xref: /openbmc/linux/drivers/net/ipa/gsi_trans.c (revision fbb6b31a)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 /* Copyright (c) 2012-2018, The Linux Foundation. All rights reserved.
4  * Copyright (C) 2019-2020 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 commands in a single structure.
26  * (A "command" 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 commands
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 a command (or set of them), a user of the GSI transaction
34  * interface allocates a transaction, indicating the number of TREs required
35  * (one per command).  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
38  * as early as possible.  All resources required to complete a transaction
39  * are allocated at transaction allocation time.
40  *
41  * Commands performed as part of a transaction are represented in an array
42  * of Linux scatterlist structures.  This array is allocated with the
43  * transaction, and its entries are initialized using standard scatterlist
44  * functions (such as sg_set_buf() or skb_to_sgvec()).
45  *
46  * Once a transaction's scatterlist structures have been initialized, the
47  * transaction is committed.  The caller is responsible for mapping buffers
48  * for DMA if necessary, and this should be done *before* allocating
49  * the transaction.  Between a successful allocation and commit of a
50  * transaction no errors should occur.
51  *
52  * Committing transfers ownership of the entire transaction to the GSI
53  * transaction core.  The GSI transaction code formats the content of
54  * the scatterlist array into the channel ring buffer and informs the
55  * hardware that new TREs are available to process.
56  *
57  * The last TRE in each transaction is marked to interrupt the AP when the
58  * GSI hardware has completed it.  Because transfers described by TREs are
59  * performed strictly in order, signaling the completion of just the last
60  * TRE in the transaction is sufficient to indicate the full transaction
61  * is complete.
62  *
63  * When a transaction is complete, ipa_gsi_trans_complete() is called by the
64  * GSI code into the IPA layer, allowing it to perform any final cleanup
65  * required before the transaction is freed.
66  */
67 
68 /* Hardware values representing a transfer element type */
69 enum gsi_tre_type {
70 	GSI_RE_XFER	= 0x2,
71 	GSI_RE_IMMD_CMD	= 0x3,
72 };
73 
74 /* An entry in a channel ring */
75 struct gsi_tre {
76 	__le64 addr;		/* DMA address */
77 	__le16 len_opcode;	/* length in bytes or enum IPA_CMD_* */
78 	__le16 reserved;
79 	__le32 flags;		/* TRE_FLAGS_* */
80 };
81 
82 /* gsi_tre->flags mask values (in CPU byte order) */
83 #define TRE_FLAGS_CHAIN_FMASK	GENMASK(0, 0)
84 #define TRE_FLAGS_IEOT_FMASK	GENMASK(9, 9)
85 #define TRE_FLAGS_BEI_FMASK	GENMASK(10, 10)
86 #define TRE_FLAGS_TYPE_FMASK	GENMASK(23, 16)
87 
88 int gsi_trans_pool_init(struct gsi_trans_pool *pool, size_t size, u32 count,
89 			u32 max_alloc)
90 {
91 	void *virt;
92 
93 	if (!size)
94 		return -EINVAL;
95 	if (count < max_alloc)
96 		return -EINVAL;
97 	if (!max_alloc)
98 		return -EINVAL;
99 
100 	/* By allocating a few extra entries in our pool (one less
101 	 * than the maximum number that will be requested in a
102 	 * single allocation), we can always satisfy requests without
103 	 * ever worrying about straddling the end of the pool array.
104 	 * If there aren't enough entries starting at the free index,
105 	 * we just allocate free entries from the beginning of the pool.
106 	 */
107 	virt = kcalloc(count + max_alloc - 1, size, GFP_KERNEL);
108 	if (!virt)
109 		return -ENOMEM;
110 
111 	pool->base = virt;
112 	/* If the allocator gave us any extra memory, use it */
113 	pool->count = ksize(pool->base) / size;
114 	pool->free = 0;
115 	pool->max_alloc = max_alloc;
116 	pool->size = size;
117 	pool->addr = 0;		/* Only used for DMA pools */
118 
119 	return 0;
120 }
121 
122 void gsi_trans_pool_exit(struct gsi_trans_pool *pool)
123 {
124 	kfree(pool->base);
125 	memset(pool, 0, sizeof(*pool));
126 }
127 
128 /* Allocate the requested number of (zeroed) entries from the pool */
129 /* Home-grown DMA pool.  This way we can preallocate and use the tre_count
130  * to guarantee allocations will succeed.  Even though we specify max_alloc
131  * (and it can be more than one), we only allow allocation of a single
132  * element from a DMA pool.
133  */
134 int gsi_trans_pool_init_dma(struct device *dev, struct gsi_trans_pool *pool,
135 			    size_t size, u32 count, u32 max_alloc)
136 {
137 	size_t total_size;
138 	dma_addr_t addr;
139 	void *virt;
140 
141 	if (!size)
142 		return -EINVAL;
143 	if (count < max_alloc)
144 		return -EINVAL;
145 	if (!max_alloc)
146 		return -EINVAL;
147 
148 	/* Don't let allocations cross a power-of-two boundary */
149 	size = __roundup_pow_of_two(size);
150 	total_size = (count + max_alloc - 1) * size;
151 
152 	/* The allocator will give us a power-of-2 number of pages
153 	 * sufficient to satisfy our request.  Round up our requested
154 	 * size to avoid any unused space in the allocation.  This way
155 	 * gsi_trans_pool_exit_dma() can assume the total allocated
156 	 * size is exactly (count * size).
157 	 */
158 	total_size = get_order(total_size) << PAGE_SHIFT;
159 
160 	virt = dma_alloc_coherent(dev, total_size, &addr, GFP_KERNEL);
161 	if (!virt)
162 		return -ENOMEM;
163 
164 	pool->base = virt;
165 	pool->count = total_size / size;
166 	pool->free = 0;
167 	pool->size = size;
168 	pool->max_alloc = max_alloc;
169 	pool->addr = addr;
170 
171 	return 0;
172 }
173 
174 void gsi_trans_pool_exit_dma(struct device *dev, struct gsi_trans_pool *pool)
175 {
176 	size_t total_size = pool->count * pool->size;
177 
178 	dma_free_coherent(dev, total_size, pool->base, pool->addr);
179 	memset(pool, 0, sizeof(*pool));
180 }
181 
182 /* Return the byte offset of the next free entry in the pool */
183 static u32 gsi_trans_pool_alloc_common(struct gsi_trans_pool *pool, u32 count)
184 {
185 	u32 offset;
186 
187 	WARN_ON(!count);
188 	WARN_ON(count > pool->max_alloc);
189 
190 	/* Allocate from beginning if wrap would occur */
191 	if (count > pool->count - pool->free)
192 		pool->free = 0;
193 
194 	offset = pool->free * pool->size;
195 	pool->free += count;
196 	memset(pool->base + offset, 0, count * pool->size);
197 
198 	return offset;
199 }
200 
201 /* Allocate a contiguous block of zeroed entries from a pool */
202 void *gsi_trans_pool_alloc(struct gsi_trans_pool *pool, u32 count)
203 {
204 	return pool->base + gsi_trans_pool_alloc_common(pool, count);
205 }
206 
207 /* Allocate a single zeroed entry from a DMA pool */
208 void *gsi_trans_pool_alloc_dma(struct gsi_trans_pool *pool, dma_addr_t *addr)
209 {
210 	u32 offset = gsi_trans_pool_alloc_common(pool, 1);
211 
212 	*addr = pool->addr + offset;
213 
214 	return pool->base + offset;
215 }
216 
217 /* Return the pool element that immediately follows the one given.
218  * This only works done if elements are allocated one at a time.
219  */
220 void *gsi_trans_pool_next(struct gsi_trans_pool *pool, void *element)
221 {
222 	void *end = pool->base + pool->count * pool->size;
223 
224 	WARN_ON(element < pool->base);
225 	WARN_ON(element >= end);
226 	WARN_ON(pool->max_alloc != 1);
227 
228 	element += pool->size;
229 
230 	return element < end ? element : pool->base;
231 }
232 
233 /* Map a given ring entry index to the transaction associated with it */
234 static void gsi_channel_trans_map(struct gsi_channel *channel, u32 index,
235 				  struct gsi_trans *trans)
236 {
237 	/* Note: index *must* be used modulo the ring count here */
238 	channel->trans_info.map[index % channel->tre_ring.count] = trans;
239 }
240 
241 /* Return the transaction mapped to a given ring entry */
242 struct gsi_trans *
243 gsi_channel_trans_mapped(struct gsi_channel *channel, u32 index)
244 {
245 	/* Note: index *must* be used modulo the ring count here */
246 	return channel->trans_info.map[index % channel->tre_ring.count];
247 }
248 
249 /* Return the oldest completed transaction for a channel (or null) */
250 struct gsi_trans *gsi_channel_trans_complete(struct gsi_channel *channel)
251 {
252 	return list_first_entry_or_null(&channel->trans_info.complete,
253 					struct gsi_trans, links);
254 }
255 
256 /* Move a transaction from the allocated list to the pending list */
257 static void gsi_trans_move_pending(struct gsi_trans *trans)
258 {
259 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
260 	struct gsi_trans_info *trans_info = &channel->trans_info;
261 
262 	spin_lock_bh(&trans_info->spinlock);
263 
264 	list_move_tail(&trans->links, &trans_info->pending);
265 
266 	spin_unlock_bh(&trans_info->spinlock);
267 }
268 
269 /* Move a transaction and all of its predecessors from the pending list
270  * to the completed list.
271  */
272 void gsi_trans_move_complete(struct gsi_trans *trans)
273 {
274 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
275 	struct gsi_trans_info *trans_info = &channel->trans_info;
276 	struct list_head list;
277 
278 	spin_lock_bh(&trans_info->spinlock);
279 
280 	/* Move this transaction and all predecessors to completed list */
281 	list_cut_position(&list, &trans_info->pending, &trans->links);
282 	list_splice_tail(&list, &trans_info->complete);
283 
284 	spin_unlock_bh(&trans_info->spinlock);
285 }
286 
287 /* Move a transaction from the completed list to the polled list */
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 	spin_lock_bh(&trans_info->spinlock);
294 
295 	list_move_tail(&trans->links, &trans_info->polled);
296 
297 	spin_unlock_bh(&trans_info->spinlock);
298 }
299 
300 /* Reserve some number of TREs on a channel.  Returns true if successful */
301 static bool
302 gsi_trans_tre_reserve(struct gsi_trans_info *trans_info, u32 tre_count)
303 {
304 	int avail = atomic_read(&trans_info->tre_avail);
305 	int new;
306 
307 	do {
308 		new = avail - (int)tre_count;
309 		if (unlikely(new < 0))
310 			return false;
311 	} while (!atomic_try_cmpxchg(&trans_info->tre_avail, &avail, new));
312 
313 	return true;
314 }
315 
316 /* Release previously-reserved TRE entries to a channel */
317 static void
318 gsi_trans_tre_release(struct gsi_trans_info *trans_info, u32 tre_count)
319 {
320 	atomic_add(tre_count, &trans_info->tre_avail);
321 }
322 
323 /* Return true if no transactions are allocated, false otherwise */
324 bool gsi_channel_trans_idle(struct gsi *gsi, u32 channel_id)
325 {
326 	u32 tre_max = gsi_channel_tre_max(gsi, channel_id);
327 	struct gsi_trans_info *trans_info;
328 
329 	trans_info = &gsi->channel[channel_id].trans_info;
330 
331 	return atomic_read(&trans_info->tre_avail) == tre_max;
332 }
333 
334 /* Allocate a GSI transaction on a channel */
335 struct gsi_trans *gsi_channel_trans_alloc(struct gsi *gsi, u32 channel_id,
336 					  u32 tre_count,
337 					  enum dma_data_direction direction)
338 {
339 	struct gsi_channel *channel = &gsi->channel[channel_id];
340 	struct gsi_trans_info *trans_info;
341 	struct gsi_trans *trans;
342 
343 	if (WARN_ON(tre_count > gsi_channel_trans_tre_max(gsi, channel_id)))
344 		return NULL;
345 
346 	trans_info = &channel->trans_info;
347 
348 	/* We reserve the TREs now, but consume them at commit time.
349 	 * If there aren't enough available, we're done.
350 	 */
351 	if (!gsi_trans_tre_reserve(trans_info, tre_count))
352 		return NULL;
353 
354 	/* Allocate and initialize non-zero fields in the the transaction */
355 	trans = gsi_trans_pool_alloc(&trans_info->pool, 1);
356 	trans->gsi = gsi;
357 	trans->channel_id = channel_id;
358 	trans->tre_count = tre_count;
359 	init_completion(&trans->completion);
360 
361 	/* Allocate the scatterlist and (if requested) info entries. */
362 	trans->sgl = gsi_trans_pool_alloc(&trans_info->sg_pool, tre_count);
363 	sg_init_marker(trans->sgl, tre_count);
364 
365 	trans->direction = direction;
366 
367 	spin_lock_bh(&trans_info->spinlock);
368 
369 	list_add_tail(&trans->links, &trans_info->alloc);
370 
371 	spin_unlock_bh(&trans_info->spinlock);
372 
373 	refcount_set(&trans->refcount, 1);
374 
375 	return trans;
376 }
377 
378 /* Free a previously-allocated transaction */
379 void gsi_trans_free(struct gsi_trans *trans)
380 {
381 	refcount_t *refcount = &trans->refcount;
382 	struct gsi_trans_info *trans_info;
383 	bool last;
384 
385 	/* We must hold the lock to release the last reference */
386 	if (refcount_dec_not_one(refcount))
387 		return;
388 
389 	trans_info = &trans->gsi->channel[trans->channel_id].trans_info;
390 
391 	spin_lock_bh(&trans_info->spinlock);
392 
393 	/* Reference might have been added before we got the lock */
394 	last = refcount_dec_and_test(refcount);
395 	if (last)
396 		list_del(&trans->links);
397 
398 	spin_unlock_bh(&trans_info->spinlock);
399 
400 	if (!last)
401 		return;
402 
403 	ipa_gsi_trans_release(trans);
404 
405 	/* Releasing the reserved TREs implicitly frees the sgl[] and
406 	 * (if present) info[] arrays, plus the transaction itself.
407 	 */
408 	gsi_trans_tre_release(trans_info, trans->tre_count);
409 }
410 
411 /* Add an immediate command to a transaction */
412 void gsi_trans_cmd_add(struct gsi_trans *trans, void *buf, u32 size,
413 		       dma_addr_t addr, enum dma_data_direction direction,
414 		       enum ipa_cmd_opcode opcode)
415 {
416 	struct ipa_cmd_info *info;
417 	u32 which = trans->used++;
418 	struct scatterlist *sg;
419 
420 	WARN_ON(which >= trans->tre_count);
421 
422 	/* Commands are quite different from data transfer requests.
423 	 * Their payloads come from a pool whose memory is allocated
424 	 * using dma_alloc_coherent().  We therefore do *not* map them
425 	 * for DMA (unlike what we do for pages and skbs).
426 	 *
427 	 * When a transaction completes, the SGL is normally unmapped.
428 	 * A command transaction has direction DMA_NONE, which tells
429 	 * gsi_trans_complete() to skip the unmapping step.
430 	 *
431 	 * The only things we use directly in a command scatter/gather
432 	 * entry are the DMA address and length.  We still need the SG
433 	 * table flags to be maintained though, so assign a NULL page
434 	 * pointer for that purpose.
435 	 */
436 	sg = &trans->sgl[which];
437 	sg_assign_page(sg, NULL);
438 	sg_dma_address(sg) = addr;
439 	sg_dma_len(sg) = size;
440 
441 	info = &trans->info[which];
442 	info->opcode = opcode;
443 	info->direction = direction;
444 }
445 
446 /* Add a page transfer to a transaction.  It will fill the only TRE. */
447 int gsi_trans_page_add(struct gsi_trans *trans, struct page *page, u32 size,
448 		       u32 offset)
449 {
450 	struct scatterlist *sg = &trans->sgl[0];
451 	int ret;
452 
453 	if (WARN_ON(trans->tre_count != 1))
454 		return -EINVAL;
455 	if (WARN_ON(trans->used))
456 		return -EINVAL;
457 
458 	sg_set_page(sg, page, size, offset);
459 	ret = dma_map_sg(trans->gsi->dev, sg, 1, trans->direction);
460 	if (!ret)
461 		return -ENOMEM;
462 
463 	trans->used++;	/* Transaction now owns the (DMA mapped) page */
464 
465 	return 0;
466 }
467 
468 /* Add an SKB transfer to a transaction.  No other TREs will be used. */
469 int gsi_trans_skb_add(struct gsi_trans *trans, struct sk_buff *skb)
470 {
471 	struct scatterlist *sg = &trans->sgl[0];
472 	u32 used;
473 	int ret;
474 
475 	if (WARN_ON(trans->tre_count != 1))
476 		return -EINVAL;
477 	if (WARN_ON(trans->used))
478 		return -EINVAL;
479 
480 	/* skb->len will not be 0 (checked early) */
481 	ret = skb_to_sgvec(skb, sg, 0, skb->len);
482 	if (ret < 0)
483 		return ret;
484 	used = ret;
485 
486 	ret = dma_map_sg(trans->gsi->dev, sg, used, trans->direction);
487 	if (!ret)
488 		return -ENOMEM;
489 
490 	trans->used += used;	/* Transaction now owns the (DMA mapped) skb */
491 
492 	return 0;
493 }
494 
495 /* Compute the length/opcode value to use for a TRE */
496 static __le16 gsi_tre_len_opcode(enum ipa_cmd_opcode opcode, u32 len)
497 {
498 	return opcode == IPA_CMD_NONE ? cpu_to_le16((u16)len)
499 				      : cpu_to_le16((u16)opcode);
500 }
501 
502 /* Compute the flags value to use for a given TRE */
503 static __le32 gsi_tre_flags(bool last_tre, bool bei, enum ipa_cmd_opcode opcode)
504 {
505 	enum gsi_tre_type tre_type;
506 	u32 tre_flags;
507 
508 	tre_type = opcode == IPA_CMD_NONE ? GSI_RE_XFER : GSI_RE_IMMD_CMD;
509 	tre_flags = u32_encode_bits(tre_type, TRE_FLAGS_TYPE_FMASK);
510 
511 	/* Last TRE contains interrupt flags */
512 	if (last_tre) {
513 		/* All transactions end in a transfer completion interrupt */
514 		tre_flags |= TRE_FLAGS_IEOT_FMASK;
515 		/* Don't interrupt when outbound commands are acknowledged */
516 		if (bei)
517 			tre_flags |= TRE_FLAGS_BEI_FMASK;
518 	} else {	/* All others indicate there's more to come */
519 		tre_flags |= TRE_FLAGS_CHAIN_FMASK;
520 	}
521 
522 	return cpu_to_le32(tre_flags);
523 }
524 
525 static void gsi_trans_tre_fill(struct gsi_tre *dest_tre, dma_addr_t addr,
526 			       u32 len, bool last_tre, bool bei,
527 			       enum ipa_cmd_opcode opcode)
528 {
529 	struct gsi_tre tre;
530 
531 	tre.addr = cpu_to_le64(addr);
532 	tre.len_opcode = gsi_tre_len_opcode(opcode, len);
533 	tre.reserved = 0;
534 	tre.flags = gsi_tre_flags(last_tre, bei, opcode);
535 
536 	/* ARM64 can write 16 bytes as a unit with a single instruction.
537 	 * Doing the assignment this way is an attempt to make that happen.
538 	 */
539 	*dest_tre = tre;
540 }
541 
542 /**
543  * __gsi_trans_commit() - Common GSI transaction commit code
544  * @trans:	Transaction to commit
545  * @ring_db:	Whether to tell the hardware about these queued transfers
546  *
547  * Formats channel ring TRE entries based on the content of the scatterlist.
548  * Maps a transaction pointer to the last ring entry used for the transaction,
549  * so it can be recovered when it completes.  Moves the transaction to the
550  * pending list.  Finally, updates the channel ring pointer and optionally
551  * rings the doorbell.
552  */
553 static void __gsi_trans_commit(struct gsi_trans *trans, bool ring_db)
554 {
555 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
556 	struct gsi_ring *ring = &channel->tre_ring;
557 	enum ipa_cmd_opcode opcode = IPA_CMD_NONE;
558 	bool bei = channel->toward_ipa;
559 	struct ipa_cmd_info *info;
560 	struct gsi_tre *dest_tre;
561 	struct scatterlist *sg;
562 	u32 byte_count = 0;
563 	u32 avail;
564 	u32 i;
565 
566 	WARN_ON(!trans->used);
567 
568 	/* Consume the entries.  If we cross the end of the ring while
569 	 * filling them we'll switch to the beginning to finish.
570 	 * If there is no info array we're doing a simple data
571 	 * transfer request, whose opcode is IPA_CMD_NONE.
572 	 */
573 	info = trans->info ? &trans->info[0] : NULL;
574 	avail = ring->count - ring->index % ring->count;
575 	dest_tre = gsi_ring_virt(ring, ring->index);
576 	for_each_sg(trans->sgl, sg, trans->used, i) {
577 		bool last_tre = i == trans->used - 1;
578 		dma_addr_t addr = sg_dma_address(sg);
579 		u32 len = sg_dma_len(sg);
580 
581 		byte_count += len;
582 		if (!avail--)
583 			dest_tre = gsi_ring_virt(ring, 0);
584 		if (info)
585 			opcode = info++->opcode;
586 
587 		gsi_trans_tre_fill(dest_tre, addr, len, last_tre, bei, opcode);
588 		dest_tre++;
589 	}
590 	ring->index += trans->used;
591 
592 	if (channel->toward_ipa) {
593 		/* We record TX bytes when they are sent */
594 		trans->len = byte_count;
595 		trans->trans_count = channel->trans_count;
596 		trans->byte_count = channel->byte_count;
597 		channel->trans_count++;
598 		channel->byte_count += byte_count;
599 	}
600 
601 	/* Associate the last TRE with the transaction */
602 	gsi_channel_trans_map(channel, ring->index - 1, trans);
603 
604 	gsi_trans_move_pending(trans);
605 
606 	/* Ring doorbell if requested, or if all TREs are allocated */
607 	if (ring_db || !atomic_read(&channel->trans_info.tre_avail)) {
608 		/* Report what we're handing off to hardware for TX channels */
609 		if (channel->toward_ipa)
610 			gsi_channel_tx_queued(channel);
611 		gsi_channel_doorbell(channel);
612 	}
613 }
614 
615 /* Commit a GSI transaction */
616 void gsi_trans_commit(struct gsi_trans *trans, bool ring_db)
617 {
618 	if (trans->used)
619 		__gsi_trans_commit(trans, ring_db);
620 	else
621 		gsi_trans_free(trans);
622 }
623 
624 /* Commit a GSI transaction and wait for it to complete */
625 void gsi_trans_commit_wait(struct gsi_trans *trans)
626 {
627 	if (!trans->used)
628 		goto out_trans_free;
629 
630 	refcount_inc(&trans->refcount);
631 
632 	__gsi_trans_commit(trans, true);
633 
634 	wait_for_completion(&trans->completion);
635 
636 out_trans_free:
637 	gsi_trans_free(trans);
638 }
639 
640 /* Commit a GSI transaction and wait for it to complete, with timeout */
641 int gsi_trans_commit_wait_timeout(struct gsi_trans *trans,
642 				  unsigned long timeout)
643 {
644 	unsigned long timeout_jiffies = msecs_to_jiffies(timeout);
645 	unsigned long remaining = 1;	/* In case of empty transaction */
646 
647 	if (!trans->used)
648 		goto out_trans_free;
649 
650 	refcount_inc(&trans->refcount);
651 
652 	__gsi_trans_commit(trans, true);
653 
654 	remaining = wait_for_completion_timeout(&trans->completion,
655 						timeout_jiffies);
656 out_trans_free:
657 	gsi_trans_free(trans);
658 
659 	return remaining ? 0 : -ETIMEDOUT;
660 }
661 
662 /* Process the completion of a transaction; called while polling */
663 void gsi_trans_complete(struct gsi_trans *trans)
664 {
665 	/* If the entire SGL was mapped when added, unmap it now */
666 	if (trans->direction != DMA_NONE)
667 		dma_unmap_sg(trans->gsi->dev, trans->sgl, trans->used,
668 			     trans->direction);
669 
670 	ipa_gsi_trans_complete(trans);
671 
672 	complete(&trans->completion);
673 
674 	gsi_trans_free(trans);
675 }
676 
677 /* Cancel a channel's pending transactions */
678 void gsi_channel_trans_cancel_pending(struct gsi_channel *channel)
679 {
680 	struct gsi_trans_info *trans_info = &channel->trans_info;
681 	struct gsi_trans *trans;
682 	bool cancelled;
683 
684 	/* channel->gsi->mutex is held by caller */
685 	spin_lock_bh(&trans_info->spinlock);
686 
687 	cancelled = !list_empty(&trans_info->pending);
688 	list_for_each_entry(trans, &trans_info->pending, links)
689 		trans->cancelled = true;
690 
691 	list_splice_tail_init(&trans_info->pending, &trans_info->complete);
692 
693 	spin_unlock_bh(&trans_info->spinlock);
694 
695 	/* Schedule NAPI polling to complete the cancelled transactions */
696 	if (cancelled)
697 		napi_schedule(&channel->napi);
698 }
699 
700 /* Issue a command to read a single byte from a channel */
701 int gsi_trans_read_byte(struct gsi *gsi, u32 channel_id, dma_addr_t addr)
702 {
703 	struct gsi_channel *channel = &gsi->channel[channel_id];
704 	struct gsi_ring *ring = &channel->tre_ring;
705 	struct gsi_trans_info *trans_info;
706 	struct gsi_tre *dest_tre;
707 
708 	trans_info = &channel->trans_info;
709 
710 	/* First reserve the TRE, if possible */
711 	if (!gsi_trans_tre_reserve(trans_info, 1))
712 		return -EBUSY;
713 
714 	/* Now fill the the reserved TRE and tell the hardware */
715 
716 	dest_tre = gsi_ring_virt(ring, ring->index);
717 	gsi_trans_tre_fill(dest_tre, addr, 1, true, false, IPA_CMD_NONE);
718 
719 	ring->index++;
720 	gsi_channel_doorbell(channel);
721 
722 	return 0;
723 }
724 
725 /* Mark a gsi_trans_read_byte() request done */
726 void gsi_trans_read_byte_done(struct gsi *gsi, u32 channel_id)
727 {
728 	struct gsi_channel *channel = &gsi->channel[channel_id];
729 
730 	gsi_trans_tre_release(&channel->trans_info, 1);
731 }
732 
733 /* Initialize a channel's GSI transaction info */
734 int gsi_channel_trans_init(struct gsi *gsi, u32 channel_id)
735 {
736 	struct gsi_channel *channel = &gsi->channel[channel_id];
737 	struct gsi_trans_info *trans_info;
738 	u32 tre_max;
739 	int ret;
740 
741 	/* Ensure the size of a channel element is what's expected */
742 	BUILD_BUG_ON(sizeof(struct gsi_tre) != GSI_RING_ELEMENT_SIZE);
743 
744 	/* The map array is used to determine what transaction is associated
745 	 * with a TRE that the hardware reports has completed.  We need one
746 	 * map entry per TRE.
747 	 */
748 	trans_info = &channel->trans_info;
749 	trans_info->map = kcalloc(channel->tre_count, sizeof(*trans_info->map),
750 				  GFP_KERNEL);
751 	if (!trans_info->map)
752 		return -ENOMEM;
753 
754 	/* We can't use more TREs than there are available in the ring.
755 	 * This limits the number of transactions that can be oustanding.
756 	 * Worst case is one TRE per transaction (but we actually limit
757 	 * it to something a little less than that).  We allocate resources
758 	 * for transactions (including transaction structures) based on
759 	 * this maximum number.
760 	 */
761 	tre_max = gsi_channel_tre_max(channel->gsi, channel_id);
762 
763 	/* Transactions are allocated one at a time. */
764 	ret = gsi_trans_pool_init(&trans_info->pool, sizeof(struct gsi_trans),
765 				  tre_max, 1);
766 	if (ret)
767 		goto err_kfree;
768 
769 	/* A transaction uses a scatterlist array to represent the data
770 	 * transfers implemented by the transaction.  Each scatterlist
771 	 * element is used to fill a single TRE when the transaction is
772 	 * committed.  So we need as many scatterlist elements as the
773 	 * maximum number of TREs that can be outstanding.
774 	 *
775 	 * All TREs in a transaction must fit within the channel's TLV FIFO.
776 	 * A transaction on a channel can allocate as many TREs as that but
777 	 * no more.
778 	 */
779 	ret = gsi_trans_pool_init(&trans_info->sg_pool,
780 				  sizeof(struct scatterlist),
781 				  tre_max, channel->tlv_count);
782 	if (ret)
783 		goto err_trans_pool_exit;
784 
785 	/* Finally, the tre_avail field is what ultimately limits the number
786 	 * of outstanding transactions and their resources.  A transaction
787 	 * allocation succeeds only if the TREs available are sufficient for
788 	 * what the transaction might need.  Transaction resource pools are
789 	 * sized based on the maximum number of outstanding TREs, so there
790 	 * will always be resources available if there are TREs available.
791 	 */
792 	atomic_set(&trans_info->tre_avail, tre_max);
793 
794 	spin_lock_init(&trans_info->spinlock);
795 	INIT_LIST_HEAD(&trans_info->alloc);
796 	INIT_LIST_HEAD(&trans_info->pending);
797 	INIT_LIST_HEAD(&trans_info->complete);
798 	INIT_LIST_HEAD(&trans_info->polled);
799 
800 	return 0;
801 
802 err_trans_pool_exit:
803 	gsi_trans_pool_exit(&trans_info->pool);
804 err_kfree:
805 	kfree(trans_info->map);
806 
807 	dev_err(gsi->dev, "error %d initializing channel %u transactions\n",
808 		ret, channel_id);
809 
810 	return ret;
811 }
812 
813 /* Inverse of gsi_channel_trans_init() */
814 void gsi_channel_trans_exit(struct gsi_channel *channel)
815 {
816 	struct gsi_trans_info *trans_info = &channel->trans_info;
817 
818 	gsi_trans_pool_exit(&trans_info->sg_pool);
819 	gsi_trans_pool_exit(&trans_info->pool);
820 	kfree(trans_info->map);
821 }
822