1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2020 - 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 "mmu_rb.h"
28 #include "user_sdma.h"
29 #include "verbs.h"  /* for the headers */
30 #include "common.h" /* for struct hfi1_tid_info */
31 #include "trace.h"
32 
33 static uint hfi1_sdma_comp_ring_size = 128;
34 module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO);
35 MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128");
36 
37 static unsigned initial_pkt_count = 8;
38 
39 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts);
40 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status);
41 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq);
42 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin);
43 static int pin_vector_pages(struct user_sdma_request *req,
44 			    struct user_sdma_iovec *iovec);
45 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
46 			       unsigned start, unsigned npages);
47 static int check_header_template(struct user_sdma_request *req,
48 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
49 				 u32 datalen);
50 static int set_txreq_header(struct user_sdma_request *req,
51 			    struct user_sdma_txreq *tx, u32 datalen);
52 static int set_txreq_header_ahg(struct user_sdma_request *req,
53 				struct user_sdma_txreq *tx, u32 len);
54 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
55 				  struct hfi1_user_sdma_comp_q *cq,
56 				  u16 idx, enum hfi1_sdma_comp_state state,
57 				  int ret);
58 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags);
59 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
60 
61 static int defer_packet_queue(
62 	struct sdma_engine *sde,
63 	struct iowait_work *wait,
64 	struct sdma_txreq *txreq,
65 	uint seq,
66 	bool pkts_sent);
67 static void activate_packet_queue(struct iowait *wait, int reason);
68 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
69 			   unsigned long len);
70 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode);
71 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
72 			 void *arg2, bool *stop);
73 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode);
74 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
75 
76 static struct mmu_rb_ops sdma_rb_ops = {
77 	.filter = sdma_rb_filter,
78 	.insert = sdma_rb_insert,
79 	.evict = sdma_rb_evict,
80 	.remove = sdma_rb_remove,
81 	.invalidate = sdma_rb_invalidate
82 };
83 
84 static int defer_packet_queue(
85 	struct sdma_engine *sde,
86 	struct iowait_work *wait,
87 	struct sdma_txreq *txreq,
88 	uint seq,
89 	bool pkts_sent)
90 {
91 	struct hfi1_user_sdma_pkt_q *pq =
92 		container_of(wait->iow, struct hfi1_user_sdma_pkt_q, busy);
93 
94 	write_seqlock(&sde->waitlock);
95 	trace_hfi1_usdma_defer(pq, sde, &pq->busy);
96 	if (sdma_progress(sde, seq, txreq))
97 		goto eagain;
98 	/*
99 	 * We are assuming that if the list is enqueued somewhere, it
100 	 * is to the dmawait list since that is the only place where
101 	 * it is supposed to be enqueued.
102 	 */
103 	xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
104 	if (list_empty(&pq->busy.list)) {
105 		pq->busy.lock = &sde->waitlock;
106 		iowait_get_priority(&pq->busy);
107 		iowait_queue(pkts_sent, &pq->busy, &sde->dmawait);
108 	}
109 	write_sequnlock(&sde->waitlock);
110 	return -EBUSY;
111 eagain:
112 	write_sequnlock(&sde->waitlock);
113 	return -EAGAIN;
114 }
115 
116 static void activate_packet_queue(struct iowait *wait, int reason)
117 {
118 	struct hfi1_user_sdma_pkt_q *pq =
119 		container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
120 
121 	trace_hfi1_usdma_activate(pq, wait, reason);
122 	xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
123 	wake_up(&wait->wait_dma);
124 };
125 
126 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt,
127 				struct hfi1_filedata *fd)
128 {
129 	int ret = -ENOMEM;
130 	char buf[64];
131 	struct hfi1_devdata *dd;
132 	struct hfi1_user_sdma_comp_q *cq;
133 	struct hfi1_user_sdma_pkt_q *pq;
134 
135 	if (!uctxt || !fd)
136 		return -EBADF;
137 
138 	if (!hfi1_sdma_comp_ring_size)
139 		return -EINVAL;
140 
141 	dd = uctxt->dd;
142 
143 	pq = kzalloc(sizeof(*pq), GFP_KERNEL);
144 	if (!pq)
145 		return -ENOMEM;
146 	pq->dd = dd;
147 	pq->ctxt = uctxt->ctxt;
148 	pq->subctxt = fd->subctxt;
149 	pq->n_max_reqs = hfi1_sdma_comp_ring_size;
150 	atomic_set(&pq->n_reqs, 0);
151 	init_waitqueue_head(&pq->wait);
152 	atomic_set(&pq->n_locked, 0);
153 
154 	iowait_init(&pq->busy, 0, NULL, NULL, defer_packet_queue,
155 		    activate_packet_queue, NULL, NULL);
156 	pq->reqidx = 0;
157 
158 	pq->reqs = kcalloc(hfi1_sdma_comp_ring_size,
159 			   sizeof(*pq->reqs),
160 			   GFP_KERNEL);
161 	if (!pq->reqs)
162 		goto pq_reqs_nomem;
163 
164 	pq->req_in_use = kcalloc(BITS_TO_LONGS(hfi1_sdma_comp_ring_size),
165 				 sizeof(*pq->req_in_use),
166 				 GFP_KERNEL);
167 	if (!pq->req_in_use)
168 		goto pq_reqs_no_in_use;
169 
170 	snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
171 		 fd->subctxt);
172 	pq->txreq_cache = kmem_cache_create(buf,
173 					    sizeof(struct user_sdma_txreq),
174 					    L1_CACHE_BYTES,
175 					    SLAB_HWCACHE_ALIGN,
176 					    NULL);
177 	if (!pq->txreq_cache) {
178 		dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
179 			   uctxt->ctxt);
180 		goto pq_txreq_nomem;
181 	}
182 
183 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
184 	if (!cq)
185 		goto cq_nomem;
186 
187 	cq->comps = vmalloc_user(PAGE_ALIGN(sizeof(*cq->comps)
188 				 * hfi1_sdma_comp_ring_size));
189 	if (!cq->comps)
190 		goto cq_comps_nomem;
191 
192 	cq->nentries = hfi1_sdma_comp_ring_size;
193 
194 	ret = hfi1_mmu_rb_register(pq, &sdma_rb_ops, dd->pport->hfi1_wq,
195 				   &pq->handler);
196 	if (ret) {
197 		dd_dev_err(dd, "Failed to register with MMU %d", ret);
198 		goto pq_mmu_fail;
199 	}
200 
201 	rcu_assign_pointer(fd->pq, pq);
202 	fd->cq = cq;
203 
204 	return 0;
205 
206 pq_mmu_fail:
207 	vfree(cq->comps);
208 cq_comps_nomem:
209 	kfree(cq);
210 cq_nomem:
211 	kmem_cache_destroy(pq->txreq_cache);
212 pq_txreq_nomem:
213 	kfree(pq->req_in_use);
214 pq_reqs_no_in_use:
215 	kfree(pq->reqs);
216 pq_reqs_nomem:
217 	kfree(pq);
218 
219 	return ret;
220 }
221 
222 static void flush_pq_iowait(struct hfi1_user_sdma_pkt_q *pq)
223 {
224 	unsigned long flags;
225 	seqlock_t *lock = pq->busy.lock;
226 
227 	if (!lock)
228 		return;
229 	write_seqlock_irqsave(lock, flags);
230 	if (!list_empty(&pq->busy.list)) {
231 		list_del_init(&pq->busy.list);
232 		pq->busy.lock = NULL;
233 	}
234 	write_sequnlock_irqrestore(lock, flags);
235 }
236 
237 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd,
238 			       struct hfi1_ctxtdata *uctxt)
239 {
240 	struct hfi1_user_sdma_pkt_q *pq;
241 
242 	trace_hfi1_sdma_user_free_queues(uctxt->dd, uctxt->ctxt, fd->subctxt);
243 
244 	spin_lock(&fd->pq_rcu_lock);
245 	pq = srcu_dereference_check(fd->pq, &fd->pq_srcu,
246 				    lockdep_is_held(&fd->pq_rcu_lock));
247 	if (pq) {
248 		rcu_assign_pointer(fd->pq, NULL);
249 		spin_unlock(&fd->pq_rcu_lock);
250 		synchronize_srcu(&fd->pq_srcu);
251 		/* at this point there can be no more new requests */
252 		if (pq->handler)
253 			hfi1_mmu_rb_unregister(pq->handler);
254 		iowait_sdma_drain(&pq->busy);
255 		/* Wait until all requests have been freed. */
256 		wait_event_interruptible(
257 			pq->wait,
258 			!atomic_read(&pq->n_reqs));
259 		kfree(pq->reqs);
260 		kfree(pq->req_in_use);
261 		kmem_cache_destroy(pq->txreq_cache);
262 		flush_pq_iowait(pq);
263 		kfree(pq);
264 	} else {
265 		spin_unlock(&fd->pq_rcu_lock);
266 	}
267 	if (fd->cq) {
268 		vfree(fd->cq->comps);
269 		kfree(fd->cq);
270 		fd->cq = NULL;
271 	}
272 	return 0;
273 }
274 
275 static u8 dlid_to_selector(u16 dlid)
276 {
277 	static u8 mapping[256];
278 	static int initialized;
279 	static u8 next;
280 	int hash;
281 
282 	if (!initialized) {
283 		memset(mapping, 0xFF, 256);
284 		initialized = 1;
285 	}
286 
287 	hash = ((dlid >> 8) ^ dlid) & 0xFF;
288 	if (mapping[hash] == 0xFF) {
289 		mapping[hash] = next;
290 		next = (next + 1) & 0x7F;
291 	}
292 
293 	return mapping[hash];
294 }
295 
296 /**
297  * hfi1_user_sdma_process_request() - Process and start a user sdma request
298  * @fd: valid file descriptor
299  * @iovec: array of io vectors to process
300  * @dim: overall iovec array size
301  * @count: number of io vector array entries processed
302  */
303 int hfi1_user_sdma_process_request(struct hfi1_filedata *fd,
304 				   struct iovec *iovec, unsigned long dim,
305 				   unsigned long *count)
306 {
307 	int ret = 0, i;
308 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
309 	struct hfi1_user_sdma_pkt_q *pq =
310 		srcu_dereference(fd->pq, &fd->pq_srcu);
311 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
312 	struct hfi1_devdata *dd = pq->dd;
313 	unsigned long idx = 0;
314 	u8 pcount = initial_pkt_count;
315 	struct sdma_req_info info;
316 	struct user_sdma_request *req;
317 	u8 opcode, sc, vl;
318 	u16 pkey;
319 	u32 slid;
320 	u16 dlid;
321 	u32 selector;
322 
323 	if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
324 		hfi1_cdbg(
325 		   SDMA,
326 		   "[%u:%u:%u] First vector not big enough for header %lu/%lu",
327 		   dd->unit, uctxt->ctxt, fd->subctxt,
328 		   iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
329 		return -EINVAL;
330 	}
331 	ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
332 	if (ret) {
333 		hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
334 			  dd->unit, uctxt->ctxt, fd->subctxt, ret);
335 		return -EFAULT;
336 	}
337 
338 	trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
339 				     (u16 *)&info);
340 	if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
341 		hfi1_cdbg(SDMA,
342 			  "[%u:%u:%u:%u] Invalid comp index",
343 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
344 		return -EINVAL;
345 	}
346 
347 	/*
348 	 * Sanity check the header io vector count.  Need at least 1 vector
349 	 * (header) and cannot be larger than the actual io vector count.
350 	 */
351 	if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
352 		hfi1_cdbg(SDMA,
353 			  "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
354 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
355 			  req_iovcnt(info.ctrl), dim);
356 		return -EINVAL;
357 	}
358 
359 	if (!info.fragsize) {
360 		hfi1_cdbg(SDMA,
361 			  "[%u:%u:%u:%u] Request does not specify fragsize",
362 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
363 		return -EINVAL;
364 	}
365 
366 	/* Try to claim the request. */
367 	if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
368 		hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
369 			  dd->unit, uctxt->ctxt, fd->subctxt,
370 			  info.comp_idx);
371 		return -EBADSLT;
372 	}
373 	/*
374 	 * All safety checks have been done and this request has been claimed.
375 	 */
376 	trace_hfi1_sdma_user_process_request(dd, uctxt->ctxt, fd->subctxt,
377 					     info.comp_idx);
378 	req = pq->reqs + info.comp_idx;
379 	req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
380 	req->data_len  = 0;
381 	req->pq = pq;
382 	req->cq = cq;
383 	req->ahg_idx = -1;
384 	req->iov_idx = 0;
385 	req->sent = 0;
386 	req->seqnum = 0;
387 	req->seqcomp = 0;
388 	req->seqsubmitted = 0;
389 	req->tids = NULL;
390 	req->has_error = 0;
391 	INIT_LIST_HEAD(&req->txps);
392 
393 	memcpy(&req->info, &info, sizeof(info));
394 
395 	/* The request is initialized, count it */
396 	atomic_inc(&pq->n_reqs);
397 
398 	if (req_opcode(info.ctrl) == EXPECTED) {
399 		/* expected must have a TID info and at least one data vector */
400 		if (req->data_iovs < 2) {
401 			SDMA_DBG(req,
402 				 "Not enough vectors for expected request");
403 			ret = -EINVAL;
404 			goto free_req;
405 		}
406 		req->data_iovs--;
407 	}
408 
409 	if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
410 		SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
411 			 MAX_VECTORS_PER_REQ);
412 		ret = -EINVAL;
413 		goto free_req;
414 	}
415 	/* Copy the header from the user buffer */
416 	ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
417 			     sizeof(req->hdr));
418 	if (ret) {
419 		SDMA_DBG(req, "Failed to copy header template (%d)", ret);
420 		ret = -EFAULT;
421 		goto free_req;
422 	}
423 
424 	/* If Static rate control is not enabled, sanitize the header. */
425 	if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
426 		req->hdr.pbc[2] = 0;
427 
428 	/* Validate the opcode. Do not trust packets from user space blindly. */
429 	opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
430 	if ((opcode & USER_OPCODE_CHECK_MASK) !=
431 	     USER_OPCODE_CHECK_VAL) {
432 		SDMA_DBG(req, "Invalid opcode (%d)", opcode);
433 		ret = -EINVAL;
434 		goto free_req;
435 	}
436 	/*
437 	 * Validate the vl. Do not trust packets from user space blindly.
438 	 * VL comes from PBC, SC comes from LRH, and the VL needs to
439 	 * match the SC look up.
440 	 */
441 	vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
442 	sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
443 	      (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
444 	if (vl >= dd->pport->vls_operational ||
445 	    vl != sc_to_vlt(dd, sc)) {
446 		SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
447 		ret = -EINVAL;
448 		goto free_req;
449 	}
450 
451 	/* Checking P_KEY for requests from user-space */
452 	pkey = (u16)be32_to_cpu(req->hdr.bth[0]);
453 	slid = be16_to_cpu(req->hdr.lrh[3]);
454 	if (egress_pkey_check(dd->pport, slid, pkey, sc, PKEY_CHECK_INVALID)) {
455 		ret = -EINVAL;
456 		goto free_req;
457 	}
458 
459 	/*
460 	 * Also should check the BTH.lnh. If it says the next header is GRH then
461 	 * the RXE parsing will be off and will land in the middle of the KDETH
462 	 * or miss it entirely.
463 	 */
464 	if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
465 		SDMA_DBG(req, "User tried to pass in a GRH");
466 		ret = -EINVAL;
467 		goto free_req;
468 	}
469 
470 	req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
471 	/*
472 	 * Calculate the initial TID offset based on the values of
473 	 * KDETH.OFFSET and KDETH.OM that are passed in.
474 	 */
475 	req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
476 		(KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
477 		 KDETH_OM_LARGE : KDETH_OM_SMALL);
478 	trace_hfi1_sdma_user_initial_tidoffset(dd, uctxt->ctxt, fd->subctxt,
479 					       info.comp_idx, req->tidoffset);
480 	idx++;
481 
482 	/* Save all the IO vector structures */
483 	for (i = 0; i < req->data_iovs; i++) {
484 		req->iovs[i].offset = 0;
485 		INIT_LIST_HEAD(&req->iovs[i].list);
486 		memcpy(&req->iovs[i].iov,
487 		       iovec + idx++,
488 		       sizeof(req->iovs[i].iov));
489 		ret = pin_vector_pages(req, &req->iovs[i]);
490 		if (ret) {
491 			req->data_iovs = i;
492 			goto free_req;
493 		}
494 		req->data_len += req->iovs[i].iov.iov_len;
495 	}
496 	trace_hfi1_sdma_user_data_length(dd, uctxt->ctxt, fd->subctxt,
497 					 info.comp_idx, req->data_len);
498 	if (pcount > req->info.npkts)
499 		pcount = req->info.npkts;
500 	/*
501 	 * Copy any TID info
502 	 * User space will provide the TID info only when the
503 	 * request type is EXPECTED. This is true even if there is
504 	 * only one packet in the request and the header is already
505 	 * setup. The reason for the singular TID case is that the
506 	 * driver needs to perform safety checks.
507 	 */
508 	if (req_opcode(req->info.ctrl) == EXPECTED) {
509 		u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
510 		u32 *tmp;
511 
512 		if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
513 			ret = -EINVAL;
514 			goto free_req;
515 		}
516 
517 		/*
518 		 * We have to copy all of the tids because they may vary
519 		 * in size and, therefore, the TID count might not be
520 		 * equal to the pkt count. However, there is no way to
521 		 * tell at this point.
522 		 */
523 		tmp = memdup_user(iovec[idx].iov_base,
524 				  ntids * sizeof(*req->tids));
525 		if (IS_ERR(tmp)) {
526 			ret = PTR_ERR(tmp);
527 			SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
528 				 ntids, ret);
529 			goto free_req;
530 		}
531 		req->tids = tmp;
532 		req->n_tids = ntids;
533 		req->tididx = 0;
534 		idx++;
535 	}
536 
537 	dlid = be16_to_cpu(req->hdr.lrh[1]);
538 	selector = dlid_to_selector(dlid);
539 	selector += uctxt->ctxt + fd->subctxt;
540 	req->sde = sdma_select_user_engine(dd, selector, vl);
541 
542 	if (!req->sde || !sdma_running(req->sde)) {
543 		ret = -ECOMM;
544 		goto free_req;
545 	}
546 
547 	/* We don't need an AHG entry if the request contains only one packet */
548 	if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG))
549 		req->ahg_idx = sdma_ahg_alloc(req->sde);
550 
551 	set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
552 	pq->state = SDMA_PKT_Q_ACTIVE;
553 
554 	/*
555 	 * This is a somewhat blocking send implementation.
556 	 * The driver will block the caller until all packets of the
557 	 * request have been submitted to the SDMA engine. However, it
558 	 * will not wait for send completions.
559 	 */
560 	while (req->seqsubmitted != req->info.npkts) {
561 		ret = user_sdma_send_pkts(req, pcount);
562 		if (ret < 0) {
563 			int we_ret;
564 
565 			if (ret != -EBUSY)
566 				goto free_req;
567 			we_ret = wait_event_interruptible_timeout(
568 				pq->busy.wait_dma,
569 				pq->state == SDMA_PKT_Q_ACTIVE,
570 				msecs_to_jiffies(
571 					SDMA_IOWAIT_TIMEOUT));
572 			trace_hfi1_usdma_we(pq, we_ret);
573 			if (we_ret <= 0)
574 				flush_pq_iowait(pq);
575 		}
576 	}
577 	*count += idx;
578 	return 0;
579 free_req:
580 	/*
581 	 * If the submitted seqsubmitted == npkts, the completion routine
582 	 * controls the final state.  If sequbmitted < npkts, wait for any
583 	 * outstanding packets to finish before cleaning up.
584 	 */
585 	if (req->seqsubmitted < req->info.npkts) {
586 		if (req->seqsubmitted)
587 			wait_event(pq->busy.wait_dma,
588 				   (req->seqcomp == req->seqsubmitted - 1));
589 		user_sdma_free_request(req, true);
590 		pq_update(pq);
591 		set_comp_state(pq, cq, info.comp_idx, ERROR, ret);
592 	}
593 	return ret;
594 }
595 
596 static inline u32 compute_data_length(struct user_sdma_request *req,
597 				      struct user_sdma_txreq *tx)
598 {
599 	/*
600 	 * Determine the proper size of the packet data.
601 	 * The size of the data of the first packet is in the header
602 	 * template. However, it includes the header and ICRC, which need
603 	 * to be subtracted.
604 	 * The minimum representable packet data length in a header is 4 bytes,
605 	 * therefore, when the data length request is less than 4 bytes, there's
606 	 * only one packet, and the packet data length is equal to that of the
607 	 * request data length.
608 	 * The size of the remaining packets is the minimum of the frag
609 	 * size (MTU) or remaining data in the request.
610 	 */
611 	u32 len;
612 
613 	if (!req->seqnum) {
614 		if (req->data_len < sizeof(u32))
615 			len = req->data_len;
616 		else
617 			len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
618 			       (sizeof(tx->hdr) - 4));
619 	} else if (req_opcode(req->info.ctrl) == EXPECTED) {
620 		u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
621 			PAGE_SIZE;
622 		/*
623 		 * Get the data length based on the remaining space in the
624 		 * TID pair.
625 		 */
626 		len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
627 		/* If we've filled up the TID pair, move to the next one. */
628 		if (unlikely(!len) && ++req->tididx < req->n_tids &&
629 		    req->tids[req->tididx]) {
630 			tidlen = EXP_TID_GET(req->tids[req->tididx],
631 					     LEN) * PAGE_SIZE;
632 			req->tidoffset = 0;
633 			len = min_t(u32, tidlen, req->info.fragsize);
634 		}
635 		/*
636 		 * Since the TID pairs map entire pages, make sure that we
637 		 * are not going to try to send more data that we have
638 		 * remaining.
639 		 */
640 		len = min(len, req->data_len - req->sent);
641 	} else {
642 		len = min(req->data_len - req->sent, (u32)req->info.fragsize);
643 	}
644 	trace_hfi1_sdma_user_compute_length(req->pq->dd,
645 					    req->pq->ctxt,
646 					    req->pq->subctxt,
647 					    req->info.comp_idx,
648 					    len);
649 	return len;
650 }
651 
652 static inline u32 pad_len(u32 len)
653 {
654 	if (len & (sizeof(u32) - 1))
655 		len += sizeof(u32) - (len & (sizeof(u32) - 1));
656 	return len;
657 }
658 
659 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
660 {
661 	/* (Size of complete header - size of PBC) + 4B ICRC + data length */
662 	return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
663 }
664 
665 static int user_sdma_txadd_ahg(struct user_sdma_request *req,
666 			       struct user_sdma_txreq *tx,
667 			       u32 datalen)
668 {
669 	int ret;
670 	u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
671 	u32 lrhlen = get_lrh_len(req->hdr, pad_len(datalen));
672 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
673 
674 	/*
675 	 * Copy the request header into the tx header
676 	 * because the HW needs a cacheline-aligned
677 	 * address.
678 	 * This copy can be optimized out if the hdr
679 	 * member of user_sdma_request were also
680 	 * cacheline aligned.
681 	 */
682 	memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
683 	if (PBC2LRH(pbclen) != lrhlen) {
684 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
685 		tx->hdr.pbc[0] = cpu_to_le16(pbclen);
686 	}
687 	ret = check_header_template(req, &tx->hdr, lrhlen, datalen);
688 	if (ret)
689 		return ret;
690 	ret = sdma_txinit_ahg(&tx->txreq, SDMA_TXREQ_F_AHG_COPY,
691 			      sizeof(tx->hdr) + datalen, req->ahg_idx,
692 			      0, NULL, 0, user_sdma_txreq_cb);
693 	if (ret)
694 		return ret;
695 	ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq, &tx->hdr, sizeof(tx->hdr));
696 	if (ret)
697 		sdma_txclean(pq->dd, &tx->txreq);
698 	return ret;
699 }
700 
701 static int user_sdma_txadd(struct user_sdma_request *req,
702 			   struct user_sdma_txreq *tx,
703 			   struct user_sdma_iovec *iovec, u32 datalen,
704 			   u32 *queued_ptr, u32 *data_sent_ptr,
705 			   u64 *iov_offset_ptr)
706 {
707 	int ret;
708 	unsigned int pageidx, len;
709 	unsigned long base, offset;
710 	u64 iov_offset = *iov_offset_ptr;
711 	u32 queued = *queued_ptr, data_sent = *data_sent_ptr;
712 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
713 
714 	base = (unsigned long)iovec->iov.iov_base;
715 	offset = offset_in_page(base + iovec->offset + iov_offset);
716 	pageidx = (((iovec->offset + iov_offset + base) - (base & PAGE_MASK)) >>
717 		   PAGE_SHIFT);
718 	len = offset + req->info.fragsize > PAGE_SIZE ?
719 		PAGE_SIZE - offset : req->info.fragsize;
720 	len = min((datalen - queued), len);
721 	ret = sdma_txadd_page(pq->dd, &tx->txreq, iovec->pages[pageidx],
722 			      offset, len);
723 	if (ret) {
724 		SDMA_DBG(req, "SDMA txreq add page failed %d\n", ret);
725 		return ret;
726 	}
727 	iov_offset += len;
728 	queued += len;
729 	data_sent += len;
730 	if (unlikely(queued < datalen && pageidx == iovec->npages &&
731 		     req->iov_idx < req->data_iovs - 1)) {
732 		iovec->offset += iov_offset;
733 		iovec = &req->iovs[++req->iov_idx];
734 		iov_offset = 0;
735 	}
736 
737 	*queued_ptr = queued;
738 	*data_sent_ptr = data_sent;
739 	*iov_offset_ptr = iov_offset;
740 	return ret;
741 }
742 
743 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts)
744 {
745 	int ret = 0;
746 	u16 count;
747 	unsigned npkts = 0;
748 	struct user_sdma_txreq *tx = NULL;
749 	struct hfi1_user_sdma_pkt_q *pq = NULL;
750 	struct user_sdma_iovec *iovec = NULL;
751 
752 	if (!req->pq)
753 		return -EINVAL;
754 
755 	pq = req->pq;
756 
757 	/* If tx completion has reported an error, we are done. */
758 	if (READ_ONCE(req->has_error))
759 		return -EFAULT;
760 
761 	/*
762 	 * Check if we might have sent the entire request already
763 	 */
764 	if (unlikely(req->seqnum == req->info.npkts)) {
765 		if (!list_empty(&req->txps))
766 			goto dosend;
767 		return ret;
768 	}
769 
770 	if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
771 		maxpkts = req->info.npkts - req->seqnum;
772 
773 	while (npkts < maxpkts) {
774 		u32 datalen = 0, queued = 0, data_sent = 0;
775 		u64 iov_offset = 0;
776 
777 		/*
778 		 * Check whether any of the completions have come back
779 		 * with errors. If so, we are not going to process any
780 		 * more packets from this request.
781 		 */
782 		if (READ_ONCE(req->has_error))
783 			return -EFAULT;
784 
785 		tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
786 		if (!tx)
787 			return -ENOMEM;
788 
789 		tx->flags = 0;
790 		tx->req = req;
791 		INIT_LIST_HEAD(&tx->list);
792 
793 		/*
794 		 * For the last packet set the ACK request
795 		 * and disable header suppression.
796 		 */
797 		if (req->seqnum == req->info.npkts - 1)
798 			tx->flags |= (TXREQ_FLAGS_REQ_ACK |
799 				      TXREQ_FLAGS_REQ_DISABLE_SH);
800 
801 		/*
802 		 * Calculate the payload size - this is min of the fragment
803 		 * (MTU) size or the remaining bytes in the request but only
804 		 * if we have payload data.
805 		 */
806 		if (req->data_len) {
807 			iovec = &req->iovs[req->iov_idx];
808 			if (READ_ONCE(iovec->offset) == iovec->iov.iov_len) {
809 				if (++req->iov_idx == req->data_iovs) {
810 					ret = -EFAULT;
811 					goto free_tx;
812 				}
813 				iovec = &req->iovs[req->iov_idx];
814 				WARN_ON(iovec->offset);
815 			}
816 
817 			datalen = compute_data_length(req, tx);
818 
819 			/*
820 			 * Disable header suppression for the payload <= 8DWS.
821 			 * If there is an uncorrectable error in the receive
822 			 * data FIFO when the received payload size is less than
823 			 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
824 			 * not reported.There is set RHF.EccErr if the header
825 			 * is not suppressed.
826 			 */
827 			if (!datalen) {
828 				SDMA_DBG(req,
829 					 "Request has data but pkt len is 0");
830 				ret = -EFAULT;
831 				goto free_tx;
832 			} else if (datalen <= 32) {
833 				tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
834 			}
835 		}
836 
837 		if (req->ahg_idx >= 0) {
838 			if (!req->seqnum) {
839 				ret = user_sdma_txadd_ahg(req, tx, datalen);
840 				if (ret)
841 					goto free_tx;
842 			} else {
843 				int changes;
844 
845 				changes = set_txreq_header_ahg(req, tx,
846 							       datalen);
847 				if (changes < 0) {
848 					ret = changes;
849 					goto free_tx;
850 				}
851 			}
852 		} else {
853 			ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
854 					  datalen, user_sdma_txreq_cb);
855 			if (ret)
856 				goto free_tx;
857 			/*
858 			 * Modify the header for this packet. This only needs
859 			 * to be done if we are not going to use AHG. Otherwise,
860 			 * the HW will do it based on the changes we gave it
861 			 * during sdma_txinit_ahg().
862 			 */
863 			ret = set_txreq_header(req, tx, datalen);
864 			if (ret)
865 				goto free_txreq;
866 		}
867 
868 		/*
869 		 * If the request contains any data vectors, add up to
870 		 * fragsize bytes to the descriptor.
871 		 */
872 		while (queued < datalen &&
873 		       (req->sent + data_sent) < req->data_len) {
874 			ret = user_sdma_txadd(req, tx, iovec, datalen,
875 					      &queued, &data_sent, &iov_offset);
876 			if (ret)
877 				goto free_txreq;
878 		}
879 		/*
880 		 * The txreq was submitted successfully so we can update
881 		 * the counters.
882 		 */
883 		req->koffset += datalen;
884 		if (req_opcode(req->info.ctrl) == EXPECTED)
885 			req->tidoffset += datalen;
886 		req->sent += data_sent;
887 		if (req->data_len)
888 			iovec->offset += iov_offset;
889 		list_add_tail(&tx->txreq.list, &req->txps);
890 		/*
891 		 * It is important to increment this here as it is used to
892 		 * generate the BTH.PSN and, therefore, can't be bulk-updated
893 		 * outside of the loop.
894 		 */
895 		tx->seqnum = req->seqnum++;
896 		npkts++;
897 	}
898 dosend:
899 	ret = sdma_send_txlist(req->sde,
900 			       iowait_get_ib_work(&pq->busy),
901 			       &req->txps, &count);
902 	req->seqsubmitted += count;
903 	if (req->seqsubmitted == req->info.npkts) {
904 		/*
905 		 * The txreq has already been submitted to the HW queue
906 		 * so we can free the AHG entry now. Corruption will not
907 		 * happen due to the sequential manner in which
908 		 * descriptors are processed.
909 		 */
910 		if (req->ahg_idx >= 0)
911 			sdma_ahg_free(req->sde, req->ahg_idx);
912 	}
913 	return ret;
914 
915 free_txreq:
916 	sdma_txclean(pq->dd, &tx->txreq);
917 free_tx:
918 	kmem_cache_free(pq->txreq_cache, tx);
919 	return ret;
920 }
921 
922 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
923 {
924 	struct evict_data evict_data;
925 
926 	evict_data.cleared = 0;
927 	evict_data.target = npages;
928 	hfi1_mmu_rb_evict(pq->handler, &evict_data);
929 	return evict_data.cleared;
930 }
931 
932 static int pin_sdma_pages(struct user_sdma_request *req,
933 			  struct user_sdma_iovec *iovec,
934 			  struct sdma_mmu_node *node,
935 			  int npages)
936 {
937 	int pinned, cleared;
938 	struct page **pages;
939 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
940 
941 	pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
942 	if (!pages)
943 		return -ENOMEM;
944 	memcpy(pages, node->pages, node->npages * sizeof(*pages));
945 
946 	npages -= node->npages;
947 retry:
948 	if (!hfi1_can_pin_pages(pq->dd, current->mm,
949 				atomic_read(&pq->n_locked), npages)) {
950 		cleared = sdma_cache_evict(pq, npages);
951 		if (cleared >= npages)
952 			goto retry;
953 	}
954 	pinned = hfi1_acquire_user_pages(current->mm,
955 					 ((unsigned long)iovec->iov.iov_base +
956 					 (node->npages * PAGE_SIZE)), npages, 0,
957 					 pages + node->npages);
958 	if (pinned < 0) {
959 		kfree(pages);
960 		return pinned;
961 	}
962 	if (pinned != npages) {
963 		unpin_vector_pages(current->mm, pages, node->npages, pinned);
964 		return -EFAULT;
965 	}
966 	kfree(node->pages);
967 	node->rb.len = iovec->iov.iov_len;
968 	node->pages = pages;
969 	atomic_add(pinned, &pq->n_locked);
970 	return pinned;
971 }
972 
973 static void unpin_sdma_pages(struct sdma_mmu_node *node)
974 {
975 	if (node->npages) {
976 		unpin_vector_pages(mm_from_sdma_node(node), node->pages, 0,
977 				   node->npages);
978 		atomic_sub(node->npages, &node->pq->n_locked);
979 	}
980 }
981 
982 static int pin_vector_pages(struct user_sdma_request *req,
983 			    struct user_sdma_iovec *iovec)
984 {
985 	int ret = 0, pinned, npages;
986 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
987 	struct sdma_mmu_node *node = NULL;
988 	struct mmu_rb_node *rb_node;
989 	struct iovec *iov;
990 	bool extracted;
991 
992 	extracted =
993 		hfi1_mmu_rb_remove_unless_exact(pq->handler,
994 						(unsigned long)
995 						iovec->iov.iov_base,
996 						iovec->iov.iov_len, &rb_node);
997 	if (rb_node) {
998 		node = container_of(rb_node, struct sdma_mmu_node, rb);
999 		if (!extracted) {
1000 			atomic_inc(&node->refcount);
1001 			iovec->pages = node->pages;
1002 			iovec->npages = node->npages;
1003 			iovec->node = node;
1004 			return 0;
1005 		}
1006 	}
1007 
1008 	if (!node) {
1009 		node = kzalloc(sizeof(*node), GFP_KERNEL);
1010 		if (!node)
1011 			return -ENOMEM;
1012 
1013 		node->rb.addr = (unsigned long)iovec->iov.iov_base;
1014 		node->pq = pq;
1015 		atomic_set(&node->refcount, 0);
1016 	}
1017 
1018 	iov = &iovec->iov;
1019 	npages = num_user_pages((unsigned long)iov->iov_base, iov->iov_len);
1020 	if (node->npages < npages) {
1021 		pinned = pin_sdma_pages(req, iovec, node, npages);
1022 		if (pinned < 0) {
1023 			ret = pinned;
1024 			goto bail;
1025 		}
1026 		node->npages += pinned;
1027 		npages = node->npages;
1028 	}
1029 	iovec->pages = node->pages;
1030 	iovec->npages = npages;
1031 	iovec->node = node;
1032 
1033 	ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1034 	if (ret) {
1035 		iovec->node = NULL;
1036 		goto bail;
1037 	}
1038 	return 0;
1039 bail:
1040 	unpin_sdma_pages(node);
1041 	kfree(node);
1042 	return ret;
1043 }
1044 
1045 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1046 			       unsigned start, unsigned npages)
1047 {
1048 	hfi1_release_user_pages(mm, pages + start, npages, false);
1049 	kfree(pages);
1050 }
1051 
1052 static int check_header_template(struct user_sdma_request *req,
1053 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
1054 				 u32 datalen)
1055 {
1056 	/*
1057 	 * Perform safety checks for any type of packet:
1058 	 *    - transfer size is multiple of 64bytes
1059 	 *    - packet length is multiple of 4 bytes
1060 	 *    - packet length is not larger than MTU size
1061 	 *
1062 	 * These checks are only done for the first packet of the
1063 	 * transfer since the header is "given" to us by user space.
1064 	 * For the remainder of the packets we compute the values.
1065 	 */
1066 	if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1067 	    lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1068 		return -EINVAL;
1069 
1070 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1071 		/*
1072 		 * The header is checked only on the first packet. Furthermore,
1073 		 * we ensure that at least one TID entry is copied when the
1074 		 * request is submitted. Therefore, we don't have to verify that
1075 		 * tididx points to something sane.
1076 		 */
1077 		u32 tidval = req->tids[req->tididx],
1078 			tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1079 			tididx = EXP_TID_GET(tidval, IDX),
1080 			tidctrl = EXP_TID_GET(tidval, CTRL),
1081 			tidoff;
1082 		__le32 kval = hdr->kdeth.ver_tid_offset;
1083 
1084 		tidoff = KDETH_GET(kval, OFFSET) *
1085 			  (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1086 			   KDETH_OM_LARGE : KDETH_OM_SMALL);
1087 		/*
1088 		 * Expected receive packets have the following
1089 		 * additional checks:
1090 		 *     - offset is not larger than the TID size
1091 		 *     - TIDCtrl values match between header and TID array
1092 		 *     - TID indexes match between header and TID array
1093 		 */
1094 		if ((tidoff + datalen > tidlen) ||
1095 		    KDETH_GET(kval, TIDCTRL) != tidctrl ||
1096 		    KDETH_GET(kval, TID) != tididx)
1097 			return -EINVAL;
1098 	}
1099 	return 0;
1100 }
1101 
1102 /*
1103  * Correctly set the BTH.PSN field based on type of
1104  * transfer - eager packets can just increment the PSN but
1105  * expected packets encode generation and sequence in the
1106  * BTH.PSN field so just incrementing will result in errors.
1107  */
1108 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1109 {
1110 	u32 val = be32_to_cpu(bthpsn),
1111 		mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1112 			0xffffffull),
1113 		psn = val & mask;
1114 	if (expct)
1115 		psn = (psn & ~HFI1_KDETH_BTH_SEQ_MASK) |
1116 			((psn + frags) & HFI1_KDETH_BTH_SEQ_MASK);
1117 	else
1118 		psn = psn + frags;
1119 	return psn & mask;
1120 }
1121 
1122 static int set_txreq_header(struct user_sdma_request *req,
1123 			    struct user_sdma_txreq *tx, u32 datalen)
1124 {
1125 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1126 	struct hfi1_pkt_header *hdr = &tx->hdr;
1127 	u8 omfactor; /* KDETH.OM */
1128 	u16 pbclen;
1129 	int ret;
1130 	u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1131 
1132 	/* Copy the header template to the request before modification */
1133 	memcpy(hdr, &req->hdr, sizeof(*hdr));
1134 
1135 	/*
1136 	 * Check if the PBC and LRH length are mismatched. If so
1137 	 * adjust both in the header.
1138 	 */
1139 	pbclen = le16_to_cpu(hdr->pbc[0]);
1140 	if (PBC2LRH(pbclen) != lrhlen) {
1141 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1142 		hdr->pbc[0] = cpu_to_le16(pbclen);
1143 		hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1144 		/*
1145 		 * Third packet
1146 		 * This is the first packet in the sequence that has
1147 		 * a "static" size that can be used for the rest of
1148 		 * the packets (besides the last one).
1149 		 */
1150 		if (unlikely(req->seqnum == 2)) {
1151 			/*
1152 			 * From this point on the lengths in both the
1153 			 * PBC and LRH are the same until the last
1154 			 * packet.
1155 			 * Adjust the template so we don't have to update
1156 			 * every packet
1157 			 */
1158 			req->hdr.pbc[0] = hdr->pbc[0];
1159 			req->hdr.lrh[2] = hdr->lrh[2];
1160 		}
1161 	}
1162 	/*
1163 	 * We only have to modify the header if this is not the
1164 	 * first packet in the request. Otherwise, we use the
1165 	 * header given to us.
1166 	 */
1167 	if (unlikely(!req->seqnum)) {
1168 		ret = check_header_template(req, hdr, lrhlen, datalen);
1169 		if (ret)
1170 			return ret;
1171 		goto done;
1172 	}
1173 
1174 	hdr->bth[2] = cpu_to_be32(
1175 		set_pkt_bth_psn(hdr->bth[2],
1176 				(req_opcode(req->info.ctrl) == EXPECTED),
1177 				req->seqnum));
1178 
1179 	/* Set ACK request on last packet */
1180 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1181 		hdr->bth[2] |= cpu_to_be32(1UL << 31);
1182 
1183 	/* Set the new offset */
1184 	hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1185 	/* Expected packets have to fill in the new TID information */
1186 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1187 		tidval = req->tids[req->tididx];
1188 		/*
1189 		 * If the offset puts us at the end of the current TID,
1190 		 * advance everything.
1191 		 */
1192 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1193 					 PAGE_SIZE)) {
1194 			req->tidoffset = 0;
1195 			/*
1196 			 * Since we don't copy all the TIDs, all at once,
1197 			 * we have to check again.
1198 			 */
1199 			if (++req->tididx > req->n_tids - 1 ||
1200 			    !req->tids[req->tididx]) {
1201 				return -EINVAL;
1202 			}
1203 			tidval = req->tids[req->tididx];
1204 		}
1205 		omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1206 			KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE_SHIFT :
1207 			KDETH_OM_SMALL_SHIFT;
1208 		/* Set KDETH.TIDCtrl based on value for this TID. */
1209 		KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1210 			  EXP_TID_GET(tidval, CTRL));
1211 		/* Set KDETH.TID based on value for this TID */
1212 		KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1213 			  EXP_TID_GET(tidval, IDX));
1214 		/* Clear KDETH.SH when DISABLE_SH flag is set */
1215 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1216 			KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1217 		/*
1218 		 * Set the KDETH.OFFSET and KDETH.OM based on size of
1219 		 * transfer.
1220 		 */
1221 		trace_hfi1_sdma_user_tid_info(
1222 			pq->dd, pq->ctxt, pq->subctxt, req->info.comp_idx,
1223 			req->tidoffset, req->tidoffset >> omfactor,
1224 			omfactor != KDETH_OM_SMALL_SHIFT);
1225 		KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1226 			  req->tidoffset >> omfactor);
1227 		KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1228 			  omfactor != KDETH_OM_SMALL_SHIFT);
1229 	}
1230 done:
1231 	trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1232 				    req->info.comp_idx, hdr, tidval);
1233 	return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1234 }
1235 
1236 static int set_txreq_header_ahg(struct user_sdma_request *req,
1237 				struct user_sdma_txreq *tx, u32 datalen)
1238 {
1239 	u32 ahg[AHG_KDETH_ARRAY_SIZE];
1240 	int idx = 0;
1241 	u8 omfactor; /* KDETH.OM */
1242 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1243 	struct hfi1_pkt_header *hdr = &req->hdr;
1244 	u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1245 	u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1246 	size_t array_size = ARRAY_SIZE(ahg);
1247 
1248 	if (PBC2LRH(pbclen) != lrhlen) {
1249 		/* PBC.PbcLengthDWs */
1250 		idx = ahg_header_set(ahg, idx, array_size, 0, 0, 12,
1251 				     (__force u16)cpu_to_le16(LRH2PBC(lrhlen)));
1252 		if (idx < 0)
1253 			return idx;
1254 		/* LRH.PktLen (we need the full 16 bits due to byte swap) */
1255 		idx = ahg_header_set(ahg, idx, array_size, 3, 0, 16,
1256 				     (__force u16)cpu_to_be16(lrhlen >> 2));
1257 		if (idx < 0)
1258 			return idx;
1259 	}
1260 
1261 	/*
1262 	 * Do the common updates
1263 	 */
1264 	/* BTH.PSN and BTH.A */
1265 	val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1266 		(HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1267 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1268 		val32 |= 1UL << 31;
1269 	idx = ahg_header_set(ahg, idx, array_size, 6, 0, 16,
1270 			     (__force u16)cpu_to_be16(val32 >> 16));
1271 	if (idx < 0)
1272 		return idx;
1273 	idx = ahg_header_set(ahg, idx, array_size, 6, 16, 16,
1274 			     (__force u16)cpu_to_be16(val32 & 0xffff));
1275 	if (idx < 0)
1276 		return idx;
1277 	/* KDETH.Offset */
1278 	idx = ahg_header_set(ahg, idx, array_size, 15, 0, 16,
1279 			     (__force u16)cpu_to_le16(req->koffset & 0xffff));
1280 	if (idx < 0)
1281 		return idx;
1282 	idx = ahg_header_set(ahg, idx, array_size, 15, 16, 16,
1283 			     (__force u16)cpu_to_le16(req->koffset >> 16));
1284 	if (idx < 0)
1285 		return idx;
1286 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1287 		__le16 val;
1288 
1289 		tidval = req->tids[req->tididx];
1290 
1291 		/*
1292 		 * If the offset puts us at the end of the current TID,
1293 		 * advance everything.
1294 		 */
1295 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1296 					 PAGE_SIZE)) {
1297 			req->tidoffset = 0;
1298 			/*
1299 			 * Since we don't copy all the TIDs, all at once,
1300 			 * we have to check again.
1301 			 */
1302 			if (++req->tididx > req->n_tids - 1 ||
1303 			    !req->tids[req->tididx])
1304 				return -EINVAL;
1305 			tidval = req->tids[req->tididx];
1306 		}
1307 		omfactor = ((EXP_TID_GET(tidval, LEN) *
1308 				  PAGE_SIZE) >=
1309 				 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE_SHIFT :
1310 				 KDETH_OM_SMALL_SHIFT;
1311 		/* KDETH.OM and KDETH.OFFSET (TID) */
1312 		idx = ahg_header_set(
1313 				ahg, idx, array_size, 7, 0, 16,
1314 				((!!(omfactor - KDETH_OM_SMALL_SHIFT)) << 15 |
1315 				((req->tidoffset >> omfactor)
1316 				& 0x7fff)));
1317 		if (idx < 0)
1318 			return idx;
1319 		/* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1320 		val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1321 				   (EXP_TID_GET(tidval, IDX) & 0x3ff));
1322 
1323 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1324 			val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1325 						      INTR) <<
1326 					    AHG_KDETH_INTR_SHIFT));
1327 		} else {
1328 			val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1329 			       cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1330 			       cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1331 						      INTR) <<
1332 					     AHG_KDETH_INTR_SHIFT));
1333 		}
1334 
1335 		idx = ahg_header_set(ahg, idx, array_size,
1336 				     7, 16, 14, (__force u16)val);
1337 		if (idx < 0)
1338 			return idx;
1339 	}
1340 
1341 	trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1342 					req->info.comp_idx, req->sde->this_idx,
1343 					req->ahg_idx, ahg, idx, tidval);
1344 	sdma_txinit_ahg(&tx->txreq,
1345 			SDMA_TXREQ_F_USE_AHG,
1346 			datalen, req->ahg_idx, idx,
1347 			ahg, sizeof(req->hdr),
1348 			user_sdma_txreq_cb);
1349 
1350 	return idx;
1351 }
1352 
1353 /**
1354  * user_sdma_txreq_cb() - SDMA tx request completion callback.
1355  * @txreq: valid sdma tx request
1356  * @status: success/failure of request
1357  *
1358  * Called when the SDMA progress state machine gets notification that
1359  * the SDMA descriptors for this tx request have been processed by the
1360  * DMA engine. Called in interrupt context.
1361  * Only do work on completed sequences.
1362  */
1363 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1364 {
1365 	struct user_sdma_txreq *tx =
1366 		container_of(txreq, struct user_sdma_txreq, txreq);
1367 	struct user_sdma_request *req;
1368 	struct hfi1_user_sdma_pkt_q *pq;
1369 	struct hfi1_user_sdma_comp_q *cq;
1370 	enum hfi1_sdma_comp_state state = COMPLETE;
1371 
1372 	if (!tx->req)
1373 		return;
1374 
1375 	req = tx->req;
1376 	pq = req->pq;
1377 	cq = req->cq;
1378 
1379 	if (status != SDMA_TXREQ_S_OK) {
1380 		SDMA_DBG(req, "SDMA completion with error %d",
1381 			 status);
1382 		WRITE_ONCE(req->has_error, 1);
1383 		state = ERROR;
1384 	}
1385 
1386 	req->seqcomp = tx->seqnum;
1387 	kmem_cache_free(pq->txreq_cache, tx);
1388 
1389 	/* sequence isn't complete?  We are done */
1390 	if (req->seqcomp != req->info.npkts - 1)
1391 		return;
1392 
1393 	user_sdma_free_request(req, false);
1394 	set_comp_state(pq, cq, req->info.comp_idx, state, status);
1395 	pq_update(pq);
1396 }
1397 
1398 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1399 {
1400 	if (atomic_dec_and_test(&pq->n_reqs))
1401 		wake_up(&pq->wait);
1402 }
1403 
1404 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1405 {
1406 	int i;
1407 
1408 	if (!list_empty(&req->txps)) {
1409 		struct sdma_txreq *t, *p;
1410 
1411 		list_for_each_entry_safe(t, p, &req->txps, list) {
1412 			struct user_sdma_txreq *tx =
1413 				container_of(t, struct user_sdma_txreq, txreq);
1414 			list_del_init(&t->list);
1415 			sdma_txclean(req->pq->dd, t);
1416 			kmem_cache_free(req->pq->txreq_cache, tx);
1417 		}
1418 	}
1419 
1420 	for (i = 0; i < req->data_iovs; i++) {
1421 		struct sdma_mmu_node *node = req->iovs[i].node;
1422 
1423 		if (!node)
1424 			continue;
1425 
1426 		req->iovs[i].node = NULL;
1427 
1428 		if (unpin)
1429 			hfi1_mmu_rb_remove(req->pq->handler,
1430 					   &node->rb);
1431 		else
1432 			atomic_dec(&node->refcount);
1433 	}
1434 
1435 	kfree(req->tids);
1436 	clear_bit(req->info.comp_idx, req->pq->req_in_use);
1437 }
1438 
1439 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1440 				  struct hfi1_user_sdma_comp_q *cq,
1441 				  u16 idx, enum hfi1_sdma_comp_state state,
1442 				  int ret)
1443 {
1444 	if (state == ERROR)
1445 		cq->comps[idx].errcode = -ret;
1446 	smp_wmb(); /* make sure errcode is visible first */
1447 	cq->comps[idx].status = state;
1448 	trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1449 					idx, state, ret);
1450 }
1451 
1452 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1453 			   unsigned long len)
1454 {
1455 	return (bool)(node->addr == addr);
1456 }
1457 
1458 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1459 {
1460 	struct sdma_mmu_node *node =
1461 		container_of(mnode, struct sdma_mmu_node, rb);
1462 
1463 	atomic_inc(&node->refcount);
1464 	return 0;
1465 }
1466 
1467 /*
1468  * Return 1 to remove the node from the rb tree and call the remove op.
1469  *
1470  * Called with the rb tree lock held.
1471  */
1472 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1473 			 void *evict_arg, bool *stop)
1474 {
1475 	struct sdma_mmu_node *node =
1476 		container_of(mnode, struct sdma_mmu_node, rb);
1477 	struct evict_data *evict_data = evict_arg;
1478 
1479 	/* is this node still being used? */
1480 	if (atomic_read(&node->refcount))
1481 		return 0; /* keep this node */
1482 
1483 	/* this node will be evicted, add its pages to our count */
1484 	evict_data->cleared += node->npages;
1485 
1486 	/* have enough pages been cleared? */
1487 	if (evict_data->cleared >= evict_data->target)
1488 		*stop = true;
1489 
1490 	return 1; /* remove this node */
1491 }
1492 
1493 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1494 {
1495 	struct sdma_mmu_node *node =
1496 		container_of(mnode, struct sdma_mmu_node, rb);
1497 
1498 	unpin_sdma_pages(node);
1499 	kfree(node);
1500 }
1501 
1502 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1503 {
1504 	struct sdma_mmu_node *node =
1505 		container_of(mnode, struct sdma_mmu_node, rb);
1506 
1507 	if (!atomic_read(&node->refcount))
1508 		return 1;
1509 	return 0;
1510 }
1511