ring_buffer.c

  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#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 = kcalloc(page_cnt * 2 - 1,
200				   sizeof(struct page *),
201				   GFP_KERNEL);
202	if (!pages_wraparound)
203		return -ENOMEM;
204
205	pages_wraparound[0] = pages;
206	for (i = 0; i < 2 * (page_cnt - 1); i++)
207		pages_wraparound[i + 1] =
208			&pages[i % (page_cnt - 1) + 1];
209
210	ring_info->ring_buffer = (struct hv_ring_buffer *)
211		vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP,
212			confidential ? PAGE_KERNEL : pgprot_decrypted(PAGE_KERNEL));
213
214	kfree(pages_wraparound);
215	if (!ring_info->ring_buffer)
216		return -ENOMEM;
217
218	/*
219	 * Ensure the header page is zero'ed since
220	 * encryption status may have changed.
221	 */
222	memset(ring_info->ring_buffer, 0, HV_HYP_PAGE_SIZE);
223
224	ring_info->ring_buffer->read_index =
225		ring_info->ring_buffer->write_index = 0;
226
227	/* Set the feature bit for enabling flow control. */
228	ring_info->ring_buffer->feature_bits.value = 1;
229
230	ring_info->ring_size = page_cnt << PAGE_SHIFT;
231	ring_info->ring_size_div10_reciprocal =
232		reciprocal_value(ring_info->ring_size / 10);
233	ring_info->ring_datasize = ring_info->ring_size -
234		sizeof(struct hv_ring_buffer);
235	ring_info->priv_read_index = 0;
236
237	/* Initialize buffer that holds copies of incoming packets */
238	if (max_pkt_size) {
239		ring_info->pkt_buffer = kzalloc(max_pkt_size, GFP_KERNEL);
240		if (!ring_info->pkt_buffer)
241			return -ENOMEM;
242		ring_info->pkt_buffer_size = max_pkt_size;
243	}
244
245	spin_lock_init(&ring_info->ring_lock);
246
247	return 0;
248}
249
250/* Cleanup the ring buffer. */
251void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
252{
253	mutex_lock(&ring_info->ring_buffer_mutex);
254	vunmap(ring_info->ring_buffer);
255	ring_info->ring_buffer = NULL;
256	mutex_unlock(&ring_info->ring_buffer_mutex);
257
258	kfree(ring_info->pkt_buffer);
259	ring_info->pkt_buffer = NULL;
260	ring_info->pkt_buffer_size = 0;
261}
262
263/*
264 * Check if the ring buffer spinlock is available to take or not; used on
265 * atomic contexts, like panic path (see the Hyper-V framebuffer driver).
266 */
267
268bool hv_ringbuffer_spinlock_busy(struct vmbus_channel *channel)
269{
270	struct hv_ring_buffer_info *rinfo = &channel->outbound;
271
272	return spin_is_locked(&rinfo->ring_lock);
273}
274EXPORT_SYMBOL_GPL(hv_ringbuffer_spinlock_busy);
275
276/* Write to the ring buffer. */
277int hv_ringbuffer_write(struct vmbus_channel *channel,
278			const struct kvec *kv_list, u32 kv_count,
279			u64 requestid, u64 *trans_id)
280{
281	int i;
282	u32 bytes_avail_towrite;
283	u32 totalbytes_towrite = sizeof(u64);
284	u32 next_write_location;
285	u32 old_write;
286	u64 prev_indices;
287	unsigned long flags;
288	struct hv_ring_buffer_info *outring_info = &channel->outbound;
289	struct vmpacket_descriptor *desc = kv_list[0].iov_base;
290	u64 __trans_id, rqst_id = VMBUS_NO_RQSTOR;
291
292	if (channel->rescind)
293		return -ENODEV;
294
295	for (i = 0; i < kv_count; i++)
296		totalbytes_towrite += kv_list[i].iov_len;
297
298	spin_lock_irqsave(&outring_info->ring_lock, flags);
299
300	bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
301
302	/*
303	 * If there is only room for the packet, assume it is full.
304	 * Otherwise, the next time around, we think the ring buffer
305	 * is empty since the read index == write index.
306	 */
307	if (bytes_avail_towrite <= totalbytes_towrite) {
308		++channel->out_full_total;
309
310		if (!channel->out_full_flag) {
311			++channel->out_full_first;
312			channel->out_full_flag = true;
313		}
314
315		spin_unlock_irqrestore(&outring_info->ring_lock, flags);
316		return -EAGAIN;
317	}
318
319	channel->out_full_flag = false;
320
321	/* Write to the ring buffer */
322	next_write_location = hv_get_next_write_location(outring_info);
323
324	old_write = next_write_location;
325
326	for (i = 0; i < kv_count; i++) {
327		next_write_location = hv_copyto_ringbuffer(outring_info,
328						     next_write_location,
329						     kv_list[i].iov_base,
330						     kv_list[i].iov_len);
331	}
332
333	/*
334	 * Allocate the request ID after the data has been copied into the
335	 * ring buffer.  Once this request ID is allocated, the completion
336	 * path could find the data and free it.
337	 */
338
339	if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {
340		if (channel->next_request_id_callback != NULL) {
341			rqst_id = channel->next_request_id_callback(channel, requestid);
342			if (rqst_id == VMBUS_RQST_ERROR) {
343				spin_unlock_irqrestore(&outring_info->ring_lock, flags);
344				return -EAGAIN;
345			}
346		}
347	}
348	desc = hv_get_ring_buffer(outring_info) + old_write;
349	__trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;
350	/*
351	 * Ensure the compiler doesn't generate code that reads the value of
352	 * the transaction ID from the ring buffer, which is shared with the
353	 * Hyper-V host and subject to being changed at any time.
354	 */
355	WRITE_ONCE(desc->trans_id, __trans_id);
356	if (trans_id)
357		*trans_id = __trans_id;
358
359	/* Set previous packet start */
360	prev_indices = hv_get_ring_bufferindices(outring_info);
361
362	next_write_location = hv_copyto_ringbuffer(outring_info,
363					     next_write_location,
364					     &prev_indices,
365					     sizeof(u64));
366
367	/* Issue a full memory barrier before updating the write index */
368	virt_mb();
369
370	/* Now, update the write location */
371	hv_set_next_write_location(outring_info, next_write_location);
372
373
374	spin_unlock_irqrestore(&outring_info->ring_lock, flags);
375
376	hv_signal_on_write(old_write, channel);
377
378	if (channel->rescind) {
379		if (rqst_id != VMBUS_NO_RQSTOR) {
380			/* Reclaim request ID to avoid leak of IDs */
381			if (channel->request_addr_callback != NULL)
382				channel->request_addr_callback(channel, rqst_id);
383		}
384		return -ENODEV;
385	}
386
387	return 0;
388}
389
390int hv_ringbuffer_read(struct vmbus_channel *channel,
391		       void *buffer, u32 buflen, u32 *buffer_actual_len,
392		       u64 *requestid, bool raw)
393{
394	struct vmpacket_descriptor *desc;
395	u32 packetlen, offset;
396
397	if (unlikely(buflen == 0))
398		return -EINVAL;
399
400	*buffer_actual_len = 0;
401	*requestid = 0;
402
403	/* Make sure there is something to read */
404	desc = hv_pkt_iter_first(channel);
405	if (desc == NULL) {
406		/*
407		 * No error is set when there is even no header, drivers are
408		 * supposed to analyze buffer_actual_len.
409		 */
410		return 0;
411	}
412
413	offset = raw ? 0 : (desc->offset8 << 3);
414	packetlen = (desc->len8 << 3) - offset;
415	*buffer_actual_len = packetlen;
416	*requestid = desc->trans_id;
417
418	if (unlikely(packetlen > buflen))
419		return -ENOBUFS;
420
421	/* since ring is double mapped, only one copy is necessary */
422	memcpy(buffer, (const char *)desc + offset, packetlen);
423
424	/* Advance ring index to next packet descriptor */
425	__hv_pkt_iter_next(channel, desc);
426
427	/* Notify host of update */
428	hv_pkt_iter_close(channel);
429
430	return 0;
431}
432
433/*
434 * Determine number of bytes available in ring buffer after
435 * the current iterator (priv_read_index) location.
436 *
437 * This is similar to hv_get_bytes_to_read but with private
438 * read index instead.
439 */
440static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
441{
442	u32 priv_read_loc = rbi->priv_read_index;
443	u32 write_loc;
444
445	/*
446	 * The Hyper-V host writes the packet data, then uses
447	 * store_release() to update the write_index.  Use load_acquire()
448	 * here to prevent loads of the packet data from being re-ordered
449	 * before the read of the write_index and potentially getting
450	 * stale data.
451	 */
452	write_loc = virt_load_acquire(&rbi->ring_buffer->write_index);
453
454	if (write_loc >= priv_read_loc)
455		return write_loc - priv_read_loc;
456	else
457		return (rbi->ring_datasize - priv_read_loc) + write_loc;
458}
459
460/*
461 * Get first vmbus packet from ring buffer after read_index
462 *
463 * If ring buffer is empty, returns NULL and no other action needed.
464 */
465struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
466{
467	struct hv_ring_buffer_info *rbi = &channel->inbound;
468	struct vmpacket_descriptor *desc, *desc_copy;
469	u32 bytes_avail, pkt_len, pkt_offset;
470
471	hv_debug_delay_test(channel, MESSAGE_DELAY);
472
473	bytes_avail = hv_pkt_iter_avail(rbi);
474	if (bytes_avail < sizeof(struct vmpacket_descriptor))
475		return NULL;
476	bytes_avail = min(rbi->pkt_buffer_size, bytes_avail);
477
478	desc = (struct vmpacket_descriptor *)(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
479
480	/*
481	 * Ensure the compiler does not use references to incoming Hyper-V values (which
482	 * could change at any moment) when reading local variables later in the code
483	 */
484	pkt_len = READ_ONCE(desc->len8) << 3;
485	pkt_offset = READ_ONCE(desc->offset8) << 3;
486
487	/*
488	 * If pkt_len is invalid, set it to the smaller of hv_pkt_iter_avail() and
489	 * rbi->pkt_buffer_size
490	 */
491	if (pkt_len < sizeof(struct vmpacket_descriptor) || pkt_len > bytes_avail)
492		pkt_len = bytes_avail;
493
494	/*
495	 * If pkt_offset is invalid, arbitrarily set it to
496	 * the size of vmpacket_descriptor
497	 */
498	if (pkt_offset < sizeof(struct vmpacket_descriptor) || pkt_offset > pkt_len)
499		pkt_offset = sizeof(struct vmpacket_descriptor);
500
501	/* Copy the Hyper-V packet out of the ring buffer */
502	desc_copy = (struct vmpacket_descriptor *)rbi->pkt_buffer;
503	memcpy(desc_copy, desc, pkt_len);
504
505	/*
506	 * Hyper-V could still change len8 and offset8 after the earlier read.
507	 * Ensure that desc_copy has legal values for len8 and offset8 that
508	 * are consistent with the copy we just made
509	 */
510	desc_copy->len8 = pkt_len >> 3;
511	desc_copy->offset8 = pkt_offset >> 3;
512
513	return desc_copy;
514}
515EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
516
517/*
518 * Get next vmbus packet from ring buffer.
519 *
520 * Advances the current location (priv_read_index) and checks for more
521 * data. If the end of the ring buffer is reached, then return NULL.
522 */
523struct vmpacket_descriptor *
524__hv_pkt_iter_next(struct vmbus_channel *channel,
525		   const struct vmpacket_descriptor *desc)
526{
527	struct hv_ring_buffer_info *rbi = &channel->inbound;
528	u32 packetlen = desc->len8 << 3;
529	u32 dsize = rbi->ring_datasize;
530
531	hv_debug_delay_test(channel, MESSAGE_DELAY);
532	/* bump offset to next potential packet */
533	rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
534	if (rbi->priv_read_index >= dsize)
535		rbi->priv_read_index -= dsize;
536
537	/* more data? */
538	return hv_pkt_iter_first(channel);
539}
540EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
541
542/* How many bytes were read in this iterator cycle */
543static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
544					u32 start_read_index)
545{
546	if (rbi->priv_read_index >= start_read_index)
547		return rbi->priv_read_index - start_read_index;
548	else
549		return rbi->ring_datasize - start_read_index +
550			rbi->priv_read_index;
551}
552
553/*
554 * Update host ring buffer after iterating over packets. If the host has
555 * stopped queuing new entries because it found the ring buffer full, and
556 * sufficient space is being freed up, signal the host. But be careful to
557 * only signal the host when necessary, both for performance reasons and
558 * because Hyper-V protects itself by throttling guests that signal
559 * inappropriately.
560 *
561 * Determining when to signal is tricky. There are three key data inputs
562 * that must be handled in this order to avoid race conditions:
563 *
564 * 1. Update the read_index
565 * 2. Read the pending_send_sz
566 * 3. Read the current write_index
567 *
568 * The interrupt_mask is not used to determine when to signal. The
569 * interrupt_mask is used only on the guest->host ring buffer when
570 * sending requests to the host. The host does not use it on the host->
571 * guest ring buffer to indicate whether it should be signaled.
572 */
573void hv_pkt_iter_close(struct vmbus_channel *channel)
574{
575	struct hv_ring_buffer_info *rbi = &channel->inbound;
576	u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
577
578	/*
579	 * Make sure all reads are done before we update the read index since
580	 * the writer may start writing to the read area once the read index
581	 * is updated.
582	 */
583	virt_rmb();
584	start_read_index = rbi->ring_buffer->read_index;
585	rbi->ring_buffer->read_index = rbi->priv_read_index;
586
587	/*
588	 * Older versions of Hyper-V (before WS2102 and Win8) do not
589	 * implement pending_send_sz and simply poll if the host->guest
590	 * ring buffer is full.  No signaling is needed or expected.
591	 */
592	if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
593		return;
594
595	/*
596	 * Issue a full memory barrier before making the signaling decision.
597	 * If reading pending_send_sz were to be reordered and happen
598	 * before we commit the new read_index, a race could occur.  If the
599	 * host were to set the pending_send_sz after we have sampled
600	 * pending_send_sz, and the ring buffer blocks before we commit the
601	 * read index, we could miss sending the interrupt. Issue a full
602	 * memory barrier to address this.
603	 */
604	virt_mb();
605
606	/*
607	 * If the pending_send_sz is zero, then the ring buffer is not
608	 * blocked and there is no need to signal.  This is far by the
609	 * most common case, so exit quickly for best performance.
610	 */
611	pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
612	if (!pending_sz)
613		return;
614
615	/*
616	 * Ensure the read of write_index in hv_get_bytes_to_write()
617	 * happens after the read of pending_send_sz.
618	 */
619	virt_rmb();
620	curr_write_sz = hv_get_bytes_to_write(rbi);
621	bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
622
623	/*
624	 * We want to signal the host only if we're transitioning
625	 * from a "not enough free space" state to a "enough free
626	 * space" state.  For example, it's possible that this function
627	 * could run and free up enough space to signal the host, and then
628	 * run again and free up additional space before the host has a
629	 * chance to clear the pending_send_sz.  The 2nd invocation would
630	 * be a null transition from "enough free space" to "enough free
631	 * space", which doesn't warrant a signal.
632	 *
633	 * Exactly filling the ring buffer is treated as "not enough
634	 * space". The ring buffer always must have at least one byte
635	 * empty so the empty and full conditions are distinguishable.
636	 * hv_get_bytes_to_write() doesn't fully tell the truth in
637	 * this regard.
638	 *
639	 * So first check if we were in the "enough free space" state
640	 * before we began the iteration. If so, the host was not
641	 * blocked, and there's no need to signal.
642	 */
643	if (curr_write_sz - bytes_read > pending_sz)
644		return;
645
646	/*
647	 * Similarly, if the new state is "not enough space", then
648	 * there's no need to signal.
649	 */
650	if (curr_write_sz <= pending_sz)
651		return;
652
653	++channel->intr_in_full;
654	vmbus_setevent(channel);
655}
656EXPORT_SYMBOL_GPL(hv_pkt_iter_close);