hw/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 "vring.h"
  19 #include "qemu/error-report.h"
  20
  21 /* Map the guest's vring to host memory */
  22 bool vring_setup(Vring *vring, VirtIODevice *vdev, int n)
  23 {
  24     hwaddr vring_addr = virtio_queue_get_ring_addr(vdev, n);
  25     hwaddr vring_size = virtio_queue_get_ring_size(vdev, n);
  26     void *vring_ptr;
  27
  28     vring->broken = false;
  29
  30     hostmem_init(&vring->hostmem);
  31     vring_ptr = hostmem_lookup(&vring->hostmem, vring_addr, vring_size, true);
  32     if (!vring_ptr) {
  33         error_report("Failed to map vring "
  34                      "addr %#" HWADDR_PRIx " size %" HWADDR_PRIu,
  35                      vring_addr, vring_size);
  36         vring->broken = true;
  37         return false;
  38     }
  39
  40     vring_init(&vring->vr, virtio_queue_get_num(vdev, n), vring_ptr, 4096);
  41
  42     vring->last_avail_idx = 0;
  43     vring->last_used_idx = 0;
  44     vring->signalled_used = 0;
  45     vring->signalled_used_valid = false;
  46
  47     trace_vring_setup(virtio_queue_get_ring_addr(vdev, n),
  48                       vring->vr.desc, vring->vr.avail, vring->vr.used);
  49     return true;
  50 }
  51
  52 void vring_teardown(Vring *vring)
  53 {
  54     hostmem_finalize(&vring->hostmem);
  55 }
  56
  57 /* Disable guest->host notifies */
  58 void vring_disable_notification(VirtIODevice *vdev, Vring *vring)
  59 {
  60     if (!(vdev->guest_features & (1 << VIRTIO_RING_F_EVENT_IDX))) {
  61         vring->vr.used->flags |= VRING_USED_F_NO_NOTIFY;
  62     }
  63 }
  64
  65 /* Enable guest->host notifies
  66  *
  67  * Return true if the vring is empty, false if there are more requests.
  68  */
  69 bool vring_enable_notification(VirtIODevice *vdev, Vring *vring)
  70 {
  71     if (vdev->guest_features & (1 << VIRTIO_RING_F_EVENT_IDX)) {
  72         vring_avail_event(&vring->vr) = vring->vr.avail->idx;
  73     } else {
  74         vring->vr.used->flags &= ~VRING_USED_F_NO_NOTIFY;
  75     }
  76     smp_mb(); /* ensure update is seen before reading avail_idx */
  77     return !vring_more_avail(vring);
  78 }
  79
  80 /* This is stolen from linux/drivers/vhost/vhost.c:vhost_notify() */
  81 bool vring_should_notify(VirtIODevice *vdev, Vring *vring)
  82 {
  83     uint16_t old, new;
  84     bool v;
  85     /* Flush out used index updates. This is paired
  86      * with the barrier that the Guest executes when enabling
  87      * interrupts. */
  88     smp_mb();
  89
  90     if ((vdev->guest_features & VIRTIO_F_NOTIFY_ON_EMPTY) &&
  91         unlikely(vring->vr.avail->idx == vring->last_avail_idx)) {
  92         return true;
  93     }
  94
  95     if (!(vdev->guest_features & VIRTIO_RING_F_EVENT_IDX)) {
  96         return !(vring->vr.avail->flags & VRING_AVAIL_F_NO_INTERRUPT);
  97     }
  98     old = vring->signalled_used;
  99     v = vring->signalled_used_valid;
 100     new = vring->signalled_used = vring->last_used_idx;
 101     vring->signalled_used_valid = true;
 102
 103     if (unlikely(!v)) {
 104         return true;
 105     }
 106
 107     return vring_need_event(vring_used_event(&vring->vr), new, old);
 108 }
 109
 110 /* This is stolen from linux/drivers/vhost/vhost.c. */
 111 static int get_indirect(Vring *vring,
 112                         struct iovec iov[], struct iovec *iov_end,
 113                         unsigned int *out_num, unsigned int *in_num,
 114                         struct vring_desc *indirect)
 115 {
 116     struct vring_desc desc;
 117     unsigned int i = 0, count, found = 0;
 118
 119     /* Sanity check */
 120     if (unlikely(indirect->len % sizeof(desc))) {
 121         error_report("Invalid length in indirect descriptor: "
 122                      "len %#x not multiple of %#zx",
 123                      indirect->len, sizeof(desc));
 124         vring->broken = true;
 125         return -EFAULT;
 126     }
 127
 128     count = indirect->len / sizeof(desc);
 129     /* Buffers are chained via a 16 bit next field, so
 130      * we can have at most 2^16 of these. */
 131     if (unlikely(count > USHRT_MAX + 1)) {
 132         error_report("Indirect buffer length too big: %d", indirect->len);
 133         vring->broken = true;
 134         return -EFAULT;
 135     }
 136
 137     do {
 138         struct vring_desc *desc_ptr;
 139
 140         /* Translate indirect descriptor */
 141         desc_ptr = hostmem_lookup(&vring->hostmem,
 142                                   indirect->addr + found * sizeof(desc),
 143                                   sizeof(desc), false);
 144         if (!desc_ptr) {
 145             error_report("Failed to map indirect descriptor "
 146                          "addr %#" PRIx64 " len %zu",
 147                          (uint64_t)indirect->addr + found * sizeof(desc),
 148                          sizeof(desc));
 149             vring->broken = true;
 150             return -EFAULT;
 151         }
 152         desc = *desc_ptr;
 153
 154         /* Ensure descriptor has been loaded before accessing fields */
 155         barrier(); /* read_barrier_depends(); */
 156
 157         if (unlikely(++found > count)) {
 158             error_report("Loop detected: last one at %u "
 159                          "indirect size %u", i, count);
 160             vring->broken = true;
 161             return -EFAULT;
 162         }
 163
 164         if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
 165             error_report("Nested indirect descriptor");
 166             vring->broken = true;
 167             return -EFAULT;
 168         }
 169
 170         /* Stop for now if there are not enough iovecs available. */
 171         if (iov >= iov_end) {
 172             return -ENOBUFS;
 173         }
 174
 175         iov->iov_base = hostmem_lookup(&vring->hostmem, desc.addr, desc.len,
 176                                        desc.flags & VRING_DESC_F_WRITE);
 177         if (!iov->iov_base) {
 178             error_report("Failed to map indirect descriptor"
 179                          "addr %#" PRIx64 " len %u",
 180                          (uint64_t)desc.addr, desc.len);
 181             vring->broken = true;
 182             return -EFAULT;
 183         }
 184         iov->iov_len = desc.len;
 185         iov++;
 186
 187         /* If this is an input descriptor, increment that count. */
 188         if (desc.flags & VRING_DESC_F_WRITE) {
 189             *in_num += 1;
 190         } else {
 191             /* If it's an output descriptor, they're all supposed
 192              * to come before any input descriptors. */
 193             if (unlikely(*in_num)) {
 194                 error_report("Indirect descriptor "
 195                              "has out after in: idx %u", i);
 196                 vring->broken = true;
 197                 return -EFAULT;
 198             }
 199             *out_num += 1;
 200         }
 201         i = desc.next;
 202     } while (desc.flags & VRING_DESC_F_NEXT);
 203     return 0;
 204 }
 205
 206 /* This looks in the virtqueue and for the first available buffer, and converts
 207  * it to an iovec for convenient access.  Since descriptors consist of some
 208  * number of output then some number of input descriptors, it's actually two
 209  * iovecs, but we pack them into one and note how many of each there were.
 210  *
 211  * This function returns the descriptor number found, or vq->num (which is
 212  * never a valid descriptor number) if none was found.  A negative code is
 213  * returned on error.
 214  *
 215  * Stolen from linux/drivers/vhost/vhost.c.
 216  */
 217 int vring_pop(VirtIODevice *vdev, Vring *vring,
 218               struct iovec iov[], struct iovec *iov_end,
 219               unsigned int *out_num, unsigned int *in_num)
 220 {
 221     struct vring_desc desc;
 222     unsigned int i, head, found = 0, num = vring->vr.num;
 223     uint16_t avail_idx, last_avail_idx;
 224
 225     /* If there was a fatal error then refuse operation */
 226     if (vring->broken) {
 227         return -EFAULT;
 228     }
 229
 230     /* Check it isn't doing very strange things with descriptor numbers. */
 231     last_avail_idx = vring->last_avail_idx;
 232     avail_idx = vring->vr.avail->idx;
 233     barrier(); /* load indices now and not again later */
 234
 235     if (unlikely((uint16_t)(avail_idx - last_avail_idx) > num)) {
 236         error_report("Guest moved used index from %u to %u",
 237                      last_avail_idx, avail_idx);
 238         vring->broken = true;
 239         return -EFAULT;
 240     }
 241
 242     /* If there's nothing new since last we looked. */
 243     if (avail_idx == last_avail_idx) {
 244         return -EAGAIN;
 245     }
 246
 247     /* Only get avail ring entries after they have been exposed by guest. */
 248     smp_rmb();
 249
 250     /* Grab the next descriptor number they're advertising, and increment
 251      * the index we've seen. */
 252     head = vring->vr.avail->ring[last_avail_idx % num];
 253
 254     /* If their number is silly, that's an error. */
 255     if (unlikely(head >= num)) {
 256         error_report("Guest says index %u > %u is available", head, num);
 257         vring->broken = true;
 258         return -EFAULT;
 259     }
 260
 261     if (vdev->guest_features & (1 << VIRTIO_RING_F_EVENT_IDX)) {
 262         vring_avail_event(&vring->vr) = vring->vr.avail->idx;
 263     }
 264
 265     /* When we start there are none of either input nor output. */
 266     *out_num = *in_num = 0;
 267
 268     i = head;
 269     do {
 270         if (unlikely(i >= num)) {
 271             error_report("Desc index is %u > %u, head = %u", i, num, head);
 272             vring->broken = true;
 273             return -EFAULT;
 274         }
 275         if (unlikely(++found > num)) {
 276             error_report("Loop detected: last one at %u vq size %u head %u",
 277                          i, num, head);
 278             vring->broken = true;
 279             return -EFAULT;
 280         }
 281         desc = vring->vr.desc[i];
 282
 283         /* Ensure descriptor is loaded before accessing fields */
 284         barrier();
 285
 286         if (desc.flags & VRING_DESC_F_INDIRECT) {
 287             int ret = get_indirect(vring, iov, iov_end, out_num, in_num, &desc);
 288             if (ret < 0) {
 289                 return ret;
 290             }
 291             continue;
 292         }
 293
 294         /* If there are not enough iovecs left, stop for now.  The caller
 295          * should check if there are more descs available once they have dealt
 296          * with the current set.
 297          */
 298         if (iov >= iov_end) {
 299             return -ENOBUFS;
 300         }
 301
 302         /* TODO handle non-contiguous memory across region boundaries */
 303         iov->iov_base = hostmem_lookup(&vring->hostmem, desc.addr, desc.len,
 304                                        desc.flags & VRING_DESC_F_WRITE);
 305         if (!iov->iov_base) {
 306             error_report("Failed to map vring desc addr %#" PRIx64 " len %u",
 307                          (uint64_t)desc.addr, desc.len);
 308             vring->broken = true;
 309             return -EFAULT;
 310         }
 311         iov->iov_len  = desc.len;
 312         iov++;
 313
 314         if (desc.flags & VRING_DESC_F_WRITE) {
 315             /* If this is an input descriptor,
 316              * increment that count. */
 317             *in_num += 1;
 318         } else {
 319             /* If it's an output descriptor, they're all supposed
 320              * to come before any input descriptors. */
 321             if (unlikely(*in_num)) {
 322                 error_report("Descriptor has out after in: idx %d", i);
 323                 vring->broken = true;
 324                 return -EFAULT;
 325             }
 326             *out_num += 1;
 327         }
 328         i = desc.next;
 329     } while (desc.flags & VRING_DESC_F_NEXT);
 330
 331     /* On success, increment avail index. */
 332     vring->last_avail_idx++;
 333     return head;
 334 }
 335
 336 /* After we've used one of their buffers, we tell them about it.
 337  *
 338  * Stolen from linux/drivers/vhost/vhost.c.
 339  */
 340 void vring_push(Vring *vring, unsigned int head, int len)
 341 {
 342     struct vring_used_elem *used;
 343     uint16_t new;
 344
 345     /* Don't touch vring if a fatal error occurred */
 346     if (vring->broken) {
 347         return;
 348     }
 349
 350     /* The virtqueue contains a ring of used buffers.  Get a pointer to the
 351      * next entry in that used ring. */
 352     used = &vring->vr.used->ring[vring->last_used_idx % vring->vr.num];
 353     used->id = head;
 354     used->len = len;
 355
 356     /* Make sure buffer is written before we update index. */
 357     smp_wmb();
 358
 359     new = vring->vr.used->idx = ++vring->last_used_idx;
 360     if (unlikely((int16_t)(new - vring->signalled_used) < (uint16_t)1)) {
 361         vring->signalled_used_valid = false;
 362     }
 363 }