xref: /openbmc/linux/drivers/infiniband/hw/qib/qib_user_sdma.c (revision baa7eb025ab14f3cba2e35c0a8648f9c9f01d24f)
1 /*
2  * Copyright (c) 2007, 2008, 2009 QLogic Corporation. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #include <linux/mm.h>
33 #include <linux/types.h>
34 #include <linux/device.h>
35 #include <linux/dmapool.h>
36 #include <linux/slab.h>
37 #include <linux/list.h>
38 #include <linux/highmem.h>
39 #include <linux/io.h>
40 #include <linux/uio.h>
41 #include <linux/rbtree.h>
42 #include <linux/spinlock.h>
43 #include <linux/delay.h>
44 
45 #include "qib.h"
46 #include "qib_user_sdma.h"
47 
48 /* minimum size of header */
49 #define QIB_USER_SDMA_MIN_HEADER_LENGTH 64
50 /* expected size of headers (for dma_pool) */
51 #define QIB_USER_SDMA_EXP_HEADER_LENGTH 64
52 /* attempt to drain the queue for 5secs */
53 #define QIB_USER_SDMA_DRAIN_TIMEOUT 500
54 
55 struct qib_user_sdma_pkt {
56 	u8 naddr;               /* dimension of addr (1..3) ... */
57 	u32 counter;            /* sdma pkts queued counter for this entry */
58 	u64 added;              /* global descq number of entries */
59 
60 	struct {
61 		u32 offset;                     /* offset for kvaddr, addr */
62 		u32 length;                     /* length in page */
63 		u8  put_page;                   /* should we put_page? */
64 		u8  dma_mapped;                 /* is page dma_mapped? */
65 		struct page *page;              /* may be NULL (coherent mem) */
66 		void *kvaddr;                   /* FIXME: only for pio hack */
67 		dma_addr_t addr;
68 	} addr[4];   /* max pages, any more and we coalesce */
69 	struct list_head list;  /* list element */
70 };
71 
72 struct qib_user_sdma_queue {
73 	/*
74 	 * pkts sent to dma engine are queued on this
75 	 * list head.  the type of the elements of this
76 	 * list are struct qib_user_sdma_pkt...
77 	 */
78 	struct list_head sent;
79 
80 	/* headers with expected length are allocated from here... */
81 	char header_cache_name[64];
82 	struct dma_pool *header_cache;
83 
84 	/* packets are allocated from the slab cache... */
85 	char pkt_slab_name[64];
86 	struct kmem_cache *pkt_slab;
87 
88 	/* as packets go on the queued queue, they are counted... */
89 	u32 counter;
90 	u32 sent_counter;
91 
92 	/* dma page table */
93 	struct rb_root dma_pages_root;
94 
95 	/* protect everything above... */
96 	struct mutex lock;
97 };
98 
99 struct qib_user_sdma_queue *
100 qib_user_sdma_queue_create(struct device *dev, int unit, int ctxt, int sctxt)
101 {
102 	struct qib_user_sdma_queue *pq =
103 		kmalloc(sizeof(struct qib_user_sdma_queue), GFP_KERNEL);
104 
105 	if (!pq)
106 		goto done;
107 
108 	pq->counter = 0;
109 	pq->sent_counter = 0;
110 	INIT_LIST_HEAD(&pq->sent);
111 
112 	mutex_init(&pq->lock);
113 
114 	snprintf(pq->pkt_slab_name, sizeof(pq->pkt_slab_name),
115 		 "qib-user-sdma-pkts-%u-%02u.%02u", unit, ctxt, sctxt);
116 	pq->pkt_slab = kmem_cache_create(pq->pkt_slab_name,
117 					 sizeof(struct qib_user_sdma_pkt),
118 					 0, 0, NULL);
119 
120 	if (!pq->pkt_slab)
121 		goto err_kfree;
122 
123 	snprintf(pq->header_cache_name, sizeof(pq->header_cache_name),
124 		 "qib-user-sdma-headers-%u-%02u.%02u", unit, ctxt, sctxt);
125 	pq->header_cache = dma_pool_create(pq->header_cache_name,
126 					   dev,
127 					   QIB_USER_SDMA_EXP_HEADER_LENGTH,
128 					   4, 0);
129 	if (!pq->header_cache)
130 		goto err_slab;
131 
132 	pq->dma_pages_root = RB_ROOT;
133 
134 	goto done;
135 
136 err_slab:
137 	kmem_cache_destroy(pq->pkt_slab);
138 err_kfree:
139 	kfree(pq);
140 	pq = NULL;
141 
142 done:
143 	return pq;
144 }
145 
146 static void qib_user_sdma_init_frag(struct qib_user_sdma_pkt *pkt,
147 				    int i, size_t offset, size_t len,
148 				    int put_page, int dma_mapped,
149 				    struct page *page,
150 				    void *kvaddr, dma_addr_t dma_addr)
151 {
152 	pkt->addr[i].offset = offset;
153 	pkt->addr[i].length = len;
154 	pkt->addr[i].put_page = put_page;
155 	pkt->addr[i].dma_mapped = dma_mapped;
156 	pkt->addr[i].page = page;
157 	pkt->addr[i].kvaddr = kvaddr;
158 	pkt->addr[i].addr = dma_addr;
159 }
160 
161 static void qib_user_sdma_init_header(struct qib_user_sdma_pkt *pkt,
162 				      u32 counter, size_t offset,
163 				      size_t len, int dma_mapped,
164 				      struct page *page,
165 				      void *kvaddr, dma_addr_t dma_addr)
166 {
167 	pkt->naddr = 1;
168 	pkt->counter = counter;
169 	qib_user_sdma_init_frag(pkt, 0, offset, len, 0, dma_mapped, page,
170 				kvaddr, dma_addr);
171 }
172 
173 /* we've too many pages in the iovec, coalesce to a single page */
174 static int qib_user_sdma_coalesce(const struct qib_devdata *dd,
175 				  struct qib_user_sdma_pkt *pkt,
176 				  const struct iovec *iov,
177 				  unsigned long niov)
178 {
179 	int ret = 0;
180 	struct page *page = alloc_page(GFP_KERNEL);
181 	void *mpage_save;
182 	char *mpage;
183 	int i;
184 	int len = 0;
185 	dma_addr_t dma_addr;
186 
187 	if (!page) {
188 		ret = -ENOMEM;
189 		goto done;
190 	}
191 
192 	mpage = kmap(page);
193 	mpage_save = mpage;
194 	for (i = 0; i < niov; i++) {
195 		int cfur;
196 
197 		cfur = copy_from_user(mpage,
198 				      iov[i].iov_base, iov[i].iov_len);
199 		if (cfur) {
200 			ret = -EFAULT;
201 			goto free_unmap;
202 		}
203 
204 		mpage += iov[i].iov_len;
205 		len += iov[i].iov_len;
206 	}
207 
208 	dma_addr = dma_map_page(&dd->pcidev->dev, page, 0, len,
209 				DMA_TO_DEVICE);
210 	if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
211 		ret = -ENOMEM;
212 		goto free_unmap;
213 	}
214 
215 	qib_user_sdma_init_frag(pkt, 1, 0, len, 0, 1, page, mpage_save,
216 				dma_addr);
217 	pkt->naddr = 2;
218 
219 	goto done;
220 
221 free_unmap:
222 	kunmap(page);
223 	__free_page(page);
224 done:
225 	return ret;
226 }
227 
228 /*
229  * How many pages in this iovec element?
230  */
231 static int qib_user_sdma_num_pages(const struct iovec *iov)
232 {
233 	const unsigned long addr  = (unsigned long) iov->iov_base;
234 	const unsigned long  len  = iov->iov_len;
235 	const unsigned long spage = addr & PAGE_MASK;
236 	const unsigned long epage = (addr + len - 1) & PAGE_MASK;
237 
238 	return 1 + ((epage - spage) >> PAGE_SHIFT);
239 }
240 
241 /*
242  * Truncate length to page boundry.
243  */
244 static int qib_user_sdma_page_length(unsigned long addr, unsigned long len)
245 {
246 	const unsigned long offset = addr & ~PAGE_MASK;
247 
248 	return ((offset + len) > PAGE_SIZE) ? (PAGE_SIZE - offset) : len;
249 }
250 
251 static void qib_user_sdma_free_pkt_frag(struct device *dev,
252 					struct qib_user_sdma_queue *pq,
253 					struct qib_user_sdma_pkt *pkt,
254 					int frag)
255 {
256 	const int i = frag;
257 
258 	if (pkt->addr[i].page) {
259 		if (pkt->addr[i].dma_mapped)
260 			dma_unmap_page(dev,
261 				       pkt->addr[i].addr,
262 				       pkt->addr[i].length,
263 				       DMA_TO_DEVICE);
264 
265 		if (pkt->addr[i].kvaddr)
266 			kunmap(pkt->addr[i].page);
267 
268 		if (pkt->addr[i].put_page)
269 			put_page(pkt->addr[i].page);
270 		else
271 			__free_page(pkt->addr[i].page);
272 	} else if (pkt->addr[i].kvaddr)
273 		/* free coherent mem from cache... */
274 		dma_pool_free(pq->header_cache,
275 			      pkt->addr[i].kvaddr, pkt->addr[i].addr);
276 }
277 
278 /* return number of pages pinned... */
279 static int qib_user_sdma_pin_pages(const struct qib_devdata *dd,
280 				   struct qib_user_sdma_pkt *pkt,
281 				   unsigned long addr, int tlen, int npages)
282 {
283 	struct page *pages[2];
284 	int j;
285 	int ret;
286 
287 	ret = get_user_pages(current, current->mm, addr,
288 			     npages, 0, 1, pages, NULL);
289 
290 	if (ret != npages) {
291 		int i;
292 
293 		for (i = 0; i < ret; i++)
294 			put_page(pages[i]);
295 
296 		ret = -ENOMEM;
297 		goto done;
298 	}
299 
300 	for (j = 0; j < npages; j++) {
301 		/* map the pages... */
302 		const int flen = qib_user_sdma_page_length(addr, tlen);
303 		dma_addr_t dma_addr =
304 			dma_map_page(&dd->pcidev->dev,
305 				     pages[j], 0, flen, DMA_TO_DEVICE);
306 		unsigned long fofs = addr & ~PAGE_MASK;
307 
308 		if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
309 			ret = -ENOMEM;
310 			goto done;
311 		}
312 
313 		qib_user_sdma_init_frag(pkt, pkt->naddr, fofs, flen, 1, 1,
314 					pages[j], kmap(pages[j]), dma_addr);
315 
316 		pkt->naddr++;
317 		addr += flen;
318 		tlen -= flen;
319 	}
320 
321 done:
322 	return ret;
323 }
324 
325 static int qib_user_sdma_pin_pkt(const struct qib_devdata *dd,
326 				 struct qib_user_sdma_queue *pq,
327 				 struct qib_user_sdma_pkt *pkt,
328 				 const struct iovec *iov,
329 				 unsigned long niov)
330 {
331 	int ret = 0;
332 	unsigned long idx;
333 
334 	for (idx = 0; idx < niov; idx++) {
335 		const int npages = qib_user_sdma_num_pages(iov + idx);
336 		const unsigned long addr = (unsigned long) iov[idx].iov_base;
337 
338 		ret = qib_user_sdma_pin_pages(dd, pkt, addr,
339 					      iov[idx].iov_len, npages);
340 		if (ret < 0)
341 			goto free_pkt;
342 	}
343 
344 	goto done;
345 
346 free_pkt:
347 	for (idx = 0; idx < pkt->naddr; idx++)
348 		qib_user_sdma_free_pkt_frag(&dd->pcidev->dev, pq, pkt, idx);
349 
350 done:
351 	return ret;
352 }
353 
354 static int qib_user_sdma_init_payload(const struct qib_devdata *dd,
355 				      struct qib_user_sdma_queue *pq,
356 				      struct qib_user_sdma_pkt *pkt,
357 				      const struct iovec *iov,
358 				      unsigned long niov, int npages)
359 {
360 	int ret = 0;
361 
362 	if (npages >= ARRAY_SIZE(pkt->addr))
363 		ret = qib_user_sdma_coalesce(dd, pkt, iov, niov);
364 	else
365 		ret = qib_user_sdma_pin_pkt(dd, pq, pkt, iov, niov);
366 
367 	return ret;
368 }
369 
370 /* free a packet list -- return counter value of last packet */
371 static void qib_user_sdma_free_pkt_list(struct device *dev,
372 					struct qib_user_sdma_queue *pq,
373 					struct list_head *list)
374 {
375 	struct qib_user_sdma_pkt *pkt, *pkt_next;
376 
377 	list_for_each_entry_safe(pkt, pkt_next, list, list) {
378 		int i;
379 
380 		for (i = 0; i < pkt->naddr; i++)
381 			qib_user_sdma_free_pkt_frag(dev, pq, pkt, i);
382 
383 		kmem_cache_free(pq->pkt_slab, pkt);
384 	}
385 }
386 
387 /*
388  * copy headers, coalesce etc -- pq->lock must be held
389  *
390  * we queue all the packets to list, returning the
391  * number of bytes total.  list must be empty initially,
392  * as, if there is an error we clean it...
393  */
394 static int qib_user_sdma_queue_pkts(const struct qib_devdata *dd,
395 				    struct qib_user_sdma_queue *pq,
396 				    struct list_head *list,
397 				    const struct iovec *iov,
398 				    unsigned long niov,
399 				    int maxpkts)
400 {
401 	unsigned long idx = 0;
402 	int ret = 0;
403 	int npkts = 0;
404 	struct page *page = NULL;
405 	__le32 *pbc;
406 	dma_addr_t dma_addr;
407 	struct qib_user_sdma_pkt *pkt = NULL;
408 	size_t len;
409 	size_t nw;
410 	u32 counter = pq->counter;
411 	int dma_mapped = 0;
412 
413 	while (idx < niov && npkts < maxpkts) {
414 		const unsigned long addr = (unsigned long) iov[idx].iov_base;
415 		const unsigned long idx_save = idx;
416 		unsigned pktnw;
417 		unsigned pktnwc;
418 		int nfrags = 0;
419 		int npages = 0;
420 		int cfur;
421 
422 		dma_mapped = 0;
423 		len = iov[idx].iov_len;
424 		nw = len >> 2;
425 		page = NULL;
426 
427 		pkt = kmem_cache_alloc(pq->pkt_slab, GFP_KERNEL);
428 		if (!pkt) {
429 			ret = -ENOMEM;
430 			goto free_list;
431 		}
432 
433 		if (len < QIB_USER_SDMA_MIN_HEADER_LENGTH ||
434 		    len > PAGE_SIZE || len & 3 || addr & 3) {
435 			ret = -EINVAL;
436 			goto free_pkt;
437 		}
438 
439 		if (len == QIB_USER_SDMA_EXP_HEADER_LENGTH)
440 			pbc = dma_pool_alloc(pq->header_cache, GFP_KERNEL,
441 					     &dma_addr);
442 		else
443 			pbc = NULL;
444 
445 		if (!pbc) {
446 			page = alloc_page(GFP_KERNEL);
447 			if (!page) {
448 				ret = -ENOMEM;
449 				goto free_pkt;
450 			}
451 			pbc = kmap(page);
452 		}
453 
454 		cfur = copy_from_user(pbc, iov[idx].iov_base, len);
455 		if (cfur) {
456 			ret = -EFAULT;
457 			goto free_pbc;
458 		}
459 
460 		/*
461 		 * This assignment is a bit strange.  it's because the
462 		 * the pbc counts the number of 32 bit words in the full
463 		 * packet _except_ the first word of the pbc itself...
464 		 */
465 		pktnwc = nw - 1;
466 
467 		/*
468 		 * pktnw computation yields the number of 32 bit words
469 		 * that the caller has indicated in the PBC.  note that
470 		 * this is one less than the total number of words that
471 		 * goes to the send DMA engine as the first 32 bit word
472 		 * of the PBC itself is not counted.  Armed with this count,
473 		 * we can verify that the packet is consistent with the
474 		 * iovec lengths.
475 		 */
476 		pktnw = le32_to_cpu(*pbc) & QIB_PBC_LENGTH_MASK;
477 		if (pktnw < pktnwc || pktnw > pktnwc + (PAGE_SIZE >> 2)) {
478 			ret = -EINVAL;
479 			goto free_pbc;
480 		}
481 
482 		idx++;
483 		while (pktnwc < pktnw && idx < niov) {
484 			const size_t slen = iov[idx].iov_len;
485 			const unsigned long faddr =
486 				(unsigned long) iov[idx].iov_base;
487 
488 			if (slen & 3 || faddr & 3 || !slen ||
489 			    slen > PAGE_SIZE) {
490 				ret = -EINVAL;
491 				goto free_pbc;
492 			}
493 
494 			npages++;
495 			if ((faddr & PAGE_MASK) !=
496 			    ((faddr + slen - 1) & PAGE_MASK))
497 				npages++;
498 
499 			pktnwc += slen >> 2;
500 			idx++;
501 			nfrags++;
502 		}
503 
504 		if (pktnwc != pktnw) {
505 			ret = -EINVAL;
506 			goto free_pbc;
507 		}
508 
509 		if (page) {
510 			dma_addr = dma_map_page(&dd->pcidev->dev,
511 						page, 0, len, DMA_TO_DEVICE);
512 			if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
513 				ret = -ENOMEM;
514 				goto free_pbc;
515 			}
516 
517 			dma_mapped = 1;
518 		}
519 
520 		qib_user_sdma_init_header(pkt, counter, 0, len, dma_mapped,
521 					  page, pbc, dma_addr);
522 
523 		if (nfrags) {
524 			ret = qib_user_sdma_init_payload(dd, pq, pkt,
525 							 iov + idx_save + 1,
526 							 nfrags, npages);
527 			if (ret < 0)
528 				goto free_pbc_dma;
529 		}
530 
531 		counter++;
532 		npkts++;
533 
534 		list_add_tail(&pkt->list, list);
535 	}
536 
537 	ret = idx;
538 	goto done;
539 
540 free_pbc_dma:
541 	if (dma_mapped)
542 		dma_unmap_page(&dd->pcidev->dev, dma_addr, len, DMA_TO_DEVICE);
543 free_pbc:
544 	if (page) {
545 		kunmap(page);
546 		__free_page(page);
547 	} else
548 		dma_pool_free(pq->header_cache, pbc, dma_addr);
549 free_pkt:
550 	kmem_cache_free(pq->pkt_slab, pkt);
551 free_list:
552 	qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, list);
553 done:
554 	return ret;
555 }
556 
557 static void qib_user_sdma_set_complete_counter(struct qib_user_sdma_queue *pq,
558 					       u32 c)
559 {
560 	pq->sent_counter = c;
561 }
562 
563 /* try to clean out queue -- needs pq->lock */
564 static int qib_user_sdma_queue_clean(struct qib_pportdata *ppd,
565 				     struct qib_user_sdma_queue *pq)
566 {
567 	struct qib_devdata *dd = ppd->dd;
568 	struct list_head free_list;
569 	struct qib_user_sdma_pkt *pkt;
570 	struct qib_user_sdma_pkt *pkt_prev;
571 	int ret = 0;
572 
573 	INIT_LIST_HEAD(&free_list);
574 
575 	list_for_each_entry_safe(pkt, pkt_prev, &pq->sent, list) {
576 		s64 descd = ppd->sdma_descq_removed - pkt->added;
577 
578 		if (descd < 0)
579 			break;
580 
581 		list_move_tail(&pkt->list, &free_list);
582 
583 		/* one more packet cleaned */
584 		ret++;
585 	}
586 
587 	if (!list_empty(&free_list)) {
588 		u32 counter;
589 
590 		pkt = list_entry(free_list.prev,
591 				 struct qib_user_sdma_pkt, list);
592 		counter = pkt->counter;
593 
594 		qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
595 		qib_user_sdma_set_complete_counter(pq, counter);
596 	}
597 
598 	return ret;
599 }
600 
601 void qib_user_sdma_queue_destroy(struct qib_user_sdma_queue *pq)
602 {
603 	if (!pq)
604 		return;
605 
606 	kmem_cache_destroy(pq->pkt_slab);
607 	dma_pool_destroy(pq->header_cache);
608 	kfree(pq);
609 }
610 
611 /* clean descriptor queue, returns > 0 if some elements cleaned */
612 static int qib_user_sdma_hwqueue_clean(struct qib_pportdata *ppd)
613 {
614 	int ret;
615 	unsigned long flags;
616 
617 	spin_lock_irqsave(&ppd->sdma_lock, flags);
618 	ret = qib_sdma_make_progress(ppd);
619 	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
620 
621 	return ret;
622 }
623 
624 /* we're in close, drain packets so that we can cleanup successfully... */
625 void qib_user_sdma_queue_drain(struct qib_pportdata *ppd,
626 			       struct qib_user_sdma_queue *pq)
627 {
628 	struct qib_devdata *dd = ppd->dd;
629 	int i;
630 
631 	if (!pq)
632 		return;
633 
634 	for (i = 0; i < QIB_USER_SDMA_DRAIN_TIMEOUT; i++) {
635 		mutex_lock(&pq->lock);
636 		if (list_empty(&pq->sent)) {
637 			mutex_unlock(&pq->lock);
638 			break;
639 		}
640 		qib_user_sdma_hwqueue_clean(ppd);
641 		qib_user_sdma_queue_clean(ppd, pq);
642 		mutex_unlock(&pq->lock);
643 		msleep(10);
644 	}
645 
646 	if (!list_empty(&pq->sent)) {
647 		struct list_head free_list;
648 
649 		qib_dev_err(dd, "user sdma lists not empty: forcing!\n");
650 		INIT_LIST_HEAD(&free_list);
651 		mutex_lock(&pq->lock);
652 		list_splice_init(&pq->sent, &free_list);
653 		qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
654 		mutex_unlock(&pq->lock);
655 	}
656 }
657 
658 static inline __le64 qib_sdma_make_desc0(struct qib_pportdata *ppd,
659 					 u64 addr, u64 dwlen, u64 dwoffset)
660 {
661 	u8 tmpgen;
662 
663 	tmpgen = ppd->sdma_generation;
664 
665 	return cpu_to_le64(/* SDmaPhyAddr[31:0] */
666 			   ((addr & 0xfffffffcULL) << 32) |
667 			   /* SDmaGeneration[1:0] */
668 			   ((tmpgen & 3ULL) << 30) |
669 			   /* SDmaDwordCount[10:0] */
670 			   ((dwlen & 0x7ffULL) << 16) |
671 			   /* SDmaBufOffset[12:2] */
672 			   (dwoffset & 0x7ffULL));
673 }
674 
675 static inline __le64 qib_sdma_make_first_desc0(__le64 descq)
676 {
677 	return descq | cpu_to_le64(1ULL << 12);
678 }
679 
680 static inline __le64 qib_sdma_make_last_desc0(__le64 descq)
681 {
682 					      /* last */  /* dma head */
683 	return descq | cpu_to_le64(1ULL << 11 | 1ULL << 13);
684 }
685 
686 static inline __le64 qib_sdma_make_desc1(u64 addr)
687 {
688 	/* SDmaPhyAddr[47:32] */
689 	return cpu_to_le64(addr >> 32);
690 }
691 
692 static void qib_user_sdma_send_frag(struct qib_pportdata *ppd,
693 				    struct qib_user_sdma_pkt *pkt, int idx,
694 				    unsigned ofs, u16 tail)
695 {
696 	const u64 addr = (u64) pkt->addr[idx].addr +
697 		(u64) pkt->addr[idx].offset;
698 	const u64 dwlen = (u64) pkt->addr[idx].length / 4;
699 	__le64 *descqp;
700 	__le64 descq0;
701 
702 	descqp = &ppd->sdma_descq[tail].qw[0];
703 
704 	descq0 = qib_sdma_make_desc0(ppd, addr, dwlen, ofs);
705 	if (idx == 0)
706 		descq0 = qib_sdma_make_first_desc0(descq0);
707 	if (idx == pkt->naddr - 1)
708 		descq0 = qib_sdma_make_last_desc0(descq0);
709 
710 	descqp[0] = descq0;
711 	descqp[1] = qib_sdma_make_desc1(addr);
712 }
713 
714 /* pq->lock must be held, get packets on the wire... */
715 static int qib_user_sdma_push_pkts(struct qib_pportdata *ppd,
716 				   struct qib_user_sdma_queue *pq,
717 				   struct list_head *pktlist)
718 {
719 	struct qib_devdata *dd = ppd->dd;
720 	int ret = 0;
721 	unsigned long flags;
722 	u16 tail;
723 	u8 generation;
724 	u64 descq_added;
725 
726 	if (list_empty(pktlist))
727 		return 0;
728 
729 	if (unlikely(!(ppd->lflags & QIBL_LINKACTIVE)))
730 		return -ECOMM;
731 
732 	spin_lock_irqsave(&ppd->sdma_lock, flags);
733 
734 	/* keep a copy for restoring purposes in case of problems */
735 	generation = ppd->sdma_generation;
736 	descq_added = ppd->sdma_descq_added;
737 
738 	if (unlikely(!__qib_sdma_running(ppd))) {
739 		ret = -ECOMM;
740 		goto unlock;
741 	}
742 
743 	tail = ppd->sdma_descq_tail;
744 	while (!list_empty(pktlist)) {
745 		struct qib_user_sdma_pkt *pkt =
746 			list_entry(pktlist->next, struct qib_user_sdma_pkt,
747 				   list);
748 		int i;
749 		unsigned ofs = 0;
750 		u16 dtail = tail;
751 
752 		if (pkt->naddr > qib_sdma_descq_freecnt(ppd))
753 			goto unlock_check_tail;
754 
755 		for (i = 0; i < pkt->naddr; i++) {
756 			qib_user_sdma_send_frag(ppd, pkt, i, ofs, tail);
757 			ofs += pkt->addr[i].length >> 2;
758 
759 			if (++tail == ppd->sdma_descq_cnt) {
760 				tail = 0;
761 				++ppd->sdma_generation;
762 			}
763 		}
764 
765 		if ((ofs << 2) > ppd->ibmaxlen) {
766 			ret = -EMSGSIZE;
767 			goto unlock;
768 		}
769 
770 		/*
771 		 * If the packet is >= 2KB mtu equivalent, we have to use
772 		 * the large buffers, and have to mark each descriptor as
773 		 * part of a large buffer packet.
774 		 */
775 		if (ofs > dd->piosize2kmax_dwords) {
776 			for (i = 0; i < pkt->naddr; i++) {
777 				ppd->sdma_descq[dtail].qw[0] |=
778 					cpu_to_le64(1ULL << 14);
779 				if (++dtail == ppd->sdma_descq_cnt)
780 					dtail = 0;
781 			}
782 		}
783 
784 		ppd->sdma_descq_added += pkt->naddr;
785 		pkt->added = ppd->sdma_descq_added;
786 		list_move_tail(&pkt->list, &pq->sent);
787 		ret++;
788 	}
789 
790 unlock_check_tail:
791 	/* advance the tail on the chip if necessary */
792 	if (ppd->sdma_descq_tail != tail)
793 		dd->f_sdma_update_tail(ppd, tail);
794 
795 unlock:
796 	if (unlikely(ret < 0)) {
797 		ppd->sdma_generation = generation;
798 		ppd->sdma_descq_added = descq_added;
799 	}
800 	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
801 
802 	return ret;
803 }
804 
805 int qib_user_sdma_writev(struct qib_ctxtdata *rcd,
806 			 struct qib_user_sdma_queue *pq,
807 			 const struct iovec *iov,
808 			 unsigned long dim)
809 {
810 	struct qib_devdata *dd = rcd->dd;
811 	struct qib_pportdata *ppd = rcd->ppd;
812 	int ret = 0;
813 	struct list_head list;
814 	int npkts = 0;
815 
816 	INIT_LIST_HEAD(&list);
817 
818 	mutex_lock(&pq->lock);
819 
820 	/* why not -ECOMM like qib_user_sdma_push_pkts() below? */
821 	if (!qib_sdma_running(ppd))
822 		goto done_unlock;
823 
824 	if (ppd->sdma_descq_added != ppd->sdma_descq_removed) {
825 		qib_user_sdma_hwqueue_clean(ppd);
826 		qib_user_sdma_queue_clean(ppd, pq);
827 	}
828 
829 	while (dim) {
830 		const int mxp = 8;
831 
832 		down_write(&current->mm->mmap_sem);
833 		ret = qib_user_sdma_queue_pkts(dd, pq, &list, iov, dim, mxp);
834 		up_write(&current->mm->mmap_sem);
835 
836 		if (ret <= 0)
837 			goto done_unlock;
838 		else {
839 			dim -= ret;
840 			iov += ret;
841 		}
842 
843 		/* force packets onto the sdma hw queue... */
844 		if (!list_empty(&list)) {
845 			/*
846 			 * Lazily clean hw queue.  the 4 is a guess of about
847 			 * how many sdma descriptors a packet will take (it
848 			 * doesn't have to be perfect).
849 			 */
850 			if (qib_sdma_descq_freecnt(ppd) < ret * 4) {
851 				qib_user_sdma_hwqueue_clean(ppd);
852 				qib_user_sdma_queue_clean(ppd, pq);
853 			}
854 
855 			ret = qib_user_sdma_push_pkts(ppd, pq, &list);
856 			if (ret < 0)
857 				goto done_unlock;
858 			else {
859 				npkts += ret;
860 				pq->counter += ret;
861 
862 				if (!list_empty(&list))
863 					goto done_unlock;
864 			}
865 		}
866 	}
867 
868 done_unlock:
869 	if (!list_empty(&list))
870 		qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &list);
871 	mutex_unlock(&pq->lock);
872 
873 	return (ret < 0) ? ret : npkts;
874 }
875 
876 int qib_user_sdma_make_progress(struct qib_pportdata *ppd,
877 				struct qib_user_sdma_queue *pq)
878 {
879 	int ret = 0;
880 
881 	mutex_lock(&pq->lock);
882 	qib_user_sdma_hwqueue_clean(ppd);
883 	ret = qib_user_sdma_queue_clean(ppd, pq);
884 	mutex_unlock(&pq->lock);
885 
886 	return ret;
887 }
888 
889 u32 qib_user_sdma_complete_counter(const struct qib_user_sdma_queue *pq)
890 {
891 	return pq ? pq->sent_counter : 0;
892 }
893 
894 u32 qib_user_sdma_inflight_counter(struct qib_user_sdma_queue *pq)
895 {
896 	return pq ? pq->counter : 0;
897 }
898