xref: /openbmc/linux/drivers/net/ipa/gsi_trans.c (revision c4a7b9b5)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 /* Copyright (c) 2012-2018, The Linux Foundation. All rights reserved.
4  * Copyright (C) 2019-2022 Linaro Ltd.
5  */
6 
7 #include <linux/types.h>
8 #include <linux/bits.h>
9 #include <linux/bitfield.h>
10 #include <linux/refcount.h>
11 #include <linux/scatterlist.h>
12 #include <linux/dma-direction.h>
13 
14 #include "gsi.h"
15 #include "gsi_private.h"
16 #include "gsi_trans.h"
17 #include "ipa_gsi.h"
18 #include "ipa_data.h"
19 #include "ipa_cmd.h"
20 
21 /**
22  * DOC: GSI Transactions
23  *
24  * A GSI transaction abstracts the behavior of a GSI channel by representing
25  * everything about a related group of IPA operations in a single structure.
26  * (A "operation" in this sense is either a data transfer or an IPA immediate
27  * command.)  Most details of interaction with the GSI hardware are managed
28  * by the GSI transaction core, allowing users to simply describe operations
29  * to be performed.  When a transaction has completed a callback function
30  * (dependent on the type of endpoint associated with the channel) allows
31  * cleanup of resources associated with the transaction.
32  *
33  * To perform an operation (or set of them), a user of the GSI transaction
34  * interface allocates a transaction, indicating the number of TREs required
35  * (one per operation).  If sufficient TREs are available, they are reserved
36  * for use in the transaction and the allocation succeeds.  This way
37  * exhaustion of the available TREs in a channel ring is detected as early
38  * as possible.  Any other resources that might be needed to complete a
39  * transaction are also allocated when the transaction is allocated.
40  *
41  * Operations performed as part of a transaction are represented in an array
42  * of Linux scatterlist structures, allocated with the transaction.  These
43  * scatterlist structures are initialized by "adding" operations to the
44  * transaction.  If a buffer in an operation must be mapped for DMA, this is
45  * done at the time it is added to the transaction.  It is possible for a
46  * mapping error to occur when an operation is added.  In this case the
47  * transaction should simply be freed; this correctly releases resources
48  * associated with the transaction.
49  *
50  * Once all operations have been successfully added to a transaction, the
51  * transaction is committed.  Committing transfers ownership of the entire
52  * transaction to the GSI transaction core.  The GSI transaction code
53  * formats the content of the scatterlist array into the channel ring
54  * buffer and informs the hardware that new TREs are available to process.
55  *
56  * The last TRE in each transaction is marked to interrupt the AP when the
57  * GSI hardware has completed it.  Because transfers described by TREs are
58  * performed strictly in order, signaling the completion of just the last
59  * TRE in the transaction is sufficient to indicate the full transaction
60  * is complete.
61  *
62  * When a transaction is complete, ipa_gsi_trans_complete() is called by the
63  * GSI code into the IPA layer, allowing it to perform any final cleanup
64  * required before the transaction is freed.
65  */
66 
67 /* Hardware values representing a transfer element type */
68 enum gsi_tre_type {
69 	GSI_RE_XFER	= 0x2,
70 	GSI_RE_IMMD_CMD	= 0x3,
71 };
72 
73 /* An entry in a channel ring */
74 struct gsi_tre {
75 	__le64 addr;		/* DMA address */
76 	__le16 len_opcode;	/* length in bytes or enum IPA_CMD_* */
77 	__le16 reserved;
78 	__le32 flags;		/* TRE_FLAGS_* */
79 };
80 
81 /* gsi_tre->flags mask values (in CPU byte order) */
82 #define TRE_FLAGS_CHAIN_FMASK	GENMASK(0, 0)
83 #define TRE_FLAGS_IEOT_FMASK	GENMASK(9, 9)
84 #define TRE_FLAGS_BEI_FMASK	GENMASK(10, 10)
85 #define TRE_FLAGS_TYPE_FMASK	GENMASK(23, 16)
86 
87 int gsi_trans_pool_init(struct gsi_trans_pool *pool, size_t size, u32 count,
88 			u32 max_alloc)
89 {
90 	void *virt;
91 
92 	if (!size)
93 		return -EINVAL;
94 	if (count < max_alloc)
95 		return -EINVAL;
96 	if (!max_alloc)
97 		return -EINVAL;
98 
99 	/* By allocating a few extra entries in our pool (one less
100 	 * than the maximum number that will be requested in a
101 	 * single allocation), we can always satisfy requests without
102 	 * ever worrying about straddling the end of the pool array.
103 	 * If there aren't enough entries starting at the free index,
104 	 * we just allocate free entries from the beginning of the pool.
105 	 */
106 	virt = kcalloc(count + max_alloc - 1, size, GFP_KERNEL);
107 	if (!virt)
108 		return -ENOMEM;
109 
110 	pool->base = virt;
111 	/* If the allocator gave us any extra memory, use it */
112 	pool->count = ksize(pool->base) / size;
113 	pool->free = 0;
114 	pool->max_alloc = max_alloc;
115 	pool->size = size;
116 	pool->addr = 0;		/* Only used for DMA pools */
117 
118 	return 0;
119 }
120 
121 void gsi_trans_pool_exit(struct gsi_trans_pool *pool)
122 {
123 	kfree(pool->base);
124 	memset(pool, 0, sizeof(*pool));
125 }
126 
127 /* Home-grown DMA pool.  This way we can preallocate the pool, and guarantee
128  * allocations will succeed.  The immediate commands in a transaction can
129  * require up to max_alloc elements from the pool.  But we only allow
130  * allocation of a single element from a DMA pool at a time.
131  */
132 int gsi_trans_pool_init_dma(struct device *dev, struct gsi_trans_pool *pool,
133 			    size_t size, u32 count, u32 max_alloc)
134 {
135 	size_t total_size;
136 	dma_addr_t addr;
137 	void *virt;
138 
139 	if (!size)
140 		return -EINVAL;
141 	if (count < max_alloc)
142 		return -EINVAL;
143 	if (!max_alloc)
144 		return -EINVAL;
145 
146 	/* Don't let allocations cross a power-of-two boundary */
147 	size = __roundup_pow_of_two(size);
148 	total_size = (count + max_alloc - 1) * size;
149 
150 	/* The allocator will give us a power-of-2 number of pages
151 	 * sufficient to satisfy our request.  Round up our requested
152 	 * size to avoid any unused space in the allocation.  This way
153 	 * gsi_trans_pool_exit_dma() can assume the total allocated
154 	 * size is exactly (count * size).
155 	 */
156 	total_size = get_order(total_size) << PAGE_SHIFT;
157 
158 	virt = dma_alloc_coherent(dev, total_size, &addr, GFP_KERNEL);
159 	if (!virt)
160 		return -ENOMEM;
161 
162 	pool->base = virt;
163 	pool->count = total_size / size;
164 	pool->free = 0;
165 	pool->size = size;
166 	pool->max_alloc = max_alloc;
167 	pool->addr = addr;
168 
169 	return 0;
170 }
171 
172 void gsi_trans_pool_exit_dma(struct device *dev, struct gsi_trans_pool *pool)
173 {
174 	size_t total_size = pool->count * pool->size;
175 
176 	dma_free_coherent(dev, total_size, pool->base, pool->addr);
177 	memset(pool, 0, sizeof(*pool));
178 }
179 
180 /* Return the byte offset of the next free entry in the pool */
181 static u32 gsi_trans_pool_alloc_common(struct gsi_trans_pool *pool, u32 count)
182 {
183 	u32 offset;
184 
185 	WARN_ON(!count);
186 	WARN_ON(count > pool->max_alloc);
187 
188 	/* Allocate from beginning if wrap would occur */
189 	if (count > pool->count - pool->free)
190 		pool->free = 0;
191 
192 	offset = pool->free * pool->size;
193 	pool->free += count;
194 	memset(pool->base + offset, 0, count * pool->size);
195 
196 	return offset;
197 }
198 
199 /* Allocate a contiguous block of zeroed entries from a pool */
200 void *gsi_trans_pool_alloc(struct gsi_trans_pool *pool, u32 count)
201 {
202 	return pool->base + gsi_trans_pool_alloc_common(pool, count);
203 }
204 
205 /* Allocate a single zeroed entry from a DMA pool */
206 void *gsi_trans_pool_alloc_dma(struct gsi_trans_pool *pool, dma_addr_t *addr)
207 {
208 	u32 offset = gsi_trans_pool_alloc_common(pool, 1);
209 
210 	*addr = pool->addr + offset;
211 
212 	return pool->base + offset;
213 }
214 
215 /* Map a TRE ring entry index to the transaction it is associated with */
216 static void gsi_trans_map(struct gsi_trans *trans, u32 index)
217 {
218 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
219 
220 	/* The completion event will indicate the last TRE used */
221 	index += trans->used_count - 1;
222 
223 	/* Note: index *must* be used modulo the ring count here */
224 	channel->trans_info.map[index % channel->tre_ring.count] = trans;
225 }
226 
227 /* Return the transaction mapped to a given ring entry */
228 struct gsi_trans *
229 gsi_channel_trans_mapped(struct gsi_channel *channel, u32 index)
230 {
231 	/* Note: index *must* be used modulo the ring count here */
232 	return channel->trans_info.map[index % channel->tre_ring.count];
233 }
234 
235 /* Return the oldest completed transaction for a channel (or null) */
236 struct gsi_trans *gsi_channel_trans_complete(struct gsi_channel *channel)
237 {
238 	struct gsi_trans_info *trans_info = &channel->trans_info;
239 	u16 trans_id = trans_info->completed_id;
240 
241 	if (trans_id == trans_info->pending_id) {
242 		gsi_channel_update(channel);
243 		if (trans_id == trans_info->pending_id)
244 			return NULL;
245 	}
246 
247 	return &trans_info->trans[trans_id %= channel->tre_count];
248 }
249 
250 /* Move a transaction from allocated to committed state */
251 static void gsi_trans_move_committed(struct gsi_trans *trans)
252 {
253 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
254 	struct gsi_trans_info *trans_info = &channel->trans_info;
255 
256 	/* This allocated transaction is now committed */
257 	trans_info->allocated_id++;
258 }
259 
260 /* Move committed transactions to pending state */
261 static void gsi_trans_move_pending(struct gsi_trans *trans)
262 {
263 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
264 	struct gsi_trans_info *trans_info = &channel->trans_info;
265 	u16 trans_index = trans - &trans_info->trans[0];
266 	u16 delta;
267 
268 	/* These committed transactions are now pending */
269 	delta = trans_index - trans_info->committed_id + 1;
270 	trans_info->committed_id += delta % channel->tre_count;
271 }
272 
273 /* Move pending transactions to completed state */
274 void gsi_trans_move_complete(struct gsi_trans *trans)
275 {
276 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
277 	struct gsi_trans_info *trans_info = &channel->trans_info;
278 	u16 trans_index = trans - trans_info->trans;
279 	u16 delta;
280 
281 	/* These pending transactions are now completed */
282 	delta = trans_index - trans_info->pending_id + 1;
283 	delta %= channel->tre_count;
284 	trans_info->pending_id += delta;
285 }
286 
287 /* Move a transaction from completed to polled state */
288 void gsi_trans_move_polled(struct gsi_trans *trans)
289 {
290 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
291 	struct gsi_trans_info *trans_info = &channel->trans_info;
292 
293 	/* This completed transaction is now polled */
294 	trans_info->completed_id++;
295 }
296 
297 /* Reserve some number of TREs on a channel.  Returns true if successful */
298 static bool
299 gsi_trans_tre_reserve(struct gsi_trans_info *trans_info, u32 tre_count)
300 {
301 	int avail = atomic_read(&trans_info->tre_avail);
302 	int new;
303 
304 	do {
305 		new = avail - (int)tre_count;
306 		if (unlikely(new < 0))
307 			return false;
308 	} while (!atomic_try_cmpxchg(&trans_info->tre_avail, &avail, new));
309 
310 	return true;
311 }
312 
313 /* Release previously-reserved TRE entries to a channel */
314 static void
315 gsi_trans_tre_release(struct gsi_trans_info *trans_info, u32 tre_count)
316 {
317 	atomic_add(tre_count, &trans_info->tre_avail);
318 }
319 
320 /* Return true if no transactions are allocated, false otherwise */
321 bool gsi_channel_trans_idle(struct gsi *gsi, u32 channel_id)
322 {
323 	u32 tre_max = gsi_channel_tre_max(gsi, channel_id);
324 	struct gsi_trans_info *trans_info;
325 
326 	trans_info = &gsi->channel[channel_id].trans_info;
327 
328 	return atomic_read(&trans_info->tre_avail) == tre_max;
329 }
330 
331 /* Allocate a GSI transaction on a channel */
332 struct gsi_trans *gsi_channel_trans_alloc(struct gsi *gsi, u32 channel_id,
333 					  u32 tre_count,
334 					  enum dma_data_direction direction)
335 {
336 	struct gsi_channel *channel = &gsi->channel[channel_id];
337 	struct gsi_trans_info *trans_info;
338 	struct gsi_trans *trans;
339 	u16 trans_index;
340 
341 	if (WARN_ON(tre_count > channel->trans_tre_max))
342 		return NULL;
343 
344 	trans_info = &channel->trans_info;
345 
346 	/* If we can't reserve the TREs for the transaction, we're done */
347 	if (!gsi_trans_tre_reserve(trans_info, tre_count))
348 		return NULL;
349 
350 	trans_index = trans_info->free_id % channel->tre_count;
351 	trans = &trans_info->trans[trans_index];
352 	memset(trans, 0, sizeof(*trans));
353 
354 	/* Initialize non-zero fields in the transaction */
355 	trans->gsi = gsi;
356 	trans->channel_id = channel_id;
357 	trans->rsvd_count = tre_count;
358 	init_completion(&trans->completion);
359 
360 	/* Allocate the scatterlist */
361 	trans->sgl = gsi_trans_pool_alloc(&trans_info->sg_pool, tre_count);
362 	sg_init_marker(trans->sgl, tre_count);
363 
364 	trans->direction = direction;
365 	refcount_set(&trans->refcount, 1);
366 
367 	/* This free transaction is now allocated */
368 	trans_info->free_id++;
369 
370 	return trans;
371 }
372 
373 /* Free a previously-allocated transaction */
374 void gsi_trans_free(struct gsi_trans *trans)
375 {
376 	struct gsi_trans_info *trans_info;
377 
378 	if (!refcount_dec_and_test(&trans->refcount))
379 		return;
380 
381 	/* Unused transactions are allocated but never committed, pending,
382 	 * completed, or polled.
383 	 */
384 	trans_info = &trans->gsi->channel[trans->channel_id].trans_info;
385 	if (!trans->used_count) {
386 		trans_info->allocated_id++;
387 		trans_info->committed_id++;
388 		trans_info->pending_id++;
389 		trans_info->completed_id++;
390 	} else {
391 		ipa_gsi_trans_release(trans);
392 	}
393 
394 	/* This transaction is now free */
395 	trans_info->polled_id++;
396 
397 	/* Releasing the reserved TREs implicitly frees the sgl[] and
398 	 * (if present) info[] arrays, plus the transaction itself.
399 	 */
400 	gsi_trans_tre_release(trans_info, trans->rsvd_count);
401 }
402 
403 /* Add an immediate command to a transaction */
404 void gsi_trans_cmd_add(struct gsi_trans *trans, void *buf, u32 size,
405 		       dma_addr_t addr, enum ipa_cmd_opcode opcode)
406 {
407 	u32 which = trans->used_count++;
408 	struct scatterlist *sg;
409 
410 	WARN_ON(which >= trans->rsvd_count);
411 
412 	/* Commands are quite different from data transfer requests.
413 	 * Their payloads come from a pool whose memory is allocated
414 	 * using dma_alloc_coherent().  We therefore do *not* map them
415 	 * for DMA (unlike what we do for pages and skbs).
416 	 *
417 	 * When a transaction completes, the SGL is normally unmapped.
418 	 * A command transaction has direction DMA_NONE, which tells
419 	 * gsi_trans_complete() to skip the unmapping step.
420 	 *
421 	 * The only things we use directly in a command scatter/gather
422 	 * entry are the DMA address and length.  We still need the SG
423 	 * table flags to be maintained though, so assign a NULL page
424 	 * pointer for that purpose.
425 	 */
426 	sg = &trans->sgl[which];
427 	sg_assign_page(sg, NULL);
428 	sg_dma_address(sg) = addr;
429 	sg_dma_len(sg) = size;
430 
431 	trans->cmd_opcode[which] = opcode;
432 }
433 
434 /* Add a page transfer to a transaction.  It will fill the only TRE. */
435 int gsi_trans_page_add(struct gsi_trans *trans, struct page *page, u32 size,
436 		       u32 offset)
437 {
438 	struct scatterlist *sg = &trans->sgl[0];
439 	int ret;
440 
441 	if (WARN_ON(trans->rsvd_count != 1))
442 		return -EINVAL;
443 	if (WARN_ON(trans->used_count))
444 		return -EINVAL;
445 
446 	sg_set_page(sg, page, size, offset);
447 	ret = dma_map_sg(trans->gsi->dev, sg, 1, trans->direction);
448 	if (!ret)
449 		return -ENOMEM;
450 
451 	trans->used_count++;	/* Transaction now owns the (DMA mapped) page */
452 
453 	return 0;
454 }
455 
456 /* Add an SKB transfer to a transaction.  No other TREs will be used. */
457 int gsi_trans_skb_add(struct gsi_trans *trans, struct sk_buff *skb)
458 {
459 	struct scatterlist *sg = &trans->sgl[0];
460 	u32 used_count;
461 	int ret;
462 
463 	if (WARN_ON(trans->rsvd_count != 1))
464 		return -EINVAL;
465 	if (WARN_ON(trans->used_count))
466 		return -EINVAL;
467 
468 	/* skb->len will not be 0 (checked early) */
469 	ret = skb_to_sgvec(skb, sg, 0, skb->len);
470 	if (ret < 0)
471 		return ret;
472 	used_count = ret;
473 
474 	ret = dma_map_sg(trans->gsi->dev, sg, used_count, trans->direction);
475 	if (!ret)
476 		return -ENOMEM;
477 
478 	/* Transaction now owns the (DMA mapped) skb */
479 	trans->used_count += used_count;
480 
481 	return 0;
482 }
483 
484 /* Compute the length/opcode value to use for a TRE */
485 static __le16 gsi_tre_len_opcode(enum ipa_cmd_opcode opcode, u32 len)
486 {
487 	return opcode == IPA_CMD_NONE ? cpu_to_le16((u16)len)
488 				      : cpu_to_le16((u16)opcode);
489 }
490 
491 /* Compute the flags value to use for a given TRE */
492 static __le32 gsi_tre_flags(bool last_tre, bool bei, enum ipa_cmd_opcode opcode)
493 {
494 	enum gsi_tre_type tre_type;
495 	u32 tre_flags;
496 
497 	tre_type = opcode == IPA_CMD_NONE ? GSI_RE_XFER : GSI_RE_IMMD_CMD;
498 	tre_flags = u32_encode_bits(tre_type, TRE_FLAGS_TYPE_FMASK);
499 
500 	/* Last TRE contains interrupt flags */
501 	if (last_tre) {
502 		/* All transactions end in a transfer completion interrupt */
503 		tre_flags |= TRE_FLAGS_IEOT_FMASK;
504 		/* Don't interrupt when outbound commands are acknowledged */
505 		if (bei)
506 			tre_flags |= TRE_FLAGS_BEI_FMASK;
507 	} else {	/* All others indicate there's more to come */
508 		tre_flags |= TRE_FLAGS_CHAIN_FMASK;
509 	}
510 
511 	return cpu_to_le32(tre_flags);
512 }
513 
514 static void gsi_trans_tre_fill(struct gsi_tre *dest_tre, dma_addr_t addr,
515 			       u32 len, bool last_tre, bool bei,
516 			       enum ipa_cmd_opcode opcode)
517 {
518 	struct gsi_tre tre;
519 
520 	tre.addr = cpu_to_le64(addr);
521 	tre.len_opcode = gsi_tre_len_opcode(opcode, len);
522 	tre.reserved = 0;
523 	tre.flags = gsi_tre_flags(last_tre, bei, opcode);
524 
525 	/* ARM64 can write 16 bytes as a unit with a single instruction.
526 	 * Doing the assignment this way is an attempt to make that happen.
527 	 */
528 	*dest_tre = tre;
529 }
530 
531 /**
532  * __gsi_trans_commit() - Common GSI transaction commit code
533  * @trans:	Transaction to commit
534  * @ring_db:	Whether to tell the hardware about these queued transfers
535  *
536  * Formats channel ring TRE entries based on the content of the scatterlist.
537  * Maps a transaction pointer to the last ring entry used for the transaction,
538  * so it can be recovered when it completes.  Moves the transaction to
539  * pending state.  Finally, updates the channel ring pointer and optionally
540  * rings the doorbell.
541  */
542 static void __gsi_trans_commit(struct gsi_trans *trans, bool ring_db)
543 {
544 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
545 	struct gsi_ring *tre_ring = &channel->tre_ring;
546 	enum ipa_cmd_opcode opcode = IPA_CMD_NONE;
547 	bool bei = channel->toward_ipa;
548 	struct gsi_tre *dest_tre;
549 	struct scatterlist *sg;
550 	u32 byte_count = 0;
551 	u8 *cmd_opcode;
552 	u32 avail;
553 	u32 i;
554 
555 	WARN_ON(!trans->used_count);
556 
557 	/* Consume the entries.  If we cross the end of the ring while
558 	 * filling them we'll switch to the beginning to finish.
559 	 * If there is no info array we're doing a simple data
560 	 * transfer request, whose opcode is IPA_CMD_NONE.
561 	 */
562 	cmd_opcode = channel->command ? &trans->cmd_opcode[0] : NULL;
563 	avail = tre_ring->count - tre_ring->index % tre_ring->count;
564 	dest_tre = gsi_ring_virt(tre_ring, tre_ring->index);
565 	for_each_sg(trans->sgl, sg, trans->used_count, i) {
566 		bool last_tre = i == trans->used_count - 1;
567 		dma_addr_t addr = sg_dma_address(sg);
568 		u32 len = sg_dma_len(sg);
569 
570 		byte_count += len;
571 		if (!avail--)
572 			dest_tre = gsi_ring_virt(tre_ring, 0);
573 		if (cmd_opcode)
574 			opcode = *cmd_opcode++;
575 
576 		gsi_trans_tre_fill(dest_tre, addr, len, last_tre, bei, opcode);
577 		dest_tre++;
578 	}
579 	/* Associate the TRE with the transaction */
580 	gsi_trans_map(trans, tre_ring->index);
581 
582 	tre_ring->index += trans->used_count;
583 
584 	trans->len = byte_count;
585 	if (channel->toward_ipa)
586 		gsi_trans_tx_committed(trans);
587 
588 	gsi_trans_move_committed(trans);
589 
590 	/* Ring doorbell if requested, or if all TREs are allocated */
591 	if (ring_db || !atomic_read(&channel->trans_info.tre_avail)) {
592 		/* Report what we're handing off to hardware for TX channels */
593 		if (channel->toward_ipa)
594 			gsi_trans_tx_queued(trans);
595 		gsi_trans_move_pending(trans);
596 		gsi_channel_doorbell(channel);
597 	}
598 }
599 
600 /* Commit a GSI transaction */
601 void gsi_trans_commit(struct gsi_trans *trans, bool ring_db)
602 {
603 	if (trans->used_count)
604 		__gsi_trans_commit(trans, ring_db);
605 	else
606 		gsi_trans_free(trans);
607 }
608 
609 /* Commit a GSI transaction and wait for it to complete */
610 void gsi_trans_commit_wait(struct gsi_trans *trans)
611 {
612 	if (!trans->used_count)
613 		goto out_trans_free;
614 
615 	refcount_inc(&trans->refcount);
616 
617 	__gsi_trans_commit(trans, true);
618 
619 	wait_for_completion(&trans->completion);
620 
621 out_trans_free:
622 	gsi_trans_free(trans);
623 }
624 
625 /* Process the completion of a transaction; called while polling */
626 void gsi_trans_complete(struct gsi_trans *trans)
627 {
628 	/* If the entire SGL was mapped when added, unmap it now */
629 	if (trans->direction != DMA_NONE)
630 		dma_unmap_sg(trans->gsi->dev, trans->sgl, trans->used_count,
631 			     trans->direction);
632 
633 	ipa_gsi_trans_complete(trans);
634 
635 	complete(&trans->completion);
636 
637 	gsi_trans_free(trans);
638 }
639 
640 /* Cancel a channel's pending transactions */
641 void gsi_channel_trans_cancel_pending(struct gsi_channel *channel)
642 {
643 	struct gsi_trans_info *trans_info = &channel->trans_info;
644 	u16 trans_id = trans_info->pending_id;
645 
646 	/* channel->gsi->mutex is held by caller */
647 
648 	/* If there are no pending transactions, we're done */
649 	if (trans_id == trans_info->committed_id)
650 		return;
651 
652 	/* Mark all pending transactions cancelled */
653 	do {
654 		struct gsi_trans *trans;
655 
656 		trans = &trans_info->trans[trans_id % channel->tre_count];
657 		trans->cancelled = true;
658 	} while (++trans_id != trans_info->committed_id);
659 
660 	/* All pending transactions are now completed */
661 	trans_info->pending_id = trans_info->committed_id;
662 
663 	/* Schedule NAPI polling to complete the cancelled transactions */
664 	napi_schedule(&channel->napi);
665 }
666 
667 /* Issue a command to read a single byte from a channel */
668 int gsi_trans_read_byte(struct gsi *gsi, u32 channel_id, dma_addr_t addr)
669 {
670 	struct gsi_channel *channel = &gsi->channel[channel_id];
671 	struct gsi_ring *tre_ring = &channel->tre_ring;
672 	struct gsi_trans_info *trans_info;
673 	struct gsi_tre *dest_tre;
674 
675 	trans_info = &channel->trans_info;
676 
677 	/* First reserve the TRE, if possible */
678 	if (!gsi_trans_tre_reserve(trans_info, 1))
679 		return -EBUSY;
680 
681 	/* Now fill the reserved TRE and tell the hardware */
682 
683 	dest_tre = gsi_ring_virt(tre_ring, tre_ring->index);
684 	gsi_trans_tre_fill(dest_tre, addr, 1, true, false, IPA_CMD_NONE);
685 
686 	tre_ring->index++;
687 	gsi_channel_doorbell(channel);
688 
689 	return 0;
690 }
691 
692 /* Mark a gsi_trans_read_byte() request done */
693 void gsi_trans_read_byte_done(struct gsi *gsi, u32 channel_id)
694 {
695 	struct gsi_channel *channel = &gsi->channel[channel_id];
696 
697 	gsi_trans_tre_release(&channel->trans_info, 1);
698 }
699 
700 /* Initialize a channel's GSI transaction info */
701 int gsi_channel_trans_init(struct gsi *gsi, u32 channel_id)
702 {
703 	struct gsi_channel *channel = &gsi->channel[channel_id];
704 	u32 tre_count = channel->tre_count;
705 	struct gsi_trans_info *trans_info;
706 	u32 tre_max;
707 	int ret;
708 
709 	/* Ensure the size of a channel element is what's expected */
710 	BUILD_BUG_ON(sizeof(struct gsi_tre) != GSI_RING_ELEMENT_SIZE);
711 
712 	trans_info = &channel->trans_info;
713 
714 	/* The tre_avail field is what ultimately limits the number of
715 	 * outstanding transactions and their resources.  A transaction
716 	 * allocation succeeds only if the TREs available are sufficient
717 	 * for what the transaction might need.
718 	 */
719 	tre_max = gsi_channel_tre_max(channel->gsi, channel_id);
720 	atomic_set(&trans_info->tre_avail, tre_max);
721 
722 	/* We can't use more TREs than the number available in the ring.
723 	 * This limits the number of transactions that can be outstanding.
724 	 * Worst case is one TRE per transaction (but we actually limit
725 	 * it to something a little less than that).  By allocating a
726 	 * power-of-two number of transactions we can use an index
727 	 * modulo that number to determine the next one that's free.
728 	 * Transactions are allocated one at a time.
729 	 */
730 	trans_info->trans = kcalloc(tre_count, sizeof(*trans_info->trans),
731 				    GFP_KERNEL);
732 	if (!trans_info->trans)
733 		return -ENOMEM;
734 	trans_info->free_id = 0;	/* all modulo channel->tre_count */
735 	trans_info->allocated_id = 0;
736 	trans_info->committed_id = 0;
737 	trans_info->pending_id = 0;
738 	trans_info->completed_id = 0;
739 	trans_info->polled_id = 0;
740 
741 	/* A completion event contains a pointer to the TRE that caused
742 	 * the event (which will be the last one used by the transaction).
743 	 * Each entry in this map records the transaction associated
744 	 * with a corresponding completed TRE.
745 	 */
746 	trans_info->map = kcalloc(tre_count, sizeof(*trans_info->map),
747 				  GFP_KERNEL);
748 	if (!trans_info->map) {
749 		ret = -ENOMEM;
750 		goto err_trans_free;
751 	}
752 
753 	/* A transaction uses a scatterlist array to represent the data
754 	 * transfers implemented by the transaction.  Each scatterlist
755 	 * element is used to fill a single TRE when the transaction is
756 	 * committed.  So we need as many scatterlist elements as the
757 	 * maximum number of TREs that can be outstanding.
758 	 */
759 	ret = gsi_trans_pool_init(&trans_info->sg_pool,
760 				  sizeof(struct scatterlist),
761 				  tre_max, channel->trans_tre_max);
762 	if (ret)
763 		goto err_map_free;
764 
765 
766 	return 0;
767 
768 err_map_free:
769 	kfree(trans_info->map);
770 err_trans_free:
771 	kfree(trans_info->trans);
772 
773 	dev_err(gsi->dev, "error %d initializing channel %u transactions\n",
774 		ret, channel_id);
775 
776 	return ret;
777 }
778 
779 /* Inverse of gsi_channel_trans_init() */
780 void gsi_channel_trans_exit(struct gsi_channel *channel)
781 {
782 	struct gsi_trans_info *trans_info = &channel->trans_info;
783 
784 	gsi_trans_pool_exit(&trans_info->sg_pool);
785 	kfree(trans_info->trans);
786 	kfree(trans_info->map);
787 }
788