1 /*
2  * Copyright(c) 2015 - 2017 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 #include <linux/mm.h>
48 #include <linux/types.h>
49 #include <linux/device.h>
50 #include <linux/dmapool.h>
51 #include <linux/slab.h>
52 #include <linux/list.h>
53 #include <linux/highmem.h>
54 #include <linux/io.h>
55 #include <linux/uio.h>
56 #include <linux/rbtree.h>
57 #include <linux/spinlock.h>
58 #include <linux/delay.h>
59 #include <linux/kthread.h>
60 #include <linux/mmu_context.h>
61 #include <linux/module.h>
62 #include <linux/vmalloc.h>
63 #include <linux/string.h>
64 
65 #include "hfi.h"
66 #include "sdma.h"
67 #include "mmu_rb.h"
68 #include "user_sdma.h"
69 #include "verbs.h"  /* for the headers */
70 #include "common.h" /* for struct hfi1_tid_info */
71 #include "trace.h"
72 
73 static uint hfi1_sdma_comp_ring_size = 128;
74 module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO);
75 MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128");
76 
77 static unsigned initial_pkt_count = 8;
78 
79 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts);
80 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status);
81 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq);
82 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin);
83 static int pin_vector_pages(struct user_sdma_request *req,
84 			    struct user_sdma_iovec *iovec);
85 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
86 			       unsigned start, unsigned npages);
87 static int check_header_template(struct user_sdma_request *req,
88 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
89 				 u32 datalen);
90 static int set_txreq_header(struct user_sdma_request *req,
91 			    struct user_sdma_txreq *tx, u32 datalen);
92 static int set_txreq_header_ahg(struct user_sdma_request *req,
93 				struct user_sdma_txreq *tx, u32 len);
94 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
95 				  struct hfi1_user_sdma_comp_q *cq,
96 				  u16 idx, enum hfi1_sdma_comp_state state,
97 				  int ret);
98 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags);
99 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
100 
101 static int defer_packet_queue(
102 	struct sdma_engine *sde,
103 	struct iowait *wait,
104 	struct sdma_txreq *txreq,
105 	uint seq,
106 	bool pkts_sent);
107 static void activate_packet_queue(struct iowait *wait, int reason);
108 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
109 			   unsigned long len);
110 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode);
111 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
112 			 void *arg2, bool *stop);
113 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode);
114 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
115 
116 static struct mmu_rb_ops sdma_rb_ops = {
117 	.filter = sdma_rb_filter,
118 	.insert = sdma_rb_insert,
119 	.evict = sdma_rb_evict,
120 	.remove = sdma_rb_remove,
121 	.invalidate = sdma_rb_invalidate
122 };
123 
124 static int defer_packet_queue(
125 	struct sdma_engine *sde,
126 	struct iowait *wait,
127 	struct sdma_txreq *txreq,
128 	uint seq,
129 	bool pkts_sent)
130 {
131 	struct hfi1_user_sdma_pkt_q *pq =
132 		container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
133 	struct hfi1_ibdev *dev = &pq->dd->verbs_dev;
134 	struct user_sdma_txreq *tx =
135 		container_of(txreq, struct user_sdma_txreq, txreq);
136 
137 	if (sdma_progress(sde, seq, txreq)) {
138 		if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
139 			goto eagain;
140 	}
141 	/*
142 	 * We are assuming that if the list is enqueued somewhere, it
143 	 * is to the dmawait list since that is the only place where
144 	 * it is supposed to be enqueued.
145 	 */
146 	xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
147 	write_seqlock(&dev->iowait_lock);
148 	if (list_empty(&pq->busy.list))
149 		iowait_queue(pkts_sent, &pq->busy, &sde->dmawait);
150 	write_sequnlock(&dev->iowait_lock);
151 	return -EBUSY;
152 eagain:
153 	return -EAGAIN;
154 }
155 
156 static void activate_packet_queue(struct iowait *wait, int reason)
157 {
158 	struct hfi1_user_sdma_pkt_q *pq =
159 		container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
160 	xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
161 	wake_up(&wait->wait_dma);
162 };
163 
164 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt,
165 				struct hfi1_filedata *fd)
166 {
167 	int ret = -ENOMEM;
168 	char buf[64];
169 	struct hfi1_devdata *dd;
170 	struct hfi1_user_sdma_comp_q *cq;
171 	struct hfi1_user_sdma_pkt_q *pq;
172 
173 	if (!uctxt || !fd)
174 		return -EBADF;
175 
176 	if (!hfi1_sdma_comp_ring_size)
177 		return -EINVAL;
178 
179 	dd = uctxt->dd;
180 
181 	pq = kzalloc(sizeof(*pq), GFP_KERNEL);
182 	if (!pq)
183 		return -ENOMEM;
184 
185 	pq->dd = dd;
186 	pq->ctxt = uctxt->ctxt;
187 	pq->subctxt = fd->subctxt;
188 	pq->n_max_reqs = hfi1_sdma_comp_ring_size;
189 	atomic_set(&pq->n_reqs, 0);
190 	init_waitqueue_head(&pq->wait);
191 	atomic_set(&pq->n_locked, 0);
192 	pq->mm = fd->mm;
193 
194 	iowait_init(&pq->busy, 0, NULL, defer_packet_queue,
195 		    activate_packet_queue, NULL);
196 	pq->reqidx = 0;
197 
198 	pq->reqs = kcalloc(hfi1_sdma_comp_ring_size,
199 			   sizeof(*pq->reqs),
200 			   GFP_KERNEL);
201 	if (!pq->reqs)
202 		goto pq_reqs_nomem;
203 
204 	pq->req_in_use = kcalloc(BITS_TO_LONGS(hfi1_sdma_comp_ring_size),
205 				 sizeof(*pq->req_in_use),
206 				 GFP_KERNEL);
207 	if (!pq->req_in_use)
208 		goto pq_reqs_no_in_use;
209 
210 	snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
211 		 fd->subctxt);
212 	pq->txreq_cache = kmem_cache_create(buf,
213 					    sizeof(struct user_sdma_txreq),
214 					    L1_CACHE_BYTES,
215 					    SLAB_HWCACHE_ALIGN,
216 					    NULL);
217 	if (!pq->txreq_cache) {
218 		dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
219 			   uctxt->ctxt);
220 		goto pq_txreq_nomem;
221 	}
222 
223 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
224 	if (!cq)
225 		goto cq_nomem;
226 
227 	cq->comps = vmalloc_user(PAGE_ALIGN(sizeof(*cq->comps)
228 				 * hfi1_sdma_comp_ring_size));
229 	if (!cq->comps)
230 		goto cq_comps_nomem;
231 
232 	cq->nentries = hfi1_sdma_comp_ring_size;
233 
234 	ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq,
235 				   &pq->handler);
236 	if (ret) {
237 		dd_dev_err(dd, "Failed to register with MMU %d", ret);
238 		goto pq_mmu_fail;
239 	}
240 
241 	fd->pq = pq;
242 	fd->cq = cq;
243 
244 	return 0;
245 
246 pq_mmu_fail:
247 	vfree(cq->comps);
248 cq_comps_nomem:
249 	kfree(cq);
250 cq_nomem:
251 	kmem_cache_destroy(pq->txreq_cache);
252 pq_txreq_nomem:
253 	kfree(pq->req_in_use);
254 pq_reqs_no_in_use:
255 	kfree(pq->reqs);
256 pq_reqs_nomem:
257 	kfree(pq);
258 
259 	return ret;
260 }
261 
262 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd,
263 			       struct hfi1_ctxtdata *uctxt)
264 {
265 	struct hfi1_user_sdma_pkt_q *pq;
266 
267 	trace_hfi1_sdma_user_free_queues(uctxt->dd, uctxt->ctxt, fd->subctxt);
268 
269 	pq = fd->pq;
270 	if (pq) {
271 		if (pq->handler)
272 			hfi1_mmu_rb_unregister(pq->handler);
273 		iowait_sdma_drain(&pq->busy);
274 		/* Wait until all requests have been freed. */
275 		wait_event_interruptible(
276 			pq->wait,
277 			!atomic_read(&pq->n_reqs));
278 		kfree(pq->reqs);
279 		kfree(pq->req_in_use);
280 		kmem_cache_destroy(pq->txreq_cache);
281 		kfree(pq);
282 		fd->pq = NULL;
283 	}
284 	if (fd->cq) {
285 		vfree(fd->cq->comps);
286 		kfree(fd->cq);
287 		fd->cq = NULL;
288 	}
289 	return 0;
290 }
291 
292 static u8 dlid_to_selector(u16 dlid)
293 {
294 	static u8 mapping[256];
295 	static int initialized;
296 	static u8 next;
297 	int hash;
298 
299 	if (!initialized) {
300 		memset(mapping, 0xFF, 256);
301 		initialized = 1;
302 	}
303 
304 	hash = ((dlid >> 8) ^ dlid) & 0xFF;
305 	if (mapping[hash] == 0xFF) {
306 		mapping[hash] = next;
307 		next = (next + 1) & 0x7F;
308 	}
309 
310 	return mapping[hash];
311 }
312 
313 /**
314  * hfi1_user_sdma_process_request() - Process and start a user sdma request
315  * @fd: valid file descriptor
316  * @iovec: array of io vectors to process
317  * @dim: overall iovec array size
318  * @count: number of io vector array entries processed
319  */
320 int hfi1_user_sdma_process_request(struct hfi1_filedata *fd,
321 				   struct iovec *iovec, unsigned long dim,
322 				   unsigned long *count)
323 {
324 	int ret = 0, i;
325 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
326 	struct hfi1_user_sdma_pkt_q *pq = fd->pq;
327 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
328 	struct hfi1_devdata *dd = pq->dd;
329 	unsigned long idx = 0;
330 	u8 pcount = initial_pkt_count;
331 	struct sdma_req_info info;
332 	struct user_sdma_request *req;
333 	u8 opcode, sc, vl;
334 	u16 pkey;
335 	u32 slid;
336 	u16 dlid;
337 	u32 selector;
338 
339 	if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
340 		hfi1_cdbg(
341 		   SDMA,
342 		   "[%u:%u:%u] First vector not big enough for header %lu/%lu",
343 		   dd->unit, uctxt->ctxt, fd->subctxt,
344 		   iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
345 		return -EINVAL;
346 	}
347 	ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
348 	if (ret) {
349 		hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
350 			  dd->unit, uctxt->ctxt, fd->subctxt, ret);
351 		return -EFAULT;
352 	}
353 
354 	trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
355 				     (u16 *)&info);
356 	if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
357 		hfi1_cdbg(SDMA,
358 			  "[%u:%u:%u:%u] Invalid comp index",
359 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
360 		return -EINVAL;
361 	}
362 
363 	/*
364 	 * Sanity check the header io vector count.  Need at least 1 vector
365 	 * (header) and cannot be larger than the actual io vector count.
366 	 */
367 	if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
368 		hfi1_cdbg(SDMA,
369 			  "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
370 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
371 			  req_iovcnt(info.ctrl), dim);
372 		return -EINVAL;
373 	}
374 
375 	if (!info.fragsize) {
376 		hfi1_cdbg(SDMA,
377 			  "[%u:%u:%u:%u] Request does not specify fragsize",
378 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
379 		return -EINVAL;
380 	}
381 
382 	/* Try to claim the request. */
383 	if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
384 		hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
385 			  dd->unit, uctxt->ctxt, fd->subctxt,
386 			  info.comp_idx);
387 		return -EBADSLT;
388 	}
389 	/*
390 	 * All safety checks have been done and this request has been claimed.
391 	 */
392 	trace_hfi1_sdma_user_process_request(dd, uctxt->ctxt, fd->subctxt,
393 					     info.comp_idx);
394 	req = pq->reqs + info.comp_idx;
395 	req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
396 	req->data_len  = 0;
397 	req->pq = pq;
398 	req->cq = cq;
399 	req->ahg_idx = -1;
400 	req->iov_idx = 0;
401 	req->sent = 0;
402 	req->seqnum = 0;
403 	req->seqcomp = 0;
404 	req->seqsubmitted = 0;
405 	req->tids = NULL;
406 	req->has_error = 0;
407 	INIT_LIST_HEAD(&req->txps);
408 
409 	memcpy(&req->info, &info, sizeof(info));
410 
411 	/* The request is initialized, count it */
412 	atomic_inc(&pq->n_reqs);
413 
414 	if (req_opcode(info.ctrl) == EXPECTED) {
415 		/* expected must have a TID info and at least one data vector */
416 		if (req->data_iovs < 2) {
417 			SDMA_DBG(req,
418 				 "Not enough vectors for expected request");
419 			ret = -EINVAL;
420 			goto free_req;
421 		}
422 		req->data_iovs--;
423 	}
424 
425 	if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
426 		SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
427 			 MAX_VECTORS_PER_REQ);
428 		ret = -EINVAL;
429 		goto free_req;
430 	}
431 	/* Copy the header from the user buffer */
432 	ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
433 			     sizeof(req->hdr));
434 	if (ret) {
435 		SDMA_DBG(req, "Failed to copy header template (%d)", ret);
436 		ret = -EFAULT;
437 		goto free_req;
438 	}
439 
440 	/* If Static rate control is not enabled, sanitize the header. */
441 	if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
442 		req->hdr.pbc[2] = 0;
443 
444 	/* Validate the opcode. Do not trust packets from user space blindly. */
445 	opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
446 	if ((opcode & USER_OPCODE_CHECK_MASK) !=
447 	     USER_OPCODE_CHECK_VAL) {
448 		SDMA_DBG(req, "Invalid opcode (%d)", opcode);
449 		ret = -EINVAL;
450 		goto free_req;
451 	}
452 	/*
453 	 * Validate the vl. Do not trust packets from user space blindly.
454 	 * VL comes from PBC, SC comes from LRH, and the VL needs to
455 	 * match the SC look up.
456 	 */
457 	vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
458 	sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
459 	      (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
460 	if (vl >= dd->pport->vls_operational ||
461 	    vl != sc_to_vlt(dd, sc)) {
462 		SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
463 		ret = -EINVAL;
464 		goto free_req;
465 	}
466 
467 	/* Checking P_KEY for requests from user-space */
468 	pkey = (u16)be32_to_cpu(req->hdr.bth[0]);
469 	slid = be16_to_cpu(req->hdr.lrh[3]);
470 	if (egress_pkey_check(dd->pport, slid, pkey, sc, PKEY_CHECK_INVALID)) {
471 		ret = -EINVAL;
472 		goto free_req;
473 	}
474 
475 	/*
476 	 * Also should check the BTH.lnh. If it says the next header is GRH then
477 	 * the RXE parsing will be off and will land in the middle of the KDETH
478 	 * or miss it entirely.
479 	 */
480 	if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
481 		SDMA_DBG(req, "User tried to pass in a GRH");
482 		ret = -EINVAL;
483 		goto free_req;
484 	}
485 
486 	req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
487 	/*
488 	 * Calculate the initial TID offset based on the values of
489 	 * KDETH.OFFSET and KDETH.OM that are passed in.
490 	 */
491 	req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
492 		(KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
493 		 KDETH_OM_LARGE : KDETH_OM_SMALL);
494 	trace_hfi1_sdma_user_initial_tidoffset(dd, uctxt->ctxt, fd->subctxt,
495 					       info.comp_idx, req->tidoffset);
496 	idx++;
497 
498 	/* Save all the IO vector structures */
499 	for (i = 0; i < req->data_iovs; i++) {
500 		req->iovs[i].offset = 0;
501 		INIT_LIST_HEAD(&req->iovs[i].list);
502 		memcpy(&req->iovs[i].iov,
503 		       iovec + idx++,
504 		       sizeof(req->iovs[i].iov));
505 		ret = pin_vector_pages(req, &req->iovs[i]);
506 		if (ret) {
507 			req->data_iovs = i;
508 			goto free_req;
509 		}
510 		req->data_len += req->iovs[i].iov.iov_len;
511 	}
512 	trace_hfi1_sdma_user_data_length(dd, uctxt->ctxt, fd->subctxt,
513 					 info.comp_idx, req->data_len);
514 	if (pcount > req->info.npkts)
515 		pcount = req->info.npkts;
516 	/*
517 	 * Copy any TID info
518 	 * User space will provide the TID info only when the
519 	 * request type is EXPECTED. This is true even if there is
520 	 * only one packet in the request and the header is already
521 	 * setup. The reason for the singular TID case is that the
522 	 * driver needs to perform safety checks.
523 	 */
524 	if (req_opcode(req->info.ctrl) == EXPECTED) {
525 		u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
526 		u32 *tmp;
527 
528 		if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
529 			ret = -EINVAL;
530 			goto free_req;
531 		}
532 
533 		/*
534 		 * We have to copy all of the tids because they may vary
535 		 * in size and, therefore, the TID count might not be
536 		 * equal to the pkt count. However, there is no way to
537 		 * tell at this point.
538 		 */
539 		tmp = memdup_user(iovec[idx].iov_base,
540 				  ntids * sizeof(*req->tids));
541 		if (IS_ERR(tmp)) {
542 			ret = PTR_ERR(tmp);
543 			SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
544 				 ntids, ret);
545 			goto free_req;
546 		}
547 		req->tids = tmp;
548 		req->n_tids = ntids;
549 		req->tididx = 0;
550 		idx++;
551 	}
552 
553 	dlid = be16_to_cpu(req->hdr.lrh[1]);
554 	selector = dlid_to_selector(dlid);
555 	selector += uctxt->ctxt + fd->subctxt;
556 	req->sde = sdma_select_user_engine(dd, selector, vl);
557 
558 	if (!req->sde || !sdma_running(req->sde)) {
559 		ret = -ECOMM;
560 		goto free_req;
561 	}
562 
563 	/* We don't need an AHG entry if the request contains only one packet */
564 	if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG))
565 		req->ahg_idx = sdma_ahg_alloc(req->sde);
566 
567 	set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
568 	pq->state = SDMA_PKT_Q_ACTIVE;
569 	/* Send the first N packets in the request to buy us some time */
570 	ret = user_sdma_send_pkts(req, pcount);
571 	if (unlikely(ret < 0 && ret != -EBUSY))
572 		goto free_req;
573 
574 	/*
575 	 * This is a somewhat blocking send implementation.
576 	 * The driver will block the caller until all packets of the
577 	 * request have been submitted to the SDMA engine. However, it
578 	 * will not wait for send completions.
579 	 */
580 	while (req->seqsubmitted != req->info.npkts) {
581 		ret = user_sdma_send_pkts(req, pcount);
582 		if (ret < 0) {
583 			if (ret != -EBUSY)
584 				goto free_req;
585 			wait_event_interruptible_timeout(
586 				pq->busy.wait_dma,
587 				(pq->state == SDMA_PKT_Q_ACTIVE),
588 				msecs_to_jiffies(
589 					SDMA_IOWAIT_TIMEOUT));
590 		}
591 	}
592 	*count += idx;
593 	return 0;
594 free_req:
595 	/*
596 	 * If the submitted seqsubmitted == npkts, the completion routine
597 	 * controls the final state.  If sequbmitted < npkts, wait for any
598 	 * outstanding packets to finish before cleaning up.
599 	 */
600 	if (req->seqsubmitted < req->info.npkts) {
601 		if (req->seqsubmitted)
602 			wait_event(pq->busy.wait_dma,
603 				   (req->seqcomp == req->seqsubmitted - 1));
604 		user_sdma_free_request(req, true);
605 		pq_update(pq);
606 		set_comp_state(pq, cq, info.comp_idx, ERROR, ret);
607 	}
608 	return ret;
609 }
610 
611 static inline u32 compute_data_length(struct user_sdma_request *req,
612 				      struct user_sdma_txreq *tx)
613 {
614 	/*
615 	 * Determine the proper size of the packet data.
616 	 * The size of the data of the first packet is in the header
617 	 * template. However, it includes the header and ICRC, which need
618 	 * to be subtracted.
619 	 * The minimum representable packet data length in a header is 4 bytes,
620 	 * therefore, when the data length request is less than 4 bytes, there's
621 	 * only one packet, and the packet data length is equal to that of the
622 	 * request data length.
623 	 * The size of the remaining packets is the minimum of the frag
624 	 * size (MTU) or remaining data in the request.
625 	 */
626 	u32 len;
627 
628 	if (!req->seqnum) {
629 		if (req->data_len < sizeof(u32))
630 			len = req->data_len;
631 		else
632 			len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
633 			       (sizeof(tx->hdr) - 4));
634 	} else if (req_opcode(req->info.ctrl) == EXPECTED) {
635 		u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
636 			PAGE_SIZE;
637 		/*
638 		 * Get the data length based on the remaining space in the
639 		 * TID pair.
640 		 */
641 		len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
642 		/* If we've filled up the TID pair, move to the next one. */
643 		if (unlikely(!len) && ++req->tididx < req->n_tids &&
644 		    req->tids[req->tididx]) {
645 			tidlen = EXP_TID_GET(req->tids[req->tididx],
646 					     LEN) * PAGE_SIZE;
647 			req->tidoffset = 0;
648 			len = min_t(u32, tidlen, req->info.fragsize);
649 		}
650 		/*
651 		 * Since the TID pairs map entire pages, make sure that we
652 		 * are not going to try to send more data that we have
653 		 * remaining.
654 		 */
655 		len = min(len, req->data_len - req->sent);
656 	} else {
657 		len = min(req->data_len - req->sent, (u32)req->info.fragsize);
658 	}
659 	trace_hfi1_sdma_user_compute_length(req->pq->dd,
660 					    req->pq->ctxt,
661 					    req->pq->subctxt,
662 					    req->info.comp_idx,
663 					    len);
664 	return len;
665 }
666 
667 static inline u32 pad_len(u32 len)
668 {
669 	if (len & (sizeof(u32) - 1))
670 		len += sizeof(u32) - (len & (sizeof(u32) - 1));
671 	return len;
672 }
673 
674 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
675 {
676 	/* (Size of complete header - size of PBC) + 4B ICRC + data length */
677 	return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
678 }
679 
680 static int user_sdma_txadd_ahg(struct user_sdma_request *req,
681 			       struct user_sdma_txreq *tx,
682 			       u32 datalen)
683 {
684 	int ret;
685 	u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
686 	u32 lrhlen = get_lrh_len(req->hdr, pad_len(datalen));
687 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
688 
689 	/*
690 	 * Copy the request header into the tx header
691 	 * because the HW needs a cacheline-aligned
692 	 * address.
693 	 * This copy can be optimized out if the hdr
694 	 * member of user_sdma_request were also
695 	 * cacheline aligned.
696 	 */
697 	memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
698 	if (PBC2LRH(pbclen) != lrhlen) {
699 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
700 		tx->hdr.pbc[0] = cpu_to_le16(pbclen);
701 	}
702 	ret = check_header_template(req, &tx->hdr, lrhlen, datalen);
703 	if (ret)
704 		return ret;
705 	ret = sdma_txinit_ahg(&tx->txreq, SDMA_TXREQ_F_AHG_COPY,
706 			      sizeof(tx->hdr) + datalen, req->ahg_idx,
707 			      0, NULL, 0, user_sdma_txreq_cb);
708 	if (ret)
709 		return ret;
710 	ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq, &tx->hdr, sizeof(tx->hdr));
711 	if (ret)
712 		sdma_txclean(pq->dd, &tx->txreq);
713 	return ret;
714 }
715 
716 static int user_sdma_txadd(struct user_sdma_request *req,
717 			   struct user_sdma_txreq *tx,
718 			   struct user_sdma_iovec *iovec, u32 datalen,
719 			   u32 *queued_ptr, u32 *data_sent_ptr,
720 			   u64 *iov_offset_ptr)
721 {
722 	int ret;
723 	unsigned int pageidx, len;
724 	unsigned long base, offset;
725 	u64 iov_offset = *iov_offset_ptr;
726 	u32 queued = *queued_ptr, data_sent = *data_sent_ptr;
727 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
728 
729 	base = (unsigned long)iovec->iov.iov_base;
730 	offset = offset_in_page(base + iovec->offset + iov_offset);
731 	pageidx = (((iovec->offset + iov_offset + base) - (base & PAGE_MASK)) >>
732 		   PAGE_SHIFT);
733 	len = offset + req->info.fragsize > PAGE_SIZE ?
734 		PAGE_SIZE - offset : req->info.fragsize;
735 	len = min((datalen - queued), len);
736 	ret = sdma_txadd_page(pq->dd, &tx->txreq, iovec->pages[pageidx],
737 			      offset, len);
738 	if (ret) {
739 		SDMA_DBG(req, "SDMA txreq add page failed %d\n", ret);
740 		return ret;
741 	}
742 	iov_offset += len;
743 	queued += len;
744 	data_sent += len;
745 	if (unlikely(queued < datalen && pageidx == iovec->npages &&
746 		     req->iov_idx < req->data_iovs - 1)) {
747 		iovec->offset += iov_offset;
748 		iovec = &req->iovs[++req->iov_idx];
749 		iov_offset = 0;
750 	}
751 
752 	*queued_ptr = queued;
753 	*data_sent_ptr = data_sent;
754 	*iov_offset_ptr = iov_offset;
755 	return ret;
756 }
757 
758 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts)
759 {
760 	int ret = 0;
761 	u16 count;
762 	unsigned npkts = 0;
763 	struct user_sdma_txreq *tx = NULL;
764 	struct hfi1_user_sdma_pkt_q *pq = NULL;
765 	struct user_sdma_iovec *iovec = NULL;
766 
767 	if (!req->pq)
768 		return -EINVAL;
769 
770 	pq = req->pq;
771 
772 	/* If tx completion has reported an error, we are done. */
773 	if (READ_ONCE(req->has_error))
774 		return -EFAULT;
775 
776 	/*
777 	 * Check if we might have sent the entire request already
778 	 */
779 	if (unlikely(req->seqnum == req->info.npkts)) {
780 		if (!list_empty(&req->txps))
781 			goto dosend;
782 		return ret;
783 	}
784 
785 	if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
786 		maxpkts = req->info.npkts - req->seqnum;
787 
788 	while (npkts < maxpkts) {
789 		u32 datalen = 0, queued = 0, data_sent = 0;
790 		u64 iov_offset = 0;
791 
792 		/*
793 		 * Check whether any of the completions have come back
794 		 * with errors. If so, we are not going to process any
795 		 * more packets from this request.
796 		 */
797 		if (READ_ONCE(req->has_error))
798 			return -EFAULT;
799 
800 		tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
801 		if (!tx)
802 			return -ENOMEM;
803 
804 		tx->flags = 0;
805 		tx->req = req;
806 		tx->busycount = 0;
807 		INIT_LIST_HEAD(&tx->list);
808 
809 		/*
810 		 * For the last packet set the ACK request
811 		 * and disable header suppression.
812 		 */
813 		if (req->seqnum == req->info.npkts - 1)
814 			tx->flags |= (TXREQ_FLAGS_REQ_ACK |
815 				      TXREQ_FLAGS_REQ_DISABLE_SH);
816 
817 		/*
818 		 * Calculate the payload size - this is min of the fragment
819 		 * (MTU) size or the remaining bytes in the request but only
820 		 * if we have payload data.
821 		 */
822 		if (req->data_len) {
823 			iovec = &req->iovs[req->iov_idx];
824 			if (READ_ONCE(iovec->offset) == iovec->iov.iov_len) {
825 				if (++req->iov_idx == req->data_iovs) {
826 					ret = -EFAULT;
827 					goto free_txreq;
828 				}
829 				iovec = &req->iovs[req->iov_idx];
830 				WARN_ON(iovec->offset);
831 			}
832 
833 			datalen = compute_data_length(req, tx);
834 
835 			/*
836 			 * Disable header suppression for the payload <= 8DWS.
837 			 * If there is an uncorrectable error in the receive
838 			 * data FIFO when the received payload size is less than
839 			 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
840 			 * not reported.There is set RHF.EccErr if the header
841 			 * is not suppressed.
842 			 */
843 			if (!datalen) {
844 				SDMA_DBG(req,
845 					 "Request has data but pkt len is 0");
846 				ret = -EFAULT;
847 				goto free_tx;
848 			} else if (datalen <= 32) {
849 				tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
850 			}
851 		}
852 
853 		if (req->ahg_idx >= 0) {
854 			if (!req->seqnum) {
855 				ret = user_sdma_txadd_ahg(req, tx, datalen);
856 				if (ret)
857 					goto free_tx;
858 			} else {
859 				int changes;
860 
861 				changes = set_txreq_header_ahg(req, tx,
862 							       datalen);
863 				if (changes < 0)
864 					goto free_tx;
865 			}
866 		} else {
867 			ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
868 					  datalen, user_sdma_txreq_cb);
869 			if (ret)
870 				goto free_tx;
871 			/*
872 			 * Modify the header for this packet. This only needs
873 			 * to be done if we are not going to use AHG. Otherwise,
874 			 * the HW will do it based on the changes we gave it
875 			 * during sdma_txinit_ahg().
876 			 */
877 			ret = set_txreq_header(req, tx, datalen);
878 			if (ret)
879 				goto free_txreq;
880 		}
881 
882 		/*
883 		 * If the request contains any data vectors, add up to
884 		 * fragsize bytes to the descriptor.
885 		 */
886 		while (queued < datalen &&
887 		       (req->sent + data_sent) < req->data_len) {
888 			ret = user_sdma_txadd(req, tx, iovec, datalen,
889 					      &queued, &data_sent, &iov_offset);
890 			if (ret)
891 				goto free_txreq;
892 		}
893 		/*
894 		 * The txreq was submitted successfully so we can update
895 		 * the counters.
896 		 */
897 		req->koffset += datalen;
898 		if (req_opcode(req->info.ctrl) == EXPECTED)
899 			req->tidoffset += datalen;
900 		req->sent += data_sent;
901 		if (req->data_len)
902 			iovec->offset += iov_offset;
903 		list_add_tail(&tx->txreq.list, &req->txps);
904 		/*
905 		 * It is important to increment this here as it is used to
906 		 * generate the BTH.PSN and, therefore, can't be bulk-updated
907 		 * outside of the loop.
908 		 */
909 		tx->seqnum = req->seqnum++;
910 		npkts++;
911 	}
912 dosend:
913 	ret = sdma_send_txlist(req->sde, &pq->busy, &req->txps, &count);
914 	req->seqsubmitted += count;
915 	if (req->seqsubmitted == req->info.npkts) {
916 		/*
917 		 * The txreq has already been submitted to the HW queue
918 		 * so we can free the AHG entry now. Corruption will not
919 		 * happen due to the sequential manner in which
920 		 * descriptors are processed.
921 		 */
922 		if (req->ahg_idx >= 0)
923 			sdma_ahg_free(req->sde, req->ahg_idx);
924 	}
925 	return ret;
926 
927 free_txreq:
928 	sdma_txclean(pq->dd, &tx->txreq);
929 free_tx:
930 	kmem_cache_free(pq->txreq_cache, tx);
931 	return ret;
932 }
933 
934 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
935 {
936 	struct evict_data evict_data;
937 
938 	evict_data.cleared = 0;
939 	evict_data.target = npages;
940 	hfi1_mmu_rb_evict(pq->handler, &evict_data);
941 	return evict_data.cleared;
942 }
943 
944 static int pin_sdma_pages(struct user_sdma_request *req,
945 			  struct user_sdma_iovec *iovec,
946 			  struct sdma_mmu_node *node,
947 			  int npages)
948 {
949 	int pinned, cleared;
950 	struct page **pages;
951 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
952 
953 	pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
954 	if (!pages)
955 		return -ENOMEM;
956 	memcpy(pages, node->pages, node->npages * sizeof(*pages));
957 
958 	npages -= node->npages;
959 retry:
960 	if (!hfi1_can_pin_pages(pq->dd, pq->mm,
961 				atomic_read(&pq->n_locked), npages)) {
962 		cleared = sdma_cache_evict(pq, npages);
963 		if (cleared >= npages)
964 			goto retry;
965 	}
966 	pinned = hfi1_acquire_user_pages(pq->mm,
967 					 ((unsigned long)iovec->iov.iov_base +
968 					 (node->npages * PAGE_SIZE)), npages, 0,
969 					 pages + node->npages);
970 	if (pinned < 0) {
971 		kfree(pages);
972 		return pinned;
973 	}
974 	if (pinned != npages) {
975 		unpin_vector_pages(pq->mm, pages, node->npages, pinned);
976 		return -EFAULT;
977 	}
978 	kfree(node->pages);
979 	node->rb.len = iovec->iov.iov_len;
980 	node->pages = pages;
981 	atomic_add(pinned, &pq->n_locked);
982 	return pinned;
983 }
984 
985 static void unpin_sdma_pages(struct sdma_mmu_node *node)
986 {
987 	if (node->npages) {
988 		unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages);
989 		atomic_sub(node->npages, &node->pq->n_locked);
990 	}
991 }
992 
993 static int pin_vector_pages(struct user_sdma_request *req,
994 			    struct user_sdma_iovec *iovec)
995 {
996 	int ret = 0, pinned, npages;
997 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
998 	struct sdma_mmu_node *node = NULL;
999 	struct mmu_rb_node *rb_node;
1000 	struct iovec *iov;
1001 	bool extracted;
1002 
1003 	extracted =
1004 		hfi1_mmu_rb_remove_unless_exact(pq->handler,
1005 						(unsigned long)
1006 						iovec->iov.iov_base,
1007 						iovec->iov.iov_len, &rb_node);
1008 	if (rb_node) {
1009 		node = container_of(rb_node, struct sdma_mmu_node, rb);
1010 		if (!extracted) {
1011 			atomic_inc(&node->refcount);
1012 			iovec->pages = node->pages;
1013 			iovec->npages = node->npages;
1014 			iovec->node = node;
1015 			return 0;
1016 		}
1017 	}
1018 
1019 	if (!node) {
1020 		node = kzalloc(sizeof(*node), GFP_KERNEL);
1021 		if (!node)
1022 			return -ENOMEM;
1023 
1024 		node->rb.addr = (unsigned long)iovec->iov.iov_base;
1025 		node->pq = pq;
1026 		atomic_set(&node->refcount, 0);
1027 	}
1028 
1029 	iov = &iovec->iov;
1030 	npages = num_user_pages((unsigned long)iov->iov_base, iov->iov_len);
1031 	if (node->npages < npages) {
1032 		pinned = pin_sdma_pages(req, iovec, node, npages);
1033 		if (pinned < 0) {
1034 			ret = pinned;
1035 			goto bail;
1036 		}
1037 		node->npages += pinned;
1038 		npages = node->npages;
1039 	}
1040 	iovec->pages = node->pages;
1041 	iovec->npages = npages;
1042 	iovec->node = node;
1043 
1044 	ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1045 	if (ret) {
1046 		iovec->node = NULL;
1047 		goto bail;
1048 	}
1049 	return 0;
1050 bail:
1051 	unpin_sdma_pages(node);
1052 	kfree(node);
1053 	return ret;
1054 }
1055 
1056 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1057 			       unsigned start, unsigned npages)
1058 {
1059 	hfi1_release_user_pages(mm, pages + start, npages, false);
1060 	kfree(pages);
1061 }
1062 
1063 static int check_header_template(struct user_sdma_request *req,
1064 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
1065 				 u32 datalen)
1066 {
1067 	/*
1068 	 * Perform safety checks for any type of packet:
1069 	 *    - transfer size is multiple of 64bytes
1070 	 *    - packet length is multiple of 4 bytes
1071 	 *    - packet length is not larger than MTU size
1072 	 *
1073 	 * These checks are only done for the first packet of the
1074 	 * transfer since the header is "given" to us by user space.
1075 	 * For the remainder of the packets we compute the values.
1076 	 */
1077 	if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1078 	    lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1079 		return -EINVAL;
1080 
1081 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1082 		/*
1083 		 * The header is checked only on the first packet. Furthermore,
1084 		 * we ensure that at least one TID entry is copied when the
1085 		 * request is submitted. Therefore, we don't have to verify that
1086 		 * tididx points to something sane.
1087 		 */
1088 		u32 tidval = req->tids[req->tididx],
1089 			tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1090 			tididx = EXP_TID_GET(tidval, IDX),
1091 			tidctrl = EXP_TID_GET(tidval, CTRL),
1092 			tidoff;
1093 		__le32 kval = hdr->kdeth.ver_tid_offset;
1094 
1095 		tidoff = KDETH_GET(kval, OFFSET) *
1096 			  (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1097 			   KDETH_OM_LARGE : KDETH_OM_SMALL);
1098 		/*
1099 		 * Expected receive packets have the following
1100 		 * additional checks:
1101 		 *     - offset is not larger than the TID size
1102 		 *     - TIDCtrl values match between header and TID array
1103 		 *     - TID indexes match between header and TID array
1104 		 */
1105 		if ((tidoff + datalen > tidlen) ||
1106 		    KDETH_GET(kval, TIDCTRL) != tidctrl ||
1107 		    KDETH_GET(kval, TID) != tididx)
1108 			return -EINVAL;
1109 	}
1110 	return 0;
1111 }
1112 
1113 /*
1114  * Correctly set the BTH.PSN field based on type of
1115  * transfer - eager packets can just increment the PSN but
1116  * expected packets encode generation and sequence in the
1117  * BTH.PSN field so just incrementing will result in errors.
1118  */
1119 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1120 {
1121 	u32 val = be32_to_cpu(bthpsn),
1122 		mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1123 			0xffffffull),
1124 		psn = val & mask;
1125 	if (expct)
1126 		psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
1127 	else
1128 		psn = psn + frags;
1129 	return psn & mask;
1130 }
1131 
1132 static int set_txreq_header(struct user_sdma_request *req,
1133 			    struct user_sdma_txreq *tx, u32 datalen)
1134 {
1135 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1136 	struct hfi1_pkt_header *hdr = &tx->hdr;
1137 	u8 omfactor; /* KDETH.OM */
1138 	u16 pbclen;
1139 	int ret;
1140 	u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1141 
1142 	/* Copy the header template to the request before modification */
1143 	memcpy(hdr, &req->hdr, sizeof(*hdr));
1144 
1145 	/*
1146 	 * Check if the PBC and LRH length are mismatched. If so
1147 	 * adjust both in the header.
1148 	 */
1149 	pbclen = le16_to_cpu(hdr->pbc[0]);
1150 	if (PBC2LRH(pbclen) != lrhlen) {
1151 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1152 		hdr->pbc[0] = cpu_to_le16(pbclen);
1153 		hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1154 		/*
1155 		 * Third packet
1156 		 * This is the first packet in the sequence that has
1157 		 * a "static" size that can be used for the rest of
1158 		 * the packets (besides the last one).
1159 		 */
1160 		if (unlikely(req->seqnum == 2)) {
1161 			/*
1162 			 * From this point on the lengths in both the
1163 			 * PBC and LRH are the same until the last
1164 			 * packet.
1165 			 * Adjust the template so we don't have to update
1166 			 * every packet
1167 			 */
1168 			req->hdr.pbc[0] = hdr->pbc[0];
1169 			req->hdr.lrh[2] = hdr->lrh[2];
1170 		}
1171 	}
1172 	/*
1173 	 * We only have to modify the header if this is not the
1174 	 * first packet in the request. Otherwise, we use the
1175 	 * header given to us.
1176 	 */
1177 	if (unlikely(!req->seqnum)) {
1178 		ret = check_header_template(req, hdr, lrhlen, datalen);
1179 		if (ret)
1180 			return ret;
1181 		goto done;
1182 	}
1183 
1184 	hdr->bth[2] = cpu_to_be32(
1185 		set_pkt_bth_psn(hdr->bth[2],
1186 				(req_opcode(req->info.ctrl) == EXPECTED),
1187 				req->seqnum));
1188 
1189 	/* Set ACK request on last packet */
1190 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1191 		hdr->bth[2] |= cpu_to_be32(1UL << 31);
1192 
1193 	/* Set the new offset */
1194 	hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1195 	/* Expected packets have to fill in the new TID information */
1196 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1197 		tidval = req->tids[req->tididx];
1198 		/*
1199 		 * If the offset puts us at the end of the current TID,
1200 		 * advance everything.
1201 		 */
1202 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1203 					 PAGE_SIZE)) {
1204 			req->tidoffset = 0;
1205 			/*
1206 			 * Since we don't copy all the TIDs, all at once,
1207 			 * we have to check again.
1208 			 */
1209 			if (++req->tididx > req->n_tids - 1 ||
1210 			    !req->tids[req->tididx]) {
1211 				return -EINVAL;
1212 			}
1213 			tidval = req->tids[req->tididx];
1214 		}
1215 		omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1216 			KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE_SHIFT :
1217 			KDETH_OM_SMALL_SHIFT;
1218 		/* Set KDETH.TIDCtrl based on value for this TID. */
1219 		KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1220 			  EXP_TID_GET(tidval, CTRL));
1221 		/* Set KDETH.TID based on value for this TID */
1222 		KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1223 			  EXP_TID_GET(tidval, IDX));
1224 		/* Clear KDETH.SH when DISABLE_SH flag is set */
1225 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1226 			KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1227 		/*
1228 		 * Set the KDETH.OFFSET and KDETH.OM based on size of
1229 		 * transfer.
1230 		 */
1231 		trace_hfi1_sdma_user_tid_info(
1232 			pq->dd, pq->ctxt, pq->subctxt, req->info.comp_idx,
1233 			req->tidoffset, req->tidoffset >> omfactor,
1234 			omfactor != KDETH_OM_SMALL_SHIFT);
1235 		KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1236 			  req->tidoffset >> omfactor);
1237 		KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1238 			  omfactor != KDETH_OM_SMALL_SHIFT);
1239 	}
1240 done:
1241 	trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1242 				    req->info.comp_idx, hdr, tidval);
1243 	return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1244 }
1245 
1246 static int set_txreq_header_ahg(struct user_sdma_request *req,
1247 				struct user_sdma_txreq *tx, u32 datalen)
1248 {
1249 	u32 ahg[AHG_KDETH_ARRAY_SIZE];
1250 	int idx = 0;
1251 	u8 omfactor; /* KDETH.OM */
1252 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1253 	struct hfi1_pkt_header *hdr = &req->hdr;
1254 	u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1255 	u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1256 	size_t array_size = ARRAY_SIZE(ahg);
1257 
1258 	if (PBC2LRH(pbclen) != lrhlen) {
1259 		/* PBC.PbcLengthDWs */
1260 		idx = ahg_header_set(ahg, idx, array_size, 0, 0, 12,
1261 				     (__force u16)cpu_to_le16(LRH2PBC(lrhlen)));
1262 		if (idx < 0)
1263 			return idx;
1264 		/* LRH.PktLen (we need the full 16 bits due to byte swap) */
1265 		idx = ahg_header_set(ahg, idx, array_size, 3, 0, 16,
1266 				     (__force u16)cpu_to_be16(lrhlen >> 2));
1267 		if (idx < 0)
1268 			return idx;
1269 	}
1270 
1271 	/*
1272 	 * Do the common updates
1273 	 */
1274 	/* BTH.PSN and BTH.A */
1275 	val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1276 		(HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1277 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1278 		val32 |= 1UL << 31;
1279 	idx = ahg_header_set(ahg, idx, array_size, 6, 0, 16,
1280 			     (__force u16)cpu_to_be16(val32 >> 16));
1281 	if (idx < 0)
1282 		return idx;
1283 	idx = ahg_header_set(ahg, idx, array_size, 6, 16, 16,
1284 			     (__force u16)cpu_to_be16(val32 & 0xffff));
1285 	if (idx < 0)
1286 		return idx;
1287 	/* KDETH.Offset */
1288 	idx = ahg_header_set(ahg, idx, array_size, 15, 0, 16,
1289 			     (__force u16)cpu_to_le16(req->koffset & 0xffff));
1290 	if (idx < 0)
1291 		return idx;
1292 	idx = ahg_header_set(ahg, idx, array_size, 15, 16, 16,
1293 			     (__force u16)cpu_to_le16(req->koffset >> 16));
1294 	if (idx < 0)
1295 		return idx;
1296 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1297 		__le16 val;
1298 
1299 		tidval = req->tids[req->tididx];
1300 
1301 		/*
1302 		 * If the offset puts us at the end of the current TID,
1303 		 * advance everything.
1304 		 */
1305 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1306 					 PAGE_SIZE)) {
1307 			req->tidoffset = 0;
1308 			/*
1309 			 * Since we don't copy all the TIDs, all at once,
1310 			 * we have to check again.
1311 			 */
1312 			if (++req->tididx > req->n_tids - 1 ||
1313 			    !req->tids[req->tididx])
1314 				return -EINVAL;
1315 			tidval = req->tids[req->tididx];
1316 		}
1317 		omfactor = ((EXP_TID_GET(tidval, LEN) *
1318 				  PAGE_SIZE) >=
1319 				 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE_SHIFT :
1320 				 KDETH_OM_SMALL_SHIFT;
1321 		/* KDETH.OM and KDETH.OFFSET (TID) */
1322 		idx = ahg_header_set(
1323 				ahg, idx, array_size, 7, 0, 16,
1324 				((!!(omfactor - KDETH_OM_SMALL_SHIFT)) << 15 |
1325 				((req->tidoffset >> omfactor)
1326 				& 0x7fff)));
1327 		if (idx < 0)
1328 			return idx;
1329 		/* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1330 		val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1331 				   (EXP_TID_GET(tidval, IDX) & 0x3ff));
1332 
1333 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1334 			val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1335 						      INTR) <<
1336 					    AHG_KDETH_INTR_SHIFT));
1337 		} else {
1338 			val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1339 			       cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1340 			       cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1341 						      INTR) <<
1342 					     AHG_KDETH_INTR_SHIFT));
1343 		}
1344 
1345 		idx = ahg_header_set(ahg, idx, array_size,
1346 				     7, 16, 14, (__force u16)val);
1347 		if (idx < 0)
1348 			return idx;
1349 	}
1350 
1351 	trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1352 					req->info.comp_idx, req->sde->this_idx,
1353 					req->ahg_idx, ahg, idx, tidval);
1354 	sdma_txinit_ahg(&tx->txreq,
1355 			SDMA_TXREQ_F_USE_AHG,
1356 			datalen, req->ahg_idx, idx,
1357 			ahg, sizeof(req->hdr),
1358 			user_sdma_txreq_cb);
1359 
1360 	return idx;
1361 }
1362 
1363 /**
1364  * user_sdma_txreq_cb() - SDMA tx request completion callback.
1365  * @txreq: valid sdma tx request
1366  * @status: success/failure of request
1367  *
1368  * Called when the SDMA progress state machine gets notification that
1369  * the SDMA descriptors for this tx request have been processed by the
1370  * DMA engine. Called in interrupt context.
1371  * Only do work on completed sequences.
1372  */
1373 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1374 {
1375 	struct user_sdma_txreq *tx =
1376 		container_of(txreq, struct user_sdma_txreq, txreq);
1377 	struct user_sdma_request *req;
1378 	struct hfi1_user_sdma_pkt_q *pq;
1379 	struct hfi1_user_sdma_comp_q *cq;
1380 	enum hfi1_sdma_comp_state state = COMPLETE;
1381 
1382 	if (!tx->req)
1383 		return;
1384 
1385 	req = tx->req;
1386 	pq = req->pq;
1387 	cq = req->cq;
1388 
1389 	if (status != SDMA_TXREQ_S_OK) {
1390 		SDMA_DBG(req, "SDMA completion with error %d",
1391 			 status);
1392 		WRITE_ONCE(req->has_error, 1);
1393 		state = ERROR;
1394 	}
1395 
1396 	req->seqcomp = tx->seqnum;
1397 	kmem_cache_free(pq->txreq_cache, tx);
1398 
1399 	/* sequence isn't complete?  We are done */
1400 	if (req->seqcomp != req->info.npkts - 1)
1401 		return;
1402 
1403 	user_sdma_free_request(req, false);
1404 	set_comp_state(pq, cq, req->info.comp_idx, state, status);
1405 	pq_update(pq);
1406 }
1407 
1408 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1409 {
1410 	if (atomic_dec_and_test(&pq->n_reqs))
1411 		wake_up(&pq->wait);
1412 }
1413 
1414 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1415 {
1416 	int i;
1417 
1418 	if (!list_empty(&req->txps)) {
1419 		struct sdma_txreq *t, *p;
1420 
1421 		list_for_each_entry_safe(t, p, &req->txps, list) {
1422 			struct user_sdma_txreq *tx =
1423 				container_of(t, struct user_sdma_txreq, txreq);
1424 			list_del_init(&t->list);
1425 			sdma_txclean(req->pq->dd, t);
1426 			kmem_cache_free(req->pq->txreq_cache, tx);
1427 		}
1428 	}
1429 
1430 	for (i = 0; i < req->data_iovs; i++) {
1431 		struct sdma_mmu_node *node = req->iovs[i].node;
1432 
1433 		if (!node)
1434 			continue;
1435 
1436 		req->iovs[i].node = NULL;
1437 
1438 		if (unpin)
1439 			hfi1_mmu_rb_remove(req->pq->handler,
1440 					   &node->rb);
1441 		else
1442 			atomic_dec(&node->refcount);
1443 	}
1444 
1445 	kfree(req->tids);
1446 	clear_bit(req->info.comp_idx, req->pq->req_in_use);
1447 }
1448 
1449 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1450 				  struct hfi1_user_sdma_comp_q *cq,
1451 				  u16 idx, enum hfi1_sdma_comp_state state,
1452 				  int ret)
1453 {
1454 	if (state == ERROR)
1455 		cq->comps[idx].errcode = -ret;
1456 	smp_wmb(); /* make sure errcode is visible first */
1457 	cq->comps[idx].status = state;
1458 	trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1459 					idx, state, ret);
1460 }
1461 
1462 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1463 			   unsigned long len)
1464 {
1465 	return (bool)(node->addr == addr);
1466 }
1467 
1468 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1469 {
1470 	struct sdma_mmu_node *node =
1471 		container_of(mnode, struct sdma_mmu_node, rb);
1472 
1473 	atomic_inc(&node->refcount);
1474 	return 0;
1475 }
1476 
1477 /*
1478  * Return 1 to remove the node from the rb tree and call the remove op.
1479  *
1480  * Called with the rb tree lock held.
1481  */
1482 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1483 			 void *evict_arg, bool *stop)
1484 {
1485 	struct sdma_mmu_node *node =
1486 		container_of(mnode, struct sdma_mmu_node, rb);
1487 	struct evict_data *evict_data = evict_arg;
1488 
1489 	/* is this node still being used? */
1490 	if (atomic_read(&node->refcount))
1491 		return 0; /* keep this node */
1492 
1493 	/* this node will be evicted, add its pages to our count */
1494 	evict_data->cleared += node->npages;
1495 
1496 	/* have enough pages been cleared? */
1497 	if (evict_data->cleared >= evict_data->target)
1498 		*stop = true;
1499 
1500 	return 1; /* remove this node */
1501 }
1502 
1503 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1504 {
1505 	struct sdma_mmu_node *node =
1506 		container_of(mnode, struct sdma_mmu_node, rb);
1507 
1508 	unpin_sdma_pages(node);
1509 	kfree(node);
1510 }
1511 
1512 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1513 {
1514 	struct sdma_mmu_node *node =
1515 		container_of(mnode, struct sdma_mmu_node, rb);
1516 
1517 	if (!atomic_read(&node->refcount))
1518 		return 1;
1519 	return 0;
1520 }
1521