1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright (c) 2009, Microsoft Corporation.
7 * Haiyang Zhang <haiyangz@microsoft.com>
8 * Hank Janssen <hjanssen@microsoft.com>
9 * K. Y. Srinivasan <kys@microsoft.com>
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 #include <linux/kernel.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>
21 #include "hyperv_vmbus.h"
23 #define VMBUS_PKT_TRAILER 8
26 * When we write to the ring buffer, check if the host needs to
27 * be signaled. Here is the details of this protocol:
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
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
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
46 * So, base the signaling decision solely on the ring state until the
47 * host logic is fixed.
50 static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)
52 struct hv_ring_buffer_info *rbi = &channel->outbound;
55 if (READ_ONCE(rbi->ring_buffer->interrupt_mask))
58 /* check interrupt_mask before read_index */
61 * This is the only case we need to signal when the
62 * ring transitions from being empty to non-empty.
64 if (old_write == READ_ONCE(rbi->ring_buffer->read_index)) {
65 ++channel->intr_out_empty;
66 vmbus_setevent(channel);
70 /* Get the next write location for the specified ring buffer. */
72 hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
74 u32 next = ring_info->ring_buffer->write_index;
79 /* Set the next write location for the specified ring buffer. */
81 hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
82 u32 next_write_location)
84 ring_info->ring_buffer->write_index = next_write_location;
87 /* Get the size of the ring buffer. */
89 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
91 return ring_info->ring_datasize;
94 /* Get the read and write indices as u64 of the specified ring buffer. */
96 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
98 return (u64)ring_info->ring_buffer->write_index << 32;
102 * Helper routine to copy from source to ring buffer.
103 * Assume there is enough room. Handles wrap-around in dest case only!!
105 static u32 hv_copyto_ringbuffer(
106 struct hv_ring_buffer_info *ring_info,
107 u32 start_write_offset,
111 void *ring_buffer = hv_get_ring_buffer(ring_info);
112 u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
114 memcpy(ring_buffer + start_write_offset, src, srclen);
116 start_write_offset += srclen;
117 if (start_write_offset >= ring_buffer_size)
118 start_write_offset -= ring_buffer_size;
120 return start_write_offset;
125 * hv_get_ringbuffer_availbytes()
127 * Get number of bytes available to read and to write to
128 * for the specified ring buffer
131 hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
132 u32 *read, u32 *write)
134 u32 read_loc, write_loc, dsize;
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;
141 *write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
142 read_loc - write_loc;
143 *read = dsize - *write;
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)
150 u32 bytes_avail_towrite;
151 u32 bytes_avail_toread;
153 mutex_lock(&ring_info->ring_buffer_mutex);
155 if (!ring_info->ring_buffer) {
156 mutex_unlock(&ring_info->ring_buffer_mutex);
160 hv_get_ringbuffer_availbytes(ring_info,
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);
173 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
175 /* Initialize a channel's ring buffer info mutex locks */
176 void hv_ringbuffer_pre_init(struct vmbus_channel *channel)
178 mutex_init(&channel->inbound.ring_buffer_mutex);
179 mutex_init(&channel->outbound.ring_buffer_mutex);
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)
187 struct page **pages_wraparound;
189 BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
192 * First page holds struct hv_ring_buffer, do wraparound mapping for
195 pages_wraparound = kcalloc(page_cnt * 2 - 1, sizeof(struct page *),
197 if (!pages_wraparound)
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];
204 ring_info->ring_buffer = (struct hv_ring_buffer *)
205 vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP, PAGE_KERNEL);
207 kfree(pages_wraparound);
210 if (!ring_info->ring_buffer)
213 ring_info->ring_buffer->read_index =
214 ring_info->ring_buffer->write_index = 0;
216 /* Set the feature bit for enabling flow control. */
217 ring_info->ring_buffer->feature_bits.value = 1;
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;
226 /* Initialize buffer that holds copies of incoming packets */
228 ring_info->pkt_buffer = kzalloc(max_pkt_size, GFP_KERNEL);
229 if (!ring_info->pkt_buffer)
231 ring_info->pkt_buffer_size = max_pkt_size;
234 spin_lock_init(&ring_info->ring_lock);
239 /* Cleanup the ring buffer. */
240 void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
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);
247 kfree(ring_info->pkt_buffer);
248 ring_info->pkt_buffer_size = 0;
251 /* Write to the ring buffer. */
252 int hv_ringbuffer_write(struct vmbus_channel *channel,
253 const struct kvec *kv_list, u32 kv_count,
257 u32 bytes_avail_towrite;
258 u32 totalbytes_towrite = sizeof(u64);
259 u32 next_write_location;
263 struct hv_ring_buffer_info *outring_info = &channel->outbound;
264 struct vmpacket_descriptor *desc = kv_list[0].iov_base;
265 u64 rqst_id = VMBUS_NO_RQSTOR;
267 if (channel->rescind)
270 for (i = 0; i < kv_count; i++)
271 totalbytes_towrite += kv_list[i].iov_len;
273 spin_lock_irqsave(&outring_info->ring_lock, flags);
275 bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
278 * If there is only room for the packet, assume it is full.
279 * Otherwise, the next time around, we think the ring buffer
280 * is empty since the read index == write index.
282 if (bytes_avail_towrite <= totalbytes_towrite) {
283 ++channel->out_full_total;
285 if (!channel->out_full_flag) {
286 ++channel->out_full_first;
287 channel->out_full_flag = true;
290 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
294 channel->out_full_flag = false;
296 /* Write to the ring buffer */
297 next_write_location = hv_get_next_write_location(outring_info);
299 old_write = next_write_location;
301 for (i = 0; i < kv_count; i++) {
302 next_write_location = hv_copyto_ringbuffer(outring_info,
309 * Allocate the request ID after the data has been copied into the
310 * ring buffer. Once this request ID is allocated, the completion
311 * path could find the data and free it.
314 if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {
315 rqst_id = vmbus_next_request_id(&channel->requestor, requestid);
316 if (rqst_id == VMBUS_RQST_ERROR) {
317 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
321 desc = hv_get_ring_buffer(outring_info) + old_write;
322 desc->trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;
324 /* Set previous packet start */
325 prev_indices = hv_get_ring_bufferindices(outring_info);
327 next_write_location = hv_copyto_ringbuffer(outring_info,
332 /* Issue a full memory barrier before updating the write index */
335 /* Now, update the write location */
336 hv_set_next_write_location(outring_info, next_write_location);
339 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
341 hv_signal_on_write(old_write, channel);
343 if (channel->rescind) {
344 if (rqst_id != VMBUS_NO_RQSTOR) {
345 /* Reclaim request ID to avoid leak of IDs */
346 vmbus_request_addr(&channel->requestor, rqst_id);
354 int hv_ringbuffer_read(struct vmbus_channel *channel,
355 void *buffer, u32 buflen, u32 *buffer_actual_len,
356 u64 *requestid, bool raw)
358 struct vmpacket_descriptor *desc;
359 u32 packetlen, offset;
361 if (unlikely(buflen == 0))
364 *buffer_actual_len = 0;
367 /* Make sure there is something to read */
368 desc = hv_pkt_iter_first(channel);
371 * No error is set when there is even no header, drivers are
372 * supposed to analyze buffer_actual_len.
377 offset = raw ? 0 : (desc->offset8 << 3);
378 packetlen = (desc->len8 << 3) - offset;
379 *buffer_actual_len = packetlen;
380 *requestid = desc->trans_id;
382 if (unlikely(packetlen > buflen))
385 /* since ring is double mapped, only one copy is necessary */
386 memcpy(buffer, (const char *)desc + offset, packetlen);
388 /* Advance ring index to next packet descriptor */
389 __hv_pkt_iter_next(channel, desc, true);
391 /* Notify host of update */
392 hv_pkt_iter_close(channel);
398 * Determine number of bytes available in ring buffer after
399 * the current iterator (priv_read_index) location.
401 * This is similar to hv_get_bytes_to_read but with private
402 * read index instead.
404 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
406 u32 priv_read_loc = rbi->priv_read_index;
407 u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
409 if (write_loc >= priv_read_loc)
410 return write_loc - priv_read_loc;
412 return (rbi->ring_datasize - priv_read_loc) + write_loc;
416 * Get first vmbus packet without copying it out of the ring buffer
418 struct vmpacket_descriptor *hv_pkt_iter_first_raw(struct vmbus_channel *channel)
420 struct hv_ring_buffer_info *rbi = &channel->inbound;
422 hv_debug_delay_test(channel, MESSAGE_DELAY);
424 if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
427 return (struct vmpacket_descriptor *)(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
429 EXPORT_SYMBOL_GPL(hv_pkt_iter_first_raw);
432 * Get first vmbus packet from ring buffer after read_index
434 * If ring buffer is empty, returns NULL and no other action needed.
436 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
438 struct hv_ring_buffer_info *rbi = &channel->inbound;
439 struct vmpacket_descriptor *desc, *desc_copy;
440 u32 bytes_avail, pkt_len, pkt_offset;
442 desc = hv_pkt_iter_first_raw(channel);
446 bytes_avail = min(rbi->pkt_buffer_size, hv_pkt_iter_avail(rbi));
449 * Ensure the compiler does not use references to incoming Hyper-V values (which
450 * could change at any moment) when reading local variables later in the code
452 pkt_len = READ_ONCE(desc->len8) << 3;
453 pkt_offset = READ_ONCE(desc->offset8) << 3;
456 * If pkt_len is invalid, set it to the smaller of hv_pkt_iter_avail() and
457 * rbi->pkt_buffer_size
459 if (pkt_len < sizeof(struct vmpacket_descriptor) || pkt_len > bytes_avail)
460 pkt_len = bytes_avail;
463 * If pkt_offset is invalid, arbitrarily set it to
464 * the size of vmpacket_descriptor
466 if (pkt_offset < sizeof(struct vmpacket_descriptor) || pkt_offset > pkt_len)
467 pkt_offset = sizeof(struct vmpacket_descriptor);
469 /* Copy the Hyper-V packet out of the ring buffer */
470 desc_copy = (struct vmpacket_descriptor *)rbi->pkt_buffer;
471 memcpy(desc_copy, desc, pkt_len);
474 * Hyper-V could still change len8 and offset8 after the earlier read.
475 * Ensure that desc_copy has legal values for len8 and offset8 that
476 * are consistent with the copy we just made
478 desc_copy->len8 = pkt_len >> 3;
479 desc_copy->offset8 = pkt_offset >> 3;
483 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
486 * Get next vmbus packet from ring buffer.
488 * Advances the current location (priv_read_index) and checks for more
489 * data. If the end of the ring buffer is reached, then return NULL.
491 struct vmpacket_descriptor *
492 __hv_pkt_iter_next(struct vmbus_channel *channel,
493 const struct vmpacket_descriptor *desc,
496 struct hv_ring_buffer_info *rbi = &channel->inbound;
497 u32 packetlen = desc->len8 << 3;
498 u32 dsize = rbi->ring_datasize;
500 hv_debug_delay_test(channel, MESSAGE_DELAY);
501 /* bump offset to next potential packet */
502 rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
503 if (rbi->priv_read_index >= dsize)
504 rbi->priv_read_index -= dsize;
507 return copy ? hv_pkt_iter_first(channel) : hv_pkt_iter_first_raw(channel);
509 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
511 /* How many bytes were read in this iterator cycle */
512 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
513 u32 start_read_index)
515 if (rbi->priv_read_index >= start_read_index)
516 return rbi->priv_read_index - start_read_index;
518 return rbi->ring_datasize - start_read_index +
519 rbi->priv_read_index;
523 * Update host ring buffer after iterating over packets. If the host has
524 * stopped queuing new entries because it found the ring buffer full, and
525 * sufficient space is being freed up, signal the host. But be careful to
526 * only signal the host when necessary, both for performance reasons and
527 * because Hyper-V protects itself by throttling guests that signal
530 * Determining when to signal is tricky. There are three key data inputs
531 * that must be handled in this order to avoid race conditions:
533 * 1. Update the read_index
534 * 2. Read the pending_send_sz
535 * 3. Read the current write_index
537 * The interrupt_mask is not used to determine when to signal. The
538 * interrupt_mask is used only on the guest->host ring buffer when
539 * sending requests to the host. The host does not use it on the host->
540 * guest ring buffer to indicate whether it should be signaled.
542 void hv_pkt_iter_close(struct vmbus_channel *channel)
544 struct hv_ring_buffer_info *rbi = &channel->inbound;
545 u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
548 * Make sure all reads are done before we update the read index since
549 * the writer may start writing to the read area once the read index
553 start_read_index = rbi->ring_buffer->read_index;
554 rbi->ring_buffer->read_index = rbi->priv_read_index;
557 * Older versions of Hyper-V (before WS2102 and Win8) do not
558 * implement pending_send_sz and simply poll if the host->guest
559 * ring buffer is full. No signaling is needed or expected.
561 if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
565 * Issue a full memory barrier before making the signaling decision.
566 * If reading pending_send_sz were to be reordered and happen
567 * before we commit the new read_index, a race could occur. If the
568 * host were to set the pending_send_sz after we have sampled
569 * pending_send_sz, and the ring buffer blocks before we commit the
570 * read index, we could miss sending the interrupt. Issue a full
571 * memory barrier to address this.
576 * If the pending_send_sz is zero, then the ring buffer is not
577 * blocked and there is no need to signal. This is far by the
578 * most common case, so exit quickly for best performance.
580 pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
585 * Ensure the read of write_index in hv_get_bytes_to_write()
586 * happens after the read of pending_send_sz.
589 curr_write_sz = hv_get_bytes_to_write(rbi);
590 bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
593 * We want to signal the host only if we're transitioning
594 * from a "not enough free space" state to a "enough free
595 * space" state. For example, it's possible that this function
596 * could run and free up enough space to signal the host, and then
597 * run again and free up additional space before the host has a
598 * chance to clear the pending_send_sz. The 2nd invocation would
599 * be a null transition from "enough free space" to "enough free
600 * space", which doesn't warrant a signal.
602 * Exactly filling the ring buffer is treated as "not enough
603 * space". The ring buffer always must have at least one byte
604 * empty so the empty and full conditions are distinguishable.
605 * hv_get_bytes_to_write() doesn't fully tell the truth in
608 * So first check if we were in the "enough free space" state
609 * before we began the iteration. If so, the host was not
610 * blocked, and there's no need to signal.
612 if (curr_write_sz - bytes_read > pending_sz)
616 * Similarly, if the new state is "not enough space", then
617 * there's no need to signal.
619 if (curr_write_sz <= pending_sz)
622 ++channel->intr_in_full;
623 vmbus_setevent(channel);
625 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);