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