3 * Copyright (c) 2009, Microsoft Corporation.
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.
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
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.
19 * Haiyang Zhang <haiyangz@microsoft.com>
20 * Hank Janssen <hjanssen@microsoft.com>
21 * K. Y. Srinivasan <kys@microsoft.com>
24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
26 #include <linux/kernel.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>
34 #include "hyperv_vmbus.h"
36 #define VMBUS_PKT_TRAILER 8
39 * When we write to the ring buffer, check if the host needs to
40 * be signaled. Here is the details of this protocol:
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
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
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
59 * So, base the signaling decision solely on the ring state until the
60 * host logic is fixed.
63 static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)
65 struct hv_ring_buffer_info *rbi = &channel->outbound;
68 if (READ_ONCE(rbi->ring_buffer->interrupt_mask))
71 /* check interrupt_mask before read_index */
74 * This is the only case we need to signal when the
75 * ring transitions from being empty to non-empty.
77 if (old_write == READ_ONCE(rbi->ring_buffer->read_index))
78 vmbus_setevent(channel);
81 /* Get the next write location for the specified ring buffer. */
83 hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
85 u32 next = ring_info->ring_buffer->write_index;
90 /* Set the next write location for the specified ring buffer. */
92 hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
93 u32 next_write_location)
95 ring_info->ring_buffer->write_index = next_write_location;
98 /* Set the next read location for the specified ring buffer. */
100 hv_set_next_read_location(struct hv_ring_buffer_info *ring_info,
101 u32 next_read_location)
103 ring_info->ring_buffer->read_index = next_read_location;
104 ring_info->priv_read_index = next_read_location;
107 /* Get the size of the ring buffer. */
109 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
111 return ring_info->ring_datasize;
114 /* Get the read and write indices as u64 of the specified ring buffer. */
116 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
118 return (u64)ring_info->ring_buffer->write_index << 32;
122 * Helper routine to copy from source to ring buffer.
123 * Assume there is enough room. Handles wrap-around in dest case only!!
125 static u32 hv_copyto_ringbuffer(
126 struct hv_ring_buffer_info *ring_info,
127 u32 start_write_offset,
131 void *ring_buffer = hv_get_ring_buffer(ring_info);
132 u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
134 memcpy(ring_buffer + start_write_offset, src, srclen);
136 start_write_offset += srclen;
137 if (start_write_offset >= ring_buffer_size)
138 start_write_offset -= ring_buffer_size;
140 return start_write_offset;
145 * hv_get_ringbuffer_availbytes()
147 * Get number of bytes available to read and to write to
148 * for the specified ring buffer
151 hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
152 u32 *read, u32 *write)
154 u32 read_loc, write_loc, dsize;
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;
161 *write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
162 read_loc - write_loc;
163 *read = dsize - *write;
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)
170 u32 bytes_avail_towrite;
171 u32 bytes_avail_toread;
173 if (ring_info->ring_buffer) {
174 hv_get_ringbuffer_availbytes(ring_info,
176 &bytes_avail_towrite);
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;
188 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
190 /* Initialize the ring buffer. */
191 int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
192 struct page *pages, u32 page_cnt)
195 struct page **pages_wraparound;
197 BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
199 memset(ring_info, 0, sizeof(struct hv_ring_buffer_info));
202 * First page holds struct hv_ring_buffer, do wraparound mapping for
205 pages_wraparound = kcalloc(page_cnt * 2 - 1, sizeof(struct page *),
207 if (!pages_wraparound)
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];
214 ring_info->ring_buffer = (struct hv_ring_buffer *)
215 vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP, PAGE_KERNEL);
217 kfree(pages_wraparound);
220 if (!ring_info->ring_buffer)
223 ring_info->ring_buffer->read_index =
224 ring_info->ring_buffer->write_index = 0;
226 /* Set the feature bit for enabling flow control. */
227 ring_info->ring_buffer->feature_bits.value = 1;
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);
235 spin_lock_init(&ring_info->ring_lock);
240 /* Cleanup the ring buffer. */
241 void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
243 vunmap(ring_info->ring_buffer);
244 ring_info->ring_buffer = NULL;
247 /* Write to the ring buffer. */
248 int hv_ringbuffer_write(struct vmbus_channel *channel,
249 const struct kvec *kv_list, u32 kv_count)
252 u32 bytes_avail_towrite;
253 u32 totalbytes_towrite = sizeof(u64);
254 u32 next_write_location;
258 struct hv_ring_buffer_info *outring_info = &channel->outbound;
260 if (channel->rescind)
263 for (i = 0; i < kv_count; i++)
264 totalbytes_towrite += kv_list[i].iov_len;
266 spin_lock_irqsave(&outring_info->ring_lock, flags);
268 bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
271 * If there is only room for the packet, assume it is full.
272 * Otherwise, the next time around, we think the ring buffer
273 * is empty since the read index == write index.
275 if (bytes_avail_towrite <= totalbytes_towrite) {
276 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
280 /* Write to the ring buffer */
281 next_write_location = hv_get_next_write_location(outring_info);
283 old_write = next_write_location;
285 for (i = 0; i < kv_count; i++) {
286 next_write_location = hv_copyto_ringbuffer(outring_info,
292 /* Set previous packet start */
293 prev_indices = hv_get_ring_bufferindices(outring_info);
295 next_write_location = hv_copyto_ringbuffer(outring_info,
300 /* Issue a full memory barrier before updating the write index */
303 /* Now, update the write location */
304 hv_set_next_write_location(outring_info, next_write_location);
307 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
309 hv_signal_on_write(old_write, channel);
311 if (channel->rescind)
317 int hv_ringbuffer_read(struct vmbus_channel *channel,
318 void *buffer, u32 buflen, u32 *buffer_actual_len,
319 u64 *requestid, bool raw)
321 struct vmpacket_descriptor *desc;
322 u32 packetlen, offset;
324 if (unlikely(buflen == 0))
327 *buffer_actual_len = 0;
330 /* Make sure there is something to read */
331 desc = hv_pkt_iter_first(channel);
334 * No error is set when there is even no header, drivers are
335 * supposed to analyze buffer_actual_len.
340 offset = raw ? 0 : (desc->offset8 << 3);
341 packetlen = (desc->len8 << 3) - offset;
342 *buffer_actual_len = packetlen;
343 *requestid = desc->trans_id;
345 if (unlikely(packetlen > buflen))
348 /* since ring is double mapped, only one copy is necessary */
349 memcpy(buffer, (const char *)desc + offset, packetlen);
351 /* Advance ring index to next packet descriptor */
352 __hv_pkt_iter_next(channel, desc);
354 /* Notify host of update */
355 hv_pkt_iter_close(channel);
361 * Determine number of bytes available in ring buffer after
362 * the current iterator (priv_read_index) location.
364 * This is similar to hv_get_bytes_to_read but with private
365 * read index instead.
367 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
369 u32 priv_read_loc = rbi->priv_read_index;
370 u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
372 if (write_loc >= priv_read_loc)
373 return write_loc - priv_read_loc;
375 return (rbi->ring_datasize - priv_read_loc) + write_loc;
379 * Get first vmbus packet from ring buffer after read_index
381 * If ring buffer is empty, returns NULL and no other action needed.
383 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
385 struct hv_ring_buffer_info *rbi = &channel->inbound;
386 struct vmpacket_descriptor *desc;
388 if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
391 desc = hv_get_ring_buffer(rbi) + rbi->priv_read_index;
393 prefetch((char *)desc + (desc->len8 << 3));
397 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
400 * Get next vmbus packet from ring buffer.
402 * Advances the current location (priv_read_index) and checks for more
403 * data. If the end of the ring buffer is reached, then return NULL.
405 struct vmpacket_descriptor *
406 __hv_pkt_iter_next(struct vmbus_channel *channel,
407 const struct vmpacket_descriptor *desc)
409 struct hv_ring_buffer_info *rbi = &channel->inbound;
410 u32 packetlen = desc->len8 << 3;
411 u32 dsize = rbi->ring_datasize;
413 /* bump offset to next potential packet */
414 rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
415 if (rbi->priv_read_index >= dsize)
416 rbi->priv_read_index -= dsize;
419 return hv_pkt_iter_first(channel);
421 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
423 /* How many bytes were read in this iterator cycle */
424 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
425 u32 start_read_index)
427 if (rbi->priv_read_index >= start_read_index)
428 return rbi->priv_read_index - start_read_index;
430 return rbi->ring_datasize - start_read_index +
431 rbi->priv_read_index;
435 * Update host ring buffer after iterating over packets. If the host has
436 * stopped queuing new entries because it found the ring buffer full, and
437 * sufficient space is being freed up, signal the host. But be careful to
438 * only signal the host when necessary, both for performance reasons and
439 * because Hyper-V protects itself by throttling guests that signal
442 * Determining when to signal is tricky. There are three key data inputs
443 * that must be handled in this order to avoid race conditions:
445 * 1. Update the read_index
446 * 2. Read the pending_send_sz
447 * 3. Read the current write_index
449 * The interrupt_mask is not used to determine when to signal. The
450 * interrupt_mask is used only on the guest->host ring buffer when
451 * sending requests to the host. The host does not use it on the host->
452 * guest ring buffer to indicate whether it should be signaled.
454 void hv_pkt_iter_close(struct vmbus_channel *channel)
456 struct hv_ring_buffer_info *rbi = &channel->inbound;
457 u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
460 * Make sure all reads are done before we update the read index since
461 * the writer may start writing to the read area once the read index
465 start_read_index = rbi->ring_buffer->read_index;
466 rbi->ring_buffer->read_index = rbi->priv_read_index;
469 * Older versions of Hyper-V (before WS2102 and Win8) do not
470 * implement pending_send_sz and simply poll if the host->guest
471 * ring buffer is full. No signaling is needed or expected.
473 if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
477 * Issue a full memory barrier before making the signaling decision.
478 * If reading pending_send_sz were to be reordered and happen
479 * before we commit the new read_index, a race could occur. If the
480 * host were to set the pending_send_sz after we have sampled
481 * pending_send_sz, and the ring buffer blocks before we commit the
482 * read index, we could miss sending the interrupt. Issue a full
483 * memory barrier to address this.
488 * If the pending_send_sz is zero, then the ring buffer is not
489 * blocked and there is no need to signal. This is far by the
490 * most common case, so exit quickly for best performance.
492 pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
497 * Ensure the read of write_index in hv_get_bytes_to_write()
498 * happens after the read of pending_send_sz.
501 curr_write_sz = hv_get_bytes_to_write(rbi);
502 bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
505 * We want to signal the host only if we're transitioning
506 * from a "not enough free space" state to a "enough free
507 * space" state. For example, it's possible that this function
508 * could run and free up enough space to signal the host, and then
509 * run again and free up additional space before the host has a
510 * chance to clear the pending_send_sz. The 2nd invocation would
511 * be a null transition from "enough free space" to "enough free
512 * space", which doesn't warrant a signal.
514 * Exactly filling the ring buffer is treated as "not enough
515 * space". The ring buffer always must have at least one byte
516 * empty so the empty and full conditions are distinguishable.
517 * hv_get_bytes_to_write() doesn't fully tell the truth in
520 * So first check if we were in the "enough free space" state
521 * before we began the iteration. If so, the host was not
522 * blocked, and there's no need to signal.
524 if (curr_write_sz - bytes_read > pending_sz)
528 * Similarly, if the new state is "not enough space", then
529 * there's no need to signal.
531 if (curr_write_sz <= pending_sz)
534 vmbus_setevent(channel);
536 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);