xref: /linux/drivers/infiniband/hw/hfi1/user_sdma.c (revision 98838d95075a5295f3478ceba18bcccf472e30f4)
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 
119 #define PBC2LRH(x) ((((x) & 0xfff) << 2) - 4)
120 #define LRH2PBC(x) ((((x) >> 2) + 1) & 0xfff)
121 
122 #define KDETH_GET(val, field)						\
123 	(((le32_to_cpu((val))) >> KDETH_##field##_SHIFT) & KDETH_##field##_MASK)
124 #define KDETH_SET(dw, field, val) do {					\
125 		u32 dwval = le32_to_cpu(dw);				\
126 		dwval &= ~(KDETH_##field##_MASK << KDETH_##field##_SHIFT); \
127 		dwval |= (((val) & KDETH_##field##_MASK) << \
128 			  KDETH_##field##_SHIFT);			\
129 		dw = cpu_to_le32(dwval);				\
130 	} while (0)
131 
132 #define AHG_HEADER_SET(arr, idx, dw, bit, width, value)			\
133 	do {								\
134 		if ((idx) < ARRAY_SIZE((arr)))				\
135 			(arr)[(idx++)] = sdma_build_ahg_descriptor(	\
136 				(__force u16)(value), (dw), (bit),	\
137 							(width));	\
138 		else							\
139 			return -ERANGE;					\
140 	} while (0)
141 
142 /* KDETH OM multipliers and switch over point */
143 #define KDETH_OM_SMALL     4
144 #define KDETH_OM_LARGE     64
145 #define KDETH_OM_MAX_SIZE  (1 << ((KDETH_OM_LARGE / KDETH_OM_SMALL) + 1))
146 
147 /* Last packet in the request */
148 #define TXREQ_FLAGS_REQ_LAST_PKT BIT(0)
149 
150 /* SDMA request flag bits */
151 #define SDMA_REQ_FOR_THREAD 1
152 #define SDMA_REQ_SEND_DONE  2
153 #define SDMA_REQ_HAVE_AHG   3
154 #define SDMA_REQ_HAS_ERROR  4
155 #define SDMA_REQ_DONE_ERROR 5
156 
157 #define SDMA_PKT_Q_INACTIVE BIT(0)
158 #define SDMA_PKT_Q_ACTIVE   BIT(1)
159 #define SDMA_PKT_Q_DEFERRED BIT(2)
160 
161 /*
162  * Maximum retry attempts to submit a TX request
163  * before putting the process to sleep.
164  */
165 #define MAX_DEFER_RETRY_COUNT 1
166 
167 static unsigned initial_pkt_count = 8;
168 
169 #define SDMA_IOWAIT_TIMEOUT 1000 /* in milliseconds */
170 
171 struct sdma_mmu_node;
172 
173 struct user_sdma_iovec {
174 	struct list_head list;
175 	struct iovec iov;
176 	/* number of pages in this vector */
177 	unsigned npages;
178 	/* array of pinned pages for this vector */
179 	struct page **pages;
180 	/*
181 	 * offset into the virtual address space of the vector at
182 	 * which we last left off.
183 	 */
184 	u64 offset;
185 	struct sdma_mmu_node *node;
186 };
187 
188 struct sdma_mmu_node {
189 	struct mmu_rb_node rb;
190 	struct hfi1_user_sdma_pkt_q *pq;
191 	atomic_t refcount;
192 	struct page **pages;
193 	unsigned npages;
194 };
195 
196 /* evict operation argument */
197 struct evict_data {
198 	u32 cleared;	/* count evicted so far */
199 	u32 target;	/* target count to evict */
200 };
201 
202 struct user_sdma_request {
203 	struct sdma_req_info info;
204 	struct hfi1_user_sdma_pkt_q *pq;
205 	struct hfi1_user_sdma_comp_q *cq;
206 	/* This is the original header from user space */
207 	struct hfi1_pkt_header hdr;
208 	/*
209 	 * Pointer to the SDMA engine for this request.
210 	 * Since different request could be on different VLs,
211 	 * each request will need it's own engine pointer.
212 	 */
213 	struct sdma_engine *sde;
214 	u8 ahg_idx;
215 	u32 ahg[9];
216 	/*
217 	 * KDETH.Offset (Eager) field
218 	 * We need to remember the initial value so the headers
219 	 * can be updated properly.
220 	 */
221 	u32 koffset;
222 	/*
223 	 * KDETH.OFFSET (TID) field
224 	 * The offset can cover multiple packets, depending on the
225 	 * size of the TID entry.
226 	 */
227 	u32 tidoffset;
228 	/*
229 	 * KDETH.OM
230 	 * Remember this because the header template always sets it
231 	 * to 0.
232 	 */
233 	u8 omfactor;
234 	/*
235 	 * We copy the iovs for this request (based on
236 	 * info.iovcnt). These are only the data vectors
237 	 */
238 	unsigned data_iovs;
239 	/* total length of the data in the request */
240 	u32 data_len;
241 	/* progress index moving along the iovs array */
242 	unsigned iov_idx;
243 	struct user_sdma_iovec iovs[MAX_VECTORS_PER_REQ];
244 	/* number of elements copied to the tids array */
245 	u16 n_tids;
246 	/* TID array values copied from the tid_iov vector */
247 	u32 *tids;
248 	u16 tididx;
249 	u32 sent;
250 	u64 seqnum;
251 	u64 seqcomp;
252 	u64 seqsubmitted;
253 	struct list_head txps;
254 	unsigned long flags;
255 	/* status of the last txreq completed */
256 	int status;
257 };
258 
259 /*
260  * A single txreq could span up to 3 physical pages when the MTU
261  * is sufficiently large (> 4K). Each of the IOV pointers also
262  * needs it's own set of flags so the vector has been handled
263  * independently of each other.
264  */
265 struct user_sdma_txreq {
266 	/* Packet header for the txreq */
267 	struct hfi1_pkt_header hdr;
268 	struct sdma_txreq txreq;
269 	struct list_head list;
270 	struct user_sdma_request *req;
271 	u16 flags;
272 	unsigned busycount;
273 	u64 seqnum;
274 };
275 
276 #define SDMA_DBG(req, fmt, ...)				     \
277 	hfi1_cdbg(SDMA, "[%u:%u:%u:%u] " fmt, (req)->pq->dd->unit, \
278 		 (req)->pq->ctxt, (req)->pq->subctxt, (req)->info.comp_idx, \
279 		 ##__VA_ARGS__)
280 #define SDMA_Q_DBG(pq, fmt, ...)			 \
281 	hfi1_cdbg(SDMA, "[%u:%u:%u] " fmt, (pq)->dd->unit, (pq)->ctxt, \
282 		 (pq)->subctxt, ##__VA_ARGS__)
283 
284 static int user_sdma_send_pkts(struct user_sdma_request *, unsigned);
285 static int num_user_pages(const struct iovec *);
286 static void user_sdma_txreq_cb(struct sdma_txreq *, int);
287 static inline void pq_update(struct hfi1_user_sdma_pkt_q *);
288 static void user_sdma_free_request(struct user_sdma_request *, bool);
289 static int pin_vector_pages(struct user_sdma_request *,
290 			    struct user_sdma_iovec *);
291 static void unpin_vector_pages(struct mm_struct *, struct page **, unsigned,
292 			       unsigned);
293 static int check_header_template(struct user_sdma_request *,
294 				 struct hfi1_pkt_header *, u32, u32);
295 static int set_txreq_header(struct user_sdma_request *,
296 			    struct user_sdma_txreq *, u32);
297 static int set_txreq_header_ahg(struct user_sdma_request *,
298 				struct user_sdma_txreq *, u32);
299 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *,
300 				  struct hfi1_user_sdma_comp_q *,
301 				  u16, enum hfi1_sdma_comp_state, int);
302 static inline u32 set_pkt_bth_psn(__be32, u8, u32);
303 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
304 
305 static int defer_packet_queue(
306 	struct sdma_engine *,
307 	struct iowait *,
308 	struct sdma_txreq *,
309 	unsigned seq);
310 static void activate_packet_queue(struct iowait *, int);
311 static bool sdma_rb_filter(struct mmu_rb_node *, unsigned long, unsigned long);
312 static int sdma_rb_insert(void *, struct mmu_rb_node *);
313 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
314 			 void *arg2, bool *stop);
315 static void sdma_rb_remove(void *, struct mmu_rb_node *);
316 static int sdma_rb_invalidate(void *, struct mmu_rb_node *);
317 
318 static struct mmu_rb_ops sdma_rb_ops = {
319 	.filter = sdma_rb_filter,
320 	.insert = sdma_rb_insert,
321 	.evict = sdma_rb_evict,
322 	.remove = sdma_rb_remove,
323 	.invalidate = sdma_rb_invalidate
324 };
325 
326 static int defer_packet_queue(
327 	struct sdma_engine *sde,
328 	struct iowait *wait,
329 	struct sdma_txreq *txreq,
330 	unsigned seq)
331 {
332 	struct hfi1_user_sdma_pkt_q *pq =
333 		container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
334 	struct hfi1_ibdev *dev = &pq->dd->verbs_dev;
335 	struct user_sdma_txreq *tx =
336 		container_of(txreq, struct user_sdma_txreq, txreq);
337 
338 	if (sdma_progress(sde, seq, txreq)) {
339 		if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
340 			goto eagain;
341 	}
342 	/*
343 	 * We are assuming that if the list is enqueued somewhere, it
344 	 * is to the dmawait list since that is the only place where
345 	 * it is supposed to be enqueued.
346 	 */
347 	xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
348 	write_seqlock(&dev->iowait_lock);
349 	if (list_empty(&pq->busy.list))
350 		list_add_tail(&pq->busy.list, &sde->dmawait);
351 	write_sequnlock(&dev->iowait_lock);
352 	return -EBUSY;
353 eagain:
354 	return -EAGAIN;
355 }
356 
357 static void activate_packet_queue(struct iowait *wait, int reason)
358 {
359 	struct hfi1_user_sdma_pkt_q *pq =
360 		container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
361 	xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
362 	wake_up(&wait->wait_dma);
363 };
364 
365 static void sdma_kmem_cache_ctor(void *obj)
366 {
367 	struct user_sdma_txreq *tx = obj;
368 
369 	memset(tx, 0, sizeof(*tx));
370 }
371 
372 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt, struct file *fp)
373 {
374 	struct hfi1_filedata *fd;
375 	int ret = 0;
376 	unsigned memsize;
377 	char buf[64];
378 	struct hfi1_devdata *dd;
379 	struct hfi1_user_sdma_comp_q *cq;
380 	struct hfi1_user_sdma_pkt_q *pq;
381 	unsigned long flags;
382 
383 	if (!uctxt || !fp) {
384 		ret = -EBADF;
385 		goto done;
386 	}
387 
388 	fd = fp->private_data;
389 
390 	if (!hfi1_sdma_comp_ring_size) {
391 		ret = -EINVAL;
392 		goto done;
393 	}
394 
395 	dd = uctxt->dd;
396 
397 	pq = kzalloc(sizeof(*pq), GFP_KERNEL);
398 	if (!pq)
399 		goto pq_nomem;
400 
401 	memsize = sizeof(*pq->reqs) * hfi1_sdma_comp_ring_size;
402 	pq->reqs = kzalloc(memsize, GFP_KERNEL);
403 	if (!pq->reqs)
404 		goto pq_reqs_nomem;
405 
406 	memsize = BITS_TO_LONGS(hfi1_sdma_comp_ring_size) * sizeof(long);
407 	pq->req_in_use = kzalloc(memsize, GFP_KERNEL);
408 	if (!pq->req_in_use)
409 		goto pq_reqs_no_in_use;
410 
411 	INIT_LIST_HEAD(&pq->list);
412 	pq->dd = dd;
413 	pq->ctxt = uctxt->ctxt;
414 	pq->subctxt = fd->subctxt;
415 	pq->n_max_reqs = hfi1_sdma_comp_ring_size;
416 	pq->state = SDMA_PKT_Q_INACTIVE;
417 	atomic_set(&pq->n_reqs, 0);
418 	init_waitqueue_head(&pq->wait);
419 	atomic_set(&pq->n_locked, 0);
420 	pq->mm = fd->mm;
421 
422 	iowait_init(&pq->busy, 0, NULL, defer_packet_queue,
423 		    activate_packet_queue, NULL);
424 	pq->reqidx = 0;
425 	snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
426 		 fd->subctxt);
427 	pq->txreq_cache = kmem_cache_create(buf,
428 			       sizeof(struct user_sdma_txreq),
429 					    L1_CACHE_BYTES,
430 					    SLAB_HWCACHE_ALIGN,
431 					    sdma_kmem_cache_ctor);
432 	if (!pq->txreq_cache) {
433 		dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
434 			   uctxt->ctxt);
435 		goto pq_txreq_nomem;
436 	}
437 	fd->pq = pq;
438 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
439 	if (!cq)
440 		goto cq_nomem;
441 
442 	memsize = PAGE_ALIGN(sizeof(*cq->comps) * hfi1_sdma_comp_ring_size);
443 	cq->comps = vmalloc_user(memsize);
444 	if (!cq->comps)
445 		goto cq_comps_nomem;
446 
447 	cq->nentries = hfi1_sdma_comp_ring_size;
448 	fd->cq = cq;
449 
450 	ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq,
451 				   &pq->handler);
452 	if (ret) {
453 		dd_dev_err(dd, "Failed to register with MMU %d", ret);
454 		goto done;
455 	}
456 
457 	spin_lock_irqsave(&uctxt->sdma_qlock, flags);
458 	list_add(&pq->list, &uctxt->sdma_queues);
459 	spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
460 	goto done;
461 
462 cq_comps_nomem:
463 	kfree(cq);
464 cq_nomem:
465 	kmem_cache_destroy(pq->txreq_cache);
466 pq_txreq_nomem:
467 	kfree(pq->req_in_use);
468 pq_reqs_no_in_use:
469 	kfree(pq->reqs);
470 pq_reqs_nomem:
471 	kfree(pq);
472 	fd->pq = NULL;
473 pq_nomem:
474 	ret = -ENOMEM;
475 done:
476 	return ret;
477 }
478 
479 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd)
480 {
481 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
482 	struct hfi1_user_sdma_pkt_q *pq;
483 	unsigned long flags;
484 
485 	hfi1_cdbg(SDMA, "[%u:%u:%u] Freeing user SDMA queues", uctxt->dd->unit,
486 		  uctxt->ctxt, fd->subctxt);
487 	pq = fd->pq;
488 	if (pq) {
489 		if (pq->handler)
490 			hfi1_mmu_rb_unregister(pq->handler);
491 		spin_lock_irqsave(&uctxt->sdma_qlock, flags);
492 		if (!list_empty(&pq->list))
493 			list_del_init(&pq->list);
494 		spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
495 		iowait_sdma_drain(&pq->busy);
496 		/* Wait until all requests have been freed. */
497 		wait_event_interruptible(
498 			pq->wait,
499 			(ACCESS_ONCE(pq->state) == SDMA_PKT_Q_INACTIVE));
500 		kfree(pq->reqs);
501 		kfree(pq->req_in_use);
502 		kmem_cache_destroy(pq->txreq_cache);
503 		kfree(pq);
504 		fd->pq = NULL;
505 	}
506 	if (fd->cq) {
507 		vfree(fd->cq->comps);
508 		kfree(fd->cq);
509 		fd->cq = NULL;
510 	}
511 	return 0;
512 }
513 
514 static u8 dlid_to_selector(u16 dlid)
515 {
516 	static u8 mapping[256];
517 	static int initialized;
518 	static u8 next;
519 	int hash;
520 
521 	if (!initialized) {
522 		memset(mapping, 0xFF, 256);
523 		initialized = 1;
524 	}
525 
526 	hash = ((dlid >> 8) ^ dlid) & 0xFF;
527 	if (mapping[hash] == 0xFF) {
528 		mapping[hash] = next;
529 		next = (next + 1) & 0x7F;
530 	}
531 
532 	return mapping[hash];
533 }
534 
535 int hfi1_user_sdma_process_request(struct file *fp, struct iovec *iovec,
536 				   unsigned long dim, unsigned long *count)
537 {
538 	int ret = 0, i;
539 	struct hfi1_filedata *fd = fp->private_data;
540 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
541 	struct hfi1_user_sdma_pkt_q *pq = fd->pq;
542 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
543 	struct hfi1_devdata *dd = pq->dd;
544 	unsigned long idx = 0;
545 	u8 pcount = initial_pkt_count;
546 	struct sdma_req_info info;
547 	struct user_sdma_request *req;
548 	u8 opcode, sc, vl;
549 	int req_queued = 0;
550 	u16 dlid;
551 	u32 selector;
552 
553 	if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
554 		hfi1_cdbg(
555 		   SDMA,
556 		   "[%u:%u:%u] First vector not big enough for header %lu/%lu",
557 		   dd->unit, uctxt->ctxt, fd->subctxt,
558 		   iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
559 		return -EINVAL;
560 	}
561 	ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
562 	if (ret) {
563 		hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
564 			  dd->unit, uctxt->ctxt, fd->subctxt, ret);
565 		return -EFAULT;
566 	}
567 
568 	trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
569 				     (u16 *)&info);
570 
571 	if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
572 		hfi1_cdbg(SDMA,
573 			  "[%u:%u:%u:%u] Invalid comp index",
574 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
575 		return -EINVAL;
576 	}
577 
578 	/*
579 	 * Sanity check the header io vector count.  Need at least 1 vector
580 	 * (header) and cannot be larger than the actual io vector count.
581 	 */
582 	if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
583 		hfi1_cdbg(SDMA,
584 			  "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
585 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
586 			  req_iovcnt(info.ctrl), dim);
587 		return -EINVAL;
588 	}
589 
590 	if (!info.fragsize) {
591 		hfi1_cdbg(SDMA,
592 			  "[%u:%u:%u:%u] Request does not specify fragsize",
593 			  dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
594 		return -EINVAL;
595 	}
596 
597 	/* Try to claim the request. */
598 	if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
599 		hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
600 			  dd->unit, uctxt->ctxt, fd->subctxt,
601 			  info.comp_idx);
602 		return -EBADSLT;
603 	}
604 	/*
605 	 * All safety checks have been done and this request has been claimed.
606 	 */
607 	hfi1_cdbg(SDMA, "[%u:%u:%u] Using req/comp entry %u\n", dd->unit,
608 		  uctxt->ctxt, fd->subctxt, info.comp_idx);
609 	req = pq->reqs + info.comp_idx;
610 	memset(req, 0, sizeof(*req));
611 	req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
612 	req->pq = pq;
613 	req->cq = cq;
614 	req->status = -1;
615 	INIT_LIST_HEAD(&req->txps);
616 
617 	memcpy(&req->info, &info, sizeof(info));
618 
619 	if (req_opcode(info.ctrl) == EXPECTED) {
620 		/* expected must have a TID info and at least one data vector */
621 		if (req->data_iovs < 2) {
622 			SDMA_DBG(req,
623 				 "Not enough vectors for expected request");
624 			ret = -EINVAL;
625 			goto free_req;
626 		}
627 		req->data_iovs--;
628 	}
629 
630 	if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
631 		SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
632 			 MAX_VECTORS_PER_REQ);
633 		ret = -EINVAL;
634 		goto free_req;
635 	}
636 	/* Copy the header from the user buffer */
637 	ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
638 			     sizeof(req->hdr));
639 	if (ret) {
640 		SDMA_DBG(req, "Failed to copy header template (%d)", ret);
641 		ret = -EFAULT;
642 		goto free_req;
643 	}
644 
645 	/* If Static rate control is not enabled, sanitize the header. */
646 	if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
647 		req->hdr.pbc[2] = 0;
648 
649 	/* Validate the opcode. Do not trust packets from user space blindly. */
650 	opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
651 	if ((opcode & USER_OPCODE_CHECK_MASK) !=
652 	     USER_OPCODE_CHECK_VAL) {
653 		SDMA_DBG(req, "Invalid opcode (%d)", opcode);
654 		ret = -EINVAL;
655 		goto free_req;
656 	}
657 	/*
658 	 * Validate the vl. Do not trust packets from user space blindly.
659 	 * VL comes from PBC, SC comes from LRH, and the VL needs to
660 	 * match the SC look up.
661 	 */
662 	vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
663 	sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
664 	      (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
665 	if (vl >= dd->pport->vls_operational ||
666 	    vl != sc_to_vlt(dd, sc)) {
667 		SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
668 		ret = -EINVAL;
669 		goto free_req;
670 	}
671 
672 	/* Checking P_KEY for requests from user-space */
673 	if (egress_pkey_check(dd->pport, req->hdr.lrh, req->hdr.bth, sc,
674 			      PKEY_CHECK_INVALID)) {
675 		ret = -EINVAL;
676 		goto free_req;
677 	}
678 
679 	/*
680 	 * Also should check the BTH.lnh. If it says the next header is GRH then
681 	 * the RXE parsing will be off and will land in the middle of the KDETH
682 	 * or miss it entirely.
683 	 */
684 	if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
685 		SDMA_DBG(req, "User tried to pass in a GRH");
686 		ret = -EINVAL;
687 		goto free_req;
688 	}
689 
690 	req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
691 	/*
692 	 * Calculate the initial TID offset based on the values of
693 	 * KDETH.OFFSET and KDETH.OM that are passed in.
694 	 */
695 	req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
696 		(KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
697 		 KDETH_OM_LARGE : KDETH_OM_SMALL);
698 	SDMA_DBG(req, "Initial TID offset %u", req->tidoffset);
699 	idx++;
700 
701 	/* Save all the IO vector structures */
702 	for (i = 0; i < req->data_iovs; i++) {
703 		INIT_LIST_HEAD(&req->iovs[i].list);
704 		memcpy(&req->iovs[i].iov, iovec + idx++, sizeof(struct iovec));
705 		ret = pin_vector_pages(req, &req->iovs[i]);
706 		if (ret) {
707 			req->status = ret;
708 			goto free_req;
709 		}
710 		req->data_len += req->iovs[i].iov.iov_len;
711 	}
712 	SDMA_DBG(req, "total data length %u", req->data_len);
713 
714 	if (pcount > req->info.npkts)
715 		pcount = req->info.npkts;
716 	/*
717 	 * Copy any TID info
718 	 * User space will provide the TID info only when the
719 	 * request type is EXPECTED. This is true even if there is
720 	 * only one packet in the request and the header is already
721 	 * setup. The reason for the singular TID case is that the
722 	 * driver needs to perform safety checks.
723 	 */
724 	if (req_opcode(req->info.ctrl) == EXPECTED) {
725 		u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
726 
727 		if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
728 			ret = -EINVAL;
729 			goto free_req;
730 		}
731 		req->tids = kcalloc(ntids, sizeof(*req->tids), GFP_KERNEL);
732 		if (!req->tids) {
733 			ret = -ENOMEM;
734 			goto free_req;
735 		}
736 		/*
737 		 * We have to copy all of the tids because they may vary
738 		 * in size and, therefore, the TID count might not be
739 		 * equal to the pkt count. However, there is no way to
740 		 * tell at this point.
741 		 */
742 		ret = copy_from_user(req->tids, iovec[idx].iov_base,
743 				     ntids * sizeof(*req->tids));
744 		if (ret) {
745 			SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
746 				 ntids, ret);
747 			ret = -EFAULT;
748 			goto free_req;
749 		}
750 		req->n_tids = ntids;
751 		idx++;
752 	}
753 
754 	dlid = be16_to_cpu(req->hdr.lrh[1]);
755 	selector = dlid_to_selector(dlid);
756 	selector += uctxt->ctxt + fd->subctxt;
757 	req->sde = sdma_select_user_engine(dd, selector, vl);
758 
759 	if (!req->sde || !sdma_running(req->sde)) {
760 		ret = -ECOMM;
761 		goto free_req;
762 	}
763 
764 	/* We don't need an AHG entry if the request contains only one packet */
765 	if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG)) {
766 		int ahg = sdma_ahg_alloc(req->sde);
767 
768 		if (likely(ahg >= 0)) {
769 			req->ahg_idx = (u8)ahg;
770 			set_bit(SDMA_REQ_HAVE_AHG, &req->flags);
771 		}
772 	}
773 
774 	set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
775 	atomic_inc(&pq->n_reqs);
776 	req_queued = 1;
777 	/* Send the first N packets in the request to buy us some time */
778 	ret = user_sdma_send_pkts(req, pcount);
779 	if (unlikely(ret < 0 && ret != -EBUSY)) {
780 		req->status = ret;
781 		goto free_req;
782 	}
783 
784 	/*
785 	 * It is possible that the SDMA engine would have processed all the
786 	 * submitted packets by the time we get here. Therefore, only set
787 	 * packet queue state to ACTIVE if there are still uncompleted
788 	 * requests.
789 	 */
790 	if (atomic_read(&pq->n_reqs))
791 		xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
792 
793 	/*
794 	 * This is a somewhat blocking send implementation.
795 	 * The driver will block the caller until all packets of the
796 	 * request have been submitted to the SDMA engine. However, it
797 	 * will not wait for send completions.
798 	 */
799 	while (!test_bit(SDMA_REQ_SEND_DONE, &req->flags)) {
800 		ret = user_sdma_send_pkts(req, pcount);
801 		if (ret < 0) {
802 			if (ret != -EBUSY) {
803 				req->status = ret;
804 				set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
805 				if (ACCESS_ONCE(req->seqcomp) ==
806 				    req->seqsubmitted - 1)
807 					goto free_req;
808 				return ret;
809 			}
810 			wait_event_interruptible_timeout(
811 				pq->busy.wait_dma,
812 				(pq->state == SDMA_PKT_Q_ACTIVE),
813 				msecs_to_jiffies(
814 					SDMA_IOWAIT_TIMEOUT));
815 		}
816 	}
817 	*count += idx;
818 	return 0;
819 free_req:
820 	user_sdma_free_request(req, true);
821 	if (req_queued)
822 		pq_update(pq);
823 	set_comp_state(pq, cq, info.comp_idx, ERROR, req->status);
824 	return ret;
825 }
826 
827 static inline u32 compute_data_length(struct user_sdma_request *req,
828 				      struct user_sdma_txreq *tx)
829 {
830 	/*
831 	 * Determine the proper size of the packet data.
832 	 * The size of the data of the first packet is in the header
833 	 * template. However, it includes the header and ICRC, which need
834 	 * to be subtracted.
835 	 * The minimum representable packet data length in a header is 4 bytes,
836 	 * therefore, when the data length request is less than 4 bytes, there's
837 	 * only one packet, and the packet data length is equal to that of the
838 	 * request data length.
839 	 * The size of the remaining packets is the minimum of the frag
840 	 * size (MTU) or remaining data in the request.
841 	 */
842 	u32 len;
843 
844 	if (!req->seqnum) {
845 		if (req->data_len < sizeof(u32))
846 			len = req->data_len;
847 		else
848 			len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
849 			       (sizeof(tx->hdr) - 4));
850 	} else if (req_opcode(req->info.ctrl) == EXPECTED) {
851 		u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
852 			PAGE_SIZE;
853 		/*
854 		 * Get the data length based on the remaining space in the
855 		 * TID pair.
856 		 */
857 		len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
858 		/* If we've filled up the TID pair, move to the next one. */
859 		if (unlikely(!len) && ++req->tididx < req->n_tids &&
860 		    req->tids[req->tididx]) {
861 			tidlen = EXP_TID_GET(req->tids[req->tididx],
862 					     LEN) * PAGE_SIZE;
863 			req->tidoffset = 0;
864 			len = min_t(u32, tidlen, req->info.fragsize);
865 		}
866 		/*
867 		 * Since the TID pairs map entire pages, make sure that we
868 		 * are not going to try to send more data that we have
869 		 * remaining.
870 		 */
871 		len = min(len, req->data_len - req->sent);
872 	} else {
873 		len = min(req->data_len - req->sent, (u32)req->info.fragsize);
874 	}
875 	SDMA_DBG(req, "Data Length = %u", len);
876 	return len;
877 }
878 
879 static inline u32 pad_len(u32 len)
880 {
881 	if (len & (sizeof(u32) - 1))
882 		len += sizeof(u32) - (len & (sizeof(u32) - 1));
883 	return len;
884 }
885 
886 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
887 {
888 	/* (Size of complete header - size of PBC) + 4B ICRC + data length */
889 	return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
890 }
891 
892 static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
893 {
894 	int ret = 0, count;
895 	unsigned npkts = 0;
896 	struct user_sdma_txreq *tx = NULL;
897 	struct hfi1_user_sdma_pkt_q *pq = NULL;
898 	struct user_sdma_iovec *iovec = NULL;
899 
900 	if (!req->pq)
901 		return -EINVAL;
902 
903 	pq = req->pq;
904 
905 	/* If tx completion has reported an error, we are done. */
906 	if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) {
907 		set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
908 		return -EFAULT;
909 	}
910 
911 	/*
912 	 * Check if we might have sent the entire request already
913 	 */
914 	if (unlikely(req->seqnum == req->info.npkts)) {
915 		if (!list_empty(&req->txps))
916 			goto dosend;
917 		return ret;
918 	}
919 
920 	if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
921 		maxpkts = req->info.npkts - req->seqnum;
922 
923 	while (npkts < maxpkts) {
924 		u32 datalen = 0, queued = 0, data_sent = 0;
925 		u64 iov_offset = 0;
926 
927 		/*
928 		 * Check whether any of the completions have come back
929 		 * with errors. If so, we are not going to process any
930 		 * more packets from this request.
931 		 */
932 		if (test_bit(SDMA_REQ_HAS_ERROR, &req->flags)) {
933 			set_bit(SDMA_REQ_DONE_ERROR, &req->flags);
934 			return -EFAULT;
935 		}
936 
937 		tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
938 		if (!tx)
939 			return -ENOMEM;
940 
941 		tx->flags = 0;
942 		tx->req = req;
943 		tx->busycount = 0;
944 		INIT_LIST_HEAD(&tx->list);
945 
946 		if (req->seqnum == req->info.npkts - 1)
947 			tx->flags |= TXREQ_FLAGS_REQ_LAST_PKT;
948 
949 		/*
950 		 * Calculate the payload size - this is min of the fragment
951 		 * (MTU) size or the remaining bytes in the request but only
952 		 * if we have payload data.
953 		 */
954 		if (req->data_len) {
955 			iovec = &req->iovs[req->iov_idx];
956 			if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) {
957 				if (++req->iov_idx == req->data_iovs) {
958 					ret = -EFAULT;
959 					goto free_txreq;
960 				}
961 				iovec = &req->iovs[req->iov_idx];
962 				WARN_ON(iovec->offset);
963 			}
964 
965 			datalen = compute_data_length(req, tx);
966 			if (!datalen) {
967 				SDMA_DBG(req,
968 					 "Request has data but pkt len is 0");
969 				ret = -EFAULT;
970 				goto free_tx;
971 			}
972 		}
973 
974 		if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags)) {
975 			if (!req->seqnum) {
976 				u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
977 				u32 lrhlen = get_lrh_len(req->hdr,
978 							 pad_len(datalen));
979 				/*
980 				 * Copy the request header into the tx header
981 				 * because the HW needs a cacheline-aligned
982 				 * address.
983 				 * This copy can be optimized out if the hdr
984 				 * member of user_sdma_request were also
985 				 * cacheline aligned.
986 				 */
987 				memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
988 				if (PBC2LRH(pbclen) != lrhlen) {
989 					pbclen = (pbclen & 0xf000) |
990 						LRH2PBC(lrhlen);
991 					tx->hdr.pbc[0] = cpu_to_le16(pbclen);
992 				}
993 				ret = sdma_txinit_ahg(&tx->txreq,
994 						      SDMA_TXREQ_F_AHG_COPY,
995 						      sizeof(tx->hdr) + datalen,
996 						      req->ahg_idx, 0, NULL, 0,
997 						      user_sdma_txreq_cb);
998 				if (ret)
999 					goto free_tx;
1000 				ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq,
1001 							&tx->hdr,
1002 							sizeof(tx->hdr));
1003 				if (ret)
1004 					goto free_txreq;
1005 			} else {
1006 				int changes;
1007 
1008 				changes = set_txreq_header_ahg(req, tx,
1009 							       datalen);
1010 				if (changes < 0)
1011 					goto free_tx;
1012 				sdma_txinit_ahg(&tx->txreq,
1013 						SDMA_TXREQ_F_USE_AHG,
1014 						datalen, req->ahg_idx, changes,
1015 						req->ahg, sizeof(req->hdr),
1016 						user_sdma_txreq_cb);
1017 			}
1018 		} else {
1019 			ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
1020 					  datalen, user_sdma_txreq_cb);
1021 			if (ret)
1022 				goto free_tx;
1023 			/*
1024 			 * Modify the header for this packet. This only needs
1025 			 * to be done if we are not going to use AHG. Otherwise,
1026 			 * the HW will do it based on the changes we gave it
1027 			 * during sdma_txinit_ahg().
1028 			 */
1029 			ret = set_txreq_header(req, tx, datalen);
1030 			if (ret)
1031 				goto free_txreq;
1032 		}
1033 
1034 		/*
1035 		 * If the request contains any data vectors, add up to
1036 		 * fragsize bytes to the descriptor.
1037 		 */
1038 		while (queued < datalen &&
1039 		       (req->sent + data_sent) < req->data_len) {
1040 			unsigned long base, offset;
1041 			unsigned pageidx, len;
1042 
1043 			base = (unsigned long)iovec->iov.iov_base;
1044 			offset = offset_in_page(base + iovec->offset +
1045 						iov_offset);
1046 			pageidx = (((iovec->offset + iov_offset +
1047 				     base) - (base & PAGE_MASK)) >> PAGE_SHIFT);
1048 			len = offset + req->info.fragsize > PAGE_SIZE ?
1049 				PAGE_SIZE - offset : req->info.fragsize;
1050 			len = min((datalen - queued), len);
1051 			ret = sdma_txadd_page(pq->dd, &tx->txreq,
1052 					      iovec->pages[pageidx],
1053 					      offset, len);
1054 			if (ret) {
1055 				SDMA_DBG(req, "SDMA txreq add page failed %d\n",
1056 					 ret);
1057 				goto free_txreq;
1058 			}
1059 			iov_offset += len;
1060 			queued += len;
1061 			data_sent += len;
1062 			if (unlikely(queued < datalen &&
1063 				     pageidx == iovec->npages &&
1064 				     req->iov_idx < req->data_iovs - 1)) {
1065 				iovec->offset += iov_offset;
1066 				iovec = &req->iovs[++req->iov_idx];
1067 				iov_offset = 0;
1068 			}
1069 		}
1070 		/*
1071 		 * The txreq was submitted successfully so we can update
1072 		 * the counters.
1073 		 */
1074 		req->koffset += datalen;
1075 		if (req_opcode(req->info.ctrl) == EXPECTED)
1076 			req->tidoffset += datalen;
1077 		req->sent += data_sent;
1078 		if (req->data_len)
1079 			iovec->offset += iov_offset;
1080 		list_add_tail(&tx->txreq.list, &req->txps);
1081 		/*
1082 		 * It is important to increment this here as it is used to
1083 		 * generate the BTH.PSN and, therefore, can't be bulk-updated
1084 		 * outside of the loop.
1085 		 */
1086 		tx->seqnum = req->seqnum++;
1087 		npkts++;
1088 	}
1089 dosend:
1090 	ret = sdma_send_txlist(req->sde, &pq->busy, &req->txps, &count);
1091 	req->seqsubmitted += count;
1092 	if (req->seqsubmitted == req->info.npkts) {
1093 		set_bit(SDMA_REQ_SEND_DONE, &req->flags);
1094 		/*
1095 		 * The txreq has already been submitted to the HW queue
1096 		 * so we can free the AHG entry now. Corruption will not
1097 		 * happen due to the sequential manner in which
1098 		 * descriptors are processed.
1099 		 */
1100 		if (test_bit(SDMA_REQ_HAVE_AHG, &req->flags))
1101 			sdma_ahg_free(req->sde, req->ahg_idx);
1102 	}
1103 	return ret;
1104 
1105 free_txreq:
1106 	sdma_txclean(pq->dd, &tx->txreq);
1107 free_tx:
1108 	kmem_cache_free(pq->txreq_cache, tx);
1109 	return ret;
1110 }
1111 
1112 /*
1113  * How many pages in this iovec element?
1114  */
1115 static inline int num_user_pages(const struct iovec *iov)
1116 {
1117 	const unsigned long addr  = (unsigned long)iov->iov_base;
1118 	const unsigned long len   = iov->iov_len;
1119 	const unsigned long spage = addr & PAGE_MASK;
1120 	const unsigned long epage = (addr + len - 1) & PAGE_MASK;
1121 
1122 	return 1 + ((epage - spage) >> PAGE_SHIFT);
1123 }
1124 
1125 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
1126 {
1127 	struct evict_data evict_data;
1128 
1129 	evict_data.cleared = 0;
1130 	evict_data.target = npages;
1131 	hfi1_mmu_rb_evict(pq->handler, &evict_data);
1132 	return evict_data.cleared;
1133 }
1134 
1135 static int pin_vector_pages(struct user_sdma_request *req,
1136 			    struct user_sdma_iovec *iovec)
1137 {
1138 	int ret = 0, pinned, npages, cleared;
1139 	struct page **pages;
1140 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1141 	struct sdma_mmu_node *node = NULL;
1142 	struct mmu_rb_node *rb_node;
1143 
1144 	rb_node = hfi1_mmu_rb_extract(pq->handler,
1145 				      (unsigned long)iovec->iov.iov_base,
1146 				      iovec->iov.iov_len);
1147 	if (rb_node && !IS_ERR(rb_node))
1148 		node = container_of(rb_node, struct sdma_mmu_node, rb);
1149 	else
1150 		rb_node = NULL;
1151 
1152 	if (!node) {
1153 		node = kzalloc(sizeof(*node), GFP_KERNEL);
1154 		if (!node)
1155 			return -ENOMEM;
1156 
1157 		node->rb.addr = (unsigned long)iovec->iov.iov_base;
1158 		node->pq = pq;
1159 		atomic_set(&node->refcount, 0);
1160 	}
1161 
1162 	npages = num_user_pages(&iovec->iov);
1163 	if (node->npages < npages) {
1164 		pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1165 		if (!pages) {
1166 			SDMA_DBG(req, "Failed page array alloc");
1167 			ret = -ENOMEM;
1168 			goto bail;
1169 		}
1170 		memcpy(pages, node->pages, node->npages * sizeof(*pages));
1171 
1172 		npages -= node->npages;
1173 
1174 retry:
1175 		if (!hfi1_can_pin_pages(pq->dd, pq->mm,
1176 					atomic_read(&pq->n_locked), npages)) {
1177 			cleared = sdma_cache_evict(pq, npages);
1178 			if (cleared >= npages)
1179 				goto retry;
1180 		}
1181 		pinned = hfi1_acquire_user_pages(pq->mm,
1182 			((unsigned long)iovec->iov.iov_base +
1183 			 (node->npages * PAGE_SIZE)), npages, 0,
1184 			pages + node->npages);
1185 		if (pinned < 0) {
1186 			kfree(pages);
1187 			ret = pinned;
1188 			goto bail;
1189 		}
1190 		if (pinned != npages) {
1191 			unpin_vector_pages(pq->mm, pages, node->npages,
1192 					   pinned);
1193 			ret = -EFAULT;
1194 			goto bail;
1195 		}
1196 		kfree(node->pages);
1197 		node->rb.len = iovec->iov.iov_len;
1198 		node->pages = pages;
1199 		node->npages += pinned;
1200 		npages = node->npages;
1201 		atomic_add(pinned, &pq->n_locked);
1202 	}
1203 	iovec->pages = node->pages;
1204 	iovec->npages = npages;
1205 	iovec->node = node;
1206 
1207 	ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1208 	if (ret) {
1209 		atomic_sub(node->npages, &pq->n_locked);
1210 		iovec->node = NULL;
1211 		goto bail;
1212 	}
1213 	return 0;
1214 bail:
1215 	if (rb_node)
1216 		unpin_vector_pages(pq->mm, node->pages, 0, node->npages);
1217 	kfree(node);
1218 	return ret;
1219 }
1220 
1221 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1222 			       unsigned start, unsigned npages)
1223 {
1224 	hfi1_release_user_pages(mm, pages + start, npages, false);
1225 	kfree(pages);
1226 }
1227 
1228 static int check_header_template(struct user_sdma_request *req,
1229 				 struct hfi1_pkt_header *hdr, u32 lrhlen,
1230 				 u32 datalen)
1231 {
1232 	/*
1233 	 * Perform safety checks for any type of packet:
1234 	 *    - transfer size is multiple of 64bytes
1235 	 *    - packet length is multiple of 4 bytes
1236 	 *    - packet length is not larger than MTU size
1237 	 *
1238 	 * These checks are only done for the first packet of the
1239 	 * transfer since the header is "given" to us by user space.
1240 	 * For the remainder of the packets we compute the values.
1241 	 */
1242 	if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1243 	    lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1244 		return -EINVAL;
1245 
1246 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1247 		/*
1248 		 * The header is checked only on the first packet. Furthermore,
1249 		 * we ensure that at least one TID entry is copied when the
1250 		 * request is submitted. Therefore, we don't have to verify that
1251 		 * tididx points to something sane.
1252 		 */
1253 		u32 tidval = req->tids[req->tididx],
1254 			tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1255 			tididx = EXP_TID_GET(tidval, IDX),
1256 			tidctrl = EXP_TID_GET(tidval, CTRL),
1257 			tidoff;
1258 		__le32 kval = hdr->kdeth.ver_tid_offset;
1259 
1260 		tidoff = KDETH_GET(kval, OFFSET) *
1261 			  (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1262 			   KDETH_OM_LARGE : KDETH_OM_SMALL);
1263 		/*
1264 		 * Expected receive packets have the following
1265 		 * additional checks:
1266 		 *     - offset is not larger than the TID size
1267 		 *     - TIDCtrl values match between header and TID array
1268 		 *     - TID indexes match between header and TID array
1269 		 */
1270 		if ((tidoff + datalen > tidlen) ||
1271 		    KDETH_GET(kval, TIDCTRL) != tidctrl ||
1272 		    KDETH_GET(kval, TID) != tididx)
1273 			return -EINVAL;
1274 	}
1275 	return 0;
1276 }
1277 
1278 /*
1279  * Correctly set the BTH.PSN field based on type of
1280  * transfer - eager packets can just increment the PSN but
1281  * expected packets encode generation and sequence in the
1282  * BTH.PSN field so just incrementing will result in errors.
1283  */
1284 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1285 {
1286 	u32 val = be32_to_cpu(bthpsn),
1287 		mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1288 			0xffffffull),
1289 		psn = val & mask;
1290 	if (expct)
1291 		psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
1292 	else
1293 		psn = psn + frags;
1294 	return psn & mask;
1295 }
1296 
1297 static int set_txreq_header(struct user_sdma_request *req,
1298 			    struct user_sdma_txreq *tx, u32 datalen)
1299 {
1300 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1301 	struct hfi1_pkt_header *hdr = &tx->hdr;
1302 	u16 pbclen;
1303 	int ret;
1304 	u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1305 
1306 	/* Copy the header template to the request before modification */
1307 	memcpy(hdr, &req->hdr, sizeof(*hdr));
1308 
1309 	/*
1310 	 * Check if the PBC and LRH length are mismatched. If so
1311 	 * adjust both in the header.
1312 	 */
1313 	pbclen = le16_to_cpu(hdr->pbc[0]);
1314 	if (PBC2LRH(pbclen) != lrhlen) {
1315 		pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1316 		hdr->pbc[0] = cpu_to_le16(pbclen);
1317 		hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1318 		/*
1319 		 * Third packet
1320 		 * This is the first packet in the sequence that has
1321 		 * a "static" size that can be used for the rest of
1322 		 * the packets (besides the last one).
1323 		 */
1324 		if (unlikely(req->seqnum == 2)) {
1325 			/*
1326 			 * From this point on the lengths in both the
1327 			 * PBC and LRH are the same until the last
1328 			 * packet.
1329 			 * Adjust the template so we don't have to update
1330 			 * every packet
1331 			 */
1332 			req->hdr.pbc[0] = hdr->pbc[0];
1333 			req->hdr.lrh[2] = hdr->lrh[2];
1334 		}
1335 	}
1336 	/*
1337 	 * We only have to modify the header if this is not the
1338 	 * first packet in the request. Otherwise, we use the
1339 	 * header given to us.
1340 	 */
1341 	if (unlikely(!req->seqnum)) {
1342 		ret = check_header_template(req, hdr, lrhlen, datalen);
1343 		if (ret)
1344 			return ret;
1345 		goto done;
1346 	}
1347 
1348 	hdr->bth[2] = cpu_to_be32(
1349 		set_pkt_bth_psn(hdr->bth[2],
1350 				(req_opcode(req->info.ctrl) == EXPECTED),
1351 				req->seqnum));
1352 
1353 	/* Set ACK request on last packet */
1354 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
1355 		hdr->bth[2] |= cpu_to_be32(1UL << 31);
1356 
1357 	/* Set the new offset */
1358 	hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1359 	/* Expected packets have to fill in the new TID information */
1360 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1361 		tidval = req->tids[req->tididx];
1362 		/*
1363 		 * If the offset puts us at the end of the current TID,
1364 		 * advance everything.
1365 		 */
1366 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1367 					 PAGE_SIZE)) {
1368 			req->tidoffset = 0;
1369 			/*
1370 			 * Since we don't copy all the TIDs, all at once,
1371 			 * we have to check again.
1372 			 */
1373 			if (++req->tididx > req->n_tids - 1 ||
1374 			    !req->tids[req->tididx]) {
1375 				return -EINVAL;
1376 			}
1377 			tidval = req->tids[req->tididx];
1378 		}
1379 		req->omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1380 			KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE : KDETH_OM_SMALL;
1381 		/* Set KDETH.TIDCtrl based on value for this TID. */
1382 		KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1383 			  EXP_TID_GET(tidval, CTRL));
1384 		/* Set KDETH.TID based on value for this TID */
1385 		KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1386 			  EXP_TID_GET(tidval, IDX));
1387 		/* Clear KDETH.SH only on the last packet */
1388 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
1389 			KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1390 		/*
1391 		 * Set the KDETH.OFFSET and KDETH.OM based on size of
1392 		 * transfer.
1393 		 */
1394 		SDMA_DBG(req, "TID offset %ubytes %uunits om%u",
1395 			 req->tidoffset, req->tidoffset / req->omfactor,
1396 			 req->omfactor != KDETH_OM_SMALL);
1397 		KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1398 			  req->tidoffset / req->omfactor);
1399 		KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1400 			  req->omfactor != KDETH_OM_SMALL);
1401 	}
1402 done:
1403 	trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1404 				    req->info.comp_idx, hdr, tidval);
1405 	return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1406 }
1407 
1408 static int set_txreq_header_ahg(struct user_sdma_request *req,
1409 				struct user_sdma_txreq *tx, u32 len)
1410 {
1411 	int diff = 0;
1412 	struct hfi1_user_sdma_pkt_q *pq = req->pq;
1413 	struct hfi1_pkt_header *hdr = &req->hdr;
1414 	u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1415 	u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(len));
1416 
1417 	if (PBC2LRH(pbclen) != lrhlen) {
1418 		/* PBC.PbcLengthDWs */
1419 		AHG_HEADER_SET(req->ahg, diff, 0, 0, 12,
1420 			       cpu_to_le16(LRH2PBC(lrhlen)));
1421 		/* LRH.PktLen (we need the full 16 bits due to byte swap) */
1422 		AHG_HEADER_SET(req->ahg, diff, 3, 0, 16,
1423 			       cpu_to_be16(lrhlen >> 2));
1424 	}
1425 
1426 	/*
1427 	 * Do the common updates
1428 	 */
1429 	/* BTH.PSN and BTH.A */
1430 	val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1431 		(HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1432 	if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT))
1433 		val32 |= 1UL << 31;
1434 	AHG_HEADER_SET(req->ahg, diff, 6, 0, 16, cpu_to_be16(val32 >> 16));
1435 	AHG_HEADER_SET(req->ahg, diff, 6, 16, 16, cpu_to_be16(val32 & 0xffff));
1436 	/* KDETH.Offset */
1437 	AHG_HEADER_SET(req->ahg, diff, 15, 0, 16,
1438 		       cpu_to_le16(req->koffset & 0xffff));
1439 	AHG_HEADER_SET(req->ahg, diff, 15, 16, 16,
1440 		       cpu_to_le16(req->koffset >> 16));
1441 	if (req_opcode(req->info.ctrl) == EXPECTED) {
1442 		__le16 val;
1443 
1444 		tidval = req->tids[req->tididx];
1445 
1446 		/*
1447 		 * If the offset puts us at the end of the current TID,
1448 		 * advance everything.
1449 		 */
1450 		if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1451 					 PAGE_SIZE)) {
1452 			req->tidoffset = 0;
1453 			/*
1454 			 * Since we don't copy all the TIDs, all at once,
1455 			 * we have to check again.
1456 			 */
1457 			if (++req->tididx > req->n_tids - 1 ||
1458 			    !req->tids[req->tididx]) {
1459 				return -EINVAL;
1460 			}
1461 			tidval = req->tids[req->tididx];
1462 		}
1463 		req->omfactor = ((EXP_TID_GET(tidval, LEN) *
1464 				  PAGE_SIZE) >=
1465 				 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE :
1466 			KDETH_OM_SMALL;
1467 		/* KDETH.OM and KDETH.OFFSET (TID) */
1468 		AHG_HEADER_SET(req->ahg, diff, 7, 0, 16,
1469 			       ((!!(req->omfactor - KDETH_OM_SMALL)) << 15 |
1470 				((req->tidoffset / req->omfactor) & 0x7fff)));
1471 		/* KDETH.TIDCtrl, KDETH.TID */
1472 		val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1473 					(EXP_TID_GET(tidval, IDX) & 0x3ff));
1474 		/* Clear KDETH.SH on last packet */
1475 		if (unlikely(tx->flags & TXREQ_FLAGS_REQ_LAST_PKT)) {
1476 			val |= cpu_to_le16(KDETH_GET(hdr->kdeth.ver_tid_offset,
1477 						     INTR) <<
1478 					   AHG_KDETH_INTR_SHIFT);
1479 			val &= cpu_to_le16(~(1U << 13));
1480 			AHG_HEADER_SET(req->ahg, diff, 7, 16, 14, val);
1481 		} else {
1482 			AHG_HEADER_SET(req->ahg, diff, 7, 16, 12, val);
1483 		}
1484 	}
1485 
1486 	trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1487 					req->info.comp_idx, req->sde->this_idx,
1488 					req->ahg_idx, req->ahg, diff, tidval);
1489 	return diff;
1490 }
1491 
1492 /*
1493  * SDMA tx request completion callback. Called when the SDMA progress
1494  * state machine gets notification that the SDMA descriptors for this
1495  * tx request have been processed by the DMA engine. Called in
1496  * interrupt context.
1497  */
1498 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1499 {
1500 	struct user_sdma_txreq *tx =
1501 		container_of(txreq, struct user_sdma_txreq, txreq);
1502 	struct user_sdma_request *req;
1503 	struct hfi1_user_sdma_pkt_q *pq;
1504 	struct hfi1_user_sdma_comp_q *cq;
1505 	u16 idx;
1506 
1507 	if (!tx->req)
1508 		return;
1509 
1510 	req = tx->req;
1511 	pq = req->pq;
1512 	cq = req->cq;
1513 
1514 	if (status != SDMA_TXREQ_S_OK) {
1515 		SDMA_DBG(req, "SDMA completion with error %d",
1516 			 status);
1517 		set_bit(SDMA_REQ_HAS_ERROR, &req->flags);
1518 	}
1519 
1520 	req->seqcomp = tx->seqnum;
1521 	kmem_cache_free(pq->txreq_cache, tx);
1522 	tx = NULL;
1523 
1524 	idx = req->info.comp_idx;
1525 	if (req->status == -1 && status == SDMA_TXREQ_S_OK) {
1526 		if (req->seqcomp == req->info.npkts - 1) {
1527 			req->status = 0;
1528 			user_sdma_free_request(req, false);
1529 			pq_update(pq);
1530 			set_comp_state(pq, cq, idx, COMPLETE, 0);
1531 		}
1532 	} else {
1533 		if (status != SDMA_TXREQ_S_OK)
1534 			req->status = status;
1535 		if (req->seqcomp == (ACCESS_ONCE(req->seqsubmitted) - 1) &&
1536 		    (test_bit(SDMA_REQ_SEND_DONE, &req->flags) ||
1537 		     test_bit(SDMA_REQ_DONE_ERROR, &req->flags))) {
1538 			user_sdma_free_request(req, false);
1539 			pq_update(pq);
1540 			set_comp_state(pq, cq, idx, ERROR, req->status);
1541 		}
1542 	}
1543 }
1544 
1545 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1546 {
1547 	if (atomic_dec_and_test(&pq->n_reqs)) {
1548 		xchg(&pq->state, SDMA_PKT_Q_INACTIVE);
1549 		wake_up(&pq->wait);
1550 	}
1551 }
1552 
1553 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1554 {
1555 	if (!list_empty(&req->txps)) {
1556 		struct sdma_txreq *t, *p;
1557 
1558 		list_for_each_entry_safe(t, p, &req->txps, list) {
1559 			struct user_sdma_txreq *tx =
1560 				container_of(t, struct user_sdma_txreq, txreq);
1561 			list_del_init(&t->list);
1562 			sdma_txclean(req->pq->dd, t);
1563 			kmem_cache_free(req->pq->txreq_cache, tx);
1564 		}
1565 	}
1566 	if (req->data_iovs) {
1567 		struct sdma_mmu_node *node;
1568 		int i;
1569 
1570 		for (i = 0; i < req->data_iovs; i++) {
1571 			node = req->iovs[i].node;
1572 			if (!node)
1573 				continue;
1574 
1575 			if (unpin)
1576 				hfi1_mmu_rb_remove(req->pq->handler,
1577 						   &node->rb);
1578 			else
1579 				atomic_dec(&node->refcount);
1580 		}
1581 	}
1582 	kfree(req->tids);
1583 	clear_bit(req->info.comp_idx, req->pq->req_in_use);
1584 }
1585 
1586 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1587 				  struct hfi1_user_sdma_comp_q *cq,
1588 				  u16 idx, enum hfi1_sdma_comp_state state,
1589 				  int ret)
1590 {
1591 	hfi1_cdbg(SDMA, "[%u:%u:%u:%u] Setting completion status %u %d",
1592 		  pq->dd->unit, pq->ctxt, pq->subctxt, idx, state, ret);
1593 	cq->comps[idx].status = state;
1594 	if (state == ERROR)
1595 		cq->comps[idx].errcode = -ret;
1596 	trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1597 					idx, state, ret);
1598 }
1599 
1600 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1601 			   unsigned long len)
1602 {
1603 	return (bool)(node->addr == addr);
1604 }
1605 
1606 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1607 {
1608 	struct sdma_mmu_node *node =
1609 		container_of(mnode, struct sdma_mmu_node, rb);
1610 
1611 	atomic_inc(&node->refcount);
1612 	return 0;
1613 }
1614 
1615 /*
1616  * Return 1 to remove the node from the rb tree and call the remove op.
1617  *
1618  * Called with the rb tree lock held.
1619  */
1620 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1621 			 void *evict_arg, bool *stop)
1622 {
1623 	struct sdma_mmu_node *node =
1624 		container_of(mnode, struct sdma_mmu_node, rb);
1625 	struct evict_data *evict_data = evict_arg;
1626 
1627 	/* is this node still being used? */
1628 	if (atomic_read(&node->refcount))
1629 		return 0; /* keep this node */
1630 
1631 	/* this node will be evicted, add its pages to our count */
1632 	evict_data->cleared += node->npages;
1633 
1634 	/* have enough pages been cleared? */
1635 	if (evict_data->cleared >= evict_data->target)
1636 		*stop = true;
1637 
1638 	return 1; /* remove this node */
1639 }
1640 
1641 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1642 {
1643 	struct sdma_mmu_node *node =
1644 		container_of(mnode, struct sdma_mmu_node, rb);
1645 
1646 	atomic_sub(node->npages, &node->pq->n_locked);
1647 
1648 	unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages);
1649 
1650 	kfree(node);
1651 }
1652 
1653 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1654 {
1655 	struct sdma_mmu_node *node =
1656 		container_of(mnode, struct sdma_mmu_node, rb);
1657 
1658 	if (!atomic_read(&node->refcount))
1659 		return 1;
1660 	return 0;
1661 }
1662