xref: /linux/drivers/infiniband/hw/hfi1/user_sdma.c (revision 32786fdc9506aeba98278c1844d4bfb766863832)
1 /*
2  * Copyright(c) 2015, 2016 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 
64 #include "hfi.h"
65 #include "sdma.h"
66 #include "user_sdma.h"
67 #include "verbs.h"  /* for the headers */
68 #include "common.h" /* for struct hfi1_tid_info */
69 #include "trace.h"
70 #include "mmu_rb.h"
71 
72 static uint hfi1_sdma_comp_ring_size = 128;
73 module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO);
74 MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128");
75 
76 /* The maximum number of Data io vectors per message/request */
77 #define MAX_VECTORS_PER_REQ 8
78 /*
79  * Maximum number of packet to send from each message/request
80  * before moving to the next one.
81  */
82 #define MAX_PKTS_PER_QUEUE 16
83 
84 #define num_pages(x) (1 + ((((x) - 1) & PAGE_MASK) >> PAGE_SHIFT))
85 
86 #define req_opcode(x) \
87 	(((x) >> HFI1_SDMA_REQ_OPCODE_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
88 #define req_version(x) \
89 	(((x) >> HFI1_SDMA_REQ_VERSION_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
90 #define req_iovcnt(x) \
91 	(((x) >> HFI1_SDMA_REQ_IOVCNT_SHIFT) & HFI1_SDMA_REQ_IOVCNT_MASK)
92 
93 /* Number of BTH.PSN bits used for sequence number in expected rcvs */
94 #define BTH_SEQ_MASK 0x7ffull
95 
96 /*
97  * Define fields in the KDETH header so we can update the header
98  * template.
99  */
100 #define KDETH_OFFSET_SHIFT        0
101 #define KDETH_OFFSET_MASK         0x7fff
102 #define KDETH_OM_SHIFT            15
103 #define KDETH_OM_MASK             0x1
104 #define KDETH_TID_SHIFT           16
105 #define KDETH_TID_MASK            0x3ff
106 #define KDETH_TIDCTRL_SHIFT       26
107 #define KDETH_TIDCTRL_MASK        0x3
108 #define KDETH_INTR_SHIFT          28
109 #define KDETH_INTR_MASK           0x1
110 #define KDETH_SH_SHIFT            29
111 #define KDETH_SH_MASK             0x1
112 #define KDETH_HCRC_UPPER_SHIFT    16
113 #define KDETH_HCRC_UPPER_MASK     0xff
114 #define KDETH_HCRC_LOWER_SHIFT    24
115 #define KDETH_HCRC_LOWER_MASK     0xff
116 
117 #define AHG_KDETH_INTR_SHIFT 12
118 #define AHG_KDETH_SH_SHIFT   13
119 
120 #define PBC2LRH(x) ((((x) & 0xfff) << 2) - 4)
121 #define LRH2PBC(x) ((((x) >> 2) + 1) & 0xfff)
122 
123 #define KDETH_GET(val, field)						\
124 	(((le32_to_cpu((val))) >> KDETH_##field##_SHIFT) & KDETH_##field##_MASK)
125 #define KDETH_SET(dw, field, val) do {					\
126 		u32 dwval = le32_to_cpu(dw);				\
127 		dwval &= ~(KDETH_##field##_MASK << KDETH_##field##_SHIFT); \
128 		dwval |= (((val) & KDETH_##field##_MASK) << \
129 			  KDETH_##field##_SHIFT);			\
130 		dw = cpu_to_le32(dwval);				\
131 	} while (0)
132 
133 #define AHG_HEADER_SET(arr, idx, dw, bit, width, value)			\
134 	do {								\
135 		if ((idx) < ARRAY_SIZE((arr)))				\
136 			(arr)[(idx++)] = sdma_build_ahg_descriptor(	\
137 				(__force u16)(value), (dw), (bit),	\
138 							(width));	\
139 		else							\
140 			return -ERANGE;					\
141 	} while (0)
142 
143 /* KDETH OM multipliers and switch over point */
144 #define KDETH_OM_SMALL     4
145 #define KDETH_OM_LARGE     64
146 #define KDETH_OM_MAX_SIZE  (1 << ((KDETH_OM_LARGE / KDETH_OM_SMALL) + 1))
147 
148 /* Tx request flag bits */
149 #define TXREQ_FLAGS_REQ_ACK   BIT(0)      /* Set the ACK bit in the header */
150 #define TXREQ_FLAGS_REQ_DISABLE_SH BIT(1) /* Disable header suppression */
151 
152 /* SDMA request flag bits */
153 #define SDMA_REQ_FOR_THREAD 1
154 #define SDMA_REQ_SEND_DONE  2
155 #define SDMA_REQ_HAVE_AHG   3
156 #define SDMA_REQ_HAS_ERROR  4
157 #define SDMA_REQ_DONE_ERROR 5
158 
159 #define SDMA_PKT_Q_INACTIVE BIT(0)
160 #define SDMA_PKT_Q_ACTIVE   BIT(1)
161 #define SDMA_PKT_Q_DEFERRED BIT(2)
162 
163 /*
164  * Maximum retry attempts to submit a TX request
165  * before putting the process to sleep.
166  */
167 #define MAX_DEFER_RETRY_COUNT 1
168 
169 static unsigned initial_pkt_count = 8;
170 
171 #define SDMA_IOWAIT_TIMEOUT 1000 /* in milliseconds */
172 
173 struct sdma_mmu_node;
174 
175 struct user_sdma_iovec {
176 	struct list_head list;
177 	struct iovec iov;
178 	/* number of pages in this vector */
179 	unsigned npages;
180 	/* array of pinned pages for this vector */
181 	struct page **pages;
182 	/*
183 	 * offset into the virtual address space of the vector at
184 	 * which we last left off.
185 	 */
186 	u64 offset;
187 	struct sdma_mmu_node *node;
188 };
189 
190 struct sdma_mmu_node {
191 	struct mmu_rb_node rb;
192 	struct hfi1_user_sdma_pkt_q *pq;
193 	atomic_t refcount;
194 	struct page **pages;
195 	unsigned npages;
196 };
197 
198 /* evict operation argument */
199 struct evict_data {
200 	u32 cleared;	/* count evicted so far */
201 	u32 target;	/* target count to evict */
202 };
203 
204 struct user_sdma_request {
205 	struct sdma_req_info info;
206 	struct hfi1_user_sdma_pkt_q *pq;
207 	struct hfi1_user_sdma_comp_q *cq;
208 	/* This is the original header from user space */
209 	struct hfi1_pkt_header hdr;
210 	/*
211 	 * Pointer to the SDMA engine for this request.
212 	 * Since different request could be on different VLs,
213 	 * each request will need it's own engine pointer.
214 	 */
215 	struct sdma_engine *sde;
216 	u8 ahg_idx;
217 	u32 ahg[9];
218 	/*
219 	 * KDETH.Offset (Eager) field
220 	 * We need to remember the initial value so the headers
221 	 * can be updated properly.
222 	 */
223 	u32 koffset;
224 	/*
225 	 * KDETH.OFFSET (TID) field
226 	 * The offset can cover multiple packets, depending on the
227 	 * size of the TID entry.
228 	 */
229 	u32 tidoffset;
230 	/*
231 	 * KDETH.OM
232 	 * Remember this because the header template always sets it
233 	 * to 0.
234 	 */
235 	u8 omfactor;
236 	/*
237 	 * We copy the iovs for this request (based on
238 	 * info.iovcnt). These are only the data vectors
239 	 */
240 	unsigned data_iovs;
241 	/* total length of the data in the request */
242 	u32 data_len;
243 	/* progress index moving along the iovs array */
244 	unsigned iov_idx;
245 	struct user_sdma_iovec iovs[MAX_VECTORS_PER_REQ];
246 	/* number of elements copied to the tids array */
247 	u16 n_tids;
248 	/* TID array values copied from the tid_iov vector */
249 	u32 *tids;
250 	u16 tididx;
251 	u32 sent;
252 	u64 seqnum;
253 	u64 seqcomp;
254 	u64 seqsubmitted;
255 	struct list_head txps;
256 	unsigned long flags;
257 	/* status of the last txreq completed */
258 	int status;
259 };
260 
261 /*
262  * A single txreq could span up to 3 physical pages when the MTU
263  * is sufficiently large (> 4K). Each of the IOV pointers also
264  * needs it's own set of flags so the vector has been handled
265  * independently of each other.
266  */
267 struct user_sdma_txreq {
268 	/* Packet header for the txreq */
269 	struct hfi1_pkt_header hdr;
270 	struct sdma_txreq txreq;
271 	struct list_head list;
272 	struct user_sdma_request *req;
273 	u16 flags;
274 	unsigned busycount;
275 	u64 seqnum;
276 };
277 
278 #define SDMA_DBG(req, fmt, ...)				     \
279 	hfi1_cdbg(SDMA, "[%u:%u:%u:%u] " fmt, (req)->pq->dd->unit, \
280 		 (req)->pq->ctxt, (req)->pq->subctxt, (req)->info.comp_idx, \
281 		 ##__VA_ARGS__)
282 #define SDMA_Q_DBG(pq, fmt, ...)			 \
283 	hfi1_cdbg(SDMA, "[%u:%u:%u] " fmt, (pq)->dd->unit, (pq)->ctxt, \
284 		 (pq)->subctxt, ##__VA_ARGS__)
285 
286 static int user_sdma_send_pkts(struct user_sdma_request *, unsigned);
287 static int num_user_pages(const struct iovec *);
288 static void user_sdma_txreq_cb(struct sdma_txreq *, int);
289 static inline void pq_update(struct hfi1_user_sdma_pkt_q *);
290 static void user_sdma_free_request(struct user_sdma_request *, bool);
291 static int pin_vector_pages(struct user_sdma_request *,
292 			    struct user_sdma_iovec *);
293 static void unpin_vector_pages(struct mm_struct *, struct page **, unsigned,
294 			       unsigned);
295 static int check_header_template(struct user_sdma_request *,
296 				 struct hfi1_pkt_header *, u32, u32);
297 static int set_txreq_header(struct user_sdma_request *,
298 			    struct user_sdma_txreq *, u32);
299 static int set_txreq_header_ahg(struct user_sdma_request *,
300 				struct user_sdma_txreq *, u32);
301 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *,
302 				  struct hfi1_user_sdma_comp_q *,
303 				  u16, enum hfi1_sdma_comp_state, int);
304 static inline u32 set_pkt_bth_psn(__be32, u8, u32);
305 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
306 
307 static int defer_packet_queue(
308 	struct sdma_engine *,
309 	struct iowait *,
310 	struct sdma_txreq *,
311 	unsigned seq);
312 static void activate_packet_queue(struct iowait *, int);
313 static bool sdma_rb_filter(struct mmu_rb_node *, unsigned long, unsigned long);
314 static int sdma_rb_insert(void *, struct mmu_rb_node *);
315 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
316 			 void *arg2, bool *stop);
317 static void sdma_rb_remove(void *, struct mmu_rb_node *);
318 static int sdma_rb_invalidate(void *, struct mmu_rb_node *);
319 
320 static struct mmu_rb_ops sdma_rb_ops = {
321 	.filter = sdma_rb_filter,
322 	.insert = sdma_rb_insert,
323 	.evict = sdma_rb_evict,
324 	.remove = sdma_rb_remove,
325 	.invalidate = sdma_rb_invalidate
326 };
327 
328 static int defer_packet_queue(
329 	struct sdma_engine *sde,
330 	struct iowait *wait,
331 	struct sdma_txreq *txreq,
332 	unsigned seq)
333 {
334 	struct hfi1_user_sdma_pkt_q *pq =
335 		container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
336 	struct hfi1_ibdev *dev = &pq->dd->verbs_dev;
337 	struct user_sdma_txreq *tx =
338 		container_of(txreq, struct user_sdma_txreq, txreq);
339 
340 	if (sdma_progress(sde, seq, txreq)) {
341 		if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
342 			goto eagain;
343 	}
344 	/*
345 	 * We are assuming that if the list is enqueued somewhere, it
346 	 * is to the dmawait list since that is the only place where
347 	 * it is supposed to be enqueued.
348 	 */
349 	xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
350 	write_seqlock(&dev->iowait_lock);
351 	if (list_empty(&pq->busy.list))
352 		list_add_tail(&pq->busy.list, &sde->dmawait);
353 	write_sequnlock(&dev->iowait_lock);
354 	return -EBUSY;
355 eagain:
356 	return -EAGAIN;
357 }
358 
359 static void activate_packet_queue(struct iowait *wait, int reason)
360 {
361 	struct hfi1_user_sdma_pkt_q *pq =
362 		container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
363 	xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
364 	wake_up(&wait->wait_dma);
365 };
366 
367 static void sdma_kmem_cache_ctor(void *obj)
368 {
369 	struct user_sdma_txreq *tx = obj;
370 
371 	memset(tx, 0, sizeof(*tx));
372 }
373 
374 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt, struct file *fp)
375 {
376 	struct hfi1_filedata *fd;
377 	int ret = 0;
378 	unsigned memsize;
379 	char buf[64];
380 	struct hfi1_devdata *dd;
381 	struct hfi1_user_sdma_comp_q *cq;
382 	struct hfi1_user_sdma_pkt_q *pq;
383 	unsigned long flags;
384 
385 	if (!uctxt || !fp) {
386 		ret = -EBADF;
387 		goto done;
388 	}
389 
390 	fd = fp->private_data;
391 
392 	if (!hfi1_sdma_comp_ring_size) {
393 		ret = -EINVAL;
394 		goto done;
395 	}
396 
397 	dd = uctxt->dd;
398 
399 	pq = kzalloc(sizeof(*pq), GFP_KERNEL);
400 	if (!pq)
401 		goto pq_nomem;
402 
403 	memsize = sizeof(*pq->reqs) * hfi1_sdma_comp_ring_size;
404 	pq->reqs = kzalloc(memsize, GFP_KERNEL);
405 	if (!pq->reqs)
406 		goto pq_reqs_nomem;
407 
408 	memsize = BITS_TO_LONGS(hfi1_sdma_comp_ring_size) * sizeof(long);
409 	pq->req_in_use = kzalloc(memsize, GFP_KERNEL);
410 	if (!pq->req_in_use)
411 		goto pq_reqs_no_in_use;
412 
413 	INIT_LIST_HEAD(&pq->list);
414 	pq->dd = dd;
415 	pq->ctxt = uctxt->ctxt;
416 	pq->subctxt = fd->subctxt;
417 	pq->n_max_reqs = hfi1_sdma_comp_ring_size;
418 	pq->state = SDMA_PKT_Q_INACTIVE;
419 	atomic_set(&pq->n_reqs, 0);
420 	init_waitqueue_head(&pq->wait);
421 	atomic_set(&pq->n_locked, 0);
422 	pq->mm = fd->mm;
423 
424 	iowait_init(&pq->busy, 0, NULL, defer_packet_queue,
425 		    activate_packet_queue, NULL);
426 	pq->reqidx = 0;
427 	snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
428 		 fd->subctxt);
429 	pq->txreq_cache = kmem_cache_create(buf,
430 			       sizeof(struct user_sdma_txreq),
431 					    L1_CACHE_BYTES,
432 					    SLAB_HWCACHE_ALIGN,
433 					    sdma_kmem_cache_ctor);
434 	if (!pq->txreq_cache) {
435 		dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
436 			   uctxt->ctxt);
437 		goto pq_txreq_nomem;
438 	}
439 	fd->pq = pq;
440 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
441 	if (!cq)
442 		goto cq_nomem;
443 
444 	memsize = PAGE_ALIGN(sizeof(*cq->comps) * hfi1_sdma_comp_ring_size);
445 	cq->comps = vmalloc_user(memsize);
446 	if (!cq->comps)
447 		goto cq_comps_nomem;
448 
449 	cq->nentries = hfi1_sdma_comp_ring_size;
450 	fd->cq = cq;
451 
452 	ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq,
453 				   &pq->handler);
454 	if (ret) {
455 		dd_dev_err(dd, "Failed to register with MMU %d", ret);
456 		goto done;
457 	}
458 
459 	spin_lock_irqsave(&uctxt->sdma_qlock, flags);
460 	list_add(&pq->list, &uctxt->sdma_queues);
461 	spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
462 	goto done;
463 
464 cq_comps_nomem:
465 	kfree(cq);
466 cq_nomem:
467 	kmem_cache_destroy(pq->txreq_cache);
468 pq_txreq_nomem:
469 	kfree(pq->req_in_use);
470 pq_reqs_no_in_use:
471 	kfree(pq->reqs);
472 pq_reqs_nomem:
473 	kfree(pq);
474 	fd->pq = NULL;
475 pq_nomem:
476 	ret = -ENOMEM;
477 done:
478 	return ret;
479 }
480 
481 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd)
482 {
483 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
484 	struct hfi1_user_sdma_pkt_q *pq;
485 	unsigned long flags;
486 
487 	hfi1_cdbg(SDMA, "[%u:%u:%u] Freeing user SDMA queues", uctxt->dd->unit,
488 		  uctxt->ctxt, fd->subctxt);
489 	pq = fd->pq;
490 	if (pq) {
491 		if (pq->handler)
492 			hfi1_mmu_rb_unregister(pq->handler);
493 		spin_lock_irqsave(&uctxt->sdma_qlock, flags);
494 		if (!list_empty(&pq->list))
495 			list_del_init(&pq->list);
496 		spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
497 		iowait_sdma_drain(&pq->busy);
498 		/* Wait until all requests have been freed. */
499 		wait_event_interruptible(
500 			pq->wait,
501 			(ACCESS_ONCE(pq->state) == SDMA_PKT_Q_INACTIVE));
502 		kfree(pq->reqs);
503 		kfree(pq->req_in_use);
504 		kmem_cache_destroy(pq->txreq_cache);
505 		kfree(pq);
506 		fd->pq = NULL;
507 	}
508 	if (fd->cq) {
509 		vfree(fd->cq->comps);
510 		kfree(fd->cq);
511 		fd->cq = NULL;
512 	}
513 	return 0;
514 }
515 
516 static u8 dlid_to_selector(u16 dlid)
517 {
518 	static u8 mapping[256];
519 	static int initialized;
520 	static u8 next;
521 	int hash;
522 
523 	if (!initialized) {
524 		memset(mapping, 0xFF, 256);
525 		initialized = 1;
526 	}
527 
528 	hash = ((dlid >> 8) ^ dlid) & 0xFF;
529 	if (mapping[hash] == 0xFF) {
530 		mapping[hash] = next;
531 		next = (next + 1) & 0x7F;
532 	}
533 
534 	return mapping[hash];
535 }
536 
537 int hfi1_user_sdma_process_request(struct file *fp, struct iovec *iovec,
538 				   unsigned long dim, unsigned long *count)
539 {
540 	int ret = 0, i;
541 	struct hfi1_filedata *fd = fp->private_data;
542 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
543 	struct hfi1_user_sdma_pkt_q *pq = fd->pq;
544 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
545 	struct hfi1_devdata *dd = pq->dd;
546 	unsigned long idx = 0;
547 	u8 pcount = initial_pkt_count;
548 	struct sdma_req_info info;
549 	struct user_sdma_request *req;
550 	u8 opcode, sc, vl;
551 	int req_queued = 0;
552 	u16 dlid;
553 	u32 selector;
554 
555 	if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
556 		hfi1_cdbg(
557 		   SDMA,
558 		   "[%u:%u:%u] First vector not big enough for header %lu/%lu",
559 		   dd->unit, uctxt->ctxt, fd->subctxt,
560 		   iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
561 		return -EINVAL;
562 	}
563 	ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
564 	if (ret) {
565 		hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
566 			  dd->unit, uctxt->ctxt, fd->subctxt, ret);
567 		return -EFAULT;
568 	}
569 
570 	trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
571 				     (u16 *)&info);
572 
573 	if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
574 		hfi1_cdbg(SDMA,
575 			  "[%u:%u:%u:%u] Invalid comp index",
576 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
577 		return -EINVAL;
578 	}
579 
580 	/*
581 	 * Sanity check the header io vector count.  Need at least 1 vector
582 	 * (header) and cannot be larger than the actual io vector count.
583 	 */
584 	if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
585 		hfi1_cdbg(SDMA,
586 			  "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
587 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
588 			  req_iovcnt(info.ctrl), dim);
589 		return -EINVAL;
590 	}
591 
592 	if (!info.fragsize) {
593 		hfi1_cdbg(SDMA,
594 			  "[%u:%u:%u:%u] Request does not specify fragsize",
595 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
596 		return -EINVAL;
597 	}
598 
599 	/* Try to claim the request. */
600 	if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
601 		hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
602 			  dd->unit, uctxt->ctxt, fd->subctxt,
603 			  info.comp_idx);
604 		return -EBADSLT;
605 	}
606 	/*
607 	 * All safety checks have been done and this request has been claimed.
608 	 */
609 	hfi1_cdbg(SDMA, "[%u:%u:%u] Using req/comp entry %u\n", dd->unit,
610 		  uctxt->ctxt, fd->subctxt, info.comp_idx);
611 	req = pq->reqs + info.comp_idx;
612 	memset(req, 0, sizeof(*req));
613 	req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
614 	req->pq = pq;
615 	req->cq = cq;
616 	req->status = -1;
617 	INIT_LIST_HEAD(&req->txps);
618 
619 	memcpy(&req->info, &info, sizeof(info));
620 
621 	if (req_opcode(info.ctrl) == EXPECTED) {
622 		/* expected must have a TID info and at least one data vector */
623 		if (req->data_iovs < 2) {
624 			SDMA_DBG(req,
625 				 "Not enough vectors for expected request");
626 			ret = -EINVAL;
627 			goto free_req;
628 		}
629 		req->data_iovs--;
630 	}
631 
632 	if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
633 		SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
634 			 MAX_VECTORS_PER_REQ);
635 		ret = -EINVAL;
636 		goto free_req;
637 	}
638 	/* Copy the header from the user buffer */
639 	ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
640 			     sizeof(req->hdr));
641 	if (ret) {
642 		SDMA_DBG(req, "Failed to copy header template (%d)", ret);
643 		ret = -EFAULT;
644 		goto free_req;
645 	}
646 
647 	/* If Static rate control is not enabled, sanitize the header. */
648 	if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
649 		req->hdr.pbc[2] = 0;
650 
651 	/* Validate the opcode. Do not trust packets from user space blindly. */
652 	opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
653 	if ((opcode & USER_OPCODE_CHECK_MASK) !=
654 	     USER_OPCODE_CHECK_VAL) {
655 		SDMA_DBG(req, "Invalid opcode (%d)", opcode);
656 		ret = -EINVAL;
657 		goto free_req;
658 	}
659 	/*
660 	 * Validate the vl. Do not trust packets from user space blindly.
661 	 * VL comes from PBC, SC comes from LRH, and the VL needs to
662 	 * match the SC look up.
663 	 */
664 	vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
665 	sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
666 	      (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
667 	if (vl >= dd->pport->vls_operational ||
668 	    vl != sc_to_vlt(dd, sc)) {
669 		SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
670 		ret = -EINVAL;
671 		goto free_req;
672 	}
673 
674 	/* Checking P_KEY for requests from user-space */
675 	if (egress_pkey_check(dd->pport, req->hdr.lrh, req->hdr.bth, sc,
676 			      PKEY_CHECK_INVALID)) {
677 		ret = -EINVAL;
678 		goto free_req;
679 	}
680 
681 	/*
682 	 * Also should check the BTH.lnh. If it says the next header is GRH then
683 	 * the RXE parsing will be off and will land in the middle of the KDETH
684 	 * or miss it entirely.
685 	 */
686 	if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
687 		SDMA_DBG(req, "User tried to pass in a GRH");
688 		ret = -EINVAL;
689 		goto free_req;
690 	}
691 
692 	req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
693 	/*
694 	 * Calculate the initial TID offset based on the values of
695 	 * KDETH.OFFSET and KDETH.OM that are passed in.
696 	 */
697 	req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
698 		(KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
699 		 KDETH_OM_LARGE : KDETH_OM_SMALL);
700 	SDMA_DBG(req, "Initial TID offset %u", req->tidoffset);
701 	idx++;
702 
703 	/* Save all the IO vector structures */
704 	for (i = 0; i < req->data_iovs; i++) {
705 		INIT_LIST_HEAD(&req->iovs[i].list);
706 		memcpy(&req->iovs[i].iov, iovec + idx++, sizeof(struct iovec));
707 		ret = pin_vector_pages(req, &req->iovs[i]);
708 		if (ret) {
709 			req->status = ret;
710 			goto free_req;
711 		}
712 		req->data_len += req->iovs[i].iov.iov_len;
713 	}
714 	SDMA_DBG(req, "total data length %u", req->data_len);
715 
716 	if (pcount > req->info.npkts)
717 		pcount = req->info.npkts;
718 	/*
719 	 * Copy any TID info
720 	 * User space will provide the TID info only when the
721 	 * request type is EXPECTED. This is true even if there is
722 	 * only one packet in the request and the header is already
723 	 * setup. The reason for the singular TID case is that the
724 	 * driver needs to perform safety checks.
725 	 */
726 	if (req_opcode(req->info.ctrl) == EXPECTED) {
727 		u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
728 
729 		if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
730 			ret = -EINVAL;
731 			goto free_req;
732 		}
733 		req->tids = kcalloc(ntids, sizeof(*req->tids), GFP_KERNEL);
734 		if (!req->tids) {
735 			ret = -ENOMEM;
736 			goto free_req;
737 		}
738 		/*
739 		 * We have to copy all of the tids because they may vary
740 		 * in size and, therefore, the TID count might not be
741 		 * equal to the pkt count. However, there is no way to
742 		 * tell at this point.
743 		 */
744 		ret = copy_from_user(req->tids, iovec[idx].iov_base,
745 				     ntids * sizeof(*req->tids));
746 		if (ret) {
747 			SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
748 				 ntids, ret);
749 			ret = -EFAULT;
750 			goto free_req;
751 		}
752 		req->n_tids = ntids;
753 		idx++;
754 	}
755 
756 	dlid = be16_to_cpu(req->hdr.lrh[1]);
757 	selector = dlid_to_selector(dlid);
758 	selector += uctxt->ctxt + fd->subctxt;
759 	req->sde = sdma_select_user_engine(dd, selector, vl);
760 
761 	if (!req->sde || !sdma_running(req->sde)) {
762 		ret = -ECOMM;
763 		goto free_req;
764 	}
765 
766 	/* We don't need an AHG entry if the request contains only one packet */
767 	if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG)) {
768 		int ahg = sdma_ahg_alloc(req->sde);
769 
770 		if (likely(ahg >= 0)) {
771 			req->ahg_idx = (u8)ahg;
772 			set_bit(SDMA_REQ_HAVE_AHG, &req->flags);
773 		}
774 	}
775 
776 	set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
777 	atomic_inc(&pq->n_reqs);
778 	req_queued = 1;
779 	/* Send the first N packets in the request to buy us some time */
780 	ret = user_sdma_send_pkts(req, pcount);
781 	if (unlikely(ret < 0 && ret != -EBUSY)) {
782 		req->status = ret;
783 		goto free_req;
784 	}
785 
786 	/*
787 	 * It is possible that the SDMA engine would have processed all the
788 	 * submitted packets by the time we get here. Therefore, only set
789 	 * packet queue state to ACTIVE if there are still uncompleted
790 	 * requests.
791 	 */
792 	if (atomic_read(&pq->n_reqs))
793 		xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
794 
795 	/*
796 	 * This is a somewhat blocking send implementation.
797 	 * The driver will block the caller until all packets of the
798 	 * request have been submitted to the SDMA engine. However, it
799 	 * will not wait for send completions.
800 	 */
801 	while (!test_bit(SDMA_REQ_SEND_DONE, &req->flags)) {
802 		ret = user_sdma_send_pkts(req, pcount);
803 		if (ret < 0) {
804 			if (ret != -EBUSY) {
805 				req->status = ret;
806 				set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
807 				if (ACCESS_ONCE(req->seqcomp) ==
808 				    req->seqsubmitted - 1)
809 					goto free_req;
810 				return ret;
811 			}
812 			wait_event_interruptible_timeout(
813 				pq->busy.wait_dma,
814 				(pq->state == SDMA_PKT_Q_ACTIVE),
815 				msecs_to_jiffies(
816 					SDMA_IOWAIT_TIMEOUT));
817 		}
818 	}
819 	*count += idx;
820 	return 0;
821 free_req:
822 	user_sdma_free_request(req, true);
823 	if (req_queued)
824 		pq_update(pq);
825 	set_comp_state(pq, cq, info.comp_idx, ERROR, req->status);
826 	return ret;
827 }
828 
829 static inline u32 compute_data_length(struct user_sdma_request *req,
830 				      struct user_sdma_txreq *tx)
831 {
832 	/*
833 	 * Determine the proper size of the packet data.
834 	 * The size of the data of the first packet is in the header
835 	 * template. However, it includes the header and ICRC, which need
836 	 * to be subtracted.
837 	 * The minimum representable packet data length in a header is 4 bytes,
838 	 * therefore, when the data length request is less than 4 bytes, there's
839 	 * only one packet, and the packet data length is equal to that of the
840 	 * request data length.
841 	 * The size of the remaining packets is the minimum of the frag
842 	 * size (MTU) or remaining data in the request.
843 	 */
844 	u32 len;
845 
846 	if (!req->seqnum) {
847 		if (req->data_len < sizeof(u32))
848 			len = req->data_len;
849 		else
850 			len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
851 			       (sizeof(tx->hdr) - 4));
852 	} else if (req_opcode(req->info.ctrl) == EXPECTED) {
853 		u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
854 			PAGE_SIZE;
855 		/*
856 		 * Get the data length based on the remaining space in the
857 		 * TID pair.
858 		 */
859 		len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
860 		/* If we've filled up the TID pair, move to the next one. */
861 		if (unlikely(!len) && ++req->tididx < req->n_tids &&
862 		    req->tids[req->tididx]) {
863 			tidlen = EXP_TID_GET(req->tids[req->tididx],
864 					     LEN) * PAGE_SIZE;
865 			req->tidoffset = 0;
866 			len = min_t(u32, tidlen, req->info.fragsize);
867 		}
868 		/*
869 		 * Since the TID pairs map entire pages, make sure that we
870 		 * are not going to try to send more data that we have
871 		 * remaining.
872 		 */
873 		len = min(len, req->data_len - req->sent);
874 	} else {
875 		len = min(req->data_len - req->sent, (u32)req->info.fragsize);
876 	}
877 	SDMA_DBG(req, "Data Length = %u", len);
878 	return len;
879 }
880 
881 static inline u32 pad_len(u32 len)
882 {
883 	if (len & (sizeof(u32) - 1))
884 		len += sizeof(u32) - (len & (sizeof(u32) - 1));
885 	return len;
886 }
887 
888 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
889 {
890 	/* (Size of complete header - size of PBC) + 4B ICRC + data length */
891 	return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
892 }
893 
894 static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
895 {
896 	int ret = 0, count;
897 	unsigned npkts = 0;
898 	struct user_sdma_txreq *tx = NULL;
899 	struct hfi1_user_sdma_pkt_q *pq = NULL;
900 	struct user_sdma_iovec *iovec = NULL;
901 
902 	if (!req->pq)
903 		return -EINVAL;
904 
905 	pq = req->pq;
906 
907 	/* If tx completion has reported an error, we are done. */
908 	if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) {
909 		set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
910 		return -EFAULT;
911 	}
912 
913 	/*
914 	 * Check if we might have sent the entire request already
915 	 */
916 	if (unlikely(req->seqnum == req->info.npkts)) {
917 		if (!list_empty(&req->txps))
918 			goto dosend;
919 		return ret;
920 	}
921 
922 	if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
923 		maxpkts = req->info.npkts - req->seqnum;
924 
925 	while (npkts < maxpkts) {
926 		u32 datalen = 0, queued = 0, data_sent = 0;
927 		u64 iov_offset = 0;
928 
929 		/*
930 		 * Check whether any of the completions have come back
931 		 * with errors. If so, we are not going to process any
932 		 * more packets from this request.
933 		 */
934 		if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) {
935 			set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
936 			return -EFAULT;
937 		}
938 
939 		tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
940 		if (!tx)
941 			return -ENOMEM;
942 
943 		tx->flags = 0;
944 		tx->req = req;
945 		tx->busycount = 0;
946 		INIT_LIST_HEAD(&tx->list);
947 
948 		/*
949 		 * For the last packet set the ACK request
950 		 * and disable header suppression.
951 		 */
952 		if (req->seqnum == req->info.npkts - 1)
953 			tx->flags |= (TXREQ_FLAGS_REQ_ACK |
954 				      TXREQ_FLAGS_REQ_DISABLE_SH);
955 
956 		/*
957 		 * Calculate the payload size - this is min of the fragment
958 		 * (MTU) size or the remaining bytes in the request but only
959 		 * if we have payload data.
960 		 */
961 		if (req->data_len) {
962 			iovec = &req->iovs[req->iov_idx];
963 			if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) {
964 				if (++req->iov_idx == req->data_iovs) {
965 					ret = -EFAULT;
966 					goto free_txreq;
967 				}
968 				iovec = &req->iovs[req->iov_idx];
969 				WARN_ON(iovec->offset);
970 			}
971 
972 			datalen = compute_data_length(req, tx);
973 
974 			/*
975 			 * Disable header suppression for the payload <= 8DWS.
976 			 * If there is an uncorrectable error in the receive
977 			 * data FIFO when the received payload size is less than
978 			 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
979 			 * not reported.There is set RHF.EccErr if the header
980 			 * is not suppressed.
981 			 */
982 			if (!datalen) {
983 				SDMA_DBG(req,
984 					 "Request has data but pkt len is 0");
985 				ret = -EFAULT;
986 				goto free_tx;
987 			} else if (datalen <= 32) {
988 				tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
989 			}
990 		}
991 
992 		if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags)) {
993 			if (!req->seqnum) {
994 				u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
995 				u32 lrhlen = get_lrh_len(req->hdr,
996 							 pad_len(datalen));
997 				/*
998 				 * Copy the request header into the tx header
999 				 * because the HW needs a cacheline-aligned
1000 				 * address.
1001 				 * This copy can be optimized out if the hdr
1002 				 * member of user_sdma_request were also
1003 				 * cacheline aligned.
1004 				 */
1005 				memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
1006 				if (PBC2LRH(pbclen) != lrhlen) {
1007 					pbclen = (pbclen & 0xf000) |
1008 						LRH2PBC(lrhlen);
1009 					tx->hdr.pbc[0] = cpu_to_le16(pbclen);
1010 				}
1011 				ret = check_header_template(req, &tx->hdr,
1012 							    lrhlen, datalen);
1013 				if (ret)
1014 					goto free_tx;
1015 				ret = sdma_txinit_ahg(&tx->txreq,
1016 						      SDMA_TXREQ_F_AHG_COPY,
1017 						      sizeof(tx->hdr) + datalen,
1018 						      req->ahg_idx, 0, NULL, 0,
1019 						      user_sdma_txreq_cb);
1020 				if (ret)
1021 					goto free_tx;
1022 				ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq,
1023 							&tx->hdr,
1024 							sizeof(tx->hdr));
1025 				if (ret)
1026 					goto free_txreq;
1027 			} else {
1028 				int changes;
1029 
1030 				changes = set_txreq_header_ahg(req, tx,
1031 							       datalen);
1032 				if (changes < 0)
1033 					goto free_tx;
1034 				sdma_txinit_ahg(&tx->txreq,
1035 						SDMA_TXREQ_F_USE_AHG,
1036 						datalen, req->ahg_idx, changes,
1037 						req->ahg, sizeof(req->hdr),
1038 						user_sdma_txreq_cb);
1039 			}
1040 		} else {
1041 			ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
1042 					  datalen, user_sdma_txreq_cb);
1043 			if (ret)
1044 				goto free_tx;
1045 			/*
1046 			 * Modify the header for this packet. This only needs
1047 			 * to be done if we are not going to use AHG. Otherwise,
1048 			 * the HW will do it based on the changes we gave it
1049 			 * during sdma_txinit_ahg().
1050 			 */
1051 			ret = set_txreq_header(req, tx, datalen);
1052 			if (ret)
1053 				goto free_txreq;
1054 		}
1055 
1056 		/*
1057 		 * If the request contains any data vectors, add up to
1058 		 * fragsize bytes to the descriptor.
1059 		 */
1060 		while (queued < datalen &&
1061 		       (req->sent + data_sent) < req->data_len) {
1062 			unsigned long base, offset;
1063 			unsigned pageidx, len;
1064 
1065 			base = (unsigned long)iovec->iov.iov_base;
1066 			offset = offset_in_page(base + iovec->offset +
1067 						iov_offset);
1068 			pageidx = (((iovec->offset + iov_offset +
1069 				     base) - (base & PAGE_MASK)) >> PAGE_SHIFT);
1070 			len = offset + req->info.fragsize > PAGE_SIZE ?
1071 				PAGE_SIZE - offset : req->info.fragsize;
1072 			len = min((datalen - queued), len);
1073 			ret = sdma_txadd_page(pq->dd, &tx->txreq,
1074 					      iovec->pages[pageidx],
1075 					      offset, len);
1076 			if (ret) {
1077 				SDMA_DBG(req, "SDMA txreq add page failed %d\n",
1078 					 ret);
1079 				goto free_txreq;
1080 			}
1081 			iov_offset += len;
1082 			queued += len;
1083 			data_sent += len;
1084 			if (unlikely(queued < datalen &&
1085 				     pageidx == iovec->npages &&
1086 				     req->iov_idx < req->data_iovs - 1)) {
1087 				iovec->offset += iov_offset;
1088 				iovec = &req->iovs[++req->iov_idx];
1089 				iov_offset = 0;
1090 			}
1091 		}
1092 		/*
1093 		 * The txreq was submitted successfully so we can update
1094 		 * the counters.
1095 		 */
1096 		req->koffset += datalen;
1097 		if (req_opcode(req->info.ctrl) == EXPECTED)
1098 			req->tidoffset += datalen;
1099 		req->sent += data_sent;
1100 		if (req->data_len)
1101 			iovec->offset += iov_offset;
1102 		list_add_tail(&tx->txreq.list, &req->txps);
1103 		/*
1104 		 * It is important to increment this here as it is used to
1105 		 * generate the BTH.PSN and, therefore, can't be bulk-updated
1106 		 * outside of the loop.
1107 		 */
1108 		tx->seqnum = req->seqnum++;
1109 		npkts++;
1110 	}
1111 dosend:
1112 	ret = sdma_send_txlist(req->sde, &pq->busy, &req->txps, &count);
1113 	req->seqsubmitted += count;
1114 	if (req->seqsubmitted == req->info.npkts) {
1115 		set_bit(SDMA_REQ_SEND_DONE, &req->flags);
1116 		/*
1117 		 * The txreq has already been submitted to the HW queue
1118 		 * so we can free the AHG entry now. Corruption will not
1119 		 * happen due to the sequential manner in which
1120 		 * descriptors are processed.
1121 		 */
1122 		if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags))
1123 			sdma_ahg_free(req->sde, req->ahg_idx);
1124 	}
1125 	return ret;
1126 
1127 free_txreq:
1128 	sdma_txclean(pq->dd, &tx->txreq);
1129 free_tx:
1130 	kmem_cache_free(pq->txreq_cache, tx);
1131 	return ret;
1132 }
1133 
1134 /*
1135  * How many pages in this iovec element?
1136  */
1137 static inline int num_user_pages(const struct iovec *iov)
1138 {
1139 	const unsigned long addr  = (unsigned long)iov->iov_base;
1140 	const unsigned long len   = iov->iov_len;
1141 	const unsigned long spage = addr & PAGE_MASK;
1142 	const unsigned long epage = (addr + len - 1) & PAGE_MASK;
1143 
1144 	return 1 + ((epage - spage) >> PAGE_SHIFT);
1145 }
1146 
1147 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
1148 {
1149 	struct evict_data evict_data;
1150 
1151 	evict_data.cleared = 0;
1152 	evict_data.target = npages;
1153 	hfi1_mmu_rb_evict(pq->handler, &evict_data);
1154 	return evict_data.cleared;
1155 }
1156 
1157 static int pin_vector_pages(struct user_sdma_request *req,
1158 			    struct user_sdma_iovec *iovec)
1159 {
1160 	int ret = 0, pinned, npages, cleared;
1161 	struct page **pages;
1162 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1163 	struct sdma_mmu_node *node = NULL;
1164 	struct mmu_rb_node *rb_node;
1165 
1166 	rb_node = hfi1_mmu_rb_extract(pq->handler,
1167 				      (unsigned long)iovec->iov.iov_base,
1168 				      iovec->iov.iov_len);
1169 	if (rb_node)
1170 		node = container_of(rb_node, struct sdma_mmu_node, rb);
1171 	else
1172 		rb_node = NULL;
1173 
1174 	if (!node) {
1175 		node = kzalloc(sizeof(*node), GFP_KERNEL);
1176 		if (!node)
1177 			return -ENOMEM;
1178 
1179 		node->rb.addr = (unsigned long)iovec->iov.iov_base;
1180 		node->pq = pq;
1181 		atomic_set(&node->refcount, 0);
1182 	}
1183 
1184 	npages = num_user_pages(&iovec->iov);
1185 	if (node->npages < npages) {
1186 		pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1187 		if (!pages) {
1188 			SDMA_DBG(req, "Failed page array alloc");
1189 			ret = -ENOMEM;
1190 			goto bail;
1191 		}
1192 		memcpy(pages, node->pages, node->npages * sizeof(*pages));
1193 
1194 		npages -= node->npages;
1195 
1196 retry:
1197 		if (!hfi1_can_pin_pages(pq->dd, pq->mm,
1198 					atomic_read(&pq->n_locked), npages)) {
1199 			cleared = sdma_cache_evict(pq, npages);
1200 			if (cleared >= npages)
1201 				goto retry;
1202 		}
1203 		pinned = hfi1_acquire_user_pages(pq->mm,
1204 			((unsigned long)iovec->iov.iov_base +
1205 			 (node->npages * PAGE_SIZE)), npages, 0,
1206 			pages + node->npages);
1207 		if (pinned < 0) {
1208 			kfree(pages);
1209 			ret = pinned;
1210 			goto bail;
1211 		}
1212 		if (pinned != npages) {
1213 			unpin_vector_pages(pq->mm, pages, node->npages,
1214 					   pinned);
1215 			ret = -EFAULT;
1216 			goto bail;
1217 		}
1218 		kfree(node->pages);
1219 		node->rb.len = iovec->iov.iov_len;
1220 		node->pages = pages;
1221 		node->npages += pinned;
1222 		npages = node->npages;
1223 		atomic_add(pinned, &pq->n_locked);
1224 	}
1225 	iovec->pages = node->pages;
1226 	iovec->npages = npages;
1227 	iovec->node = node;
1228 
1229 	ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1230 	if (ret) {
1231 		atomic_sub(node->npages, &pq->n_locked);
1232 		iovec->node = NULL;
1233 		goto bail;
1234 	}
1235 	return 0;
1236 bail:
1237 	if (rb_node)
1238 		unpin_vector_pages(pq->mm, node->pages, 0, node->npages);
1239 	kfree(node);
1240 	return ret;
1241 }
1242 
1243 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1244 			       unsigned start, unsigned npages)
1245 {
1246 	hfi1_release_user_pages(mm, pages + start, npages, false);
1247 	kfree(pages);
1248 }
1249 
1250 static int check_header_template(struct user_sdma_request *req,
1251 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
1252 				 u32 datalen)
1253 {
1254 	/*
1255 	 * Perform safety checks for any type of packet:
1256 	 *    - transfer size is multiple of 64bytes
1257 	 *    - packet length is multiple of 4 bytes
1258 	 *    - packet length is not larger than MTU size
1259 	 *
1260 	 * These checks are only done for the first packet of the
1261 	 * transfer since the header is "given" to us by user space.
1262 	 * For the remainder of the packets we compute the values.
1263 	 */
1264 	if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1265 	    lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1266 		return -EINVAL;
1267 
1268 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1269 		/*
1270 		 * The header is checked only on the first packet. Furthermore,
1271 		 * we ensure that at least one TID entry is copied when the
1272 		 * request is submitted. Therefore, we don't have to verify that
1273 		 * tididx points to something sane.
1274 		 */
1275 		u32 tidval = req->tids[req->tididx],
1276 			tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1277 			tididx = EXP_TID_GET(tidval, IDX),
1278 			tidctrl = EXP_TID_GET(tidval, CTRL),
1279 			tidoff;
1280 		__le32 kval = hdr->kdeth.ver_tid_offset;
1281 
1282 		tidoff = KDETH_GET(kval, OFFSET) *
1283 			  (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1284 			   KDETH_OM_LARGE : KDETH_OM_SMALL);
1285 		/*
1286 		 * Expected receive packets have the following
1287 		 * additional checks:
1288 		 *     - offset is not larger than the TID size
1289 		 *     - TIDCtrl values match between header and TID array
1290 		 *     - TID indexes match between header and TID array
1291 		 */
1292 		if ((tidoff + datalen > tidlen) ||
1293 		    KDETH_GET(kval, TIDCTRL) != tidctrl ||
1294 		    KDETH_GET(kval, TID) != tididx)
1295 			return -EINVAL;
1296 	}
1297 	return 0;
1298 }
1299 
1300 /*
1301  * Correctly set the BTH.PSN field based on type of
1302  * transfer - eager packets can just increment the PSN but
1303  * expected packets encode generation and sequence in the
1304  * BTH.PSN field so just incrementing will result in errors.
1305  */
1306 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1307 {
1308 	u32 val = be32_to_cpu(bthpsn),
1309 		mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1310 			0xffffffull),
1311 		psn = val & mask;
1312 	if (expct)
1313 		psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
1314 	else
1315 		psn = psn + frags;
1316 	return psn & mask;
1317 }
1318 
1319 static int set_txreq_header(struct user_sdma_request *req,
1320 			    struct user_sdma_txreq *tx, u32 datalen)
1321 {
1322 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1323 	struct hfi1_pkt_header *hdr = &tx->hdr;
1324 	u16 pbclen;
1325 	int ret;
1326 	u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1327 
1328 	/* Copy the header template to the request before modification */
1329 	memcpy(hdr, &req->hdr, sizeof(*hdr));
1330 
1331 	/*
1332 	 * Check if the PBC and LRH length are mismatched. If so
1333 	 * adjust both in the header.
1334 	 */
1335 	pbclen = le16_to_cpu(hdr->pbc[0]);
1336 	if (PBC2LRH(pbclen) != lrhlen) {
1337 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1338 		hdr->pbc[0] = cpu_to_le16(pbclen);
1339 		hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1340 		/*
1341 		 * Third packet
1342 		 * This is the first packet in the sequence that has
1343 		 * a "static" size that can be used for the rest of
1344 		 * the packets (besides the last one).
1345 		 */
1346 		if (unlikely(req->seqnum == 2)) {
1347 			/*
1348 			 * From this point on the lengths in both the
1349 			 * PBC and LRH are the same until the last
1350 			 * packet.
1351 			 * Adjust the template so we don't have to update
1352 			 * every packet
1353 			 */
1354 			req->hdr.pbc[0] = hdr->pbc[0];
1355 			req->hdr.lrh[2] = hdr->lrh[2];
1356 		}
1357 	}
1358 	/*
1359 	 * We only have to modify the header if this is not the
1360 	 * first packet in the request. Otherwise, we use the
1361 	 * header given to us.
1362 	 */
1363 	if (unlikely(!req->seqnum)) {
1364 		ret = check_header_template(req, hdr, lrhlen, datalen);
1365 		if (ret)
1366 			return ret;
1367 		goto done;
1368 	}
1369 
1370 	hdr->bth[2] = cpu_to_be32(
1371 		set_pkt_bth_psn(hdr->bth[2],
1372 				(req_opcode(req->info.ctrl) == EXPECTED),
1373 				req->seqnum));
1374 
1375 	/* Set ACK request on last packet */
1376 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1377 		hdr->bth[2] |= cpu_to_be32(1UL << 31);
1378 
1379 	/* Set the new offset */
1380 	hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1381 	/* Expected packets have to fill in the new TID information */
1382 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1383 		tidval = req->tids[req->tididx];
1384 		/*
1385 		 * If the offset puts us at the end of the current TID,
1386 		 * advance everything.
1387 		 */
1388 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1389 					 PAGE_SIZE)) {
1390 			req->tidoffset = 0;
1391 			/*
1392 			 * Since we don't copy all the TIDs, all at once,
1393 			 * we have to check again.
1394 			 */
1395 			if (++req->tididx > req->n_tids - 1 ||
1396 			    !req->tids[req->tididx]) {
1397 				return -EINVAL;
1398 			}
1399 			tidval = req->tids[req->tididx];
1400 		}
1401 		req->omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1402 			KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE : KDETH_OM_SMALL;
1403 		/* Set KDETH.TIDCtrl based on value for this TID. */
1404 		KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1405 			  EXP_TID_GET(tidval, CTRL));
1406 		/* Set KDETH.TID based on value for this TID */
1407 		KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1408 			  EXP_TID_GET(tidval, IDX));
1409 		/* Clear KDETH.SH when DISABLE_SH flag is set */
1410 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1411 			KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1412 		/*
1413 		 * Set the KDETH.OFFSET and KDETH.OM based on size of
1414 		 * transfer.
1415 		 */
1416 		SDMA_DBG(req, "TID offset %ubytes %uunits om%u",
1417 			 req->tidoffset, req->tidoffset / req->omfactor,
1418 			 req->omfactor != KDETH_OM_SMALL);
1419 		KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1420 			  req->tidoffset / req->omfactor);
1421 		KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1422 			  req->omfactor != KDETH_OM_SMALL);
1423 	}
1424 done:
1425 	trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1426 				    req->info.comp_idx, hdr, tidval);
1427 	return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1428 }
1429 
1430 static int set_txreq_header_ahg(struct user_sdma_request *req,
1431 				struct user_sdma_txreq *tx, u32 len)
1432 {
1433 	int diff = 0;
1434 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1435 	struct hfi1_pkt_header *hdr = &req->hdr;
1436 	u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1437 	u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(len));
1438 
1439 	if (PBC2LRH(pbclen) != lrhlen) {
1440 		/* PBC.PbcLengthDWs */
1441 		AHG_HEADER_SET(req->ahg, diff, 0, 0, 12,
1442 			       cpu_to_le16(LRH2PBC(lrhlen)));
1443 		/* LRH.PktLen (we need the full 16 bits due to byte swap) */
1444 		AHG_HEADER_SET(req->ahg, diff, 3, 0, 16,
1445 			       cpu_to_be16(lrhlen >> 2));
1446 	}
1447 
1448 	/*
1449 	 * Do the common updates
1450 	 */
1451 	/* BTH.PSN and BTH.A */
1452 	val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1453 		(HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1454 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1455 		val32 |= 1UL << 31;
1456 	AHG_HEADER_SET(req->ahg, diff, 6, 0, 16, cpu_to_be16(val32 >> 16));
1457 	AHG_HEADER_SET(req->ahg, diff, 6, 16, 16, cpu_to_be16(val32 & 0xffff));
1458 	/* KDETH.Offset */
1459 	AHG_HEADER_SET(req->ahg, diff, 15, 0, 16,
1460 		       cpu_to_le16(req->koffset & 0xffff));
1461 	AHG_HEADER_SET(req->ahg, diff, 15, 16, 16,
1462 		       cpu_to_le16(req->koffset >> 16));
1463 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1464 		__le16 val;
1465 
1466 		tidval = req->tids[req->tididx];
1467 
1468 		/*
1469 		 * If the offset puts us at the end of the current TID,
1470 		 * advance everything.
1471 		 */
1472 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1473 					 PAGE_SIZE)) {
1474 			req->tidoffset = 0;
1475 			/*
1476 			 * Since we don't copy all the TIDs, all at once,
1477 			 * we have to check again.
1478 			 */
1479 			if (++req->tididx > req->n_tids - 1 ||
1480 			    !req->tids[req->tididx]) {
1481 				return -EINVAL;
1482 			}
1483 			tidval = req->tids[req->tididx];
1484 		}
1485 		req->omfactor = ((EXP_TID_GET(tidval, LEN) *
1486 				  PAGE_SIZE) >=
1487 				 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE :
1488 			KDETH_OM_SMALL;
1489 		/* KDETH.OM and KDETH.OFFSET (TID) */
1490 		AHG_HEADER_SET(req->ahg, diff, 7, 0, 16,
1491 			       ((!!(req->omfactor - KDETH_OM_SMALL)) << 15 |
1492 				((req->tidoffset / req->omfactor) & 0x7fff)));
1493 		/* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1494 		val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1495 				   (EXP_TID_GET(tidval, IDX) & 0x3ff));
1496 
1497 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1498 			val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1499 						      INTR) <<
1500 					    AHG_KDETH_INTR_SHIFT));
1501 		} else {
1502 			val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1503 			       cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1504 			       cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1505 						      INTR) <<
1506 					     AHG_KDETH_INTR_SHIFT));
1507 		}
1508 
1509 		AHG_HEADER_SET(req->ahg, diff, 7, 16, 14, val);
1510 	}
1511 
1512 	trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1513 					req->info.comp_idx, req->sde->this_idx,
1514 					req->ahg_idx, req->ahg, diff, tidval);
1515 	return diff;
1516 }
1517 
1518 /*
1519  * SDMA tx request completion callback. Called when the SDMA progress
1520  * state machine gets notification that the SDMA descriptors for this
1521  * tx request have been processed by the DMA engine. Called in
1522  * interrupt context.
1523  */
1524 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1525 {
1526 	struct user_sdma_txreq *tx =
1527 		container_of(txreq, struct user_sdma_txreq, txreq);
1528 	struct user_sdma_request *req;
1529 	struct hfi1_user_sdma_pkt_q *pq;
1530 	struct hfi1_user_sdma_comp_q *cq;
1531 	u16 idx;
1532 
1533 	if (!tx->req)
1534 		return;
1535 
1536 	req = tx->req;
1537 	pq = req->pq;
1538 	cq = req->cq;
1539 
1540 	if (status != SDMA_TXREQ_S_OK) {
1541 		SDMA_DBG(req, "SDMA completion with error %d",
1542 			 status);
1543 		set_bit(SDMA_REQ_HAS_ERROR, &req->flags);
1544 	}
1545 
1546 	req->seqcomp = tx->seqnum;
1547 	kmem_cache_free(pq->txreq_cache, tx);
1548 	tx = NULL;
1549 
1550 	idx = req->info.comp_idx;
1551 	if (req->status == -1 && status == SDMA_TXREQ_S_OK) {
1552 		if (req->seqcomp == req->info.npkts - 1) {
1553 			req->status = 0;
1554 			user_sdma_free_request(req, false);
1555 			pq_update(pq);
1556 			set_comp_state(pq, cq, idx, COMPLETE, 0);
1557 		}
1558 	} else {
1559 		if (status != SDMA_TXREQ_S_OK)
1560 			req->status = status;
1561 		if (req->seqcomp == (ACCESS_ONCE(req->seqsubmitted) - 1) &&
1562 		    (test_bit(SDMA_REQ_SEND_DONE, &req->flags) ||
1563 		     test_bit(SDMA_REQ_DONE_ERROR, &req->flags))) {
1564 			user_sdma_free_request(req, false);
1565 			pq_update(pq);
1566 			set_comp_state(pq, cq, idx, ERROR, req->status);
1567 		}
1568 	}
1569 }
1570 
1571 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1572 {
1573 	if (atomic_dec_and_test(&pq->n_reqs)) {
1574 		xchg(&pq->state, SDMA_PKT_Q_INACTIVE);
1575 		wake_up(&pq->wait);
1576 	}
1577 }
1578 
1579 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1580 {
1581 	if (!list_empty(&req->txps)) {
1582 		struct sdma_txreq *t, *p;
1583 
1584 		list_for_each_entry_safe(t, p, &req->txps, list) {
1585 			struct user_sdma_txreq *tx =
1586 				container_of(t, struct user_sdma_txreq, txreq);
1587 			list_del_init(&t->list);
1588 			sdma_txclean(req->pq->dd, t);
1589 			kmem_cache_free(req->pq->txreq_cache, tx);
1590 		}
1591 	}
1592 	if (req->data_iovs) {
1593 		struct sdma_mmu_node *node;
1594 		int i;
1595 
1596 		for (i = 0; i < req->data_iovs; i++) {
1597 			node = req->iovs[i].node;
1598 			if (!node)
1599 				continue;
1600 
1601 			if (unpin)
1602 				hfi1_mmu_rb_remove(req->pq->handler,
1603 						   &node->rb);
1604 			else
1605 				atomic_dec(&node->refcount);
1606 		}
1607 	}
1608 	kfree(req->tids);
1609 	clear_bit(req->info.comp_idx, req->pq->req_in_use);
1610 }
1611 
1612 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1613 				  struct hfi1_user_sdma_comp_q *cq,
1614 				  u16 idx, enum hfi1_sdma_comp_state state,
1615 				  int ret)
1616 {
1617 	hfi1_cdbg(SDMA, "[%u:%u:%u:%u] Setting completion status %u %d",
1618 		  pq->dd->unit, pq->ctxt, pq->subctxt, idx, state, ret);
1619 	cq->comps[idx].status = state;
1620 	if (state == ERROR)
1621 		cq->comps[idx].errcode = -ret;
1622 	trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1623 					idx, state, ret);
1624 }
1625 
1626 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1627 			   unsigned long len)
1628 {
1629 	return (bool)(node->addr == addr);
1630 }
1631 
1632 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1633 {
1634 	struct sdma_mmu_node *node =
1635 		container_of(mnode, struct sdma_mmu_node, rb);
1636 
1637 	atomic_inc(&node->refcount);
1638 	return 0;
1639 }
1640 
1641 /*
1642  * Return 1 to remove the node from the rb tree and call the remove op.
1643  *
1644  * Called with the rb tree lock held.
1645  */
1646 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1647 			 void *evict_arg, bool *stop)
1648 {
1649 	struct sdma_mmu_node *node =
1650 		container_of(mnode, struct sdma_mmu_node, rb);
1651 	struct evict_data *evict_data = evict_arg;
1652 
1653 	/* is this node still being used? */
1654 	if (atomic_read(&node->refcount))
1655 		return 0; /* keep this node */
1656 
1657 	/* this node will be evicted, add its pages to our count */
1658 	evict_data->cleared += node->npages;
1659 
1660 	/* have enough pages been cleared? */
1661 	if (evict_data->cleared >= evict_data->target)
1662 		*stop = true;
1663 
1664 	return 1; /* remove this node */
1665 }
1666 
1667 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1668 {
1669 	struct sdma_mmu_node *node =
1670 		container_of(mnode, struct sdma_mmu_node, rb);
1671 
1672 	atomic_sub(node->npages, &node->pq->n_locked);
1673 
1674 	unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages);
1675 
1676 	kfree(node);
1677 }
1678 
1679 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1680 {
1681 	struct sdma_mmu_node *node =
1682 		container_of(mnode, struct sdma_mmu_node, rb);
1683 
1684 	if (!atomic_read(&node->refcount))
1685 		return 1;
1686 	return 0;
1687 }
1688