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                 vmbus_setevent(channel);
  79 }
  80
  81 /* Get the next write location for the specified ring buffer. */
  82 static inline u32
  83 hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
  84 {
  85         u32 next = ring_info->ring_buffer->write_index;
  86
  87         return next;
  88 }
  89
  90 /* Set the next write location for the specified ring buffer. */
  91 static inline void
  92 hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
  93                      u32 next_write_location)
  94 {
  95         ring_info->ring_buffer->write_index = next_write_location;
  96 }
  97
  98 /* Set the next read location for the specified ring buffer. */
  99 static inline void
 100 hv_set_next_read_location(struct hv_ring_buffer_info *ring_info,
 101                     u32 next_read_location)
 102 {
 103         ring_info->ring_buffer->read_index = next_read_location;
 104         ring_info->priv_read_index = next_read_location;
 105 }
 106
 107 /* Get the size of the ring buffer. */
 108 static inline u32
 109 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
 110 {
 111         return ring_info->ring_datasize;
 112 }
 113
 114 /* Get the read and write indices as u64 of the specified ring buffer. */
 115 static inline u64
 116 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
 117 {
 118         return (u64)ring_info->ring_buffer->write_index << 32;
 119 }
 120
 121 /*
 122  * Helper routine to copy from source to ring buffer.
 123  * Assume there is enough room. Handles wrap-around in dest case only!!
 124  */
 125 static u32 hv_copyto_ringbuffer(
 126         struct hv_ring_buffer_info      *ring_info,
 127         u32                             start_write_offset,
 128         const void                      *src,
 129         u32                             srclen)
 130 {
 131         void *ring_buffer = hv_get_ring_buffer(ring_info);
 132         u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
 133
 134         memcpy(ring_buffer + start_write_offset, src, srclen);
 135
 136         start_write_offset += srclen;
 137         if (start_write_offset >= ring_buffer_size)
 138                 start_write_offset -= ring_buffer_size;
 139
 140         return start_write_offset;
 141 }
 142
 143 /*
 144  *
 145  * hv_get_ringbuffer_availbytes()
 146  *
 147  * Get number of bytes available to read and to write to
 148  * for the specified ring buffer
 149  */
 150 static void
 151 hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
 152                              u32 *read, u32 *write)
 153 {
 154         u32 read_loc, write_loc, dsize;
 155
 156         /* Capture the read/write indices before they changed */
 157         read_loc = READ_ONCE(rbi->ring_buffer->read_index);
 158         write_loc = READ_ONCE(rbi->ring_buffer->write_index);
 159         dsize = rbi->ring_datasize;
 160
 161         *write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
 162                 read_loc - write_loc;
 163         *read = dsize - *write;
 164 }
 165
 166 /* Get various debug metrics for the specified ring buffer. */
 167 void hv_ringbuffer_get_debuginfo(const struct hv_ring_buffer_info *ring_info,
 168                                  struct hv_ring_buffer_debug_info *debug_info)
 169 {
 170         u32 bytes_avail_towrite;
 171         u32 bytes_avail_toread;
 172
 173         if (ring_info->ring_buffer) {
 174                 hv_get_ringbuffer_availbytes(ring_info,
 175                                         &bytes_avail_toread,
 176                                         &bytes_avail_towrite);
 177
 178                 debug_info->bytes_avail_toread = bytes_avail_toread;
 179                 debug_info->bytes_avail_towrite = bytes_avail_towrite;
 180                 debug_info->current_read_index =
 181                         ring_info->ring_buffer->read_index;
 182                 debug_info->current_write_index =
 183                         ring_info->ring_buffer->write_index;
 184                 debug_info->current_interrupt_mask =
 185                         ring_info->ring_buffer->interrupt_mask;
 186         }
 187 }
 188 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
 189
 190 /* Initialize the ring buffer. */
 191 int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
 192                        struct page *pages, u32 page_cnt)
 193 {
 194         int i;
 195         struct page **pages_wraparound;
 196
 197         BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
 198
 199         memset(ring_info, 0, sizeof(struct hv_ring_buffer_info));
 200
 201         /*
 202          * First page holds struct hv_ring_buffer, do wraparound mapping for
 203          * the rest.
 204          */
 205         pages_wraparound = kcalloc(page_cnt * 2 - 1, sizeof(struct page *),
 206                                    GFP_KERNEL);
 207         if (!pages_wraparound)
 208                 return -ENOMEM;
 209
 210         pages_wraparound[0] = pages;
 211         for (i = 0; i < 2 * (page_cnt - 1); i++)
 212                 pages_wraparound[i + 1] = &pages[i % (page_cnt - 1) + 1];
 213
 214         ring_info->ring_buffer = (struct hv_ring_buffer *)
 215                 vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP, PAGE_KERNEL);
 216
 217         kfree(pages_wraparound);
 218
 219
 220         if (!ring_info->ring_buffer)
 221                 return -ENOMEM;
 222
 223         ring_info->ring_buffer->read_index =
 224                 ring_info->ring_buffer->write_index = 0;
 225
 226         /* Set the feature bit for enabling flow control. */
 227         ring_info->ring_buffer->feature_bits.value = 1;
 228
 229         ring_info->ring_size = page_cnt << PAGE_SHIFT;
 230         ring_info->ring_size_div10_reciprocal =
 231                 reciprocal_value(ring_info->ring_size / 10);
 232         ring_info->ring_datasize = ring_info->ring_size -
 233                 sizeof(struct hv_ring_buffer);
 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         vunmap(ring_info->ring_buffer);
 244 }
 245
 246 /* Write to the ring buffer. */
 247 int hv_ringbuffer_write(struct vmbus_channel *channel,
 248                         const struct kvec *kv_list, u32 kv_count)
 249 {
 250         int i;
 251         u32 bytes_avail_towrite;
 252         u32 totalbytes_towrite = sizeof(u64);
 253         u32 next_write_location;
 254         u32 old_write;
 255         u64 prev_indices;
 256         unsigned long flags;
 257         struct hv_ring_buffer_info *outring_info = &channel->outbound;
 258
 259         if (channel->rescind)
 260                 return -ENODEV;
 261
 262         for (i = 0; i < kv_count; i++)
 263                 totalbytes_towrite += kv_list[i].iov_len;
 264
 265         spin_lock_irqsave(&outring_info->ring_lock, flags);
 266
 267         bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
 268
 269         /*
 270          * If there is only room for the packet, assume it is full.
 271          * Otherwise, the next time around, we think the ring buffer
 272          * is empty since the read index == write index.
 273          */
 274         if (bytes_avail_towrite <= totalbytes_towrite) {
 275                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 276                 return -EAGAIN;
 277         }
 278
 279         /* Write to the ring buffer */
 280         next_write_location = hv_get_next_write_location(outring_info);
 281
 282         old_write = next_write_location;
 283
 284         for (i = 0; i < kv_count; i++) {
 285                 next_write_location = hv_copyto_ringbuffer(outring_info,
 286                                                      next_write_location,
 287                                                      kv_list[i].iov_base,
 288                                                      kv_list[i].iov_len);
 289         }
 290
 291         /* Set previous packet start */
 292         prev_indices = hv_get_ring_bufferindices(outring_info);
 293
 294         next_write_location = hv_copyto_ringbuffer(outring_info,
 295                                              next_write_location,
 296                                              &prev_indices,
 297                                              sizeof(u64));
 298
 299         /* Issue a full memory barrier before updating the write index */
 300         virt_mb();
 301
 302         /* Now, update the write location */
 303         hv_set_next_write_location(outring_info, next_write_location);
 304
 305
 306         spin_unlock_irqrestore(&outring_info->ring_lock, flags);
 307
 308         hv_signal_on_write(old_write, channel);
 309
 310         if (channel->rescind)
 311                 return -ENODEV;
 312
 313         return 0;
 314 }
 315
 316 int hv_ringbuffer_read(struct vmbus_channel *channel,
 317                        void *buffer, u32 buflen, u32 *buffer_actual_len,
 318                        u64 *requestid, bool raw)
 319 {
 320         struct vmpacket_descriptor *desc;
 321         u32 packetlen, offset;
 322
 323         if (unlikely(buflen == 0))
 324                 return -EINVAL;
 325
 326         *buffer_actual_len = 0;
 327         *requestid = 0;
 328
 329         /* Make sure there is something to read */
 330         desc = hv_pkt_iter_first(channel);
 331         if (desc == NULL) {
 332                 /*
 333                  * No error is set when there is even no header, drivers are
 334                  * supposed to analyze buffer_actual_len.
 335                  */
 336                 return 0;
 337         }
 338
 339         offset = raw ? 0 : (desc->offset8 << 3);
 340         packetlen = (desc->len8 << 3) - offset;
 341         *buffer_actual_len = packetlen;
 342         *requestid = desc->trans_id;
 343
 344         if (unlikely(packetlen > buflen))
 345                 return -ENOBUFS;
 346
 347         /* since ring is double mapped, only one copy is necessary */
 348         memcpy(buffer, (const char *)desc + offset, packetlen);
 349
 350         /* Advance ring index to next packet descriptor */
 351         __hv_pkt_iter_next(channel, desc);
 352
 353         /* Notify host of update */
 354         hv_pkt_iter_close(channel);
 355
 356         return 0;
 357 }
 358
 359 /*
 360  * Determine number of bytes available in ring buffer after
 361  * the current iterator (priv_read_index) location.
 362  *
 363  * This is similar to hv_get_bytes_to_read but with private
 364  * read index instead.
 365  */
 366 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
 367 {
 368         u32 priv_read_loc = rbi->priv_read_index;
 369         u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
 370
 371         if (write_loc >= priv_read_loc)
 372                 return write_loc - priv_read_loc;
 373         else
 374                 return (rbi->ring_datasize - priv_read_loc) + write_loc;
 375 }
 376
 377 /*
 378  * Get first vmbus packet from ring buffer after read_index
 379  *
 380  * If ring buffer is empty, returns NULL and no other action needed.
 381  */
 382 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
 383 {
 384         struct hv_ring_buffer_info *rbi = &channel->inbound;
 385         struct vmpacket_descriptor *desc;
 386
 387         if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
 388                 return NULL;
 389
 390         desc = hv_get_ring_buffer(rbi) + rbi->priv_read_index;
 391         if (desc)
 392                 prefetch((char *)desc + (desc->len8 << 3));
 393
 394         return desc;
 395 }
 396 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
 397
 398 /*
 399  * Get next vmbus packet from ring buffer.
 400  *
 401  * Advances the current location (priv_read_index) and checks for more
 402  * data. If the end of the ring buffer is reached, then return NULL.
 403  */
 404 struct vmpacket_descriptor *
 405 __hv_pkt_iter_next(struct vmbus_channel *channel,
 406                    const struct vmpacket_descriptor *desc)
 407 {
 408         struct hv_ring_buffer_info *rbi = &channel->inbound;
 409         u32 packetlen = desc->len8 << 3;
 410         u32 dsize = rbi->ring_datasize;
 411
 412         /* bump offset to next potential packet */
 413         rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
 414         if (rbi->priv_read_index >= dsize)
 415                 rbi->priv_read_index -= dsize;
 416
 417         /* more data? */
 418         return hv_pkt_iter_first(channel);
 419 }
 420 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
 421
 422 /* How many bytes were read in this iterator cycle */
 423 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
 424                                         u32 start_read_index)
 425 {
 426         if (rbi->priv_read_index >= start_read_index)
 427                 return rbi->priv_read_index - start_read_index;
 428         else
 429                 return rbi->ring_datasize - start_read_index +
 430                         rbi->priv_read_index;
 431 }
 432
 433 /*
 434  * Update host ring buffer after iterating over packets.
 435  */
 436 void hv_pkt_iter_close(struct vmbus_channel *channel)
 437 {
 438         struct hv_ring_buffer_info *rbi = &channel->inbound;
 439         u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
 440
 441         /*
 442          * Make sure all reads are done before we update the read index since
 443          * the writer may start writing to the read area once the read index
 444          * is updated.
 445          */
 446         virt_rmb();
 447         start_read_index = rbi->ring_buffer->read_index;
 448         rbi->ring_buffer->read_index = rbi->priv_read_index;
 449
 450         if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
 451                 return;
 452
 453         /*
 454          * Issue a full memory barrier before making the signaling decision.
 455          * Here is the reason for having this barrier:
 456          * If the reading of the pend_sz (in this function)
 457          * were to be reordered and read before we commit the new read
 458          * index (in the calling function)  we could
 459          * have a problem. If the host were to set the pending_sz after we
 460          * have sampled pending_sz and go to sleep before we commit the
 461          * read index, we could miss sending the interrupt. Issue a full
 462          * memory barrier to address this.
 463          */
 464         virt_mb();
 465
 466         pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
 467         if (!pending_sz)
 468                 return;
 469
 470         /*
 471          * Ensure the read of write_index in hv_get_bytes_to_write()
 472          * happens after the read of pending_send_sz.
 473          */
 474         virt_rmb();
 475         curr_write_sz = hv_get_bytes_to_write(rbi);
 476         bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
 477
 478         /*
 479          * If there was space before we began iteration,
 480          * then host was not blocked.
 481          */
 482
 483         if (curr_write_sz - bytes_read > pending_sz)
 484                 return;
 485
 486         /* If pending write will not fit, don't give false hope. */
 487         if (curr_write_sz <= pending_sz)
 488                 return;
 489
 490         vmbus_setevent(channel);
 491 }
 492 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);