hw/virtio/dataplane/vring.c

   1 /* Copyright 2012 Red Hat, Inc.
   2  * Copyright IBM, Corp. 2012
   3  *
   4  * Based on Linux 2.6.39 vhost code:
   5  * Copyright (C) 2009 Red Hat, Inc.
   6  * Copyright (C) 2006 Rusty Russell IBM Corporation
   7  *
   8  * Author: Michael S. Tsirkin <mst@redhat.com>
   9  *         Stefan Hajnoczi <stefanha@redhat.com>
  10  *
  11  * Inspiration, some code, and most witty comments come from
  12  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
  13  *
  14  * This work is licensed under the terms of the GNU GPL, version 2.
  15  */
  16
  17 #include "trace.h"
  18 #include "hw/hw.h"
  19 #include "exec/memory.h"
  20 #include "exec/address-spaces.h"
  21 #include "hw/virtio/virtio-access.h"
  22 #include "hw/virtio/dataplane/vring.h"
  23 #include "hw/virtio/dataplane/vring-accessors.h"
  24 #include "qemu/error-report.h"
  25
  26 /* vring_map can be coupled with vring_unmap or (if you still have the
  27  * value returned in *mr) memory_region_unref.
  28  */
  29 static void *vring_map(MemoryRegion **mr, hwaddr phys, hwaddr len,
  30                        bool is_write)
  31 {
  32     MemoryRegionSection section = memory_region_find(get_system_memory(), phys, len);
  33
  34     if (!section.mr || int128_get64(section.size) < len) {
  35         goto out;
  36     }
  37     if (is_write && section.readonly) {
  38         goto out;
  39     }
  40     if (!memory_region_is_ram(section.mr)) {
  41         goto out;
  42     }
  43
  44     /* Ignore regions with dirty logging, we cannot mark them dirty */
  45     if (memory_region_get_dirty_log_mask(section.mr)) {
  46         goto out;
  47     }
  48
  49     *mr = section.mr;
  50     return memory_region_get_ram_ptr(section.mr) + section.offset_within_region;
  51
  52 out:
  53     memory_region_unref(section.mr);
  54     *mr = NULL;
  55     return NULL;
  56 }
  57
  58 static void vring_unmap(void *buffer, bool is_write)
  59 {
  60     ram_addr_t addr;
  61     MemoryRegion *mr;
  62
  63     mr = qemu_ram_addr_from_host(buffer, &addr);
  64     memory_region_unref(mr);
  65 }
  66
  67 /* Map the guest's vring to host memory */
  68 bool vring_setup(Vring *vring, VirtIODevice *vdev, int n)
  69 {
  70     hwaddr vring_addr = virtio_queue_get_ring_addr(vdev, n);
  71     hwaddr vring_size = virtio_queue_get_ring_size(vdev, n);
  72     void *vring_ptr;
  73
  74     vring->broken = false;
  75
  76     vring_ptr = vring_map(&vring->mr, vring_addr, vring_size, true);
  77     if (!vring_ptr) {
  78         error_report("Failed to map vring "
  79                      "addr %#" HWADDR_PRIx " size %" HWADDR_PRIu,
  80                      vring_addr, vring_size);
  81         vring->broken = true;
  82         return false;
  83     }
  84
  85     vring_init(&vring->vr, virtio_queue_get_num(vdev, n), vring_ptr, 4096);
  86
  87     vring->last_avail_idx = virtio_queue_get_last_avail_idx(vdev, n);
  88     vring->last_used_idx = vring_get_used_idx(vdev, vring);
  89     vring->signalled_used = 0;
  90     vring->signalled_used_valid = false;
  91
  92     trace_vring_setup(virtio_queue_get_ring_addr(vdev, n),
  93                       vring->vr.desc, vring->vr.avail, vring->vr.used);
  94     return true;
  95 }
  96
  97 void vring_teardown(Vring *vring, VirtIODevice *vdev, int n)
  98 {
  99     virtio_queue_set_last_avail_idx(vdev, n, vring->last_avail_idx);
 100     virtio_queue_invalidate_signalled_used(vdev, n);
 101
 102     memory_region_unref(vring->mr);
 103 }
 104
 105 /* Disable guest->host notifies */
 106 void vring_disable_notification(VirtIODevice *vdev, Vring *vring)
 107 {
 108     if (!virtio_vdev_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX)) {
 109         vring_set_used_flags(vdev, vring, VRING_USED_F_NO_NOTIFY);
 110     }
 111 }
 112
 113 /* Enable guest->host notifies
 114  *
 115  * Return true if the vring is empty, false if there are more requests.
 116  */
 117 bool vring_enable_notification(VirtIODevice *vdev, Vring *vring)
 118 {
 119     if (virtio_vdev_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX)) {
 120         vring_avail_event(&vring->vr) = vring->vr.avail->idx;
 121     } else {
 122         vring_clear_used_flags(vdev, vring, VRING_USED_F_NO_NOTIFY);
 123     }
 124     smp_mb(); /* ensure update is seen before reading avail_idx */
 125     return !vring_more_avail(vdev, vring);
 126 }
 127
 128 /* This is stolen from linux/drivers/vhost/vhost.c:vhost_notify() */
 129 bool vring_should_notify(VirtIODevice *vdev, Vring *vring)
 130 {
 131     uint16_t old, new;
 132     bool v;
 133     /* Flush out used index updates. This is paired
 134      * with the barrier that the Guest executes when enabling
 135      * interrupts. */
 136     smp_mb();
 137
 138     if (virtio_vdev_has_feature(vdev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
 139         unlikely(!vring_more_avail(vdev, vring))) {
 140         return true;
 141     }
 142
 143     if (!virtio_vdev_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX)) {
 144         return !(vring_get_avail_flags(vdev, vring) &
 145                  VRING_AVAIL_F_NO_INTERRUPT);
 146     }
 147     old = vring->signalled_used;
 148     v = vring->signalled_used_valid;
 149     new = vring->signalled_used = vring->last_used_idx;
 150     vring->signalled_used_valid = true;
 151
 152     if (unlikely(!v)) {
 153         return true;
 154     }
 155
 156     return vring_need_event(virtio_tswap16(vdev, vring_used_event(&vring->vr)),
 157                             new, old);
 158 }
 159
 160
 161 static int get_desc(Vring *vring, VirtQueueElement *elem,
 162                     struct vring_desc *desc)
 163 {
 164     unsigned *num;
 165     struct iovec *iov;
 166     hwaddr *addr;
 167     MemoryRegion *mr;
 168
 169     if (desc->flags & VRING_DESC_F_WRITE) {
 170         num = &elem->in_num;
 171         iov = &elem->in_sg[*num];
 172         addr = &elem->in_addr[*num];
 173     } else {
 174         num = &elem->out_num;
 175         iov = &elem->out_sg[*num];
 176         addr = &elem->out_addr[*num];
 177
 178         /* If it's an output descriptor, they're all supposed
 179          * to come before any input descriptors. */
 180         if (unlikely(elem->in_num)) {
 181             error_report("Descriptor has out after in");
 182             return -EFAULT;
 183         }
 184     }
 185
 186     /* Stop for now if there are not enough iovecs available. */
 187     if (*num >= VIRTQUEUE_MAX_SIZE) {
 188         error_report("Invalid SG num: %u", *num);
 189         return -EFAULT;
 190     }
 191
 192     /* TODO handle non-contiguous memory across region boundaries */
 193     iov->iov_base = vring_map(&mr, desc->addr, desc->len,
 194                               desc->flags & VRING_DESC_F_WRITE);
 195     if (!iov->iov_base) {
 196         error_report("Failed to map descriptor addr %#" PRIx64 " len %u",
 197                      (uint64_t)desc->addr, desc->len);
 198         return -EFAULT;
 199     }
 200
 201     /* The MemoryRegion is looked up again and unref'ed later, leave the
 202      * ref in place.  */
 203     iov->iov_len = desc->len;
 204     *addr = desc->addr;
 205     *num += 1;
 206     return 0;
 207 }
 208
 209 static void copy_in_vring_desc(VirtIODevice *vdev,
 210                                const struct vring_desc *guest,
 211                                struct vring_desc *host)
 212 {
 213     host->addr = virtio_ldq_p(vdev, &guest->addr);
 214     host->len = virtio_ldl_p(vdev, &guest->len);
 215     host->flags = virtio_lduw_p(vdev, &guest->flags);
 216     host->next = virtio_lduw_p(vdev, &guest->next);
 217 }
 218
 219 /* This is stolen from linux/drivers/vhost/vhost.c. */
 220 static int get_indirect(VirtIODevice *vdev, Vring *vring,
 221                         VirtQueueElement *elem, struct vring_desc *indirect)
 222 {
 223     struct vring_desc desc;
 224     unsigned int i = 0, count, found = 0;
 225     int ret;
 226
 227     /* Sanity check */
 228     if (unlikely(indirect->len % sizeof(desc))) {
 229         error_report("Invalid length in indirect descriptor: "
 230                      "len %#x not multiple of %#zx",
 231                      indirect->len, sizeof(desc));
 232         vring->broken = true;
 233         return -EFAULT;
 234     }
 235
 236     count = indirect->len / sizeof(desc);
 237     /* Buffers are chained via a 16 bit next field, so
 238      * we can have at most 2^16 of these. */
 239     if (unlikely(count > USHRT_MAX + 1)) {
 240         error_report("Indirect buffer length too big: %d", indirect->len);
 241         vring->broken = true;
 242         return -EFAULT;
 243     }
 244
 245     do {
 246         struct vring_desc *desc_ptr;
 247         MemoryRegion *mr;
 248
 249         /* Translate indirect descriptor */
 250         desc_ptr = vring_map(&mr,
 251                              indirect->addr + found * sizeof(desc),
 252                              sizeof(desc), false);
 253         if (!desc_ptr) {
 254             error_report("Failed to map indirect descriptor "
 255                          "addr %#" PRIx64 " len %zu",
 256                          (uint64_t)indirect->addr + found * sizeof(desc),
 257                          sizeof(desc));
 258             vring->broken = true;
 259             return -EFAULT;
 260         }
 261         copy_in_vring_desc(vdev, desc_ptr, &desc);
 262         memory_region_unref(mr);
 263
 264         /* Ensure descriptor has been loaded before accessing fields */
 265         barrier(); /* read_barrier_depends(); */
 266
 267         if (unlikely(++found > count)) {
 268             error_report("Loop detected: last one at %u "
 269                          "indirect size %u", i, count);
 270             vring->broken = true;
 271             return -EFAULT;
 272         }
 273
 274         if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
 275             error_report("Nested indirect descriptor");
 276             vring->broken = true;
 277             return -EFAULT;
 278         }
 279
 280         ret = get_desc(vring, elem, &desc);
 281         if (ret < 0) {
 282             vring->broken |= (ret == -EFAULT);
 283             return ret;
 284         }
 285         i = desc.next;
 286     } while (desc.flags & VRING_DESC_F_NEXT);
 287     return 0;
 288 }
 289
 290 static void vring_unmap_element(VirtQueueElement *elem)
 291 {
 292     int i;
 293
 294     /* This assumes that the iovecs, if changed, are never moved past
 295      * the end of the valid area.  This is true if iovec manipulations
 296      * are done with iov_discard_front and iov_discard_back.
 297      */
 298     for (i = 0; i < elem->out_num; i++) {
 299         vring_unmap(elem->out_sg[i].iov_base, false);
 300     }
 301
 302     for (i = 0; i < elem->in_num; i++) {
 303         vring_unmap(elem->in_sg[i].iov_base, true);
 304     }
 305 }
 306
 307 /* This looks in the virtqueue and for the first available buffer, and converts
 308  * it to an iovec for convenient access.  Since descriptors consist of some
 309  * number of output then some number of input descriptors, it's actually two
 310  * iovecs, but we pack them into one and note how many of each there were.
 311  *
 312  * This function returns the descriptor number found, or vq->num (which is
 313  * never a valid descriptor number) if none was found.  A negative code is
 314  * returned on error.
 315  *
 316  * Stolen from linux/drivers/vhost/vhost.c.
 317  */
 318 int vring_pop(VirtIODevice *vdev, Vring *vring,
 319               VirtQueueElement *elem)
 320 {
 321     struct vring_desc desc;
 322     unsigned int i, head, found = 0, num = vring->vr.num;
 323     uint16_t avail_idx, last_avail_idx;
 324     int ret;
 325
 326     /* Initialize elem so it can be safely unmapped */
 327     elem->in_num = elem->out_num = 0;
 328
 329     /* If there was a fatal error then refuse operation */
 330     if (vring->broken) {
 331         ret = -EFAULT;
 332         goto out;
 333     }
 334
 335     /* Check it isn't doing very strange things with descriptor numbers. */
 336     last_avail_idx = vring->last_avail_idx;
 337     avail_idx = vring_get_avail_idx(vdev, vring);
 338     barrier(); /* load indices now and not again later */
 339
 340     if (unlikely((uint16_t)(avail_idx - last_avail_idx) > num)) {
 341         error_report("Guest moved used index from %u to %u",
 342                      last_avail_idx, avail_idx);
 343         ret = -EFAULT;
 344         goto out;
 345     }
 346
 347     /* If there's nothing new since last we looked. */
 348     if (avail_idx == last_avail_idx) {
 349         ret = -EAGAIN;
 350         goto out;
 351     }
 352
 353     /* Only get avail ring entries after they have been exposed by guest. */
 354     smp_rmb();
 355
 356     /* Grab the next descriptor number they're advertising, and increment
 357      * the index we've seen. */
 358     head = vring_get_avail_ring(vdev, vring, last_avail_idx % num);
 359
 360     elem->index = head;
 361
 362     /* If their number is silly, that's an error. */
 363     if (unlikely(head >= num)) {
 364         error_report("Guest says index %u > %u is available", head, num);
 365         ret = -EFAULT;
 366         goto out;
 367     }
 368
 369     i = head;
 370     do {
 371         if (unlikely(i >= num)) {
 372             error_report("Desc index is %u > %u, head = %u", i, num, head);
 373             ret = -EFAULT;
 374             goto out;
 375         }
 376         if (unlikely(++found > num)) {
 377             error_report("Loop detected: last one at %u vq size %u head %u",
 378                          i, num, head);
 379             ret = -EFAULT;
 380             goto out;
 381         }
 382         copy_in_vring_desc(vdev, &vring->vr.desc[i], &desc);
 383
 384         /* Ensure descriptor is loaded before accessing fields */
 385         barrier();
 386
 387         if (desc.flags & VRING_DESC_F_INDIRECT) {
 388             ret = get_indirect(vdev, vring, elem, &desc);
 389             if (ret < 0) {
 390                 goto out;
 391             }
 392             continue;
 393         }
 394
 395         ret = get_desc(vring, elem, &desc);
 396         if (ret < 0) {
 397             goto out;
 398         }
 399
 400         i = desc.next;
 401     } while (desc.flags & VRING_DESC_F_NEXT);
 402
 403     /* On success, increment avail index. */
 404     vring->last_avail_idx++;
 405     if (virtio_vdev_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX)) {
 406         vring_avail_event(&vring->vr) =
 407             virtio_tswap16(vdev, vring->last_avail_idx);
 408     }
 409
 410     return head;
 411
 412 out:
 413     assert(ret < 0);
 414     if (ret == -EFAULT) {
 415         vring->broken = true;
 416     }
 417     vring_unmap_element(elem);
 418     return ret;
 419 }
 420
 421 /* After we've used one of their buffers, we tell them about it.
 422  *
 423  * Stolen from linux/drivers/vhost/vhost.c.
 424  */
 425 void vring_push(VirtIODevice *vdev, Vring *vring, VirtQueueElement *elem,
 426                 int len)
 427 {
 428     unsigned int head = elem->index;
 429     uint16_t new;
 430
 431     vring_unmap_element(elem);
 432
 433     /* Don't touch vring if a fatal error occurred */
 434     if (vring->broken) {
 435         return;
 436     }
 437
 438     /* The virtqueue contains a ring of used buffers.  Get a pointer to the
 439      * next entry in that used ring. */
 440     vring_set_used_ring_id(vdev, vring, vring->last_used_idx % vring->vr.num,
 441                            head);
 442     vring_set_used_ring_len(vdev, vring, vring->last_used_idx % vring->vr.num,
 443                             len);
 444
 445     /* Make sure buffer is written before we update index. */
 446     smp_wmb();
 447
 448     new = ++vring->last_used_idx;
 449     vring_set_used_idx(vdev, vring, new);
 450     if (unlikely((int16_t)(new - vring->signalled_used) < (uint16_t)1)) {
 451         vring->signalled_used_valid = false;
 452     }
 453 }