1 /*
2  * Copyright(c) 2015 - 2018 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_work *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_work *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->iow, struct hfi1_user_sdma_pkt_q, busy);
133 	struct user_sdma_txreq *tx =
134 		container_of(txreq, struct user_sdma_txreq, txreq);
135 
136 	if (sdma_progress(sde, seq, txreq)) {
137 		if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
138 			goto eagain;
139 	}
140 	/*
141 	 * We are assuming that if the list is enqueued somewhere, it
142 	 * is to the dmawait list since that is the only place where
143 	 * it is supposed to be enqueued.
144 	 */
145 	xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
146 	write_seqlock(&sde->waitlock);
147 	if (list_empty(&pq->busy.list)) {
148 		iowait_get_priority(&pq->busy);
149 		iowait_queue(pkts_sent, &pq->busy, &sde->dmawait);
150 	}
151 	write_sequnlock(&sde->waitlock);
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 	atomic_set(&pq->n_reqs, 0);
191 	init_waitqueue_head(&pq->wait);
192 	atomic_set(&pq->n_locked, 0);
193 	pq->mm = fd->mm;
194 
195 	iowait_init(&pq->busy, 0, NULL, NULL, defer_packet_queue,
196 		    activate_packet_queue, NULL, NULL);
197 	pq->reqidx = 0;
198 
199 	pq->reqs = kcalloc(hfi1_sdma_comp_ring_size,
200 			   sizeof(*pq->reqs),
201 			   GFP_KERNEL);
202 	if (!pq->reqs)
203 		goto pq_reqs_nomem;
204 
205 	pq->req_in_use = kcalloc(BITS_TO_LONGS(hfi1_sdma_comp_ring_size),
206 				 sizeof(*pq->req_in_use),
207 				 GFP_KERNEL);
208 	if (!pq->req_in_use)
209 		goto pq_reqs_no_in_use;
210 
211 	snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
212 		 fd->subctxt);
213 	pq->txreq_cache = kmem_cache_create(buf,
214 					    sizeof(struct user_sdma_txreq),
215 					    L1_CACHE_BYTES,
216 					    SLAB_HWCACHE_ALIGN,
217 					    NULL);
218 	if (!pq->txreq_cache) {
219 		dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
220 			   uctxt->ctxt);
221 		goto pq_txreq_nomem;
222 	}
223 
224 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
225 	if (!cq)
226 		goto cq_nomem;
227 
228 	cq->comps = vmalloc_user(PAGE_ALIGN(sizeof(*cq->comps)
229 				 * hfi1_sdma_comp_ring_size));
230 	if (!cq->comps)
231 		goto cq_comps_nomem;
232 
233 	cq->nentries = hfi1_sdma_comp_ring_size;
234 
235 	ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq,
236 				   &pq->handler);
237 	if (ret) {
238 		dd_dev_err(dd, "Failed to register with MMU %d", ret);
239 		goto pq_mmu_fail;
240 	}
241 
242 	fd->pq = pq;
243 	fd->cq = cq;
244 
245 	return 0;
246 
247 pq_mmu_fail:
248 	vfree(cq->comps);
249 cq_comps_nomem:
250 	kfree(cq);
251 cq_nomem:
252 	kmem_cache_destroy(pq->txreq_cache);
253 pq_txreq_nomem:
254 	kfree(pq->req_in_use);
255 pq_reqs_no_in_use:
256 	kfree(pq->reqs);
257 pq_reqs_nomem:
258 	kfree(pq);
259 
260 	return ret;
261 }
262 
263 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd,
264 			       struct hfi1_ctxtdata *uctxt)
265 {
266 	struct hfi1_user_sdma_pkt_q *pq;
267 
268 	trace_hfi1_sdma_user_free_queues(uctxt->dd, uctxt->ctxt, fd->subctxt);
269 
270 	pq = fd->pq;
271 	if (pq) {
272 		if (pq->handler)
273 			hfi1_mmu_rb_unregister(pq->handler);
274 		iowait_sdma_drain(&pq->busy);
275 		/* Wait until all requests have been freed. */
276 		wait_event_interruptible(
277 			pq->wait,
278 			!atomic_read(&pq->n_reqs));
279 		kfree(pq->reqs);
280 		kfree(pq->req_in_use);
281 		kmem_cache_destroy(pq->txreq_cache);
282 		kfree(pq);
283 		fd->pq = NULL;
284 	}
285 	if (fd->cq) {
286 		vfree(fd->cq->comps);
287 		kfree(fd->cq);
288 		fd->cq = NULL;
289 	}
290 	return 0;
291 }
292 
293 static u8 dlid_to_selector(u16 dlid)
294 {
295 	static u8 mapping[256];
296 	static int initialized;
297 	static u8 next;
298 	int hash;
299 
300 	if (!initialized) {
301 		memset(mapping, 0xFF, 256);
302 		initialized = 1;
303 	}
304 
305 	hash = ((dlid >> 8) ^ dlid) & 0xFF;
306 	if (mapping[hash] == 0xFF) {
307 		mapping[hash] = next;
308 		next = (next + 1) & 0x7F;
309 	}
310 
311 	return mapping[hash];
312 }
313 
314 /**
315  * hfi1_user_sdma_process_request() - Process and start a user sdma request
316  * @fd: valid file descriptor
317  * @iovec: array of io vectors to process
318  * @dim: overall iovec array size
319  * @count: number of io vector array entries processed
320  */
321 int hfi1_user_sdma_process_request(struct hfi1_filedata *fd,
322 				   struct iovec *iovec, unsigned long dim,
323 				   unsigned long *count)
324 {
325 	int ret = 0, i;
326 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
327 	struct hfi1_user_sdma_pkt_q *pq = fd->pq;
328 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
329 	struct hfi1_devdata *dd = pq->dd;
330 	unsigned long idx = 0;
331 	u8 pcount = initial_pkt_count;
332 	struct sdma_req_info info;
333 	struct user_sdma_request *req;
334 	u8 opcode, sc, vl;
335 	u16 pkey;
336 	u32 slid;
337 	u16 dlid;
338 	u32 selector;
339 
340 	if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
341 		hfi1_cdbg(
342 		   SDMA,
343 		   "[%u:%u:%u] First vector not big enough for header %lu/%lu",
344 		   dd->unit, uctxt->ctxt, fd->subctxt,
345 		   iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
346 		return -EINVAL;
347 	}
348 	ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
349 	if (ret) {
350 		hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
351 			  dd->unit, uctxt->ctxt, fd->subctxt, ret);
352 		return -EFAULT;
353 	}
354 
355 	trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
356 				     (u16 *)&info);
357 	if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
358 		hfi1_cdbg(SDMA,
359 			  "[%u:%u:%u:%u] Invalid comp index",
360 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
361 		return -EINVAL;
362 	}
363 
364 	/*
365 	 * Sanity check the header io vector count.  Need at least 1 vector
366 	 * (header) and cannot be larger than the actual io vector count.
367 	 */
368 	if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
369 		hfi1_cdbg(SDMA,
370 			  "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
371 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
372 			  req_iovcnt(info.ctrl), dim);
373 		return -EINVAL;
374 	}
375 
376 	if (!info.fragsize) {
377 		hfi1_cdbg(SDMA,
378 			  "[%u:%u:%u:%u] Request does not specify fragsize",
379 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
380 		return -EINVAL;
381 	}
382 
383 	/* Try to claim the request. */
384 	if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
385 		hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
386 			  dd->unit, uctxt->ctxt, fd->subctxt,
387 			  info.comp_idx);
388 		return -EBADSLT;
389 	}
390 	/*
391 	 * All safety checks have been done and this request has been claimed.
392 	 */
393 	trace_hfi1_sdma_user_process_request(dd, uctxt->ctxt, fd->subctxt,
394 					     info.comp_idx);
395 	req = pq->reqs + info.comp_idx;
396 	req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
397 	req->data_len  = 0;
398 	req->pq = pq;
399 	req->cq = cq;
400 	req->ahg_idx = -1;
401 	req->iov_idx = 0;
402 	req->sent = 0;
403 	req->seqnum = 0;
404 	req->seqcomp = 0;
405 	req->seqsubmitted = 0;
406 	req->tids = NULL;
407 	req->has_error = 0;
408 	INIT_LIST_HEAD(&req->txps);
409 
410 	memcpy(&req->info, &info, sizeof(info));
411 
412 	/* The request is initialized, count it */
413 	atomic_inc(&pq->n_reqs);
414 
415 	if (req_opcode(info.ctrl) == EXPECTED) {
416 		/* expected must have a TID info and at least one data vector */
417 		if (req->data_iovs < 2) {
418 			SDMA_DBG(req,
419 				 "Not enough vectors for expected request");
420 			ret = -EINVAL;
421 			goto free_req;
422 		}
423 		req->data_iovs--;
424 	}
425 
426 	if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
427 		SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
428 			 MAX_VECTORS_PER_REQ);
429 		ret = -EINVAL;
430 		goto free_req;
431 	}
432 	/* Copy the header from the user buffer */
433 	ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
434 			     sizeof(req->hdr));
435 	if (ret) {
436 		SDMA_DBG(req, "Failed to copy header template (%d)", ret);
437 		ret = -EFAULT;
438 		goto free_req;
439 	}
440 
441 	/* If Static rate control is not enabled, sanitize the header. */
442 	if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
443 		req->hdr.pbc[2] = 0;
444 
445 	/* Validate the opcode. Do not trust packets from user space blindly. */
446 	opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
447 	if ((opcode & USER_OPCODE_CHECK_MASK) !=
448 	     USER_OPCODE_CHECK_VAL) {
449 		SDMA_DBG(req, "Invalid opcode (%d)", opcode);
450 		ret = -EINVAL;
451 		goto free_req;
452 	}
453 	/*
454 	 * Validate the vl. Do not trust packets from user space blindly.
455 	 * VL comes from PBC, SC comes from LRH, and the VL needs to
456 	 * match the SC look up.
457 	 */
458 	vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
459 	sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
460 	      (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
461 	if (vl >= dd->pport->vls_operational ||
462 	    vl != sc_to_vlt(dd, sc)) {
463 		SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
464 		ret = -EINVAL;
465 		goto free_req;
466 	}
467 
468 	/* Checking P_KEY for requests from user-space */
469 	pkey = (u16)be32_to_cpu(req->hdr.bth[0]);
470 	slid = be16_to_cpu(req->hdr.lrh[3]);
471 	if (egress_pkey_check(dd->pport, slid, pkey, sc, PKEY_CHECK_INVALID)) {
472 		ret = -EINVAL;
473 		goto free_req;
474 	}
475 
476 	/*
477 	 * Also should check the BTH.lnh. If it says the next header is GRH then
478 	 * the RXE parsing will be off and will land in the middle of the KDETH
479 	 * or miss it entirely.
480 	 */
481 	if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
482 		SDMA_DBG(req, "User tried to pass in a GRH");
483 		ret = -EINVAL;
484 		goto free_req;
485 	}
486 
487 	req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
488 	/*
489 	 * Calculate the initial TID offset based on the values of
490 	 * KDETH.OFFSET and KDETH.OM that are passed in.
491 	 */
492 	req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
493 		(KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
494 		 KDETH_OM_LARGE : KDETH_OM_SMALL);
495 	trace_hfi1_sdma_user_initial_tidoffset(dd, uctxt->ctxt, fd->subctxt,
496 					       info.comp_idx, req->tidoffset);
497 	idx++;
498 
499 	/* Save all the IO vector structures */
500 	for (i = 0; i < req->data_iovs; i++) {
501 		req->iovs[i].offset = 0;
502 		INIT_LIST_HEAD(&req->iovs[i].list);
503 		memcpy(&req->iovs[i].iov,
504 		       iovec + idx++,
505 		       sizeof(req->iovs[i].iov));
506 		ret = pin_vector_pages(req, &req->iovs[i]);
507 		if (ret) {
508 			req->data_iovs = i;
509 			goto free_req;
510 		}
511 		req->data_len += req->iovs[i].iov.iov_len;
512 	}
513 	trace_hfi1_sdma_user_data_length(dd, uctxt->ctxt, fd->subctxt,
514 					 info.comp_idx, req->data_len);
515 	if (pcount > req->info.npkts)
516 		pcount = req->info.npkts;
517 	/*
518 	 * Copy any TID info
519 	 * User space will provide the TID info only when the
520 	 * request type is EXPECTED. This is true even if there is
521 	 * only one packet in the request and the header is already
522 	 * setup. The reason for the singular TID case is that the
523 	 * driver needs to perform safety checks.
524 	 */
525 	if (req_opcode(req->info.ctrl) == EXPECTED) {
526 		u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
527 		u32 *tmp;
528 
529 		if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
530 			ret = -EINVAL;
531 			goto free_req;
532 		}
533 
534 		/*
535 		 * We have to copy all of the tids because they may vary
536 		 * in size and, therefore, the TID count might not be
537 		 * equal to the pkt count. However, there is no way to
538 		 * tell at this point.
539 		 */
540 		tmp = memdup_user(iovec[idx].iov_base,
541 				  ntids * sizeof(*req->tids));
542 		if (IS_ERR(tmp)) {
543 			ret = PTR_ERR(tmp);
544 			SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
545 				 ntids, ret);
546 			goto free_req;
547 		}
548 		req->tids = tmp;
549 		req->n_tids = ntids;
550 		req->tididx = 0;
551 		idx++;
552 	}
553 
554 	dlid = be16_to_cpu(req->hdr.lrh[1]);
555 	selector = dlid_to_selector(dlid);
556 	selector += uctxt->ctxt + fd->subctxt;
557 	req->sde = sdma_select_user_engine(dd, selector, vl);
558 
559 	if (!req->sde || !sdma_running(req->sde)) {
560 		ret = -ECOMM;
561 		goto free_req;
562 	}
563 
564 	/* We don't need an AHG entry if the request contains only one packet */
565 	if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG))
566 		req->ahg_idx = sdma_ahg_alloc(req->sde);
567 
568 	set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
569 	pq->state = SDMA_PKT_Q_ACTIVE;
570 	/* Send the first N packets in the request to buy us some time */
571 	ret = user_sdma_send_pkts(req, pcount);
572 	if (unlikely(ret < 0 && ret != -EBUSY))
573 		goto free_req;
574 
575 	/*
576 	 * This is a somewhat blocking send implementation.
577 	 * The driver will block the caller until all packets of the
578 	 * request have been submitted to the SDMA engine. However, it
579 	 * will not wait for send completions.
580 	 */
581 	while (req->seqsubmitted != req->info.npkts) {
582 		ret = user_sdma_send_pkts(req, pcount);
583 		if (ret < 0) {
584 			if (ret != -EBUSY)
585 				goto free_req;
586 			wait_event_interruptible_timeout(
587 				pq->busy.wait_dma,
588 				(pq->state == SDMA_PKT_Q_ACTIVE),
589 				msecs_to_jiffies(
590 					SDMA_IOWAIT_TIMEOUT));
591 		}
592 	}
593 	*count += idx;
594 	return 0;
595 free_req:
596 	/*
597 	 * If the submitted seqsubmitted == npkts, the completion routine
598 	 * controls the final state.  If sequbmitted < npkts, wait for any
599 	 * outstanding packets to finish before cleaning up.
600 	 */
601 	if (req->seqsubmitted < req->info.npkts) {
602 		if (req->seqsubmitted)
603 			wait_event(pq->busy.wait_dma,
604 				   (req->seqcomp == req->seqsubmitted - 1));
605 		user_sdma_free_request(req, true);
606 		pq_update(pq);
607 		set_comp_state(pq, cq, info.comp_idx, ERROR, ret);
608 	}
609 	return ret;
610 }
611 
612 static inline u32 compute_data_length(struct user_sdma_request *req,
613 				      struct user_sdma_txreq *tx)
614 {
615 	/*
616 	 * Determine the proper size of the packet data.
617 	 * The size of the data of the first packet is in the header
618 	 * template. However, it includes the header and ICRC, which need
619 	 * to be subtracted.
620 	 * The minimum representable packet data length in a header is 4 bytes,
621 	 * therefore, when the data length request is less than 4 bytes, there's
622 	 * only one packet, and the packet data length is equal to that of the
623 	 * request data length.
624 	 * The size of the remaining packets is the minimum of the frag
625 	 * size (MTU) or remaining data in the request.
626 	 */
627 	u32 len;
628 
629 	if (!req->seqnum) {
630 		if (req->data_len < sizeof(u32))
631 			len = req->data_len;
632 		else
633 			len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
634 			       (sizeof(tx->hdr) - 4));
635 	} else if (req_opcode(req->info.ctrl) == EXPECTED) {
636 		u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
637 			PAGE_SIZE;
638 		/*
639 		 * Get the data length based on the remaining space in the
640 		 * TID pair.
641 		 */
642 		len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
643 		/* If we've filled up the TID pair, move to the next one. */
644 		if (unlikely(!len) && ++req->tididx < req->n_tids &&
645 		    req->tids[req->tididx]) {
646 			tidlen = EXP_TID_GET(req->tids[req->tididx],
647 					     LEN) * PAGE_SIZE;
648 			req->tidoffset = 0;
649 			len = min_t(u32, tidlen, req->info.fragsize);
650 		}
651 		/*
652 		 * Since the TID pairs map entire pages, make sure that we
653 		 * are not going to try to send more data that we have
654 		 * remaining.
655 		 */
656 		len = min(len, req->data_len - req->sent);
657 	} else {
658 		len = min(req->data_len - req->sent, (u32)req->info.fragsize);
659 	}
660 	trace_hfi1_sdma_user_compute_length(req->pq->dd,
661 					    req->pq->ctxt,
662 					    req->pq->subctxt,
663 					    req->info.comp_idx,
664 					    len);
665 	return len;
666 }
667 
668 static inline u32 pad_len(u32 len)
669 {
670 	if (len & (sizeof(u32) - 1))
671 		len += sizeof(u32) - (len & (sizeof(u32) - 1));
672 	return len;
673 }
674 
675 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
676 {
677 	/* (Size of complete header - size of PBC) + 4B ICRC + data length */
678 	return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
679 }
680 
681 static int user_sdma_txadd_ahg(struct user_sdma_request *req,
682 			       struct user_sdma_txreq *tx,
683 			       u32 datalen)
684 {
685 	int ret;
686 	u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
687 	u32 lrhlen = get_lrh_len(req->hdr, pad_len(datalen));
688 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
689 
690 	/*
691 	 * Copy the request header into the tx header
692 	 * because the HW needs a cacheline-aligned
693 	 * address.
694 	 * This copy can be optimized out if the hdr
695 	 * member of user_sdma_request were also
696 	 * cacheline aligned.
697 	 */
698 	memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
699 	if (PBC2LRH(pbclen) != lrhlen) {
700 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
701 		tx->hdr.pbc[0] = cpu_to_le16(pbclen);
702 	}
703 	ret = check_header_template(req, &tx->hdr, lrhlen, datalen);
704 	if (ret)
705 		return ret;
706 	ret = sdma_txinit_ahg(&tx->txreq, SDMA_TXREQ_F_AHG_COPY,
707 			      sizeof(tx->hdr) + datalen, req->ahg_idx,
708 			      0, NULL, 0, user_sdma_txreq_cb);
709 	if (ret)
710 		return ret;
711 	ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq, &tx->hdr, sizeof(tx->hdr));
712 	if (ret)
713 		sdma_txclean(pq->dd, &tx->txreq);
714 	return ret;
715 }
716 
717 static int user_sdma_txadd(struct user_sdma_request *req,
718 			   struct user_sdma_txreq *tx,
719 			   struct user_sdma_iovec *iovec, u32 datalen,
720 			   u32 *queued_ptr, u32 *data_sent_ptr,
721 			   u64 *iov_offset_ptr)
722 {
723 	int ret;
724 	unsigned int pageidx, len;
725 	unsigned long base, offset;
726 	u64 iov_offset = *iov_offset_ptr;
727 	u32 queued = *queued_ptr, data_sent = *data_sent_ptr;
728 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
729 
730 	base = (unsigned long)iovec->iov.iov_base;
731 	offset = offset_in_page(base + iovec->offset + iov_offset);
732 	pageidx = (((iovec->offset + iov_offset + base) - (base & PAGE_MASK)) >>
733 		   PAGE_SHIFT);
734 	len = offset + req->info.fragsize > PAGE_SIZE ?
735 		PAGE_SIZE - offset : req->info.fragsize;
736 	len = min((datalen - queued), len);
737 	ret = sdma_txadd_page(pq->dd, &tx->txreq, iovec->pages[pageidx],
738 			      offset, len);
739 	if (ret) {
740 		SDMA_DBG(req, "SDMA txreq add page failed %d\n", ret);
741 		return ret;
742 	}
743 	iov_offset += len;
744 	queued += len;
745 	data_sent += len;
746 	if (unlikely(queued < datalen && pageidx == iovec->npages &&
747 		     req->iov_idx < req->data_iovs - 1)) {
748 		iovec->offset += iov_offset;
749 		iovec = &req->iovs[++req->iov_idx];
750 		iov_offset = 0;
751 	}
752 
753 	*queued_ptr = queued;
754 	*data_sent_ptr = data_sent;
755 	*iov_offset_ptr = iov_offset;
756 	return ret;
757 }
758 
759 static int user_sdma_send_pkts(struct user_sdma_request *req, u16 maxpkts)
760 {
761 	int ret = 0;
762 	u16 count;
763 	unsigned npkts = 0;
764 	struct user_sdma_txreq *tx = NULL;
765 	struct hfi1_user_sdma_pkt_q *pq = NULL;
766 	struct user_sdma_iovec *iovec = NULL;
767 
768 	if (!req->pq)
769 		return -EINVAL;
770 
771 	pq = req->pq;
772 
773 	/* If tx completion has reported an error, we are done. */
774 	if (READ_ONCE(req->has_error))
775 		return -EFAULT;
776 
777 	/*
778 	 * Check if we might have sent the entire request already
779 	 */
780 	if (unlikely(req->seqnum == req->info.npkts)) {
781 		if (!list_empty(&req->txps))
782 			goto dosend;
783 		return ret;
784 	}
785 
786 	if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
787 		maxpkts = req->info.npkts - req->seqnum;
788 
789 	while (npkts < maxpkts) {
790 		u32 datalen = 0, queued = 0, data_sent = 0;
791 		u64 iov_offset = 0;
792 
793 		/*
794 		 * Check whether any of the completions have come back
795 		 * with errors. If so, we are not going to process any
796 		 * more packets from this request.
797 		 */
798 		if (READ_ONCE(req->has_error))
799 			return -EFAULT;
800 
801 		tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
802 		if (!tx)
803 			return -ENOMEM;
804 
805 		tx->flags = 0;
806 		tx->req = req;
807 		tx->busycount = 0;
808 		INIT_LIST_HEAD(&tx->list);
809 
810 		/*
811 		 * For the last packet set the ACK request
812 		 * and disable header suppression.
813 		 */
814 		if (req->seqnum == req->info.npkts - 1)
815 			tx->flags |= (TXREQ_FLAGS_REQ_ACK |
816 				      TXREQ_FLAGS_REQ_DISABLE_SH);
817 
818 		/*
819 		 * Calculate the payload size - this is min of the fragment
820 		 * (MTU) size or the remaining bytes in the request but only
821 		 * if we have payload data.
822 		 */
823 		if (req->data_len) {
824 			iovec = &req->iovs[req->iov_idx];
825 			if (READ_ONCE(iovec->offset) == iovec->iov.iov_len) {
826 				if (++req->iov_idx == req->data_iovs) {
827 					ret = -EFAULT;
828 					goto free_tx;
829 				}
830 				iovec = &req->iovs[req->iov_idx];
831 				WARN_ON(iovec->offset);
832 			}
833 
834 			datalen = compute_data_length(req, tx);
835 
836 			/*
837 			 * Disable header suppression for the payload <= 8DWS.
838 			 * If there is an uncorrectable error in the receive
839 			 * data FIFO when the received payload size is less than
840 			 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
841 			 * not reported.There is set RHF.EccErr if the header
842 			 * is not suppressed.
843 			 */
844 			if (!datalen) {
845 				SDMA_DBG(req,
846 					 "Request has data but pkt len is 0");
847 				ret = -EFAULT;
848 				goto free_tx;
849 			} else if (datalen <= 32) {
850 				tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
851 			}
852 		}
853 
854 		if (req->ahg_idx >= 0) {
855 			if (!req->seqnum) {
856 				ret = user_sdma_txadd_ahg(req, tx, datalen);
857 				if (ret)
858 					goto free_tx;
859 			} else {
860 				int changes;
861 
862 				changes = set_txreq_header_ahg(req, tx,
863 							       datalen);
864 				if (changes < 0) {
865 					ret = changes;
866 					goto free_tx;
867 				}
868 			}
869 		} else {
870 			ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
871 					  datalen, user_sdma_txreq_cb);
872 			if (ret)
873 				goto free_tx;
874 			/*
875 			 * Modify the header for this packet. This only needs
876 			 * to be done if we are not going to use AHG. Otherwise,
877 			 * the HW will do it based on the changes we gave it
878 			 * during sdma_txinit_ahg().
879 			 */
880 			ret = set_txreq_header(req, tx, datalen);
881 			if (ret)
882 				goto free_txreq;
883 		}
884 
885 		/*
886 		 * If the request contains any data vectors, add up to
887 		 * fragsize bytes to the descriptor.
888 		 */
889 		while (queued < datalen &&
890 		       (req->sent + data_sent) < req->data_len) {
891 			ret = user_sdma_txadd(req, tx, iovec, datalen,
892 					      &queued, &data_sent, &iov_offset);
893 			if (ret)
894 				goto free_txreq;
895 		}
896 		/*
897 		 * The txreq was submitted successfully so we can update
898 		 * the counters.
899 		 */
900 		req->koffset += datalen;
901 		if (req_opcode(req->info.ctrl) == EXPECTED)
902 			req->tidoffset += datalen;
903 		req->sent += data_sent;
904 		if (req->data_len)
905 			iovec->offset += iov_offset;
906 		list_add_tail(&tx->txreq.list, &req->txps);
907 		/*
908 		 * It is important to increment this here as it is used to
909 		 * generate the BTH.PSN and, therefore, can't be bulk-updated
910 		 * outside of the loop.
911 		 */
912 		tx->seqnum = req->seqnum++;
913 		npkts++;
914 	}
915 dosend:
916 	ret = sdma_send_txlist(req->sde,
917 			       iowait_get_ib_work(&pq->busy),
918 			       &req->txps, &count);
919 	req->seqsubmitted += count;
920 	if (req->seqsubmitted == req->info.npkts) {
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 		return -ENOMEM;
961 	memcpy(pages, node->pages, node->npages * sizeof(*pages));
962 
963 	npages -= node->npages;
964 retry:
965 	if (!hfi1_can_pin_pages(pq->dd, pq->mm,
966 				atomic_read(&pq->n_locked), npages)) {
967 		cleared = sdma_cache_evict(pq, npages);
968 		if (cleared >= npages)
969 			goto retry;
970 	}
971 	pinned = hfi1_acquire_user_pages(pq->mm,
972 					 ((unsigned long)iovec->iov.iov_base +
973 					 (node->npages * PAGE_SIZE)), npages, 0,
974 					 pages + node->npages);
975 	if (pinned < 0) {
976 		kfree(pages);
977 		return pinned;
978 	}
979 	if (pinned != npages) {
980 		unpin_vector_pages(pq->mm, pages, node->npages, pinned);
981 		return -EFAULT;
982 	}
983 	kfree(node->pages);
984 	node->rb.len = iovec->iov.iov_len;
985 	node->pages = pages;
986 	atomic_add(pinned, &pq->n_locked);
987 	return pinned;
988 }
989 
990 static void unpin_sdma_pages(struct sdma_mmu_node *node)
991 {
992 	if (node->npages) {
993 		unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages);
994 		atomic_sub(node->npages, &node->pq->n_locked);
995 	}
996 }
997 
998 static int pin_vector_pages(struct user_sdma_request *req,
999 			    struct user_sdma_iovec *iovec)
1000 {
1001 	int ret = 0, pinned, npages;
1002 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1003 	struct sdma_mmu_node *node = NULL;
1004 	struct mmu_rb_node *rb_node;
1005 	struct iovec *iov;
1006 	bool extracted;
1007 
1008 	extracted =
1009 		hfi1_mmu_rb_remove_unless_exact(pq->handler,
1010 						(unsigned long)
1011 						iovec->iov.iov_base,
1012 						iovec->iov.iov_len, &rb_node);
1013 	if (rb_node) {
1014 		node = container_of(rb_node, struct sdma_mmu_node, rb);
1015 		if (!extracted) {
1016 			atomic_inc(&node->refcount);
1017 			iovec->pages = node->pages;
1018 			iovec->npages = node->npages;
1019 			iovec->node = node;
1020 			return 0;
1021 		}
1022 	}
1023 
1024 	if (!node) {
1025 		node = kzalloc(sizeof(*node), GFP_KERNEL);
1026 		if (!node)
1027 			return -ENOMEM;
1028 
1029 		node->rb.addr = (unsigned long)iovec->iov.iov_base;
1030 		node->pq = pq;
1031 		atomic_set(&node->refcount, 0);
1032 	}
1033 
1034 	iov = &iovec->iov;
1035 	npages = num_user_pages((unsigned long)iov->iov_base, iov->iov_len);
1036 	if (node->npages < npages) {
1037 		pinned = pin_sdma_pages(req, iovec, node, npages);
1038 		if (pinned < 0) {
1039 			ret = pinned;
1040 			goto bail;
1041 		}
1042 		node->npages += pinned;
1043 		npages = node->npages;
1044 	}
1045 	iovec->pages = node->pages;
1046 	iovec->npages = npages;
1047 	iovec->node = node;
1048 
1049 	ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1050 	if (ret) {
1051 		iovec->node = NULL;
1052 		goto bail;
1053 	}
1054 	return 0;
1055 bail:
1056 	unpin_sdma_pages(node);
1057 	kfree(node);
1058 	return ret;
1059 }
1060 
1061 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1062 			       unsigned start, unsigned npages)
1063 {
1064 	hfi1_release_user_pages(mm, pages + start, npages, false);
1065 	kfree(pages);
1066 }
1067 
1068 static int check_header_template(struct user_sdma_request *req,
1069 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
1070 				 u32 datalen)
1071 {
1072 	/*
1073 	 * Perform safety checks for any type of packet:
1074 	 *    - transfer size is multiple of 64bytes
1075 	 *    - packet length is multiple of 4 bytes
1076 	 *    - packet length is not larger than MTU size
1077 	 *
1078 	 * These checks are only done for the first packet of the
1079 	 * transfer since the header is "given" to us by user space.
1080 	 * For the remainder of the packets we compute the values.
1081 	 */
1082 	if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1083 	    lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1084 		return -EINVAL;
1085 
1086 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1087 		/*
1088 		 * The header is checked only on the first packet. Furthermore,
1089 		 * we ensure that at least one TID entry is copied when the
1090 		 * request is submitted. Therefore, we don't have to verify that
1091 		 * tididx points to something sane.
1092 		 */
1093 		u32 tidval = req->tids[req->tididx],
1094 			tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1095 			tididx = EXP_TID_GET(tidval, IDX),
1096 			tidctrl = EXP_TID_GET(tidval, CTRL),
1097 			tidoff;
1098 		__le32 kval = hdr->kdeth.ver_tid_offset;
1099 
1100 		tidoff = KDETH_GET(kval, OFFSET) *
1101 			  (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1102 			   KDETH_OM_LARGE : KDETH_OM_SMALL);
1103 		/*
1104 		 * Expected receive packets have the following
1105 		 * additional checks:
1106 		 *     - offset is not larger than the TID size
1107 		 *     - TIDCtrl values match between header and TID array
1108 		 *     - TID indexes match between header and TID array
1109 		 */
1110 		if ((tidoff + datalen > tidlen) ||
1111 		    KDETH_GET(kval, TIDCTRL) != tidctrl ||
1112 		    KDETH_GET(kval, TID) != tididx)
1113 			return -EINVAL;
1114 	}
1115 	return 0;
1116 }
1117 
1118 /*
1119  * Correctly set the BTH.PSN field based on type of
1120  * transfer - eager packets can just increment the PSN but
1121  * expected packets encode generation and sequence in the
1122  * BTH.PSN field so just incrementing will result in errors.
1123  */
1124 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1125 {
1126 	u32 val = be32_to_cpu(bthpsn),
1127 		mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1128 			0xffffffull),
1129 		psn = val & mask;
1130 	if (expct)
1131 		psn = (psn & ~HFI1_KDETH_BTH_SEQ_MASK) |
1132 			((psn + frags) & HFI1_KDETH_BTH_SEQ_MASK);
1133 	else
1134 		psn = psn + frags;
1135 	return psn & mask;
1136 }
1137 
1138 static int set_txreq_header(struct user_sdma_request *req,
1139 			    struct user_sdma_txreq *tx, u32 datalen)
1140 {
1141 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1142 	struct hfi1_pkt_header *hdr = &tx->hdr;
1143 	u8 omfactor; /* KDETH.OM */
1144 	u16 pbclen;
1145 	int ret;
1146 	u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1147 
1148 	/* Copy the header template to the request before modification */
1149 	memcpy(hdr, &req->hdr, sizeof(*hdr));
1150 
1151 	/*
1152 	 * Check if the PBC and LRH length are mismatched. If so
1153 	 * adjust both in the header.
1154 	 */
1155 	pbclen = le16_to_cpu(hdr->pbc[0]);
1156 	if (PBC2LRH(pbclen) != lrhlen) {
1157 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1158 		hdr->pbc[0] = cpu_to_le16(pbclen);
1159 		hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1160 		/*
1161 		 * Third packet
1162 		 * This is the first packet in the sequence that has
1163 		 * a "static" size that can be used for the rest of
1164 		 * the packets (besides the last one).
1165 		 */
1166 		if (unlikely(req->seqnum == 2)) {
1167 			/*
1168 			 * From this point on the lengths in both the
1169 			 * PBC and LRH are the same until the last
1170 			 * packet.
1171 			 * Adjust the template so we don't have to update
1172 			 * every packet
1173 			 */
1174 			req->hdr.pbc[0] = hdr->pbc[0];
1175 			req->hdr.lrh[2] = hdr->lrh[2];
1176 		}
1177 	}
1178 	/*
1179 	 * We only have to modify the header if this is not the
1180 	 * first packet in the request. Otherwise, we use the
1181 	 * header given to us.
1182 	 */
1183 	if (unlikely(!req->seqnum)) {
1184 		ret = check_header_template(req, hdr, lrhlen, datalen);
1185 		if (ret)
1186 			return ret;
1187 		goto done;
1188 	}
1189 
1190 	hdr->bth[2] = cpu_to_be32(
1191 		set_pkt_bth_psn(hdr->bth[2],
1192 				(req_opcode(req->info.ctrl) == EXPECTED),
1193 				req->seqnum));
1194 
1195 	/* Set ACK request on last packet */
1196 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1197 		hdr->bth[2] |= cpu_to_be32(1UL << 31);
1198 
1199 	/* Set the new offset */
1200 	hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1201 	/* Expected packets have to fill in the new TID information */
1202 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1203 		tidval = req->tids[req->tididx];
1204 		/*
1205 		 * If the offset puts us at the end of the current TID,
1206 		 * advance everything.
1207 		 */
1208 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1209 					 PAGE_SIZE)) {
1210 			req->tidoffset = 0;
1211 			/*
1212 			 * Since we don't copy all the TIDs, all at once,
1213 			 * we have to check again.
1214 			 */
1215 			if (++req->tididx > req->n_tids - 1 ||
1216 			    !req->tids[req->tididx]) {
1217 				return -EINVAL;
1218 			}
1219 			tidval = req->tids[req->tididx];
1220 		}
1221 		omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1222 			KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE_SHIFT :
1223 			KDETH_OM_SMALL_SHIFT;
1224 		/* Set KDETH.TIDCtrl based on value for this TID. */
1225 		KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1226 			  EXP_TID_GET(tidval, CTRL));
1227 		/* Set KDETH.TID based on value for this TID */
1228 		KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1229 			  EXP_TID_GET(tidval, IDX));
1230 		/* Clear KDETH.SH when DISABLE_SH flag is set */
1231 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1232 			KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1233 		/*
1234 		 * Set the KDETH.OFFSET and KDETH.OM based on size of
1235 		 * transfer.
1236 		 */
1237 		trace_hfi1_sdma_user_tid_info(
1238 			pq->dd, pq->ctxt, pq->subctxt, req->info.comp_idx,
1239 			req->tidoffset, req->tidoffset >> omfactor,
1240 			omfactor != KDETH_OM_SMALL_SHIFT);
1241 		KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1242 			  req->tidoffset >> omfactor);
1243 		KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1244 			  omfactor != KDETH_OM_SMALL_SHIFT);
1245 	}
1246 done:
1247 	trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1248 				    req->info.comp_idx, hdr, tidval);
1249 	return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1250 }
1251 
1252 static int set_txreq_header_ahg(struct user_sdma_request *req,
1253 				struct user_sdma_txreq *tx, u32 datalen)
1254 {
1255 	u32 ahg[AHG_KDETH_ARRAY_SIZE];
1256 	int idx = 0;
1257 	u8 omfactor; /* KDETH.OM */
1258 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1259 	struct hfi1_pkt_header *hdr = &req->hdr;
1260 	u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1261 	u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1262 	size_t array_size = ARRAY_SIZE(ahg);
1263 
1264 	if (PBC2LRH(pbclen) != lrhlen) {
1265 		/* PBC.PbcLengthDWs */
1266 		idx = ahg_header_set(ahg, idx, array_size, 0, 0, 12,
1267 				     (__force u16)cpu_to_le16(LRH2PBC(lrhlen)));
1268 		if (idx < 0)
1269 			return idx;
1270 		/* LRH.PktLen (we need the full 16 bits due to byte swap) */
1271 		idx = ahg_header_set(ahg, idx, array_size, 3, 0, 16,
1272 				     (__force u16)cpu_to_be16(lrhlen >> 2));
1273 		if (idx < 0)
1274 			return idx;
1275 	}
1276 
1277 	/*
1278 	 * Do the common updates
1279 	 */
1280 	/* BTH.PSN and BTH.A */
1281 	val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1282 		(HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1283 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1284 		val32 |= 1UL << 31;
1285 	idx = ahg_header_set(ahg, idx, array_size, 6, 0, 16,
1286 			     (__force u16)cpu_to_be16(val32 >> 16));
1287 	if (idx < 0)
1288 		return idx;
1289 	idx = ahg_header_set(ahg, idx, array_size, 6, 16, 16,
1290 			     (__force u16)cpu_to_be16(val32 & 0xffff));
1291 	if (idx < 0)
1292 		return idx;
1293 	/* KDETH.Offset */
1294 	idx = ahg_header_set(ahg, idx, array_size, 15, 0, 16,
1295 			     (__force u16)cpu_to_le16(req->koffset & 0xffff));
1296 	if (idx < 0)
1297 		return idx;
1298 	idx = ahg_header_set(ahg, idx, array_size, 15, 16, 16,
1299 			     (__force u16)cpu_to_le16(req->koffset >> 16));
1300 	if (idx < 0)
1301 		return idx;
1302 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1303 		__le16 val;
1304 
1305 		tidval = req->tids[req->tididx];
1306 
1307 		/*
1308 		 * If the offset puts us at the end of the current TID,
1309 		 * advance everything.
1310 		 */
1311 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1312 					 PAGE_SIZE)) {
1313 			req->tidoffset = 0;
1314 			/*
1315 			 * Since we don't copy all the TIDs, all at once,
1316 			 * we have to check again.
1317 			 */
1318 			if (++req->tididx > req->n_tids - 1 ||
1319 			    !req->tids[req->tididx])
1320 				return -EINVAL;
1321 			tidval = req->tids[req->tididx];
1322 		}
1323 		omfactor = ((EXP_TID_GET(tidval, LEN) *
1324 				  PAGE_SIZE) >=
1325 				 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE_SHIFT :
1326 				 KDETH_OM_SMALL_SHIFT;
1327 		/* KDETH.OM and KDETH.OFFSET (TID) */
1328 		idx = ahg_header_set(
1329 				ahg, idx, array_size, 7, 0, 16,
1330 				((!!(omfactor - KDETH_OM_SMALL_SHIFT)) << 15 |
1331 				((req->tidoffset >> omfactor)
1332 				& 0x7fff)));
1333 		if (idx < 0)
1334 			return idx;
1335 		/* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1336 		val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1337 				   (EXP_TID_GET(tidval, IDX) & 0x3ff));
1338 
1339 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1340 			val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1341 						      INTR) <<
1342 					    AHG_KDETH_INTR_SHIFT));
1343 		} else {
1344 			val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1345 			       cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1346 			       cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1347 						      INTR) <<
1348 					     AHG_KDETH_INTR_SHIFT));
1349 		}
1350 
1351 		idx = ahg_header_set(ahg, idx, array_size,
1352 				     7, 16, 14, (__force u16)val);
1353 		if (idx < 0)
1354 			return idx;
1355 	}
1356 
1357 	trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1358 					req->info.comp_idx, req->sde->this_idx,
1359 					req->ahg_idx, ahg, idx, tidval);
1360 	sdma_txinit_ahg(&tx->txreq,
1361 			SDMA_TXREQ_F_USE_AHG,
1362 			datalen, req->ahg_idx, idx,
1363 			ahg, sizeof(req->hdr),
1364 			user_sdma_txreq_cb);
1365 
1366 	return idx;
1367 }
1368 
1369 /**
1370  * user_sdma_txreq_cb() - SDMA tx request completion callback.
1371  * @txreq: valid sdma tx request
1372  * @status: success/failure of request
1373  *
1374  * Called when the SDMA progress state machine gets notification that
1375  * the SDMA descriptors for this tx request have been processed by the
1376  * DMA engine. Called in interrupt context.
1377  * Only do work on completed sequences.
1378  */
1379 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1380 {
1381 	struct user_sdma_txreq *tx =
1382 		container_of(txreq, struct user_sdma_txreq, txreq);
1383 	struct user_sdma_request *req;
1384 	struct hfi1_user_sdma_pkt_q *pq;
1385 	struct hfi1_user_sdma_comp_q *cq;
1386 	enum hfi1_sdma_comp_state state = COMPLETE;
1387 
1388 	if (!tx->req)
1389 		return;
1390 
1391 	req = tx->req;
1392 	pq = req->pq;
1393 	cq = req->cq;
1394 
1395 	if (status != SDMA_TXREQ_S_OK) {
1396 		SDMA_DBG(req, "SDMA completion with error %d",
1397 			 status);
1398 		WRITE_ONCE(req->has_error, 1);
1399 		state = ERROR;
1400 	}
1401 
1402 	req->seqcomp = tx->seqnum;
1403 	kmem_cache_free(pq->txreq_cache, tx);
1404 
1405 	/* sequence isn't complete?  We are done */
1406 	if (req->seqcomp != req->info.npkts - 1)
1407 		return;
1408 
1409 	user_sdma_free_request(req, false);
1410 	set_comp_state(pq, cq, req->info.comp_idx, state, status);
1411 	pq_update(pq);
1412 }
1413 
1414 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1415 {
1416 	if (atomic_dec_and_test(&pq->n_reqs))
1417 		wake_up(&pq->wait);
1418 }
1419 
1420 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1421 {
1422 	int i;
1423 
1424 	if (!list_empty(&req->txps)) {
1425 		struct sdma_txreq *t, *p;
1426 
1427 		list_for_each_entry_safe(t, p, &req->txps, list) {
1428 			struct user_sdma_txreq *tx =
1429 				container_of(t, struct user_sdma_txreq, txreq);
1430 			list_del_init(&t->list);
1431 			sdma_txclean(req->pq->dd, t);
1432 			kmem_cache_free(req->pq->txreq_cache, tx);
1433 		}
1434 	}
1435 
1436 	for (i = 0; i < req->data_iovs; i++) {
1437 		struct sdma_mmu_node *node = req->iovs[i].node;
1438 
1439 		if (!node)
1440 			continue;
1441 
1442 		req->iovs[i].node = NULL;
1443 
1444 		if (unpin)
1445 			hfi1_mmu_rb_remove(req->pq->handler,
1446 					   &node->rb);
1447 		else
1448 			atomic_dec(&node->refcount);
1449 	}
1450 
1451 	kfree(req->tids);
1452 	clear_bit(req->info.comp_idx, req->pq->req_in_use);
1453 }
1454 
1455 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1456 				  struct hfi1_user_sdma_comp_q *cq,
1457 				  u16 idx, enum hfi1_sdma_comp_state state,
1458 				  int ret)
1459 {
1460 	if (state == ERROR)
1461 		cq->comps[idx].errcode = -ret;
1462 	smp_wmb(); /* make sure errcode is visible first */
1463 	cq->comps[idx].status = state;
1464 	trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1465 					idx, state, ret);
1466 }
1467 
1468 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1469 			   unsigned long len)
1470 {
1471 	return (bool)(node->addr == addr);
1472 }
1473 
1474 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1475 {
1476 	struct sdma_mmu_node *node =
1477 		container_of(mnode, struct sdma_mmu_node, rb);
1478 
1479 	atomic_inc(&node->refcount);
1480 	return 0;
1481 }
1482 
1483 /*
1484  * Return 1 to remove the node from the rb tree and call the remove op.
1485  *
1486  * Called with the rb tree lock held.
1487  */
1488 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1489 			 void *evict_arg, bool *stop)
1490 {
1491 	struct sdma_mmu_node *node =
1492 		container_of(mnode, struct sdma_mmu_node, rb);
1493 	struct evict_data *evict_data = evict_arg;
1494 
1495 	/* is this node still being used? */
1496 	if (atomic_read(&node->refcount))
1497 		return 0; /* keep this node */
1498 
1499 	/* this node will be evicted, add its pages to our count */
1500 	evict_data->cleared += node->npages;
1501 
1502 	/* have enough pages been cleared? */
1503 	if (evict_data->cleared >= evict_data->target)
1504 		*stop = true;
1505 
1506 	return 1; /* remove this node */
1507 }
1508 
1509 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1510 {
1511 	struct sdma_mmu_node *node =
1512 		container_of(mnode, struct sdma_mmu_node, rb);
1513 
1514 	unpin_sdma_pages(node);
1515 	kfree(node);
1516 }
1517 
1518 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1519 {
1520 	struct sdma_mmu_node *node =
1521 		container_of(mnode, struct sdma_mmu_node, rb);
1522 
1523 	if (!atomic_read(&node->refcount))
1524 		return 1;
1525 	return 0;
1526 }
1527