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