drivers/hv/ring_buffer.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *
   4  * Copyright (c) 2009, Microsoft Corporation.
   5  *
   6  * Authors:
   7  *   Haiyang Zhang <haiyangz@microsoft.com>
   8  *   Hank Janssen  <hjanssen@microsoft.com>
   9  *   K. Y. Srinivasan <kys@microsoft.com>
  10  */
  11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  12
  13 #include <linux/kernel.h>
  14 #include <linux/mm.h>
  15 #include <linux/hyperv.h>
  16 #include <linux/uio.h>
  17 #include <linux/vmalloc.h>
  18 #include <linux/slab.h>
  19 #include <linux/prefetch.h>
  20
  21 #include "hyperv_vmbus.h"
  22
  23 #define VMBUS_PKT_TRAILER       8
  24
  25 /*
  26  * When we write to the ring buffer, check if the host needs to
  27  * be signaled. Here is the details of this protocol:
  28  *
  29  *      1. The host guarantees that while it is draining the
  30  *         ring buffer, it will set the interrupt_mask to
  31  *         indicate it does not need to be interrupted when
  32  *         new data is placed.
  33  *
  34  *      2. The host guarantees that it will completely drain
  35  *         the ring buffer before exiting the read loop. Further,
  36  *         once the ring buffer is empty, it will clear the
  37  *         interrupt_mask and re-check to see if new data has
  38  *         arrived.
  39  *
  40  * KYS: Oct. 30, 2016:
  41  * It looks like Windows hosts have logic to deal with DOS attacks that
  42  * can be triggered if it receives interrupts when it is not expecting
  43  * the interrupt. The host expects interrupts only when the ring
  44  * transitions from empty to non-empty (or full to non full on the guest
  45  * to host ring).
  46  * So, base the signaling decision solely on the ring state until the
  47  * host logic is fixed.
  48  */
  49
  50 static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)
  51 {
  52         struct hv_ring_buffer_info *rbi = &channel->outbound;
  53
  54         virt_mb();
  55         if (READ_ONCE(rbi->ring_buffer->interrupt_mask))
  56                 return;
  57
  58         /* check interrupt_mask before read_index */
  59         virt_rmb();
  60         /*
  61          * This is the only case we need to signal when the
  62          * ring transitions from being empty to non-empty.
  63          */
  64         if (old_write == READ_ONCE(rbi->ring_buffer->read_index)) {
  65                 ++channel->intr_out_empty;
  66                 vmbus_setevent(channel);
  67         }
  68 }
  69
  70 /* Get the next write location for the specified ring buffer. */
  71 static inline u32
  72 hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
  73 {
  74         u32 next = ring_info->ring_buffer->write_index;
  75
  76         return next;
  77 }
  78
  79 /* Set the next write location for the specified ring buffer. */
  80 static inline void
  81 hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
  82                      u32 next_write_location)
  83 {
  84         ring_info->ring_buffer->write_index = next_write_location;
  85 }
  86
  87 /* Set the next read location for the specified ring buffer. */
  88 static inline void
  89 hv_set_next_read_location(struct hv_ring_buffer_info *ring_info,
  90                     u32 next_read_location)
  91 {
  92         ring_info->ring_buffer->read_index = next_read_location;
  93         ring_info->priv_read_index = next_read_location;
  94 }
  95
  96 /* Get the size of the ring buffer. */
  97 static inline u32
  98 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
  99 {
 100         return ring_info->ring_datasize;
 101 }
 102
 103 /* Get the read and write indices as u64 of the specified ring buffer. */
 104 static inline u64
 105 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
 106 {
 107         return (u64)ring_info->ring_buffer->write_index << 32;
 108 }
 109
 110 /*
 111  * Helper routine to copy from source to ring buffer.
 112  * Assume there is enough room. Handles wrap-around in dest case only!!
 113  */
 114 static u32 hv_copyto_ringbuffer(
 115         struct hv_ring_buffer_info      *ring_info,
 116         u32                             start_write_offset,
 117         const void                      *src,
 118         u32                             srclen)
 119 {
 120         void *ring_buffer = hv_get_ring_buffer(ring_info);
 121         u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
 122
 123         memcpy(ring_buffer + start_write_offset, src, srclen);
 124
 125         start_write_offset += srclen;
 126         if (start_write_offset >= ring_buffer_size)
 127                 start_write_offset -= ring_buffer_size;
 128
 129         return start_write_offset;
 130 }
 131
 132 /*
 133  *
 134  * hv_get_ringbuffer_availbytes()
 135  *
 136  * Get number of bytes available to read and to write to
 137  * for the specified ring buffer
 138  */
 139 static void
 140 hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
 141                              u32 *read, u32 *write)
 142 {
 143         u32 read_loc, write_loc, dsize;
 144
 145         /* Capture the read/write indices before they changed */
 146         read_loc = READ_ONCE(rbi->ring_buffer->read_index);
 147         write_loc = READ_ONCE(rbi->ring_buffer->write_index);
 148         dsize = rbi->ring_datasize;
 149
 150         *write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
 151                 read_loc - write_loc;
 152         *read = dsize - *write;
 153 }
 154
 155 /* Get various debug metrics for the specified ring buffer. */
 156 int hv_ringbuffer_get_debuginfo(struct hv_ring_buffer_info *ring_info,
 157                                 struct hv_ring_buffer_debug_info *debug_info)
 158 {
 159         u32 bytes_avail_towrite;
 160         u32 bytes_avail_toread;
 161
 162         mutex_lock(&ring_info->ring_buffer_mutex);
 163
 164         if (!ring_info->ring_buffer) {
 165                 mutex_unlock(&ring_info->ring_buffer_mutex);
 166                 return -EINVAL;
 167         }
 168
 169         hv_get_ringbuffer_availbytes(ring_info,
 170                                      &bytes_avail_toread,
 171                                      &bytes_avail_towrite);
 172         debug_info->bytes_avail_toread = bytes_avail_toread;
 173         debug_info->bytes_avail_towrite = bytes_avail_towrite;
 174         debug_info->current_read_index = ring_info->ring_buffer->read_index;
 175         debug_info->current_write_index = ring_info->ring_buffer->write_index;
 176         debug_info->current_interrupt_mask
 177                 = ring_info->ring_buffer->interrupt_mask;
 178         mutex_unlock(&ring_info->ring_buffer_mutex);
 179
 180         return 0;
 181 }
 182 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
 183
 184 /* Initialize a channel's ring buffer info mutex locks */
 185 void hv_ringbuffer_pre_init(struct vmbus_channel *channel)
 186 {
 187         mutex_init(&channel->inbound.ring_buffer_mutex);
 188         mutex_init(&channel->outbound.ring_buffer_mutex);
 189 }
 190
 191 /* Initialize the ring buffer. */
 192 int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
 193                        struct page *pages, u32 page_cnt)
 194 {
 195         int i;
 196         struct page **pages_wraparound;
 197
 198         BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
 199
 200         /*
 201          * First page holds struct hv_ring_buffer, do wraparound mapping for
 202          * the rest.
 203          */
 204         pages_wraparound = kcalloc(page_cnt * 2 - 1, sizeof(struct page *),
 205                                    GFP_KERNEL);
 206         if (!pages_wraparound)
 207                 return -ENOMEM;
 208
 209         pages_wraparound[0] = pages;
 210         for (i = 0; i < 2 * (page_cnt - 1); i++)
 211                 pages_wraparound[i + 1] = &pages[i % (page_cnt - 1) + 1];
 212
 213         ring_info->ring_buffer = (struct hv_ring_buffer *)
 214                 vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP, PAGE_KERNEL);
 215
 216         kfree(pages_wraparound);
 217
 218
 219         if (!ring_info->ring_buffer)
 220                 return -ENOMEM;
 221
 222         ring_info->ring_buffer->read_index =
 223                 ring_info->ring_buffer->write_index = 0;
 224
 225         /* Set the feature bit for enabling flow control. */
 226         ring_info->ring_buffer->feature_bits.value = 1;
 227
 228         ring_info->ring_size = page_cnt << PAGE_SHIFT;
 229         ring_info->ring_size_div10_reciprocal =
 230                 reciprocal_value(ring_info->ring_size / 10);
 231         ring_info->ring_datasize = ring_info->ring_size -
 232                 sizeof(struct hv_ring_buffer);
 233         ring_info->priv_read_index = 0;
 234
 235         spin_lock_init(&ring_info->ring_lock);
 236
 237         return 0;
 238 }
 239
 240 /* Cleanup the ring buffer. */
 241 void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
 242 {
 243         mutex_lock(&ring_info->ring_buffer_mutex);
 244         vunmap(ring_info->ring_buffer);
 245         ring_info->ring_buffer = NULL;
 246         mutex_unlock(&ring_info->ring_buffer_mutex);
 247 }
 248
 249 /* Write to the ring buffer. */
 250 int hv_ringbuffer_write(struct vmbus_channel *channel,
 251                         const struct kvec *kv_list, u32 kv_count)
 252 {
 253         int i;
 254         u32 bytes_avail_towrite;
 255         u32 totalbytes_towrite = sizeof(u64);
 256         u32 next_write_location;
 257         u32 old_write;
 258         u64 prev_indices;
 259         unsigned long flags;
 260         struct hv_ring_buffer_info *outring_info = &channel->outbound;
 261
 262         if (channel->rescind)
 263                 return -ENODEV;
 264
 265         for (i = 0; i < kv_count; i++)
 266                 totalbytes_towrite += kv_list[i].iov_len;
 267
 268         spin_lock_irqsave(&outring_info->ring_lock, flags);
 269
 270         bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
 271
 272         /*
 273          * If there is only room for the packet, assume it is full.
 274          * Otherwise, the next time around, we think the ring buffer
 275          * is empty since the read index == write index.
 276          */
 277         if (bytes_avail_towrite <= totalbytes_towrite) {
 278                 ++channel->out_full_total;
 279
 280                 if (!channel->out_full_flag) {
 281                         ++channel->out_full_first;
 282                         channel->out_full_flag = true;
 283                 }
 284
 285                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 286                 return -EAGAIN;
 287         }
 288
 289         channel->out_full_flag = false;
 290
 291         /* Write to the ring buffer */
 292         next_write_location = hv_get_next_write_location(outring_info);
 293
 294         old_write = next_write_location;
 295
 296         for (i = 0; i < kv_count; i++) {
 297                 next_write_location = hv_copyto_ringbuffer(outring_info,
 298                                                      next_write_location,
 299                                                      kv_list[i].iov_base,
 300                                                      kv_list[i].iov_len);
 301         }
 302
 303         /* Set previous packet start */
 304         prev_indices = hv_get_ring_bufferindices(outring_info);
 305
 306         next_write_location = hv_copyto_ringbuffer(outring_info,
 307                                              next_write_location,
 308                                              &prev_indices,
 309                                              sizeof(u64));
 310
 311         /* Issue a full memory barrier before updating the write index */
 312         virt_mb();
 313
 314         /* Now, update the write location */
 315         hv_set_next_write_location(outring_info, next_write_location);
 316
 317
 318         spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 319
 320         hv_signal_on_write(old_write, channel);
 321
 322         if (channel->rescind)
 323                 return -ENODEV;
 324
 325         return 0;
 326 }
 327
 328 int hv_ringbuffer_read(struct vmbus_channel *channel,
 329                        void *buffer, u32 buflen, u32 *buffer_actual_len,
 330                        u64 *requestid, bool raw)
 331 {
 332         struct vmpacket_descriptor *desc;
 333         u32 packetlen, offset;
 334
 335         if (unlikely(buflen == 0))
 336                 return -EINVAL;
 337
 338         *buffer_actual_len = 0;
 339         *requestid = 0;
 340
 341         /* Make sure there is something to read */
 342         desc = hv_pkt_iter_first(channel);
 343         if (desc == NULL) {
 344                 /*
 345                  * No error is set when there is even no header, drivers are
 346                  * supposed to analyze buffer_actual_len.
 347                  */
 348                 return 0;
 349         }
 350
 351         offset = raw ? 0 : (desc->offset8 << 3);
 352         packetlen = (desc->len8 << 3) - offset;
 353         *buffer_actual_len = packetlen;
 354         *requestid = desc->trans_id;
 355
 356         if (unlikely(packetlen > buflen))
 357                 return -ENOBUFS;
 358
 359         /* since ring is double mapped, only one copy is necessary */
 360         memcpy(buffer, (const char *)desc + offset, packetlen);
 361
 362         /* Advance ring index to next packet descriptor */
 363         __hv_pkt_iter_next(channel, desc);
 364
 365         /* Notify host of update */
 366         hv_pkt_iter_close(channel);
 367
 368         return 0;
 369 }
 370
 371 /*
 372  * Determine number of bytes available in ring buffer after
 373  * the current iterator (priv_read_index) location.
 374  *
 375  * This is similar to hv_get_bytes_to_read but with private
 376  * read index instead.
 377  */
 378 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
 379 {
 380         u32 priv_read_loc = rbi->priv_read_index;
 381         u32 write_loc;
 382
 383         /*
 384          * The Hyper-V host writes the packet data, then uses
 385          * store_release() to update the write_index.  Use load_acquire()
 386          * here to prevent loads of the packet data from being re-ordered
 387          * before the read of the write_index and potentially getting
 388          * stale data.
 389          */
 390         write_loc = virt_load_acquire(&rbi->ring_buffer->write_index);
 391
 392         if (write_loc >= priv_read_loc)
 393                 return write_loc - priv_read_loc;
 394         else
 395                 return (rbi->ring_datasize - priv_read_loc) + write_loc;
 396 }
 397
 398 /*
 399  * Get first vmbus packet from ring buffer after read_index
 400  *
 401  * If ring buffer is empty, returns NULL and no other action needed.
 402  */
 403 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
 404 {
 405         struct hv_ring_buffer_info *rbi = &channel->inbound;
 406         struct vmpacket_descriptor *desc;
 407
 408         hv_debug_delay_test(channel, MESSAGE_DELAY);
 409         if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
 410                 return NULL;
 411
 412         desc = hv_get_ring_buffer(rbi) + rbi->priv_read_index;
 413         if (desc)
 414                 prefetch((char *)desc + (desc->len8 << 3));
 415
 416         return desc;
 417 }
 418 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
 419
 420 /*
 421  * Get next vmbus packet from ring buffer.
 422  *
 423  * Advances the current location (priv_read_index) and checks for more
 424  * data. If the end of the ring buffer is reached, then return NULL.
 425  */
 426 struct vmpacket_descriptor *
 427 __hv_pkt_iter_next(struct vmbus_channel *channel,
 428                    const struct vmpacket_descriptor *desc)
 429 {
 430         struct hv_ring_buffer_info *rbi = &channel->inbound;
 431         u32 packetlen = desc->len8 << 3;
 432         u32 dsize = rbi->ring_datasize;
 433
 434         hv_debug_delay_test(channel, MESSAGE_DELAY);
 435         /* bump offset to next potential packet */
 436         rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
 437         if (rbi->priv_read_index >= dsize)
 438                 rbi->priv_read_index -= dsize;
 439
 440         /* more data? */
 441         return hv_pkt_iter_first(channel);
 442 }
 443 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
 444
 445 /* How many bytes were read in this iterator cycle */
 446 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
 447                                         u32 start_read_index)
 448 {
 449         if (rbi->priv_read_index >= start_read_index)
 450                 return rbi->priv_read_index - start_read_index;
 451         else
 452                 return rbi->ring_datasize - start_read_index +
 453                         rbi->priv_read_index;
 454 }
 455
 456 /*
 457  * Update host ring buffer after iterating over packets. If the host has
 458  * stopped queuing new entries because it found the ring buffer full, and
 459  * sufficient space is being freed up, signal the host. But be careful to
 460  * only signal the host when necessary, both for performance reasons and
 461  * because Hyper-V protects itself by throttling guests that signal
 462  * inappropriately.
 463  *
 464  * Determining when to signal is tricky. There are three key data inputs
 465  * that must be handled in this order to avoid race conditions:
 466  *
 467  * 1. Update the read_index
 468  * 2. Read the pending_send_sz
 469  * 3. Read the current write_index
 470  *
 471  * The interrupt_mask is not used to determine when to signal. The
 472  * interrupt_mask is used only on the guest->host ring buffer when
 473  * sending requests to the host. The host does not use it on the host->
 474  * guest ring buffer to indicate whether it should be signaled.
 475  */
 476 void hv_pkt_iter_close(struct vmbus_channel *channel)
 477 {
 478         struct hv_ring_buffer_info *rbi = &channel->inbound;
 479         u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
 480
 481         /*
 482          * Make sure all reads are done before we update the read index since
 483          * the writer may start writing to the read area once the read index
 484          * is updated.
 485          */
 486         virt_rmb();
 487         start_read_index = rbi->ring_buffer->read_index;
 488         rbi->ring_buffer->read_index = rbi->priv_read_index;
 489
 490         /*
 491          * Older versions of Hyper-V (before WS2102 and Win8) do not
 492          * implement pending_send_sz and simply poll if the host->guest
 493          * ring buffer is full.  No signaling is needed or expected.
 494          */
 495         if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
 496                 return;
 497
 498         /*
 499          * Issue a full memory barrier before making the signaling decision.
 500          * If reading pending_send_sz were to be reordered and happen
 501          * before we commit the new read_index, a race could occur.  If the
 502          * host were to set the pending_send_sz after we have sampled
 503          * pending_send_sz, and the ring buffer blocks before we commit the
 504          * read index, we could miss sending the interrupt. Issue a full
 505          * memory barrier to address this.
 506          */
 507         virt_mb();
 508
 509         /*
 510          * If the pending_send_sz is zero, then the ring buffer is not
 511          * blocked and there is no need to signal.  This is far by the
 512          * most common case, so exit quickly for best performance.
 513          */
 514         pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
 515         if (!pending_sz)
 516                 return;
 517
 518         /*
 519          * Ensure the read of write_index in hv_get_bytes_to_write()
 520          * happens after the read of pending_send_sz.
 521          */
 522         virt_rmb();
 523         curr_write_sz = hv_get_bytes_to_write(rbi);
 524         bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
 525
 526         /*
 527          * We want to signal the host only if we're transitioning
 528          * from a "not enough free space" state to a "enough free
 529          * space" state.  For example, it's possible that this function
 530          * could run and free up enough space to signal the host, and then
 531          * run again and free up additional space before the host has a
 532          * chance to clear the pending_send_sz.  The 2nd invocation would
 533          * be a null transition from "enough free space" to "enough free
 534          * space", which doesn't warrant a signal.
 535          *
 536          * Exactly filling the ring buffer is treated as "not enough
 537          * space". The ring buffer always must have at least one byte
 538          * empty so the empty and full conditions are distinguishable.
 539          * hv_get_bytes_to_write() doesn't fully tell the truth in
 540          * this regard.
 541          *
 542          * So first check if we were in the "enough free space" state
 543          * before we began the iteration. If so, the host was not
 544          * blocked, and there's no need to signal.
 545          */
 546         if (curr_write_sz - bytes_read > pending_sz)
 547                 return;
 548
 549         /*
 550          * Similarly, if the new state is "not enough space", then
 551          * there's no need to signal.
 552          */
 553         if (curr_write_sz <= pending_sz)
 554                 return;
 555
 556         ++channel->intr_in_full;
 557         vmbus_setevent(channel);
 558 }
 559 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);