xref: /linux/drivers/infiniband/hw/hfi1/user_sdma.c (revision bf4afc53b77aeaa48b5409da5c8da6bb4eff7f43)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright(c) 2020 - 2023 Cornelis Networks, Inc.
4  * Copyright(c) 2015 - 2018 Intel Corporation.
5  */
6 
7 #include <linux/mm.h>
8 #include <linux/types.h>
9 #include <linux/device.h>
10 #include <linux/dmapool.h>
11 #include <linux/slab.h>
12 #include <linux/list.h>
13 #include <linux/highmem.h>
14 #include <linux/io.h>
15 #include <linux/uio.h>
16 #include <linux/rbtree.h>
17 #include <linux/spinlock.h>
18 #include <linux/delay.h>
19 #include <linux/kthread.h>
20 #include <linux/mmu_context.h>
21 #include <linux/module.h>
22 #include <linux/vmalloc.h>
23 #include <linux/string.h>
24 
25 #include "hfi.h"
26 #include "sdma.h"
27 #include "user_sdma.h"
28 #include "verbs.h"  /* for the headers */
29 #include "common.h" /* for struct hfi1_tid_info */
30 #include "trace.h"
31 
32 static uint hfi1_sdma_comp_ring_size = 128;
33 module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO);
34 MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128");
35 
36 static unsigned initial_pkt_count = 8;
37 
38 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts);
39 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status);
40 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq);
41 static void user_sdma_free_request(struct user_sdma_request *req);
42 static int check_header_template(struct user_sdma_request *req,
43 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
44 				 u32 datalen);
45 static int set_txreq_header(struct user_sdma_request *req,
46 			    struct user_sdma_txreq *tx, u32 datalen);
47 static int set_txreq_header_ahg(struct user_sdma_request *req,
48 				struct user_sdma_txreq *tx, u32 len);
49 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
50 				  struct hfi1_user_sdma_comp_q *cq,
51 				  u16 idx, enum hfi1_sdma_comp_state state,
52 				  int ret);
53 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags);
54 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
55 
56 static int defer_packet_queue(
57 	struct sdma_engine *sde,
58 	struct iowait_work *wait,
59 	struct sdma_txreq *txreq,
60 	uint seq,
61 	bool pkts_sent);
62 static void activate_packet_queue(struct iowait *wait, int reason);
63 
defer_packet_queue(struct sdma_engine * sde,struct iowait_work * wait,struct sdma_txreq * txreq,uint seq,bool pkts_sent)64 static int defer_packet_queue(
65 	struct sdma_engine *sde,
66 	struct iowait_work *wait,
67 	struct sdma_txreq *txreq,
68 	uint seq,
69 	bool pkts_sent)
70 {
71 	struct hfi1_user_sdma_pkt_q *pq =
72 		container_of(wait->iow, struct hfi1_user_sdma_pkt_q, busy);
73 
74 	write_seqlock(&sde->waitlock);
75 	trace_hfi1_usdma_defer(pq, sde, &pq->busy);
76 	if (sdma_progress(sde, seq, txreq))
77 		goto eagain;
78 	/*
79 	 * We are assuming that if the list is enqueued somewhere, it
80 	 * is to the dmawait list since that is the only place where
81 	 * it is supposed to be enqueued.
82 	 */
83 	xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
84 	if (list_empty(&pq->busy.list)) {
85 		pq->busy.lock = &sde->waitlock;
86 		iowait_get_priority(&pq->busy);
87 		iowait_queue(pkts_sent, &pq->busy, &sde->dmawait);
88 	}
89 	write_sequnlock(&sde->waitlock);
90 	return -EBUSY;
91 eagain:
92 	write_sequnlock(&sde->waitlock);
93 	return -EAGAIN;
94 }
95 
activate_packet_queue(struct iowait * wait,int reason)96 static void activate_packet_queue(struct iowait *wait, int reason)
97 {
98 	struct hfi1_user_sdma_pkt_q *pq =
99 		container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
100 
101 	trace_hfi1_usdma_activate(pq, wait, reason);
102 	xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
103 	wake_up(&wait->wait_dma);
104 };
105 
hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata * uctxt,struct hfi1_filedata * fd)106 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt,
107 				struct hfi1_filedata *fd)
108 {
109 	int ret = -ENOMEM;
110 	char buf[64];
111 	struct hfi1_devdata *dd;
112 	struct hfi1_user_sdma_comp_q *cq;
113 	struct hfi1_user_sdma_pkt_q *pq;
114 
115 	if (!uctxt || !fd)
116 		return -EBADF;
117 
118 	if (!hfi1_sdma_comp_ring_size)
119 		return -EINVAL;
120 
121 	dd = uctxt->dd;
122 
123 	pq = kzalloc_obj(*pq);
124 	if (!pq)
125 		return -ENOMEM;
126 	pq->dd = dd;
127 	pq->ctxt = uctxt->ctxt;
128 	pq->subctxt = fd->subctxt;
129 	pq->n_max_reqs = hfi1_sdma_comp_ring_size;
130 	atomic_set(&pq->n_reqs, 0);
131 	init_waitqueue_head(&pq->wait);
132 	atomic_set(&pq->n_locked, 0);
133 
134 	iowait_init(&pq->busy, 0, NULL, NULL, defer_packet_queue,
135 		    activate_packet_queue, NULL, NULL);
136 	pq->reqidx = 0;
137 
138 	pq->reqs = kzalloc_objs(*pq->reqs, hfi1_sdma_comp_ring_size);
139 	if (!pq->reqs)
140 		goto pq_reqs_nomem;
141 
142 	pq->req_in_use = bitmap_zalloc(hfi1_sdma_comp_ring_size, GFP_KERNEL);
143 	if (!pq->req_in_use)
144 		goto pq_reqs_no_in_use;
145 
146 	snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
147 		 fd->subctxt);
148 	pq->txreq_cache = kmem_cache_create(buf,
149 					    sizeof(struct user_sdma_txreq),
150 					    L1_CACHE_BYTES,
151 					    SLAB_HWCACHE_ALIGN,
152 					    NULL);
153 	if (!pq->txreq_cache) {
154 		dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
155 			   uctxt->ctxt);
156 		goto pq_txreq_nomem;
157 	}
158 
159 	cq = kzalloc_obj(*cq);
160 	if (!cq)
161 		goto cq_nomem;
162 
163 	cq->comps = vmalloc_user(PAGE_ALIGN(sizeof(*cq->comps)
164 				 * hfi1_sdma_comp_ring_size));
165 	if (!cq->comps)
166 		goto cq_comps_nomem;
167 
168 	cq->nentries = hfi1_sdma_comp_ring_size;
169 
170 	ret = hfi1_init_system_pinning(pq);
171 	if (ret)
172 		goto pq_mmu_fail;
173 
174 	rcu_assign_pointer(fd->pq, pq);
175 	fd->cq = cq;
176 
177 	return 0;
178 
179 pq_mmu_fail:
180 	vfree(cq->comps);
181 cq_comps_nomem:
182 	kfree(cq);
183 cq_nomem:
184 	kmem_cache_destroy(pq->txreq_cache);
185 pq_txreq_nomem:
186 	bitmap_free(pq->req_in_use);
187 pq_reqs_no_in_use:
188 	kfree(pq->reqs);
189 pq_reqs_nomem:
190 	kfree(pq);
191 
192 	return ret;
193 }
194 
flush_pq_iowait(struct hfi1_user_sdma_pkt_q * pq)195 static void flush_pq_iowait(struct hfi1_user_sdma_pkt_q *pq)
196 {
197 	unsigned long flags;
198 	seqlock_t *lock = pq->busy.lock;
199 
200 	if (!lock)
201 		return;
202 	write_seqlock_irqsave(lock, flags);
203 	if (!list_empty(&pq->busy.list)) {
204 		list_del_init(&pq->busy.list);
205 		pq->busy.lock = NULL;
206 	}
207 	write_sequnlock_irqrestore(lock, flags);
208 }
209 
hfi1_user_sdma_free_queues(struct hfi1_filedata * fd,struct hfi1_ctxtdata * uctxt)210 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd,
211 			       struct hfi1_ctxtdata *uctxt)
212 {
213 	struct hfi1_user_sdma_pkt_q *pq;
214 
215 	trace_hfi1_sdma_user_free_queues(uctxt->dd, uctxt->ctxt, fd->subctxt);
216 
217 	spin_lock(&fd->pq_rcu_lock);
218 	pq = srcu_dereference_check(fd->pq, &fd->pq_srcu,
219 				    lockdep_is_held(&fd->pq_rcu_lock));
220 	if (pq) {
221 		rcu_assign_pointer(fd->pq, NULL);
222 		spin_unlock(&fd->pq_rcu_lock);
223 		synchronize_srcu(&fd->pq_srcu);
224 		/* at this point there can be no more new requests */
225 		iowait_sdma_drain(&pq->busy);
226 		/* Wait until all requests have been freed. */
227 		wait_event_interruptible(
228 			pq->wait,
229 			!atomic_read(&pq->n_reqs));
230 		kfree(pq->reqs);
231 		hfi1_free_system_pinning(pq);
232 		bitmap_free(pq->req_in_use);
233 		kmem_cache_destroy(pq->txreq_cache);
234 		flush_pq_iowait(pq);
235 		kfree(pq);
236 	} else {
237 		spin_unlock(&fd->pq_rcu_lock);
238 	}
239 	if (fd->cq) {
240 		vfree(fd->cq->comps);
241 		kfree(fd->cq);
242 		fd->cq = NULL;
243 	}
244 	return 0;
245 }
246 
dlid_to_selector(u16 dlid)247 static u8 dlid_to_selector(u16 dlid)
248 {
249 	static u8 mapping[256];
250 	static int initialized;
251 	static u8 next;
252 	int hash;
253 
254 	if (!initialized) {
255 		memset(mapping, 0xFF, 256);
256 		initialized = 1;
257 	}
258 
259 	hash = ((dlid >> 8) ^ dlid) & 0xFF;
260 	if (mapping[hash] == 0xFF) {
261 		mapping[hash] = next;
262 		next = (next + 1) & 0x7F;
263 	}
264 
265 	return mapping[hash];
266 }
267 
268 /**
269  * hfi1_user_sdma_process_request() - Process and start a user sdma request
270  * @fd: valid file descriptor
271  * @iovec: array of io vectors to process
272  * @dim: overall iovec array size
273  * @count: number of io vector array entries processed
274  */
hfi1_user_sdma_process_request(struct hfi1_filedata * fd,struct iovec * iovec,unsigned long dim,unsigned long * count)275 int hfi1_user_sdma_process_request(struct hfi1_filedata *fd,
276 				   struct iovec *iovec, unsigned long dim,
277 				   unsigned long *count)
278 {
279 	int ret = 0, i;
280 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
281 	struct hfi1_user_sdma_pkt_q *pq =
282 		srcu_dereference(fd->pq, &fd->pq_srcu);
283 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
284 	struct hfi1_devdata *dd = pq->dd;
285 	unsigned long idx = 0;
286 	u8 pcount = initial_pkt_count;
287 	struct sdma_req_info info;
288 	struct user_sdma_request *req;
289 	u8 opcode, sc, vl;
290 	u16 pkey;
291 	u32 slid;
292 	u16 dlid;
293 	u32 selector;
294 
295 	if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
296 		hfi1_cdbg(
297 		   SDMA,
298 		   "[%u:%u:%u] First vector not big enough for header %lu/%lu",
299 		   dd->unit, uctxt->ctxt, fd->subctxt,
300 		   iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
301 		return -EINVAL;
302 	}
303 	ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
304 	if (ret) {
305 		hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
306 			  dd->unit, uctxt->ctxt, fd->subctxt, ret);
307 		return -EFAULT;
308 	}
309 
310 	trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
311 				     (u16 *)&info);
312 	if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
313 		hfi1_cdbg(SDMA,
314 			  "[%u:%u:%u:%u] Invalid comp index",
315 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
316 		return -EINVAL;
317 	}
318 
319 	/*
320 	 * Sanity check the header io vector count.  Need at least 1 vector
321 	 * (header) and cannot be larger than the actual io vector count.
322 	 */
323 	if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
324 		hfi1_cdbg(SDMA,
325 			  "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
326 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
327 			  req_iovcnt(info.ctrl), dim);
328 		return -EINVAL;
329 	}
330 
331 	if (!info.fragsize) {
332 		hfi1_cdbg(SDMA,
333 			  "[%u:%u:%u:%u] Request does not specify fragsize",
334 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
335 		return -EINVAL;
336 	}
337 
338 	/* Try to claim the request. */
339 	if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
340 		hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
341 			  dd->unit, uctxt->ctxt, fd->subctxt,
342 			  info.comp_idx);
343 		return -EBADSLT;
344 	}
345 	/*
346 	 * All safety checks have been done and this request has been claimed.
347 	 */
348 	trace_hfi1_sdma_user_process_request(dd, uctxt->ctxt, fd->subctxt,
349 					     info.comp_idx);
350 	req = pq->reqs + info.comp_idx;
351 	req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
352 	req->data_len  = 0;
353 	req->pq = pq;
354 	req->cq = cq;
355 	req->ahg_idx = -1;
356 	req->iov_idx = 0;
357 	req->sent = 0;
358 	req->seqnum = 0;
359 	req->seqcomp = 0;
360 	req->seqsubmitted = 0;
361 	req->tids = NULL;
362 	req->has_error = 0;
363 	INIT_LIST_HEAD(&req->txps);
364 
365 	memcpy(&req->info, &info, sizeof(info));
366 
367 	/* The request is initialized, count it */
368 	atomic_inc(&pq->n_reqs);
369 
370 	if (req_opcode(info.ctrl) == EXPECTED) {
371 		/* expected must have a TID info and at least one data vector */
372 		if (req->data_iovs < 2) {
373 			SDMA_DBG(req,
374 				 "Not enough vectors for expected request");
375 			ret = -EINVAL;
376 			goto free_req;
377 		}
378 		req->data_iovs--;
379 	}
380 
381 	if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
382 		SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
383 			 MAX_VECTORS_PER_REQ);
384 		ret = -EINVAL;
385 		goto free_req;
386 	}
387 
388 	/* Copy the header from the user buffer */
389 	ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
390 			     sizeof(req->hdr));
391 	if (ret) {
392 		SDMA_DBG(req, "Failed to copy header template (%d)", ret);
393 		ret = -EFAULT;
394 		goto free_req;
395 	}
396 
397 	/* If Static rate control is not enabled, sanitize the header. */
398 	if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
399 		req->hdr.pbc[2] = 0;
400 
401 	/* Validate the opcode. Do not trust packets from user space blindly. */
402 	opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
403 	if ((opcode & USER_OPCODE_CHECK_MASK) !=
404 	     USER_OPCODE_CHECK_VAL) {
405 		SDMA_DBG(req, "Invalid opcode (%d)", opcode);
406 		ret = -EINVAL;
407 		goto free_req;
408 	}
409 	/*
410 	 * Validate the vl. Do not trust packets from user space blindly.
411 	 * VL comes from PBC, SC comes from LRH, and the VL needs to
412 	 * match the SC look up.
413 	 */
414 	vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
415 	sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
416 	      (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
417 	if (vl >= dd->pport->vls_operational ||
418 	    vl != sc_to_vlt(dd, sc)) {
419 		SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
420 		ret = -EINVAL;
421 		goto free_req;
422 	}
423 
424 	/* Checking P_KEY for requests from user-space */
425 	pkey = (u16)be32_to_cpu(req->hdr.bth[0]);
426 	slid = be16_to_cpu(req->hdr.lrh[3]);
427 	if (egress_pkey_check(dd->pport, slid, pkey, sc, PKEY_CHECK_INVALID)) {
428 		ret = -EINVAL;
429 		goto free_req;
430 	}
431 
432 	/*
433 	 * Also should check the BTH.lnh. If it says the next header is GRH then
434 	 * the RXE parsing will be off and will land in the middle of the KDETH
435 	 * or miss it entirely.
436 	 */
437 	if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
438 		SDMA_DBG(req, "User tried to pass in a GRH");
439 		ret = -EINVAL;
440 		goto free_req;
441 	}
442 
443 	req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
444 	/*
445 	 * Calculate the initial TID offset based on the values of
446 	 * KDETH.OFFSET and KDETH.OM that are passed in.
447 	 */
448 	req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
449 		(KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
450 		 KDETH_OM_LARGE : KDETH_OM_SMALL);
451 	trace_hfi1_sdma_user_initial_tidoffset(dd, uctxt->ctxt, fd->subctxt,
452 					       info.comp_idx, req->tidoffset);
453 	idx++;
454 
455 	/* Save all the IO vector structures */
456 	for (i = 0; i < req->data_iovs; i++) {
457 		req->iovs[i].offset = 0;
458 		INIT_LIST_HEAD(&req->iovs[i].list);
459 		memcpy(&req->iovs[i].iov,
460 		       iovec + idx++,
461 		       sizeof(req->iovs[i].iov));
462 		if (req->iovs[i].iov.iov_len == 0) {
463 			ret = -EINVAL;
464 			goto free_req;
465 		}
466 		req->data_len += req->iovs[i].iov.iov_len;
467 	}
468 	trace_hfi1_sdma_user_data_length(dd, uctxt->ctxt, fd->subctxt,
469 					 info.comp_idx, req->data_len);
470 	if (pcount > req->info.npkts)
471 		pcount = req->info.npkts;
472 	/*
473 	 * Copy any TID info
474 	 * User space will provide the TID info only when the
475 	 * request type is EXPECTED. This is true even if there is
476 	 * only one packet in the request and the header is already
477 	 * setup. The reason for the singular TID case is that the
478 	 * driver needs to perform safety checks.
479 	 */
480 	if (req_opcode(req->info.ctrl) == EXPECTED) {
481 		u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
482 		u32 *tmp;
483 
484 		if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
485 			ret = -EINVAL;
486 			goto free_req;
487 		}
488 
489 		/*
490 		 * We have to copy all of the tids because they may vary
491 		 * in size and, therefore, the TID count might not be
492 		 * equal to the pkt count. However, there is no way to
493 		 * tell at this point.
494 		 */
495 		tmp = memdup_array_user(iovec[idx].iov_base,
496 					ntids, sizeof(*req->tids));
497 		if (IS_ERR(tmp)) {
498 			ret = PTR_ERR(tmp);
499 			SDMA_DBG(req, "Failed to copy %d TIDs (%pe)", ntids,
500 				 tmp);
501 			goto free_req;
502 		}
503 		req->tids = tmp;
504 		req->n_tids = ntids;
505 		req->tididx = 0;
506 		idx++;
507 	}
508 
509 	dlid = be16_to_cpu(req->hdr.lrh[1]);
510 	selector = dlid_to_selector(dlid);
511 	selector += uctxt->ctxt + fd->subctxt;
512 	req->sde = sdma_select_user_engine(dd, selector, vl);
513 
514 	if (!req->sde || !sdma_running(req->sde)) {
515 		ret = -ECOMM;
516 		goto free_req;
517 	}
518 
519 	/* We don't need an AHG entry if the request contains only one packet */
520 	if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG))
521 		req->ahg_idx = sdma_ahg_alloc(req->sde);
522 
523 	set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
524 	pq->state = SDMA_PKT_Q_ACTIVE;
525 
526 	/*
527 	 * This is a somewhat blocking send implementation.
528 	 * The driver will block the caller until all packets of the
529 	 * request have been submitted to the SDMA engine. However, it
530 	 * will not wait for send completions.
531 	 */
532 	while (req->seqsubmitted != req->info.npkts) {
533 		ret = user_sdma_send_pkts(req, pcount);
534 		if (ret < 0) {
535 			int we_ret;
536 
537 			if (ret != -EBUSY)
538 				goto free_req;
539 			we_ret = wait_event_interruptible_timeout(
540 				pq->busy.wait_dma,
541 				pq->state == SDMA_PKT_Q_ACTIVE,
542 				msecs_to_jiffies(
543 					SDMA_IOWAIT_TIMEOUT));
544 			trace_hfi1_usdma_we(pq, we_ret);
545 			if (we_ret <= 0)
546 				flush_pq_iowait(pq);
547 		}
548 	}
549 	*count += idx;
550 	return 0;
551 free_req:
552 	/*
553 	 * If the submitted seqsubmitted == npkts, the completion routine
554 	 * controls the final state.  If sequbmitted < npkts, wait for any
555 	 * outstanding packets to finish before cleaning up.
556 	 */
557 	if (req->seqsubmitted < req->info.npkts) {
558 		if (req->seqsubmitted)
559 			wait_event(pq->busy.wait_dma,
560 				   (req->seqcomp == req->seqsubmitted - 1));
561 		user_sdma_free_request(req);
562 		pq_update(pq);
563 		set_comp_state(pq, cq, info.comp_idx, ERROR, ret);
564 	}
565 	return ret;
566 }
567 
compute_data_length(struct user_sdma_request * req,struct user_sdma_txreq * tx)568 static inline u32 compute_data_length(struct user_sdma_request *req,
569 				      struct user_sdma_txreq *tx)
570 {
571 	/*
572 	 * Determine the proper size of the packet data.
573 	 * The size of the data of the first packet is in the header
574 	 * template. However, it includes the header and ICRC, which need
575 	 * to be subtracted.
576 	 * The minimum representable packet data length in a header is 4 bytes,
577 	 * therefore, when the data length request is less than 4 bytes, there's
578 	 * only one packet, and the packet data length is equal to that of the
579 	 * request data length.
580 	 * The size of the remaining packets is the minimum of the frag
581 	 * size (MTU) or remaining data in the request.
582 	 */
583 	u32 len;
584 
585 	if (!req->seqnum) {
586 		if (req->data_len < sizeof(u32))
587 			len = req->data_len;
588 		else
589 			len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
590 			       (sizeof(tx->hdr) - 4));
591 	} else if (req_opcode(req->info.ctrl) == EXPECTED) {
592 		u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
593 			PAGE_SIZE;
594 		/*
595 		 * Get the data length based on the remaining space in the
596 		 * TID pair.
597 		 */
598 		len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
599 		/* If we've filled up the TID pair, move to the next one. */
600 		if (unlikely(!len) && ++req->tididx < req->n_tids &&
601 		    req->tids[req->tididx]) {
602 			tidlen = EXP_TID_GET(req->tids[req->tididx],
603 					     LEN) * PAGE_SIZE;
604 			req->tidoffset = 0;
605 			len = min_t(u32, tidlen, req->info.fragsize);
606 		}
607 		/*
608 		 * Since the TID pairs map entire pages, make sure that we
609 		 * are not going to try to send more data that we have
610 		 * remaining.
611 		 */
612 		len = min(len, req->data_len - req->sent);
613 	} else {
614 		len = min(req->data_len - req->sent, (u32)req->info.fragsize);
615 	}
616 	trace_hfi1_sdma_user_compute_length(req->pq->dd,
617 					    req->pq->ctxt,
618 					    req->pq->subctxt,
619 					    req->info.comp_idx,
620 					    len);
621 	return len;
622 }
623 
pad_len(u32 len)624 static inline u32 pad_len(u32 len)
625 {
626 	if (len & (sizeof(u32) - 1))
627 		len += sizeof(u32) - (len & (sizeof(u32) - 1));
628 	return len;
629 }
630 
get_lrh_len(struct hfi1_pkt_header hdr,u32 len)631 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
632 {
633 	/* (Size of complete header - size of PBC) + 4B ICRC + data length */
634 	return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
635 }
636 
user_sdma_txadd_ahg(struct user_sdma_request * req,struct user_sdma_txreq * tx,u32 datalen)637 static int user_sdma_txadd_ahg(struct user_sdma_request *req,
638 			       struct user_sdma_txreq *tx,
639 			       u32 datalen)
640 {
641 	int ret;
642 	u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
643 	u32 lrhlen = get_lrh_len(req->hdr, pad_len(datalen));
644 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
645 
646 	/*
647 	 * Copy the request header into the tx header
648 	 * because the HW needs a cacheline-aligned
649 	 * address.
650 	 * This copy can be optimized out if the hdr
651 	 * member of user_sdma_request were also
652 	 * cacheline aligned.
653 	 */
654 	memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
655 	if (PBC2LRH(pbclen) != lrhlen) {
656 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
657 		tx->hdr.pbc[0] = cpu_to_le16(pbclen);
658 	}
659 	ret = check_header_template(req, &tx->hdr, lrhlen, datalen);
660 	if (ret)
661 		return ret;
662 	ret = sdma_txinit_ahg(&tx->txreq, SDMA_TXREQ_F_AHG_COPY,
663 			      sizeof(tx->hdr) + datalen, req->ahg_idx,
664 			      0, NULL, 0, user_sdma_txreq_cb);
665 	if (ret)
666 		return ret;
667 	ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq, &tx->hdr, sizeof(tx->hdr));
668 	if (ret)
669 		sdma_txclean(pq->dd, &tx->txreq);
670 	return ret;
671 }
672 
user_sdma_send_pkts(struct user_sdma_request * req,u16 maxpkts)673 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts)
674 {
675 	int ret = 0;
676 	u16 count;
677 	unsigned npkts = 0;
678 	struct user_sdma_txreq *tx = NULL;
679 	struct hfi1_user_sdma_pkt_q *pq = NULL;
680 	struct user_sdma_iovec *iovec = NULL;
681 
682 	if (!req->pq)
683 		return -EINVAL;
684 
685 	pq = req->pq;
686 
687 	/* If tx completion has reported an error, we are done. */
688 	if (READ_ONCE(req->has_error))
689 		return -EFAULT;
690 
691 	/*
692 	 * Check if we might have sent the entire request already
693 	 */
694 	if (unlikely(req->seqnum == req->info.npkts)) {
695 		if (!list_empty(&req->txps))
696 			goto dosend;
697 		return ret;
698 	}
699 
700 	if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
701 		maxpkts = req->info.npkts - req->seqnum;
702 
703 	while (npkts < maxpkts) {
704 		u32 datalen = 0;
705 
706 		/*
707 		 * Check whether any of the completions have come back
708 		 * with errors. If so, we are not going to process any
709 		 * more packets from this request.
710 		 */
711 		if (READ_ONCE(req->has_error))
712 			return -EFAULT;
713 
714 		tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
715 		if (!tx)
716 			return -ENOMEM;
717 
718 		tx->flags = 0;
719 		tx->req = req;
720 		INIT_LIST_HEAD(&tx->list);
721 
722 		/*
723 		 * For the last packet set the ACK request
724 		 * and disable header suppression.
725 		 */
726 		if (req->seqnum == req->info.npkts - 1)
727 			tx->flags |= (TXREQ_FLAGS_REQ_ACK |
728 				      TXREQ_FLAGS_REQ_DISABLE_SH);
729 
730 		/*
731 		 * Calculate the payload size - this is min of the fragment
732 		 * (MTU) size or the remaining bytes in the request but only
733 		 * if we have payload data.
734 		 */
735 		if (req->data_len) {
736 			iovec = &req->iovs[req->iov_idx];
737 			if (READ_ONCE(iovec->offset) == iovec->iov.iov_len) {
738 				if (++req->iov_idx == req->data_iovs) {
739 					ret = -EFAULT;
740 					goto free_tx;
741 				}
742 				iovec = &req->iovs[req->iov_idx];
743 				WARN_ON(iovec->offset);
744 			}
745 
746 			datalen = compute_data_length(req, tx);
747 
748 			/*
749 			 * Disable header suppression for the payload <= 8DWS.
750 			 * If there is an uncorrectable error in the receive
751 			 * data FIFO when the received payload size is less than
752 			 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
753 			 * not reported.There is set RHF.EccErr if the header
754 			 * is not suppressed.
755 			 */
756 			if (!datalen) {
757 				SDMA_DBG(req,
758 					 "Request has data but pkt len is 0");
759 				ret = -EFAULT;
760 				goto free_tx;
761 			} else if (datalen <= 32) {
762 				tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
763 			}
764 		}
765 
766 		if (req->ahg_idx >= 0) {
767 			if (!req->seqnum) {
768 				ret = user_sdma_txadd_ahg(req, tx, datalen);
769 				if (ret)
770 					goto free_tx;
771 			} else {
772 				int changes;
773 
774 				changes = set_txreq_header_ahg(req, tx,
775 							       datalen);
776 				if (changes < 0) {
777 					ret = changes;
778 					goto free_tx;
779 				}
780 			}
781 		} else {
782 			ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
783 					  datalen, user_sdma_txreq_cb);
784 			if (ret)
785 				goto free_tx;
786 			/*
787 			 * Modify the header for this packet. This only needs
788 			 * to be done if we are not going to use AHG. Otherwise,
789 			 * the HW will do it based on the changes we gave it
790 			 * during sdma_txinit_ahg().
791 			 */
792 			ret = set_txreq_header(req, tx, datalen);
793 			if (ret)
794 				goto free_txreq;
795 		}
796 
797 		req->koffset += datalen;
798 		if (req_opcode(req->info.ctrl) == EXPECTED)
799 			req->tidoffset += datalen;
800 		req->sent += datalen;
801 		while (datalen) {
802 			ret = hfi1_add_pages_to_sdma_packet(req, tx, iovec,
803 							    &datalen);
804 			if (ret)
805 				goto free_txreq;
806 			iovec = &req->iovs[req->iov_idx];
807 		}
808 		list_add_tail(&tx->txreq.list, &req->txps);
809 		/*
810 		 * It is important to increment this here as it is used to
811 		 * generate the BTH.PSN and, therefore, can't be bulk-updated
812 		 * outside of the loop.
813 		 */
814 		tx->seqnum = req->seqnum++;
815 		npkts++;
816 	}
817 dosend:
818 	ret = sdma_send_txlist(req->sde,
819 			       iowait_get_ib_work(&pq->busy),
820 			       &req->txps, &count);
821 	req->seqsubmitted += count;
822 	if (req->seqsubmitted == req->info.npkts) {
823 		/*
824 		 * The txreq has already been submitted to the HW queue
825 		 * so we can free the AHG entry now. Corruption will not
826 		 * happen due to the sequential manner in which
827 		 * descriptors are processed.
828 		 */
829 		if (req->ahg_idx >= 0)
830 			sdma_ahg_free(req->sde, req->ahg_idx);
831 	}
832 	return ret;
833 
834 free_txreq:
835 	sdma_txclean(pq->dd, &tx->txreq);
836 free_tx:
837 	kmem_cache_free(pq->txreq_cache, tx);
838 	return ret;
839 }
840 
check_header_template(struct user_sdma_request * req,struct hfi1_pkt_header * hdr,u32 lrhlen,u32 datalen)841 static int check_header_template(struct user_sdma_request *req,
842 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
843 				 u32 datalen)
844 {
845 	/*
846 	 * Perform safety checks for any type of packet:
847 	 *    - transfer size is multiple of 64bytes
848 	 *    - packet length is multiple of 4 bytes
849 	 *    - packet length is not larger than MTU size
850 	 *
851 	 * These checks are only done for the first packet of the
852 	 * transfer since the header is "given" to us by user space.
853 	 * For the remainder of the packets we compute the values.
854 	 */
855 	if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
856 	    lrhlen > get_lrh_len(*hdr, req->info.fragsize))
857 		return -EINVAL;
858 
859 	if (req_opcode(req->info.ctrl) == EXPECTED) {
860 		/*
861 		 * The header is checked only on the first packet. Furthermore,
862 		 * we ensure that at least one TID entry is copied when the
863 		 * request is submitted. Therefore, we don't have to verify that
864 		 * tididx points to something sane.
865 		 */
866 		u32 tidval = req->tids[req->tididx],
867 			tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
868 			tididx = EXP_TID_GET(tidval, IDX),
869 			tidctrl = EXP_TID_GET(tidval, CTRL),
870 			tidoff;
871 		__le32 kval = hdr->kdeth.ver_tid_offset;
872 
873 		tidoff = KDETH_GET(kval, OFFSET) *
874 			  (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
875 			   KDETH_OM_LARGE : KDETH_OM_SMALL);
876 		/*
877 		 * Expected receive packets have the following
878 		 * additional checks:
879 		 *     - offset is not larger than the TID size
880 		 *     - TIDCtrl values match between header and TID array
881 		 *     - TID indexes match between header and TID array
882 		 */
883 		if ((tidoff + datalen > tidlen) ||
884 		    KDETH_GET(kval, TIDCTRL) != tidctrl ||
885 		    KDETH_GET(kval, TID) != tididx)
886 			return -EINVAL;
887 	}
888 	return 0;
889 }
890 
891 /*
892  * Correctly set the BTH.PSN field based on type of
893  * transfer - eager packets can just increment the PSN but
894  * expected packets encode generation and sequence in the
895  * BTH.PSN field so just incrementing will result in errors.
896  */
set_pkt_bth_psn(__be32 bthpsn,u8 expct,u32 frags)897 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
898 {
899 	u32 val = be32_to_cpu(bthpsn),
900 		mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
901 			0xffffffull),
902 		psn = val & mask;
903 	if (expct)
904 		psn = (psn & ~HFI1_KDETH_BTH_SEQ_MASK) |
905 			((psn + frags) & HFI1_KDETH_BTH_SEQ_MASK);
906 	else
907 		psn = psn + frags;
908 	return psn & mask;
909 }
910 
set_txreq_header(struct user_sdma_request * req,struct user_sdma_txreq * tx,u32 datalen)911 static int set_txreq_header(struct user_sdma_request *req,
912 			    struct user_sdma_txreq *tx, u32 datalen)
913 {
914 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
915 	struct hfi1_pkt_header *hdr = &tx->hdr;
916 	u8 omfactor; /* KDETH.OM */
917 	u16 pbclen;
918 	int ret;
919 	u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
920 
921 	/* Copy the header template to the request before modification */
922 	memcpy(hdr, &req->hdr, sizeof(*hdr));
923 
924 	/*
925 	 * Check if the PBC and LRH length are mismatched. If so
926 	 * adjust both in the header.
927 	 */
928 	pbclen = le16_to_cpu(hdr->pbc[0]);
929 	if (PBC2LRH(pbclen) != lrhlen) {
930 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
931 		hdr->pbc[0] = cpu_to_le16(pbclen);
932 		hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
933 		/*
934 		 * Third packet
935 		 * This is the first packet in the sequence that has
936 		 * a "static" size that can be used for the rest of
937 		 * the packets (besides the last one).
938 		 */
939 		if (unlikely(req->seqnum == 2)) {
940 			/*
941 			 * From this point on the lengths in both the
942 			 * PBC and LRH are the same until the last
943 			 * packet.
944 			 * Adjust the template so we don't have to update
945 			 * every packet
946 			 */
947 			req->hdr.pbc[0] = hdr->pbc[0];
948 			req->hdr.lrh[2] = hdr->lrh[2];
949 		}
950 	}
951 	/*
952 	 * We only have to modify the header if this is not the
953 	 * first packet in the request. Otherwise, we use the
954 	 * header given to us.
955 	 */
956 	if (unlikely(!req->seqnum)) {
957 		ret = check_header_template(req, hdr, lrhlen, datalen);
958 		if (ret)
959 			return ret;
960 		goto done;
961 	}
962 
963 	hdr->bth[2] = cpu_to_be32(
964 		set_pkt_bth_psn(hdr->bth[2],
965 				(req_opcode(req->info.ctrl) == EXPECTED),
966 				req->seqnum));
967 
968 	/* Set ACK request on last packet */
969 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
970 		hdr->bth[2] |= cpu_to_be32(1UL << 31);
971 
972 	/* Set the new offset */
973 	hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
974 	/* Expected packets have to fill in the new TID information */
975 	if (req_opcode(req->info.ctrl) == EXPECTED) {
976 		tidval = req->tids[req->tididx];
977 		/*
978 		 * If the offset puts us at the end of the current TID,
979 		 * advance everything.
980 		 */
981 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
982 					 PAGE_SIZE)) {
983 			req->tidoffset = 0;
984 			/*
985 			 * Since we don't copy all the TIDs, all at once,
986 			 * we have to check again.
987 			 */
988 			if (++req->tididx > req->n_tids - 1 ||
989 			    !req->tids[req->tididx]) {
990 				return -EINVAL;
991 			}
992 			tidval = req->tids[req->tididx];
993 		}
994 		omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
995 			KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE_SHIFT :
996 			KDETH_OM_SMALL_SHIFT;
997 		/* Set KDETH.TIDCtrl based on value for this TID. */
998 		KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
999 			  EXP_TID_GET(tidval, CTRL));
1000 		/* Set KDETH.TID based on value for this TID */
1001 		KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1002 			  EXP_TID_GET(tidval, IDX));
1003 		/* Clear KDETH.SH when DISABLE_SH flag is set */
1004 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1005 			KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1006 		/*
1007 		 * Set the KDETH.OFFSET and KDETH.OM based on size of
1008 		 * transfer.
1009 		 */
1010 		trace_hfi1_sdma_user_tid_info(
1011 			pq->dd, pq->ctxt, pq->subctxt, req->info.comp_idx,
1012 			req->tidoffset, req->tidoffset >> omfactor,
1013 			omfactor != KDETH_OM_SMALL_SHIFT);
1014 		KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1015 			  req->tidoffset >> omfactor);
1016 		KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1017 			  omfactor != KDETH_OM_SMALL_SHIFT);
1018 	}
1019 done:
1020 	trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1021 				    req->info.comp_idx, hdr, tidval);
1022 	return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1023 }
1024 
set_txreq_header_ahg(struct user_sdma_request * req,struct user_sdma_txreq * tx,u32 datalen)1025 static int set_txreq_header_ahg(struct user_sdma_request *req,
1026 				struct user_sdma_txreq *tx, u32 datalen)
1027 {
1028 	u32 ahg[AHG_KDETH_ARRAY_SIZE];
1029 	int idx = 0;
1030 	u8 omfactor; /* KDETH.OM */
1031 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1032 	struct hfi1_pkt_header *hdr = &req->hdr;
1033 	u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1034 	u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1035 	size_t array_size = ARRAY_SIZE(ahg);
1036 
1037 	if (PBC2LRH(pbclen) != lrhlen) {
1038 		/* PBC.PbcLengthDWs */
1039 		idx = ahg_header_set(ahg, idx, array_size, 0, 0, 12,
1040 				     (__force u16)cpu_to_le16(LRH2PBC(lrhlen)));
1041 		if (idx < 0)
1042 			return idx;
1043 		/* LRH.PktLen (we need the full 16 bits due to byte swap) */
1044 		idx = ahg_header_set(ahg, idx, array_size, 3, 0, 16,
1045 				     (__force u16)cpu_to_be16(lrhlen >> 2));
1046 		if (idx < 0)
1047 			return idx;
1048 	}
1049 
1050 	/*
1051 	 * Do the common updates
1052 	 */
1053 	/* BTH.PSN and BTH.A */
1054 	val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1055 		(HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1056 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1057 		val32 |= 1UL << 31;
1058 	idx = ahg_header_set(ahg, idx, array_size, 6, 0, 16,
1059 			     (__force u16)cpu_to_be16(val32 >> 16));
1060 	if (idx < 0)
1061 		return idx;
1062 	idx = ahg_header_set(ahg, idx, array_size, 6, 16, 16,
1063 			     (__force u16)cpu_to_be16(val32 & 0xffff));
1064 	if (idx < 0)
1065 		return idx;
1066 	/* KDETH.Offset */
1067 	idx = ahg_header_set(ahg, idx, array_size, 15, 0, 16,
1068 			     (__force u16)cpu_to_le16(req->koffset & 0xffff));
1069 	if (idx < 0)
1070 		return idx;
1071 	idx = ahg_header_set(ahg, idx, array_size, 15, 16, 16,
1072 			     (__force u16)cpu_to_le16(req->koffset >> 16));
1073 	if (idx < 0)
1074 		return idx;
1075 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1076 		__le16 val;
1077 
1078 		tidval = req->tids[req->tididx];
1079 
1080 		/*
1081 		 * If the offset puts us at the end of the current TID,
1082 		 * advance everything.
1083 		 */
1084 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1085 					 PAGE_SIZE)) {
1086 			req->tidoffset = 0;
1087 			/*
1088 			 * Since we don't copy all the TIDs, all at once,
1089 			 * we have to check again.
1090 			 */
1091 			if (++req->tididx > req->n_tids - 1 ||
1092 			    !req->tids[req->tididx])
1093 				return -EINVAL;
1094 			tidval = req->tids[req->tididx];
1095 		}
1096 		omfactor = ((EXP_TID_GET(tidval, LEN) *
1097 				  PAGE_SIZE) >=
1098 				 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE_SHIFT :
1099 				 KDETH_OM_SMALL_SHIFT;
1100 		/* KDETH.OM and KDETH.OFFSET (TID) */
1101 		idx = ahg_header_set(
1102 				ahg, idx, array_size, 7, 0, 16,
1103 				((!!(omfactor - KDETH_OM_SMALL_SHIFT)) << 15 |
1104 				((req->tidoffset >> omfactor)
1105 				& 0x7fff)));
1106 		if (idx < 0)
1107 			return idx;
1108 		/* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1109 		val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1110 				   (EXP_TID_GET(tidval, IDX) & 0x3ff));
1111 
1112 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1113 			val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1114 						      INTR) <<
1115 					    AHG_KDETH_INTR_SHIFT));
1116 		} else {
1117 			val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1118 			       cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1119 			       cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1120 						      INTR) <<
1121 					     AHG_KDETH_INTR_SHIFT));
1122 		}
1123 
1124 		idx = ahg_header_set(ahg, idx, array_size,
1125 				     7, 16, 14, (__force u16)val);
1126 		if (idx < 0)
1127 			return idx;
1128 	}
1129 
1130 	trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1131 					req->info.comp_idx, req->sde->this_idx,
1132 					req->ahg_idx, ahg, idx, tidval);
1133 	sdma_txinit_ahg(&tx->txreq,
1134 			SDMA_TXREQ_F_USE_AHG,
1135 			datalen, req->ahg_idx, idx,
1136 			ahg, sizeof(req->hdr),
1137 			user_sdma_txreq_cb);
1138 
1139 	return idx;
1140 }
1141 
1142 /**
1143  * user_sdma_txreq_cb() - SDMA tx request completion callback.
1144  * @txreq: valid sdma tx request
1145  * @status: success/failure of request
1146  *
1147  * Called when the SDMA progress state machine gets notification that
1148  * the SDMA descriptors for this tx request have been processed by the
1149  * DMA engine. Called in interrupt context.
1150  * Only do work on completed sequences.
1151  */
user_sdma_txreq_cb(struct sdma_txreq * txreq,int status)1152 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1153 {
1154 	struct user_sdma_txreq *tx =
1155 		container_of(txreq, struct user_sdma_txreq, txreq);
1156 	struct user_sdma_request *req;
1157 	struct hfi1_user_sdma_pkt_q *pq;
1158 	struct hfi1_user_sdma_comp_q *cq;
1159 	enum hfi1_sdma_comp_state state = COMPLETE;
1160 
1161 	if (!tx->req)
1162 		return;
1163 
1164 	req = tx->req;
1165 	pq = req->pq;
1166 	cq = req->cq;
1167 
1168 	if (status != SDMA_TXREQ_S_OK) {
1169 		SDMA_DBG(req, "SDMA completion with error %d",
1170 			 status);
1171 		WRITE_ONCE(req->has_error, 1);
1172 		state = ERROR;
1173 	}
1174 
1175 	req->seqcomp = tx->seqnum;
1176 	kmem_cache_free(pq->txreq_cache, tx);
1177 
1178 	/* sequence isn't complete?  We are done */
1179 	if (req->seqcomp != req->info.npkts - 1)
1180 		return;
1181 
1182 	user_sdma_free_request(req);
1183 	set_comp_state(pq, cq, req->info.comp_idx, state, status);
1184 	pq_update(pq);
1185 }
1186 
pq_update(struct hfi1_user_sdma_pkt_q * pq)1187 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1188 {
1189 	if (atomic_dec_and_test(&pq->n_reqs))
1190 		wake_up(&pq->wait);
1191 }
1192 
user_sdma_free_request(struct user_sdma_request * req)1193 static void user_sdma_free_request(struct user_sdma_request *req)
1194 {
1195 	if (!list_empty(&req->txps)) {
1196 		struct sdma_txreq *t, *p;
1197 
1198 		list_for_each_entry_safe(t, p, &req->txps, list) {
1199 			struct user_sdma_txreq *tx =
1200 				container_of(t, struct user_sdma_txreq, txreq);
1201 			list_del_init(&t->list);
1202 			sdma_txclean(req->pq->dd, t);
1203 			kmem_cache_free(req->pq->txreq_cache, tx);
1204 		}
1205 	}
1206 
1207 	kfree(req->tids);
1208 	clear_bit(req->info.comp_idx, req->pq->req_in_use);
1209 }
1210 
set_comp_state(struct hfi1_user_sdma_pkt_q * pq,struct hfi1_user_sdma_comp_q * cq,u16 idx,enum hfi1_sdma_comp_state state,int ret)1211 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1212 				  struct hfi1_user_sdma_comp_q *cq,
1213 				  u16 idx, enum hfi1_sdma_comp_state state,
1214 				  int ret)
1215 {
1216 	if (state == ERROR)
1217 		cq->comps[idx].errcode = -ret;
1218 	smp_wmb(); /* make sure errcode is visible first */
1219 	cq->comps[idx].status = state;
1220 	trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1221 					idx, state, ret);
1222 }
1223