Merge tag 'dlm-5.15' of git://git.kernel.org/pub/scm/linux/kernel/git/teigland/linux-dlm
[linux-2.6-microblaze.git] / drivers / hv / 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
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 /* Get the size of the ring buffer. */
88 static inline u32
89 hv_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. */
95 static inline u64
96 hv_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  */
105 static 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  */
130 static void
131 hv_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. */
147 int 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 }
173 EXPORT_SYMBOL_GPL(hv_ringbuffer_get_debuginfo);
174
175 /* Initialize a channel's ring buffer info mutex locks */
176 void 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. */
183 int 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. */
240 void 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_size = 0;
249 }
250
251 /* Write to the ring buffer. */
252 int hv_ringbuffer_write(struct vmbus_channel *channel,
253                         const struct kvec *kv_list, u32 kv_count,
254                         u64 requestid)
255 {
256         int i;
257         u32 bytes_avail_towrite;
258         u32 totalbytes_towrite = sizeof(u64);
259         u32 next_write_location;
260         u32 old_write;
261         u64 prev_indices;
262         unsigned long flags;
263         struct hv_ring_buffer_info *outring_info = &channel->outbound;
264         struct vmpacket_descriptor *desc = kv_list[0].iov_base;
265         u64 rqst_id = VMBUS_NO_RQSTOR;
266
267         if (channel->rescind)
268                 return -ENODEV;
269
270         for (i = 0; i < kv_count; i++)
271                 totalbytes_towrite += kv_list[i].iov_len;
272
273         spin_lock_irqsave(&outring_info->ring_lock, flags);
274
275         bytes_avail_towrite = hv_get_bytes_to_write(outring_info);
276
277         /*
278          * If there is only room for the packet, assume it is full.
279          * Otherwise, the next time around, we think the ring buffer
280          * is empty since the read index == write index.
281          */
282         if (bytes_avail_towrite <= totalbytes_towrite) {
283                 ++channel->out_full_total;
284
285                 if (!channel->out_full_flag) {
286                         ++channel->out_full_first;
287                         channel->out_full_flag = true;
288                 }
289
290                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
291                 return -EAGAIN;
292         }
293
294         channel->out_full_flag = false;
295
296         /* Write to the ring buffer */
297         next_write_location = hv_get_next_write_location(outring_info);
298
299         old_write = next_write_location;
300
301         for (i = 0; i < kv_count; i++) {
302                 next_write_location = hv_copyto_ringbuffer(outring_info,
303                                                      next_write_location,
304                                                      kv_list[i].iov_base,
305                                                      kv_list[i].iov_len);
306         }
307
308         /*
309          * Allocate the request ID after the data has been copied into the
310          * ring buffer.  Once this request ID is allocated, the completion
311          * path could find the data and free it.
312          */
313
314         if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {
315                 if (channel->next_request_id_callback != NULL) {
316                         rqst_id = channel->next_request_id_callback(channel, requestid);
317                         if (rqst_id == VMBUS_RQST_ERROR) {
318                                 spin_unlock_irqrestore(&outring_info->ring_lock, flags);
319                                 return -EAGAIN;
320                         }
321                 }
322         }
323         desc = hv_get_ring_buffer(outring_info) + old_write;
324         desc->trans_id = (rqst_id == VMBUS_NO_RQSTOR) ? requestid : rqst_id;
325
326         /* Set previous packet start */
327         prev_indices = hv_get_ring_bufferindices(outring_info);
328
329         next_write_location = hv_copyto_ringbuffer(outring_info,
330                                              next_write_location,
331                                              &prev_indices,
332                                              sizeof(u64));
333
334         /* Issue a full memory barrier before updating the write index */
335         virt_mb();
336
337         /* Now, update the write location */
338         hv_set_next_write_location(outring_info, next_write_location);
339
340
341         spin_unlock_irqrestore(&outring_info->ring_lock, flags);
342
343         hv_signal_on_write(old_write, channel);
344
345         if (channel->rescind) {
346                 if (rqst_id != VMBUS_NO_RQSTOR) {
347                         /* Reclaim request ID to avoid leak of IDs */
348                         if (channel->request_addr_callback != NULL)
349                                 channel->request_addr_callback(channel, rqst_id);
350                 }
351                 return -ENODEV;
352         }
353
354         return 0;
355 }
356
357 int hv_ringbuffer_read(struct vmbus_channel *channel,
358                        void *buffer, u32 buflen, u32 *buffer_actual_len,
359                        u64 *requestid, bool raw)
360 {
361         struct vmpacket_descriptor *desc;
362         u32 packetlen, offset;
363
364         if (unlikely(buflen == 0))
365                 return -EINVAL;
366
367         *buffer_actual_len = 0;
368         *requestid = 0;
369
370         /* Make sure there is something to read */
371         desc = hv_pkt_iter_first(channel);
372         if (desc == NULL) {
373                 /*
374                  * No error is set when there is even no header, drivers are
375                  * supposed to analyze buffer_actual_len.
376                  */
377                 return 0;
378         }
379
380         offset = raw ? 0 : (desc->offset8 << 3);
381         packetlen = (desc->len8 << 3) - offset;
382         *buffer_actual_len = packetlen;
383         *requestid = desc->trans_id;
384
385         if (unlikely(packetlen > buflen))
386                 return -ENOBUFS;
387
388         /* since ring is double mapped, only one copy is necessary */
389         memcpy(buffer, (const char *)desc + offset, packetlen);
390
391         /* Advance ring index to next packet descriptor */
392         __hv_pkt_iter_next(channel, desc, true);
393
394         /* Notify host of update */
395         hv_pkt_iter_close(channel);
396
397         return 0;
398 }
399
400 /*
401  * Determine number of bytes available in ring buffer after
402  * the current iterator (priv_read_index) location.
403  *
404  * This is similar to hv_get_bytes_to_read but with private
405  * read index instead.
406  */
407 static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
408 {
409         u32 priv_read_loc = rbi->priv_read_index;
410         u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
411
412         if (write_loc >= priv_read_loc)
413                 return write_loc - priv_read_loc;
414         else
415                 return (rbi->ring_datasize - priv_read_loc) + write_loc;
416 }
417
418 /*
419  * Get first vmbus packet without copying it out of the ring buffer
420  */
421 struct vmpacket_descriptor *hv_pkt_iter_first_raw(struct vmbus_channel *channel)
422 {
423         struct hv_ring_buffer_info *rbi = &channel->inbound;
424
425         hv_debug_delay_test(channel, MESSAGE_DELAY);
426
427         if (hv_pkt_iter_avail(rbi) < sizeof(struct vmpacket_descriptor))
428                 return NULL;
429
430         return (struct vmpacket_descriptor *)(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
431 }
432 EXPORT_SYMBOL_GPL(hv_pkt_iter_first_raw);
433
434 /*
435  * Get first vmbus packet from ring buffer after read_index
436  *
437  * If ring buffer is empty, returns NULL and no other action needed.
438  */
439 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel *channel)
440 {
441         struct hv_ring_buffer_info *rbi = &channel->inbound;
442         struct vmpacket_descriptor *desc, *desc_copy;
443         u32 bytes_avail, pkt_len, pkt_offset;
444
445         desc = hv_pkt_iter_first_raw(channel);
446         if (!desc)
447                 return NULL;
448
449         bytes_avail = min(rbi->pkt_buffer_size, hv_pkt_iter_avail(rbi));
450
451         /*
452          * Ensure the compiler does not use references to incoming Hyper-V values (which
453          * could change at any moment) when reading local variables later in the code
454          */
455         pkt_len = READ_ONCE(desc->len8) << 3;
456         pkt_offset = READ_ONCE(desc->offset8) << 3;
457
458         /*
459          * If pkt_len is invalid, set it to the smaller of hv_pkt_iter_avail() and
460          * rbi->pkt_buffer_size
461          */
462         if (pkt_len < sizeof(struct vmpacket_descriptor) || pkt_len > bytes_avail)
463                 pkt_len = bytes_avail;
464
465         /*
466          * If pkt_offset is invalid, arbitrarily set it to
467          * the size of vmpacket_descriptor
468          */
469         if (pkt_offset < sizeof(struct vmpacket_descriptor) || pkt_offset > pkt_len)
470                 pkt_offset = sizeof(struct vmpacket_descriptor);
471
472         /* Copy the Hyper-V packet out of the ring buffer */
473         desc_copy = (struct vmpacket_descriptor *)rbi->pkt_buffer;
474         memcpy(desc_copy, desc, pkt_len);
475
476         /*
477          * Hyper-V could still change len8 and offset8 after the earlier read.
478          * Ensure that desc_copy has legal values for len8 and offset8 that
479          * are consistent with the copy we just made
480          */
481         desc_copy->len8 = pkt_len >> 3;
482         desc_copy->offset8 = pkt_offset >> 3;
483
484         return desc_copy;
485 }
486 EXPORT_SYMBOL_GPL(hv_pkt_iter_first);
487
488 /*
489  * Get next vmbus packet from ring buffer.
490  *
491  * Advances the current location (priv_read_index) and checks for more
492  * data. If the end of the ring buffer is reached, then return NULL.
493  */
494 struct vmpacket_descriptor *
495 __hv_pkt_iter_next(struct vmbus_channel *channel,
496                    const struct vmpacket_descriptor *desc,
497                    bool copy)
498 {
499         struct hv_ring_buffer_info *rbi = &channel->inbound;
500         u32 packetlen = desc->len8 << 3;
501         u32 dsize = rbi->ring_datasize;
502
503         hv_debug_delay_test(channel, MESSAGE_DELAY);
504         /* bump offset to next potential packet */
505         rbi->priv_read_index += packetlen + VMBUS_PKT_TRAILER;
506         if (rbi->priv_read_index >= dsize)
507                 rbi->priv_read_index -= dsize;
508
509         /* more data? */
510         return copy ? hv_pkt_iter_first(channel) : hv_pkt_iter_first_raw(channel);
511 }
512 EXPORT_SYMBOL_GPL(__hv_pkt_iter_next);
513
514 /* How many bytes were read in this iterator cycle */
515 static u32 hv_pkt_iter_bytes_read(const struct hv_ring_buffer_info *rbi,
516                                         u32 start_read_index)
517 {
518         if (rbi->priv_read_index >= start_read_index)
519                 return rbi->priv_read_index - start_read_index;
520         else
521                 return rbi->ring_datasize - start_read_index +
522                         rbi->priv_read_index;
523 }
524
525 /*
526  * Update host ring buffer after iterating over packets. If the host has
527  * stopped queuing new entries because it found the ring buffer full, and
528  * sufficient space is being freed up, signal the host. But be careful to
529  * only signal the host when necessary, both for performance reasons and
530  * because Hyper-V protects itself by throttling guests that signal
531  * inappropriately.
532  *
533  * Determining when to signal is tricky. There are three key data inputs
534  * that must be handled in this order to avoid race conditions:
535  *
536  * 1. Update the read_index
537  * 2. Read the pending_send_sz
538  * 3. Read the current write_index
539  *
540  * The interrupt_mask is not used to determine when to signal. The
541  * interrupt_mask is used only on the guest->host ring buffer when
542  * sending requests to the host. The host does not use it on the host->
543  * guest ring buffer to indicate whether it should be signaled.
544  */
545 void hv_pkt_iter_close(struct vmbus_channel *channel)
546 {
547         struct hv_ring_buffer_info *rbi = &channel->inbound;
548         u32 curr_write_sz, pending_sz, bytes_read, start_read_index;
549
550         /*
551          * Make sure all reads are done before we update the read index since
552          * the writer may start writing to the read area once the read index
553          * is updated.
554          */
555         virt_rmb();
556         start_read_index = rbi->ring_buffer->read_index;
557         rbi->ring_buffer->read_index = rbi->priv_read_index;
558
559         /*
560          * Older versions of Hyper-V (before WS2102 and Win8) do not
561          * implement pending_send_sz and simply poll if the host->guest
562          * ring buffer is full.  No signaling is needed or expected.
563          */
564         if (!rbi->ring_buffer->feature_bits.feat_pending_send_sz)
565                 return;
566
567         /*
568          * Issue a full memory barrier before making the signaling decision.
569          * If reading pending_send_sz were to be reordered and happen
570          * before we commit the new read_index, a race could occur.  If the
571          * host were to set the pending_send_sz after we have sampled
572          * pending_send_sz, and the ring buffer blocks before we commit the
573          * read index, we could miss sending the interrupt. Issue a full
574          * memory barrier to address this.
575          */
576         virt_mb();
577
578         /*
579          * If the pending_send_sz is zero, then the ring buffer is not
580          * blocked and there is no need to signal.  This is far by the
581          * most common case, so exit quickly for best performance.
582          */
583         pending_sz = READ_ONCE(rbi->ring_buffer->pending_send_sz);
584         if (!pending_sz)
585                 return;
586
587         /*
588          * Ensure the read of write_index in hv_get_bytes_to_write()
589          * happens after the read of pending_send_sz.
590          */
591         virt_rmb();
592         curr_write_sz = hv_get_bytes_to_write(rbi);
593         bytes_read = hv_pkt_iter_bytes_read(rbi, start_read_index);
594
595         /*
596          * We want to signal the host only if we're transitioning
597          * from a "not enough free space" state to a "enough free
598          * space" state.  For example, it's possible that this function
599          * could run and free up enough space to signal the host, and then
600          * run again and free up additional space before the host has a
601          * chance to clear the pending_send_sz.  The 2nd invocation would
602          * be a null transition from "enough free space" to "enough free
603          * space", which doesn't warrant a signal.
604          *
605          * Exactly filling the ring buffer is treated as "not enough
606          * space". The ring buffer always must have at least one byte
607          * empty so the empty and full conditions are distinguishable.
608          * hv_get_bytes_to_write() doesn't fully tell the truth in
609          * this regard.
610          *
611          * So first check if we were in the "enough free space" state
612          * before we began the iteration. If so, the host was not
613          * blocked, and there's no need to signal.
614          */
615         if (curr_write_sz - bytes_read > pending_sz)
616                 return;
617
618         /*
619          * Similarly, if the new state is "not enough space", then
620          * there's no need to signal.
621          */
622         if (curr_write_sz <= pending_sz)
623                 return;
624
625         ++channel->intr_in_full;
626         vmbus_setevent(channel);
627 }
628 EXPORT_SYMBOL_GPL(hv_pkt_iter_close);