projects/linux·blob master · refs

654 lines · 19.1 KiB

1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 *
4 * Copyright (c) 2009, Microsoft Corporation.
5 *
6 * Authors:
7 * Haiyang Zhang <[email protected]>
8 * Hank Janssen <[email protected]>
9 * K. Y. Srinivasan <[email protected]>
10 */
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12
13#include <linux/kernel.h>
14#include <linux/mm.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>
20#include <linux/io.h>
21#include <linux/export.h>
22#include <asm/mshyperv.h>
23
24#include "hyperv_vmbus.h"
25
26#define VMBUS_PKT_TRAILER 8
27
28/*
29 * When we write to the ring buffer, check if the host needs to
30 * be signaled. Here is the details of this protocol:
31 *
32 * 1. The host guarantees that while it is draining the
33 * ring buffer, it will set the interrupt_mask to
34 * indicate it does not need to be interrupted when
35 * new data is placed.
36 *
37 * 2. The host guarantees that it will completely drain
38 * the ring buffer before exiting the read loop. Further,
39 * once the ring buffer is empty, it will clear the
40 * interrupt_mask and re-check to see if new data has
41 * arrived.
42 *
43 * KYS: Oct. 30, 2016:
44 * It looks like Windows hosts have logic to deal with DOS attacks that
45 * can be triggered if it receives interrupts when it is not expecting
46 * the interrupt. The host expects interrupts only when the ring
47 * transitions from empty to non-empty (or full to non full on the guest
48 * to host ring).
49 * So, base the signaling decision solely on the ring state until the
50 * host logic is fixed.
51 */
52
53static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)
54{
55 struct hv_ring_buffer_info *rbi = &channel->outbound;
56
57 virt_mb();
58 if (READ_ONCE(rbi->ring_buffer->interrupt_mask))
59 return;
60
61 /* check interrupt_mask before read_index */
62 virt_rmb();
63 /*
64 * This is the only case we need to signal when the
65 * ring transitions from being empty to non-empty.
66 */
67 if (old_write == READ_ONCE(rbi->ring_buffer->read_index)) {
68 ++channel->intr_out_empty;
69 vmbus_setevent(channel);
70 }
71}
72
73/* Get the next write location for the specified ring buffer. */
74static inline u32
75hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
76{
77 u32 next = ring_info->ring_buffer->write_index;
78
79 return next;
80}
81
82/* Set the next write location for the specified ring buffer. */
83static inline void
84hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
85 u32 next_write_location)
86{
87 ring_info->ring_buffer->write_index = next_write_location;
88}
89
90/* Get the size of the ring buffer. */
91static inline u32
92hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
93{
94 return ring_info->ring_datasize;
95}
96
97/* Get the read and write indices as u64 of the specified ring buffer. */
98static inline u64
99hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
100{
101 return (u64)ring_info->ring_buffer->write_index << 32;
102}
103
104/*
105 * Helper routine to copy from source to ring buffer.
106 * Assume there is enough room. Handles wrap-around in dest case only!!
107 */
108static u32 hv_copyto_ringbuffer(
109 struct hv_ring_buffer_info *ring_info,
110 u32 start_write_offset,
111 const void *src,
112 u32 srclen)
113{
114 void *ring_buffer = hv_get_ring_buffer(ring_info);
115 u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
116
117 memcpy(ring_buffer + start_write_offset, src, srclen);
118
119 start_write_offset += srclen;
120 if (start_write_offset >= ring_buffer_size)
121 start_write_offset -= ring_buffer_size;
122
123 return start_write_offset;
124}
125
126/*
127 *
128 * hv_get_ringbuffer_availbytes()
129 *
130 * Get number of bytes available to read and to write to
131 * for the specified ring buffer
132 */
133static void
134hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
135 u32 *read, u32 *write)
136{
137 u32 read_loc, write_loc, dsize;
138
139 /* Capture the read/write indices before they changed */
140 read_loc = READ_ONCE(rbi->ring_buffer->read_index);
141 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
142 dsize = rbi->ring_datasize;
143
144 *write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
145 read_loc - write_loc;
146 *read = dsize - *write;
147}
148
149/* Get various debug metrics for the specified ring buffer. */
150int hv_ringbuffer_get_debuginfo(struct hv_ring_buffer_info *ring_info,
151 struct hv_ring_buffer_debug_info *debug_info)
152{
153 u32 bytes_avail_towrite;
154 u32 bytes_avail_toread;
155
156 mutex_lock(&ring_info->ring_buffer_mutex);
157
158 if (!ring_info->ring_buffer) {
159 mutex_unlock(&ring_info->ring_buffer_mutex);
160 return -EINVAL;
161 }
162
163 hv_get_ringbuffer_availbytes(ring_info,
164 &bytes_avail_toread,
165 &bytes_avail_towrite);
166 debug_info->bytes_avail_toread = bytes_avail_toread;
167 debug_info->bytes_avail_towrite = bytes_avail_towrite;
168 debug_info->current_read_index = ring_info->ring_buffer->read_index;
169 debug_info->current_write_index = ring_info->ring_buffer->write_index;
170 debug_info->current_interrupt_mask
171 = ring_info->ring_buffer->interrupt_mask;
172 mutex_unlock(&ring_info->ring_buffer_mutex);
173
174 return 0;
175}
176EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
177
178/* Initialize a channel's ring buffer info mutex locks */
179void hv_ringbuffer_pre_init(struct vmbus_channel *channel)
180{
181 mutex_init(&channel->inbound.ring_buffer_mutex);
182 mutex_init(&channel->outbound.ring_buffer_mutex);
183}
184
185/* Initialize the ring buffer. */
186int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
187 struct page *pages, u32 page_cnt, u32 max_pkt_size,
188 bool confidential)
189{
190 struct page **pages_wraparound;
191 int i;
192
193 BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
194
195 /*
196 * First page holds struct hv_ring_buffer, do wraparound mapping for
197 * the rest.
198 */
199 pages_wraparound = kzalloc_objs(struct page *, page_cnt * 2 - 1);
200 if (!pages_wraparound)
201 return -ENOMEM;
202
203 pages_wraparound[0] = pages;
204 for (i = 0; i < 2 * (page_cnt - 1); i++)
205 pages_wraparound[i + 1] =
206 &pages[i % (page_cnt - 1) + 1];
207
208 ring_info->ring_buffer = (struct hv_ring_buffer *)
209 vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP,
210 confidential ? PAGE_KERNEL : pgprot_decrypted(PAGE_KERNEL));
211
212 kfree(pages_wraparound);
213 if (!ring_info->ring_buffer)
214 return -ENOMEM;
215
216 /*
217 * Ensure the header page is zero'ed since
218 * encryption status may have changed.
219 */
220 memset(ring_info->ring_buffer, 0, HV_HYP_PAGE_SIZE);
221
222 ring_info->ring_buffer->read_index =
223 ring_info->ring_buffer->write_index = 0;
224
225 /* Set the feature bit for enabling flow control. */
226 ring_info->ring_buffer->feature_bits.value = 1;
227
228 ring_info->ring_size = page_cnt << PAGE_SHIFT;
229 ring_info->ring_size_div10_reciprocal =
230 reciprocal_value(ring_info->ring_size / 10);
231 ring_info->ring_datasize = ring_info->ring_size -
232 sizeof(struct hv_ring_buffer);
233 ring_info->priv_read_index = 0;
234
235 /* Initialize buffer that holds copies of incoming packets */
236 if (max_pkt_size) {
237 ring_info->pkt_buffer = kzalloc(max_pkt_size, GFP_KERNEL);
238 if (!ring_info->pkt_buffer)
239 return -ENOMEM;
240 ring_info->pkt_buffer_size = max_pkt_size;
241 }
242
243 spin_lock_init(&ring_info->ring_lock);
244
245 return 0;
246}
247
248/* Cleanup the ring buffer. */
249void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
250{
251 mutex_lock(&ring_info->ring_buffer_mutex);
252 vunmap(ring_info->ring_buffer);
253 ring_info->ring_buffer = NULL;
254 mutex_unlock(&ring_info->ring_buffer_mutex);
255
256 kfree(ring_info->pkt_buffer);
257 ring_info->pkt_buffer = NULL;
258 ring_info->pkt_buffer_size = 0;
259}
260
261/*
262 * Check if the ring buffer spinlock is available to take or not; used on
263 * atomic contexts, like panic path (see the Hyper-V framebuffer driver).
264 */
265
266bool hv_ringbuffer_spinlock_busy(struct vmbus_channel *channel)
267{
268 struct hv_ring_buffer_info *rinfo = &channel->outbound;
269
270 return spin_is_locked(&rinfo->ring_lock);
271}
272EXPORT_SYMBOL_GPL(hv_ringbuffer_spinlock_busy);
273
274/* Write to the ring buffer. */
275int hv_ringbuffer_write(struct vmbus_channel *channel,
276 const struct kvec *kv_list, u32 kv_count,
277 u64 requestid, u64 *trans_id)
278{
279 int i;
280 u32 bytes_avail_towrite;
281 u32 totalbytes_towrite = sizeof(u64);
282 u32 next_write_location;
283 u32 old_write;
284 u64 prev_indices;
285 unsigned long flags;
286 struct hv_ring_buffer_info *outring_info = &channel->outbound;
287 struct vmpacket_descriptor *desc = kv_list[0].iov_base;
288 u64 __trans_id, rqst_id = VMBUS_NO_RQSTOR;
289
290 if (channel->rescind)
291 return -ENODEV;
292
293 for (i = 0; i < kv_count; i++)
294 totalbytes_towrite += kv_list[i].iov_len;
295
296 spin_lock_irqsave(&outring_info->ring_lock, flags);
297
298 bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
299
300 /*
301 * If there is only room for the packet, assume it is full.
302 * Otherwise, the next time around, we think the ring buffer
303 * is empty since the read index == write index.
304 */
305 if (bytes_avail_towrite <= totalbytes_towrite) {
306 ++channel->out_full_total;
307
308 if (!channel->out_full_flag) {
309 ++channel->out_full_first;
310 channel->out_full_flag = true;
311 }
312
313 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
314 return -EAGAIN;
315 }
316
317 channel->out_full_flag = false;
318
319 /* Write to the ring buffer */
320 next_write_location = hv_get_next_write_location(outring_info);
321
322 old_write = next_write_location;
323
324 for (i = 0; i < kv_count; i++) {
325 next_write_location = hv_copyto_ringbuffer(outring_info,
326 next_write_location,
327 kv_list[i].iov_base,
328 kv_list[i].iov_len);
329 }
330
331 /*
332 * Allocate the request ID after the data has been copied into the
333 * ring buffer. Once this request ID is allocated, the completion
334 * path could find the data and free it.
335 */
336
337 if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {
338 if (channel->next_request_id_callback != NULL) {
339 rqst_id = channel->next_request_id_callback(channel, requestid);
340 if (rqst_id == VMBUS_RQST_ERROR) {
341 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
342 return -EAGAIN;
343 }
344 }
345 }
346 desc = hv_get_ring_buffer(outring_info) + old_write;
347 __trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;
348 /*
349 * Ensure the compiler doesn't generate code that reads the value of
350 * the transaction ID from the ring buffer, which is shared with the
351 * Hyper-V host and subject to being changed at any time.
352 */
353 WRITE_ONCE(desc->trans_id, __trans_id);
354 if (trans_id)
355 *trans_id = __trans_id;
356
357 /* Set previous packet start */
358 prev_indices = hv_get_ring_bufferindices(outring_info);
359
360 next_write_location = hv_copyto_ringbuffer(outring_info,
361 next_write_location,
362 &prev_indices,
363 sizeof(u64));
364
365 /* Issue a full memory barrier before updating the write index */
366 virt_mb();
367
368 /* Now, update the write location */
369 hv_set_next_write_location(outring_info, next_write_location);
370
371
372 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
373
374 hv_signal_on_write(old_write, channel);
375
376 if (channel->rescind) {
377 if (rqst_id != VMBUS_NO_RQSTOR) {
378 /* Reclaim request ID to avoid leak of IDs */
379 if (channel->request_addr_callback != NULL)
380 channel->request_addr_callback(channel, rqst_id);
381 }
382 return -ENODEV;
383 }
384
385 return 0;
386}
387
388int hv_ringbuffer_read(struct vmbus_channel *channel,
389 void *buffer, u32 buflen, u32 *buffer_actual_len,
390 u64 *requestid, bool raw)
391{
392 struct vmpacket_descriptor *desc;
393 u32 packetlen, offset;
394
395 if (unlikely(buflen == 0))
396 return -EINVAL;
397
398 *buffer_actual_len = 0;
399 *requestid = 0;
400
401 /* Make sure there is something to read */
402 desc = hv_pkt_iter_first(channel);
403 if (desc == NULL) {
404 /*
405 * No error is set when there is even no header, drivers are
406 * supposed to analyze buffer_actual_len.
407 */
408 return 0;
409 }
410
411 offset = raw ? 0 : (desc->offset8 << 3);
412 packetlen = (desc->len8 << 3) - offset;
413 *buffer_actual_len = packetlen;
414 *requestid = desc->trans_id;
415
416 if (unlikely(packetlen > buflen))
417 return -ENOBUFS;
418
419 /* since ring is double mapped, only one copy is necessary */
420 memcpy(buffer, (const char *)desc + offset, packetlen);
421
422 /* Advance ring index to next packet descriptor */
423 __hv_pkt_iter_next(channel, desc);
424
425 /* Notify host of update */
426 hv_pkt_iter_close(channel);
427
428 return 0;
429}
430
431/*
432 * Determine number of bytes available in ring buffer after
433 * the current iterator (priv_read_index) location.
434 *
435 * This is similar to hv_get_bytes_to_read but with private
436 * read index instead.
437 */
438static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
439{
440 u32 priv_read_loc = rbi->priv_read_index;
441 u32 write_loc;
442
443 /*
444 * The Hyper-V host writes the packet data, then uses
445 * store_release() to update the write_index. Use load_acquire()
446 * here to prevent loads of the packet data from being re-ordered
447 * before the read of the write_index and potentially getting
448 * stale data.
449 */
450 write_loc = virt_load_acquire(&rbi->ring_buffer->write_index);
451
452 if (write_loc >= priv_read_loc)
453 return write_loc - priv_read_loc;
454 else
455 return (rbi->ring_datasize - priv_read_loc) + write_loc;
456}
457
458/*
459 * Get first vmbus packet from ring buffer after read_index
460 *
461 * If ring buffer is empty, returns NULL and no other action needed.
462 */
463struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
464{
465 struct hv_ring_buffer_info *rbi = &channel->inbound;
466 struct vmpacket_descriptor *desc, *desc_copy;
467 u32 bytes_avail, pkt_len, pkt_offset;
468
469 hv_debug_delay_test(channel, MESSAGE_DELAY);
470
471 bytes_avail = hv_pkt_iter_avail(rbi);
472 if (bytes_avail < sizeof(struct vmpacket_descriptor))
473 return NULL;
474 bytes_avail = min(rbi->pkt_buffer_size, bytes_avail);
475
476 desc = (struct vmpacket_descriptor *)(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
477
478 /*
479 * Ensure the compiler does not use references to incoming Hyper-V values (which
480 * could change at any moment) when reading local variables later in the code
481 */
482 pkt_len = READ_ONCE(desc->len8) << 3;
483 pkt_offset = READ_ONCE(desc->offset8) << 3;
484
485 /*
486 * If pkt_len is invalid, set it to the smaller of hv_pkt_iter_avail() and
487 * rbi->pkt_buffer_size
488 */
489 if (pkt_len < sizeof(struct vmpacket_descriptor) || pkt_len > bytes_avail)
490 pkt_len = bytes_avail;
491
492 /*
493 * If pkt_offset is invalid, arbitrarily set it to
494 * the size of vmpacket_descriptor
495 */
496 if (pkt_offset < sizeof(struct vmpacket_descriptor) || pkt_offset > pkt_len)
497 pkt_offset = sizeof(struct vmpacket_descriptor);
498
499 /* Copy the Hyper-V packet out of the ring buffer */
500 desc_copy = (struct vmpacket_descriptor *)rbi->pkt_buffer;
501 memcpy(desc_copy, desc, pkt_len);
502
503 /*
504 * Hyper-V could still change len8 and offset8 after the earlier read.
505 * Ensure that desc_copy has legal values for len8 and offset8 that
506 * are consistent with the copy we just made
507 */
508 desc_copy->len8 = pkt_len >> 3;
509 desc_copy->offset8 = pkt_offset >> 3;
510
511 return desc_copy;
512}
513EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
514
515/*
516 * Get next vmbus packet from ring buffer.
517 *
518 * Advances the current location (priv_read_index) and checks for more
519 * data. If the end of the ring buffer is reached, then return NULL.
520 */
521struct vmpacket_descriptor *
522__hv_pkt_iter_next(struct vmbus_channel *channel,
523 const struct vmpacket_descriptor *desc)
524{
525 struct hv_ring_buffer_info *rbi = &channel->inbound;
526 u32 packetlen = desc->len8 << 3;
527 u32 dsize = rbi->ring_datasize;
528
529 hv_debug_delay_test(channel, MESSAGE_DELAY);
530 /* bump offset to next potential packet */
531 rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
532 if (rbi->priv_read_index >= dsize)
533 rbi->priv_read_index -= dsize;
534
535 /* more data? */
536 return hv_pkt_iter_first(channel);
537}
538EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
539
540/* How many bytes were read in this iterator cycle */
541static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
542 u32 start_read_index)
543{
544 if (rbi->priv_read_index >= start_read_index)
545 return rbi->priv_read_index - start_read_index;
546 else
547 return rbi->ring_datasize - start_read_index +
548 rbi->priv_read_index;
549}
550
551/*
552 * Update host ring buffer after iterating over packets. If the host has
553 * stopped queuing new entries because it found the ring buffer full, and
554 * sufficient space is being freed up, signal the host. But be careful to
555 * only signal the host when necessary, both for performance reasons and
556 * because Hyper-V protects itself by throttling guests that signal
557 * inappropriately.
558 *
559 * Determining when to signal is tricky. There are three key data inputs
560 * that must be handled in this order to avoid race conditions:
561 *
562 * 1. Update the read_index
563 * 2. Read the pending_send_sz
564 * 3. Read the current write_index
565 *
566 * The interrupt_mask is not used to determine when to signal. The
567 * interrupt_mask is used only on the guest->host ring buffer when
568 * sending requests to the host. The host does not use it on the host->
569 * guest ring buffer to indicate whether it should be signaled.
570 */
571void hv_pkt_iter_close(struct vmbus_channel *channel)
572{
573 struct hv_ring_buffer_info *rbi = &channel->inbound;
574 u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
575
576 /*
577 * Make sure all reads are done before we update the read index since
578 * the writer may start writing to the read area once the read index
579 * is updated.
580 */
581 virt_rmb();
582 start_read_index = rbi->ring_buffer->read_index;
583 rbi->ring_buffer->read_index = rbi->priv_read_index;
584
585 /*
586 * Older versions of Hyper-V (before WS2102 and Win8) do not
587 * implement pending_send_sz and simply poll if the host->guest
588 * ring buffer is full. No signaling is needed or expected.
589 */
590 if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
591 return;
592
593 /*
594 * Issue a full memory barrier before making the signaling decision.
595 * If reading pending_send_sz were to be reordered and happen
596 * before we commit the new read_index, a race could occur. If the
597 * host were to set the pending_send_sz after we have sampled
598 * pending_send_sz, and the ring buffer blocks before we commit the
599 * read index, we could miss sending the interrupt. Issue a full
600 * memory barrier to address this.
601 */
602 virt_mb();
603
604 /*
605 * If the pending_send_sz is zero, then the ring buffer is not
606 * blocked and there is no need to signal. This is far by the
607 * most common case, so exit quickly for best performance.
608 */
609 pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
610 if (!pending_sz)
611 return;
612
613 /*
614 * Ensure the read of write_index in hv_get_bytes_to_write()
615 * happens after the read of pending_send_sz.
616 */
617 virt_rmb();
618 curr_write_sz = hv_get_bytes_to_write(rbi);
619 bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
620
621 /*
622 * We want to signal the host only if we're transitioning
623 * from a "not enough free space" state to a "enough free
624 * space" state. For example, it's possible that this function
625 * could run and free up enough space to signal the host, and then
626 * run again and free up additional space before the host has a
627 * chance to clear the pending_send_sz. The 2nd invocation would
628 * be a null transition from "enough free space" to "enough free
629 * space", which doesn't warrant a signal.
630 *
631 * Exactly filling the ring buffer is treated as "not enough
632 * space". The ring buffer always must have at least one byte
633 * empty so the empty and full conditions are distinguishable.
634 * hv_get_bytes_to_write() doesn't fully tell the truth in
635 * this regard.
636 *
637 * So first check if we were in the "enough free space" state
638 * before we began the iteration. If so, the host was not
639 * blocked, and there's no need to signal.
640 */
641 if (curr_write_sz - bytes_read > pending_sz)
642 return;
643
644 /*
645 * Similarly, if the new state is "not enough space", then
646 * there's no need to signal.
647 */
648 if (curr_write_sz <= pending_sz)
649 return;
650
651 ++channel->intr_in_full;
652 vmbus_setevent(channel);
653}
654EXPORT_SYMBOL_GPL(hv_pkt_iter_close);