drivers/hv/ring_buffer.c

   1 /*
   2  *
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  16  * Place - Suite 330, Boston, MA 02111-1307 USA.
  17  *
  18  * Authors:
  19  *   Haiyang Zhang <haiyangz@microsoft.com>
  20  *   Hank Janssen  <hjanssen@microsoft.com>
  21  *   K. Y. Srinivasan <kys@microsoft.com>
  22  *
  23  */
  24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/kernel.h>
  27 #include <linux/mm.h>
  28 #include <linux/hyperv.h>
  29 #include <linux/uio.h>
  30 #include <linux/vmalloc.h>
  31 #include <linux/slab.h>
  32
  33 #include "hyperv_vmbus.h"
  34
  35 #define VMBUS_PKT_TRAILER       8
  36
  37 /*
  38  * When we write to the ring buffer, check if the host needs to
  39  * be signaled. Here is the details of this protocol:
  40  *
  41  *      1. The host guarantees that while it is draining the
  42  *         ring buffer, it will set the interrupt_mask to
  43  *         indicate it does not need to be interrupted when
  44  *         new data is placed.
  45  *
  46  *      2. The host guarantees that it will completely drain
  47  *         the ring buffer before exiting the read loop. Further,
  48  *         once the ring buffer is empty, it will clear the
  49  *         interrupt_mask and re-check to see if new data has
  50  *         arrived.
  51  *
  52  * KYS: Oct. 30, 2016:
  53  * It looks like Windows hosts have logic to deal with DOS attacks that
  54  * can be triggered if it receives interrupts when it is not expecting
  55  * the interrupt. The host expects interrupts only when the ring
  56  * transitions from empty to non-empty (or full to non full on the guest
  57  * to host ring).
  58  * So, base the signaling decision solely on the ring state until the
  59  * host logic is fixed.
  60  */
  61
  62 static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)
  63 {
  64         struct hv_ring_buffer_info *rbi = &channel->outbound;
  65
  66         virt_mb();
  67         if (READ_ONCE(rbi->ring_buffer->interrupt_mask))
  68                 return;
  69
  70         /* check interrupt_mask before read_index */
  71         virt_rmb();
  72         /*
  73          * This is the only case we need to signal when the
  74          * ring transitions from being empty to non-empty.
  75          */
  76         if (old_write == READ_ONCE(rbi->ring_buffer->read_index))
  77                 vmbus_setevent(channel);
  78
  79         return;
  80 }
  81
  82 /* Get the next write location for the specified ring buffer. */
  83 static inline u32
  84 hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
  85 {
  86         u32 next = ring_info->ring_buffer->write_index;
  87
  88         return next;
  89 }
  90
  91 /* Set the next write location for the specified ring buffer. */
  92 static inline void
  93 hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
  94                      u32 next_write_location)
  95 {
  96         ring_info->ring_buffer->write_index = next_write_location;
  97 }
  98
  99 /* Get the next read location for the specified ring buffer. */
 100 static inline u32
 101 hv_get_next_read_location(const struct hv_ring_buffer_info *ring_info)
 102 {
 103         return ring_info->ring_buffer->read_index;
 104 }
 105
 106 /*
 107  * Get the next read location + offset for the specified ring buffer.
 108  * This allows the caller to skip.
 109  */
 110 static inline u32
 111 hv_get_next_readlocation_withoffset(const struct hv_ring_buffer_info *ring_info,
 112                                     u32 offset)
 113 {
 114         u32 next = ring_info->ring_buffer->read_index;
 115
 116         next += offset;
 117         if (next >= ring_info->ring_datasize)
 118                 next -= ring_info->ring_datasize;
 119
 120         return next;
 121 }
 122
 123 /* Set the next read location for the specified ring buffer. */
 124 static inline void
 125 hv_set_next_read_location(struct hv_ring_buffer_info *ring_info,
 126                     u32 next_read_location)
 127 {
 128         ring_info->ring_buffer->read_index = next_read_location;
 129         ring_info->priv_read_index = next_read_location;
 130 }
 131
 132 /* Get the size of the ring buffer. */
 133 static inline u32
 134 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
 135 {
 136         return ring_info->ring_datasize;
 137 }
 138
 139 /* Get the read and write indices as u64 of the specified ring buffer. */
 140 static inline u64
 141 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
 142 {
 143         return (u64)ring_info->ring_buffer->write_index << 32;
 144 }
 145
 146 /*
 147  * Helper routine to copy to source from ring buffer.
 148  * Assume there is enough room. Handles wrap-around in src case only!!
 149  */
 150 static u32 hv_copyfrom_ringbuffer(
 151         const struct hv_ring_buffer_info *ring_info,
 152         void                            *dest,
 153         u32                             destlen,
 154         u32                             start_read_offset)
 155 {
 156         void *ring_buffer = hv_get_ring_buffer(ring_info);
 157         u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
 158
 159         memcpy(dest, ring_buffer + start_read_offset, destlen);
 160
 161         start_read_offset += destlen;
 162         if (start_read_offset >= ring_buffer_size)
 163                 start_read_offset -= ring_buffer_size;
 164
 165         return start_read_offset;
 166 }
 167
 168
 169 /*
 170  * Helper routine to copy from source to ring buffer.
 171  * Assume there is enough room. Handles wrap-around in dest case only!!
 172  */
 173 static u32 hv_copyto_ringbuffer(
 174         struct hv_ring_buffer_info      *ring_info,
 175         u32                             start_write_offset,
 176         const void                      *src,
 177         u32                             srclen)
 178 {
 179         void *ring_buffer = hv_get_ring_buffer(ring_info);
 180         u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
 181
 182         memcpy(ring_buffer + start_write_offset, src, srclen);
 183
 184         start_write_offset += srclen;
 185         if (start_write_offset >= ring_buffer_size)
 186                 start_write_offset -= ring_buffer_size;
 187
 188         return start_write_offset;
 189 }
 190
 191 /* Get various debug metrics for the specified ring buffer. */
 192 void hv_ringbuffer_get_debuginfo(const struct hv_ring_buffer_info *ring_info,
 193                                  struct hv_ring_buffer_debug_info *debug_info)
 194 {
 195         u32 bytes_avail_towrite;
 196         u32 bytes_avail_toread;
 197
 198         if (ring_info->ring_buffer) {
 199                 hv_get_ringbuffer_availbytes(ring_info,
 200                                         &bytes_avail_toread,
 201                                         &bytes_avail_towrite);
 202
 203                 debug_info->bytes_avail_toread = bytes_avail_toread;
 204                 debug_info->bytes_avail_towrite = bytes_avail_towrite;
 205                 debug_info->current_read_index =
 206                         ring_info->ring_buffer->read_index;
 207                 debug_info->current_write_index =
 208                         ring_info->ring_buffer->write_index;
 209                 debug_info->current_interrupt_mask =
 210                         ring_info->ring_buffer->interrupt_mask;
 211         }
 212 }
 213
 214 /* Initialize the ring buffer. */
 215 int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
 216                        struct page *pages, u32 page_cnt)
 217 {
 218         int i;
 219         struct page **pages_wraparound;
 220
 221         BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
 222
 223         memset(ring_info, 0, sizeof(struct hv_ring_buffer_info));
 224
 225         /*
 226          * First page holds struct hv_ring_buffer, do wraparound mapping for
 227          * the rest.
 228          */
 229         pages_wraparound = kzalloc(sizeof(struct page *) * (page_cnt * 2 - 1),
 230                                    GFP_KERNEL);
 231         if (!pages_wraparound)
 232                 return -ENOMEM;
 233
 234         pages_wraparound[0] = pages;
 235         for (i = 0; i < 2 * (page_cnt - 1); i++)
 236                 pages_wraparound[i + 1] = &pages[i % (page_cnt - 1) + 1];
 237
 238         ring_info->ring_buffer = (struct hv_ring_buffer *)
 239                 vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP, PAGE_KERNEL);
 240
 241         kfree(pages_wraparound);
 242
 243
 244         if (!ring_info->ring_buffer)
 245                 return -ENOMEM;
 246
 247         ring_info->ring_buffer->read_index =
 248                 ring_info->ring_buffer->write_index = 0;
 249
 250         /* Set the feature bit for enabling flow control. */
 251         ring_info->ring_buffer->feature_bits.value = 1;
 252
 253         ring_info->ring_size = page_cnt << PAGE_SHIFT;
 254         ring_info->ring_datasize = ring_info->ring_size -
 255                 sizeof(struct hv_ring_buffer);
 256
 257         spin_lock_init(&ring_info->ring_lock);
 258
 259         return 0;
 260 }
 261
 262 /* Cleanup the ring buffer. */
 263 void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
 264 {
 265         vunmap(ring_info->ring_buffer);
 266 }
 267
 268 /* Write to the ring buffer. */
 269 int hv_ringbuffer_write(struct vmbus_channel *channel,
 270                         const struct kvec *kv_list, u32 kv_count)
 271 {
 272         int i = 0;
 273         u32 bytes_avail_towrite;
 274         u32 totalbytes_towrite = 0;
 275
 276         u32 next_write_location;
 277         u32 old_write;
 278         u64 prev_indices = 0;
 279         unsigned long flags = 0;
 280         struct hv_ring_buffer_info *outring_info = &channel->outbound;
 281
 282         if (channel->rescind)
 283                 return -ENODEV;
 284
 285         for (i = 0; i < kv_count; i++)
 286                 totalbytes_towrite += kv_list[i].iov_len;
 287
 288         totalbytes_towrite += sizeof(u64);
 289
 290         spin_lock_irqsave(&outring_info->ring_lock, flags);
 291
 292         bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
 293
 294         /*
 295          * If there is only room for the packet, assume it is full.
 296          * Otherwise, the next time around, we think the ring buffer
 297          * is empty since the read index == write index.
 298          */
 299         if (bytes_avail_towrite <= totalbytes_towrite) {
 300                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 301                 return -EAGAIN;
 302         }
 303
 304         /* Write to the ring buffer */
 305         next_write_location = hv_get_next_write_location(outring_info);
 306
 307         old_write = next_write_location;
 308
 309         for (i = 0; i < kv_count; i++) {
 310                 next_write_location = hv_copyto_ringbuffer(outring_info,
 311                                                      next_write_location,
 312                                                      kv_list[i].iov_base,
 313                                                      kv_list[i].iov_len);
 314         }
 315
 316         /* Set previous packet start */
 317         prev_indices = hv_get_ring_bufferindices(outring_info);
 318
 319         next_write_location = hv_copyto_ringbuffer(outring_info,
 320                                              next_write_location,
 321                                              &prev_indices,
 322                                              sizeof(u64));
 323
 324         /* Issue a full memory barrier before updating the write index */
 325         virt_mb();
 326
 327         /* Now, update the write location */
 328         hv_set_next_write_location(outring_info, next_write_location);
 329
 330
 331         spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 332
 333         hv_signal_on_write(old_write, channel);
 334
 335         if (channel->rescind)
 336                 return -ENODEV;
 337
 338         return 0;
 339 }
 340
 341 static inline void
 342 init_cached_read_index(struct hv_ring_buffer_info *rbi)
 343 {
 344         rbi->cached_read_index = rbi->ring_buffer->read_index;
 345 }
 346
 347 int hv_ringbuffer_read(struct vmbus_channel *channel,
 348                        void *buffer, u32 buflen, u32 *buffer_actual_len,
 349                        u64 *requestid, bool raw)
 350 {
 351         u32 bytes_avail_toread;
 352         u32 next_read_location = 0;
 353         u64 prev_indices = 0;
 354         struct vmpacket_descriptor desc;
 355         u32 offset;
 356         u32 packetlen;
 357         int ret = 0;
 358         struct hv_ring_buffer_info *inring_info = &channel->inbound;
 359
 360         if (buflen <= 0)
 361                 return -EINVAL;
 362
 363
 364         *buffer_actual_len = 0;
 365         *requestid = 0;
 366
 367         bytes_avail_toread = hv_get_bytes_to_read(inring_info);
 368         /* Make sure there is something to read */
 369         if (bytes_avail_toread < sizeof(desc)) {
 370                 /*
 371                  * No error is set when there is even no header, drivers are
 372                  * supposed to analyze buffer_actual_len.
 373                  */
 374                 return ret;
 375         }
 376
 377         init_cached_read_index(inring_info);
 378
 379         next_read_location = hv_get_next_read_location(inring_info);
 380         next_read_location = hv_copyfrom_ringbuffer(inring_info, &desc,
 381                                                     sizeof(desc),
 382                                                     next_read_location);
 383
 384         offset = raw ? 0 : (desc.offset8 << 3);
 385         packetlen = (desc.len8 << 3) - offset;
 386         *buffer_actual_len = packetlen;
 387         *requestid = desc.trans_id;
 388
 389         if (bytes_avail_toread < packetlen + offset)
 390                 return -EAGAIN;
 391
 392         if (packetlen > buflen)
 393                 return -ENOBUFS;
 394
 395         next_read_location =
 396                 hv_get_next_readlocation_withoffset(inring_info, offset);
 397
 398         next_read_location = hv_copyfrom_ringbuffer(inring_info,
 399                                                 buffer,
 400                                                 packetlen,
 401                                                 next_read_location);
 402
 403         next_read_location = hv_copyfrom_ringbuffer(inring_info,
 404                                                 &prev_indices,
 405                                                 sizeof(u64),
 406                                                 next_read_location);
 407
 408         /*
 409          * Make sure all reads are done before we update the read index since
 410          * the writer may start writing to the read area once the read index
 411          * is updated.
 412          */
 413         virt_mb();
 414
 415         /* Update the read index */
 416         hv_set_next_read_location(inring_info, next_read_location);
 417
 418         hv_signal_on_read(channel);
 419
 420         return ret;
 421 }
 422
 423 /*
 424  * Determine number of bytes available in ring buffer after
 425  * the current iterator (priv_read_index) location.
 426  *
 427  * This is similar to hv_get_bytes_to_read but with private
 428  * read index instead.
 429  */
 430 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
 431 {
 432         u32 priv_read_loc = rbi->priv_read_index;
 433         u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
 434
 435         if (write_loc >= priv_read_loc)
 436                 return write_loc - priv_read_loc;
 437         else
 438                 return (rbi->ring_datasize - priv_read_loc) + write_loc;
 439 }
 440
 441 /*
 442  * Get first vmbus packet from ring buffer after read_index
 443  *
 444  * If ring buffer is empty, returns NULL and no other action needed.
 445  */
 446 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
 447 {
 448         struct hv_ring_buffer_info *rbi = &channel->inbound;
 449
 450         /* set state for later hv_signal_on_read() */
 451         init_cached_read_index(rbi);
 452
 453         if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
 454                 return NULL;
 455
 456         return hv_get_ring_buffer(rbi) + rbi->priv_read_index;
 457 }
 458 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
 459
 460 /*
 461  * Get next vmbus packet from ring buffer.
 462  *
 463  * Advances the current location (priv_read_index) and checks for more
 464  * data. If the end of the ring buffer is reached, then return NULL.
 465  */
 466 struct vmpacket_descriptor *
 467 __hv_pkt_iter_next(struct vmbus_channel *channel,
 468                    const struct vmpacket_descriptor *desc)
 469 {
 470         struct hv_ring_buffer_info *rbi = &channel->inbound;
 471         u32 packetlen = desc->len8 << 3;
 472         u32 dsize = rbi->ring_datasize;
 473
 474         /* bump offset to next potential packet */
 475         rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
 476         if (rbi->priv_read_index >= dsize)
 477                 rbi->priv_read_index -= dsize;
 478
 479         /* more data? */
 480         if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
 481                 return NULL;
 482         else
 483                 return hv_get_ring_buffer(rbi) + rbi->priv_read_index;
 484 }
 485 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
 486
 487 /*
 488  * Update host ring buffer after iterating over packets.
 489  */
 490 void hv_pkt_iter_close(struct vmbus_channel *channel)
 491 {
 492         struct hv_ring_buffer_info *rbi = &channel->inbound;
 493
 494         /*
 495          * Make sure all reads are done before we update the read index since
 496          * the writer may start writing to the read area once the read index
 497          * is updated.
 498          */
 499         virt_rmb();
 500         rbi->ring_buffer->read_index = rbi->priv_read_index;
 501
 502         hv_signal_on_read(channel);
 503 }
 504 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);