Loading...
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
31#include "hyperv_vmbus.h"
32
33void hv_begin_read(struct hv_ring_buffer_info *rbi)
34{
35 rbi->ring_buffer->interrupt_mask = 1;
36 mb();
37}
38
39u32 hv_end_read(struct hv_ring_buffer_info *rbi)
40{
41 u32 read;
42 u32 write;
43
44 rbi->ring_buffer->interrupt_mask = 0;
45 mb();
46
47 /*
48 * Now check to see if the ring buffer is still empty.
49 * If it is not, we raced and we need to process new
50 * incoming messages.
51 */
52 hv_get_ringbuffer_availbytes(rbi, &read, &write);
53
54 return read;
55}
56
57/*
58 * When we write to the ring buffer, check if the host needs to
59 * be signaled. Here is the details of this protocol:
60 *
61 * 1. The host guarantees that while it is draining the
62 * ring buffer, it will set the interrupt_mask to
63 * indicate it does not need to be interrupted when
64 * new data is placed.
65 *
66 * 2. The host guarantees that it will completely drain
67 * the ring buffer before exiting the read loop. Further,
68 * once the ring buffer is empty, it will clear the
69 * interrupt_mask and re-check to see if new data has
70 * arrived.
71 */
72
73static bool hv_need_to_signal(u32 old_write, struct hv_ring_buffer_info *rbi)
74{
75 mb();
76 if (rbi->ring_buffer->interrupt_mask)
77 return false;
78
79 /* check interrupt_mask before read_index */
80 rmb();
81 /*
82 * This is the only case we need to signal when the
83 * ring transitions from being empty to non-empty.
84 */
85 if (old_write == rbi->ring_buffer->read_index)
86 return true;
87
88 return false;
89}
90
91/*
92 * To optimize the flow management on the send-side,
93 * when the sender is blocked because of lack of
94 * sufficient space in the ring buffer, potential the
95 * consumer of the ring buffer can signal the producer.
96 * This is controlled by the following parameters:
97 *
98 * 1. pending_send_sz: This is the size in bytes that the
99 * producer is trying to send.
100 * 2. The feature bit feat_pending_send_sz set to indicate if
101 * the consumer of the ring will signal when the ring
102 * state transitions from being full to a state where
103 * there is room for the producer to send the pending packet.
104 */
105
106static bool hv_need_to_signal_on_read(struct hv_ring_buffer_info *rbi)
107{
108 u32 cur_write_sz;
109 u32 r_size;
110 u32 write_loc;
111 u32 read_loc = rbi->ring_buffer->read_index;
112 u32 pending_sz;
113
114 /*
115 * Issue a full memory barrier before making the signaling decision.
116 * Here is the reason for having this barrier:
117 * If the reading of the pend_sz (in this function)
118 * were to be reordered and read before we commit the new read
119 * index (in the calling function) we could
120 * have a problem. If the host were to set the pending_sz after we
121 * have sampled pending_sz and go to sleep before we commit the
122 * read index, we could miss sending the interrupt. Issue a full
123 * memory barrier to address this.
124 */
125 mb();
126
127 pending_sz = rbi->ring_buffer->pending_send_sz;
128 write_loc = rbi->ring_buffer->write_index;
129 /* If the other end is not blocked on write don't bother. */
130 if (pending_sz == 0)
131 return false;
132
133 r_size = rbi->ring_datasize;
134 cur_write_sz = write_loc >= read_loc ? r_size - (write_loc - read_loc) :
135 read_loc - write_loc;
136
137 if (cur_write_sz >= pending_sz)
138 return true;
139
140 return false;
141}
142
143/* Get the next write location for the specified ring buffer. */
144static inline u32
145hv_get_next_write_location(struct hv_ring_buffer_info *ring_info)
146{
147 u32 next = ring_info->ring_buffer->write_index;
148
149 return next;
150}
151
152/* Set the next write location for the specified ring buffer. */
153static inline void
154hv_set_next_write_location(struct hv_ring_buffer_info *ring_info,
155 u32 next_write_location)
156{
157 ring_info->ring_buffer->write_index = next_write_location;
158}
159
160/* Get the next read location for the specified ring buffer. */
161static inline u32
162hv_get_next_read_location(struct hv_ring_buffer_info *ring_info)
163{
164 u32 next = ring_info->ring_buffer->read_index;
165
166 return next;
167}
168
169/*
170 * Get the next read location + offset for the specified ring buffer.
171 * This allows the caller to skip.
172 */
173static inline u32
174hv_get_next_readlocation_withoffset(struct hv_ring_buffer_info *ring_info,
175 u32 offset)
176{
177 u32 next = ring_info->ring_buffer->read_index;
178
179 next += offset;
180 next %= ring_info->ring_datasize;
181
182 return next;
183}
184
185/* Set the next read location for the specified ring buffer. */
186static inline void
187hv_set_next_read_location(struct hv_ring_buffer_info *ring_info,
188 u32 next_read_location)
189{
190 ring_info->ring_buffer->read_index = next_read_location;
191}
192
193
194/* Get the start of the ring buffer. */
195static inline void *
196hv_get_ring_buffer(struct hv_ring_buffer_info *ring_info)
197{
198 return (void *)ring_info->ring_buffer->buffer;
199}
200
201
202/* Get the size of the ring buffer. */
203static inline u32
204hv_get_ring_buffersize(struct hv_ring_buffer_info *ring_info)
205{
206 return ring_info->ring_datasize;
207}
208
209/* Get the read and write indices as u64 of the specified ring buffer. */
210static inline u64
211hv_get_ring_bufferindices(struct hv_ring_buffer_info *ring_info)
212{
213 return (u64)ring_info->ring_buffer->write_index << 32;
214}
215
216/*
217 * Helper routine to copy to source from ring buffer.
218 * Assume there is enough room. Handles wrap-around in src case only!!
219 */
220static u32 hv_copyfrom_ringbuffer(
221 struct hv_ring_buffer_info *ring_info,
222 void *dest,
223 u32 destlen,
224 u32 start_read_offset)
225{
226 void *ring_buffer = hv_get_ring_buffer(ring_info);
227 u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
228
229 u32 frag_len;
230
231 /* wrap-around detected at the src */
232 if (destlen > ring_buffer_size - start_read_offset) {
233 frag_len = ring_buffer_size - start_read_offset;
234
235 memcpy(dest, ring_buffer + start_read_offset, frag_len);
236 memcpy(dest + frag_len, ring_buffer, destlen - frag_len);
237 } else
238
239 memcpy(dest, ring_buffer + start_read_offset, destlen);
240
241
242 start_read_offset += destlen;
243 start_read_offset %= ring_buffer_size;
244
245 return start_read_offset;
246}
247
248
249/*
250 * Helper routine to copy from source to ring buffer.
251 * Assume there is enough room. Handles wrap-around in dest case only!!
252 */
253static u32 hv_copyto_ringbuffer(
254 struct hv_ring_buffer_info *ring_info,
255 u32 start_write_offset,
256 void *src,
257 u32 srclen)
258{
259 void *ring_buffer = hv_get_ring_buffer(ring_info);
260 u32 ring_buffer_size = hv_get_ring_buffersize(ring_info);
261 u32 frag_len;
262
263 /* wrap-around detected! */
264 if (srclen > ring_buffer_size - start_write_offset) {
265 frag_len = ring_buffer_size - start_write_offset;
266 memcpy(ring_buffer + start_write_offset, src, frag_len);
267 memcpy(ring_buffer, src + frag_len, srclen - frag_len);
268 } else
269 memcpy(ring_buffer + start_write_offset, src, srclen);
270
271 start_write_offset += srclen;
272 start_write_offset %= ring_buffer_size;
273
274 return start_write_offset;
275}
276
277/* Get various debug metrics for the specified ring buffer. */
278void hv_ringbuffer_get_debuginfo(struct hv_ring_buffer_info *ring_info,
279 struct hv_ring_buffer_debug_info *debug_info)
280{
281 u32 bytes_avail_towrite;
282 u32 bytes_avail_toread;
283
284 if (ring_info->ring_buffer) {
285 hv_get_ringbuffer_availbytes(ring_info,
286 &bytes_avail_toread,
287 &bytes_avail_towrite);
288
289 debug_info->bytes_avail_toread = bytes_avail_toread;
290 debug_info->bytes_avail_towrite = bytes_avail_towrite;
291 debug_info->current_read_index =
292 ring_info->ring_buffer->read_index;
293 debug_info->current_write_index =
294 ring_info->ring_buffer->write_index;
295 debug_info->current_interrupt_mask =
296 ring_info->ring_buffer->interrupt_mask;
297 }
298}
299
300/* Initialize the ring buffer. */
301int hv_ringbuffer_init(struct hv_ring_buffer_info *ring_info,
302 void *buffer, u32 buflen)
303{
304 if (sizeof(struct hv_ring_buffer) != PAGE_SIZE)
305 return -EINVAL;
306
307 memset(ring_info, 0, sizeof(struct hv_ring_buffer_info));
308
309 ring_info->ring_buffer = (struct hv_ring_buffer *)buffer;
310 ring_info->ring_buffer->read_index =
311 ring_info->ring_buffer->write_index = 0;
312
313 /* Set the feature bit for enabling flow control. */
314 ring_info->ring_buffer->feature_bits.value = 1;
315
316 ring_info->ring_size = buflen;
317 ring_info->ring_datasize = buflen - sizeof(struct hv_ring_buffer);
318
319 spin_lock_init(&ring_info->ring_lock);
320
321 return 0;
322}
323
324/* Cleanup the ring buffer. */
325void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
326{
327}
328
329/* Write to the ring buffer. */
330int hv_ringbuffer_write(struct hv_ring_buffer_info *outring_info,
331 struct kvec *kv_list, u32 kv_count, bool *signal, bool lock)
332{
333 int i = 0;
334 u32 bytes_avail_towrite;
335 u32 bytes_avail_toread;
336 u32 totalbytes_towrite = 0;
337
338 u32 next_write_location;
339 u32 old_write;
340 u64 prev_indices = 0;
341 unsigned long flags = 0;
342
343 for (i = 0; i < kv_count; i++)
344 totalbytes_towrite += kv_list[i].iov_len;
345
346 totalbytes_towrite += sizeof(u64);
347
348 if (lock)
349 spin_lock_irqsave(&outring_info->ring_lock, flags);
350
351 hv_get_ringbuffer_availbytes(outring_info,
352 &bytes_avail_toread,
353 &bytes_avail_towrite);
354
355 /*
356 * If there is only room for the packet, assume it is full.
357 * Otherwise, the next time around, we think the ring buffer
358 * is empty since the read index == write index.
359 */
360 if (bytes_avail_towrite <= totalbytes_towrite) {
361 if (lock)
362 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
363 return -EAGAIN;
364 }
365
366 /* Write to the ring buffer */
367 next_write_location = hv_get_next_write_location(outring_info);
368
369 old_write = next_write_location;
370
371 for (i = 0; i < kv_count; i++) {
372 next_write_location = hv_copyto_ringbuffer(outring_info,
373 next_write_location,
374 kv_list[i].iov_base,
375 kv_list[i].iov_len);
376 }
377
378 /* Set previous packet start */
379 prev_indices = hv_get_ring_bufferindices(outring_info);
380
381 next_write_location = hv_copyto_ringbuffer(outring_info,
382 next_write_location,
383 &prev_indices,
384 sizeof(u64));
385
386 /* Issue a full memory barrier before updating the write index */
387 mb();
388
389 /* Now, update the write location */
390 hv_set_next_write_location(outring_info, next_write_location);
391
392
393 if (lock)
394 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
395
396 *signal = hv_need_to_signal(old_write, outring_info);
397 return 0;
398}
399
400int hv_ringbuffer_read(struct hv_ring_buffer_info *inring_info,
401 void *buffer, u32 buflen, u32 *buffer_actual_len,
402 u64 *requestid, bool *signal, bool raw)
403{
404 u32 bytes_avail_towrite;
405 u32 bytes_avail_toread;
406 u32 next_read_location = 0;
407 u64 prev_indices = 0;
408 struct vmpacket_descriptor desc;
409 u32 offset;
410 u32 packetlen;
411 int ret = 0;
412
413 if (buflen <= 0)
414 return -EINVAL;
415
416
417 *buffer_actual_len = 0;
418 *requestid = 0;
419
420 hv_get_ringbuffer_availbytes(inring_info,
421 &bytes_avail_toread,
422 &bytes_avail_towrite);
423
424 /* Make sure there is something to read */
425 if (bytes_avail_toread < sizeof(desc)) {
426 /*
427 * No error is set when there is even no header, drivers are
428 * supposed to analyze buffer_actual_len.
429 */
430 return ret;
431 }
432
433 next_read_location = hv_get_next_read_location(inring_info);
434 next_read_location = hv_copyfrom_ringbuffer(inring_info, &desc,
435 sizeof(desc),
436 next_read_location);
437
438 offset = raw ? 0 : (desc.offset8 << 3);
439 packetlen = (desc.len8 << 3) - offset;
440 *buffer_actual_len = packetlen;
441 *requestid = desc.trans_id;
442
443 if (bytes_avail_toread < packetlen + offset)
444 return -EAGAIN;
445
446 if (packetlen > buflen)
447 return -ENOBUFS;
448
449 next_read_location =
450 hv_get_next_readlocation_withoffset(inring_info, offset);
451
452 next_read_location = hv_copyfrom_ringbuffer(inring_info,
453 buffer,
454 packetlen,
455 next_read_location);
456
457 next_read_location = hv_copyfrom_ringbuffer(inring_info,
458 &prev_indices,
459 sizeof(u64),
460 next_read_location);
461
462 /*
463 * Make sure all reads are done before we update the read index since
464 * the writer may start writing to the read area once the read index
465 * is updated.
466 */
467 mb();
468
469 /* Update the read index */
470 hv_set_next_read_location(inring_info, next_read_location);
471
472 *signal = hv_need_to_signal_on_read(inring_info);
473
474 return ret;
475}
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);