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