1 /*
2 * Copyright (c) 2007, 2008 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/sched.h>
37 #include <linux/slab.h>
38 #include <linux/list.h>
39 #include <linux/highmem.h>
40 #include <linux/io.h>
41 #include <linux/uio.h>
42 #include <linux/rbtree.h>
43 #include <linux/spinlock.h>
44 #include <linux/delay.h>
45
46 #include "ipath_kernel.h"
47 #include "ipath_user_sdma.h"
48
49 /* minimum size of header */
50 #define IPATH_USER_SDMA_MIN_HEADER_LENGTH 64
51 /* expected size of headers (for dma_pool) */
52 #define IPATH_USER_SDMA_EXP_HEADER_LENGTH 64
53 /* length mask in PBC (lower 11 bits) */
54 #define IPATH_PBC_LENGTH_MASK ((1 << 11) - 1)
55
56 struct ipath_user_sdma_pkt {
57 u8 naddr; /* dimension of addr (1..3) ... */
58 u32 counter; /* sdma pkts queued counter for this entry */
59 u64 added; /* global descq number of entries */
60
61 struct {
62 u32 offset; /* offset for kvaddr, addr */
63 u32 length; /* length in page */
64 u8 put_page; /* should we put_page? */
65 u8 dma_mapped; /* is page dma_mapped? */
66 struct page *page; /* may be NULL (coherent mem) */
67 void *kvaddr; /* FIXME: only for pio hack */
68 dma_addr_t addr;
69 } addr[4]; /* max pages, any more and we coalesce */
70 struct list_head list; /* list element */
71 };
72
73 struct ipath_user_sdma_queue {
74 /*
75 * pkts sent to dma engine are queued on this
76 * list head. the type of the elements of this
77 * list are struct ipath_user_sdma_pkt...
78 */
79 struct list_head sent;
80
81 /* headers with expected length are allocated from here... */
82 char header_cache_name[64];
83 struct dma_pool *header_cache;
84
85 /* packets are allocated from the slab cache... */
86 char pkt_slab_name[64];
87 struct kmem_cache *pkt_slab;
88
89 /* as packets go on the queued queue, they are counted... */
90 u32 counter;
91 u32 sent_counter;
92
93 /* dma page table */
94 struct rb_root dma_pages_root;
95
96 /* protect everything above... */
97 struct mutex lock;
98 };
99
100 struct ipath_user_sdma_queue *
ipath_user_sdma_queue_create(struct device * dev,int unit,int port,int sport)101 ipath_user_sdma_queue_create(struct device *dev, int unit, int port, int sport)
102 {
103 struct ipath_user_sdma_queue *pq =
104 kmalloc(sizeof(struct ipath_user_sdma_queue), GFP_KERNEL);
105
106 if (!pq)
107 goto done;
108
109 pq->counter = 0;
110 pq->sent_counter = 0;
111 INIT_LIST_HEAD(&pq->sent);
112
113 mutex_init(&pq->lock);
114
115 snprintf(pq->pkt_slab_name, sizeof(pq->pkt_slab_name),
116 "ipath-user-sdma-pkts-%u-%02u.%02u", unit, port, sport);
117 pq->pkt_slab = kmem_cache_create(pq->pkt_slab_name,
118 sizeof(struct ipath_user_sdma_pkt),
119 0, 0, NULL);
120
121 if (!pq->pkt_slab)
122 goto err_kfree;
123
124 snprintf(pq->header_cache_name, sizeof(pq->header_cache_name),
125 "ipath-user-sdma-headers-%u-%02u.%02u", unit, port, sport);
126 pq->header_cache = dma_pool_create(pq->header_cache_name,
127 dev,
128 IPATH_USER_SDMA_EXP_HEADER_LENGTH,
129 4, 0);
130 if (!pq->header_cache)
131 goto err_slab;
132
133 pq->dma_pages_root = RB_ROOT;
134
135 goto done;
136
137 err_slab:
138 kmem_cache_destroy(pq->pkt_slab);
139 err_kfree:
140 kfree(pq);
141 pq = NULL;
142
143 done:
144 return pq;
145 }
146
ipath_user_sdma_init_frag(struct ipath_user_sdma_pkt * pkt,int i,size_t offset,size_t len,int put_page,int dma_mapped,struct page * page,void * kvaddr,dma_addr_t dma_addr)147 static void ipath_user_sdma_init_frag(struct ipath_user_sdma_pkt *pkt,
148 int i, size_t offset, size_t len,
149 int put_page, int dma_mapped,
150 struct page *page,
151 void *kvaddr, dma_addr_t dma_addr)
152 {
153 pkt->addr[i].offset = offset;
154 pkt->addr[i].length = len;
155 pkt->addr[i].put_page = put_page;
156 pkt->addr[i].dma_mapped = dma_mapped;
157 pkt->addr[i].page = page;
158 pkt->addr[i].kvaddr = kvaddr;
159 pkt->addr[i].addr = dma_addr;
160 }
161
ipath_user_sdma_init_header(struct ipath_user_sdma_pkt * pkt,u32 counter,size_t offset,size_t len,int dma_mapped,struct page * page,void * kvaddr,dma_addr_t dma_addr)162 static void ipath_user_sdma_init_header(struct ipath_user_sdma_pkt *pkt,
163 u32 counter, size_t offset,
164 size_t len, int dma_mapped,
165 struct page *page,
166 void *kvaddr, dma_addr_t dma_addr)
167 {
168 pkt->naddr = 1;
169 pkt->counter = counter;
170 ipath_user_sdma_init_frag(pkt, 0, offset, len, 0, dma_mapped, page,
171 kvaddr, dma_addr);
172 }
173
174 /* we've too many pages in the iovec, coalesce to a single page */
ipath_user_sdma_coalesce(const struct ipath_devdata * dd,struct ipath_user_sdma_pkt * pkt,const struct iovec * iov,unsigned long niov)175 static int ipath_user_sdma_coalesce(const struct ipath_devdata *dd,
176 struct ipath_user_sdma_pkt *pkt,
177 const struct iovec *iov,
178 unsigned long niov) {
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 ipath_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 /* how many pages in this iovec element? */
ipath_user_sdma_num_pages(const struct iovec * iov)229 static int ipath_user_sdma_num_pages(const struct iovec *iov)
230 {
231 const unsigned long addr = (unsigned long) iov->iov_base;
232 const unsigned long len = iov->iov_len;
233 const unsigned long spage = addr & PAGE_MASK;
234 const unsigned long epage = (addr + len - 1) & PAGE_MASK;
235
236 return 1 + ((epage - spage) >> PAGE_SHIFT);
237 }
238
239 /* truncate length to page boundary */
ipath_user_sdma_page_length(unsigned long addr,unsigned long len)240 static int ipath_user_sdma_page_length(unsigned long addr, unsigned long len)
241 {
242 const unsigned long offset = addr & ~PAGE_MASK;
243
244 return ((offset + len) > PAGE_SIZE) ? (PAGE_SIZE - offset) : len;
245 }
246
ipath_user_sdma_free_pkt_frag(struct device * dev,struct ipath_user_sdma_queue * pq,struct ipath_user_sdma_pkt * pkt,int frag)247 static void ipath_user_sdma_free_pkt_frag(struct device *dev,
248 struct ipath_user_sdma_queue *pq,
249 struct ipath_user_sdma_pkt *pkt,
250 int frag)
251 {
252 const int i = frag;
253
254 if (pkt->addr[i].page) {
255 if (pkt->addr[i].dma_mapped)
256 dma_unmap_page(dev,
257 pkt->addr[i].addr,
258 pkt->addr[i].length,
259 DMA_TO_DEVICE);
260
261 if (pkt->addr[i].kvaddr)
262 kunmap(pkt->addr[i].page);
263
264 if (pkt->addr[i].put_page)
265 put_page(pkt->addr[i].page);
266 else
267 __free_page(pkt->addr[i].page);
268 } else if (pkt->addr[i].kvaddr)
269 /* free coherent mem from cache... */
270 dma_pool_free(pq->header_cache,
271 pkt->addr[i].kvaddr, pkt->addr[i].addr);
272 }
273
274 /* return number of pages pinned... */
ipath_user_sdma_pin_pages(const struct ipath_devdata * dd,struct ipath_user_sdma_pkt * pkt,unsigned long addr,int tlen,int npages)275 static int ipath_user_sdma_pin_pages(const struct ipath_devdata *dd,
276 struct ipath_user_sdma_pkt *pkt,
277 unsigned long addr, int tlen, int npages)
278 {
279 struct page *pages[2];
280 int j;
281 int ret;
282
283 ret = get_user_pages_fast(addr, npages, 0, pages);
284 if (ret != npages) {
285 int i;
286
287 for (i = 0; i < ret; i++)
288 put_page(pages[i]);
289
290 ret = -ENOMEM;
291 goto done;
292 }
293
294 for (j = 0; j < npages; j++) {
295 /* map the pages... */
296 const int flen =
297 ipath_user_sdma_page_length(addr, tlen);
298 dma_addr_t dma_addr =
299 dma_map_page(&dd->pcidev->dev,
300 pages[j], 0, flen, DMA_TO_DEVICE);
301 unsigned long fofs = addr & ~PAGE_MASK;
302
303 if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
304 ret = -ENOMEM;
305 goto done;
306 }
307
308 ipath_user_sdma_init_frag(pkt, pkt->naddr, fofs, flen, 1, 1,
309 pages[j], kmap(pages[j]),
310 dma_addr);
311
312 pkt->naddr++;
313 addr += flen;
314 tlen -= flen;
315 }
316
317 done:
318 return ret;
319 }
320
ipath_user_sdma_pin_pkt(const struct ipath_devdata * dd,struct ipath_user_sdma_queue * pq,struct ipath_user_sdma_pkt * pkt,const struct iovec * iov,unsigned long niov)321 static int ipath_user_sdma_pin_pkt(const struct ipath_devdata *dd,
322 struct ipath_user_sdma_queue *pq,
323 struct ipath_user_sdma_pkt *pkt,
324 const struct iovec *iov,
325 unsigned long niov)
326 {
327 int ret = 0;
328 unsigned long idx;
329
330 for (idx = 0; idx < niov; idx++) {
331 const int npages = ipath_user_sdma_num_pages(iov + idx);
332 const unsigned long addr = (unsigned long) iov[idx].iov_base;
333
334 ret = ipath_user_sdma_pin_pages(dd, pkt,
335 addr, iov[idx].iov_len,
336 npages);
337 if (ret < 0)
338 goto free_pkt;
339 }
340
341 goto done;
342
343 free_pkt:
344 for (idx = 0; idx < pkt->naddr; idx++)
345 ipath_user_sdma_free_pkt_frag(&dd->pcidev->dev, pq, pkt, idx);
346
347 done:
348 return ret;
349 }
350
ipath_user_sdma_init_payload(const struct ipath_devdata * dd,struct ipath_user_sdma_queue * pq,struct ipath_user_sdma_pkt * pkt,const struct iovec * iov,unsigned long niov,int npages)351 static int ipath_user_sdma_init_payload(const struct ipath_devdata *dd,
352 struct ipath_user_sdma_queue *pq,
353 struct ipath_user_sdma_pkt *pkt,
354 const struct iovec *iov,
355 unsigned long niov, int npages)
356 {
357 int ret = 0;
358
359 if (npages >= ARRAY_SIZE(pkt->addr))
360 ret = ipath_user_sdma_coalesce(dd, pkt, iov, niov);
361 else
362 ret = ipath_user_sdma_pin_pkt(dd, pq, pkt, iov, niov);
363
364 return ret;
365 }
366
367 /* free a packet list -- return counter value of last packet */
ipath_user_sdma_free_pkt_list(struct device * dev,struct ipath_user_sdma_queue * pq,struct list_head * list)368 static void ipath_user_sdma_free_pkt_list(struct device *dev,
369 struct ipath_user_sdma_queue *pq,
370 struct list_head *list)
371 {
372 struct ipath_user_sdma_pkt *pkt, *pkt_next;
373
374 list_for_each_entry_safe(pkt, pkt_next, list, list) {
375 int i;
376
377 for (i = 0; i < pkt->naddr; i++)
378 ipath_user_sdma_free_pkt_frag(dev, pq, pkt, i);
379
380 kmem_cache_free(pq->pkt_slab, pkt);
381 }
382 }
383
384 /*
385 * copy headers, coalesce etc -- pq->lock must be held
386 *
387 * we queue all the packets to list, returning the
388 * number of bytes total. list must be empty initially,
389 * as, if there is an error we clean it...
390 */
ipath_user_sdma_queue_pkts(const struct ipath_devdata * dd,struct ipath_user_sdma_queue * pq,struct list_head * list,const struct iovec * iov,unsigned long niov,int maxpkts)391 static int ipath_user_sdma_queue_pkts(const struct ipath_devdata *dd,
392 struct ipath_user_sdma_queue *pq,
393 struct list_head *list,
394 const struct iovec *iov,
395 unsigned long niov,
396 int maxpkts)
397 {
398 unsigned long idx = 0;
399 int ret = 0;
400 int npkts = 0;
401 struct page *page = NULL;
402 __le32 *pbc;
403 dma_addr_t dma_addr;
404 struct ipath_user_sdma_pkt *pkt = NULL;
405 size_t len;
406 size_t nw;
407 u32 counter = pq->counter;
408 int dma_mapped = 0;
409
410 while (idx < niov && npkts < maxpkts) {
411 const unsigned long addr = (unsigned long) iov[idx].iov_base;
412 const unsigned long idx_save = idx;
413 unsigned pktnw;
414 unsigned pktnwc;
415 int nfrags = 0;
416 int npages = 0;
417 int cfur;
418
419 dma_mapped = 0;
420 len = iov[idx].iov_len;
421 nw = len >> 2;
422 page = NULL;
423
424 pkt = kmem_cache_alloc(pq->pkt_slab, GFP_KERNEL);
425 if (!pkt) {
426 ret = -ENOMEM;
427 goto free_list;
428 }
429
430 if (len < IPATH_USER_SDMA_MIN_HEADER_LENGTH ||
431 len > PAGE_SIZE || len & 3 || addr & 3) {
432 ret = -EINVAL;
433 goto free_pkt;
434 }
435
436 if (len == IPATH_USER_SDMA_EXP_HEADER_LENGTH)
437 pbc = dma_pool_alloc(pq->header_cache, GFP_KERNEL,
438 &dma_addr);
439 else
440 pbc = NULL;
441
442 if (!pbc) {
443 page = alloc_page(GFP_KERNEL);
444 if (!page) {
445 ret = -ENOMEM;
446 goto free_pkt;
447 }
448 pbc = kmap(page);
449 }
450
451 cfur = copy_from_user(pbc, iov[idx].iov_base, len);
452 if (cfur) {
453 ret = -EFAULT;
454 goto free_pbc;
455 }
456
457 /*
458 * this assignment is a bit strange. it's because the
459 * the pbc counts the number of 32 bit words in the full
460 * packet _except_ the first word of the pbc itself...
461 */
462 pktnwc = nw - 1;
463
464 /*
465 * pktnw computation yields the number of 32 bit words
466 * that the caller has indicated in the PBC. note that
467 * this is one less than the total number of words that
468 * goes to the send DMA engine as the first 32 bit word
469 * of the PBC itself is not counted. Armed with this count,
470 * we can verify that the packet is consistent with the
471 * iovec lengths.
472 */
473 pktnw = le32_to_cpu(*pbc) & IPATH_PBC_LENGTH_MASK;
474 if (pktnw < pktnwc || pktnw > pktnwc + (PAGE_SIZE >> 2)) {
475 ret = -EINVAL;
476 goto free_pbc;
477 }
478
479
480 idx++;
481 while (pktnwc < pktnw && idx < niov) {
482 const size_t slen = iov[idx].iov_len;
483 const unsigned long faddr =
484 (unsigned long) iov[idx].iov_base;
485
486 if (slen & 3 || faddr & 3 || !slen ||
487 slen > PAGE_SIZE) {
488 ret = -EINVAL;
489 goto free_pbc;
490 }
491
492 npages++;
493 if ((faddr & PAGE_MASK) !=
494 ((faddr + slen - 1) & PAGE_MASK))
495 npages++;
496
497 pktnwc += slen >> 2;
498 idx++;
499 nfrags++;
500 }
501
502 if (pktnwc != pktnw) {
503 ret = -EINVAL;
504 goto free_pbc;
505 }
506
507 if (page) {
508 dma_addr = dma_map_page(&dd->pcidev->dev,
509 page, 0, len, DMA_TO_DEVICE);
510 if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
511 ret = -ENOMEM;
512 goto free_pbc;
513 }
514
515 dma_mapped = 1;
516 }
517
518 ipath_user_sdma_init_header(pkt, counter, 0, len, dma_mapped,
519 page, pbc, dma_addr);
520
521 if (nfrags) {
522 ret = ipath_user_sdma_init_payload(dd, pq, pkt,
523 iov + idx_save + 1,
524 nfrags, npages);
525 if (ret < 0)
526 goto free_pbc_dma;
527 }
528
529 counter++;
530 npkts++;
531
532 list_add_tail(&pkt->list, list);
533 }
534
535 ret = idx;
536 goto done;
537
538 free_pbc_dma:
539 if (dma_mapped)
540 dma_unmap_page(&dd->pcidev->dev, dma_addr, len, DMA_TO_DEVICE);
541 free_pbc:
542 if (page) {
543 kunmap(page);
544 __free_page(page);
545 } else
546 dma_pool_free(pq->header_cache, pbc, dma_addr);
547 free_pkt:
548 kmem_cache_free(pq->pkt_slab, pkt);
549 free_list:
550 ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, list);
551 done:
552 return ret;
553 }
554
ipath_user_sdma_set_complete_counter(struct ipath_user_sdma_queue * pq,u32 c)555 static void ipath_user_sdma_set_complete_counter(struct ipath_user_sdma_queue *pq,
556 u32 c)
557 {
558 pq->sent_counter = c;
559 }
560
561 /* try to clean out queue -- needs pq->lock */
ipath_user_sdma_queue_clean(const struct ipath_devdata * dd,struct ipath_user_sdma_queue * pq)562 static int ipath_user_sdma_queue_clean(const struct ipath_devdata *dd,
563 struct ipath_user_sdma_queue *pq)
564 {
565 struct list_head free_list;
566 struct ipath_user_sdma_pkt *pkt;
567 struct ipath_user_sdma_pkt *pkt_prev;
568 int ret = 0;
569
570 INIT_LIST_HEAD(&free_list);
571
572 list_for_each_entry_safe(pkt, pkt_prev, &pq->sent, list) {
573 s64 descd = dd->ipath_sdma_descq_removed - pkt->added;
574
575 if (descd < 0)
576 break;
577
578 list_move_tail(&pkt->list, &free_list);
579
580 /* one more packet cleaned */
581 ret++;
582 }
583
584 if (!list_empty(&free_list)) {
585 u32 counter;
586
587 pkt = list_entry(free_list.prev,
588 struct ipath_user_sdma_pkt, list);
589 counter = pkt->counter;
590
591 ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
592 ipath_user_sdma_set_complete_counter(pq, counter);
593 }
594
595 return ret;
596 }
597
ipath_user_sdma_queue_destroy(struct ipath_user_sdma_queue * pq)598 void ipath_user_sdma_queue_destroy(struct ipath_user_sdma_queue *pq)
599 {
600 if (!pq)
601 return;
602
603 kmem_cache_destroy(pq->pkt_slab);
604 dma_pool_destroy(pq->header_cache);
605 kfree(pq);
606 }
607
608 /* clean descriptor queue, returns > 0 if some elements cleaned */
ipath_user_sdma_hwqueue_clean(struct ipath_devdata * dd)609 static int ipath_user_sdma_hwqueue_clean(struct ipath_devdata *dd)
610 {
611 int ret;
612 unsigned long flags;
613
614 spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
615 ret = ipath_sdma_make_progress(dd);
616 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
617
618 return ret;
619 }
620
621 /* we're in close, drain packets so that we can cleanup successfully... */
ipath_user_sdma_queue_drain(struct ipath_devdata * dd,struct ipath_user_sdma_queue * pq)622 void ipath_user_sdma_queue_drain(struct ipath_devdata *dd,
623 struct ipath_user_sdma_queue *pq)
624 {
625 int i;
626
627 if (!pq)
628 return;
629
630 for (i = 0; i < 100; i++) {
631 mutex_lock(&pq->lock);
632 if (list_empty(&pq->sent)) {
633 mutex_unlock(&pq->lock);
634 break;
635 }
636 ipath_user_sdma_hwqueue_clean(dd);
637 ipath_user_sdma_queue_clean(dd, pq);
638 mutex_unlock(&pq->lock);
639 msleep(10);
640 }
641
642 if (!list_empty(&pq->sent)) {
643 struct list_head free_list;
644
645 printk(KERN_INFO "drain: lists not empty: forcing!\n");
646 INIT_LIST_HEAD(&free_list);
647 mutex_lock(&pq->lock);
648 list_splice_init(&pq->sent, &free_list);
649 ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
650 mutex_unlock(&pq->lock);
651 }
652 }
653
ipath_sdma_make_desc0(struct ipath_devdata * dd,u64 addr,u64 dwlen,u64 dwoffset)654 static inline __le64 ipath_sdma_make_desc0(struct ipath_devdata *dd,
655 u64 addr, u64 dwlen, u64 dwoffset)
656 {
657 return cpu_to_le64(/* SDmaPhyAddr[31:0] */
658 ((addr & 0xfffffffcULL) << 32) |
659 /* SDmaGeneration[1:0] */
660 ((dd->ipath_sdma_generation & 3ULL) << 30) |
661 /* SDmaDwordCount[10:0] */
662 ((dwlen & 0x7ffULL) << 16) |
663 /* SDmaBufOffset[12:2] */
664 (dwoffset & 0x7ffULL));
665 }
666
ipath_sdma_make_first_desc0(__le64 descq)667 static inline __le64 ipath_sdma_make_first_desc0(__le64 descq)
668 {
669 return descq | cpu_to_le64(1ULL << 12);
670 }
671
ipath_sdma_make_last_desc0(__le64 descq)672 static inline __le64 ipath_sdma_make_last_desc0(__le64 descq)
673 {
674 /* last */ /* dma head */
675 return descq | cpu_to_le64(1ULL << 11 | 1ULL << 13);
676 }
677
ipath_sdma_make_desc1(u64 addr)678 static inline __le64 ipath_sdma_make_desc1(u64 addr)
679 {
680 /* SDmaPhyAddr[47:32] */
681 return cpu_to_le64(addr >> 32);
682 }
683
ipath_user_sdma_send_frag(struct ipath_devdata * dd,struct ipath_user_sdma_pkt * pkt,int idx,unsigned ofs,u16 tail)684 static void ipath_user_sdma_send_frag(struct ipath_devdata *dd,
685 struct ipath_user_sdma_pkt *pkt, int idx,
686 unsigned ofs, u16 tail)
687 {
688 const u64 addr = (u64) pkt->addr[idx].addr +
689 (u64) pkt->addr[idx].offset;
690 const u64 dwlen = (u64) pkt->addr[idx].length / 4;
691 __le64 *descqp;
692 __le64 descq0;
693
694 descqp = &dd->ipath_sdma_descq[tail].qw[0];
695
696 descq0 = ipath_sdma_make_desc0(dd, addr, dwlen, ofs);
697 if (idx == 0)
698 descq0 = ipath_sdma_make_first_desc0(descq0);
699 if (idx == pkt->naddr - 1)
700 descq0 = ipath_sdma_make_last_desc0(descq0);
701
702 descqp[0] = descq0;
703 descqp[1] = ipath_sdma_make_desc1(addr);
704 }
705
706 /* pq->lock must be held, get packets on the wire... */
ipath_user_sdma_push_pkts(struct ipath_devdata * dd,struct ipath_user_sdma_queue * pq,struct list_head * pktlist)707 static int ipath_user_sdma_push_pkts(struct ipath_devdata *dd,
708 struct ipath_user_sdma_queue *pq,
709 struct list_head *pktlist)
710 {
711 int ret = 0;
712 unsigned long flags;
713 u16 tail;
714
715 if (list_empty(pktlist))
716 return 0;
717
718 if (unlikely(!(dd->ipath_flags & IPATH_LINKACTIVE)))
719 return -ECOMM;
720
721 spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
722
723 if (unlikely(dd->ipath_sdma_status & IPATH_SDMA_ABORT_MASK)) {
724 ret = -ECOMM;
725 goto unlock;
726 }
727
728 tail = dd->ipath_sdma_descq_tail;
729 while (!list_empty(pktlist)) {
730 struct ipath_user_sdma_pkt *pkt =
731 list_entry(pktlist->next, struct ipath_user_sdma_pkt,
732 list);
733 int i;
734 unsigned ofs = 0;
735 u16 dtail = tail;
736
737 if (pkt->naddr > ipath_sdma_descq_freecnt(dd))
738 goto unlock_check_tail;
739
740 for (i = 0; i < pkt->naddr; i++) {
741 ipath_user_sdma_send_frag(dd, pkt, i, ofs, tail);
742 ofs += pkt->addr[i].length >> 2;
743
744 if (++tail == dd->ipath_sdma_descq_cnt) {
745 tail = 0;
746 ++dd->ipath_sdma_generation;
747 }
748 }
749
750 if ((ofs<<2) > dd->ipath_ibmaxlen) {
751 ipath_dbg("packet size %X > ibmax %X, fail\n",
752 ofs<<2, dd->ipath_ibmaxlen);
753 ret = -EMSGSIZE;
754 goto unlock;
755 }
756
757 /*
758 * if the packet is >= 2KB mtu equivalent, we have to use
759 * the large buffers, and have to mark each descriptor as
760 * part of a large buffer packet.
761 */
762 if (ofs >= IPATH_SMALLBUF_DWORDS) {
763 for (i = 0; i < pkt->naddr; i++) {
764 dd->ipath_sdma_descq[dtail].qw[0] |=
765 cpu_to_le64(1ULL << 14);
766 if (++dtail == dd->ipath_sdma_descq_cnt)
767 dtail = 0;
768 }
769 }
770
771 dd->ipath_sdma_descq_added += pkt->naddr;
772 pkt->added = dd->ipath_sdma_descq_added;
773 list_move_tail(&pkt->list, &pq->sent);
774 ret++;
775 }
776
777 unlock_check_tail:
778 /* advance the tail on the chip if necessary */
779 if (dd->ipath_sdma_descq_tail != tail) {
780 wmb();
781 ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, tail);
782 dd->ipath_sdma_descq_tail = tail;
783 }
784
785 unlock:
786 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
787
788 return ret;
789 }
790
ipath_user_sdma_writev(struct ipath_devdata * dd,struct ipath_user_sdma_queue * pq,const struct iovec * iov,unsigned long dim)791 int ipath_user_sdma_writev(struct ipath_devdata *dd,
792 struct ipath_user_sdma_queue *pq,
793 const struct iovec *iov,
794 unsigned long dim)
795 {
796 int ret = 0;
797 struct list_head list;
798 int npkts = 0;
799
800 INIT_LIST_HEAD(&list);
801
802 mutex_lock(&pq->lock);
803
804 if (dd->ipath_sdma_descq_added != dd->ipath_sdma_descq_removed) {
805 ipath_user_sdma_hwqueue_clean(dd);
806 ipath_user_sdma_queue_clean(dd, pq);
807 }
808
809 while (dim) {
810 const int mxp = 8;
811
812 ret = ipath_user_sdma_queue_pkts(dd, pq, &list, iov, dim, mxp);
813 if (ret <= 0)
814 goto done_unlock;
815 else {
816 dim -= ret;
817 iov += ret;
818 }
819
820 /* force packets onto the sdma hw queue... */
821 if (!list_empty(&list)) {
822 /*
823 * lazily clean hw queue. the 4 is a guess of about
824 * how many sdma descriptors a packet will take (it
825 * doesn't have to be perfect).
826 */
827 if (ipath_sdma_descq_freecnt(dd) < ret * 4) {
828 ipath_user_sdma_hwqueue_clean(dd);
829 ipath_user_sdma_queue_clean(dd, pq);
830 }
831
832 ret = ipath_user_sdma_push_pkts(dd, pq, &list);
833 if (ret < 0)
834 goto done_unlock;
835 else {
836 npkts += ret;
837 pq->counter += ret;
838
839 if (!list_empty(&list))
840 goto done_unlock;
841 }
842 }
843 }
844
845 done_unlock:
846 if (!list_empty(&list))
847 ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &list);
848 mutex_unlock(&pq->lock);
849
850 return (ret < 0) ? ret : npkts;
851 }
852
ipath_user_sdma_make_progress(struct ipath_devdata * dd,struct ipath_user_sdma_queue * pq)853 int ipath_user_sdma_make_progress(struct ipath_devdata *dd,
854 struct ipath_user_sdma_queue *pq)
855 {
856 int ret = 0;
857
858 mutex_lock(&pq->lock);
859 ipath_user_sdma_hwqueue_clean(dd);
860 ret = ipath_user_sdma_queue_clean(dd, pq);
861 mutex_unlock(&pq->lock);
862
863 return ret;
864 }
865
ipath_user_sdma_complete_counter(const struct ipath_user_sdma_queue * pq)866 u32 ipath_user_sdma_complete_counter(const struct ipath_user_sdma_queue *pq)
867 {
868 return pq->sent_counter;
869 }
870
ipath_user_sdma_inflight_counter(struct ipath_user_sdma_queue * pq)871 u32 ipath_user_sdma_inflight_counter(struct ipath_user_sdma_queue *pq)
872 {
873 return pq->counter;
874 }
875
876