109c434b8SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 2f87d0fbbSRusty Russell /* 3f87d0fbbSRusty Russell * Helpers for the host side of a virtio ring. 4f87d0fbbSRusty Russell * 5f87d0fbbSRusty Russell * Since these may be in userspace, we use (inline) accessors. 6f87d0fbbSRusty Russell */ 79d1b972fSMark Rutland #include <linux/compiler.h> 8f558a845SDave Jones #include <linux/module.h> 9f87d0fbbSRusty Russell #include <linux/vringh.h> 10f87d0fbbSRusty Russell #include <linux/virtio_ring.h> 11f87d0fbbSRusty Russell #include <linux/kernel.h> 12f87d0fbbSRusty Russell #include <linux/ratelimit.h> 13f87d0fbbSRusty Russell #include <linux/uaccess.h> 14f87d0fbbSRusty Russell #include <linux/slab.h> 15f87d0fbbSRusty Russell #include <linux/export.h> 163302363aSMichael S. Tsirkin #if IS_REACHABLE(CONFIG_VHOST_IOTLB) 179ad9c49cSJason Wang #include <linux/bvec.h> 189ad9c49cSJason Wang #include <linux/highmem.h> 199ad9c49cSJason Wang #include <linux/vhost_iotlb.h> 203302363aSMichael S. Tsirkin #endif 21b9f7ac8cSMichael S. Tsirkin #include <uapi/linux/virtio_config.h> 22f87d0fbbSRusty Russell 23f87d0fbbSRusty Russell static __printf(1,2) __cold void vringh_bad(const char *fmt, ...) 24f87d0fbbSRusty Russell { 25f87d0fbbSRusty Russell static DEFINE_RATELIMIT_STATE(vringh_rs, 26f87d0fbbSRusty Russell DEFAULT_RATELIMIT_INTERVAL, 27f87d0fbbSRusty Russell DEFAULT_RATELIMIT_BURST); 28f87d0fbbSRusty Russell if (__ratelimit(&vringh_rs)) { 29f87d0fbbSRusty Russell va_list ap; 30f87d0fbbSRusty Russell va_start(ap, fmt); 31f87d0fbbSRusty Russell printk(KERN_NOTICE "vringh:"); 32f87d0fbbSRusty Russell vprintk(fmt, ap); 33f87d0fbbSRusty Russell va_end(ap); 34f87d0fbbSRusty Russell } 35f87d0fbbSRusty Russell } 36f87d0fbbSRusty Russell 37f87d0fbbSRusty Russell /* Returns vring->num if empty, -ve on error. */ 38f87d0fbbSRusty Russell static inline int __vringh_get_head(const struct vringh *vrh, 39b9f7ac8cSMichael S. Tsirkin int (*getu16)(const struct vringh *vrh, 40b9f7ac8cSMichael S. Tsirkin u16 *val, const __virtio16 *p), 41f87d0fbbSRusty Russell u16 *last_avail_idx) 42f87d0fbbSRusty Russell { 43f87d0fbbSRusty Russell u16 avail_idx, i, head; 44f87d0fbbSRusty Russell int err; 45f87d0fbbSRusty Russell 46b9f7ac8cSMichael S. Tsirkin err = getu16(vrh, &avail_idx, &vrh->vring.avail->idx); 47f87d0fbbSRusty Russell if (err) { 48f87d0fbbSRusty Russell vringh_bad("Failed to access avail idx at %p", 49f87d0fbbSRusty Russell &vrh->vring.avail->idx); 50f87d0fbbSRusty Russell return err; 51f87d0fbbSRusty Russell } 52f87d0fbbSRusty Russell 53f87d0fbbSRusty Russell if (*last_avail_idx == avail_idx) 54f87d0fbbSRusty Russell return vrh->vring.num; 55f87d0fbbSRusty Russell 56f87d0fbbSRusty Russell /* Only get avail ring entries after they have been exposed by guest. */ 57f87d0fbbSRusty Russell virtio_rmb(vrh->weak_barriers); 58f87d0fbbSRusty Russell 59f87d0fbbSRusty Russell i = *last_avail_idx & (vrh->vring.num - 1); 60f87d0fbbSRusty Russell 61b9f7ac8cSMichael S. Tsirkin err = getu16(vrh, &head, &vrh->vring.avail->ring[i]); 62f87d0fbbSRusty Russell if (err) { 63f87d0fbbSRusty Russell vringh_bad("Failed to read head: idx %d address %p", 64f87d0fbbSRusty Russell *last_avail_idx, &vrh->vring.avail->ring[i]); 65f87d0fbbSRusty Russell return err; 66f87d0fbbSRusty Russell } 67f87d0fbbSRusty Russell 68f87d0fbbSRusty Russell if (head >= vrh->vring.num) { 69f87d0fbbSRusty Russell vringh_bad("Guest says index %u > %u is available", 70f87d0fbbSRusty Russell head, vrh->vring.num); 71f87d0fbbSRusty Russell return -EINVAL; 72f87d0fbbSRusty Russell } 73f87d0fbbSRusty Russell 74f87d0fbbSRusty Russell (*last_avail_idx)++; 75f87d0fbbSRusty Russell return head; 76f87d0fbbSRusty Russell } 77f87d0fbbSRusty Russell 78f87d0fbbSRusty Russell /* Copy some bytes to/from the iovec. Returns num copied. */ 799ad9c49cSJason Wang static inline ssize_t vringh_iov_xfer(struct vringh *vrh, 809ad9c49cSJason Wang struct vringh_kiov *iov, 81f87d0fbbSRusty Russell void *ptr, size_t len, 829ad9c49cSJason Wang int (*xfer)(const struct vringh *vrh, 839ad9c49cSJason Wang void *addr, void *ptr, 84f87d0fbbSRusty Russell size_t len)) 85f87d0fbbSRusty Russell { 86f87d0fbbSRusty Russell int err, done = 0; 87f87d0fbbSRusty Russell 88f87d0fbbSRusty Russell while (len && iov->i < iov->used) { 89f87d0fbbSRusty Russell size_t partlen; 90f87d0fbbSRusty Russell 91f87d0fbbSRusty Russell partlen = min(iov->iov[iov->i].iov_len, len); 929ad9c49cSJason Wang err = xfer(vrh, iov->iov[iov->i].iov_base, ptr, partlen); 93f87d0fbbSRusty Russell if (err) 94f87d0fbbSRusty Russell return err; 95f87d0fbbSRusty Russell done += partlen; 96f87d0fbbSRusty Russell len -= partlen; 97f87d0fbbSRusty Russell ptr += partlen; 98f87d0fbbSRusty Russell iov->consumed += partlen; 99f87d0fbbSRusty Russell iov->iov[iov->i].iov_len -= partlen; 100f87d0fbbSRusty Russell iov->iov[iov->i].iov_base += partlen; 101f87d0fbbSRusty Russell 102f87d0fbbSRusty Russell if (!iov->iov[iov->i].iov_len) { 103f87d0fbbSRusty Russell /* Fix up old iov element then increment. */ 104f87d0fbbSRusty Russell iov->iov[iov->i].iov_len = iov->consumed; 105f87d0fbbSRusty Russell iov->iov[iov->i].iov_base -= iov->consumed; 106f87d0fbbSRusty Russell 1079ad9c49cSJason Wang 108f87d0fbbSRusty Russell iov->consumed = 0; 109f87d0fbbSRusty Russell iov->i++; 110f87d0fbbSRusty Russell } 111f87d0fbbSRusty Russell } 112f87d0fbbSRusty Russell return done; 113f87d0fbbSRusty Russell } 114f87d0fbbSRusty Russell 115f87d0fbbSRusty Russell /* May reduce *len if range is shorter. */ 116f87d0fbbSRusty Russell static inline bool range_check(struct vringh *vrh, u64 addr, size_t *len, 117f87d0fbbSRusty Russell struct vringh_range *range, 118f87d0fbbSRusty Russell bool (*getrange)(struct vringh *, 119f87d0fbbSRusty Russell u64, struct vringh_range *)) 120f87d0fbbSRusty Russell { 121f87d0fbbSRusty Russell if (addr < range->start || addr > range->end_incl) { 122f87d0fbbSRusty Russell if (!getrange(vrh, addr, range)) 123f87d0fbbSRusty Russell return false; 124f87d0fbbSRusty Russell } 125f87d0fbbSRusty Russell BUG_ON(addr < range->start || addr > range->end_incl); 126f87d0fbbSRusty Russell 127f87d0fbbSRusty Russell /* To end of memory? */ 128f87d0fbbSRusty Russell if (unlikely(addr + *len == 0)) { 129f87d0fbbSRusty Russell if (range->end_incl == -1ULL) 130f87d0fbbSRusty Russell return true; 131f87d0fbbSRusty Russell goto truncate; 132f87d0fbbSRusty Russell } 133f87d0fbbSRusty Russell 134f87d0fbbSRusty Russell /* Otherwise, don't wrap. */ 135f87d0fbbSRusty Russell if (addr + *len < addr) { 136f87d0fbbSRusty Russell vringh_bad("Wrapping descriptor %zu@0x%llx", 137f87d0fbbSRusty Russell *len, (unsigned long long)addr); 138f87d0fbbSRusty Russell return false; 139f87d0fbbSRusty Russell } 140f87d0fbbSRusty Russell 141f87d0fbbSRusty Russell if (unlikely(addr + *len - 1 > range->end_incl)) 142f87d0fbbSRusty Russell goto truncate; 143f87d0fbbSRusty Russell return true; 144f87d0fbbSRusty Russell 145f87d0fbbSRusty Russell truncate: 146f87d0fbbSRusty Russell *len = range->end_incl + 1 - addr; 147f87d0fbbSRusty Russell return true; 148f87d0fbbSRusty Russell } 149f87d0fbbSRusty Russell 150f87d0fbbSRusty Russell static inline bool no_range_check(struct vringh *vrh, u64 addr, size_t *len, 151f87d0fbbSRusty Russell struct vringh_range *range, 152f87d0fbbSRusty Russell bool (*getrange)(struct vringh *, 153f87d0fbbSRusty Russell u64, struct vringh_range *)) 154f87d0fbbSRusty Russell { 155f87d0fbbSRusty Russell return true; 156f87d0fbbSRusty Russell } 157f87d0fbbSRusty Russell 158f87d0fbbSRusty Russell /* No reason for this code to be inline. */ 159b9f7ac8cSMichael S. Tsirkin static int move_to_indirect(const struct vringh *vrh, 160b9f7ac8cSMichael S. Tsirkin int *up_next, u16 *i, void *addr, 161f87d0fbbSRusty Russell const struct vring_desc *desc, 162f87d0fbbSRusty Russell struct vring_desc **descs, int *desc_max) 163f87d0fbbSRusty Russell { 164b9f7ac8cSMichael S. Tsirkin u32 len; 165b9f7ac8cSMichael S. Tsirkin 166f87d0fbbSRusty Russell /* Indirect tables can't have indirect. */ 167f87d0fbbSRusty Russell if (*up_next != -1) { 168f87d0fbbSRusty Russell vringh_bad("Multilevel indirect %u->%u", *up_next, *i); 169f87d0fbbSRusty Russell return -EINVAL; 170f87d0fbbSRusty Russell } 171f87d0fbbSRusty Russell 172b9f7ac8cSMichael S. Tsirkin len = vringh32_to_cpu(vrh, desc->len); 173b9f7ac8cSMichael S. Tsirkin if (unlikely(len % sizeof(struct vring_desc))) { 174f87d0fbbSRusty Russell vringh_bad("Strange indirect len %u", desc->len); 175f87d0fbbSRusty Russell return -EINVAL; 176f87d0fbbSRusty Russell } 177f87d0fbbSRusty Russell 178f87d0fbbSRusty Russell /* We will check this when we follow it! */ 179b9f7ac8cSMichael S. Tsirkin if (desc->flags & cpu_to_vringh16(vrh, VRING_DESC_F_NEXT)) 180b9f7ac8cSMichael S. Tsirkin *up_next = vringh16_to_cpu(vrh, desc->next); 181f87d0fbbSRusty Russell else 182f87d0fbbSRusty Russell *up_next = -2; 183f87d0fbbSRusty Russell *descs = addr; 184b9f7ac8cSMichael S. Tsirkin *desc_max = len / sizeof(struct vring_desc); 185f87d0fbbSRusty Russell 186f87d0fbbSRusty Russell /* Now, start at the first indirect. */ 187f87d0fbbSRusty Russell *i = 0; 188f87d0fbbSRusty Russell return 0; 189f87d0fbbSRusty Russell } 190f87d0fbbSRusty Russell 191f87d0fbbSRusty Russell static int resize_iovec(struct vringh_kiov *iov, gfp_t gfp) 192f87d0fbbSRusty Russell { 193f87d0fbbSRusty Russell struct kvec *new; 194f87d0fbbSRusty Russell unsigned int flag, new_num = (iov->max_num & ~VRINGH_IOV_ALLOCATED) * 2; 195f87d0fbbSRusty Russell 196f87d0fbbSRusty Russell if (new_num < 8) 197f87d0fbbSRusty Russell new_num = 8; 198f87d0fbbSRusty Russell 199f87d0fbbSRusty Russell flag = (iov->max_num & VRINGH_IOV_ALLOCATED); 200f87d0fbbSRusty Russell if (flag) 2013a999748SBartosz Golaszewski new = krealloc_array(iov->iov, new_num, 2023a999748SBartosz Golaszewski sizeof(struct iovec), gfp); 203f87d0fbbSRusty Russell else { 2046da2ec56SKees Cook new = kmalloc_array(new_num, sizeof(struct iovec), gfp); 205f87d0fbbSRusty Russell if (new) { 206f87d0fbbSRusty Russell memcpy(new, iov->iov, 207f87d0fbbSRusty Russell iov->max_num * sizeof(struct iovec)); 208f87d0fbbSRusty Russell flag = VRINGH_IOV_ALLOCATED; 209f87d0fbbSRusty Russell } 210f87d0fbbSRusty Russell } 211f87d0fbbSRusty Russell if (!new) 212f87d0fbbSRusty Russell return -ENOMEM; 213f87d0fbbSRusty Russell iov->iov = new; 214f87d0fbbSRusty Russell iov->max_num = (new_num | flag); 215f87d0fbbSRusty Russell return 0; 216f87d0fbbSRusty Russell } 217f87d0fbbSRusty Russell 218f87d0fbbSRusty Russell static u16 __cold return_from_indirect(const struct vringh *vrh, int *up_next, 219f87d0fbbSRusty Russell struct vring_desc **descs, int *desc_max) 220f87d0fbbSRusty Russell { 221f87d0fbbSRusty Russell u16 i = *up_next; 222f87d0fbbSRusty Russell 223f87d0fbbSRusty Russell *up_next = -1; 224f87d0fbbSRusty Russell *descs = vrh->vring.desc; 225f87d0fbbSRusty Russell *desc_max = vrh->vring.num; 226f87d0fbbSRusty Russell return i; 227f87d0fbbSRusty Russell } 228f87d0fbbSRusty Russell 229f87d0fbbSRusty Russell static int slow_copy(struct vringh *vrh, void *dst, const void *src, 230f87d0fbbSRusty Russell bool (*rcheck)(struct vringh *vrh, u64 addr, size_t *len, 231f87d0fbbSRusty Russell struct vringh_range *range, 232f87d0fbbSRusty Russell bool (*getrange)(struct vringh *vrh, 233f87d0fbbSRusty Russell u64, 234f87d0fbbSRusty Russell struct vringh_range *)), 235f87d0fbbSRusty Russell bool (*getrange)(struct vringh *vrh, 236f87d0fbbSRusty Russell u64 addr, 237f87d0fbbSRusty Russell struct vringh_range *r), 238f87d0fbbSRusty Russell struct vringh_range *range, 2399ad9c49cSJason Wang int (*copy)(const struct vringh *vrh, 2409ad9c49cSJason Wang void *dst, const void *src, size_t len)) 241f87d0fbbSRusty Russell { 242f87d0fbbSRusty Russell size_t part, len = sizeof(struct vring_desc); 243f87d0fbbSRusty Russell 244f87d0fbbSRusty Russell do { 245f87d0fbbSRusty Russell u64 addr; 246f87d0fbbSRusty Russell int err; 247f87d0fbbSRusty Russell 248f87d0fbbSRusty Russell part = len; 249f87d0fbbSRusty Russell addr = (u64)(unsigned long)src - range->offset; 250f87d0fbbSRusty Russell 251f87d0fbbSRusty Russell if (!rcheck(vrh, addr, &part, range, getrange)) 252f87d0fbbSRusty Russell return -EINVAL; 253f87d0fbbSRusty Russell 2549ad9c49cSJason Wang err = copy(vrh, dst, src, part); 255f87d0fbbSRusty Russell if (err) 256f87d0fbbSRusty Russell return err; 257f87d0fbbSRusty Russell 258f87d0fbbSRusty Russell dst += part; 259f87d0fbbSRusty Russell src += part; 260f87d0fbbSRusty Russell len -= part; 261f87d0fbbSRusty Russell } while (len); 262f87d0fbbSRusty Russell return 0; 263f87d0fbbSRusty Russell } 264f87d0fbbSRusty Russell 265f87d0fbbSRusty Russell static inline int 266f87d0fbbSRusty Russell __vringh_iov(struct vringh *vrh, u16 i, 267f87d0fbbSRusty Russell struct vringh_kiov *riov, 268f87d0fbbSRusty Russell struct vringh_kiov *wiov, 269f87d0fbbSRusty Russell bool (*rcheck)(struct vringh *vrh, u64 addr, size_t *len, 270f87d0fbbSRusty Russell struct vringh_range *range, 271f87d0fbbSRusty Russell bool (*getrange)(struct vringh *, u64, 272f87d0fbbSRusty Russell struct vringh_range *)), 273f87d0fbbSRusty Russell bool (*getrange)(struct vringh *, u64, struct vringh_range *), 274f87d0fbbSRusty Russell gfp_t gfp, 2759ad9c49cSJason Wang int (*copy)(const struct vringh *vrh, 2769ad9c49cSJason Wang void *dst, const void *src, size_t len)) 277f87d0fbbSRusty Russell { 278f87d0fbbSRusty Russell int err, count = 0, up_next, desc_max; 279f87d0fbbSRusty Russell struct vring_desc desc, *descs; 280f87d0fbbSRusty Russell struct vringh_range range = { -1ULL, 0 }, slowrange; 281f87d0fbbSRusty Russell bool slow = false; 282f87d0fbbSRusty Russell 283f87d0fbbSRusty Russell /* We start traversing vring's descriptor table. */ 284f87d0fbbSRusty Russell descs = vrh->vring.desc; 285f87d0fbbSRusty Russell desc_max = vrh->vring.num; 286f87d0fbbSRusty Russell up_next = -1; 287f87d0fbbSRusty Russell 2885745bcfbSStefano Garzarella /* You must want something! */ 2895745bcfbSStefano Garzarella if (WARN_ON(!riov && !wiov)) 2905745bcfbSStefano Garzarella return -EINVAL; 2915745bcfbSStefano Garzarella 292f87d0fbbSRusty Russell if (riov) 293bbc2c372SStefano Garzarella riov->i = riov->used = riov->consumed = 0; 2945745bcfbSStefano Garzarella if (wiov) 295bbc2c372SStefano Garzarella wiov->i = wiov->used = wiov->consumed = 0; 296f87d0fbbSRusty Russell 297f87d0fbbSRusty Russell for (;;) { 298f87d0fbbSRusty Russell void *addr; 299f87d0fbbSRusty Russell struct vringh_kiov *iov; 300f87d0fbbSRusty Russell size_t len; 301f87d0fbbSRusty Russell 302f87d0fbbSRusty Russell if (unlikely(slow)) 303f87d0fbbSRusty Russell err = slow_copy(vrh, &desc, &descs[i], rcheck, getrange, 304f87d0fbbSRusty Russell &slowrange, copy); 305f87d0fbbSRusty Russell else 3069ad9c49cSJason Wang err = copy(vrh, &desc, &descs[i], sizeof(desc)); 307f87d0fbbSRusty Russell if (unlikely(err)) 308f87d0fbbSRusty Russell goto fail; 309f87d0fbbSRusty Russell 310b9f7ac8cSMichael S. Tsirkin if (unlikely(desc.flags & 311b9f7ac8cSMichael S. Tsirkin cpu_to_vringh16(vrh, VRING_DESC_F_INDIRECT))) { 312b9f7ac8cSMichael S. Tsirkin u64 a = vringh64_to_cpu(vrh, desc.addr); 313b9f7ac8cSMichael S. Tsirkin 314f87d0fbbSRusty Russell /* Make sure it's OK, and get offset. */ 315b9f7ac8cSMichael S. Tsirkin len = vringh32_to_cpu(vrh, desc.len); 316b9f7ac8cSMichael S. Tsirkin if (!rcheck(vrh, a, &len, &range, getrange)) { 317f87d0fbbSRusty Russell err = -EINVAL; 318f87d0fbbSRusty Russell goto fail; 319f87d0fbbSRusty Russell } 320f87d0fbbSRusty Russell 321b9f7ac8cSMichael S. Tsirkin if (unlikely(len != vringh32_to_cpu(vrh, desc.len))) { 322f87d0fbbSRusty Russell slow = true; 323f87d0fbbSRusty Russell /* We need to save this range to use offset */ 324f87d0fbbSRusty Russell slowrange = range; 325f87d0fbbSRusty Russell } 326f87d0fbbSRusty Russell 327b9f7ac8cSMichael S. Tsirkin addr = (void *)(long)(a + range.offset); 328b9f7ac8cSMichael S. Tsirkin err = move_to_indirect(vrh, &up_next, &i, addr, &desc, 329f87d0fbbSRusty Russell &descs, &desc_max); 330f87d0fbbSRusty Russell if (err) 331f87d0fbbSRusty Russell goto fail; 332f87d0fbbSRusty Russell continue; 333f87d0fbbSRusty Russell } 334f87d0fbbSRusty Russell 335f87d0fbbSRusty Russell if (count++ == vrh->vring.num) { 336f87d0fbbSRusty Russell vringh_bad("Descriptor loop in %p", descs); 337f87d0fbbSRusty Russell err = -ELOOP; 338f87d0fbbSRusty Russell goto fail; 339f87d0fbbSRusty Russell } 340f87d0fbbSRusty Russell 341b9f7ac8cSMichael S. Tsirkin if (desc.flags & cpu_to_vringh16(vrh, VRING_DESC_F_WRITE)) 342f87d0fbbSRusty Russell iov = wiov; 343f87d0fbbSRusty Russell else { 344f87d0fbbSRusty Russell iov = riov; 345f87d0fbbSRusty Russell if (unlikely(wiov && wiov->i)) { 346f87d0fbbSRusty Russell vringh_bad("Readable desc %p after writable", 347f87d0fbbSRusty Russell &descs[i]); 348f87d0fbbSRusty Russell err = -EINVAL; 349f87d0fbbSRusty Russell goto fail; 350f87d0fbbSRusty Russell } 351f87d0fbbSRusty Russell } 352f87d0fbbSRusty Russell 353f87d0fbbSRusty Russell if (!iov) { 354f87d0fbbSRusty Russell vringh_bad("Unexpected %s desc", 355f87d0fbbSRusty Russell !wiov ? "writable" : "readable"); 356f87d0fbbSRusty Russell err = -EPROTO; 357f87d0fbbSRusty Russell goto fail; 358f87d0fbbSRusty Russell } 359f87d0fbbSRusty Russell 360f87d0fbbSRusty Russell again: 361f87d0fbbSRusty Russell /* Make sure it's OK, and get offset. */ 362b9f7ac8cSMichael S. Tsirkin len = vringh32_to_cpu(vrh, desc.len); 363b9f7ac8cSMichael S. Tsirkin if (!rcheck(vrh, vringh64_to_cpu(vrh, desc.addr), &len, &range, 364b9f7ac8cSMichael S. Tsirkin getrange)) { 365f87d0fbbSRusty Russell err = -EINVAL; 366f87d0fbbSRusty Russell goto fail; 367f87d0fbbSRusty Russell } 368b9f7ac8cSMichael S. Tsirkin addr = (void *)(unsigned long)(vringh64_to_cpu(vrh, desc.addr) + 369b9f7ac8cSMichael S. Tsirkin range.offset); 370f87d0fbbSRusty Russell 371f87d0fbbSRusty Russell if (unlikely(iov->used == (iov->max_num & ~VRINGH_IOV_ALLOCATED))) { 372f87d0fbbSRusty Russell err = resize_iovec(iov, gfp); 373f87d0fbbSRusty Russell if (err) 374f87d0fbbSRusty Russell goto fail; 375f87d0fbbSRusty Russell } 376f87d0fbbSRusty Russell 377f87d0fbbSRusty Russell iov->iov[iov->used].iov_base = addr; 378f87d0fbbSRusty Russell iov->iov[iov->used].iov_len = len; 379f87d0fbbSRusty Russell iov->used++; 380f87d0fbbSRusty Russell 381b9f7ac8cSMichael S. Tsirkin if (unlikely(len != vringh32_to_cpu(vrh, desc.len))) { 382b9f7ac8cSMichael S. Tsirkin desc.len = cpu_to_vringh32(vrh, 383b9f7ac8cSMichael S. Tsirkin vringh32_to_cpu(vrh, desc.len) - len); 384b9f7ac8cSMichael S. Tsirkin desc.addr = cpu_to_vringh64(vrh, 385b9f7ac8cSMichael S. Tsirkin vringh64_to_cpu(vrh, desc.addr) + len); 386f87d0fbbSRusty Russell goto again; 387f87d0fbbSRusty Russell } 388f87d0fbbSRusty Russell 389b9f7ac8cSMichael S. Tsirkin if (desc.flags & cpu_to_vringh16(vrh, VRING_DESC_F_NEXT)) { 390b9f7ac8cSMichael S. Tsirkin i = vringh16_to_cpu(vrh, desc.next); 391f87d0fbbSRusty Russell } else { 392f87d0fbbSRusty Russell /* Just in case we need to finish traversing above. */ 393f87d0fbbSRusty Russell if (unlikely(up_next > 0)) { 394f87d0fbbSRusty Russell i = return_from_indirect(vrh, &up_next, 395f87d0fbbSRusty Russell &descs, &desc_max); 396f87d0fbbSRusty Russell slow = false; 397f87d0fbbSRusty Russell } else 398f87d0fbbSRusty Russell break; 399f87d0fbbSRusty Russell } 400f87d0fbbSRusty Russell 401f87d0fbbSRusty Russell if (i >= desc_max) { 402f87d0fbbSRusty Russell vringh_bad("Chained index %u > %u", i, desc_max); 403f87d0fbbSRusty Russell err = -EINVAL; 404f87d0fbbSRusty Russell goto fail; 405f87d0fbbSRusty Russell } 406f87d0fbbSRusty Russell } 407f87d0fbbSRusty Russell 408f87d0fbbSRusty Russell return 0; 409f87d0fbbSRusty Russell 410f87d0fbbSRusty Russell fail: 411f87d0fbbSRusty Russell return err; 412f87d0fbbSRusty Russell } 413f87d0fbbSRusty Russell 414f87d0fbbSRusty Russell static inline int __vringh_complete(struct vringh *vrh, 415f87d0fbbSRusty Russell const struct vring_used_elem *used, 416f87d0fbbSRusty Russell unsigned int num_used, 417b9f7ac8cSMichael S. Tsirkin int (*putu16)(const struct vringh *vrh, 418b9f7ac8cSMichael S. Tsirkin __virtio16 *p, u16 val), 4199ad9c49cSJason Wang int (*putused)(const struct vringh *vrh, 4209ad9c49cSJason Wang struct vring_used_elem *dst, 421f87d0fbbSRusty Russell const struct vring_used_elem 422f87d0fbbSRusty Russell *src, unsigned num)) 423f87d0fbbSRusty Russell { 424f87d0fbbSRusty Russell struct vring_used *used_ring; 425f87d0fbbSRusty Russell int err; 426f87d0fbbSRusty Russell u16 used_idx, off; 427f87d0fbbSRusty Russell 428f87d0fbbSRusty Russell used_ring = vrh->vring.used; 429f87d0fbbSRusty Russell used_idx = vrh->last_used_idx + vrh->completed; 430f87d0fbbSRusty Russell 431f87d0fbbSRusty Russell off = used_idx % vrh->vring.num; 432f87d0fbbSRusty Russell 433f87d0fbbSRusty Russell /* Compiler knows num_used == 1 sometimes, hence extra check */ 434f87d0fbbSRusty Russell if (num_used > 1 && unlikely(off + num_used >= vrh->vring.num)) { 435f87d0fbbSRusty Russell u16 part = vrh->vring.num - off; 4369ad9c49cSJason Wang err = putused(vrh, &used_ring->ring[off], used, part); 437f87d0fbbSRusty Russell if (!err) 4389ad9c49cSJason Wang err = putused(vrh, &used_ring->ring[0], used + part, 439f87d0fbbSRusty Russell num_used - part); 440f87d0fbbSRusty Russell } else 4419ad9c49cSJason Wang err = putused(vrh, &used_ring->ring[off], used, num_used); 442f87d0fbbSRusty Russell 443f87d0fbbSRusty Russell if (err) { 444f87d0fbbSRusty Russell vringh_bad("Failed to write %u used entries %u at %p", 445f87d0fbbSRusty Russell num_used, off, &used_ring->ring[off]); 446f87d0fbbSRusty Russell return err; 447f87d0fbbSRusty Russell } 448f87d0fbbSRusty Russell 449f87d0fbbSRusty Russell /* Make sure buffer is written before we update index. */ 450f87d0fbbSRusty Russell virtio_wmb(vrh->weak_barriers); 451f87d0fbbSRusty Russell 452b9f7ac8cSMichael S. Tsirkin err = putu16(vrh, &vrh->vring.used->idx, used_idx + num_used); 453f87d0fbbSRusty Russell if (err) { 454f87d0fbbSRusty Russell vringh_bad("Failed to update used index at %p", 455f87d0fbbSRusty Russell &vrh->vring.used->idx); 456f87d0fbbSRusty Russell return err; 457f87d0fbbSRusty Russell } 458f87d0fbbSRusty Russell 459f87d0fbbSRusty Russell vrh->completed += num_used; 460f87d0fbbSRusty Russell return 0; 461f87d0fbbSRusty Russell } 462f87d0fbbSRusty Russell 463f87d0fbbSRusty Russell 464f87d0fbbSRusty Russell static inline int __vringh_need_notify(struct vringh *vrh, 465b9f7ac8cSMichael S. Tsirkin int (*getu16)(const struct vringh *vrh, 466b9f7ac8cSMichael S. Tsirkin u16 *val, 467b9f7ac8cSMichael S. Tsirkin const __virtio16 *p)) 468f87d0fbbSRusty Russell { 469f87d0fbbSRusty Russell bool notify; 470f87d0fbbSRusty Russell u16 used_event; 471f87d0fbbSRusty Russell int err; 472f87d0fbbSRusty Russell 473f87d0fbbSRusty Russell /* Flush out used index update. This is paired with the 474f87d0fbbSRusty Russell * barrier that the Guest executes when enabling 475f87d0fbbSRusty Russell * interrupts. */ 476f87d0fbbSRusty Russell virtio_mb(vrh->weak_barriers); 477f87d0fbbSRusty Russell 478f87d0fbbSRusty Russell /* Old-style, without event indices. */ 479f87d0fbbSRusty Russell if (!vrh->event_indices) { 480f87d0fbbSRusty Russell u16 flags; 481b9f7ac8cSMichael S. Tsirkin err = getu16(vrh, &flags, &vrh->vring.avail->flags); 482f87d0fbbSRusty Russell if (err) { 483f87d0fbbSRusty Russell vringh_bad("Failed to get flags at %p", 484f87d0fbbSRusty Russell &vrh->vring.avail->flags); 485f87d0fbbSRusty Russell return err; 486f87d0fbbSRusty Russell } 487f87d0fbbSRusty Russell return (!(flags & VRING_AVAIL_F_NO_INTERRUPT)); 488f87d0fbbSRusty Russell } 489f87d0fbbSRusty Russell 490f87d0fbbSRusty Russell /* Modern: we know when other side wants to know. */ 491b9f7ac8cSMichael S. Tsirkin err = getu16(vrh, &used_event, &vring_used_event(&vrh->vring)); 492f87d0fbbSRusty Russell if (err) { 493f87d0fbbSRusty Russell vringh_bad("Failed to get used event idx at %p", 494f87d0fbbSRusty Russell &vring_used_event(&vrh->vring)); 495f87d0fbbSRusty Russell return err; 496f87d0fbbSRusty Russell } 497f87d0fbbSRusty Russell 498f87d0fbbSRusty Russell /* Just in case we added so many that we wrap. */ 499f87d0fbbSRusty Russell if (unlikely(vrh->completed > 0xffff)) 500f87d0fbbSRusty Russell notify = true; 501f87d0fbbSRusty Russell else 502f87d0fbbSRusty Russell notify = vring_need_event(used_event, 503f87d0fbbSRusty Russell vrh->last_used_idx + vrh->completed, 504f87d0fbbSRusty Russell vrh->last_used_idx); 505f87d0fbbSRusty Russell 506f87d0fbbSRusty Russell vrh->last_used_idx += vrh->completed; 507f87d0fbbSRusty Russell vrh->completed = 0; 508f87d0fbbSRusty Russell return notify; 509f87d0fbbSRusty Russell } 510f87d0fbbSRusty Russell 511f87d0fbbSRusty Russell static inline bool __vringh_notify_enable(struct vringh *vrh, 512b9f7ac8cSMichael S. Tsirkin int (*getu16)(const struct vringh *vrh, 513b9f7ac8cSMichael S. Tsirkin u16 *val, const __virtio16 *p), 514b9f7ac8cSMichael S. Tsirkin int (*putu16)(const struct vringh *vrh, 515b9f7ac8cSMichael S. Tsirkin __virtio16 *p, u16 val)) 516f87d0fbbSRusty Russell { 517f87d0fbbSRusty Russell u16 avail; 518f87d0fbbSRusty Russell 519f87d0fbbSRusty Russell if (!vrh->event_indices) { 520f87d0fbbSRusty Russell /* Old-school; update flags. */ 521b9f7ac8cSMichael S. Tsirkin if (putu16(vrh, &vrh->vring.used->flags, 0) != 0) { 522f87d0fbbSRusty Russell vringh_bad("Clearing used flags %p", 523f87d0fbbSRusty Russell &vrh->vring.used->flags); 524f87d0fbbSRusty Russell return true; 525f87d0fbbSRusty Russell } 526f87d0fbbSRusty Russell } else { 527b9f7ac8cSMichael S. Tsirkin if (putu16(vrh, &vring_avail_event(&vrh->vring), 528f87d0fbbSRusty Russell vrh->last_avail_idx) != 0) { 529f87d0fbbSRusty Russell vringh_bad("Updating avail event index %p", 530f87d0fbbSRusty Russell &vring_avail_event(&vrh->vring)); 531f87d0fbbSRusty Russell return true; 532f87d0fbbSRusty Russell } 533f87d0fbbSRusty Russell } 534f87d0fbbSRusty Russell 535f87d0fbbSRusty Russell /* They could have slipped one in as we were doing that: make 536f87d0fbbSRusty Russell * sure it's written, then check again. */ 537f87d0fbbSRusty Russell virtio_mb(vrh->weak_barriers); 538f87d0fbbSRusty Russell 539b9f7ac8cSMichael S. Tsirkin if (getu16(vrh, &avail, &vrh->vring.avail->idx) != 0) { 540f87d0fbbSRusty Russell vringh_bad("Failed to check avail idx at %p", 541f87d0fbbSRusty Russell &vrh->vring.avail->idx); 542f87d0fbbSRusty Russell return true; 543f87d0fbbSRusty Russell } 544f87d0fbbSRusty Russell 545f87d0fbbSRusty Russell /* This is unlikely, so we just leave notifications enabled 546f87d0fbbSRusty Russell * (if we're using event_indices, we'll only get one 547f87d0fbbSRusty Russell * notification anyway). */ 548f87d0fbbSRusty Russell return avail == vrh->last_avail_idx; 549f87d0fbbSRusty Russell } 550f87d0fbbSRusty Russell 551f87d0fbbSRusty Russell static inline void __vringh_notify_disable(struct vringh *vrh, 552b9f7ac8cSMichael S. Tsirkin int (*putu16)(const struct vringh *vrh, 553b9f7ac8cSMichael S. Tsirkin __virtio16 *p, u16 val)) 554f87d0fbbSRusty Russell { 555f87d0fbbSRusty Russell if (!vrh->event_indices) { 556f87d0fbbSRusty Russell /* Old-school; update flags. */ 557b9f7ac8cSMichael S. Tsirkin if (putu16(vrh, &vrh->vring.used->flags, 558b9f7ac8cSMichael S. Tsirkin VRING_USED_F_NO_NOTIFY)) { 559f87d0fbbSRusty Russell vringh_bad("Setting used flags %p", 560f87d0fbbSRusty Russell &vrh->vring.used->flags); 561f87d0fbbSRusty Russell } 562f87d0fbbSRusty Russell } 563f87d0fbbSRusty Russell } 564f87d0fbbSRusty Russell 565f87d0fbbSRusty Russell /* Userspace access helpers: in this case, addresses are really userspace. */ 566b9f7ac8cSMichael S. Tsirkin static inline int getu16_user(const struct vringh *vrh, u16 *val, const __virtio16 *p) 567f87d0fbbSRusty Russell { 568b9f7ac8cSMichael S. Tsirkin __virtio16 v = 0; 569b9f7ac8cSMichael S. Tsirkin int rc = get_user(v, (__force __virtio16 __user *)p); 570b9f7ac8cSMichael S. Tsirkin *val = vringh16_to_cpu(vrh, v); 571b9f7ac8cSMichael S. Tsirkin return rc; 572f87d0fbbSRusty Russell } 573f87d0fbbSRusty Russell 574b9f7ac8cSMichael S. Tsirkin static inline int putu16_user(const struct vringh *vrh, __virtio16 *p, u16 val) 575f87d0fbbSRusty Russell { 576b9f7ac8cSMichael S. Tsirkin __virtio16 v = cpu_to_vringh16(vrh, val); 577b9f7ac8cSMichael S. Tsirkin return put_user(v, (__force __virtio16 __user *)p); 578f87d0fbbSRusty Russell } 579f87d0fbbSRusty Russell 5809ad9c49cSJason Wang static inline int copydesc_user(const struct vringh *vrh, 5819ad9c49cSJason Wang void *dst, const void *src, size_t len) 582f87d0fbbSRusty Russell { 583f87d0fbbSRusty Russell return copy_from_user(dst, (__force void __user *)src, len) ? 584f87d0fbbSRusty Russell -EFAULT : 0; 585f87d0fbbSRusty Russell } 586f87d0fbbSRusty Russell 5879ad9c49cSJason Wang static inline int putused_user(const struct vringh *vrh, 5889ad9c49cSJason Wang struct vring_used_elem *dst, 589f87d0fbbSRusty Russell const struct vring_used_elem *src, 590f87d0fbbSRusty Russell unsigned int num) 591f87d0fbbSRusty Russell { 592f87d0fbbSRusty Russell return copy_to_user((__force void __user *)dst, src, 593f87d0fbbSRusty Russell sizeof(*dst) * num) ? -EFAULT : 0; 594f87d0fbbSRusty Russell } 595f87d0fbbSRusty Russell 5969ad9c49cSJason Wang static inline int xfer_from_user(const struct vringh *vrh, void *src, 5979ad9c49cSJason Wang void *dst, size_t len) 598f87d0fbbSRusty Russell { 599f87d0fbbSRusty Russell return copy_from_user(dst, (__force void __user *)src, len) ? 600f87d0fbbSRusty Russell -EFAULT : 0; 601f87d0fbbSRusty Russell } 602f87d0fbbSRusty Russell 6039ad9c49cSJason Wang static inline int xfer_to_user(const struct vringh *vrh, 6049ad9c49cSJason Wang void *dst, void *src, size_t len) 605f87d0fbbSRusty Russell { 606f87d0fbbSRusty Russell return copy_to_user((__force void __user *)dst, src, len) ? 607f87d0fbbSRusty Russell -EFAULT : 0; 608f87d0fbbSRusty Russell } 609f87d0fbbSRusty Russell 610f87d0fbbSRusty Russell /** 611f87d0fbbSRusty Russell * vringh_init_user - initialize a vringh for a userspace vring. 612f87d0fbbSRusty Russell * @vrh: the vringh to initialize. 613f87d0fbbSRusty Russell * @features: the feature bits for this ring. 614f87d0fbbSRusty Russell * @num: the number of elements. 615f87d0fbbSRusty Russell * @weak_barriers: true if we only need memory barriers, not I/O. 616f87d0fbbSRusty Russell * @desc: the userpace descriptor pointer. 617f87d0fbbSRusty Russell * @avail: the userpace avail pointer. 618f87d0fbbSRusty Russell * @used: the userpace used pointer. 619f87d0fbbSRusty Russell * 620f87d0fbbSRusty Russell * Returns an error if num is invalid: you should check pointers 621f87d0fbbSRusty Russell * yourself! 622f87d0fbbSRusty Russell */ 623b97a8a90SMichael S. Tsirkin int vringh_init_user(struct vringh *vrh, u64 features, 624f87d0fbbSRusty Russell unsigned int num, bool weak_barriers, 625a865e420SMichael S. Tsirkin vring_desc_t __user *desc, 626a865e420SMichael S. Tsirkin vring_avail_t __user *avail, 627a865e420SMichael S. Tsirkin vring_used_t __user *used) 628f87d0fbbSRusty Russell { 629f87d0fbbSRusty Russell /* Sane power of 2 please! */ 630f87d0fbbSRusty Russell if (!num || num > 0xffff || (num & (num - 1))) { 631f87d0fbbSRusty Russell vringh_bad("Bad ring size %u", num); 632f87d0fbbSRusty Russell return -EINVAL; 633f87d0fbbSRusty Russell } 634f87d0fbbSRusty Russell 635b9f7ac8cSMichael S. Tsirkin vrh->little_endian = (features & (1ULL << VIRTIO_F_VERSION_1)); 636f87d0fbbSRusty Russell vrh->event_indices = (features & (1 << VIRTIO_RING_F_EVENT_IDX)); 637f87d0fbbSRusty Russell vrh->weak_barriers = weak_barriers; 638f87d0fbbSRusty Russell vrh->completed = 0; 639f87d0fbbSRusty Russell vrh->last_avail_idx = 0; 640f87d0fbbSRusty Russell vrh->last_used_idx = 0; 641f87d0fbbSRusty Russell vrh->vring.num = num; 642f87d0fbbSRusty Russell /* vring expects kernel addresses, but only used via accessors. */ 643f87d0fbbSRusty Russell vrh->vring.desc = (__force struct vring_desc *)desc; 644f87d0fbbSRusty Russell vrh->vring.avail = (__force struct vring_avail *)avail; 645f87d0fbbSRusty Russell vrh->vring.used = (__force struct vring_used *)used; 646f87d0fbbSRusty Russell return 0; 647f87d0fbbSRusty Russell } 648f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_init_user); 649f87d0fbbSRusty Russell 650f87d0fbbSRusty Russell /** 651f87d0fbbSRusty Russell * vringh_getdesc_user - get next available descriptor from userspace ring. 652f87d0fbbSRusty Russell * @vrh: the userspace vring. 653f87d0fbbSRusty Russell * @riov: where to put the readable descriptors (or NULL) 654f87d0fbbSRusty Russell * @wiov: where to put the writable descriptors (or NULL) 655f87d0fbbSRusty Russell * @getrange: function to call to check ranges. 656f87d0fbbSRusty Russell * @head: head index we received, for passing to vringh_complete_user(). 657f87d0fbbSRusty Russell * 658f87d0fbbSRusty Russell * Returns 0 if there was no descriptor, 1 if there was, or -errno. 659f87d0fbbSRusty Russell * 660f87d0fbbSRusty Russell * Note that on error return, you can tell the difference between an 661f87d0fbbSRusty Russell * invalid ring and a single invalid descriptor: in the former case, 662f87d0fbbSRusty Russell * *head will be vrh->vring.num. You may be able to ignore an invalid 663f87d0fbbSRusty Russell * descriptor, but there's not much you can do with an invalid ring. 664f87d0fbbSRusty Russell * 665*69c13c58SStefano Garzarella * Note that you can reuse riov and wiov with subsequent calls. Content is 666*69c13c58SStefano Garzarella * overwritten and memory reallocated if more space is needed. 667*69c13c58SStefano Garzarella * When you don't have to use riov and wiov anymore, you should clean up them 668*69c13c58SStefano Garzarella * calling vringh_iov_cleanup() to release the memory, even on error! 669f87d0fbbSRusty Russell */ 670f87d0fbbSRusty Russell int vringh_getdesc_user(struct vringh *vrh, 671f87d0fbbSRusty Russell struct vringh_iov *riov, 672f87d0fbbSRusty Russell struct vringh_iov *wiov, 673f87d0fbbSRusty Russell bool (*getrange)(struct vringh *vrh, 674f87d0fbbSRusty Russell u64 addr, struct vringh_range *r), 675f87d0fbbSRusty Russell u16 *head) 676f87d0fbbSRusty Russell { 677f87d0fbbSRusty Russell int err; 678f87d0fbbSRusty Russell 679f87d0fbbSRusty Russell *head = vrh->vring.num; 680f87d0fbbSRusty Russell err = __vringh_get_head(vrh, getu16_user, &vrh->last_avail_idx); 681f87d0fbbSRusty Russell if (err < 0) 682f87d0fbbSRusty Russell return err; 683f87d0fbbSRusty Russell 684f87d0fbbSRusty Russell /* Empty... */ 685f87d0fbbSRusty Russell if (err == vrh->vring.num) 686f87d0fbbSRusty Russell return 0; 687f87d0fbbSRusty Russell 688f87d0fbbSRusty Russell /* We need the layouts to be the identical for this to work */ 689f87d0fbbSRusty Russell BUILD_BUG_ON(sizeof(struct vringh_kiov) != sizeof(struct vringh_iov)); 690f87d0fbbSRusty Russell BUILD_BUG_ON(offsetof(struct vringh_kiov, iov) != 691f87d0fbbSRusty Russell offsetof(struct vringh_iov, iov)); 692f87d0fbbSRusty Russell BUILD_BUG_ON(offsetof(struct vringh_kiov, i) != 693f87d0fbbSRusty Russell offsetof(struct vringh_iov, i)); 694f87d0fbbSRusty Russell BUILD_BUG_ON(offsetof(struct vringh_kiov, used) != 695f87d0fbbSRusty Russell offsetof(struct vringh_iov, used)); 696f87d0fbbSRusty Russell BUILD_BUG_ON(offsetof(struct vringh_kiov, max_num) != 697f87d0fbbSRusty Russell offsetof(struct vringh_iov, max_num)); 698f87d0fbbSRusty Russell BUILD_BUG_ON(sizeof(struct iovec) != sizeof(struct kvec)); 699f87d0fbbSRusty Russell BUILD_BUG_ON(offsetof(struct iovec, iov_base) != 700f87d0fbbSRusty Russell offsetof(struct kvec, iov_base)); 701f87d0fbbSRusty Russell BUILD_BUG_ON(offsetof(struct iovec, iov_len) != 702f87d0fbbSRusty Russell offsetof(struct kvec, iov_len)); 703f87d0fbbSRusty Russell BUILD_BUG_ON(sizeof(((struct iovec *)NULL)->iov_base) 704f87d0fbbSRusty Russell != sizeof(((struct kvec *)NULL)->iov_base)); 705f87d0fbbSRusty Russell BUILD_BUG_ON(sizeof(((struct iovec *)NULL)->iov_len) 706f87d0fbbSRusty Russell != sizeof(((struct kvec *)NULL)->iov_len)); 707f87d0fbbSRusty Russell 708f87d0fbbSRusty Russell *head = err; 709f87d0fbbSRusty Russell err = __vringh_iov(vrh, *head, (struct vringh_kiov *)riov, 710f87d0fbbSRusty Russell (struct vringh_kiov *)wiov, 711f87d0fbbSRusty Russell range_check, getrange, GFP_KERNEL, copydesc_user); 712f87d0fbbSRusty Russell if (err) 713f87d0fbbSRusty Russell return err; 714f87d0fbbSRusty Russell 715f87d0fbbSRusty Russell return 1; 716f87d0fbbSRusty Russell } 717f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_getdesc_user); 718f87d0fbbSRusty Russell 719f87d0fbbSRusty Russell /** 720f87d0fbbSRusty Russell * vringh_iov_pull_user - copy bytes from vring_iov. 721f87d0fbbSRusty Russell * @riov: the riov as passed to vringh_getdesc_user() (updated as we consume) 722f87d0fbbSRusty Russell * @dst: the place to copy. 723f87d0fbbSRusty Russell * @len: the maximum length to copy. 724f87d0fbbSRusty Russell * 725f87d0fbbSRusty Russell * Returns the bytes copied <= len or a negative errno. 726f87d0fbbSRusty Russell */ 727f87d0fbbSRusty Russell ssize_t vringh_iov_pull_user(struct vringh_iov *riov, void *dst, size_t len) 728f87d0fbbSRusty Russell { 7299ad9c49cSJason Wang return vringh_iov_xfer(NULL, (struct vringh_kiov *)riov, 730f87d0fbbSRusty Russell dst, len, xfer_from_user); 731f87d0fbbSRusty Russell } 732f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_iov_pull_user); 733f87d0fbbSRusty Russell 734f87d0fbbSRusty Russell /** 735f87d0fbbSRusty Russell * vringh_iov_push_user - copy bytes into vring_iov. 736f87d0fbbSRusty Russell * @wiov: the wiov as passed to vringh_getdesc_user() (updated as we consume) 7378009b0f4SStefano Garzarella * @src: the place to copy from. 738f87d0fbbSRusty Russell * @len: the maximum length to copy. 739f87d0fbbSRusty Russell * 740f87d0fbbSRusty Russell * Returns the bytes copied <= len or a negative errno. 741f87d0fbbSRusty Russell */ 742f87d0fbbSRusty Russell ssize_t vringh_iov_push_user(struct vringh_iov *wiov, 743f87d0fbbSRusty Russell const void *src, size_t len) 744f87d0fbbSRusty Russell { 7459ad9c49cSJason Wang return vringh_iov_xfer(NULL, (struct vringh_kiov *)wiov, 746f87d0fbbSRusty Russell (void *)src, len, xfer_to_user); 747f87d0fbbSRusty Russell } 748f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_iov_push_user); 749f87d0fbbSRusty Russell 750f87d0fbbSRusty Russell /** 751f87d0fbbSRusty Russell * vringh_abandon_user - we've decided not to handle the descriptor(s). 752f87d0fbbSRusty Russell * @vrh: the vring. 753f87d0fbbSRusty Russell * @num: the number of descriptors to put back (ie. num 754f87d0fbbSRusty Russell * vringh_get_user() to undo). 755f87d0fbbSRusty Russell * 756f87d0fbbSRusty Russell * The next vringh_get_user() will return the old descriptor(s) again. 757f87d0fbbSRusty Russell */ 758f87d0fbbSRusty Russell void vringh_abandon_user(struct vringh *vrh, unsigned int num) 759f87d0fbbSRusty Russell { 760f87d0fbbSRusty Russell /* We only update vring_avail_event(vr) when we want to be notified, 761f87d0fbbSRusty Russell * so we haven't changed that yet. */ 762f87d0fbbSRusty Russell vrh->last_avail_idx -= num; 763f87d0fbbSRusty Russell } 764f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_abandon_user); 765f87d0fbbSRusty Russell 766f87d0fbbSRusty Russell /** 767f87d0fbbSRusty Russell * vringh_complete_user - we've finished with descriptor, publish it. 768f87d0fbbSRusty Russell * @vrh: the vring. 769f87d0fbbSRusty Russell * @head: the head as filled in by vringh_getdesc_user. 770f87d0fbbSRusty Russell * @len: the length of data we have written. 771f87d0fbbSRusty Russell * 772f87d0fbbSRusty Russell * You should check vringh_need_notify_user() after one or more calls 773f87d0fbbSRusty Russell * to this function. 774f87d0fbbSRusty Russell */ 775f87d0fbbSRusty Russell int vringh_complete_user(struct vringh *vrh, u16 head, u32 len) 776f87d0fbbSRusty Russell { 777f87d0fbbSRusty Russell struct vring_used_elem used; 778f87d0fbbSRusty Russell 779b9f7ac8cSMichael S. Tsirkin used.id = cpu_to_vringh32(vrh, head); 780b9f7ac8cSMichael S. Tsirkin used.len = cpu_to_vringh32(vrh, len); 781f87d0fbbSRusty Russell return __vringh_complete(vrh, &used, 1, putu16_user, putused_user); 782f87d0fbbSRusty Russell } 783f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_complete_user); 784f87d0fbbSRusty Russell 785f87d0fbbSRusty Russell /** 786f87d0fbbSRusty Russell * vringh_complete_multi_user - we've finished with many descriptors. 787f87d0fbbSRusty Russell * @vrh: the vring. 788f87d0fbbSRusty Russell * @used: the head, length pairs. 789f87d0fbbSRusty Russell * @num_used: the number of used elements. 790f87d0fbbSRusty Russell * 791f87d0fbbSRusty Russell * You should check vringh_need_notify_user() after one or more calls 792f87d0fbbSRusty Russell * to this function. 793f87d0fbbSRusty Russell */ 794f87d0fbbSRusty Russell int vringh_complete_multi_user(struct vringh *vrh, 795f87d0fbbSRusty Russell const struct vring_used_elem used[], 796f87d0fbbSRusty Russell unsigned num_used) 797f87d0fbbSRusty Russell { 798f87d0fbbSRusty Russell return __vringh_complete(vrh, used, num_used, 799f87d0fbbSRusty Russell putu16_user, putused_user); 800f87d0fbbSRusty Russell } 801f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_complete_multi_user); 802f87d0fbbSRusty Russell 803f87d0fbbSRusty Russell /** 804f87d0fbbSRusty Russell * vringh_notify_enable_user - we want to know if something changes. 805f87d0fbbSRusty Russell * @vrh: the vring. 806f87d0fbbSRusty Russell * 807f87d0fbbSRusty Russell * This always enables notifications, but returns false if there are 808f87d0fbbSRusty Russell * now more buffers available in the vring. 809f87d0fbbSRusty Russell */ 810f87d0fbbSRusty Russell bool vringh_notify_enable_user(struct vringh *vrh) 811f87d0fbbSRusty Russell { 812f87d0fbbSRusty Russell return __vringh_notify_enable(vrh, getu16_user, putu16_user); 813f87d0fbbSRusty Russell } 814f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_notify_enable_user); 815f87d0fbbSRusty Russell 816f87d0fbbSRusty Russell /** 817f87d0fbbSRusty Russell * vringh_notify_disable_user - don't tell us if something changes. 818f87d0fbbSRusty Russell * @vrh: the vring. 819f87d0fbbSRusty Russell * 820f87d0fbbSRusty Russell * This is our normal running state: we disable and then only enable when 821f87d0fbbSRusty Russell * we're going to sleep. 822f87d0fbbSRusty Russell */ 823f87d0fbbSRusty Russell void vringh_notify_disable_user(struct vringh *vrh) 824f87d0fbbSRusty Russell { 825f87d0fbbSRusty Russell __vringh_notify_disable(vrh, putu16_user); 826f87d0fbbSRusty Russell } 827f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_notify_disable_user); 828f87d0fbbSRusty Russell 829f87d0fbbSRusty Russell /** 830f87d0fbbSRusty Russell * vringh_need_notify_user - must we tell the other side about used buffers? 831f87d0fbbSRusty Russell * @vrh: the vring we've called vringh_complete_user() on. 832f87d0fbbSRusty Russell * 833f87d0fbbSRusty Russell * Returns -errno or 0 if we don't need to tell the other side, 1 if we do. 834f87d0fbbSRusty Russell */ 835f87d0fbbSRusty Russell int vringh_need_notify_user(struct vringh *vrh) 836f87d0fbbSRusty Russell { 837f87d0fbbSRusty Russell return __vringh_need_notify(vrh, getu16_user); 838f87d0fbbSRusty Russell } 839f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_need_notify_user); 840f87d0fbbSRusty Russell 841f87d0fbbSRusty Russell /* Kernelspace access helpers. */ 842b9f7ac8cSMichael S. Tsirkin static inline int getu16_kern(const struct vringh *vrh, 843b9f7ac8cSMichael S. Tsirkin u16 *val, const __virtio16 *p) 844f87d0fbbSRusty Russell { 8459d1b972fSMark Rutland *val = vringh16_to_cpu(vrh, READ_ONCE(*p)); 846f87d0fbbSRusty Russell return 0; 847f87d0fbbSRusty Russell } 848f87d0fbbSRusty Russell 849b9f7ac8cSMichael S. Tsirkin static inline int putu16_kern(const struct vringh *vrh, __virtio16 *p, u16 val) 850f87d0fbbSRusty Russell { 8519d1b972fSMark Rutland WRITE_ONCE(*p, cpu_to_vringh16(vrh, val)); 852f87d0fbbSRusty Russell return 0; 853f87d0fbbSRusty Russell } 854f87d0fbbSRusty Russell 8559ad9c49cSJason Wang static inline int copydesc_kern(const struct vringh *vrh, 8569ad9c49cSJason Wang void *dst, const void *src, size_t len) 857f87d0fbbSRusty Russell { 858f87d0fbbSRusty Russell memcpy(dst, src, len); 859f87d0fbbSRusty Russell return 0; 860f87d0fbbSRusty Russell } 861f87d0fbbSRusty Russell 8629ad9c49cSJason Wang static inline int putused_kern(const struct vringh *vrh, 8639ad9c49cSJason Wang struct vring_used_elem *dst, 864f87d0fbbSRusty Russell const struct vring_used_elem *src, 865f87d0fbbSRusty Russell unsigned int num) 866f87d0fbbSRusty Russell { 867f87d0fbbSRusty Russell memcpy(dst, src, num * sizeof(*dst)); 868f87d0fbbSRusty Russell return 0; 869f87d0fbbSRusty Russell } 870f87d0fbbSRusty Russell 8719ad9c49cSJason Wang static inline int xfer_kern(const struct vringh *vrh, void *src, 8729ad9c49cSJason Wang void *dst, size_t len) 873f87d0fbbSRusty Russell { 874f87d0fbbSRusty Russell memcpy(dst, src, len); 875f87d0fbbSRusty Russell return 0; 876f87d0fbbSRusty Russell } 877f87d0fbbSRusty Russell 8789ad9c49cSJason Wang static inline int kern_xfer(const struct vringh *vrh, void *dst, 8799ad9c49cSJason Wang void *src, size_t len) 880b3683deeSJason Wang { 881b3683deeSJason Wang memcpy(dst, src, len); 882b3683deeSJason Wang return 0; 883b3683deeSJason Wang } 884b3683deeSJason Wang 885f87d0fbbSRusty Russell /** 886f87d0fbbSRusty Russell * vringh_init_kern - initialize a vringh for a kernelspace vring. 887f87d0fbbSRusty Russell * @vrh: the vringh to initialize. 888f87d0fbbSRusty Russell * @features: the feature bits for this ring. 889f87d0fbbSRusty Russell * @num: the number of elements. 890f87d0fbbSRusty Russell * @weak_barriers: true if we only need memory barriers, not I/O. 891f87d0fbbSRusty Russell * @desc: the userpace descriptor pointer. 892f87d0fbbSRusty Russell * @avail: the userpace avail pointer. 893f87d0fbbSRusty Russell * @used: the userpace used pointer. 894f87d0fbbSRusty Russell * 895f87d0fbbSRusty Russell * Returns an error if num is invalid. 896f87d0fbbSRusty Russell */ 897b97a8a90SMichael S. Tsirkin int vringh_init_kern(struct vringh *vrh, u64 features, 898f87d0fbbSRusty Russell unsigned int num, bool weak_barriers, 899f87d0fbbSRusty Russell struct vring_desc *desc, 900f87d0fbbSRusty Russell struct vring_avail *avail, 901f87d0fbbSRusty Russell struct vring_used *used) 902f87d0fbbSRusty Russell { 903f87d0fbbSRusty Russell /* Sane power of 2 please! */ 904f87d0fbbSRusty Russell if (!num || num > 0xffff || (num & (num - 1))) { 905f87d0fbbSRusty Russell vringh_bad("Bad ring size %u", num); 906f87d0fbbSRusty Russell return -EINVAL; 907f87d0fbbSRusty Russell } 908f87d0fbbSRusty Russell 909b9f7ac8cSMichael S. Tsirkin vrh->little_endian = (features & (1ULL << VIRTIO_F_VERSION_1)); 910f87d0fbbSRusty Russell vrh->event_indices = (features & (1 << VIRTIO_RING_F_EVENT_IDX)); 911f87d0fbbSRusty Russell vrh->weak_barriers = weak_barriers; 912f87d0fbbSRusty Russell vrh->completed = 0; 913f87d0fbbSRusty Russell vrh->last_avail_idx = 0; 914f87d0fbbSRusty Russell vrh->last_used_idx = 0; 915f87d0fbbSRusty Russell vrh->vring.num = num; 916f87d0fbbSRusty Russell vrh->vring.desc = desc; 917f87d0fbbSRusty Russell vrh->vring.avail = avail; 918f87d0fbbSRusty Russell vrh->vring.used = used; 919f87d0fbbSRusty Russell return 0; 920f87d0fbbSRusty Russell } 921f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_init_kern); 922f87d0fbbSRusty Russell 923f87d0fbbSRusty Russell /** 924f87d0fbbSRusty Russell * vringh_getdesc_kern - get next available descriptor from kernelspace ring. 925f87d0fbbSRusty Russell * @vrh: the kernelspace vring. 926f87d0fbbSRusty Russell * @riov: where to put the readable descriptors (or NULL) 927f87d0fbbSRusty Russell * @wiov: where to put the writable descriptors (or NULL) 928f87d0fbbSRusty Russell * @head: head index we received, for passing to vringh_complete_kern(). 929f87d0fbbSRusty Russell * @gfp: flags for allocating larger riov/wiov. 930f87d0fbbSRusty Russell * 931f87d0fbbSRusty Russell * Returns 0 if there was no descriptor, 1 if there was, or -errno. 932f87d0fbbSRusty Russell * 933f87d0fbbSRusty Russell * Note that on error return, you can tell the difference between an 934f87d0fbbSRusty Russell * invalid ring and a single invalid descriptor: in the former case, 935f87d0fbbSRusty Russell * *head will be vrh->vring.num. You may be able to ignore an invalid 936f87d0fbbSRusty Russell * descriptor, but there's not much you can do with an invalid ring. 937f87d0fbbSRusty Russell * 938*69c13c58SStefano Garzarella * Note that you can reuse riov and wiov with subsequent calls. Content is 939*69c13c58SStefano Garzarella * overwritten and memory reallocated if more space is needed. 940*69c13c58SStefano Garzarella * When you don't have to use riov and wiov anymore, you should clean up them 941*69c13c58SStefano Garzarella * calling vringh_kiov_cleanup() to release the memory, even on error! 942f87d0fbbSRusty Russell */ 943f87d0fbbSRusty Russell int vringh_getdesc_kern(struct vringh *vrh, 944f87d0fbbSRusty Russell struct vringh_kiov *riov, 945f87d0fbbSRusty Russell struct vringh_kiov *wiov, 946f87d0fbbSRusty Russell u16 *head, 947f87d0fbbSRusty Russell gfp_t gfp) 948f87d0fbbSRusty Russell { 949f87d0fbbSRusty Russell int err; 950f87d0fbbSRusty Russell 951f87d0fbbSRusty Russell err = __vringh_get_head(vrh, getu16_kern, &vrh->last_avail_idx); 952f87d0fbbSRusty Russell if (err < 0) 953f87d0fbbSRusty Russell return err; 954f87d0fbbSRusty Russell 955f87d0fbbSRusty Russell /* Empty... */ 956f87d0fbbSRusty Russell if (err == vrh->vring.num) 957f87d0fbbSRusty Russell return 0; 958f87d0fbbSRusty Russell 959f87d0fbbSRusty Russell *head = err; 960f87d0fbbSRusty Russell err = __vringh_iov(vrh, *head, riov, wiov, no_range_check, NULL, 961f87d0fbbSRusty Russell gfp, copydesc_kern); 962f87d0fbbSRusty Russell if (err) 963f87d0fbbSRusty Russell return err; 964f87d0fbbSRusty Russell 965f87d0fbbSRusty Russell return 1; 966f87d0fbbSRusty Russell } 967f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_getdesc_kern); 968f87d0fbbSRusty Russell 969f87d0fbbSRusty Russell /** 970f87d0fbbSRusty Russell * vringh_iov_pull_kern - copy bytes from vring_iov. 971f87d0fbbSRusty Russell * @riov: the riov as passed to vringh_getdesc_kern() (updated as we consume) 972f87d0fbbSRusty Russell * @dst: the place to copy. 973f87d0fbbSRusty Russell * @len: the maximum length to copy. 974f87d0fbbSRusty Russell * 975f87d0fbbSRusty Russell * Returns the bytes copied <= len or a negative errno. 976f87d0fbbSRusty Russell */ 977f87d0fbbSRusty Russell ssize_t vringh_iov_pull_kern(struct vringh_kiov *riov, void *dst, size_t len) 978f87d0fbbSRusty Russell { 9799ad9c49cSJason Wang return vringh_iov_xfer(NULL, riov, dst, len, xfer_kern); 980f87d0fbbSRusty Russell } 981f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_iov_pull_kern); 982f87d0fbbSRusty Russell 983f87d0fbbSRusty Russell /** 984f87d0fbbSRusty Russell * vringh_iov_push_kern - copy bytes into vring_iov. 985f87d0fbbSRusty Russell * @wiov: the wiov as passed to vringh_getdesc_kern() (updated as we consume) 9868009b0f4SStefano Garzarella * @src: the place to copy from. 987f87d0fbbSRusty Russell * @len: the maximum length to copy. 988f87d0fbbSRusty Russell * 989f87d0fbbSRusty Russell * Returns the bytes copied <= len or a negative errno. 990f87d0fbbSRusty Russell */ 991f87d0fbbSRusty Russell ssize_t vringh_iov_push_kern(struct vringh_kiov *wiov, 992f87d0fbbSRusty Russell const void *src, size_t len) 993f87d0fbbSRusty Russell { 9949ad9c49cSJason Wang return vringh_iov_xfer(NULL, wiov, (void *)src, len, kern_xfer); 995f87d0fbbSRusty Russell } 996f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_iov_push_kern); 997f87d0fbbSRusty Russell 998f87d0fbbSRusty Russell /** 999f87d0fbbSRusty Russell * vringh_abandon_kern - we've decided not to handle the descriptor(s). 1000f87d0fbbSRusty Russell * @vrh: the vring. 1001f87d0fbbSRusty Russell * @num: the number of descriptors to put back (ie. num 1002f87d0fbbSRusty Russell * vringh_get_kern() to undo). 1003f87d0fbbSRusty Russell * 1004f87d0fbbSRusty Russell * The next vringh_get_kern() will return the old descriptor(s) again. 1005f87d0fbbSRusty Russell */ 1006f87d0fbbSRusty Russell void vringh_abandon_kern(struct vringh *vrh, unsigned int num) 1007f87d0fbbSRusty Russell { 1008f87d0fbbSRusty Russell /* We only update vring_avail_event(vr) when we want to be notified, 1009f87d0fbbSRusty Russell * so we haven't changed that yet. */ 1010f87d0fbbSRusty Russell vrh->last_avail_idx -= num; 1011f87d0fbbSRusty Russell } 1012f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_abandon_kern); 1013f87d0fbbSRusty Russell 1014f87d0fbbSRusty Russell /** 1015f87d0fbbSRusty Russell * vringh_complete_kern - we've finished with descriptor, publish it. 1016f87d0fbbSRusty Russell * @vrh: the vring. 1017f87d0fbbSRusty Russell * @head: the head as filled in by vringh_getdesc_kern. 1018f87d0fbbSRusty Russell * @len: the length of data we have written. 1019f87d0fbbSRusty Russell * 1020f87d0fbbSRusty Russell * You should check vringh_need_notify_kern() after one or more calls 1021f87d0fbbSRusty Russell * to this function. 1022f87d0fbbSRusty Russell */ 1023f87d0fbbSRusty Russell int vringh_complete_kern(struct vringh *vrh, u16 head, u32 len) 1024f87d0fbbSRusty Russell { 1025f87d0fbbSRusty Russell struct vring_used_elem used; 1026f87d0fbbSRusty Russell 1027b9f7ac8cSMichael S. Tsirkin used.id = cpu_to_vringh32(vrh, head); 1028b9f7ac8cSMichael S. Tsirkin used.len = cpu_to_vringh32(vrh, len); 1029f87d0fbbSRusty Russell 1030f87d0fbbSRusty Russell return __vringh_complete(vrh, &used, 1, putu16_kern, putused_kern); 1031f87d0fbbSRusty Russell } 1032f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_complete_kern); 1033f87d0fbbSRusty Russell 1034f87d0fbbSRusty Russell /** 1035f87d0fbbSRusty Russell * vringh_notify_enable_kern - we want to know if something changes. 1036f87d0fbbSRusty Russell * @vrh: the vring. 1037f87d0fbbSRusty Russell * 1038f87d0fbbSRusty Russell * This always enables notifications, but returns false if there are 1039f87d0fbbSRusty Russell * now more buffers available in the vring. 1040f87d0fbbSRusty Russell */ 1041f87d0fbbSRusty Russell bool vringh_notify_enable_kern(struct vringh *vrh) 1042f87d0fbbSRusty Russell { 1043f87d0fbbSRusty Russell return __vringh_notify_enable(vrh, getu16_kern, putu16_kern); 1044f87d0fbbSRusty Russell } 1045f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_notify_enable_kern); 1046f87d0fbbSRusty Russell 1047f87d0fbbSRusty Russell /** 1048f87d0fbbSRusty Russell * vringh_notify_disable_kern - don't tell us if something changes. 1049f87d0fbbSRusty Russell * @vrh: the vring. 1050f87d0fbbSRusty Russell * 1051f87d0fbbSRusty Russell * This is our normal running state: we disable and then only enable when 1052f87d0fbbSRusty Russell * we're going to sleep. 1053f87d0fbbSRusty Russell */ 1054f87d0fbbSRusty Russell void vringh_notify_disable_kern(struct vringh *vrh) 1055f87d0fbbSRusty Russell { 1056f87d0fbbSRusty Russell __vringh_notify_disable(vrh, putu16_kern); 1057f87d0fbbSRusty Russell } 1058f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_notify_disable_kern); 1059f87d0fbbSRusty Russell 1060f87d0fbbSRusty Russell /** 1061f87d0fbbSRusty Russell * vringh_need_notify_kern - must we tell the other side about used buffers? 1062f87d0fbbSRusty Russell * @vrh: the vring we've called vringh_complete_kern() on. 1063f87d0fbbSRusty Russell * 1064f87d0fbbSRusty Russell * Returns -errno or 0 if we don't need to tell the other side, 1 if we do. 1065f87d0fbbSRusty Russell */ 1066f87d0fbbSRusty Russell int vringh_need_notify_kern(struct vringh *vrh) 1067f87d0fbbSRusty Russell { 1068f87d0fbbSRusty Russell return __vringh_need_notify(vrh, getu16_kern); 1069f87d0fbbSRusty Russell } 1070f87d0fbbSRusty Russell EXPORT_SYMBOL(vringh_need_notify_kern); 1071f558a845SDave Jones 10723302363aSMichael S. Tsirkin #if IS_REACHABLE(CONFIG_VHOST_IOTLB) 10733302363aSMichael S. Tsirkin 10749ad9c49cSJason Wang static int iotlb_translate(const struct vringh *vrh, 10759ad9c49cSJason Wang u64 addr, u64 len, struct bio_vec iov[], 10769ad9c49cSJason Wang int iov_size, u32 perm) 10779ad9c49cSJason Wang { 10789ad9c49cSJason Wang struct vhost_iotlb_map *map; 10799ad9c49cSJason Wang struct vhost_iotlb *iotlb = vrh->iotlb; 10809ad9c49cSJason Wang int ret = 0; 10819ad9c49cSJason Wang u64 s = 0; 10829ad9c49cSJason Wang 1083f53d9910SStefano Garzarella spin_lock(vrh->iotlb_lock); 1084f53d9910SStefano Garzarella 10859ad9c49cSJason Wang while (len > s) { 10869ad9c49cSJason Wang u64 size, pa, pfn; 10879ad9c49cSJason Wang 10889ad9c49cSJason Wang if (unlikely(ret >= iov_size)) { 10899ad9c49cSJason Wang ret = -ENOBUFS; 10909ad9c49cSJason Wang break; 10919ad9c49cSJason Wang } 10929ad9c49cSJason Wang 10939ad9c49cSJason Wang map = vhost_iotlb_itree_first(iotlb, addr, 10949ad9c49cSJason Wang addr + len - 1); 10959ad9c49cSJason Wang if (!map || map->start > addr) { 10969ad9c49cSJason Wang ret = -EINVAL; 10979ad9c49cSJason Wang break; 10989ad9c49cSJason Wang } else if (!(map->perm & perm)) { 10999ad9c49cSJason Wang ret = -EPERM; 11009ad9c49cSJason Wang break; 11019ad9c49cSJason Wang } 11029ad9c49cSJason Wang 11039ad9c49cSJason Wang size = map->size - addr + map->start; 11049ad9c49cSJason Wang pa = map->addr + addr - map->start; 11059ad9c49cSJason Wang pfn = pa >> PAGE_SHIFT; 11069ad9c49cSJason Wang iov[ret].bv_page = pfn_to_page(pfn); 11079ad9c49cSJason Wang iov[ret].bv_len = min(len - s, size); 11089ad9c49cSJason Wang iov[ret].bv_offset = pa & (PAGE_SIZE - 1); 11099ad9c49cSJason Wang s += size; 11109ad9c49cSJason Wang addr += size; 11119ad9c49cSJason Wang ++ret; 11129ad9c49cSJason Wang } 11139ad9c49cSJason Wang 1114f53d9910SStefano Garzarella spin_unlock(vrh->iotlb_lock); 1115f53d9910SStefano Garzarella 11169ad9c49cSJason Wang return ret; 11179ad9c49cSJason Wang } 11189ad9c49cSJason Wang 11199ad9c49cSJason Wang static inline int copy_from_iotlb(const struct vringh *vrh, void *dst, 11209ad9c49cSJason Wang void *src, size_t len) 11219ad9c49cSJason Wang { 11229ad9c49cSJason Wang struct iov_iter iter; 11239ad9c49cSJason Wang struct bio_vec iov[16]; 11249ad9c49cSJason Wang int ret; 11259ad9c49cSJason Wang 11269ad9c49cSJason Wang ret = iotlb_translate(vrh, (u64)(uintptr_t)src, 11279ad9c49cSJason Wang len, iov, 16, VHOST_MAP_RO); 11289ad9c49cSJason Wang if (ret < 0) 11299ad9c49cSJason Wang return ret; 11309ad9c49cSJason Wang 11319ad9c49cSJason Wang iov_iter_bvec(&iter, READ, iov, ret, len); 11329ad9c49cSJason Wang 11339ad9c49cSJason Wang ret = copy_from_iter(dst, len, &iter); 11349ad9c49cSJason Wang 11359ad9c49cSJason Wang return ret; 11369ad9c49cSJason Wang } 11379ad9c49cSJason Wang 11389ad9c49cSJason Wang static inline int copy_to_iotlb(const struct vringh *vrh, void *dst, 11399ad9c49cSJason Wang void *src, size_t len) 11409ad9c49cSJason Wang { 11419ad9c49cSJason Wang struct iov_iter iter; 11429ad9c49cSJason Wang struct bio_vec iov[16]; 11439ad9c49cSJason Wang int ret; 11449ad9c49cSJason Wang 11459ad9c49cSJason Wang ret = iotlb_translate(vrh, (u64)(uintptr_t)dst, 11469ad9c49cSJason Wang len, iov, 16, VHOST_MAP_WO); 11479ad9c49cSJason Wang if (ret < 0) 11489ad9c49cSJason Wang return ret; 11499ad9c49cSJason Wang 11509ad9c49cSJason Wang iov_iter_bvec(&iter, WRITE, iov, ret, len); 11519ad9c49cSJason Wang 11529ad9c49cSJason Wang return copy_to_iter(src, len, &iter); 11539ad9c49cSJason Wang } 11549ad9c49cSJason Wang 11559ad9c49cSJason Wang static inline int getu16_iotlb(const struct vringh *vrh, 11569ad9c49cSJason Wang u16 *val, const __virtio16 *p) 11579ad9c49cSJason Wang { 11589ad9c49cSJason Wang struct bio_vec iov; 11599ad9c49cSJason Wang void *kaddr, *from; 11609ad9c49cSJason Wang int ret; 11619ad9c49cSJason Wang 11629ad9c49cSJason Wang /* Atomic read is needed for getu16 */ 11639ad9c49cSJason Wang ret = iotlb_translate(vrh, (u64)(uintptr_t)p, sizeof(*p), 11649ad9c49cSJason Wang &iov, 1, VHOST_MAP_RO); 11659ad9c49cSJason Wang if (ret < 0) 11669ad9c49cSJason Wang return ret; 11679ad9c49cSJason Wang 11689ad9c49cSJason Wang kaddr = kmap_atomic(iov.bv_page); 11699ad9c49cSJason Wang from = kaddr + iov.bv_offset; 11709ad9c49cSJason Wang *val = vringh16_to_cpu(vrh, READ_ONCE(*(__virtio16 *)from)); 11719ad9c49cSJason Wang kunmap_atomic(kaddr); 11729ad9c49cSJason Wang 11739ad9c49cSJason Wang return 0; 11749ad9c49cSJason Wang } 11759ad9c49cSJason Wang 11769ad9c49cSJason Wang static inline int putu16_iotlb(const struct vringh *vrh, 11779ad9c49cSJason Wang __virtio16 *p, u16 val) 11789ad9c49cSJason Wang { 11799ad9c49cSJason Wang struct bio_vec iov; 11809ad9c49cSJason Wang void *kaddr, *to; 11819ad9c49cSJason Wang int ret; 11829ad9c49cSJason Wang 11839ad9c49cSJason Wang /* Atomic write is needed for putu16 */ 11849ad9c49cSJason Wang ret = iotlb_translate(vrh, (u64)(uintptr_t)p, sizeof(*p), 11859ad9c49cSJason Wang &iov, 1, VHOST_MAP_WO); 11869ad9c49cSJason Wang if (ret < 0) 11879ad9c49cSJason Wang return ret; 11889ad9c49cSJason Wang 11899ad9c49cSJason Wang kaddr = kmap_atomic(iov.bv_page); 11909ad9c49cSJason Wang to = kaddr + iov.bv_offset; 11919ad9c49cSJason Wang WRITE_ONCE(*(__virtio16 *)to, cpu_to_vringh16(vrh, val)); 11929ad9c49cSJason Wang kunmap_atomic(kaddr); 11939ad9c49cSJason Wang 11949ad9c49cSJason Wang return 0; 11959ad9c49cSJason Wang } 11969ad9c49cSJason Wang 11979ad9c49cSJason Wang static inline int copydesc_iotlb(const struct vringh *vrh, 11989ad9c49cSJason Wang void *dst, const void *src, size_t len) 11999ad9c49cSJason Wang { 12009ad9c49cSJason Wang int ret; 12019ad9c49cSJason Wang 12029ad9c49cSJason Wang ret = copy_from_iotlb(vrh, dst, (void *)src, len); 12039ad9c49cSJason Wang if (ret != len) 12049ad9c49cSJason Wang return -EFAULT; 12059ad9c49cSJason Wang 12069ad9c49cSJason Wang return 0; 12079ad9c49cSJason Wang } 12089ad9c49cSJason Wang 12099ad9c49cSJason Wang static inline int xfer_from_iotlb(const struct vringh *vrh, void *src, 12109ad9c49cSJason Wang void *dst, size_t len) 12119ad9c49cSJason Wang { 12129ad9c49cSJason Wang int ret; 12139ad9c49cSJason Wang 12149ad9c49cSJason Wang ret = copy_from_iotlb(vrh, dst, src, len); 12159ad9c49cSJason Wang if (ret != len) 12169ad9c49cSJason Wang return -EFAULT; 12179ad9c49cSJason Wang 12189ad9c49cSJason Wang return 0; 12199ad9c49cSJason Wang } 12209ad9c49cSJason Wang 12219ad9c49cSJason Wang static inline int xfer_to_iotlb(const struct vringh *vrh, 12229ad9c49cSJason Wang void *dst, void *src, size_t len) 12239ad9c49cSJason Wang { 12249ad9c49cSJason Wang int ret; 12259ad9c49cSJason Wang 12269ad9c49cSJason Wang ret = copy_to_iotlb(vrh, dst, src, len); 12279ad9c49cSJason Wang if (ret != len) 12289ad9c49cSJason Wang return -EFAULT; 12299ad9c49cSJason Wang 12309ad9c49cSJason Wang return 0; 12319ad9c49cSJason Wang } 12329ad9c49cSJason Wang 12339ad9c49cSJason Wang static inline int putused_iotlb(const struct vringh *vrh, 12349ad9c49cSJason Wang struct vring_used_elem *dst, 12359ad9c49cSJason Wang const struct vring_used_elem *src, 12369ad9c49cSJason Wang unsigned int num) 12379ad9c49cSJason Wang { 12389ad9c49cSJason Wang int size = num * sizeof(*dst); 12399ad9c49cSJason Wang int ret; 12409ad9c49cSJason Wang 12419ad9c49cSJason Wang ret = copy_to_iotlb(vrh, dst, (void *)src, num * sizeof(*dst)); 12429ad9c49cSJason Wang if (ret != size) 12439ad9c49cSJason Wang return -EFAULT; 12449ad9c49cSJason Wang 12459ad9c49cSJason Wang return 0; 12469ad9c49cSJason Wang } 12479ad9c49cSJason Wang 12489ad9c49cSJason Wang /** 12499ad9c49cSJason Wang * vringh_init_iotlb - initialize a vringh for a ring with IOTLB. 12509ad9c49cSJason Wang * @vrh: the vringh to initialize. 12519ad9c49cSJason Wang * @features: the feature bits for this ring. 12529ad9c49cSJason Wang * @num: the number of elements. 12539ad9c49cSJason Wang * @weak_barriers: true if we only need memory barriers, not I/O. 12549ad9c49cSJason Wang * @desc: the userpace descriptor pointer. 12559ad9c49cSJason Wang * @avail: the userpace avail pointer. 12569ad9c49cSJason Wang * @used: the userpace used pointer. 12579ad9c49cSJason Wang * 12589ad9c49cSJason Wang * Returns an error if num is invalid. 12599ad9c49cSJason Wang */ 12609ad9c49cSJason Wang int vringh_init_iotlb(struct vringh *vrh, u64 features, 12619ad9c49cSJason Wang unsigned int num, bool weak_barriers, 12629ad9c49cSJason Wang struct vring_desc *desc, 12639ad9c49cSJason Wang struct vring_avail *avail, 12649ad9c49cSJason Wang struct vring_used *used) 12659ad9c49cSJason Wang { 12669ad9c49cSJason Wang return vringh_init_kern(vrh, features, num, weak_barriers, 12679ad9c49cSJason Wang desc, avail, used); 12689ad9c49cSJason Wang } 12699ad9c49cSJason Wang EXPORT_SYMBOL(vringh_init_iotlb); 12709ad9c49cSJason Wang 12719ad9c49cSJason Wang /** 12729ad9c49cSJason Wang * vringh_set_iotlb - initialize a vringh for a ring with IOTLB. 12739ad9c49cSJason Wang * @vrh: the vring 12749ad9c49cSJason Wang * @iotlb: iotlb associated with this vring 1275f53d9910SStefano Garzarella * @iotlb_lock: spinlock to synchronize the iotlb accesses 12769ad9c49cSJason Wang */ 1277f53d9910SStefano Garzarella void vringh_set_iotlb(struct vringh *vrh, struct vhost_iotlb *iotlb, 1278f53d9910SStefano Garzarella spinlock_t *iotlb_lock) 12799ad9c49cSJason Wang { 12809ad9c49cSJason Wang vrh->iotlb = iotlb; 1281f53d9910SStefano Garzarella vrh->iotlb_lock = iotlb_lock; 12829ad9c49cSJason Wang } 12839ad9c49cSJason Wang EXPORT_SYMBOL(vringh_set_iotlb); 12849ad9c49cSJason Wang 12859ad9c49cSJason Wang /** 12869ad9c49cSJason Wang * vringh_getdesc_iotlb - get next available descriptor from ring with 12879ad9c49cSJason Wang * IOTLB. 12889ad9c49cSJason Wang * @vrh: the kernelspace vring. 12899ad9c49cSJason Wang * @riov: where to put the readable descriptors (or NULL) 12909ad9c49cSJason Wang * @wiov: where to put the writable descriptors (or NULL) 12919ad9c49cSJason Wang * @head: head index we received, for passing to vringh_complete_iotlb(). 12929ad9c49cSJason Wang * @gfp: flags for allocating larger riov/wiov. 12939ad9c49cSJason Wang * 12949ad9c49cSJason Wang * Returns 0 if there was no descriptor, 1 if there was, or -errno. 12959ad9c49cSJason Wang * 12969ad9c49cSJason Wang * Note that on error return, you can tell the difference between an 12979ad9c49cSJason Wang * invalid ring and a single invalid descriptor: in the former case, 12989ad9c49cSJason Wang * *head will be vrh->vring.num. You may be able to ignore an invalid 12999ad9c49cSJason Wang * descriptor, but there's not much you can do with an invalid ring. 13009ad9c49cSJason Wang * 1301*69c13c58SStefano Garzarella * Note that you can reuse riov and wiov with subsequent calls. Content is 1302*69c13c58SStefano Garzarella * overwritten and memory reallocated if more space is needed. 1303*69c13c58SStefano Garzarella * When you don't have to use riov and wiov anymore, you should clean up them 1304*69c13c58SStefano Garzarella * calling vringh_kiov_cleanup() to release the memory, even on error! 13059ad9c49cSJason Wang */ 13069ad9c49cSJason Wang int vringh_getdesc_iotlb(struct vringh *vrh, 13079ad9c49cSJason Wang struct vringh_kiov *riov, 13089ad9c49cSJason Wang struct vringh_kiov *wiov, 13099ad9c49cSJason Wang u16 *head, 13109ad9c49cSJason Wang gfp_t gfp) 13119ad9c49cSJason Wang { 13129ad9c49cSJason Wang int err; 13139ad9c49cSJason Wang 13149ad9c49cSJason Wang err = __vringh_get_head(vrh, getu16_iotlb, &vrh->last_avail_idx); 13159ad9c49cSJason Wang if (err < 0) 13169ad9c49cSJason Wang return err; 13179ad9c49cSJason Wang 13189ad9c49cSJason Wang /* Empty... */ 13199ad9c49cSJason Wang if (err == vrh->vring.num) 13209ad9c49cSJason Wang return 0; 13219ad9c49cSJason Wang 13229ad9c49cSJason Wang *head = err; 13239ad9c49cSJason Wang err = __vringh_iov(vrh, *head, riov, wiov, no_range_check, NULL, 13249ad9c49cSJason Wang gfp, copydesc_iotlb); 13259ad9c49cSJason Wang if (err) 13269ad9c49cSJason Wang return err; 13279ad9c49cSJason Wang 13289ad9c49cSJason Wang return 1; 13299ad9c49cSJason Wang } 13309ad9c49cSJason Wang EXPORT_SYMBOL(vringh_getdesc_iotlb); 13319ad9c49cSJason Wang 13329ad9c49cSJason Wang /** 13339ad9c49cSJason Wang * vringh_iov_pull_iotlb - copy bytes from vring_iov. 13349ad9c49cSJason Wang * @vrh: the vring. 13359ad9c49cSJason Wang * @riov: the riov as passed to vringh_getdesc_iotlb() (updated as we consume) 13369ad9c49cSJason Wang * @dst: the place to copy. 13379ad9c49cSJason Wang * @len: the maximum length to copy. 13389ad9c49cSJason Wang * 13399ad9c49cSJason Wang * Returns the bytes copied <= len or a negative errno. 13409ad9c49cSJason Wang */ 13419ad9c49cSJason Wang ssize_t vringh_iov_pull_iotlb(struct vringh *vrh, 13429ad9c49cSJason Wang struct vringh_kiov *riov, 13439ad9c49cSJason Wang void *dst, size_t len) 13449ad9c49cSJason Wang { 13459ad9c49cSJason Wang return vringh_iov_xfer(vrh, riov, dst, len, xfer_from_iotlb); 13469ad9c49cSJason Wang } 13479ad9c49cSJason Wang EXPORT_SYMBOL(vringh_iov_pull_iotlb); 13489ad9c49cSJason Wang 13499ad9c49cSJason Wang /** 13509ad9c49cSJason Wang * vringh_iov_push_iotlb - copy bytes into vring_iov. 13519ad9c49cSJason Wang * @vrh: the vring. 13529ad9c49cSJason Wang * @wiov: the wiov as passed to vringh_getdesc_iotlb() (updated as we consume) 13538009b0f4SStefano Garzarella * @src: the place to copy from. 13549ad9c49cSJason Wang * @len: the maximum length to copy. 13559ad9c49cSJason Wang * 13569ad9c49cSJason Wang * Returns the bytes copied <= len or a negative errno. 13579ad9c49cSJason Wang */ 13589ad9c49cSJason Wang ssize_t vringh_iov_push_iotlb(struct vringh *vrh, 13599ad9c49cSJason Wang struct vringh_kiov *wiov, 13609ad9c49cSJason Wang const void *src, size_t len) 13619ad9c49cSJason Wang { 13629ad9c49cSJason Wang return vringh_iov_xfer(vrh, wiov, (void *)src, len, xfer_to_iotlb); 13639ad9c49cSJason Wang } 13649ad9c49cSJason Wang EXPORT_SYMBOL(vringh_iov_push_iotlb); 13659ad9c49cSJason Wang 13669ad9c49cSJason Wang /** 13679ad9c49cSJason Wang * vringh_abandon_iotlb - we've decided not to handle the descriptor(s). 13689ad9c49cSJason Wang * @vrh: the vring. 13699ad9c49cSJason Wang * @num: the number of descriptors to put back (ie. num 13709ad9c49cSJason Wang * vringh_get_iotlb() to undo). 13719ad9c49cSJason Wang * 13729ad9c49cSJason Wang * The next vringh_get_iotlb() will return the old descriptor(s) again. 13739ad9c49cSJason Wang */ 13749ad9c49cSJason Wang void vringh_abandon_iotlb(struct vringh *vrh, unsigned int num) 13759ad9c49cSJason Wang { 13769ad9c49cSJason Wang /* We only update vring_avail_event(vr) when we want to be notified, 13779ad9c49cSJason Wang * so we haven't changed that yet. 13789ad9c49cSJason Wang */ 13799ad9c49cSJason Wang vrh->last_avail_idx -= num; 13809ad9c49cSJason Wang } 13819ad9c49cSJason Wang EXPORT_SYMBOL(vringh_abandon_iotlb); 13829ad9c49cSJason Wang 13839ad9c49cSJason Wang /** 13849ad9c49cSJason Wang * vringh_complete_iotlb - we've finished with descriptor, publish it. 13859ad9c49cSJason Wang * @vrh: the vring. 13869ad9c49cSJason Wang * @head: the head as filled in by vringh_getdesc_iotlb. 13879ad9c49cSJason Wang * @len: the length of data we have written. 13889ad9c49cSJason Wang * 13899ad9c49cSJason Wang * You should check vringh_need_notify_iotlb() after one or more calls 13909ad9c49cSJason Wang * to this function. 13919ad9c49cSJason Wang */ 13929ad9c49cSJason Wang int vringh_complete_iotlb(struct vringh *vrh, u16 head, u32 len) 13939ad9c49cSJason Wang { 13949ad9c49cSJason Wang struct vring_used_elem used; 13959ad9c49cSJason Wang 13969ad9c49cSJason Wang used.id = cpu_to_vringh32(vrh, head); 13979ad9c49cSJason Wang used.len = cpu_to_vringh32(vrh, len); 13989ad9c49cSJason Wang 13999ad9c49cSJason Wang return __vringh_complete(vrh, &used, 1, putu16_iotlb, putused_iotlb); 14009ad9c49cSJason Wang } 14019ad9c49cSJason Wang EXPORT_SYMBOL(vringh_complete_iotlb); 14029ad9c49cSJason Wang 14039ad9c49cSJason Wang /** 14049ad9c49cSJason Wang * vringh_notify_enable_iotlb - we want to know if something changes. 14059ad9c49cSJason Wang * @vrh: the vring. 14069ad9c49cSJason Wang * 14079ad9c49cSJason Wang * This always enables notifications, but returns false if there are 14089ad9c49cSJason Wang * now more buffers available in the vring. 14099ad9c49cSJason Wang */ 14109ad9c49cSJason Wang bool vringh_notify_enable_iotlb(struct vringh *vrh) 14119ad9c49cSJason Wang { 14129ad9c49cSJason Wang return __vringh_notify_enable(vrh, getu16_iotlb, putu16_iotlb); 14139ad9c49cSJason Wang } 14149ad9c49cSJason Wang EXPORT_SYMBOL(vringh_notify_enable_iotlb); 14159ad9c49cSJason Wang 14169ad9c49cSJason Wang /** 14179ad9c49cSJason Wang * vringh_notify_disable_iotlb - don't tell us if something changes. 14189ad9c49cSJason Wang * @vrh: the vring. 14199ad9c49cSJason Wang * 14209ad9c49cSJason Wang * This is our normal running state: we disable and then only enable when 14219ad9c49cSJason Wang * we're going to sleep. 14229ad9c49cSJason Wang */ 14239ad9c49cSJason Wang void vringh_notify_disable_iotlb(struct vringh *vrh) 14249ad9c49cSJason Wang { 14259ad9c49cSJason Wang __vringh_notify_disable(vrh, putu16_iotlb); 14269ad9c49cSJason Wang } 14279ad9c49cSJason Wang EXPORT_SYMBOL(vringh_notify_disable_iotlb); 14289ad9c49cSJason Wang 14299ad9c49cSJason Wang /** 14309ad9c49cSJason Wang * vringh_need_notify_iotlb - must we tell the other side about used buffers? 14319ad9c49cSJason Wang * @vrh: the vring we've called vringh_complete_iotlb() on. 14329ad9c49cSJason Wang * 14339ad9c49cSJason Wang * Returns -errno or 0 if we don't need to tell the other side, 1 if we do. 14349ad9c49cSJason Wang */ 14359ad9c49cSJason Wang int vringh_need_notify_iotlb(struct vringh *vrh) 14369ad9c49cSJason Wang { 14379ad9c49cSJason Wang return __vringh_need_notify(vrh, getu16_iotlb); 14389ad9c49cSJason Wang } 14399ad9c49cSJason Wang EXPORT_SYMBOL(vringh_need_notify_iotlb); 14409ad9c49cSJason Wang 14413302363aSMichael S. Tsirkin #endif 14429ad9c49cSJason Wang 1443f558a845SDave Jones MODULE_LICENSE("GPL"); 1444