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 <asm/mshyperv.h>
23 #include "hyperv_vmbus.h"
25 #define VMBUS_PKT_TRAILER 8
28 * When we write to the ring buffer, check if the host needs to
29 * be signaled. Here is the details of this protocol:
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
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
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
48 * So, base the signaling decision solely on the ring state until the
49 * host logic is fixed.
52 static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)
54 struct hv_ring_buffer_info *rbi = &channel->outbound;
57 if (READ_ONCE(rbi->ring_buffer->interrupt_mask))
60 /* check interrupt_mask before read_index */
63 * This is the only case we need to signal when the
64 * ring transitions from being empty to non-empty.
66 if (old_write == READ_ONCE(rbi->ring_buffer->read_index)) {
67 ++channel->intr_out_empty;
68 vmbus_setevent(channel);
72 /* Get the next write location for the specified ring buffer. */
74 hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
76 u32 next = ring_info->ring_buffer->write_index;
81 /* Set the next write location for the specified ring buffer. */
83 hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
84 u32 next_write_location)
86 ring_info->ring_buffer->write_index = next_write_location;
89 /* Get the size of the ring buffer. */
91 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
93 return ring_info->ring_datasize;
96 /* Get the read and write indices as u64 of the specified ring buffer. */
98 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
100 return (u64)ring_info->ring_buffer->write_index << 32;
104 * Helper routine to copy from source to ring buffer.
105 * Assume there is enough room. Handles wrap-around in dest case only!!
107 static u32 hv_copyto_ringbuffer(
108 struct hv_ring_buffer_info *ring_info,
109 u32 start_write_offset,
113 void *ring_buffer = hv_get_ring_buffer(ring_info);
114 u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
116 memcpy(ring_buffer + start_write_offset, src, srclen);
118 start_write_offset += srclen;
119 if (start_write_offset >= ring_buffer_size)
120 start_write_offset -= ring_buffer_size;
122 return start_write_offset;
127 * hv_get_ringbuffer_availbytes()
129 * Get number of bytes available to read and to write to
130 * for the specified ring buffer
133 hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
134 u32 *read, u32 *write)
136 u32 read_loc, write_loc, dsize;
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;
143 *write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
144 read_loc - write_loc;
145 *read = dsize - *write;
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)
152 u32 bytes_avail_towrite;
153 u32 bytes_avail_toread;
155 mutex_lock(&ring_info->ring_buffer_mutex);
157 if (!ring_info->ring_buffer) {
158 mutex_unlock(&ring_info->ring_buffer_mutex);
162 hv_get_ringbuffer_availbytes(ring_info,
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);
175 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
177 /* Initialize a channel's ring buffer info mutex locks */
178 void hv_ringbuffer_pre_init(struct vmbus_channel *channel)
180 mutex_init(&channel->inbound.ring_buffer_mutex);
181 mutex_init(&channel->outbound.ring_buffer_mutex);
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)
188 struct page **pages_wraparound;
189 unsigned long *pfns_wraparound;
193 BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
196 * First page holds struct hv_ring_buffer, do wraparound mapping for
199 if (hv_isolation_type_snp()) {
200 pfn = page_to_pfn(pages) +
201 PFN_DOWN(ms_hyperv.shared_gpa_boundary);
203 pfns_wraparound = kcalloc(page_cnt * 2 - 1,
204 sizeof(unsigned long), GFP_KERNEL);
205 if (!pfns_wraparound)
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;
212 ring_info->ring_buffer = (struct hv_ring_buffer *)
213 vmap_pfn(pfns_wraparound, page_cnt * 2 - 1,
215 kfree(pfns_wraparound);
217 if (!ring_info->ring_buffer)
220 /* Zero ring buffer after setting memory host visibility. */
221 memset(ring_info->ring_buffer, 0x00, PAGE_SIZE * page_cnt);
223 pages_wraparound = kcalloc(page_cnt * 2 - 1,
224 sizeof(struct page *),
226 if (!pages_wraparound)
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];
234 ring_info->ring_buffer = (struct hv_ring_buffer *)
235 vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP,
238 kfree(pages_wraparound);
239 if (!ring_info->ring_buffer)
244 ring_info->ring_buffer->read_index =
245 ring_info->ring_buffer->write_index = 0;
247 /* Set the feature bit for enabling flow control. */
248 ring_info->ring_buffer->feature_bits.value = 1;
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;
257 /* Initialize buffer that holds copies of incoming packets */
259 ring_info->pkt_buffer = kzalloc(max_pkt_size, GFP_KERNEL);
260 if (!ring_info->pkt_buffer)
262 ring_info->pkt_buffer_size = max_pkt_size;
265 spin_lock_init(&ring_info->ring_lock);
270 /* Cleanup the ring buffer. */
271 void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
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);
278 kfree(ring_info->pkt_buffer);
279 ring_info->pkt_buffer = NULL;
280 ring_info->pkt_buffer_size = 0;
283 /* Write to the ring buffer. */
284 int hv_ringbuffer_write(struct vmbus_channel *channel,
285 const struct kvec *kv_list, u32 kv_count,
289 u32 bytes_avail_towrite;
290 u32 totalbytes_towrite = sizeof(u64);
291 u32 next_write_location;
295 struct hv_ring_buffer_info *outring_info = &channel->outbound;
296 struct vmpacket_descriptor *desc = kv_list[0].iov_base;
297 u64 rqst_id = VMBUS_NO_RQSTOR;
299 if (channel->rescind)
302 for (i = 0; i < kv_count; i++)
303 totalbytes_towrite += kv_list[i].iov_len;
305 spin_lock_irqsave(&outring_info->ring_lock, flags);
307 bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
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.
314 if (bytes_avail_towrite <= totalbytes_towrite) {
315 ++channel->out_full_total;
317 if (!channel->out_full_flag) {
318 ++channel->out_full_first;
319 channel->out_full_flag = true;
322 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
326 channel->out_full_flag = false;
328 /* Write to the ring buffer */
329 next_write_location = hv_get_next_write_location(outring_info);
331 old_write = next_write_location;
333 for (i = 0; i < kv_count; i++) {
334 next_write_location = hv_copyto_ringbuffer(outring_info,
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.
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);
355 desc = hv_get_ring_buffer(outring_info) + old_write;
356 desc->trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;
358 /* Set previous packet start */
359 prev_indices = hv_get_ring_bufferindices(outring_info);
361 next_write_location = hv_copyto_ringbuffer(outring_info,
366 /* Issue a full memory barrier before updating the write index */
369 /* Now, update the write location */
370 hv_set_next_write_location(outring_info, next_write_location);
373 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
375 hv_signal_on_write(old_write, channel);
377 if (channel->rescind) {
378 if (rqst_id != VMBUS_NO_RQSTOR) {
379 /* Reclaim request ID to avoid leak of IDs */
380 if (channel->request_addr_callback != NULL)
381 channel->request_addr_callback(channel, rqst_id);
389 int hv_ringbuffer_read(struct vmbus_channel *channel,
390 void *buffer, u32 buflen, u32 *buffer_actual_len,
391 u64 *requestid, bool raw)
393 struct vmpacket_descriptor *desc;
394 u32 packetlen, offset;
396 if (unlikely(buflen == 0))
399 *buffer_actual_len = 0;
402 /* Make sure there is something to read */
403 desc = hv_pkt_iter_first(channel);
406 * No error is set when there is even no header, drivers are
407 * supposed to analyze buffer_actual_len.
412 offset = raw ? 0 : (desc->offset8 << 3);
413 packetlen = (desc->len8 << 3) - offset;
414 *buffer_actual_len = packetlen;
415 *requestid = desc->trans_id;
417 if (unlikely(packetlen > buflen))
420 /* since ring is double mapped, only one copy is necessary */
421 memcpy(buffer, (const char *)desc + offset, packetlen);
423 /* Advance ring index to next packet descriptor */
424 __hv_pkt_iter_next(channel, desc, true);
426 /* Notify host of update */
427 hv_pkt_iter_close(channel);
433 * Determine number of bytes available in ring buffer after
434 * the current iterator (priv_read_index) location.
436 * This is similar to hv_get_bytes_to_read but with private
437 * read index instead.
439 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
441 u32 priv_read_loc = rbi->priv_read_index;
442 u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
444 if (write_loc >= priv_read_loc)
445 return write_loc - priv_read_loc;
447 return (rbi->ring_datasize - priv_read_loc) + write_loc;
451 * Get first vmbus packet without copying it out of the ring buffer
453 struct vmpacket_descriptor *hv_pkt_iter_first_raw(struct vmbus_channel *channel)
455 struct hv_ring_buffer_info *rbi = &channel->inbound;
457 hv_debug_delay_test(channel, MESSAGE_DELAY);
459 if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
462 return (struct vmpacket_descriptor *)(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
464 EXPORT_SYMBOL_GPL(hv_pkt_iter_first_raw);
467 * Get first vmbus packet from ring buffer after read_index
469 * If ring buffer is empty, returns NULL and no other action needed.
471 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
473 struct hv_ring_buffer_info *rbi = &channel->inbound;
474 struct vmpacket_descriptor *desc, *desc_copy;
475 u32 bytes_avail, pkt_len, pkt_offset;
477 desc = hv_pkt_iter_first_raw(channel);
481 bytes_avail = min(rbi->pkt_buffer_size, hv_pkt_iter_avail(rbi));
484 * Ensure the compiler does not use references to incoming Hyper-V values (which
485 * could change at any moment) when reading local variables later in the code
487 pkt_len = READ_ONCE(desc->len8) << 3;
488 pkt_offset = READ_ONCE(desc->offset8) << 3;
491 * If pkt_len is invalid, set it to the smaller of hv_pkt_iter_avail() and
492 * rbi->pkt_buffer_size
494 if (pkt_len < sizeof(struct vmpacket_descriptor) || pkt_len > bytes_avail)
495 pkt_len = bytes_avail;
498 * If pkt_offset is invalid, arbitrarily set it to
499 * the size of vmpacket_descriptor
501 if (pkt_offset < sizeof(struct vmpacket_descriptor) || pkt_offset > pkt_len)
502 pkt_offset = sizeof(struct vmpacket_descriptor);
504 /* Copy the Hyper-V packet out of the ring buffer */
505 desc_copy = (struct vmpacket_descriptor *)rbi->pkt_buffer;
506 memcpy(desc_copy, desc, pkt_len);
509 * Hyper-V could still change len8 and offset8 after the earlier read.
510 * Ensure that desc_copy has legal values for len8 and offset8 that
511 * are consistent with the copy we just made
513 desc_copy->len8 = pkt_len >> 3;
514 desc_copy->offset8 = pkt_offset >> 3;
518 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
521 * Get next vmbus packet from ring buffer.
523 * Advances the current location (priv_read_index) and checks for more
524 * data. If the end of the ring buffer is reached, then return NULL.
526 struct vmpacket_descriptor *
527 __hv_pkt_iter_next(struct vmbus_channel *channel,
528 const struct vmpacket_descriptor *desc,
531 struct hv_ring_buffer_info *rbi = &channel->inbound;
532 u32 packetlen = desc->len8 << 3;
533 u32 dsize = rbi->ring_datasize;
535 hv_debug_delay_test(channel, MESSAGE_DELAY);
536 /* bump offset to next potential packet */
537 rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
538 if (rbi->priv_read_index >= dsize)
539 rbi->priv_read_index -= dsize;
542 return copy ? hv_pkt_iter_first(channel) : hv_pkt_iter_first_raw(channel);
544 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
546 /* How many bytes were read in this iterator cycle */
547 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
548 u32 start_read_index)
550 if (rbi->priv_read_index >= start_read_index)
551 return rbi->priv_read_index - start_read_index;
553 return rbi->ring_datasize - start_read_index +
554 rbi->priv_read_index;
558 * Update host ring buffer after iterating over packets. If the host has
559 * stopped queuing new entries because it found the ring buffer full, and
560 * sufficient space is being freed up, signal the host. But be careful to
561 * only signal the host when necessary, both for performance reasons and
562 * because Hyper-V protects itself by throttling guests that signal
565 * Determining when to signal is tricky. There are three key data inputs
566 * that must be handled in this order to avoid race conditions:
568 * 1. Update the read_index
569 * 2. Read the pending_send_sz
570 * 3. Read the current write_index
572 * The interrupt_mask is not used to determine when to signal. The
573 * interrupt_mask is used only on the guest->host ring buffer when
574 * sending requests to the host. The host does not use it on the host->
575 * guest ring buffer to indicate whether it should be signaled.
577 void hv_pkt_iter_close(struct vmbus_channel *channel)
579 struct hv_ring_buffer_info *rbi = &channel->inbound;
580 u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
583 * Make sure all reads are done before we update the read index since
584 * the writer may start writing to the read area once the read index
588 start_read_index = rbi->ring_buffer->read_index;
589 rbi->ring_buffer->read_index = rbi->priv_read_index;
592 * Older versions of Hyper-V (before WS2102 and Win8) do not
593 * implement pending_send_sz and simply poll if the host->guest
594 * ring buffer is full. No signaling is needed or expected.
596 if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
600 * Issue a full memory barrier before making the signaling decision.
601 * If reading pending_send_sz were to be reordered and happen
602 * before we commit the new read_index, a race could occur. If the
603 * host were to set the pending_send_sz after we have sampled
604 * pending_send_sz, and the ring buffer blocks before we commit the
605 * read index, we could miss sending the interrupt. Issue a full
606 * memory barrier to address this.
611 * If the pending_send_sz is zero, then the ring buffer is not
612 * blocked and there is no need to signal. This is far by the
613 * most common case, so exit quickly for best performance.
615 pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
620 * Ensure the read of write_index in hv_get_bytes_to_write()
621 * happens after the read of pending_send_sz.
624 curr_write_sz = hv_get_bytes_to_write(rbi);
625 bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
628 * We want to signal the host only if we're transitioning
629 * from a "not enough free space" state to a "enough free
630 * space" state. For example, it's possible that this function
631 * could run and free up enough space to signal the host, and then
632 * run again and free up additional space before the host has a
633 * chance to clear the pending_send_sz. The 2nd invocation would
634 * be a null transition from "enough free space" to "enough free
635 * space", which doesn't warrant a signal.
637 * Exactly filling the ring buffer is treated as "not enough
638 * space". The ring buffer always must have at least one byte
639 * empty so the empty and full conditions are distinguishable.
640 * hv_get_bytes_to_write() doesn't fully tell the truth in
643 * So first check if we were in the "enough free space" state
644 * before we began the iteration. If so, the host was not
645 * blocked, and there's no need to signal.
647 if (curr_write_sz - bytes_read > pending_sz)
651 * Similarly, if the new state is "not enough space", then
652 * there's no need to signal.
654 if (curr_write_sz <= pending_sz)
657 ++channel->intr_in_full;
658 vmbus_setevent(channel);
660 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);