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 #include <linux/io.h>
  21 #include <asm/mshyperv.h>
  22
  23 #include "hyperv_vmbus.h"
  24
  25 #define VMBUS_PKT_TRAILER       8
  26
  27 /*
  28  * When we write to the ring buffer, check if the host needs to
  29  * be signaled. Here is the details of this protocol:
  30  *
  31  *      1. The host guarantees that while it is draining the
  32  *         ring buffer, it will set the interrupt_mask to
  33  *         indicate it does not need to be interrupted when
  34  *         new data is placed.
  35  *
  36  *      2. The host guarantees that it will completely drain
  37  *         the ring buffer before exiting the read loop. Further,
  38  *         once the ring buffer is empty, it will clear the
  39  *         interrupt_mask and re-check to see if new data has
  40  *         arrived.
  41  *
  42  * KYS: Oct. 30, 2016:
  43  * It looks like Windows hosts have logic to deal with DOS attacks that
  44  * can be triggered if it receives interrupts when it is not expecting
  45  * the interrupt. The host expects interrupts only when the ring
  46  * transitions from empty to non-empty (or full to non full on the guest
  47  * to host ring).
  48  * So, base the signaling decision solely on the ring state until the
  49  * host logic is fixed.
  50  */
  51
  52 static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)
  53 {
  54         struct hv_ring_buffer_info *rbi = &channel->outbound;
  55
  56         virt_mb();
  57         if (READ_ONCE(rbi->ring_buffer->interrupt_mask))
  58                 return;
  59
  60         /* check interrupt_mask before read_index */
  61         virt_rmb();
  62         /*
  63          * This is the only case we need to signal when the
  64          * ring transitions from being empty to non-empty.
  65          */
  66         if (old_write == READ_ONCE(rbi->ring_buffer->read_index)) {
  67                 ++channel->intr_out_empty;
  68                 vmbus_setevent(channel);
  69         }
  70 }
  71
  72 /* Get the next write location for the specified ring buffer. */
  73 static inline u32
  74 hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
  75 {
  76         u32 next = ring_info->ring_buffer->write_index;
  77
  78         return next;
  79 }
  80
  81 /* Set the next write location for the specified ring buffer. */
  82 static inline void
  83 hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
  84                      u32 next_write_location)
  85 {
  86         ring_info->ring_buffer->write_index = next_write_location;
  87 }
  88
  89 /* Get the size of the ring buffer. */
  90 static inline u32
  91 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
  92 {
  93         return ring_info->ring_datasize;
  94 }
  95
  96 /* Get the read and write indices as u64 of the specified ring buffer. */
  97 static inline u64
  98 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
  99 {
 100         return (u64)ring_info->ring_buffer->write_index << 32;
 101 }
 102
 103 /*
 104  * Helper routine to copy from source to ring buffer.
 105  * Assume there is enough room. Handles wrap-around in dest case only!!
 106  */
 107 static u32 hv_copyto_ringbuffer(
 108         struct hv_ring_buffer_info      *ring_info,
 109         u32                             start_write_offset,
 110         const void                      *src,
 111         u32                             srclen)
 112 {
 113         void *ring_buffer = hv_get_ring_buffer(ring_info);
 114         u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
 115
 116         memcpy(ring_buffer + start_write_offset, src, srclen);
 117
 118         start_write_offset += srclen;
 119         if (start_write_offset >= ring_buffer_size)
 120                 start_write_offset -= ring_buffer_size;
 121
 122         return start_write_offset;
 123 }
 124
 125 /*
 126  *
 127  * hv_get_ringbuffer_availbytes()
 128  *
 129  * Get number of bytes available to read and to write to
 130  * for the specified ring buffer
 131  */
 132 static void
 133 hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
 134                              u32 *read, u32 *write)
 135 {
 136         u32 read_loc, write_loc, dsize;
 137
 138         /* Capture the read/write indices before they changed */
 139         read_loc = READ_ONCE(rbi->ring_buffer->read_index);
 140         write_loc = READ_ONCE(rbi->ring_buffer->write_index);
 141         dsize = rbi->ring_datasize;
 142
 143         *write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
 144                 read_loc - write_loc;
 145         *read = dsize - *write;
 146 }
 147
 148 /* Get various debug metrics for the specified ring buffer. */
 149 int hv_ringbuffer_get_debuginfo(struct hv_ring_buffer_info *ring_info,
 150                                 struct hv_ring_buffer_debug_info *debug_info)
 151 {
 152         u32 bytes_avail_towrite;
 153         u32 bytes_avail_toread;
 154
 155         mutex_lock(&ring_info->ring_buffer_mutex);
 156
 157         if (!ring_info->ring_buffer) {
 158                 mutex_unlock(&ring_info->ring_buffer_mutex);
 159                 return -EINVAL;
 160         }
 161
 162         hv_get_ringbuffer_availbytes(ring_info,
 163                                      &bytes_avail_toread,
 164                                      &bytes_avail_towrite);
 165         debug_info->bytes_avail_toread = bytes_avail_toread;
 166         debug_info->bytes_avail_towrite = bytes_avail_towrite;
 167         debug_info->current_read_index = ring_info->ring_buffer->read_index;
 168         debug_info->current_write_index = ring_info->ring_buffer->write_index;
 169         debug_info->current_interrupt_mask
 170                 = ring_info->ring_buffer->interrupt_mask;
 171         mutex_unlock(&ring_info->ring_buffer_mutex);
 172
 173         return 0;
 174 }
 175 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
 176
 177 /* Initialize a channel's ring buffer info mutex locks */
 178 void hv_ringbuffer_pre_init(struct vmbus_channel *channel)
 179 {
 180         mutex_init(&channel->inbound.ring_buffer_mutex);
 181         mutex_init(&channel->outbound.ring_buffer_mutex);
 182 }
 183
 184 /* Initialize the ring buffer. */
 185 int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
 186                        struct page *pages, u32 page_cnt, u32 max_pkt_size)
 187 {
 188         struct page **pages_wraparound;
 189         unsigned long *pfns_wraparound;
 190         u64 pfn;
 191         int i;
 192
 193         BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
 194
 195         /*
 196          * First page holds struct hv_ring_buffer, do wraparound mapping for
 197          * the rest.
 198          */
 199         if (hv_isolation_type_snp()) {
 200                 pfn = page_to_pfn(pages) +
 201                         PFN_DOWN(ms_hyperv.shared_gpa_boundary);
 202
 203                 pfns_wraparound = kcalloc(page_cnt * 2 - 1,
 204                         sizeof(unsigned long), GFP_KERNEL);
 205                 if (!pfns_wraparound)
 206                         return -ENOMEM;
 207
 208                 pfns_wraparound[0] = pfn;
 209                 for (i = 0; i < 2 * (page_cnt - 1); i++)
 210                         pfns_wraparound[i + 1] = pfn + i % (page_cnt - 1) + 1;
 211
 212                 ring_info->ring_buffer = (struct hv_ring_buffer *)
 213                         vmap_pfn(pfns_wraparound, page_cnt * 2 - 1,
 214                                  PAGE_KERNEL);
 215                 kfree(pfns_wraparound);
 216
 217                 if (!ring_info->ring_buffer)
 218                         return -ENOMEM;
 219
 220                 /* Zero ring buffer after setting memory host visibility. */
 221                 memset(ring_info->ring_buffer, 0x00, PAGE_SIZE * page_cnt);
 222         } else {
 223                 pages_wraparound = kcalloc(page_cnt * 2 - 1,
 224                                            sizeof(struct page *),
 225                                            GFP_KERNEL);
 226                 if (!pages_wraparound)
 227                         return -ENOMEM;
 228
 229                 pages_wraparound[0] = pages;
 230                 for (i = 0; i < 2 * (page_cnt - 1); i++)
 231                         pages_wraparound[i + 1] =
 232                                 &pages[i % (page_cnt - 1) + 1];
 233
 234                 ring_info->ring_buffer = (struct hv_ring_buffer *)
 235                         vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP,
 236                                 PAGE_KERNEL);
 237
 238                 kfree(pages_wraparound);
 239                 if (!ring_info->ring_buffer)
 240                         return -ENOMEM;
 241         }
 242
 243
 244         ring_info->ring_buffer->read_index =
 245                 ring_info->ring_buffer->write_index = 0;
 246
 247         /* Set the feature bit for enabling flow control. */
 248         ring_info->ring_buffer->feature_bits.value = 1;
 249
 250         ring_info->ring_size = page_cnt << PAGE_SHIFT;
 251         ring_info->ring_size_div10_reciprocal =
 252                 reciprocal_value(ring_info->ring_size / 10);
 253         ring_info->ring_datasize = ring_info->ring_size -
 254                 sizeof(struct hv_ring_buffer);
 255         ring_info->priv_read_index = 0;
 256
 257         /* Initialize buffer that holds copies of incoming packets */
 258         if (max_pkt_size) {
 259                 ring_info->pkt_buffer = kzalloc(max_pkt_size, GFP_KERNEL);
 260                 if (!ring_info->pkt_buffer)
 261                         return -ENOMEM;
 262                 ring_info->pkt_buffer_size = max_pkt_size;
 263         }
 264
 265         spin_lock_init(&ring_info->ring_lock);
 266
 267         return 0;
 268 }
 269
 270 /* Cleanup the ring buffer. */
 271 void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
 272 {
 273         mutex_lock(&ring_info->ring_buffer_mutex);
 274         vunmap(ring_info->ring_buffer);
 275         ring_info->ring_buffer = NULL;
 276         mutex_unlock(&ring_info->ring_buffer_mutex);
 277
 278         kfree(ring_info->pkt_buffer);
 279         ring_info->pkt_buffer = NULL;
 280         ring_info->pkt_buffer_size = 0;
 281 }
 282
 283 /* Write to the ring buffer. */
 284 int hv_ringbuffer_write(struct vmbus_channel *channel,
 285                         const struct kvec *kv_list, u32 kv_count,
 286                         u64 requestid, u64 *trans_id)
 287 {
 288         int i;
 289         u32 bytes_avail_towrite;
 290         u32 totalbytes_towrite = sizeof(u64);
 291         u32 next_write_location;
 292         u32 old_write;
 293         u64 prev_indices;
 294         unsigned long flags;
 295         struct hv_ring_buffer_info *outring_info = &channel->outbound;
 296         struct vmpacket_descriptor *desc = kv_list[0].iov_base;
 297         u64 __trans_id, rqst_id = VMBUS_NO_RQSTOR;
 298
 299         if (channel->rescind)
 300                 return -ENODEV;
 301
 302         for (i = 0; i < kv_count; i++)
 303                 totalbytes_towrite += kv_list[i].iov_len;
 304
 305         spin_lock_irqsave(&outring_info->ring_lock, flags);
 306
 307         bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
 308
 309         /*
 310          * If there is only room for the packet, assume it is full.
 311          * Otherwise, the next time around, we think the ring buffer
 312          * is empty since the read index == write index.
 313          */
 314         if (bytes_avail_towrite <= totalbytes_towrite) {
 315                 ++channel->out_full_total;
 316
 317                 if (!channel->out_full_flag) {
 318                         ++channel->out_full_first;
 319                         channel->out_full_flag = true;
 320                 }
 321
 322                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 323                 return -EAGAIN;
 324         }
 325
 326         channel->out_full_flag = false;
 327
 328         /* Write to the ring buffer */
 329         next_write_location = hv_get_next_write_location(outring_info);
 330
 331         old_write = next_write_location;
 332
 333         for (i = 0; i < kv_count; i++) {
 334                 next_write_location = hv_copyto_ringbuffer(outring_info,
 335                                                      next_write_location,
 336                                                      kv_list[i].iov_base,
 337                                                      kv_list[i].iov_len);
 338         }
 339
 340         /*
 341          * Allocate the request ID after the data has been copied into the
 342          * ring buffer.  Once this request ID is allocated, the completion
 343          * path could find the data and free it.
 344          */
 345
 346         if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {
 347                 if (channel->next_request_id_callback != NULL) {
 348                         rqst_id = channel->next_request_id_callback(channel, requestid);
 349                         if (rqst_id == VMBUS_RQST_ERROR) {
 350                                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 351                                 return -EAGAIN;
 352                         }
 353                 }
 354         }
 355         desc = hv_get_ring_buffer(outring_info) + old_write;
 356         __trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;
 357         /*
 358          * Ensure the compiler doesn't generate code that reads the value of
 359          * the transaction ID from the ring buffer, which is shared with the
 360          * Hyper-V host and subject to being changed at any time.
 361          */
 362         WRITE_ONCE(desc->trans_id, __trans_id);
 363         if (trans_id)
 364                 *trans_id = __trans_id;
 365
 366         /* Set previous packet start */
 367         prev_indices = hv_get_ring_bufferindices(outring_info);
 368
 369         next_write_location = hv_copyto_ringbuffer(outring_info,
 370                                              next_write_location,
 371                                              &prev_indices,
 372                                              sizeof(u64));
 373
 374         /* Issue a full memory barrier before updating the write index */
 375         virt_mb();
 376
 377         /* Now, update the write location */
 378         hv_set_next_write_location(outring_info, next_write_location);
 379
 380
 381         spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 382
 383         hv_signal_on_write(old_write, channel);
 384
 385         if (channel->rescind) {
 386                 if (rqst_id != VMBUS_NO_RQSTOR) {
 387                         /* Reclaim request ID to avoid leak of IDs */
 388                         if (channel->request_addr_callback != NULL)
 389                                 channel->request_addr_callback(channel, rqst_id);
 390                 }
 391                 return -ENODEV;
 392         }
 393
 394         return 0;
 395 }
 396
 397 int hv_ringbuffer_read(struct vmbus_channel *channel,
 398                        void *buffer, u32 buflen, u32 *buffer_actual_len,
 399                        u64 *requestid, bool raw)
 400 {
 401         struct vmpacket_descriptor *desc;
 402         u32 packetlen, offset;
 403
 404         if (unlikely(buflen == 0))
 405                 return -EINVAL;
 406
 407         *buffer_actual_len = 0;
 408         *requestid = 0;
 409
 410         /* Make sure there is something to read */
 411         desc = hv_pkt_iter_first(channel);
 412         if (desc == NULL) {
 413                 /*
 414                  * No error is set when there is even no header, drivers are
 415                  * supposed to analyze buffer_actual_len.
 416                  */
 417                 return 0;
 418         }
 419
 420         offset = raw ? 0 : (desc->offset8 << 3);
 421         packetlen = (desc->len8 << 3) - offset;
 422         *buffer_actual_len = packetlen;
 423         *requestid = desc->trans_id;
 424
 425         if (unlikely(packetlen > buflen))
 426                 return -ENOBUFS;
 427
 428         /* since ring is double mapped, only one copy is necessary */
 429         memcpy(buffer, (const char *)desc + offset, packetlen);
 430
 431         /* Advance ring index to next packet descriptor */
 432         __hv_pkt_iter_next(channel, desc);
 433
 434         /* Notify host of update */
 435         hv_pkt_iter_close(channel);
 436
 437         return 0;
 438 }
 439
 440 /*
 441  * Determine number of bytes available in ring buffer after
 442  * the current iterator (priv_read_index) location.
 443  *
 444  * This is similar to hv_get_bytes_to_read but with private
 445  * read index instead.
 446  */
 447 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
 448 {
 449         u32 priv_read_loc = rbi->priv_read_index;
 450         u32 write_loc;
 451
 452         /*
 453          * The Hyper-V host writes the packet data, then uses
 454          * store_release() to update the write_index.  Use load_acquire()
 455          * here to prevent loads of the packet data from being re-ordered
 456          * before the read of the write_index and potentially getting
 457          * stale data.
 458          */
 459         write_loc = virt_load_acquire(&rbi->ring_buffer->write_index);
 460
 461         if (write_loc >= priv_read_loc)
 462                 return write_loc - priv_read_loc;
 463         else
 464                 return (rbi->ring_datasize - priv_read_loc) + write_loc;
 465 }
 466
 467 /*
 468  * Get first vmbus packet from ring buffer after read_index
 469  *
 470  * If ring buffer is empty, returns NULL and no other action needed.
 471  */
 472 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
 473 {
 474         struct hv_ring_buffer_info *rbi = &channel->inbound;
 475         struct vmpacket_descriptor *desc, *desc_copy;
 476         u32 bytes_avail, pkt_len, pkt_offset;
 477
 478         hv_debug_delay_test(channel, MESSAGE_DELAY);
 479
 480         bytes_avail = hv_pkt_iter_avail(rbi);
 481         if (bytes_avail < sizeof(struct vmpacket_descriptor))
 482                 return NULL;
 483         bytes_avail = min(rbi->pkt_buffer_size, bytes_avail);
 484
 485         desc = (struct vmpacket_descriptor *)(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
 486
 487         /*
 488          * Ensure the compiler does not use references to incoming Hyper-V values (which
 489          * could change at any moment) when reading local variables later in the code
 490          */
 491         pkt_len = READ_ONCE(desc->len8) << 3;
 492         pkt_offset = READ_ONCE(desc->offset8) << 3;
 493
 494         /*
 495          * If pkt_len is invalid, set it to the smaller of hv_pkt_iter_avail() and
 496          * rbi->pkt_buffer_size
 497          */
 498         if (pkt_len < sizeof(struct vmpacket_descriptor) || pkt_len > bytes_avail)
 499                 pkt_len = bytes_avail;
 500
 501         /*
 502          * If pkt_offset is invalid, arbitrarily set it to
 503          * the size of vmpacket_descriptor
 504          */
 505         if (pkt_offset < sizeof(struct vmpacket_descriptor) || pkt_offset > pkt_len)
 506                 pkt_offset = sizeof(struct vmpacket_descriptor);
 507
 508         /* Copy the Hyper-V packet out of the ring buffer */
 509         desc_copy = (struct vmpacket_descriptor *)rbi->pkt_buffer;
 510         memcpy(desc_copy, desc, pkt_len);
 511
 512         /*
 513          * Hyper-V could still change len8 and offset8 after the earlier read.
 514          * Ensure that desc_copy has legal values for len8 and offset8 that
 515          * are consistent with the copy we just made
 516          */
 517         desc_copy->len8 = pkt_len >> 3;
 518         desc_copy->offset8 = pkt_offset >> 3;
 519
 520         return desc_copy;
 521 }
 522 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
 523
 524 /*
 525  * Get next vmbus packet from ring buffer.
 526  *
 527  * Advances the current location (priv_read_index) and checks for more
 528  * data. If the end of the ring buffer is reached, then return NULL.
 529  */
 530 struct vmpacket_descriptor *
 531 __hv_pkt_iter_next(struct vmbus_channel *channel,
 532                    const struct vmpacket_descriptor *desc)
 533 {
 534         struct hv_ring_buffer_info *rbi = &channel->inbound;
 535         u32 packetlen = desc->len8 << 3;
 536         u32 dsize = rbi->ring_datasize;
 537
 538         hv_debug_delay_test(channel, MESSAGE_DELAY);
 539         /* bump offset to next potential packet */
 540         rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
 541         if (rbi->priv_read_index >= dsize)
 542                 rbi->priv_read_index -= dsize;
 543
 544         /* more data? */
 545         return hv_pkt_iter_first(channel);
 546 }
 547 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
 548
 549 /* How many bytes were read in this iterator cycle */
 550 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
 551                                         u32 start_read_index)
 552 {
 553         if (rbi->priv_read_index >= start_read_index)
 554                 return rbi->priv_read_index - start_read_index;
 555         else
 556                 return rbi->ring_datasize - start_read_index +
 557                         rbi->priv_read_index;
 558 }
 559
 560 /*
 561  * Update host ring buffer after iterating over packets. If the host has
 562  * stopped queuing new entries because it found the ring buffer full, and
 563  * sufficient space is being freed up, signal the host. But be careful to
 564  * only signal the host when necessary, both for performance reasons and
 565  * because Hyper-V protects itself by throttling guests that signal
 566  * inappropriately.
 567  *
 568  * Determining when to signal is tricky. There are three key data inputs
 569  * that must be handled in this order to avoid race conditions:
 570  *
 571  * 1. Update the read_index
 572  * 2. Read the pending_send_sz
 573  * 3. Read the current write_index
 574  *
 575  * The interrupt_mask is not used to determine when to signal. The
 576  * interrupt_mask is used only on the guest->host ring buffer when
 577  * sending requests to the host. The host does not use it on the host->
 578  * guest ring buffer to indicate whether it should be signaled.
 579  */
 580 void hv_pkt_iter_close(struct vmbus_channel *channel)
 581 {
 582         struct hv_ring_buffer_info *rbi = &channel->inbound;
 583         u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
 584
 585         /*
 586          * Make sure all reads are done before we update the read index since
 587          * the writer may start writing to the read area once the read index
 588          * is updated.
 589          */
 590         virt_rmb();
 591         start_read_index = rbi->ring_buffer->read_index;
 592         rbi->ring_buffer->read_index = rbi->priv_read_index;
 593
 594         /*
 595          * Older versions of Hyper-V (before WS2102 and Win8) do not
 596          * implement pending_send_sz and simply poll if the host->guest
 597          * ring buffer is full.  No signaling is needed or expected.
 598          */
 599         if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
 600                 return;
 601
 602         /*
 603          * Issue a full memory barrier before making the signaling decision.
 604          * If reading pending_send_sz were to be reordered and happen
 605          * before we commit the new read_index, a race could occur.  If the
 606          * host were to set the pending_send_sz after we have sampled
 607          * pending_send_sz, and the ring buffer blocks before we commit the
 608          * read index, we could miss sending the interrupt. Issue a full
 609          * memory barrier to address this.
 610          */
 611         virt_mb();
 612
 613         /*
 614          * If the pending_send_sz is zero, then the ring buffer is not
 615          * blocked and there is no need to signal.  This is far by the
 616          * most common case, so exit quickly for best performance.
 617          */
 618         pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
 619         if (!pending_sz)
 620                 return;
 621
 622         /*
 623          * Ensure the read of write_index in hv_get_bytes_to_write()
 624          * happens after the read of pending_send_sz.
 625          */
 626         virt_rmb();
 627         curr_write_sz = hv_get_bytes_to_write(rbi);
 628         bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
 629
 630         /*
 631          * We want to signal the host only if we're transitioning
 632          * from a "not enough free space" state to a "enough free
 633          * space" state.  For example, it's possible that this function
 634          * could run and free up enough space to signal the host, and then
 635          * run again and free up additional space before the host has a
 636          * chance to clear the pending_send_sz.  The 2nd invocation would
 637          * be a null transition from "enough free space" to "enough free
 638          * space", which doesn't warrant a signal.
 639          *
 640          * Exactly filling the ring buffer is treated as "not enough
 641          * space". The ring buffer always must have at least one byte
 642          * empty so the empty and full conditions are distinguishable.
 643          * hv_get_bytes_to_write() doesn't fully tell the truth in
 644          * this regard.
 645          *
 646          * So first check if we were in the "enough free space" state
 647          * before we began the iteration. If so, the host was not
 648          * blocked, and there's no need to signal.
 649          */
 650         if (curr_write_sz - bytes_read > pending_sz)
 651                 return;
 652
 653         /*
 654          * Similarly, if the new state is "not enough space", then
 655          * there's no need to signal.
 656          */
 657         if (curr_write_sz <= pending_sz)
 658                 return;
 659
 660         ++channel->intr_in_full;
 661         vmbus_setevent(channel);
 662 }
 663 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);