xref: /linux/drivers/hv/ring_buffer.c (revision c79c3c34f75d72a066e292b10aa50fc758c97c89)
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 
50 static 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. */
71 static inline u32
72 hv_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. */
80 static inline void
81 hv_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. */
88 static inline void
89 hv_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. */
97 static inline u32
98 hv_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. */
104 static inline u64
105 hv_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  */
114 static 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  */
139 static void
140 hv_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. */
156 int 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 }
182 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
183 
184 /* Initialize a channel's ring buffer info mutex locks */
185 void 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. */
192 int 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. */
241 void 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. */
250 int hv_ringbuffer_write(struct vmbus_channel *channel,
251 			const struct kvec *kv_list, u32 kv_count,
252 			u64 requestid)
253 {
254 	int i;
255 	u32 bytes_avail_towrite;
256 	u32 totalbytes_towrite = sizeof(u64);
257 	u32 next_write_location;
258 	u32 old_write;
259 	u64 prev_indices;
260 	unsigned long flags;
261 	struct hv_ring_buffer_info *outring_info = &channel->outbound;
262 	struct vmpacket_descriptor *desc = kv_list[0].iov_base;
263 	u64 rqst_id = VMBUS_NO_RQSTOR;
264 
265 	if (channel->rescind)
266 		return -ENODEV;
267 
268 	for (i = 0; i < kv_count; i++)
269 		totalbytes_towrite += kv_list[i].iov_len;
270 
271 	spin_lock_irqsave(&outring_info->ring_lock, flags);
272 
273 	bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
274 
275 	/*
276 	 * If there is only room for the packet, assume it is full.
277 	 * Otherwise, the next time around, we think the ring buffer
278 	 * is empty since the read index == write index.
279 	 */
280 	if (bytes_avail_towrite <= totalbytes_towrite) {
281 		++channel->out_full_total;
282 
283 		if (!channel->out_full_flag) {
284 			++channel->out_full_first;
285 			channel->out_full_flag = true;
286 		}
287 
288 		spin_unlock_irqrestore(&outring_info->ring_lock, flags);
289 		return -EAGAIN;
290 	}
291 
292 	channel->out_full_flag = false;
293 
294 	/* Write to the ring buffer */
295 	next_write_location = hv_get_next_write_location(outring_info);
296 
297 	old_write = next_write_location;
298 
299 	for (i = 0; i < kv_count; i++) {
300 		next_write_location = hv_copyto_ringbuffer(outring_info,
301 						     next_write_location,
302 						     kv_list[i].iov_base,
303 						     kv_list[i].iov_len);
304 	}
305 
306 	/*
307 	 * Allocate the request ID after the data has been copied into the
308 	 * ring buffer.  Once this request ID is allocated, the completion
309 	 * path could find the data and free it.
310 	 */
311 
312 	if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {
313 		rqst_id = vmbus_next_request_id(&channel->requestor, requestid);
314 		if (rqst_id == VMBUS_RQST_ERROR) {
315 			spin_unlock_irqrestore(&outring_info->ring_lock, flags);
316 			pr_err("No request id available\n");
317 			return -EAGAIN;
318 		}
319 	}
320 	desc = hv_get_ring_buffer(outring_info) + old_write;
321 	desc->trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;
322 
323 	/* Set previous packet start */
324 	prev_indices = hv_get_ring_bufferindices(outring_info);
325 
326 	next_write_location = hv_copyto_ringbuffer(outring_info,
327 					     next_write_location,
328 					     &prev_indices,
329 					     sizeof(u64));
330 
331 	/* Issue a full memory barrier before updating the write index */
332 	virt_mb();
333 
334 	/* Now, update the write location */
335 	hv_set_next_write_location(outring_info, next_write_location);
336 
337 
338 	spin_unlock_irqrestore(&outring_info->ring_lock, flags);
339 
340 	hv_signal_on_write(old_write, channel);
341 
342 	if (channel->rescind) {
343 		if (rqst_id != VMBUS_NO_RQSTOR) {
344 			/* Reclaim request ID to avoid leak of IDs */
345 			vmbus_request_addr(&channel->requestor, rqst_id);
346 		}
347 		return -ENODEV;
348 	}
349 
350 	return 0;
351 }
352 
353 int hv_ringbuffer_read(struct vmbus_channel *channel,
354 		       void *buffer, u32 buflen, u32 *buffer_actual_len,
355 		       u64 *requestid, bool raw)
356 {
357 	struct vmpacket_descriptor *desc;
358 	u32 packetlen, offset;
359 
360 	if (unlikely(buflen == 0))
361 		return -EINVAL;
362 
363 	*buffer_actual_len = 0;
364 	*requestid = 0;
365 
366 	/* Make sure there is something to read */
367 	desc = hv_pkt_iter_first(channel);
368 	if (desc == NULL) {
369 		/*
370 		 * No error is set when there is even no header, drivers are
371 		 * supposed to analyze buffer_actual_len.
372 		 */
373 		return 0;
374 	}
375 
376 	offset = raw ? 0 : (desc->offset8 << 3);
377 	packetlen = (desc->len8 << 3) - offset;
378 	*buffer_actual_len = packetlen;
379 	*requestid = desc->trans_id;
380 
381 	if (unlikely(packetlen > buflen))
382 		return -ENOBUFS;
383 
384 	/* since ring is double mapped, only one copy is necessary */
385 	memcpy(buffer, (const char *)desc + offset, packetlen);
386 
387 	/* Advance ring index to next packet descriptor */
388 	__hv_pkt_iter_next(channel, desc);
389 
390 	/* Notify host of update */
391 	hv_pkt_iter_close(channel);
392 
393 	return 0;
394 }
395 
396 /*
397  * Determine number of bytes available in ring buffer after
398  * the current iterator (priv_read_index) location.
399  *
400  * This is similar to hv_get_bytes_to_read but with private
401  * read index instead.
402  */
403 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
404 {
405 	u32 priv_read_loc = rbi->priv_read_index;
406 	u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
407 
408 	if (write_loc >= priv_read_loc)
409 		return write_loc - priv_read_loc;
410 	else
411 		return (rbi->ring_datasize - priv_read_loc) + write_loc;
412 }
413 
414 /*
415  * Get first vmbus packet from ring buffer after read_index
416  *
417  * If ring buffer is empty, returns NULL and no other action needed.
418  */
419 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
420 {
421 	struct hv_ring_buffer_info *rbi = &channel->inbound;
422 	struct vmpacket_descriptor *desc;
423 
424 	hv_debug_delay_test(channel, MESSAGE_DELAY);
425 	if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
426 		return NULL;
427 
428 	desc = hv_get_ring_buffer(rbi) + rbi->priv_read_index;
429 	if (desc)
430 		prefetch((char *)desc + (desc->len8 << 3));
431 
432 	return desc;
433 }
434 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
435 
436 /*
437  * Get next vmbus packet from ring buffer.
438  *
439  * Advances the current location (priv_read_index) and checks for more
440  * data. If the end of the ring buffer is reached, then return NULL.
441  */
442 struct vmpacket_descriptor *
443 __hv_pkt_iter_next(struct vmbus_channel *channel,
444 		   const struct vmpacket_descriptor *desc)
445 {
446 	struct hv_ring_buffer_info *rbi = &channel->inbound;
447 	u32 packetlen = desc->len8 << 3;
448 	u32 dsize = rbi->ring_datasize;
449 
450 	hv_debug_delay_test(channel, MESSAGE_DELAY);
451 	/* bump offset to next potential packet */
452 	rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
453 	if (rbi->priv_read_index >= dsize)
454 		rbi->priv_read_index -= dsize;
455 
456 	/* more data? */
457 	return hv_pkt_iter_first(channel);
458 }
459 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
460 
461 /* How many bytes were read in this iterator cycle */
462 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
463 					u32 start_read_index)
464 {
465 	if (rbi->priv_read_index >= start_read_index)
466 		return rbi->priv_read_index - start_read_index;
467 	else
468 		return rbi->ring_datasize - start_read_index +
469 			rbi->priv_read_index;
470 }
471 
472 /*
473  * Update host ring buffer after iterating over packets. If the host has
474  * stopped queuing new entries because it found the ring buffer full, and
475  * sufficient space is being freed up, signal the host. But be careful to
476  * only signal the host when necessary, both for performance reasons and
477  * because Hyper-V protects itself by throttling guests that signal
478  * inappropriately.
479  *
480  * Determining when to signal is tricky. There are three key data inputs
481  * that must be handled in this order to avoid race conditions:
482  *
483  * 1. Update the read_index
484  * 2. Read the pending_send_sz
485  * 3. Read the current write_index
486  *
487  * The interrupt_mask is not used to determine when to signal. The
488  * interrupt_mask is used only on the guest->host ring buffer when
489  * sending requests to the host. The host does not use it on the host->
490  * guest ring buffer to indicate whether it should be signaled.
491  */
492 void hv_pkt_iter_close(struct vmbus_channel *channel)
493 {
494 	struct hv_ring_buffer_info *rbi = &channel->inbound;
495 	u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
496 
497 	/*
498 	 * Make sure all reads are done before we update the read index since
499 	 * the writer may start writing to the read area once the read index
500 	 * is updated.
501 	 */
502 	virt_rmb();
503 	start_read_index = rbi->ring_buffer->read_index;
504 	rbi->ring_buffer->read_index = rbi->priv_read_index;
505 
506 	/*
507 	 * Older versions of Hyper-V (before WS2102 and Win8) do not
508 	 * implement pending_send_sz and simply poll if the host->guest
509 	 * ring buffer is full.  No signaling is needed or expected.
510 	 */
511 	if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
512 		return;
513 
514 	/*
515 	 * Issue a full memory barrier before making the signaling decision.
516 	 * If reading pending_send_sz were to be reordered and happen
517 	 * before we commit the new read_index, a race could occur.  If the
518 	 * host were to set the pending_send_sz after we have sampled
519 	 * pending_send_sz, and the ring buffer blocks before we commit the
520 	 * read index, we could miss sending the interrupt. Issue a full
521 	 * memory barrier to address this.
522 	 */
523 	virt_mb();
524 
525 	/*
526 	 * If the pending_send_sz is zero, then the ring buffer is not
527 	 * blocked and there is no need to signal.  This is far by the
528 	 * most common case, so exit quickly for best performance.
529 	 */
530 	pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
531 	if (!pending_sz)
532 		return;
533 
534 	/*
535 	 * Ensure the read of write_index in hv_get_bytes_to_write()
536 	 * happens after the read of pending_send_sz.
537 	 */
538 	virt_rmb();
539 	curr_write_sz = hv_get_bytes_to_write(rbi);
540 	bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
541 
542 	/*
543 	 * We want to signal the host only if we're transitioning
544 	 * from a "not enough free space" state to a "enough free
545 	 * space" state.  For example, it's possible that this function
546 	 * could run and free up enough space to signal the host, and then
547 	 * run again and free up additional space before the host has a
548 	 * chance to clear the pending_send_sz.  The 2nd invocation would
549 	 * be a null transition from "enough free space" to "enough free
550 	 * space", which doesn't warrant a signal.
551 	 *
552 	 * Exactly filling the ring buffer is treated as "not enough
553 	 * space". The ring buffer always must have at least one byte
554 	 * empty so the empty and full conditions are distinguishable.
555 	 * hv_get_bytes_to_write() doesn't fully tell the truth in
556 	 * this regard.
557 	 *
558 	 * So first check if we were in the "enough free space" state
559 	 * before we began the iteration. If so, the host was not
560 	 * blocked, and there's no need to signal.
561 	 */
562 	if (curr_write_sz - bytes_read > pending_sz)
563 		return;
564 
565 	/*
566 	 * Similarly, if the new state is "not enough space", then
567 	 * there's no need to signal.
568 	 */
569 	if (curr_write_sz <= pending_sz)
570 		return;
571 
572 	++channel->intr_in_full;
573 	vmbus_setevent(channel);
574 }
575 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);
576