xref: /openbmc/linux/drivers/net/ipa/gsi_trans.c (revision 278d3ba6)
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 /* Map a TRE ring entry index to the transaction it is associated with */
218 static void gsi_trans_map(struct gsi_trans *trans, u32 index)
219 {
220 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
221 
222 	/* The completion event will indicate the last TRE used */
223 	index += trans->used_count - 1;
224 
225 	/* Note: index *must* be used modulo the ring count here */
226 	channel->trans_info.map[index % channel->tre_ring.count] = trans;
227 }
228 
229 /* Return the transaction mapped to a given ring entry */
230 struct gsi_trans *
231 gsi_channel_trans_mapped(struct gsi_channel *channel, u32 index)
232 {
233 	/* Note: index *must* be used modulo the ring count here */
234 	return channel->trans_info.map[index % channel->tre_ring.count];
235 }
236 
237 /* Return the oldest completed transaction for a channel (or null) */
238 struct gsi_trans *gsi_channel_trans_complete(struct gsi_channel *channel)
239 {
240 	return list_first_entry_or_null(&channel->trans_info.complete,
241 					struct gsi_trans, links);
242 }
243 
244 /* Move a transaction from the allocated list to the committed list */
245 static void gsi_trans_move_committed(struct gsi_trans *trans)
246 {
247 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
248 	struct gsi_trans_info *trans_info = &channel->trans_info;
249 
250 	spin_lock_bh(&trans_info->spinlock);
251 
252 	list_move_tail(&trans->links, &trans_info->committed);
253 
254 	spin_unlock_bh(&trans_info->spinlock);
255 }
256 
257 /* Move transactions from the committed list to the pending list */
258 static void gsi_trans_move_pending(struct gsi_trans *trans)
259 {
260 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
261 	struct gsi_trans_info *trans_info = &channel->trans_info;
262 	struct list_head list;
263 
264 	spin_lock_bh(&trans_info->spinlock);
265 
266 	/* Move this transaction and all predecessors to the pending list */
267 	list_cut_position(&list, &trans_info->committed, &trans->links);
268 	list_splice_tail(&list, &trans_info->pending);
269 
270 	spin_unlock_bh(&trans_info->spinlock);
271 }
272 
273 /* Move a transaction and all of its predecessors from the pending list
274  * to the completed list.
275  */
276 void gsi_trans_move_complete(struct gsi_trans *trans)
277 {
278 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
279 	struct gsi_trans_info *trans_info = &channel->trans_info;
280 	struct list_head list;
281 
282 	spin_lock_bh(&trans_info->spinlock);
283 
284 	/* Move this transaction and all predecessors to completed list */
285 	list_cut_position(&list, &trans_info->pending, &trans->links);
286 	list_splice_tail(&list, &trans_info->complete);
287 
288 	spin_unlock_bh(&trans_info->spinlock);
289 }
290 
291 /* Move a transaction from the completed list to the polled list */
292 void gsi_trans_move_polled(struct gsi_trans *trans)
293 {
294 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
295 	struct gsi_trans_info *trans_info = &channel->trans_info;
296 
297 	spin_lock_bh(&trans_info->spinlock);
298 
299 	list_move_tail(&trans->links, &trans_info->polled);
300 
301 	spin_unlock_bh(&trans_info->spinlock);
302 }
303 
304 /* Reserve some number of TREs on a channel.  Returns true if successful */
305 static bool
306 gsi_trans_tre_reserve(struct gsi_trans_info *trans_info, u32 tre_count)
307 {
308 	int avail = atomic_read(&trans_info->tre_avail);
309 	int new;
310 
311 	do {
312 		new = avail - (int)tre_count;
313 		if (unlikely(new < 0))
314 			return false;
315 	} while (!atomic_try_cmpxchg(&trans_info->tre_avail, &avail, new));
316 
317 	return true;
318 }
319 
320 /* Release previously-reserved TRE entries to a channel */
321 static void
322 gsi_trans_tre_release(struct gsi_trans_info *trans_info, u32 tre_count)
323 {
324 	atomic_add(tre_count, &trans_info->tre_avail);
325 }
326 
327 /* Return true if no transactions are allocated, false otherwise */
328 bool gsi_channel_trans_idle(struct gsi *gsi, u32 channel_id)
329 {
330 	u32 tre_max = gsi_channel_tre_max(gsi, channel_id);
331 	struct gsi_trans_info *trans_info;
332 
333 	trans_info = &gsi->channel[channel_id].trans_info;
334 
335 	return atomic_read(&trans_info->tre_avail) == tre_max;
336 }
337 
338 /* Allocate a GSI transaction on a channel */
339 struct gsi_trans *gsi_channel_trans_alloc(struct gsi *gsi, u32 channel_id,
340 					  u32 tre_count,
341 					  enum dma_data_direction direction)
342 {
343 	struct gsi_channel *channel = &gsi->channel[channel_id];
344 	struct gsi_trans_info *trans_info;
345 	struct gsi_trans *trans;
346 
347 	if (WARN_ON(tre_count > channel->trans_tre_max))
348 		return NULL;
349 
350 	trans_info = &channel->trans_info;
351 
352 	/* We reserve the TREs now, but consume them at commit time.
353 	 * If there aren't enough available, we're done.
354 	 */
355 	if (!gsi_trans_tre_reserve(trans_info, tre_count))
356 		return NULL;
357 
358 	/* Allocate and initialize non-zero fields in the transaction */
359 	trans = gsi_trans_pool_alloc(&trans_info->pool, 1);
360 	trans->gsi = gsi;
361 	trans->channel_id = channel_id;
362 	trans->rsvd_count = tre_count;
363 	init_completion(&trans->completion);
364 
365 	/* Allocate the scatterlist */
366 	trans->sgl = gsi_trans_pool_alloc(&trans_info->sg_pool, tre_count);
367 	sg_init_marker(trans->sgl, tre_count);
368 
369 	trans->direction = direction;
370 
371 	spin_lock_bh(&trans_info->spinlock);
372 
373 	list_add_tail(&trans->links, &trans_info->alloc);
374 
375 	spin_unlock_bh(&trans_info->spinlock);
376 
377 	refcount_set(&trans->refcount, 1);
378 
379 	return trans;
380 }
381 
382 /* Free a previously-allocated transaction */
383 void gsi_trans_free(struct gsi_trans *trans)
384 {
385 	refcount_t *refcount = &trans->refcount;
386 	struct gsi_trans_info *trans_info;
387 	bool last;
388 
389 	/* We must hold the lock to release the last reference */
390 	if (refcount_dec_not_one(refcount))
391 		return;
392 
393 	trans_info = &trans->gsi->channel[trans->channel_id].trans_info;
394 
395 	spin_lock_bh(&trans_info->spinlock);
396 
397 	/* Reference might have been added before we got the lock */
398 	last = refcount_dec_and_test(refcount);
399 	if (last)
400 		list_del(&trans->links);
401 
402 	spin_unlock_bh(&trans_info->spinlock);
403 
404 	if (!last)
405 		return;
406 
407 	if (trans->used_count)
408 		ipa_gsi_trans_release(trans);
409 
410 	/* Releasing the reserved TREs implicitly frees the sgl[] and
411 	 * (if present) info[] arrays, plus the transaction itself.
412 	 */
413 	gsi_trans_tre_release(trans_info, trans->rsvd_count);
414 }
415 
416 /* Add an immediate command to a transaction */
417 void gsi_trans_cmd_add(struct gsi_trans *trans, void *buf, u32 size,
418 		       dma_addr_t addr, enum ipa_cmd_opcode opcode)
419 {
420 	u32 which = trans->used_count++;
421 	struct scatterlist *sg;
422 
423 	WARN_ON(which >= trans->rsvd_count);
424 
425 	/* Commands are quite different from data transfer requests.
426 	 * Their payloads come from a pool whose memory is allocated
427 	 * using dma_alloc_coherent().  We therefore do *not* map them
428 	 * for DMA (unlike what we do for pages and skbs).
429 	 *
430 	 * When a transaction completes, the SGL is normally unmapped.
431 	 * A command transaction has direction DMA_NONE, which tells
432 	 * gsi_trans_complete() to skip the unmapping step.
433 	 *
434 	 * The only things we use directly in a command scatter/gather
435 	 * entry are the DMA address and length.  We still need the SG
436 	 * table flags to be maintained though, so assign a NULL page
437 	 * pointer for that purpose.
438 	 */
439 	sg = &trans->sgl[which];
440 	sg_assign_page(sg, NULL);
441 	sg_dma_address(sg) = addr;
442 	sg_dma_len(sg) = size;
443 
444 	trans->cmd_opcode[which] = opcode;
445 }
446 
447 /* Add a page transfer to a transaction.  It will fill the only TRE. */
448 int gsi_trans_page_add(struct gsi_trans *trans, struct page *page, u32 size,
449 		       u32 offset)
450 {
451 	struct scatterlist *sg = &trans->sgl[0];
452 	int ret;
453 
454 	if (WARN_ON(trans->rsvd_count != 1))
455 		return -EINVAL;
456 	if (WARN_ON(trans->used_count))
457 		return -EINVAL;
458 
459 	sg_set_page(sg, page, size, offset);
460 	ret = dma_map_sg(trans->gsi->dev, sg, 1, trans->direction);
461 	if (!ret)
462 		return -ENOMEM;
463 
464 	trans->used_count++;	/* Transaction now owns the (DMA mapped) page */
465 
466 	return 0;
467 }
468 
469 /* Add an SKB transfer to a transaction.  No other TREs will be used. */
470 int gsi_trans_skb_add(struct gsi_trans *trans, struct sk_buff *skb)
471 {
472 	struct scatterlist *sg = &trans->sgl[0];
473 	u32 used_count;
474 	int ret;
475 
476 	if (WARN_ON(trans->rsvd_count != 1))
477 		return -EINVAL;
478 	if (WARN_ON(trans->used_count))
479 		return -EINVAL;
480 
481 	/* skb->len will not be 0 (checked early) */
482 	ret = skb_to_sgvec(skb, sg, 0, skb->len);
483 	if (ret < 0)
484 		return ret;
485 	used_count = ret;
486 
487 	ret = dma_map_sg(trans->gsi->dev, sg, used_count, trans->direction);
488 	if (!ret)
489 		return -ENOMEM;
490 
491 	/* Transaction now owns the (DMA mapped) skb */
492 	trans->used_count += used_count;
493 
494 	return 0;
495 }
496 
497 /* Compute the length/opcode value to use for a TRE */
498 static __le16 gsi_tre_len_opcode(enum ipa_cmd_opcode opcode, u32 len)
499 {
500 	return opcode == IPA_CMD_NONE ? cpu_to_le16((u16)len)
501 				      : cpu_to_le16((u16)opcode);
502 }
503 
504 /* Compute the flags value to use for a given TRE */
505 static __le32 gsi_tre_flags(bool last_tre, bool bei, enum ipa_cmd_opcode opcode)
506 {
507 	enum gsi_tre_type tre_type;
508 	u32 tre_flags;
509 
510 	tre_type = opcode == IPA_CMD_NONE ? GSI_RE_XFER : GSI_RE_IMMD_CMD;
511 	tre_flags = u32_encode_bits(tre_type, TRE_FLAGS_TYPE_FMASK);
512 
513 	/* Last TRE contains interrupt flags */
514 	if (last_tre) {
515 		/* All transactions end in a transfer completion interrupt */
516 		tre_flags |= TRE_FLAGS_IEOT_FMASK;
517 		/* Don't interrupt when outbound commands are acknowledged */
518 		if (bei)
519 			tre_flags |= TRE_FLAGS_BEI_FMASK;
520 	} else {	/* All others indicate there's more to come */
521 		tre_flags |= TRE_FLAGS_CHAIN_FMASK;
522 	}
523 
524 	return cpu_to_le32(tre_flags);
525 }
526 
527 static void gsi_trans_tre_fill(struct gsi_tre *dest_tre, dma_addr_t addr,
528 			       u32 len, bool last_tre, bool bei,
529 			       enum ipa_cmd_opcode opcode)
530 {
531 	struct gsi_tre tre;
532 
533 	tre.addr = cpu_to_le64(addr);
534 	tre.len_opcode = gsi_tre_len_opcode(opcode, len);
535 	tre.reserved = 0;
536 	tre.flags = gsi_tre_flags(last_tre, bei, opcode);
537 
538 	/* ARM64 can write 16 bytes as a unit with a single instruction.
539 	 * Doing the assignment this way is an attempt to make that happen.
540 	 */
541 	*dest_tre = tre;
542 }
543 
544 /**
545  * __gsi_trans_commit() - Common GSI transaction commit code
546  * @trans:	Transaction to commit
547  * @ring_db:	Whether to tell the hardware about these queued transfers
548  *
549  * Formats channel ring TRE entries based on the content of the scatterlist.
550  * Maps a transaction pointer to the last ring entry used for the transaction,
551  * so it can be recovered when it completes.  Moves the transaction to the
552  * pending list.  Finally, updates the channel ring pointer and optionally
553  * rings the doorbell.
554  */
555 static void __gsi_trans_commit(struct gsi_trans *trans, bool ring_db)
556 {
557 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
558 	struct gsi_ring *tre_ring = &channel->tre_ring;
559 	enum ipa_cmd_opcode opcode = IPA_CMD_NONE;
560 	bool bei = channel->toward_ipa;
561 	struct gsi_tre *dest_tre;
562 	struct scatterlist *sg;
563 	u32 byte_count = 0;
564 	u8 *cmd_opcode;
565 	u32 avail;
566 	u32 i;
567 
568 	WARN_ON(!trans->used_count);
569 
570 	/* Consume the entries.  If we cross the end of the ring while
571 	 * filling them we'll switch to the beginning to finish.
572 	 * If there is no info array we're doing a simple data
573 	 * transfer request, whose opcode is IPA_CMD_NONE.
574 	 */
575 	cmd_opcode = channel->command ? &trans->cmd_opcode[0] : NULL;
576 	avail = tre_ring->count - tre_ring->index % tre_ring->count;
577 	dest_tre = gsi_ring_virt(tre_ring, tre_ring->index);
578 	for_each_sg(trans->sgl, sg, trans->used_count, i) {
579 		bool last_tre = i == trans->used_count - 1;
580 		dma_addr_t addr = sg_dma_address(sg);
581 		u32 len = sg_dma_len(sg);
582 
583 		byte_count += len;
584 		if (!avail--)
585 			dest_tre = gsi_ring_virt(tre_ring, 0);
586 		if (cmd_opcode)
587 			opcode = *cmd_opcode++;
588 
589 		gsi_trans_tre_fill(dest_tre, addr, len, last_tre, bei, opcode);
590 		dest_tre++;
591 	}
592 	/* Associate the TRE with the transaction */
593 	gsi_trans_map(trans, tre_ring->index);
594 
595 	tre_ring->index += trans->used_count;
596 
597 	trans->len = byte_count;
598 	if (channel->toward_ipa)
599 		gsi_trans_tx_committed(trans);
600 
601 	gsi_trans_move_committed(trans);
602 
603 	/* Ring doorbell if requested, or if all TREs are allocated */
604 	if (ring_db || !atomic_read(&channel->trans_info.tre_avail)) {
605 		/* Report what we're handing off to hardware for TX channels */
606 		if (channel->toward_ipa)
607 			gsi_trans_tx_queued(trans);
608 		gsi_trans_move_pending(trans);
609 		gsi_channel_doorbell(channel);
610 	}
611 }
612 
613 /* Commit a GSI transaction */
614 void gsi_trans_commit(struct gsi_trans *trans, bool ring_db)
615 {
616 	if (trans->used_count)
617 		__gsi_trans_commit(trans, ring_db);
618 	else
619 		gsi_trans_free(trans);
620 }
621 
622 /* Commit a GSI transaction and wait for it to complete */
623 void gsi_trans_commit_wait(struct gsi_trans *trans)
624 {
625 	if (!trans->used_count)
626 		goto out_trans_free;
627 
628 	refcount_inc(&trans->refcount);
629 
630 	__gsi_trans_commit(trans, true);
631 
632 	wait_for_completion(&trans->completion);
633 
634 out_trans_free:
635 	gsi_trans_free(trans);
636 }
637 
638 /* Process the completion of a transaction; called while polling */
639 void gsi_trans_complete(struct gsi_trans *trans)
640 {
641 	/* If the entire SGL was mapped when added, unmap it now */
642 	if (trans->direction != DMA_NONE)
643 		dma_unmap_sg(trans->gsi->dev, trans->sgl, trans->used_count,
644 			     trans->direction);
645 
646 	ipa_gsi_trans_complete(trans);
647 
648 	complete(&trans->completion);
649 
650 	gsi_trans_free(trans);
651 }
652 
653 /* Cancel a channel's pending transactions */
654 void gsi_channel_trans_cancel_pending(struct gsi_channel *channel)
655 {
656 	struct gsi_trans_info *trans_info = &channel->trans_info;
657 	struct gsi_trans *trans;
658 	bool cancelled;
659 
660 	/* channel->gsi->mutex is held by caller */
661 	spin_lock_bh(&trans_info->spinlock);
662 
663 	cancelled = !list_empty(&trans_info->pending);
664 	list_for_each_entry(trans, &trans_info->pending, links)
665 		trans->cancelled = true;
666 
667 	list_splice_tail_init(&trans_info->pending, &trans_info->complete);
668 
669 	spin_unlock_bh(&trans_info->spinlock);
670 
671 	/* Schedule NAPI polling to complete the cancelled transactions */
672 	if (cancelled)
673 		napi_schedule(&channel->napi);
674 }
675 
676 /* Issue a command to read a single byte from a channel */
677 int gsi_trans_read_byte(struct gsi *gsi, u32 channel_id, dma_addr_t addr)
678 {
679 	struct gsi_channel *channel = &gsi->channel[channel_id];
680 	struct gsi_ring *tre_ring = &channel->tre_ring;
681 	struct gsi_trans_info *trans_info;
682 	struct gsi_tre *dest_tre;
683 
684 	trans_info = &channel->trans_info;
685 
686 	/* First reserve the TRE, if possible */
687 	if (!gsi_trans_tre_reserve(trans_info, 1))
688 		return -EBUSY;
689 
690 	/* Now fill the reserved TRE and tell the hardware */
691 
692 	dest_tre = gsi_ring_virt(tre_ring, tre_ring->index);
693 	gsi_trans_tre_fill(dest_tre, addr, 1, true, false, IPA_CMD_NONE);
694 
695 	tre_ring->index++;
696 	gsi_channel_doorbell(channel);
697 
698 	return 0;
699 }
700 
701 /* Mark a gsi_trans_read_byte() request done */
702 void gsi_trans_read_byte_done(struct gsi *gsi, u32 channel_id)
703 {
704 	struct gsi_channel *channel = &gsi->channel[channel_id];
705 
706 	gsi_trans_tre_release(&channel->trans_info, 1);
707 }
708 
709 /* Initialize a channel's GSI transaction info */
710 int gsi_channel_trans_init(struct gsi *gsi, u32 channel_id)
711 {
712 	struct gsi_channel *channel = &gsi->channel[channel_id];
713 	u32 tre_count = channel->tre_count;
714 	struct gsi_trans_info *trans_info;
715 	u32 tre_max;
716 	int ret;
717 
718 	/* Ensure the size of a channel element is what's expected */
719 	BUILD_BUG_ON(sizeof(struct gsi_tre) != GSI_RING_ELEMENT_SIZE);
720 
721 	trans_info = &channel->trans_info;
722 
723 	/* The tre_avail field is what ultimately limits the number of
724 	 * outstanding transactions and their resources.  A transaction
725 	 * allocation succeeds only if the TREs available are sufficient
726 	 * for what the transaction might need.
727 	 */
728 	tre_max = gsi_channel_tre_max(channel->gsi, channel_id);
729 	atomic_set(&trans_info->tre_avail, tre_max);
730 
731 	/* We can't use more TREs than the number available in the ring.
732 	 * This limits the number of transactions that can be outstanding.
733 	 * Worst case is one TRE per transaction (but we actually limit
734 	 * it to something a little less than that).  By allocating a
735 	 * power-of-two number of transactions we can use an index
736 	 * modulo that number to determine the next one that's free.
737 	 * Transactions are allocated one at a time.
738 	 */
739 	ret = gsi_trans_pool_init(&trans_info->pool, sizeof(struct gsi_trans),
740 				  tre_max, 1);
741 	if (ret)
742 		return -ENOMEM;
743 
744 	/* A completion event contains a pointer to the TRE that caused
745 	 * the event (which will be the last one used by the transaction).
746 	 * Each entry in this map records the transaction associated
747 	 * with a corresponding completed TRE.
748 	 */
749 	trans_info->map = kcalloc(tre_count, sizeof(*trans_info->map),
750 				  GFP_KERNEL);
751 	if (!trans_info->map) {
752 		ret = -ENOMEM;
753 		goto err_trans_free;
754 	}
755 
756 	/* A transaction uses a scatterlist array to represent the data
757 	 * transfers implemented by the transaction.  Each scatterlist
758 	 * element is used to fill a single TRE when the transaction is
759 	 * committed.  So we need as many scatterlist elements as the
760 	 * maximum number of TREs that can be outstanding.
761 	 */
762 	ret = gsi_trans_pool_init(&trans_info->sg_pool,
763 				  sizeof(struct scatterlist),
764 				  tre_max, channel->trans_tre_max);
765 	if (ret)
766 		goto err_map_free;
767 
768 	spin_lock_init(&trans_info->spinlock);
769 	INIT_LIST_HEAD(&trans_info->alloc);
770 	INIT_LIST_HEAD(&trans_info->committed);
771 	INIT_LIST_HEAD(&trans_info->pending);
772 	INIT_LIST_HEAD(&trans_info->complete);
773 	INIT_LIST_HEAD(&trans_info->polled);
774 
775 	return 0;
776 
777 err_map_free:
778 	kfree(trans_info->map);
779 err_trans_free:
780 	gsi_trans_pool_exit(&trans_info->pool);
781 
782 	dev_err(gsi->dev, "error %d initializing channel %u transactions\n",
783 		ret, channel_id);
784 
785 	return ret;
786 }
787 
788 /* Inverse of gsi_channel_trans_init() */
789 void gsi_channel_trans_exit(struct gsi_channel *channel)
790 {
791 	struct gsi_trans_info *trans_info = &channel->trans_info;
792 
793 	gsi_trans_pool_exit(&trans_info->sg_pool);
794 	gsi_trans_pool_exit(&trans_info->pool);
795 	kfree(trans_info->map);
796 }
797