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