xref: /openbmc/linux/drivers/hv/ring_buffer.c (revision 82e6fdd6)
1 /*
2  *
3  * Copyright (c) 2009, Microsoft Corporation.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  * You should have received a copy of the GNU General Public License along with
15  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
16  * Place - Suite 330, Boston, MA 02111-1307 USA.
17  *
18  * Authors:
19  *   Haiyang Zhang <haiyangz@microsoft.com>
20  *   Hank Janssen  <hjanssen@microsoft.com>
21  *   K. Y. Srinivasan <kys@microsoft.com>
22  *
23  */
24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
25 
26 #include <linux/kernel.h>
27 #include <linux/mm.h>
28 #include <linux/hyperv.h>
29 #include <linux/uio.h>
30 #include <linux/vmalloc.h>
31 #include <linux/slab.h>
32 #include <linux/prefetch.h>
33 
34 #include "hyperv_vmbus.h"
35 
36 #define VMBUS_PKT_TRAILER	8
37 
38 /*
39  * When we write to the ring buffer, check if the host needs to
40  * be signaled. Here is the details of this protocol:
41  *
42  *	1. The host guarantees that while it is draining the
43  *	   ring buffer, it will set the interrupt_mask to
44  *	   indicate it does not need to be interrupted when
45  *	   new data is placed.
46  *
47  *	2. The host guarantees that it will completely drain
48  *	   the ring buffer before exiting the read loop. Further,
49  *	   once the ring buffer is empty, it will clear the
50  *	   interrupt_mask and re-check to see if new data has
51  *	   arrived.
52  *
53  * KYS: Oct. 30, 2016:
54  * It looks like Windows hosts have logic to deal with DOS attacks that
55  * can be triggered if it receives interrupts when it is not expecting
56  * the interrupt. The host expects interrupts only when the ring
57  * transitions from empty to non-empty (or full to non full on the guest
58  * to host ring).
59  * So, base the signaling decision solely on the ring state until the
60  * host logic is fixed.
61  */
62 
63 static void hv_signal_on_write(u32 old_write, struct vmbus_channel *channel)
64 {
65 	struct hv_ring_buffer_info *rbi = &channel->outbound;
66 
67 	virt_mb();
68 	if (READ_ONCE(rbi->ring_buffer->interrupt_mask))
69 		return;
70 
71 	/* check interrupt_mask before read_index */
72 	virt_rmb();
73 	/*
74 	 * This is the only case we need to signal when the
75 	 * ring transitions from being empty to non-empty.
76 	 */
77 	if (old_write == READ_ONCE(rbi->ring_buffer->read_index))
78 		vmbus_setevent(channel);
79 }
80 
81 /* Get the next write location for the specified ring buffer. */
82 static inline u32
83 hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
84 {
85 	u32 next = ring_info->ring_buffer->write_index;
86 
87 	return next;
88 }
89 
90 /* Set the next write location for the specified ring buffer. */
91 static inline void
92 hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
93 		     u32 next_write_location)
94 {
95 	ring_info->ring_buffer->write_index = next_write_location;
96 }
97 
98 /* Set the next read location for the specified ring buffer. */
99 static inline void
100 hv_set_next_read_location(struct hv_ring_buffer_info *ring_info,
101 		    u32 next_read_location)
102 {
103 	ring_info->ring_buffer->read_index = next_read_location;
104 	ring_info->priv_read_index = next_read_location;
105 }
106 
107 /* Get the size of the ring buffer. */
108 static inline u32
109 hv_get_ring_buffersize(const struct hv_ring_buffer_info *ring_info)
110 {
111 	return ring_info->ring_datasize;
112 }
113 
114 /* Get the read and write indices as u64 of the specified ring buffer. */
115 static inline u64
116 hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
117 {
118 	return (u64)ring_info->ring_buffer->write_index << 32;
119 }
120 
121 /*
122  * Helper routine to copy from source to ring buffer.
123  * Assume there is enough room. Handles wrap-around in dest case only!!
124  */
125 static u32 hv_copyto_ringbuffer(
126 	struct hv_ring_buffer_info	*ring_info,
127 	u32				start_write_offset,
128 	const void			*src,
129 	u32				srclen)
130 {
131 	void *ring_buffer = hv_get_ring_buffer(ring_info);
132 	u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
133 
134 	memcpy(ring_buffer + start_write_offset, src, srclen);
135 
136 	start_write_offset += srclen;
137 	if (start_write_offset >= ring_buffer_size)
138 		start_write_offset -= ring_buffer_size;
139 
140 	return start_write_offset;
141 }
142 
143 /*
144  *
145  * hv_get_ringbuffer_availbytes()
146  *
147  * Get number of bytes available to read and to write to
148  * for the specified ring buffer
149  */
150 static void
151 hv_get_ringbuffer_availbytes(const struct hv_ring_buffer_info *rbi,
152 			     u32 *read, u32 *write)
153 {
154 	u32 read_loc, write_loc, dsize;
155 
156 	/* Capture the read/write indices before they changed */
157 	read_loc = READ_ONCE(rbi->ring_buffer->read_index);
158 	write_loc = READ_ONCE(rbi->ring_buffer->write_index);
159 	dsize = rbi->ring_datasize;
160 
161 	*write = write_loc >= read_loc ? dsize - (write_loc - read_loc) :
162 		read_loc - write_loc;
163 	*read = dsize - *write;
164 }
165 
166 /* Get various debug metrics for the specified ring buffer. */
167 void hv_ringbuffer_get_debuginfo(const struct hv_ring_buffer_info *ring_info,
168 				 struct hv_ring_buffer_debug_info *debug_info)
169 {
170 	u32 bytes_avail_towrite;
171 	u32 bytes_avail_toread;
172 
173 	if (ring_info->ring_buffer) {
174 		hv_get_ringbuffer_availbytes(ring_info,
175 					&bytes_avail_toread,
176 					&bytes_avail_towrite);
177 
178 		debug_info->bytes_avail_toread = bytes_avail_toread;
179 		debug_info->bytes_avail_towrite = bytes_avail_towrite;
180 		debug_info->current_read_index =
181 			ring_info->ring_buffer->read_index;
182 		debug_info->current_write_index =
183 			ring_info->ring_buffer->write_index;
184 		debug_info->current_interrupt_mask =
185 			ring_info->ring_buffer->interrupt_mask;
186 	}
187 }
188 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
189 
190 /* Initialize the ring buffer. */
191 int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
192 		       struct page *pages, u32 page_cnt)
193 {
194 	int i;
195 	struct page **pages_wraparound;
196 
197 	BUILD_BUG_ON((sizeof(struct hv_ring_buffer) != PAGE_SIZE));
198 
199 	memset(ring_info, 0, sizeof(struct hv_ring_buffer_info));
200 
201 	/*
202 	 * First page holds struct hv_ring_buffer, do wraparound mapping for
203 	 * the rest.
204 	 */
205 	pages_wraparound = kzalloc(sizeof(struct page *) * (page_cnt * 2 - 1),
206 				   GFP_KERNEL);
207 	if (!pages_wraparound)
208 		return -ENOMEM;
209 
210 	pages_wraparound[0] = pages;
211 	for (i = 0; i < 2 * (page_cnt - 1); i++)
212 		pages_wraparound[i + 1] = &pages[i % (page_cnt - 1) + 1];
213 
214 	ring_info->ring_buffer = (struct hv_ring_buffer *)
215 		vmap(pages_wraparound, page_cnt * 2 - 1, VM_MAP, PAGE_KERNEL);
216 
217 	kfree(pages_wraparound);
218 
219 
220 	if (!ring_info->ring_buffer)
221 		return -ENOMEM;
222 
223 	ring_info->ring_buffer->read_index =
224 		ring_info->ring_buffer->write_index = 0;
225 
226 	/* Set the feature bit for enabling flow control. */
227 	ring_info->ring_buffer->feature_bits.value = 1;
228 
229 	ring_info->ring_size = page_cnt << PAGE_SHIFT;
230 	ring_info->ring_datasize = ring_info->ring_size -
231 		sizeof(struct hv_ring_buffer);
232 
233 	spin_lock_init(&ring_info->ring_lock);
234 
235 	return 0;
236 }
237 
238 /* Cleanup the ring buffer. */
239 void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
240 {
241 	vunmap(ring_info->ring_buffer);
242 }
243 
244 /* Write to the ring buffer. */
245 int hv_ringbuffer_write(struct vmbus_channel *channel,
246 			const struct kvec *kv_list, u32 kv_count)
247 {
248 	int i;
249 	u32 bytes_avail_towrite;
250 	u32 totalbytes_towrite = sizeof(u64);
251 	u32 next_write_location;
252 	u32 old_write;
253 	u64 prev_indices;
254 	unsigned long flags;
255 	struct hv_ring_buffer_info *outring_info = &channel->outbound;
256 
257 	if (channel->rescind)
258 		return -ENODEV;
259 
260 	for (i = 0; i < kv_count; i++)
261 		totalbytes_towrite += kv_list[i].iov_len;
262 
263 	spin_lock_irqsave(&outring_info->ring_lock, flags);
264 
265 	bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
266 
267 	/*
268 	 * If there is only room for the packet, assume it is full.
269 	 * Otherwise, the next time around, we think the ring buffer
270 	 * is empty since the read index == write index.
271 	 */
272 	if (bytes_avail_towrite <= totalbytes_towrite) {
273 		spin_unlock_irqrestore(&outring_info->ring_lock, flags);
274 		return -EAGAIN;
275 	}
276 
277 	/* Write to the ring buffer */
278 	next_write_location = hv_get_next_write_location(outring_info);
279 
280 	old_write = next_write_location;
281 
282 	for (i = 0; i < kv_count; i++) {
283 		next_write_location = hv_copyto_ringbuffer(outring_info,
284 						     next_write_location,
285 						     kv_list[i].iov_base,
286 						     kv_list[i].iov_len);
287 	}
288 
289 	/* Set previous packet start */
290 	prev_indices = hv_get_ring_bufferindices(outring_info);
291 
292 	next_write_location = hv_copyto_ringbuffer(outring_info,
293 					     next_write_location,
294 					     &prev_indices,
295 					     sizeof(u64));
296 
297 	/* Issue a full memory barrier before updating the write index */
298 	virt_mb();
299 
300 	/* Now, update the write location */
301 	hv_set_next_write_location(outring_info, next_write_location);
302 
303 
304 	spin_unlock_irqrestore(&outring_info->ring_lock, flags);
305 
306 	hv_signal_on_write(old_write, channel);
307 
308 	if (channel->rescind)
309 		return -ENODEV;
310 
311 	return 0;
312 }
313 
314 int hv_ringbuffer_read(struct vmbus_channel *channel,
315 		       void *buffer, u32 buflen, u32 *buffer_actual_len,
316 		       u64 *requestid, bool raw)
317 {
318 	struct vmpacket_descriptor *desc;
319 	u32 packetlen, offset;
320 
321 	if (unlikely(buflen == 0))
322 		return -EINVAL;
323 
324 	*buffer_actual_len = 0;
325 	*requestid = 0;
326 
327 	/* Make sure there is something to read */
328 	desc = hv_pkt_iter_first(channel);
329 	if (desc == NULL) {
330 		/*
331 		 * No error is set when there is even no header, drivers are
332 		 * supposed to analyze buffer_actual_len.
333 		 */
334 		return 0;
335 	}
336 
337 	offset = raw ? 0 : (desc->offset8 << 3);
338 	packetlen = (desc->len8 << 3) - offset;
339 	*buffer_actual_len = packetlen;
340 	*requestid = desc->trans_id;
341 
342 	if (unlikely(packetlen > buflen))
343 		return -ENOBUFS;
344 
345 	/* since ring is double mapped, only one copy is necessary */
346 	memcpy(buffer, (const char *)desc + offset, packetlen);
347 
348 	/* Advance ring index to next packet descriptor */
349 	__hv_pkt_iter_next(channel, desc);
350 
351 	/* Notify host of update */
352 	hv_pkt_iter_close(channel);
353 
354 	return 0;
355 }
356 
357 /*
358  * Determine number of bytes available in ring buffer after
359  * the current iterator (priv_read_index) location.
360  *
361  * This is similar to hv_get_bytes_to_read but with private
362  * read index instead.
363  */
364 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
365 {
366 	u32 priv_read_loc = rbi->priv_read_index;
367 	u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
368 
369 	if (write_loc >= priv_read_loc)
370 		return write_loc - priv_read_loc;
371 	else
372 		return (rbi->ring_datasize - priv_read_loc) + write_loc;
373 }
374 
375 /*
376  * Get first vmbus packet from ring buffer after read_index
377  *
378  * If ring buffer is empty, returns NULL and no other action needed.
379  */
380 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
381 {
382 	struct hv_ring_buffer_info *rbi = &channel->inbound;
383 	struct vmpacket_descriptor *desc;
384 
385 	if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
386 		return NULL;
387 
388 	desc = hv_get_ring_buffer(rbi) + rbi->priv_read_index;
389 	if (desc)
390 		prefetch((char *)desc + (desc->len8 << 3));
391 
392 	return desc;
393 }
394 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
395 
396 /*
397  * Get next vmbus packet from ring buffer.
398  *
399  * Advances the current location (priv_read_index) and checks for more
400  * data. If the end of the ring buffer is reached, then return NULL.
401  */
402 struct vmpacket_descriptor *
403 __hv_pkt_iter_next(struct vmbus_channel *channel,
404 		   const struct vmpacket_descriptor *desc)
405 {
406 	struct hv_ring_buffer_info *rbi = &channel->inbound;
407 	u32 packetlen = desc->len8 << 3;
408 	u32 dsize = rbi->ring_datasize;
409 
410 	/* bump offset to next potential packet */
411 	rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
412 	if (rbi->priv_read_index >= dsize)
413 		rbi->priv_read_index -= dsize;
414 
415 	/* more data? */
416 	return hv_pkt_iter_first(channel);
417 }
418 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
419 
420 /*
421  * Update host ring buffer after iterating over packets.
422  */
423 void hv_pkt_iter_close(struct vmbus_channel *channel)
424 {
425 	struct hv_ring_buffer_info *rbi = &channel->inbound;
426 	u32 orig_write_sz = hv_get_bytes_to_write(rbi);
427 
428 	/*
429 	 * Make sure all reads are done before we update the read index since
430 	 * the writer may start writing to the read area once the read index
431 	 * is updated.
432 	 */
433 	virt_rmb();
434 	rbi->ring_buffer->read_index = rbi->priv_read_index;
435 
436 	/*
437 	 * Issue a full memory barrier before making the signaling decision.
438 	 * Here is the reason for having this barrier:
439 	 * If the reading of the pend_sz (in this function)
440 	 * were to be reordered and read before we commit the new read
441 	 * index (in the calling function)  we could
442 	 * have a problem. If the host were to set the pending_sz after we
443 	 * have sampled pending_sz and go to sleep before we commit the
444 	 * read index, we could miss sending the interrupt. Issue a full
445 	 * memory barrier to address this.
446 	 */
447 	virt_mb();
448 
449 	/* If host has disabled notifications then skip */
450 	if (rbi->ring_buffer->interrupt_mask)
451 		return;
452 
453 	if (rbi->ring_buffer->feature_bits.feat_pending_send_sz) {
454 		u32 pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
455 
456 		/*
457 		 * If there was space before we began iteration,
458 		 * then host was not blocked. Also handles case where
459 		 * pending_sz is zero then host has nothing pending
460 		 * and does not need to be signaled.
461 		 */
462 		if (orig_write_sz > pending_sz)
463 			return;
464 
465 		/* If pending write will not fit, don't give false hope. */
466 		if (hv_get_bytes_to_write(rbi) < pending_sz)
467 			return;
468 	}
469 
470 	vmbus_setevent(channel);
471 }
472 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);
473