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 /* Get the size of the ring buffer. */
  88 static inline u32
  89 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
  90 {
  91         return ring_info->ring_datasize;
  92 }
  93
  94 /* Get the read and write indices as u64 of the specified ring buffer. */
  95 static inline u64
  96 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
  97 {
  98         return (u64)ring_info->ring_buffer->write_index << 32;
  99 }
 100
 101 /*
 102  * Helper routine to copy from source to ring buffer.
 103  * Assume there is enough room. Handles wrap-around in dest case only!!
 104  */
 105 static u32 hv_copyto_ringbuffer(
 106         struct hv_ring_buffer_info      *ring_info,
 107         u32                             start_write_offset,
 108         const void                      *src,
 109         u32                             srclen)
 110 {
 111         void *ring_buffer = hv_get_ring_buffer(ring_info);
 112         u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
 113
 114         memcpy(ring_buffer + start_write_offset, src, srclen);
 115
 116         start_write_offset += srclen;
 117         if (start_write_offset >= ring_buffer_size)
 118                 start_write_offset -= ring_buffer_size;
 119
 120         return start_write_offset;
 121 }
 122
 123 /*
 124  *
 125  * hv_get_ringbuffer_availbytes()
 126  *
 127  * Get number of bytes available to read and to write to
 128  * for the specified ring buffer
 129  */
 130 static void
 131 hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
 132                              u32 *read, u32 *write)
 133 {
 134         u32 read_loc, write_loc, dsize;
 135
 136         /* Capture the read/write indices before they changed */
 137         read_loc = READ_ONCE(rbi->ring_buffer->read_index);
 138         write_loc = READ_ONCE(rbi->ring_buffer->write_index);
 139         dsize = rbi->ring_datasize;
 140
 141         *write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
 142                 read_loc - write_loc;
 143         *read = dsize - *write;
 144 }
 145
 146 /* Get various debug metrics for the specified ring buffer. */
 147 int hv_ringbuffer_get_debuginfo(struct hv_ring_buffer_info *ring_info,
 148                                 struct hv_ring_buffer_debug_info *debug_info)
 149 {
 150         u32 bytes_avail_towrite;
 151         u32 bytes_avail_toread;
 152
 153         mutex_lock(&ring_info->ring_buffer_mutex);
 154
 155         if (!ring_info->ring_buffer) {
 156                 mutex_unlock(&ring_info->ring_buffer_mutex);
 157                 return -EINVAL;
 158         }
 159
 160         hv_get_ringbuffer_availbytes(ring_info,
 161                                      &bytes_avail_toread,
 162                                      &bytes_avail_towrite);
 163         debug_info->bytes_avail_toread = bytes_avail_toread;
 164         debug_info->bytes_avail_towrite = bytes_avail_towrite;
 165         debug_info->current_read_index = ring_info->ring_buffer->read_index;
 166         debug_info->current_write_index = ring_info->ring_buffer->write_index;
 167         debug_info->current_interrupt_mask
 168                 = ring_info->ring_buffer->interrupt_mask;
 169         mutex_unlock(&ring_info->ring_buffer_mutex);
 170
 171         return 0;
 172 }
 173 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
 174
 175 /* Initialize a channel's ring buffer info mutex locks */
 176 void hv_ringbuffer_pre_init(struct vmbus_channel *channel)
 177 {
 178         mutex_init(&channel->inbound.ring_buffer_mutex);
 179         mutex_init(&channel->outbound.ring_buffer_mutex);
 180 }
 181
 182 /* Initialize the ring buffer. */
 183 int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
 184                        struct page *pages, u32 page_cnt, u32 max_pkt_size)
 185 {
 186         int i;
 187         struct page **pages_wraparound;
 188
 189         BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
 190
 191         /*
 192          * First page holds struct hv_ring_buffer, do wraparound mapping for
 193          * the rest.
 194          */
 195         pages_wraparound = kcalloc(page_cnt * 2 - 1, sizeof(struct page *),
 196                                    GFP_KERNEL);
 197         if (!pages_wraparound)
 198                 return -ENOMEM;
 199
 200         pages_wraparound[0] = pages;
 201         for (i = 0; i < 2 * (page_cnt - 1); i++)
 202                 pages_wraparound[i + 1] = &pages[i % (page_cnt - 1) + 1];
 203
 204         ring_info->ring_buffer = (struct hv_ring_buffer *)
 205                 vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP, PAGE_KERNEL);
 206
 207         kfree(pages_wraparound);
 208
 209
 210         if (!ring_info->ring_buffer)
 211                 return -ENOMEM;
 212
 213         ring_info->ring_buffer->read_index =
 214                 ring_info->ring_buffer->write_index = 0;
 215
 216         /* Set the feature bit for enabling flow control. */
 217         ring_info->ring_buffer->feature_bits.value = 1;
 218
 219         ring_info->ring_size = page_cnt << PAGE_SHIFT;
 220         ring_info->ring_size_div10_reciprocal =
 221                 reciprocal_value(ring_info->ring_size / 10);
 222         ring_info->ring_datasize = ring_info->ring_size -
 223                 sizeof(struct hv_ring_buffer);
 224         ring_info->priv_read_index = 0;
 225
 226         /* Initialize buffer that holds copies of incoming packets */
 227         if (max_pkt_size) {
 228                 ring_info->pkt_buffer = kzalloc(max_pkt_size, GFP_KERNEL);
 229                 if (!ring_info->pkt_buffer)
 230                         return -ENOMEM;
 231                 ring_info->pkt_buffer_size = max_pkt_size;
 232         }
 233
 234         spin_lock_init(&ring_info->ring_lock);
 235
 236         return 0;
 237 }
 238
 239 /* Cleanup the ring buffer. */
 240 void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
 241 {
 242         mutex_lock(&ring_info->ring_buffer_mutex);
 243         vunmap(ring_info->ring_buffer);
 244         ring_info->ring_buffer = NULL;
 245         mutex_unlock(&ring_info->ring_buffer_mutex);
 246
 247         kfree(ring_info->pkt_buffer);
 248         ring_info->pkt_buffer = NULL;
 249         ring_info->pkt_buffer_size = 0;
 250 }
 251
 252 /*
 253  * Check if the ring buffer spinlock is available to take or not; used on
 254  * atomic contexts, like panic path (see the Hyper-V framebuffer driver).
 255  */
 256
 257 bool hv_ringbuffer_spinlock_busy(struct vmbus_channel *channel)
 258 {
 259         struct hv_ring_buffer_info *rinfo = &channel->outbound;
 260
 261         return spin_is_locked(&rinfo->ring_lock);
 262 }
 263 EXPORT_SYMBOL_GPL(hv_ringbuffer_spinlock_busy);
 264
 265 /* Write to the ring buffer. */
 266 int hv_ringbuffer_write(struct vmbus_channel *channel,
 267                         const struct kvec *kv_list, u32 kv_count,
 268                         u64 requestid)
 269 {
 270         int i;
 271         u32 bytes_avail_towrite;
 272         u32 totalbytes_towrite = sizeof(u64);
 273         u32 next_write_location;
 274         u32 old_write;
 275         u64 prev_indices;
 276         unsigned long flags;
 277         struct hv_ring_buffer_info *outring_info = &channel->outbound;
 278         struct vmpacket_descriptor *desc = kv_list[0].iov_base;
 279         u64 rqst_id = VMBUS_NO_RQSTOR;
 280
 281         if (channel->rescind)
 282                 return -ENODEV;
 283
 284         for (i = 0; i < kv_count; i++)
 285                 totalbytes_towrite += kv_list[i].iov_len;
 286
 287         spin_lock_irqsave(&outring_info->ring_lock, flags);
 288
 289         bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
 290
 291         /*
 292          * If there is only room for the packet, assume it is full.
 293          * Otherwise, the next time around, we think the ring buffer
 294          * is empty since the read index == write index.
 295          */
 296         if (bytes_avail_towrite <= totalbytes_towrite) {
 297                 ++channel->out_full_total;
 298
 299                 if (!channel->out_full_flag) {
 300                         ++channel->out_full_first;
 301                         channel->out_full_flag = true;
 302                 }
 303
 304                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 305                 return -EAGAIN;
 306         }
 307
 308         channel->out_full_flag = false;
 309
 310         /* Write to the ring buffer */
 311         next_write_location = hv_get_next_write_location(outring_info);
 312
 313         old_write = next_write_location;
 314
 315         for (i = 0; i < kv_count; i++) {
 316                 next_write_location = hv_copyto_ringbuffer(outring_info,
 317                                                      next_write_location,
 318                                                      kv_list[i].iov_base,
 319                                                      kv_list[i].iov_len);
 320         }
 321
 322         /*
 323          * Allocate the request ID after the data has been copied into the
 324          * ring buffer.  Once this request ID is allocated, the completion
 325          * path could find the data and free it.
 326          */
 327
 328         if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {
 329                 if (channel->next_request_id_callback != NULL) {
 330                         rqst_id = channel->next_request_id_callback(channel, requestid);
 331                         if (rqst_id == VMBUS_RQST_ERROR) {
 332                                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 333                                 return -EAGAIN;
 334                         }
 335                 }
 336         }
 337         desc = hv_get_ring_buffer(outring_info) + old_write;
 338         desc->trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;
 339
 340         /* Set previous packet start */
 341         prev_indices = hv_get_ring_bufferindices(outring_info);
 342
 343         next_write_location = hv_copyto_ringbuffer(outring_info,
 344                                              next_write_location,
 345                                              &prev_indices,
 346                                              sizeof(u64));
 347
 348         /* Issue a full memory barrier before updating the write index */
 349         virt_mb();
 350
 351         /* Now, update the write location */
 352         hv_set_next_write_location(outring_info, next_write_location);
 353
 354
 355         spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 356
 357         hv_signal_on_write(old_write, channel);
 358
 359         if (channel->rescind) {
 360                 if (rqst_id != VMBUS_NO_RQSTOR) {
 361                         /* Reclaim request ID to avoid leak of IDs */
 362                         if (channel->request_addr_callback != NULL)
 363                                 channel->request_addr_callback(channel, rqst_id);
 364                 }
 365                 return -ENODEV;
 366         }
 367
 368         return 0;
 369 }
 370
 371 int hv_ringbuffer_read(struct vmbus_channel *channel,
 372                        void *buffer, u32 buflen, u32 *buffer_actual_len,
 373                        u64 *requestid, bool raw)
 374 {
 375         struct vmpacket_descriptor *desc;
 376         u32 packetlen, offset;
 377
 378         if (unlikely(buflen == 0))
 379                 return -EINVAL;
 380
 381         *buffer_actual_len = 0;
 382         *requestid = 0;
 383
 384         /* Make sure there is something to read */
 385         desc = hv_pkt_iter_first(channel);
 386         if (desc == NULL) {
 387                 /*
 388                  * No error is set when there is even no header, drivers are
 389                  * supposed to analyze buffer_actual_len.
 390                  */
 391                 return 0;
 392         }
 393
 394         offset = raw ? 0 : (desc->offset8 << 3);
 395         packetlen = (desc->len8 << 3) - offset;
 396         *buffer_actual_len = packetlen;
 397         *requestid = desc->trans_id;
 398
 399         if (unlikely(packetlen > buflen))
 400                 return -ENOBUFS;
 401
 402         /* since ring is double mapped, only one copy is necessary */
 403         memcpy(buffer, (const char *)desc + offset, packetlen);
 404
 405         /* Advance ring index to next packet descriptor */
 406         __hv_pkt_iter_next(channel, desc, true);
 407
 408         /* Notify host of update */
 409         hv_pkt_iter_close(channel);
 410
 411         return 0;
 412 }
 413
 414 /*
 415  * Determine number of bytes available in ring buffer after
 416  * the current iterator (priv_read_index) location.
 417  *
 418  * This is similar to hv_get_bytes_to_read but with private
 419  * read index instead.
 420  */
 421 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
 422 {
 423         u32 priv_read_loc = rbi->priv_read_index;
 424         u32 write_loc;
 425
 426         /*
 427          * The Hyper-V host writes the packet data, then uses
 428          * store_release() to update the write_index.  Use load_acquire()
 429          * here to prevent loads of the packet data from being re-ordered
 430          * before the read of the write_index and potentially getting
 431          * stale data.
 432          */
 433         write_loc = virt_load_acquire(&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 without copying it out of the ring buffer
 443  */
 444 struct vmpacket_descriptor *hv_pkt_iter_first_raw(struct vmbus_channel *channel)
 445 {
 446         struct hv_ring_buffer_info *rbi = &channel->inbound;
 447
 448         hv_debug_delay_test(channel, MESSAGE_DELAY);
 449
 450         if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
 451                 return NULL;
 452
 453         return (struct vmpacket_descriptor *)(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
 454 }
 455 EXPORT_SYMBOL_GPL(hv_pkt_iter_first_raw);
 456
 457 /*
 458  * Get first vmbus packet from ring buffer after read_index
 459  *
 460  * If ring buffer is empty, returns NULL and no other action needed.
 461  */
 462 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
 463 {
 464         struct hv_ring_buffer_info *rbi = &channel->inbound;
 465         struct vmpacket_descriptor *desc, *desc_copy;
 466         u32 bytes_avail, pkt_len, pkt_offset;
 467
 468         desc = hv_pkt_iter_first_raw(channel);
 469         if (!desc)
 470                 return NULL;
 471
 472         bytes_avail = min(rbi->pkt_buffer_size, hv_pkt_iter_avail(rbi));
 473
 474         /*
 475          * Ensure the compiler does not use references to incoming Hyper-V values (which
 476          * could change at any moment) when reading local variables later in the code
 477          */
 478         pkt_len = READ_ONCE(desc->len8) << 3;
 479         pkt_offset = READ_ONCE(desc->offset8) << 3;
 480
 481         /*
 482          * If pkt_len is invalid, set it to the smaller of hv_pkt_iter_avail() and
 483          * rbi->pkt_buffer_size
 484          */
 485         if (pkt_len < sizeof(struct vmpacket_descriptor) || pkt_len > bytes_avail)
 486                 pkt_len = bytes_avail;
 487
 488         /*
 489          * If pkt_offset is invalid, arbitrarily set it to
 490          * the size of vmpacket_descriptor
 491          */
 492         if (pkt_offset < sizeof(struct vmpacket_descriptor) || pkt_offset > pkt_len)
 493                 pkt_offset = sizeof(struct vmpacket_descriptor);
 494
 495         /* Copy the Hyper-V packet out of the ring buffer */
 496         desc_copy = (struct vmpacket_descriptor *)rbi->pkt_buffer;
 497         memcpy(desc_copy, desc, pkt_len);
 498
 499         /*
 500          * Hyper-V could still change len8 and offset8 after the earlier read.
 501          * Ensure that desc_copy has legal values for len8 and offset8 that
 502          * are consistent with the copy we just made
 503          */
 504         desc_copy->len8 = pkt_len >> 3;
 505         desc_copy->offset8 = pkt_offset >> 3;
 506
 507         return desc_copy;
 508 }
 509 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
 510
 511 /*
 512  * Get next vmbus packet from ring buffer.
 513  *
 514  * Advances the current location (priv_read_index) and checks for more
 515  * data. If the end of the ring buffer is reached, then return NULL.
 516  */
 517 struct vmpacket_descriptor *
 518 __hv_pkt_iter_next(struct vmbus_channel *channel,
 519                    const struct vmpacket_descriptor *desc,
 520                    bool copy)
 521 {
 522         struct hv_ring_buffer_info *rbi = &channel->inbound;
 523         u32 packetlen = desc->len8 << 3;
 524         u32 dsize = rbi->ring_datasize;
 525
 526         hv_debug_delay_test(channel, MESSAGE_DELAY);
 527         /* bump offset to next potential packet */
 528         rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
 529         if (rbi->priv_read_index >= dsize)
 530                 rbi->priv_read_index -= dsize;
 531
 532         /* more data? */
 533         return copy ? hv_pkt_iter_first(channel) : hv_pkt_iter_first_raw(channel);
 534 }
 535 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
 536
 537 /* How many bytes were read in this iterator cycle */
 538 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
 539                                         u32 start_read_index)
 540 {
 541         if (rbi->priv_read_index >= start_read_index)
 542                 return rbi->priv_read_index - start_read_index;
 543         else
 544                 return rbi->ring_datasize - start_read_index +
 545                         rbi->priv_read_index;
 546 }
 547
 548 /*
 549  * Update host ring buffer after iterating over packets. If the host has
 550  * stopped queuing new entries because it found the ring buffer full, and
 551  * sufficient space is being freed up, signal the host. But be careful to
 552  * only signal the host when necessary, both for performance reasons and
 553  * because Hyper-V protects itself by throttling guests that signal
 554  * inappropriately.
 555  *
 556  * Determining when to signal is tricky. There are three key data inputs
 557  * that must be handled in this order to avoid race conditions:
 558  *
 559  * 1. Update the read_index
 560  * 2. Read the pending_send_sz
 561  * 3. Read the current write_index
 562  *
 563  * The interrupt_mask is not used to determine when to signal. The
 564  * interrupt_mask is used only on the guest->host ring buffer when
 565  * sending requests to the host. The host does not use it on the host->
 566  * guest ring buffer to indicate whether it should be signaled.
 567  */
 568 void hv_pkt_iter_close(struct vmbus_channel *channel)
 569 {
 570         struct hv_ring_buffer_info *rbi = &channel->inbound;
 571         u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
 572
 573         /*
 574          * Make sure all reads are done before we update the read index since
 575          * the writer may start writing to the read area once the read index
 576          * is updated.
 577          */
 578         virt_rmb();
 579         start_read_index = rbi->ring_buffer->read_index;
 580         rbi->ring_buffer->read_index = rbi->priv_read_index;
 581
 582         /*
 583          * Older versions of Hyper-V (before WS2102 and Win8) do not
 584          * implement pending_send_sz and simply poll if the host->guest
 585          * ring buffer is full.  No signaling is needed or expected.
 586          */
 587         if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
 588                 return;
 589
 590         /*
 591          * Issue a full memory barrier before making the signaling decision.
 592          * If reading pending_send_sz were to be reordered and happen
 593          * before we commit the new read_index, a race could occur.  If the
 594          * host were to set the pending_send_sz after we have sampled
 595          * pending_send_sz, and the ring buffer blocks before we commit the
 596          * read index, we could miss sending the interrupt. Issue a full
 597          * memory barrier to address this.
 598          */
 599         virt_mb();
 600
 601         /*
 602          * If the pending_send_sz is zero, then the ring buffer is not
 603          * blocked and there is no need to signal.  This is far by the
 604          * most common case, so exit quickly for best performance.
 605          */
 606         pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
 607         if (!pending_sz)
 608                 return;
 609
 610         /*
 611          * Ensure the read of write_index in hv_get_bytes_to_write()
 612          * happens after the read of pending_send_sz.
 613          */
 614         virt_rmb();
 615         curr_write_sz = hv_get_bytes_to_write(rbi);
 616         bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
 617
 618         /*
 619          * We want to signal the host only if we're transitioning
 620          * from a "not enough free space" state to a "enough free
 621          * space" state.  For example, it's possible that this function
 622          * could run and free up enough space to signal the host, and then
 623          * run again and free up additional space before the host has a
 624          * chance to clear the pending_send_sz.  The 2nd invocation would
 625          * be a null transition from "enough free space" to "enough free
 626          * space", which doesn't warrant a signal.
 627          *
 628          * Exactly filling the ring buffer is treated as "not enough
 629          * space". The ring buffer always must have at least one byte
 630          * empty so the empty and full conditions are distinguishable.
 631          * hv_get_bytes_to_write() doesn't fully tell the truth in
 632          * this regard.
 633          *
 634          * So first check if we were in the "enough free space" state
 635          * before we began the iteration. If so, the host was not
 636          * blocked, and there's no need to signal.
 637          */
 638         if (curr_write_sz - bytes_read > pending_sz)
 639                 return;
 640
 641         /*
 642          * Similarly, if the new state is "not enough space", then
 643          * there's no need to signal.
 644          */
 645         if (curr_write_sz <= pending_sz)
 646                 return;
 647
 648         ++channel->intr_in_full;
 649         vmbus_setevent(channel);
 650 }
 651 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);