372         vu_panic(dev, "Guest says index %u is available", *head);
2373         return false;
2374     }
2375 
2376     return true;
2377 }
2378 
2379 static int
2380 virtqueue_read_indirect_desc(VuDev *dev, struct vring_desc *desc,
2381                              uint64_t addr, size_t len)
2382 {
2383     struct vring_desc *ori_desc;
2384     uint64_t read_len;
2385 
2386     if (len > (VIRTQUEUE_MAX_SIZE * sizeof(struct vring_desc))) {
2387         return -1;
2388     }
2389 
2390     if (len == 0) {
2391         return -1;
2392     }
2393 
2394     while (len) {
2395         read_len = len;
2396         ori_desc = vu_gpa_to_va(dev, &read_len, addr);
2397         if (!ori_desc) {
2398             return -1;
2399         }
2400 
2401         memcpy(desc, ori_desc, read_len);
2402         len -= read_len;
2403         addr += read_len;
2404         desc += read_len;
2405     }
2406 
2407     return 0;
2408 }
2409 
2410 enum {
2411     VIRTQUEUE_READ_DESC_ERROR = -1,
2412     VIRTQUEUE_READ_DESC_DONE = 0,   /* end of chain */
2413     VIRTQUEUE_READ_DESC_MORE = 1,   /* more buffers in chain */
2414 };
2415 
2416 static int
2417 virtqueue_read_next_desc(VuDev *dev, struct vring_desc *desc,
2418                          int i, unsigned int max, unsigned int *next)
2419 {
2420     /* If this descriptor says it doesn't chain, we're done. */
2421     if (!(le16toh(desc[i].flags) & VRING_DESC_F_NEXT)) {
2422         return VIRTQUEUE_READ_DESC_DONE;
2423     }
2424 
2425     /* Check they're not leading us off end of descriptors. */
2426     *next = le16toh(desc[i].next);
2427     /* Make sure compiler knows to grab that: we don't want it changing! */
2428     smp_wmb();
2429 
2430     if (*next >= max) {
2431         vu_panic(dev, "Desc next is %u", *next);
2432         return VIRTQUEUE_READ_DESC_ERROR;
2433     }
2434 
2435     return VIRTQUEUE_READ_DESC_MORE;
2436 }
2437 
2438 void
2439 vu_queue_get_avail_bytes(VuDev *dev, VuVirtq *vq, unsigned int *in_bytes,
2440                          unsigned int *out_bytes,
2441                          unsigned max_in_bytes, unsigned max_out_bytes)
2442 {
2443     unsigned int idx;
2444     unsigned int total_bufs, in_total, out_total;
2445     int rc;
2446 
2447     idx = vq->last_avail_idx;
2448 
2449     total_bufs = in_total = out_total = 0;
2450     if (!vu_is_vq_usable(dev, vq)) {
2451         goto done;
2452     }
2453 
2454     while ((rc = virtqueue_num_heads(dev, vq, idx)) > 0) {
2455         unsigned int max, desc_len, num_bufs, indirect = 0;
2456         uint64_t desc_addr, read_len;
2457         struct vring_desc *desc;
2458         struct vring_desc desc_buf[VIRTQUEUE_MAX_SIZE];
2459         unsigned int i;
2460 
2461         max = vq->vring.num;
2462         num_bufs = total_bufs;
2463         if (!virtqueue_get_head(dev, vq, idx++, &i)) {
2464             goto err;
2465         }
2466         desc = vq->vring.desc;
2467 
2468         if (le16toh(desc[i].flags) & VRING_DESC_F_INDIRECT) {
2469             if (le32toh(desc[i].len) % sizeof(struct vring_desc)) {
2470                 vu_panic(dev, "Invalid size for indirect buffer table");
2471                 goto err;
2472             }
2473 
2474             /* If we've got too many, that implies a descriptor loop. */
2475             if (num_bufs >= max) {
2476                 vu_panic(dev, "Looped descriptor");
2477                 goto err;
2478             }
2479 
2480             /* loop over the indirect descriptor table */
2481             indirect = 1;
2482             desc_addr = le64toh(desc[i].addr);
2483             desc_len = le32toh(desc[i].len);
2484             max = desc_len / sizeof(struct vring_desc);
2485             read_len = desc_len;
2486             desc = vu_gpa_to_va(dev, &read_len, desc_addr);
2487             if (unlikely(desc && read_len != desc_len)) {
2488                 /* Failed to use zero copy */
2489                 desc = NULL;
2490                 if (!virtqueue_read_indirect_desc(dev, desc_buf,
2491                                                   desc_addr,
2492                                                   desc_len)) {
2493                     desc = desc_buf;
2494                 }
2495             }
2496             if (!desc) {
2497                 vu_panic(dev, "Invalid indirect buffer table");
2498                 goto err;
2499             }
2500             num_bufs = i = 0;
2501         }
2502 
2503         do {
2504             /* If we've got too many, that implies a descriptor loop. */
2505             if (++num_bufs > max) {
2506                 vu_panic(dev, "Looped descriptor");
2507                 goto err;
2508             }
2509 
2510             if (le16toh(desc[i].flags) & VRING_DESC_F_WRITE) {
2511                 in_total += le32toh(desc[i].len);
2512             } else {
2513                 out_total += le32toh(desc[i].len);
2514             }
2515             if (in_total >= max_in_bytes && out_total >= max_out_bytes) {
2516                 goto done;
2517             }
2518             rc = virtqueue_read_next_desc(dev, desc, i, max, &i);
2519         } while (rc == VIRTQUEUE_READ_DESC_MORE);
2520 
2521         if (rc == VIRTQUEUE_READ_DESC_ERROR) {
2522             goto err;
2523         }
2524 
2525         if (!indirect) {
2526             total_bufs = num_bufs;
2527         } else {
2528             total_bufs++;
2529         }
2530     }
2531     if (rc < 0) {
2532         goto err;
2533     }
2534 done:
2535     if (in_bytes) {
2536         *in_bytes = in_total;
2537     }
2538     if (out_bytes) {
2539         *out_bytes = out_total;
2540     }
2541     return;
2542 
2543 err:
2544     in_total = out_total = 0;
2545     goto done;
2546 }
2547 
2548 bool
2549 vu_queue_avail_bytes(VuDev *dev, VuVirtq *vq, unsigned int in_bytes,
2550                      unsigned int out_bytes)
2551 {
2552     unsigned int in_total, out_total;
2553 
2554     vu_queue_get_avail_bytes(dev, vq, &in_total, &out_total,
2555                              in_bytes, out_bytes);
2556 
2557     return in_bytes <= in_total && out_bytes <= out_total;
2558 }
2559 
2560 /* Fetch avail_idx from VQ memory only when we really need to know if
2561  * guest has added some buffers. */
2562 bool
2563 vu_queue_empty(VuDev *dev, VuVirtq *vq)
2564 {
2565     if (!vu_is_vq_usable(dev, vq)) {
2566         return true;
2567     }
2568 
2569     if (vq->shadow_avail_idx != vq->last_avail_idx) {
2570         return false;
2571     }
2572 
2573     return vring_avail_idx(vq) == vq->last_avail_idx;
2574 }
2575 
2576 static bool
2577 vring_notify(VuDev *dev, VuVirtq *vq)
2578 {
2579     uint16_t old, new;
2580     bool v;
2581 
2582     /* We need to expose used array entries before checking used event. */
2583     smp_mb();
2584 
2585     /* Always notify when queue is empty (when feature acknowledge) */
2586     if (vu_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2587         !vq->inuse && vu_queue_empty(dev, vq)) {
2588         return true;
2589     }
2590 
2591     if (!vu_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
2592         return !(vring_avail_flags(vq) & VRING_AVAIL_F_NO_INTERRUPT);
2593     }
2594 
2595     v = vq->signalled_used_valid;
2596     vq->signalled_used_valid = true;
2597     old = vq->signalled_used;
2598     new = vq->signalled_used = vq->used_idx;
2599     return !v || vring_need_event(vring_get_used_event(vq), new, old);
2600 }
2601 
2602 static void _vu_queue_notify(VuDev *dev, VuVirtq *vq, bool sync)
2603 {
2604     if (!vu_is_vq_usable(dev, vq)) {
2605         return;
2606     }
2607 
2608     if (!vring_notify(dev, vq)) {
2609         DPRINT("skipped notify...\n");
2610         return;
2611     }
2612 
2613     if (vq->call_fd < 0 &&
2614         vu_has_protocol_feature(dev,
2615                                 VHOST_USER_PROTOCOL_F_INBAND_NOTIFICATIONS) &&
2616         vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_BACKEND_REQ)) {
2617         VhostUserMsg vmsg = {
2618             .request = VHOST_USER_BACKEND_VRING_CALL,
2619             .flags = VHOST_USER_VERSION,
2620             .size = sizeof(vmsg.payload.state),
2621             .payload.state = {
2622                 .index = vq - dev->vq,
2623             },
2624         };
2625         bool ack = sync &&
2626                    vu_has_protocol_feature(dev,
2627                                            VHOST_USER_PROTOCOL_F_REPLY_ACK);
2628 
2629         if (ack) {
2630             vmsg.flags |= VHOST_USER_NEED_REPLY_MASK;
2631         }
2632 
2633         vu_message_write(dev, dev->backend_fd, &vmsg);
2634         if (ack) {
2635             vu_message_read_default(dev, dev->backend_fd, &vmsg);
2636         }
2637         return;
2638     }
2639 
2640     if (eventfd_write(vq->call_fd, 1) < 0) {
2641         vu_panic(dev, "Error writing eventfd: %s", strerror(errno));
2642     }
2643 }
2644 
2645 void vu_queue_notify(VuDev *dev, VuVirtq *vq)
2646 {
2647     _vu_queue_notify(dev, vq, false);
2648 }
2649 
2650 void vu_queue_notify_sync(VuDev *dev, VuVirtq *vq)
2651 {
2652     _vu_queue_notify(dev, vq, true);
2653 }
2654 
2655 void vu_config_change_msg(VuDev *dev)
2656 {
2657     VhostUserMsg vmsg = {
2658         .request = VHOST_USER_BACKEND_CONFIG_CHANGE_MSG,
2659         .flags = VHOST_USER_VERSION,
2660     };
2661 
2662     vu_message_write(dev, dev->backend_fd, &vmsg);
2663 }
2664 
2665 static inline void
2666 vring_used_flags_set_bit(VuVirtq *vq, int mask)
2667 {
2668     uint16_t *flags;
2669 
2670     flags = (uint16_t *)((char*)vq->vring.used +
2671                          offsetof(struct vring_used, flags));
2672     *flags = htole16(le16toh(*flags) | mask);
2673 }
2674 
2675 static inline void
2676 vring_used_flags_unset_bit(VuVirtq *vq, int mask)
2677 {
2678     uint16_t *flags;
2679 
2680     flags = (uint16_t *)((char*)vq->vring.used +
2681                          offsetof(struct vring_used, flags));
2682     *flags = htole16(le16toh(*flags) & ~mask);
2683 }
2684 
2685 static inline void
2686 vring_set_avail_event(VuVirtq *vq, uint16_t val)
2687 {
2688     uint16_t val_le = htole16(val);
2689 
2690     if (!vq->notification) {
2691         return;
2692     }
2693 
2694     memcpy(&vq->vring.used->ring[vq->vring.num], &val_le, sizeof(uint16_t));
2695 }
2696 
2697 void
2698 vu_queue_set_notification(VuDev *dev, VuVirtq *vq, int enable)
2699 {
2700     vq->notification = enable;
2701     if (vu_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
2702         vring_set_avail_event(vq, vring_avail_idx(vq));
2703     } else if (enable) {
2704         vring_used_flags_unset_bit(vq, VRING_USED_F_NO_NOTIFY);
2705     } else {
2706         vring_used_flags_set_bit(vq, VRING_USED_F_NO_NOTIFY);
2707     }
2708     if (enable) {
2709         /* Expose avail event/used flags before caller checks the avail idx. */
2710         smp_mb();
2711     }
2712 }
2713 
2714 static bool
2715 virtqueue_map_desc(VuDev *dev,
2716                    unsigned int *p_num_sg, struct iovec *iov,
2717                    unsigned int max_num_sg, bool is_write,
2718                    uint64_t pa, size_t sz)
2719 {
2720     unsigned num_sg = *p_num_sg;
2721 
2722     assert(num_sg <= max_num_sg);
2723 
2724     if (!sz) {
2725         vu_panic(dev, "virtio: zero sized buffers are not allowed");
2726         return false;
2727     }
2728 
2729     while (sz) {
2730         uint64_t len = sz;
2731 
2732         if (num_sg == max_num_sg) {
2733             vu_panic(dev, "virtio: too many descriptors in indirect table");
2734             return false;
2735         }
2736 
2737         iov[num_sg].iov_base = vu_gpa_to_va(dev, &len, pa);
2738         if (iov[num_sg].iov_base == NULL) {
2739             vu_panic(dev, "virtio: invalid address for buffers");
2740             return false;
2741         }
2742         iov[num_sg].iov_len = len;
2743         num_sg++;
2744         sz -= len;
2745         pa += len;
2746     }
2747 
2748     *p_num_sg = num_sg;
2749     return true;
2750 }
2751 
2752 static void *
2753 virtqueue_alloc_element(size_t sz,
2754                                      unsigned out_num, unsigned in_num)
2755 {
2756     VuVirtqElement *elem;
2757     size_t in_sg_ofs = ALIGN_UP(sz, __alignof__(elem->in_sg[0]));
2758     size_t out_sg_ofs = in_sg_ofs + in_num * sizeof(elem->in_sg[0]);
2759     size_t out_sg_end = out_sg_ofs + out_num * sizeof(elem->out_sg[0]);
2760 
2761     assert(sz >= sizeof(VuVirtqElement));
2762     elem = malloc(out_sg_end);
2763     if (!elem) {
2764         DPRINT("%s: failed to malloc virtqueue element\n", __func__);
2765         return NULL;
2766     }
2767     elem->out_num = out_num;
2768     elem->in_num = in_num;
2769     elem->in_sg = (void *)elem + in_sg_ofs;
2770     elem->out_sg = (void *)elem + out_sg_ofs;
2771     return elem;
2772 }
2773 
2774 static void *
2775 vu_queue_map_desc(VuDev *dev, VuVirtq *vq, unsigned int idx, size_t sz)
2776 {
2777     struct vring_desc *desc = vq->vring.desc;
2778     uint64_t desc_addr, read_len;
2779     unsigned int desc_len;
2780     unsigned int max = vq->vring.num;
2781     unsigned int i = idx;
2782     VuVirtqElement *elem;
2783     unsigned int out_num = 0, in_num = 0;
2784     struct iovec iov[VIRTQUEUE_MAX_SIZE];
2785     struct vring_desc desc_buf[VIRTQUEUE_MAX_SIZE];
2786     int rc;
2787 
2788     if (le16toh(desc[i].flags) & VRING_DESC_F_INDIRECT) {
2789         if (le32toh(desc[i].len) % sizeof(struct vring_desc)) {
2790             vu_panic(dev, "Invalid size for indirect buffer table");
2791             return NULL;
2792         }
2793 
2794         /* loop over the indirect descriptor table */
2795         desc_addr = le64toh(desc[i].addr);
2796         desc_len = le32toh(desc[i].len);
2797         max = desc_len / sizeof(struct vring_desc);
2798         read_len = desc_len;
2799         desc = vu_gpa_to_va(dev, &read_len, desc_addr);
2800         if (unlikely(desc && read_len != desc_len)) {
2801             /* Failed to use zero copy */
2802             desc = NULL;
2803             if (!virtqueue_read_indirect_desc(dev, desc_buf,
2804                                               desc_addr,
2805                                               desc_len)) {
2806                 desc = desc_buf;
2807             }
2808         }
2809         if (!desc) {
2810             vu_panic(dev, "Invalid indirect buffer table");
2811             return NULL;
2812         }
2813         i = 0;
2814     }
2815 
2816     /* Collect all the descriptors */
2817     do {
2818         if (le16toh(desc[i].flags) & VRING_DESC_F_WRITE) {
2819             if (!virtqueue_map_desc(dev, &in_num, iov + out_num,
2820                                VIRTQUEUE_MAX_SIZE - out_num, true,
2821                                le64toh(desc[i].addr),
2822                                le32toh(desc[i].len))) {
2823                 return NULL;
2824             }
2825         } else {
2826             if (in_num) {
2827                 vu_panic(dev, "Incorrect order for descriptors");
2828                 return NULL;
2829             }
2830             if (!virtqueue_map_desc(dev, &out_num, iov,
2831                                VIRTQUEUE_MAX_SIZE, false,
2832                                le64toh(desc[i].addr),
2833                                le32toh(desc[i].len))) {
2834                 return NULL;
2835             }
2836         }
2837 
2838         /* If we've got too many, that implies a descriptor loop. */
2839         if ((in_num + out_num) > max) {
2840             vu_panic(dev, "Looped descriptor");
2841             return NULL;
2842         }
2843         rc = virtqueue_read_next_desc(dev, desc, i, max, &i);
2844     } while (rc == VIRTQUEUE_READ_DESC_MORE);
2845 
2846     if (rc == VIRTQUEUE_READ_DESC_ERROR) {
2847         vu_panic(dev, "read descriptor error");
2848         return NULL;
2849     }
2850 
2851     /* Now copy what we have collected and mapped */
2852     elem = virtqueue_alloc_element(sz, out_num, in_num);
2853     if (!elem) {
2854         return NULL;
2855     }
2856     elem->index = idx;
2857     for (i = 0; i < out_num; i++) {
2858         elem->out_sg[i] = iov[i];
2859     }
2860     for (i = 0; i < in_num; i++) {
2861         elem->in_sg[i] = iov[out_num + i];
2862     }
2863 
2864     return elem;
2865 }
2866 
2867 static int
2868 vu_queue_inflight_get(VuDev *dev, VuVirtq *vq, int desc_idx)
2869 {
2870     if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)) {
2871         return 0;
2872     }
2873 
2874     if (unlikely(!vq->inflight)) {
2875         return -1;
2876     }
2877 
2878     vq->inflight->desc[desc_idx].counter = vq->counter++;
2879     vq->inflight->desc[desc_idx].inflight = 1;
2880 
2881     return 0;
2882 }
2883 
2884 static int
2885 vu_queue_inflight_pre_put(VuDev *dev, VuVirtq *vq, int desc_idx)
2886 {
2887     if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)) {
2888         return 0;
2889     }
2890 
2891     if (unlikely(!vq->inflight)) {
2892         return -1;
2893     }
2894 
2895     vq->inflight->last_batch_head = desc_idx;
2896 
2897     return 0;
2898 }
2899 
2900 static int
2901 vu_queue_inflight_post_put(VuDev *dev, VuVirtq *vq, int desc_idx)
2902 {
2903     if (!vu_has_protocol_feature(dev, VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)) {
2904         return 0;
2905     }
2906 
2907     if (unlikely(!vq->inflight)) {
2908         return -1;
2909     }
2910 
2911     barrier();
2912 
2913     vq->inflight->desc[desc_idx].inflight = 0;
2914 
2915     barrier();
2916 
2917     vq->inflight->used_idx = vq->used_idx;
2918 
2919     return 0;
2920 }
2921 
2922 void *
2923 vu_queue_pop(VuDev *dev, VuVirtq *vq, size_t sz)
2924 {
2925     int i;
2926     unsigned int head;
2927     VuVirtqElement *elem;
2928 
2929     if (!vu_is_vq_usable(dev, vq)) {
2930         return NULL;
2931     }
2932 
2933     if (unlikely(vq->resubmit_list && vq->resubmit_num > 0)) {
2934         i = (--vq->resubmit_num);
2935         elem = vu_queue_map_desc(dev, vq, vq->resubmit_list[i].index, sz);
2936 
2937         if (!vq->resubmit_num) {
2938             free(vq->resubmit_list);
2939             vq->resubmit_list = NULL;
2940         }
2941 
2942         return elem;
2943     }
2944 
2945     if (vu_queue_empty(dev, vq)) {
2946         return NULL;
2947     }
2948     /*
2949      * Needed after virtio_queue_empty(), see comment in
2950      * virtqueue_num_heads().
2951      */
2952     smp_rmb();
2953 
2954     if (vq->inuse >= vq->vring.num) {
2955         vu_panic(dev, "Virtqueue size exceeded");
2956         return NULL;
2957     }
2958 
2959     if (!virtqueue_get_head(dev, vq, vq->last_avail_idx++, &head)) {
2960         return NULL;
2961     }
2962 
2963     if (vu_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
2964         vring_set_avail_event(vq, vq->last_avail_idx);
2965     }
2966 
2967     elem = vu_queue_map_desc(dev, vq, head, sz);
2968 
2969     if (!elem) {
2970         return NULL;
2971     }
2972 
2973     vq->inuse++;
2974 
2975     vu_queue_inflight_get(dev, vq, head);
2976 
2977     return elem;
2978 }
2979 
2980 static void
2981 vu_queue_detach_element(VuDev *dev, VuVirtq *vq, VuVirtqElement *elem,
2982                         size_t len)
2983 {
2984     vq->inuse--;
2985     /* unmap, when DMA support is added */
2986 }
2987 
2988 void
2989 vu_queue_unpop(VuDev *dev, VuVirtq *vq, VuVirtqElement *elem,
2990                size_t len)
2991 {
2992     vq->last_avail_idx--;
2993     vu_queue_detach_element(dev, vq, elem, len);
2994 }
2995 
2996 bool
2997 vu_queue_rewind(VuDev *dev, VuVirtq *vq, unsigned int num)
2998 {
2999     if (num > vq->inuse) {
3000         return false;
3001     }
3002     vq->last_avail_idx -= num;
3003     vq->inuse -= num;
3004     return true;
3005 }
3006 
3007 static inline
3008 void vring_used_write(VuDev *dev, VuVirtq *vq,
3009                       struct vring_used_elem *uelem, int i)
3010 {
3011     struct vring_used *used = vq->vring.used;
3012 
3013     used->ring[i] = *uelem;
3014     vu_log_write(dev, vq->vring.log_guest_addr +
3015                  offsetof(struct vring_used, ring[i]),
3016                  sizeof(used->ring[i]));
3017 }
3018 
3019 
3020 static void
3021 vu_log_queue_fill(VuDev *dev, VuVirtq *vq,
3022                   const VuVirtqElement *elem,
3023                   unsigned int len)
3024 {
3025     struct vring_desc *desc = vq->vring.desc;
3026     unsigned int i, max, min, desc_len;
3027     uint64_t desc_addr, read_len;
3028     struct vring_desc desc_buf[VIRTQUEUE_MAX_SIZE];
3029     unsigned num_bufs = 0;
3030 
3031     max = vq->vring.num;
3032     i = elem->index;
3033 
3034     if (le16toh(desc[i].flags) & VRING_DESC_F_INDIRECT) {
3035         if (le32toh(desc[i].len) % sizeof(struct vring_desc)) {
3036             vu_panic(dev, "Invalid size for indirect buffer table");
3037             return;
3038         }
3039 
3040         /* loop over the indirect descriptor table */
3041         desc_addr = le64toh(desc[i].addr);
3042         desc_len = le32toh(desc[i].len);
3043         max = desc_len / sizeof(struct vring_desc);
3044         read_len = desc_len;
3045         desc = vu_gpa_to_va(dev, &read_len, desc_addr);
3046         if (unlikely(desc && read_len != desc_len)) {
3047             /* Failed to use zero copy */
3048             desc = NULL;
3049             if (!virtqueue_read_indirect_desc(dev, desc_buf,
3050                                               desc_addr,
3051                                               desc_len)) {
3052                 desc = desc_buf;
3053             }
3054         }
3055         if (!desc) {
3056             vu_panic(dev, "Invalid indirect buffer table");
3057             return;
3058         }
3059         i = 0;
3060     }
3061 
3062     do {
3063         if (++num_bufs > max) {
3064             vu_panic(dev, "Looped descriptor");
3065             return;
3066         }
3067 
3068         if (le16toh(desc[i].flags) & VRING_DESC_F_WRITE) {
3069             min = MIN(le32toh(desc[i].len), len);
3070             vu_log_write(dev, le64toh(desc[i].addr), min);
3071             len -= min;
3072         }
3073 
3074     } while (len > 0 &&
3075              (virtqueue_read_next_desc(dev, desc, i, max, &i)
3076               == VIRTQUEUE_READ_DESC_MORE));
3077 }
3078 
3079 void
3080 vu_queue_fill(VuDev *dev, VuVirtq *vq,
3081               const VuVirtqElement *elem,
3082               unsigned int len, unsigned int idx)
3083 {
3084     struct vring_used_elem uelem;
3085 
3086     if (!vu_is_vq_usable(dev, vq)) {
3087         return;
3088     }
3089 
3090     vu_log_queue_fill(dev, vq, elem, len);
3091 
3092     idx = (idx + vq->used_idx) % vq->vring.num;
3093 
3094     uelem.id = htole32(elem->index);
3095     uelem.len = htole32(len);
3096     vring_used_write(dev, vq, &uelem, idx);
3097 }
3098 
3099 static inline
3100 void vring_used_idx_set(VuDev *dev, VuVirtq *vq, uint16_t val)
3101 {
3102     vq->vring.used->idx = htole16(val);
3103     vu_log_write(dev,
3104                  vq->vring.log_guest_addr + offsetof(struct vring_used, idx),
3105                  sizeof(vq->vring.used->idx));
3106 
3107     vq->used_idx = val;
3108 }
3109 
3110 void
3111 vu_queue_flush(VuDev *dev, VuVirtq *vq, unsigned int count)
3112 {
3113     uint16_t old, new;
3114 
3115     if (!vu_is_vq_usable(dev, vq)) {
3116         return;
3117     }
3118 
3119     /* Make sure buffer is written before we update index. */
3120     smp_wmb();
3121 
3122     old = vq->used_idx;
3123     new = old + count;
3124     vring_used_idx_set(dev, vq, new);
3125     vq->inuse -= count;
3126     if (unlikely((int16_t)(new - vq->signalled_used) < (uint16_t)(new - old))) {
3127         vq->signalled_used_valid = false;
3128     }
3129 }
3130 
3131 void
3132 vu_queue_push(VuDev *dev, VuVirtq *vq,
3133               const VuVirtqElement *elem, unsigned int len)
3134 {
3135     vu_queue_fill(dev, vq, elem, len, 0);
3136     vu_queue_inflight_pre_put(dev, vq, elem->index);
3137     vu_queue_flush(dev, vq, 1);
3138     vu_queue_inflight_post_put(dev, vq, elem->index);
3139 }
3140