Commit | Line | Data |
---|---|---|
800f879a AT |
1 | /* |
2 | * Copyright 2010-2017 Intel Corporation. | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License, version 2, | |
6 | * as published by the Free Software Foundation. | |
7 | * | |
8 | * This program is distributed in the hope that it will be useful, | |
9 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
10 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
11 | * General Public License for more details. | |
12 | * | |
13 | * Disclaimer: The codes contained in these modules may be specific to | |
14 | * the Intel Software Development Platform codenamed Knights Ferry, | |
15 | * and the Intel product codenamed Knights Corner, and are not backward | |
16 | * compatible with other Intel products. Additionally, Intel will NOT | |
17 | * support the codes or instruction set in future products. | |
18 | * | |
19 | * Intel offers no warranty of any kind regarding the code. This code is | |
20 | * licensed on an "AS IS" basis and Intel is not obligated to provide | |
21 | * any support, assistance, installation, training, or other services | |
22 | * of any kind. Intel is also not obligated to provide any updates, | |
23 | * enhancements or extensions. Intel specifically disclaims any warranty | |
24 | * of merchantability, non-infringement, fitness for any particular | |
25 | * purpose, and any other warranty. | |
26 | * | |
27 | * Further, Intel disclaims all liability of any kind, including but | |
28 | * not limited to liability for infringement of any proprietary rights, | |
29 | * relating to the use of the code, even if Intel is notified of the | |
30 | * possibility of such liability. Except as expressly stated in an Intel | |
31 | * license agreement provided with this code and agreed upon with Intel, | |
32 | * no license, express or implied, by estoppel or otherwise, to any | |
33 | * intellectual property rights is granted herein. | |
34 | */ | |
35 | ||
36 | #include "mic/micscif.h" | |
37 | #include "mic/micscif_smpt.h" | |
38 | #include "mic/micscif_kmem_cache.h" | |
39 | #include "mic/micscif_rma_list.h" | |
40 | #ifndef _MIC_SCIF_ | |
41 | #include "mic_common.h" | |
42 | #endif | |
43 | #include "mic/mic_dma_api.h" | |
44 | #include "mic/micscif_map.h" | |
45 | ||
46 | bool mic_reg_cache_enable = 0; | |
47 | ||
48 | bool mic_huge_page_enable = 1; | |
49 | ||
50 | #ifdef _MIC_SCIF_ | |
51 | mic_dma_handle_t mic_dma_handle; | |
52 | #endif | |
53 | static inline | |
54 | void micscif_rma_destroy_tcw(struct rma_mmu_notifier *mmn, | |
55 | struct endpt *ep, bool inrange, | |
56 | uint64_t start, uint64_t len); | |
57 | #ifdef CONFIG_MMU_NOTIFIER | |
58 | static void scif_mmu_notifier_release(struct mmu_notifier *mn, | |
59 | struct mm_struct *mm); | |
60 | static void scif_mmu_notifier_invalidate_page(struct mmu_notifier *mn, | |
61 | struct mm_struct *mm, | |
62 | unsigned long address); | |
63 | static void scif_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, | |
64 | struct mm_struct *mm, | |
65 | unsigned long start, unsigned long end); | |
66 | static void scif_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, | |
67 | struct mm_struct *mm, | |
68 | unsigned long start, unsigned long end); | |
69 | static const struct mmu_notifier_ops scif_mmu_notifier_ops = { | |
70 | .release = scif_mmu_notifier_release, | |
71 | .clear_flush_young = NULL, | |
72 | .change_pte = NULL,/*TODO*/ | |
73 | .invalidate_page = scif_mmu_notifier_invalidate_page, | |
74 | .invalidate_range_start = scif_mmu_notifier_invalidate_range_start, | |
75 | .invalidate_range_end = scif_mmu_notifier_invalidate_range_end}; | |
76 | ||
77 | static void scif_mmu_notifier_release(struct mmu_notifier *mn, | |
78 | struct mm_struct *mm) | |
79 | { | |
80 | struct endpt *ep; | |
81 | struct rma_mmu_notifier *mmn; | |
82 | mmn = container_of(mn, struct rma_mmu_notifier, ep_mmu_notifier); | |
83 | ep = mmn->ep; | |
84 | micscif_rma_destroy_tcw(mmn, ep, false, 0, 0); | |
85 | pr_debug("%s\n", __func__); | |
86 | return; | |
87 | } | |
88 | ||
89 | static void scif_mmu_notifier_invalidate_page(struct mmu_notifier *mn, | |
90 | struct mm_struct *mm, | |
91 | unsigned long address) | |
92 | { | |
93 | struct endpt *ep; | |
94 | struct rma_mmu_notifier *mmn; | |
95 | mmn = container_of(mn, struct rma_mmu_notifier, ep_mmu_notifier); | |
96 | ep = mmn->ep; | |
97 | micscif_rma_destroy_tcw(mmn, ep, true, address, PAGE_SIZE); | |
98 | pr_debug("%s address 0x%lx\n", __func__, address); | |
99 | return; | |
100 | } | |
101 | ||
102 | static void scif_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, | |
103 | struct mm_struct *mm, | |
104 | unsigned long start, unsigned long end) | |
105 | { | |
106 | struct endpt *ep; | |
107 | struct rma_mmu_notifier *mmn; | |
108 | mmn = container_of(mn, struct rma_mmu_notifier, ep_mmu_notifier); | |
109 | ep = mmn->ep; | |
110 | micscif_rma_destroy_tcw(mmn, ep, true, (uint64_t)start, (uint64_t)(end - start)); | |
111 | pr_debug("%s start=%lx, end=%lx\n", __func__, start, end); | |
112 | return; | |
113 | } | |
114 | ||
115 | static void scif_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, | |
116 | struct mm_struct *mm, | |
117 | unsigned long start, unsigned long end) | |
118 | { | |
119 | /* Nothing to do here, everything needed was done in invalidate_range_start */ | |
120 | pr_debug("%s\n", __func__); | |
121 | return; | |
122 | } | |
123 | #endif | |
124 | ||
125 | #ifdef CONFIG_MMU_NOTIFIER | |
126 | void ep_unregister_mmu_notifier(struct endpt *ep) | |
127 | { | |
128 | struct endpt_rma_info *rma = &ep->rma_info; | |
129 | struct rma_mmu_notifier *mmn = NULL; | |
130 | struct list_head *item, *tmp; | |
131 | mutex_lock(&ep->rma_info.mmn_lock); | |
132 | list_for_each_safe(item, tmp, &rma->mmn_list) { | |
133 | mmn = list_entry(item, | |
134 | struct rma_mmu_notifier, list_member); | |
135 | mmu_notifier_unregister(&mmn->ep_mmu_notifier, mmn->mm); | |
136 | #ifdef RMA_DEBUG | |
137 | BUG_ON(atomic_long_sub_return(1, &ms_info.mmu_notif_cnt) < 0); | |
138 | #endif | |
139 | list_del(item); | |
140 | kfree(mmn); | |
141 | } | |
142 | mutex_unlock(&ep->rma_info.mmn_lock); | |
143 | } | |
144 | ||
145 | static void init_mmu_notifier(struct rma_mmu_notifier *mmn, struct mm_struct *mm, struct endpt *ep) | |
146 | { | |
147 | mmn->ep = ep; | |
148 | mmn->mm = mm; | |
149 | mmn->ep_mmu_notifier.ops = &scif_mmu_notifier_ops; | |
150 | INIT_LIST_HEAD(&mmn->list_member); | |
151 | INIT_LIST_HEAD(&mmn->tc_reg_list); | |
152 | } | |
153 | ||
154 | static struct rma_mmu_notifier *find_mmu_notifier(struct mm_struct *mm, struct endpt_rma_info *rma) | |
155 | { | |
156 | struct rma_mmu_notifier *mmn; | |
157 | struct list_head *item; | |
158 | list_for_each(item, &rma->mmn_list) { | |
159 | mmn = list_entry(item, | |
160 | struct rma_mmu_notifier, list_member); | |
161 | if (mmn->mm == mm) | |
162 | return mmn; | |
163 | } | |
164 | return NULL; | |
165 | } | |
166 | #endif | |
167 | ||
168 | /** | |
169 | * micscif_rma_ep_init: | |
170 | * @ep: end point | |
171 | * | |
172 | * Initialize RMA per EP data structures. | |
173 | */ | |
174 | int micscif_rma_ep_init(struct endpt *ep) | |
175 | { | |
176 | int ret; | |
177 | struct endpt_rma_info *rma = &ep->rma_info; | |
178 | ||
179 | mutex_init (&rma->rma_lock); | |
180 | if ((ret = va_gen_init(&rma->va_gen, | |
181 | VA_GEN_MIN, VA_GEN_RANGE)) < 0) | |
182 | goto init_err; | |
183 | spin_lock_init(&rma->tc_lock); | |
184 | mutex_init (&rma->mmn_lock); | |
185 | mutex_init (&rma->va_lock); | |
186 | INIT_LIST_HEAD(&rma->reg_list); | |
187 | INIT_LIST_HEAD(&rma->remote_reg_list); | |
188 | atomic_set(&rma->tw_refcount, 0); | |
189 | atomic_set(&rma->tw_total_pages, 0); | |
190 | atomic_set(&rma->tcw_refcount, 0); | |
191 | atomic_set(&rma->tcw_total_pages, 0); | |
192 | init_waitqueue_head(&rma->fence_wq); | |
193 | rma->fence_refcount = 0; | |
194 | rma->async_list_del = 0; | |
195 | rma->dma_chan = NULL; | |
196 | INIT_LIST_HEAD(&rma->mmn_list); | |
197 | INIT_LIST_HEAD(&rma->task_list); | |
198 | init_err: | |
199 | return ret; | |
200 | } | |
201 | ||
202 | /** | |
203 | * micscif_rma_ep_can_uninit: | |
204 | * @ep: end point | |
205 | * | |
206 | * Returns 1 if an endpoint can be uninitialized and 0 otherwise. | |
207 | */ | |
208 | int micscif_rma_ep_can_uninit(struct endpt *ep) | |
209 | { | |
210 | int ret = 0; | |
211 | ||
212 | /* Destroy RMA Info only if both lists are empty */ | |
213 | if (list_empty(&ep->rma_info.reg_list) && | |
214 | list_empty(&ep->rma_info.remote_reg_list) && | |
215 | #ifdef CONFIG_MMU_NOTIFIER | |
216 | list_empty(&ep->rma_info.mmn_list) && | |
217 | #endif | |
218 | !atomic_read(&ep->rma_info.tw_refcount) && | |
219 | !atomic_read(&ep->rma_info.tcw_refcount)) | |
220 | ret = 1; | |
221 | return ret; | |
222 | } | |
223 | ||
224 | #ifdef _MIC_SCIF_ | |
225 | /** | |
226 | * __micscif_setup_proxy_dma: | |
227 | * @ep: SCIF endpoint descriptor. | |
228 | * | |
229 | * Sets up data structures for P2P Proxy DMAs. | |
230 | */ | |
231 | static int __micscif_setup_proxy_dma(struct endpt *ep) | |
232 | { | |
233 | struct endpt_rma_info *rma = &ep->rma_info; | |
234 | int err = 0; | |
235 | uint64_t *tmp = NULL; | |
236 | ||
237 | mutex_lock(&rma->rma_lock); | |
238 | if (is_p2p_scifdev(ep->remote_dev) && !rma->proxy_dma_va) { | |
239 | if (!(tmp = scif_zalloc(PAGE_SIZE))) { | |
240 | err = -ENOMEM; | |
241 | goto error; | |
242 | } | |
243 | if ((err = map_virt_into_aperture(&rma->proxy_dma_phys, | |
244 | tmp, | |
245 | ep->remote_dev, PAGE_SIZE))) { | |
246 | scif_free(tmp, PAGE_SIZE); | |
247 | goto error; | |
248 | } | |
249 | *tmp = OP_IDLE; | |
250 | rma->proxy_dma_va = tmp; | |
251 | } | |
252 | error: | |
253 | mutex_unlock(&rma->rma_lock); | |
254 | return err; | |
255 | } | |
256 | ||
257 | static __always_inline int micscif_setup_proxy_dma(struct endpt *ep) | |
258 | { | |
259 | if (ep->rma_info.proxy_dma_va) | |
260 | return 0; | |
261 | ||
262 | return __micscif_setup_proxy_dma(ep); | |
263 | } | |
264 | ||
265 | /** | |
266 | * micscif_teardown_proxy_dma: | |
267 | * @ep: SCIF endpoint descriptor. | |
268 | * | |
269 | * Tears down data structures setup for P2P Proxy DMAs. | |
270 | */ | |
271 | void micscif_teardown_proxy_dma(struct endpt *ep) | |
272 | { | |
273 | struct endpt_rma_info *rma = &ep->rma_info; | |
274 | mutex_lock(&rma->rma_lock); | |
275 | if (rma->proxy_dma_va) { | |
276 | unmap_from_aperture(rma->proxy_dma_phys, ep->remote_dev, PAGE_SIZE); | |
277 | scif_free(rma->proxy_dma_va, PAGE_SIZE); | |
278 | rma->proxy_dma_va = NULL; | |
279 | } | |
280 | mutex_unlock(&rma->rma_lock); | |
281 | } | |
282 | ||
283 | /** | |
284 | * micscif_proxy_dma: | |
285 | * @ep: SCIF endpoint descriptor. | |
286 | * @copy_work: DMA copy work information. | |
287 | * | |
288 | * This API does the following: | |
289 | * 1) Sends the peer a SCIF Node QP message with the information | |
290 | * required to program a proxy DMA to covert a P2P Read to a Write | |
291 | * which will initiate a DMA transfer from the peer card to self. | |
292 | * The reason for this special code path is KNF and KNC P2P read | |
293 | * performance being much lower than P2P write performance on Crown | |
294 | * Pass platforms. | |
295 | * 2) Poll for an update of the known proxy dma VA to OP_COMPLETED | |
296 | * via a SUD by the peer. | |
297 | */ | |
298 | static int micscif_proxy_dma(scif_epd_t epd, struct mic_copy_work *work) | |
299 | { | |
300 | struct endpt *ep = (struct endpt *)epd; | |
301 | struct nodemsg msg; | |
302 | unsigned long ts = jiffies; | |
303 | struct endpt_rma_info *rma = &ep->rma_info; | |
304 | int err; | |
305 | volatile uint64_t *proxy_dma_va = rma->proxy_dma_va; | |
306 | ||
307 | mutex_lock(&ep->rma_info.rma_lock); | |
308 | /* | |
309 | * Bail out if there is a Proxy DMA already in progress | |
310 | * for this endpoint. The callee will fallback on self | |
311 | * DMAs upon an error. | |
312 | */ | |
313 | if (*proxy_dma_va != OP_IDLE) { | |
314 | mutex_unlock(&ep->rma_info.rma_lock); | |
315 | err = -EBUSY; | |
316 | goto error; | |
317 | } | |
318 | *proxy_dma_va = OP_IN_PROGRESS; | |
319 | mutex_unlock(&ep->rma_info.rma_lock); | |
320 | ||
321 | msg.src = ep->port; | |
322 | msg.uop = work->ordered ? SCIF_PROXY_ORDERED_DMA : SCIF_PROXY_DMA; | |
323 | msg.payload[0] = ep->remote_ep; | |
324 | msg.payload[1] = work->src_offset; | |
325 | msg.payload[2] = work->dst_offset; | |
326 | msg.payload[3] = work->len; | |
327 | ||
328 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) | |
329 | goto error_init_va; | |
330 | ||
331 | while (*proxy_dma_va != OP_COMPLETED) { | |
332 | schedule(); | |
333 | if (time_after(jiffies, | |
334 | ts + NODE_ALIVE_TIMEOUT)) { | |
335 | err = -EBUSY; | |
336 | goto error_init_va; | |
337 | } | |
338 | } | |
339 | err = 0; | |
340 | error_init_va: | |
341 | *proxy_dma_va = OP_IDLE; | |
342 | error: | |
343 | return err; | |
344 | } | |
345 | #endif | |
346 | ||
347 | /** | |
348 | * micscif_create_pinned_pages: | |
349 | * @nr_pages: number of pages in window | |
350 | * @prot: read/write protection | |
351 | * | |
352 | * Allocate and prepare a set of pinned pages. | |
353 | */ | |
354 | struct scif_pinned_pages *micscif_create_pinned_pages(int nr_pages, int prot) | |
355 | { | |
356 | struct scif_pinned_pages *pinned_pages; | |
357 | ||
358 | might_sleep(); | |
359 | if (!(pinned_pages = scif_zalloc(sizeof(*pinned_pages)))) | |
360 | goto error; | |
361 | ||
362 | if (!(pinned_pages->pages = scif_zalloc(nr_pages * | |
363 | sizeof(*(pinned_pages->pages))))) | |
364 | goto error_free_pinned_pages; | |
365 | ||
366 | if (!(pinned_pages->num_pages = scif_zalloc(nr_pages * | |
367 | sizeof(*(pinned_pages->num_pages))))) | |
368 | goto error_free_pages; | |
369 | ||
370 | #if !defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HUGETLB_PAGE) && !defined(_MIC_SCIF_) | |
371 | if (!(pinned_pages->vma = scif_zalloc(nr_pages * | |
372 | sizeof(*(pinned_pages->vma))))) | |
373 | goto error_free_num_pages; | |
374 | #endif | |
375 | ||
376 | pinned_pages->prot = prot; | |
377 | pinned_pages->magic = SCIFEP_MAGIC; | |
378 | pinned_pages->nr_contig_chunks = 0; | |
379 | return pinned_pages; | |
380 | ||
381 | #if !defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HUGETLB_PAGE) && !defined(_MIC_SCIF_) | |
382 | error_free_num_pages: | |
383 | scif_free(pinned_pages->num_pages, | |
384 | pinned_pages->nr_pages * sizeof(*(pinned_pages->num_pages))); | |
385 | #endif | |
386 | error_free_pages: | |
387 | scif_free(pinned_pages->pages, | |
388 | pinned_pages->nr_pages * sizeof(*(pinned_pages->pages))); | |
389 | error_free_pinned_pages: | |
390 | scif_free(pinned_pages, sizeof(*pinned_pages)); | |
391 | error: | |
392 | return NULL; | |
393 | } | |
394 | ||
395 | /** | |
396 | * micscif_destroy_pinned_pages: | |
397 | * @pinned_pages: A set of pinned pages. | |
398 | * | |
399 | * Deallocate resources for pinned pages. | |
400 | */ | |
401 | int micscif_destroy_pinned_pages(struct scif_pinned_pages *pinned_pages) | |
402 | { | |
403 | int j; | |
404 | int writeable = pinned_pages->prot & SCIF_PROT_WRITE; | |
405 | int kernel = SCIF_MAP_KERNEL & pinned_pages->map_flags; | |
406 | ||
407 | for (j = 0; j < pinned_pages->nr_pages; j++) { | |
408 | if (pinned_pages->pages[j]) { | |
409 | if (!kernel) { | |
410 | if (writeable) | |
411 | SetPageDirty(pinned_pages->pages[j]); | |
412 | #ifdef RMA_DEBUG | |
413 | BUG_ON(!page_count(pinned_pages->pages[j])); | |
414 | BUG_ON(atomic_long_sub_return(1, &ms_info.rma_pin_cnt) < 0); | |
415 | #endif | |
60589c21 | 416 | put_page(pinned_pages->pages[j]); |
800f879a AT |
417 | } |
418 | } | |
419 | } | |
420 | ||
421 | #if !defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HUGETLB_PAGE) && !defined(_MIC_SCIF_) | |
422 | scif_free(pinned_pages->vma, | |
423 | pinned_pages->nr_pages * sizeof(*(pinned_pages->vma))); | |
424 | #endif | |
425 | scif_free(pinned_pages->pages, | |
426 | pinned_pages->nr_pages * sizeof(*(pinned_pages->pages))); | |
427 | scif_free(pinned_pages->num_pages, | |
428 | pinned_pages->nr_pages * sizeof(*(pinned_pages->num_pages))); | |
429 | scif_free(pinned_pages, sizeof(*pinned_pages)); | |
430 | return 0; | |
431 | } | |
432 | ||
433 | /* | |
434 | * micscif_create_window: | |
435 | * @ep: end point | |
436 | * @pinned_pages: Set of pinned pages which wil back this window. | |
437 | * @offset: offset hint | |
438 | * | |
439 | * Allocate and prepare a self registration window. | |
440 | */ | |
441 | struct reg_range_t *micscif_create_window(struct endpt *ep, | |
442 | int64_t nr_pages, uint64_t offset, bool temp) | |
443 | { | |
444 | struct reg_range_t *window; | |
445 | ||
446 | might_sleep(); | |
447 | if (!(window = scif_zalloc(sizeof(struct reg_range_t)))) | |
448 | goto error; | |
449 | ||
450 | #ifdef CONFIG_ML1OM | |
451 | if (!temp) { | |
452 | if (!(window->phys_addr = scif_zalloc(nr_pages * | |
453 | sizeof(*(window->phys_addr))))) | |
454 | goto error_free_window; | |
455 | ||
456 | if (!(window->temp_phys_addr = scif_zalloc(nr_pages * | |
457 | sizeof(*(window->temp_phys_addr))))) | |
458 | goto error_free_window; | |
459 | } | |
460 | #endif | |
461 | ||
462 | if (!(window->dma_addr = scif_zalloc(nr_pages * | |
463 | sizeof(*(window->dma_addr))))) | |
464 | goto error_free_window; | |
465 | ||
466 | if (!(window->num_pages = scif_zalloc(nr_pages * | |
467 | sizeof(*(window->num_pages))))) | |
468 | goto error_free_window; | |
469 | ||
470 | window->offset = offset; | |
471 | window->ep = (uint64_t)ep; | |
472 | window->magic = SCIFEP_MAGIC; | |
473 | window->reg_state = OP_IDLE; | |
474 | init_waitqueue_head(&window->regwq); | |
475 | window->unreg_state = OP_IDLE; | |
476 | init_waitqueue_head(&window->unregwq); | |
477 | INIT_LIST_HEAD(&window->list_member); | |
478 | window->type = RMA_WINDOW_SELF; | |
479 | window->temp = temp; | |
480 | #ifdef _MIC_SCIF_ | |
481 | micscif_setup_proxy_dma(ep); | |
482 | #endif | |
483 | return window; | |
484 | ||
485 | error_free_window: | |
486 | if (window->dma_addr) | |
487 | scif_free(window->dma_addr, nr_pages * sizeof(*(window->dma_addr))); | |
488 | #ifdef CONFIG_ML1OM | |
489 | if (window->temp_phys_addr) | |
490 | scif_free(window->temp_phys_addr, nr_pages * sizeof(*(window->temp_phys_addr))); | |
491 | if (window->phys_addr) | |
492 | scif_free(window->phys_addr, nr_pages * sizeof(*(window->phys_addr))); | |
493 | #endif | |
494 | scif_free(window, sizeof(*window)); | |
495 | error: | |
496 | return NULL; | |
497 | } | |
498 | ||
499 | /** | |
500 | * micscif_destroy_incomplete_window: | |
501 | * @ep: end point | |
502 | * @window: registration window | |
503 | * | |
504 | * Deallocate resources for self window. | |
505 | */ | |
506 | int micscif_destroy_incomplete_window(struct endpt *ep, struct reg_range_t *window) | |
507 | { | |
508 | int err; | |
509 | int64_t nr_pages = window->nr_pages; | |
510 | struct allocmsg *alloc = &window->alloc_handle; | |
511 | struct nodemsg msg; | |
512 | ||
513 | RMA_MAGIC(window); | |
514 | retry: | |
515 | err = wait_event_timeout(alloc->allocwq, alloc->state != OP_IN_PROGRESS, NODE_ALIVE_TIMEOUT); | |
516 | if (!err && scifdev_alive(ep)) | |
517 | goto retry; | |
518 | ||
519 | if (OP_COMPLETED == alloc->state) { | |
520 | msg.uop = SCIF_FREE_VIRT; | |
521 | msg.src = ep->port; | |
522 | msg.payload[0] = ep->remote_ep; | |
523 | msg.payload[1] = (uint64_t)window->alloc_handle.vaddr; | |
524 | msg.payload[2] = (uint64_t)window; | |
525 | msg.payload[3] = SCIF_REGISTER; | |
526 | micscif_nodeqp_send(ep->remote_dev, &msg, ep); | |
527 | } | |
528 | ||
529 | micscif_free_window_offset(ep, window->offset, | |
530 | window->nr_pages << PAGE_SHIFT); | |
531 | if (window->dma_addr) | |
532 | scif_free(window->dma_addr, nr_pages * | |
533 | sizeof(*(window->dma_addr))); | |
534 | if (window->num_pages) | |
535 | scif_free(window->num_pages, nr_pages * | |
536 | sizeof(*(window->num_pages))); | |
537 | #ifdef CONFIG_ML1OM | |
538 | if (window->phys_addr) | |
539 | scif_free(window->phys_addr, window->nr_pages * | |
540 | sizeof(*(window->phys_addr))); | |
541 | if (window->temp_phys_addr) | |
542 | scif_free(window->temp_phys_addr, nr_pages * | |
543 | sizeof(*(window->temp_phys_addr))); | |
544 | #endif | |
545 | scif_free(window, sizeof(*window)); | |
546 | return 0; | |
547 | } | |
548 | ||
549 | /** | |
550 | * micscif_destroy_window: | |
551 | * @ep: end point | |
552 | * @window: registration window | |
553 | * | |
554 | * Deallocate resources for self window. | |
555 | */ | |
556 | int micscif_destroy_window(struct endpt *ep, struct reg_range_t *window) | |
557 | { | |
558 | int j; | |
559 | struct scif_pinned_pages *pinned_pages = window->pinned_pages; | |
560 | int64_t nr_pages = window->nr_pages; | |
561 | ||
562 | might_sleep(); | |
563 | RMA_MAGIC(window); | |
564 | if (!window->temp && window->mm) { | |
565 | __scif_dec_pinned_vm_lock(window->mm, window->nr_pages, 0); | |
566 | __scif_release_mm(window->mm); | |
567 | window->mm = NULL; | |
568 | } | |
569 | ||
570 | if (!window->offset_freed) | |
571 | micscif_free_window_offset(ep, window->offset, | |
572 | window->nr_pages << PAGE_SHIFT); | |
573 | for (j = 0; j < window->nr_contig_chunks; j++) { | |
574 | if (window->dma_addr[j]) { | |
575 | unmap_from_aperture( | |
576 | window->dma_addr[j], | |
577 | ep->remote_dev, | |
578 | window->num_pages[j] << PAGE_SHIFT); | |
579 | } | |
580 | } | |
581 | ||
582 | /* | |
583 | * Decrement references for this set of pinned pages from | |
584 | * this window. | |
585 | */ | |
586 | j = atomic_sub_return((int32_t)pinned_pages->nr_pages, | |
587 | &pinned_pages->ref_count); | |
588 | BUG_ON(j < 0); | |
589 | /* | |
590 | * If the ref count for pinned_pages is zero then someone | |
591 | * has already called scif_unpin_pages() for it and we should | |
592 | * destroy the page cache. | |
593 | */ | |
594 | if (!j) | |
595 | micscif_destroy_pinned_pages(window->pinned_pages); | |
596 | if (window->dma_addr) | |
597 | scif_free(window->dma_addr, nr_pages * | |
598 | sizeof(*(window->dma_addr))); | |
599 | if (window->num_pages) | |
600 | scif_free(window->num_pages, nr_pages * | |
601 | sizeof(*(window->num_pages))); | |
602 | #ifdef CONFIG_ML1OM | |
603 | if (window->phys_addr) | |
604 | scif_free(window->phys_addr, window->nr_pages * | |
605 | sizeof(*(window->phys_addr))); | |
606 | if (window->temp_phys_addr) | |
607 | scif_free(window->temp_phys_addr, nr_pages * | |
608 | sizeof(*(window->temp_phys_addr))); | |
609 | #endif | |
610 | window->magic = 0; | |
611 | scif_free(window, sizeof(*window)); | |
612 | return 0; | |
613 | } | |
614 | ||
615 | /** | |
616 | * micscif_create_remote_lookup: | |
617 | * @ep: end point | |
618 | * @window: remote window | |
619 | * | |
620 | * Allocate and prepare lookup entries for the remote | |
621 | * end to copy over the physical addresses. | |
622 | * Returns 0 on success and appropriate errno on failure. | |
623 | */ | |
624 | int micscif_create_remote_lookup(struct endpt *ep, struct reg_range_t *window) | |
625 | { | |
626 | int i, j, err = 0; | |
627 | int64_t nr_pages = window->nr_pages; | |
628 | bool vmalloc_dma_phys; | |
629 | #ifdef CONFIG_ML1OM | |
630 | bool vmalloc_temp_phys = false; | |
631 | bool vmalloc_phys = false; | |
632 | #endif | |
633 | might_sleep(); | |
634 | ||
635 | /* Map window */ | |
636 | err = map_virt_into_aperture(&window->mapped_offset, | |
637 | window, ep->remote_dev, sizeof(*window)); | |
638 | if (err) | |
639 | goto error_window; | |
640 | ||
641 | /* Compute the number of lookup entries. 21 == 2MB Shift */ | |
642 | window->nr_lookup = ALIGN(nr_pages * PAGE_SIZE, | |
643 | ((2) * 1024 * 1024)) >> 21; | |
644 | ||
645 | if (!(window->dma_addr_lookup.lookup = | |
646 | scif_zalloc(window->nr_lookup * | |
647 | sizeof(*(window->dma_addr_lookup.lookup))))) | |
648 | goto error_window; | |
649 | ||
650 | /* Map DMA physical addess lookup array */ | |
651 | err = map_virt_into_aperture(&window->dma_addr_lookup.offset, | |
652 | window->dma_addr_lookup.lookup, ep->remote_dev, | |
653 | window->nr_lookup * | |
654 | sizeof(*window->dma_addr_lookup.lookup)); | |
655 | if (err) | |
656 | goto error_window; | |
657 | ||
658 | vmalloc_dma_phys = is_vmalloc_addr(&window->dma_addr[0]); | |
659 | ||
660 | #ifdef CONFIG_ML1OM | |
661 | if (ep->remote_dev != &scif_dev[SCIF_HOST_NODE] && !is_self_scifdev(ep->remote_dev)) { | |
662 | if (!(window->temp_phys_addr_lookup.lookup = | |
663 | scif_zalloc(window->nr_lookup * | |
664 | sizeof(*(window->temp_phys_addr_lookup.lookup))))) | |
665 | goto error_window; | |
666 | ||
667 | /* Map physical addess lookup array */ | |
668 | err = map_virt_into_aperture(&window->temp_phys_addr_lookup.offset, | |
669 | window->temp_phys_addr_lookup.lookup, ep->remote_dev, | |
670 | window->nr_lookup * | |
671 | sizeof(*window->temp_phys_addr_lookup.lookup)); | |
672 | if (err) | |
673 | goto error_window; | |
674 | ||
675 | if (!(window->phys_addr_lookup.lookup = | |
676 | scif_zalloc(window->nr_lookup * | |
677 | sizeof(*(window->phys_addr_lookup.lookup))))) | |
678 | goto error_window; | |
679 | ||
680 | /* Map physical addess lookup array */ | |
681 | err = map_virt_into_aperture(&window->phys_addr_lookup.offset, | |
682 | window->phys_addr_lookup.lookup, ep->remote_dev, | |
683 | window->nr_lookup * | |
684 | sizeof(*window->phys_addr_lookup.lookup)); | |
685 | if (err) | |
686 | goto error_window; | |
687 | ||
688 | vmalloc_phys = is_vmalloc_addr(&window->phys_addr[0]); | |
689 | vmalloc_temp_phys = is_vmalloc_addr(&window->temp_phys_addr[0]); | |
690 | } | |
691 | #endif | |
692 | ||
693 | /* Now map each of the pages containing physical addresses */ | |
694 | for (i = 0, j = 0; i < nr_pages; i += NR_PHYS_ADDR_IN_PAGE, j++) { | |
695 | #ifdef CONFIG_ML1OM | |
696 | if (ep->remote_dev != &scif_dev[SCIF_HOST_NODE] && !is_self_scifdev(ep->remote_dev)) { | |
697 | err = map_page_into_aperture( | |
698 | &window->temp_phys_addr_lookup.lookup[j], | |
699 | vmalloc_temp_phys ? | |
700 | vmalloc_to_page(&window->temp_phys_addr[i]) : | |
701 | virt_to_page(&window->temp_phys_addr[i]), | |
702 | ep->remote_dev); | |
703 | if (err) | |
704 | goto error_window; | |
705 | ||
706 | err = map_page_into_aperture( | |
707 | &window->phys_addr_lookup.lookup[j], | |
708 | vmalloc_phys ? | |
709 | vmalloc_to_page(&window->phys_addr[i]) : | |
710 | virt_to_page(&window->phys_addr[i]), | |
711 | ep->remote_dev); | |
712 | if (err) | |
713 | goto error_window; | |
714 | } | |
715 | #endif | |
716 | err = map_page_into_aperture( | |
717 | &window->dma_addr_lookup.lookup[j], | |
718 | vmalloc_dma_phys ? | |
719 | vmalloc_to_page(&window->dma_addr[i]) : | |
720 | virt_to_page(&window->dma_addr[i]), | |
721 | ep->remote_dev); | |
722 | if (err) | |
723 | goto error_window; | |
724 | } | |
725 | return 0; | |
726 | error_window: | |
727 | return err; | |
728 | } | |
729 | ||
730 | /** | |
731 | * micscif_destroy_remote_lookup: | |
732 | * @ep: end point | |
733 | * @window: remote window | |
734 | * | |
735 | * Destroy lookup entries used for the remote | |
736 | * end to copy over the physical addresses. | |
737 | */ | |
738 | void micscif_destroy_remote_lookup(struct endpt *ep, struct reg_range_t *window) | |
739 | { | |
740 | int i, j; | |
741 | ||
742 | RMA_MAGIC(window); | |
743 | if (window->nr_lookup) { | |
744 | for (i = 0, j = 0; i < window->nr_pages; | |
745 | i += NR_PHYS_ADDR_IN_PAGE, j++) { | |
746 | if (window->dma_addr_lookup.lookup && | |
747 | window->dma_addr_lookup.lookup[j]) { | |
748 | unmap_from_aperture( | |
749 | window->dma_addr_lookup.lookup[j], | |
750 | ep->remote_dev, PAGE_SIZE); | |
751 | } | |
752 | } | |
753 | if (window->dma_addr_lookup.offset) { | |
754 | unmap_from_aperture( | |
755 | window->dma_addr_lookup.offset, | |
756 | ep->remote_dev, window->nr_lookup * | |
757 | sizeof(*window->dma_addr_lookup.lookup)); | |
758 | } | |
759 | if (window->dma_addr_lookup.lookup) | |
760 | scif_free(window->dma_addr_lookup.lookup, window->nr_lookup * | |
761 | sizeof(*(window->dma_addr_lookup.lookup))); | |
762 | if (window->mapped_offset) { | |
763 | unmap_from_aperture(window->mapped_offset, | |
764 | ep->remote_dev, sizeof(*window)); | |
765 | } | |
766 | window->nr_lookup = 0; | |
767 | } | |
768 | } | |
769 | ||
770 | /** | |
771 | * micscif_create_remote_window: | |
772 | * @ep: end point | |
773 | * @nr_pages: number of pages in window | |
774 | * | |
775 | * Allocate and prepare a remote registration window. | |
776 | */ | |
777 | struct reg_range_t *micscif_create_remote_window(struct endpt *ep, int nr_pages) | |
778 | { | |
779 | struct reg_range_t *window; | |
780 | ||
781 | might_sleep(); | |
782 | if (!(window = scif_zalloc(sizeof(struct reg_range_t)))) | |
783 | goto error_ret; | |
784 | ||
785 | window->magic = SCIFEP_MAGIC; | |
786 | window->nr_pages = nr_pages; | |
787 | ||
788 | #if !defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) | |
789 | if (!(window->page_ref_count = scif_zalloc(nr_pages * | |
790 | sizeof(*(window->page_ref_count))))) | |
791 | goto error_window; | |
792 | #endif | |
793 | ||
794 | if (!(window->dma_addr = scif_zalloc(nr_pages * | |
795 | sizeof(*(window->dma_addr))))) | |
796 | goto error_window; | |
797 | ||
798 | if (!(window->num_pages = scif_zalloc(nr_pages * | |
799 | sizeof(*(window->num_pages))))) | |
800 | goto error_window; | |
801 | ||
802 | #ifdef CONFIG_ML1OM | |
803 | if (!(window->phys_addr = scif_zalloc(nr_pages * | |
804 | sizeof(*(window->phys_addr))))) | |
805 | goto error_window; | |
806 | ||
807 | if (!(window->temp_phys_addr = scif_zalloc(nr_pages * | |
808 | sizeof(*(window->temp_phys_addr))))) | |
809 | goto error_window; | |
810 | #endif | |
811 | ||
812 | if (micscif_create_remote_lookup(ep, window)) | |
813 | goto error_window; | |
814 | ||
815 | window->ep = (uint64_t)ep; | |
816 | window->type = RMA_WINDOW_PEER; | |
817 | set_window_ref_count(window, nr_pages); | |
818 | window->get_put_ref_count = 0; | |
819 | window->unreg_state = OP_IDLE; | |
820 | #if !defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) | |
821 | window->gttmap_state = OP_IDLE; | |
822 | init_waitqueue_head(&window->gttmapwq); | |
823 | #endif | |
824 | #ifdef _MIC_SCIF_ | |
825 | micscif_setup_proxy_dma(ep); | |
826 | window->proxy_dma_phys = ep->rma_info.proxy_dma_phys; | |
827 | #endif | |
828 | return window; | |
829 | error_window: | |
830 | micscif_destroy_remote_window(ep, window); | |
831 | error_ret: | |
832 | return NULL; | |
833 | } | |
834 | ||
835 | /** | |
836 | * micscif_destroy_remote_window: | |
837 | * @ep: end point | |
838 | * @window: remote registration window | |
839 | * | |
840 | * Deallocate resources for remote window. | |
841 | */ | |
842 | void micscif_destroy_remote_window(struct endpt *ep, struct reg_range_t *window) | |
843 | { | |
844 | RMA_MAGIC(window); | |
845 | micscif_destroy_remote_lookup(ep, window); | |
846 | if (window->dma_addr) | |
847 | scif_free(window->dma_addr, window->nr_pages * | |
848 | sizeof(*(window->dma_addr))); | |
849 | if (window->num_pages) | |
850 | scif_free(window->num_pages, window->nr_pages * | |
851 | sizeof(*(window->num_pages))); | |
852 | #ifdef CONFIG_ML1OM | |
853 | if (window->phys_addr) | |
854 | scif_free(window->phys_addr, window->nr_pages * | |
855 | sizeof(*(window->phys_addr))); | |
856 | if (window->temp_phys_addr) | |
857 | scif_free(window->temp_phys_addr, window->nr_pages * | |
858 | sizeof(*(window->temp_phys_addr))); | |
859 | #endif | |
860 | ||
861 | #if !defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) | |
862 | if (window->page_ref_count) | |
863 | scif_free(window->page_ref_count, window->nr_pages * | |
864 | sizeof(*(window->page_ref_count))); | |
865 | #endif | |
866 | window->magic = 0; | |
867 | scif_free(window, sizeof(*window)); | |
868 | } | |
869 | ||
870 | /** | |
871 | * micscif_map_window_pages: | |
872 | * @ep: end point | |
873 | * @window: self registration window | |
874 | * @tmp_wnd: is a temporary window? | |
875 | * | |
876 | * Map pages of a window into the aperture/PCI. | |
877 | * Also compute physical addresses required for DMA. | |
878 | */ | |
879 | int micscif_map_window_pages(struct endpt *ep, struct reg_range_t *window, bool tmp_wnd) | |
880 | { | |
881 | int j, i, err = 0, nr_pages; | |
882 | scif_pinned_pages_t pinned_pages; | |
883 | ||
884 | might_sleep(); | |
885 | RMA_MAGIC(window); | |
886 | ||
887 | pinned_pages = window->pinned_pages; | |
888 | for (j = 0, i = 0; j < window->nr_contig_chunks; j++, i += nr_pages) { | |
889 | nr_pages = pinned_pages->num_pages[i]; | |
890 | #ifdef _MIC_SCIF_ | |
891 | #ifdef CONFIG_ML1OM | |
892 | /* phys_addr[] holds addresses as seen from the remote node | |
893 | * these addressed are then copied into the remote card's | |
894 | * window structure | |
895 | * when the remote node is the host and the card is knf | |
896 | * these addresses are only created at the point of mapping | |
897 | * the card physical address into gtt (for the KNC the | |
898 | * the gtt code path returns the local address) | |
899 | * when the remote node is loopback - the address remains | |
900 | * the same | |
901 | * when the remote node is a kn* - the base address of the local | |
902 | * card as seen from the remote node is added in | |
903 | */ | |
904 | if (!tmp_wnd) { | |
905 | if(ep->remote_dev != &scif_dev[SCIF_HOST_NODE]) { | |
906 | if ((err = map_virt_into_aperture( | |
907 | &window->temp_phys_addr[j], | |
908 | phys_to_virt(page_to_phys(pinned_pages->pages[i])), | |
909 | ep->remote_dev, | |
910 | nr_pages << PAGE_SHIFT))) { | |
911 | int k,l; | |
912 | ||
913 | for (l = k = 0; k < i; l++) { | |
914 | nr_pages = pinned_pages->num_pages[k]; | |
915 | window->temp_phys_addr[l] | |
916 | &= ~RMA_HUGE_NR_PAGE_MASK; | |
917 | unmap_from_aperture( | |
918 | window->temp_phys_addr[l], | |
919 | ep->remote_dev, | |
920 | nr_pages << PAGE_SHIFT); | |
921 | k += nr_pages; | |
922 | window->temp_phys_addr[l] = 0; | |
923 | } | |
924 | return err; | |
925 | } | |
926 | if (!tmp_wnd) | |
927 | RMA_SET_NR_PAGES(window->temp_phys_addr[j], nr_pages); | |
928 | } | |
929 | } | |
930 | #endif | |
931 | window->dma_addr[j] = | |
932 | page_to_phys(pinned_pages->pages[i]); | |
933 | if (!tmp_wnd) | |
934 | RMA_SET_NR_PAGES(window->dma_addr[j], nr_pages); | |
935 | #else | |
936 | err = map_virt_into_aperture(&window->dma_addr[j], | |
937 | phys_to_virt(page_to_phys(pinned_pages->pages[i])), | |
938 | ep->remote_dev, nr_pages << PAGE_SHIFT); | |
939 | if (err) | |
940 | return err; | |
941 | if (!tmp_wnd) | |
942 | RMA_SET_NR_PAGES(window->dma_addr[j], nr_pages); | |
943 | #endif | |
944 | window->num_pages[j] = nr_pages; | |
945 | } | |
946 | return err; | |
947 | } | |
948 | ||
949 | ||
950 | /** | |
951 | * micscif_unregister_window: | |
952 | * @window: self registration window | |
953 | * | |
954 | * Send an unregistration request and wait for a response. | |
955 | */ | |
956 | int micscif_unregister_window(struct reg_range_t *window) | |
957 | { | |
958 | int err = 0; | |
959 | struct endpt *ep = (struct endpt *)window->ep; | |
960 | bool send_msg = false; | |
961 | ||
962 | might_sleep(); | |
963 | BUG_ON(!mutex_is_locked(&ep->rma_info.rma_lock)); | |
964 | ||
965 | switch (window->unreg_state) { | |
966 | case OP_IDLE: | |
967 | { | |
968 | window->unreg_state = OP_IN_PROGRESS; | |
969 | send_msg = true; | |
970 | /* fall through */ | |
971 | } | |
972 | case OP_IN_PROGRESS: | |
973 | { | |
974 | get_window_ref_count(window, 1); | |
975 | mutex_unlock(&ep->rma_info.rma_lock); | |
976 | if (send_msg && (err = micscif_send_scif_unregister(ep, window))) { | |
977 | window->unreg_state = OP_COMPLETED; | |
978 | goto done; | |
979 | } | |
980 | retry: | |
981 | err = wait_event_timeout(window->unregwq, | |
982 | window->unreg_state != OP_IN_PROGRESS, NODE_ALIVE_TIMEOUT); | |
983 | if (!err && scifdev_alive(ep)) | |
984 | goto retry; | |
985 | if (!err) { | |
986 | err = -ENODEV; | |
987 | window->unreg_state = OP_COMPLETED; | |
988 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); | |
989 | } | |
990 | if (err > 0) | |
991 | err = 0; | |
992 | done: | |
993 | mutex_lock(&ep->rma_info.rma_lock); | |
994 | put_window_ref_count(window, 1); | |
995 | break; | |
996 | } | |
997 | case OP_FAILED: | |
998 | { | |
999 | if (!scifdev_alive(ep)) { | |
1000 | err = -ENODEV; | |
1001 | window->unreg_state = OP_COMPLETED; | |
1002 | } | |
1003 | break; | |
1004 | } | |
1005 | case OP_COMPLETED: | |
1006 | break; | |
1007 | default: | |
1008 | /* Invalid opcode? */ | |
1009 | BUG_ON(1); | |
1010 | } | |
1011 | ||
1012 | if (OP_COMPLETED == window->unreg_state && | |
1013 | window->ref_count) | |
1014 | put_window_ref_count(window, window->nr_pages); | |
1015 | ||
1016 | if (!window->ref_count) { | |
1017 | atomic_inc(&ep->rma_info.tw_refcount); | |
1018 | atomic_add_return((int32_t)window->nr_pages, &ep->rma_info.tw_total_pages); | |
1019 | list_del(&window->list_member); | |
1020 | micscif_free_window_offset(ep, window->offset, | |
1021 | window->nr_pages << PAGE_SHIFT); | |
1022 | window->offset_freed = true; | |
1023 | mutex_unlock(&ep->rma_info.rma_lock); | |
1024 | if ((!!(window->pinned_pages->map_flags & SCIF_MAP_KERNEL)) | |
1025 | && scifdev_alive(ep)) { | |
1026 | drain_dma_intr(ep->rma_info.dma_chan); | |
1027 | } else { | |
1028 | if (!__scif_dec_pinned_vm_lock(window->mm, | |
1029 | window->nr_pages, 1)) { | |
1030 | __scif_release_mm(window->mm); | |
1031 | window->mm = NULL; | |
1032 | } | |
1033 | } | |
1034 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); | |
1035 | mutex_lock(&ep->rma_info.rma_lock); | |
1036 | } | |
1037 | return err; | |
1038 | } | |
1039 | ||
1040 | /** | |
1041 | * micscif_send_alloc_request: | |
1042 | * @ep: end point | |
1043 | * @window: self registration window | |
1044 | * | |
1045 | * Send a remote window allocation request | |
1046 | */ | |
1047 | int micscif_send_alloc_request(struct endpt *ep, struct reg_range_t *window) | |
1048 | { | |
1049 | struct nodemsg msg; | |
1050 | struct allocmsg *alloc = &window->alloc_handle; | |
1051 | ||
1052 | /* Set up the Alloc Handle */ | |
1053 | alloc->uop = SCIF_REGISTER; | |
1054 | alloc->state = OP_IN_PROGRESS; | |
1055 | init_waitqueue_head(&alloc->allocwq); | |
1056 | ||
1057 | /* Send out an allocation request */ | |
1058 | msg.uop = SCIF_ALLOC_REQ; | |
1059 | msg.src = ep->port; | |
1060 | msg.payload[0] = ep->remote_ep; | |
1061 | msg.payload[1] = window->nr_pages; | |
1062 | msg.payload[2] = (uint64_t)&window->alloc_handle; | |
1063 | msg.payload[3] = SCIF_REGISTER; | |
1064 | return micscif_nodeqp_send(ep->remote_dev, &msg, ep); | |
1065 | } | |
1066 | ||
1067 | /** | |
1068 | * micscif_prep_remote_window: | |
1069 | * @ep: end point | |
1070 | * @window: self registration window | |
1071 | * | |
1072 | * Send a remote window allocation request, wait for an allocation response, | |
1073 | * prepare the remote window and notify the peer to unmap it once done. | |
1074 | */ | |
1075 | int micscif_prep_remote_window(struct endpt *ep, struct reg_range_t *window) | |
1076 | { | |
1077 | struct nodemsg msg; | |
1078 | struct reg_range_t *remote_window; | |
1079 | struct allocmsg *alloc = &window->alloc_handle; | |
1080 | dma_addr_t *dma_phys_lookup, *tmp; | |
1081 | int i = 0, j = 0; | |
1082 | int nr_contig_chunks, loop_nr_contig_chunks, remaining_nr_contig_chunks, nr_lookup; | |
1083 | #if defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) | |
1084 | dma_addr_t *phys_lookup = 0; | |
1085 | #endif | |
1086 | int err, map_err; | |
1087 | ||
1088 | nr_contig_chunks = remaining_nr_contig_chunks = (int)window->nr_contig_chunks; | |
1089 | ||
1090 | if ((map_err = micscif_map_window_pages(ep, window, false))) { | |
1091 | printk(KERN_ERR "%s %d map_err %d\n", __func__, __LINE__, map_err); | |
1092 | } | |
1093 | retry: | |
1094 | /* Now wait for the response */ | |
1095 | err = wait_event_timeout(alloc->allocwq, alloc->state != OP_IN_PROGRESS, NODE_ALIVE_TIMEOUT); | |
1096 | if (!err && scifdev_alive(ep)) | |
1097 | goto retry; | |
1098 | ||
1099 | if (!err) | |
1100 | err = -ENODEV; | |
1101 | ||
1102 | if (err > 0) | |
1103 | err = 0; | |
1104 | else | |
1105 | return err; | |
1106 | ||
1107 | /* Bail out. The remote end rejected this request */ | |
1108 | if (OP_FAILED == alloc->state) | |
1109 | return -ENOMEM; | |
1110 | ||
1111 | if (map_err) { | |
1112 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, map_err); | |
1113 | msg.uop = SCIF_FREE_VIRT; | |
1114 | msg.src = ep->port; | |
1115 | msg.payload[0] = ep->remote_ep; | |
1116 | msg.payload[1] = (uint64_t)window->alloc_handle.vaddr; | |
1117 | msg.payload[2] = (uint64_t)window; | |
1118 | msg.payload[3] = SCIF_REGISTER; | |
1119 | if (!(err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) | |
1120 | err = -ENOTCONN; | |
1121 | else | |
1122 | err = map_err; | |
1123 | return err; | |
1124 | } | |
1125 | ||
1126 | ||
1127 | remote_window = scif_ioremap(alloc->phys_addr, | |
1128 | sizeof(*window), ep->remote_dev); | |
1129 | ||
1130 | RMA_MAGIC(remote_window); | |
1131 | ||
1132 | /* Compute the number of lookup entries. 21 == 2MB Shift */ | |
1133 | nr_lookup = ALIGN(nr_contig_chunks * PAGE_SIZE, ((2) * 1024 * 1024)) >> 21; | |
1134 | #if defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) | |
1135 | if (is_p2p_scifdev(ep->remote_dev)) | |
1136 | phys_lookup = scif_ioremap(remote_window->temp_phys_addr_lookup.offset, | |
1137 | nr_lookup * | |
1138 | sizeof(*remote_window->temp_phys_addr_lookup.lookup), | |
1139 | ep->remote_dev); | |
1140 | #endif | |
1141 | ||
1142 | dma_phys_lookup = scif_ioremap(remote_window->dma_addr_lookup.offset, | |
1143 | nr_lookup * | |
1144 | sizeof(*remote_window->dma_addr_lookup.lookup), | |
1145 | ep->remote_dev); | |
1146 | ||
1147 | while (remaining_nr_contig_chunks) { | |
1148 | loop_nr_contig_chunks = min(remaining_nr_contig_chunks, (int)NR_PHYS_ADDR_IN_PAGE); | |
1149 | /* #1/2 - Copy physical addresses over to the remote side */ | |
1150 | ||
1151 | #if defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) | |
1152 | /* If the remote dev is self or is any node except the host | |
1153 | * its OK to copy the bus address to the remote window | |
1154 | * in the case of the host (for KNF only) the bus address | |
1155 | * is generated at the time of mmap(..) into card memory | |
1156 | * and does not exist at this time | |
1157 | */ | |
1158 | /* Note: | |
1159 | * the phys_addr[] holds MIC address for remote cards | |
1160 | * -> GTT offset for the host (KNF) | |
1161 | * -> local address for the host (KNC) | |
1162 | * -> local address for loopback | |
1163 | * this is done in map_window_pages(..) except for GTT | |
1164 | * offset for KNF | |
1165 | */ | |
1166 | if (is_p2p_scifdev(ep->remote_dev)) { | |
1167 | tmp = scif_ioremap(phys_lookup[j], | |
1168 | loop_nr_contig_chunks * sizeof(*window->temp_phys_addr), | |
1169 | ep->remote_dev); | |
1170 | memcpy_toio(tmp, &window->temp_phys_addr[i], | |
1171 | loop_nr_contig_chunks * sizeof(*window->temp_phys_addr)); | |
1172 | serializing_request(tmp); | |
1173 | smp_mb(); | |
1174 | scif_iounmap(tmp, PAGE_SIZE, ep->remote_dev); | |
1175 | } | |
1176 | #endif | |
1177 | /* #2/2 - Copy DMA addresses (addresses that are fed into the DMA engine) | |
1178 | * We transfer bus addresses which are then converted into a MIC physical | |
1179 | * address on the remote side if it is a MIC, if the remote node is a host | |
1180 | * we transfer the MIC physical address | |
1181 | */ | |
1182 | tmp = scif_ioremap( | |
1183 | dma_phys_lookup[j], | |
1184 | loop_nr_contig_chunks * sizeof(*window->dma_addr), | |
1185 | ep->remote_dev); | |
1186 | #ifdef _MIC_SCIF_ | |
1187 | if (is_p2p_scifdev(ep->remote_dev)) { | |
1188 | /* knf: | |
1189 | * send the address as mapped through the GTT (the remote node's | |
1190 | * base address for this node is already added in) | |
1191 | * knc: | |
1192 | * add remote node's base address for this node to convert it | |
1193 | * into a MIC address | |
1194 | */ | |
1195 | int m; | |
1196 | dma_addr_t dma_addr; | |
1197 | for (m = 0; m < loop_nr_contig_chunks; m++) { | |
1198 | #ifdef CONFIG_ML1OM | |
1199 | dma_addr = window->temp_phys_addr[i + m]; | |
1200 | #else | |
1201 | dma_addr = window->dma_addr[i + m] + | |
1202 | ep->remote_dev->sd_base_addr; | |
1203 | #endif | |
1204 | writeq(dma_addr, &tmp[m]); | |
1205 | } | |
1206 | } else | |
1207 | /* Host node or loopback - transfer DMA addresses as is, this is | |
1208 | * the same as a MIC physical address (we use the dma_addr | |
1209 | * and not the phys_addr array since the phys_addr is only setup | |
1210 | * if there is a mmap() request from the host) | |
1211 | */ | |
1212 | memcpy_toio(tmp, &window->dma_addr[i], | |
1213 | loop_nr_contig_chunks * sizeof(*window->dma_addr)); | |
1214 | #else | |
1215 | /* Transfer the physical address array - this is the MIC address | |
1216 | * as seen by the card | |
1217 | */ | |
1218 | memcpy_toio(tmp, &window->dma_addr[i], | |
1219 | loop_nr_contig_chunks * sizeof(*window->dma_addr)); | |
1220 | #endif | |
1221 | remaining_nr_contig_chunks -= loop_nr_contig_chunks; | |
1222 | i += loop_nr_contig_chunks; | |
1223 | j++; | |
1224 | serializing_request(tmp); | |
1225 | smp_mb(); | |
1226 | scif_iounmap(tmp, PAGE_SIZE, ep->remote_dev); | |
1227 | } | |
1228 | ||
1229 | /* Prepare the remote window for the peer */ | |
1230 | remote_window->peer_window = (uint64_t)window; | |
1231 | remote_window->offset = window->offset; | |
1232 | remote_window->prot = window->prot; | |
1233 | remote_window->nr_contig_chunks = nr_contig_chunks; | |
1234 | #ifdef _MIC_SCIF_ | |
1235 | if (!ep->rma_info.proxy_dma_peer_phys) | |
1236 | ep->rma_info.proxy_dma_peer_phys = remote_window->proxy_dma_phys; | |
1237 | #endif | |
1238 | #if defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) | |
1239 | if (is_p2p_scifdev(ep->remote_dev)) | |
1240 | scif_iounmap(phys_lookup, | |
1241 | nr_lookup * | |
1242 | sizeof(*remote_window->temp_phys_addr_lookup.lookup), | |
1243 | ep->remote_dev); | |
1244 | #endif | |
1245 | scif_iounmap(dma_phys_lookup, | |
1246 | nr_lookup * | |
1247 | sizeof(*remote_window->dma_addr_lookup.lookup), | |
1248 | ep->remote_dev); | |
1249 | scif_iounmap(remote_window, sizeof(*remote_window), ep->remote_dev); | |
1250 | window->peer_window = (uint64_t)alloc->vaddr; | |
1251 | return err; | |
1252 | } | |
1253 | ||
1254 | /** | |
1255 | * micscif_send_scif_register: | |
1256 | * @ep: end point | |
1257 | * @window: self registration window | |
1258 | * | |
1259 | * Send a SCIF_REGISTER message if EP is connected and wait for a | |
1260 | * SCIF_REGISTER_(N)ACK message else send a SCIF_FREE_VIRT | |
1261 | * message so that the peer can free its remote window allocated earlier. | |
1262 | */ | |
1263 | int micscif_send_scif_register(struct endpt *ep, struct reg_range_t *window) | |
1264 | { | |
1265 | int err = 0; | |
1266 | struct nodemsg msg; | |
1267 | ||
1268 | msg.src = ep->port; | |
1269 | msg.payload[0] = ep->remote_ep; | |
1270 | msg.payload[1] = (uint64_t)window->alloc_handle.vaddr; | |
1271 | msg.payload[2] = (uint64_t)window; | |
1272 | if (SCIFEP_CONNECTED == ep->state) { | |
1273 | msg.uop = SCIF_REGISTER; | |
1274 | window->reg_state = OP_IN_PROGRESS; | |
1275 | if (!(err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) { | |
1276 | micscif_set_nr_pages(ep->remote_dev, window); | |
1277 | retry: | |
1278 | err = wait_event_timeout(window->regwq, | |
1279 | window->reg_state != OP_IN_PROGRESS, NODE_ALIVE_TIMEOUT); | |
1280 | if (!err && scifdev_alive(ep)) | |
1281 | goto retry; | |
1282 | if (!err) | |
1283 | err = -ENODEV; | |
1284 | if (err > 0) | |
1285 | err = 0; | |
1286 | if (OP_FAILED == window->reg_state) | |
1287 | err = -ENOTCONN; | |
1288 | } else { | |
1289 | micscif_set_nr_pages(ep->remote_dev, window); | |
1290 | } | |
1291 | } else { | |
1292 | msg.uop = SCIF_FREE_VIRT; | |
1293 | msg.payload[3] = SCIF_REGISTER; | |
1294 | if (!(err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) | |
1295 | err = -ENOTCONN; | |
1296 | micscif_set_nr_pages(ep->remote_dev, window); | |
1297 | } | |
1298 | return err; | |
1299 | } | |
1300 | ||
1301 | /** | |
1302 | * micscif_send_scif_unregister: | |
1303 | * @ep: end point | |
1304 | * @window: self registration window | |
1305 | * | |
1306 | * Send a SCIF_UNREGISTER message. | |
1307 | */ | |
1308 | int micscif_send_scif_unregister(struct endpt *ep, struct reg_range_t *window) | |
1309 | { | |
1310 | struct nodemsg msg; | |
1311 | ||
1312 | RMA_MAGIC(window); | |
1313 | msg.uop = SCIF_UNREGISTER; | |
1314 | msg.src = ep->port; | |
1315 | msg.payload[0] = (uint64_t)window->alloc_handle.vaddr; | |
1316 | msg.payload[1] = (uint64_t)window; | |
1317 | return micscif_nodeqp_send(ep->remote_dev, &msg, ep); | |
1318 | } | |
1319 | ||
1320 | /** | |
1321 | * micscif_get_window_offset: | |
1322 | * @epd: end point descriptor | |
1323 | * @flags: flags | |
1324 | * @offset: offset hint | |
1325 | * @len: length of range | |
1326 | * @out_offset: computed offset returned by reference. | |
1327 | * | |
1328 | * Compute/Claim a new offset for this EP. The callee is supposed to grab | |
1329 | * the RMA mutex before calling this API. | |
1330 | */ | |
1331 | int micscif_get_window_offset(struct endpt *ep, int flags, | |
1332 | uint64_t offset, size_t len, uint64_t *out_offset) | |
1333 | { | |
1334 | uint64_t computed_offset; | |
1335 | int err = 0; | |
1336 | ||
1337 | might_sleep(); | |
1338 | mutex_lock(&ep->rma_info.va_lock); | |
1339 | if (flags & SCIF_MAP_FIXED) { | |
1340 | computed_offset = va_gen_claim(&ep->rma_info.va_gen, | |
1341 | (uint64_t)offset, len); | |
1342 | if (INVALID_VA_GEN_ADDRESS == computed_offset) | |
1343 | err = -EADDRINUSE; | |
1344 | } else { | |
1345 | computed_offset = va_gen_alloc(&ep->rma_info.va_gen, | |
1346 | len, PAGE_SIZE); | |
1347 | if (INVALID_VA_GEN_ADDRESS == computed_offset) | |
1348 | err = -ENOMEM; | |
1349 | } | |
1350 | *out_offset = computed_offset; | |
1351 | mutex_unlock(&ep->rma_info.va_lock); | |
1352 | return err; | |
1353 | } | |
1354 | ||
1355 | /** | |
1356 | * micscif_free_window_offset: | |
1357 | * @offset: offset hint | |
1358 | * @len: length of range | |
1359 | * | |
1360 | * Free offset for this EP. The callee is supposed to grab | |
1361 | * the RMA mutex before calling this API. | |
1362 | */ | |
1363 | void micscif_free_window_offset(struct endpt *ep, | |
1364 | uint64_t offset, size_t len) | |
1365 | { | |
1366 | mutex_lock(&ep->rma_info.va_lock); | |
1367 | va_gen_free(&ep->rma_info.va_gen, offset, len); | |
1368 | mutex_unlock(&ep->rma_info.va_lock); | |
1369 | } | |
1370 | ||
1371 | /** | |
1372 | * scif_register_temp: | |
1373 | * @epd: End Point Descriptor. | |
1374 | * @addr: virtual address to/from which to copy | |
1375 | * @len: length of range to copy | |
1376 | * @out_offset: computed offset returned by reference. | |
1377 | * @out_window: allocated registered window returned by reference. | |
1378 | * | |
1379 | * Create a temporary registered window. The peer will not know about this | |
1380 | * window. This API is used for scif_vreadfrom()/scif_vwriteto() API's. | |
1381 | */ | |
1382 | static int | |
1383 | micscif_register_temp(scif_epd_t epd, void *addr, size_t len, int prot, | |
1384 | off_t *out_offset, struct reg_range_t **out_window) | |
1385 | { | |
1386 | struct endpt *ep = (struct endpt *)epd; | |
1387 | int err; | |
1388 | scif_pinned_pages_t pinned_pages; | |
1389 | size_t aligned_len; | |
1390 | ||
1391 | aligned_len = ALIGN(len, PAGE_SIZE); | |
1392 | ||
1393 | if ((err = __scif_pin_pages((void *)((uint64_t)addr & | |
1394 | PAGE_MASK), | |
1395 | aligned_len, &prot, 0, &pinned_pages))) | |
1396 | return err; | |
1397 | ||
1398 | pinned_pages->prot = prot; | |
1399 | ||
1400 | /* Compute the offset for this registration */ | |
1401 | if ((err = micscif_get_window_offset(ep, 0, 0, | |
1402 | aligned_len, (uint64_t *)out_offset))) | |
1403 | goto error_unpin; | |
1404 | ||
1405 | /* Allocate and prepare self registration window */ | |
1406 | if (!(*out_window = micscif_create_window(ep, aligned_len >> PAGE_SHIFT, | |
1407 | *out_offset, true))) { | |
1408 | micscif_free_window_offset(ep, *out_offset, aligned_len); | |
1409 | err = -ENOMEM; | |
1410 | goto error_unpin; | |
1411 | } | |
1412 | ||
1413 | (*out_window)->pinned_pages = pinned_pages; | |
1414 | (*out_window)->nr_pages = pinned_pages->nr_pages; | |
1415 | (*out_window)->nr_contig_chunks = pinned_pages->nr_contig_chunks; | |
1416 | (*out_window)->prot = pinned_pages->prot; | |
1417 | ||
1418 | (*out_window)->va_for_temp = (void*)((uint64_t)addr & PAGE_MASK); | |
1419 | if ((err = micscif_map_window_pages(ep, *out_window, true))) { | |
1420 | /* Something went wrong! Rollback */ | |
1421 | micscif_destroy_window(ep, *out_window); | |
1422 | *out_window = NULL; | |
1423 | } else | |
1424 | *out_offset |= ((uint64_t)addr & ~PAGE_MASK); | |
1425 | ||
1426 | return err; | |
1427 | error_unpin: | |
1428 | if (err) | |
1429 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); | |
1430 | scif_unpin_pages(pinned_pages); | |
1431 | return err; | |
1432 | } | |
1433 | ||
1434 | /** | |
1435 | * micscif_rma_completion_cb: | |
1436 | * @data: RMA cookie | |
1437 | * | |
1438 | * RMA interrupt completion callback. | |
1439 | */ | |
1440 | void micscif_rma_completion_cb(uint64_t data) | |
1441 | { | |
1442 | struct dma_completion_cb *comp_cb = (struct dma_completion_cb *)data; | |
1443 | #ifndef _MIC_SCIF_ | |
1444 | struct pci_dev *pdev; | |
1445 | #endif | |
1446 | ||
1447 | /* Free DMA Completion CB. */ | |
1448 | if (comp_cb && comp_cb->temp_buf) { | |
1449 | if (comp_cb->dst_window) { | |
1450 | micscif_rma_local_cpu_copy(comp_cb->dst_offset, | |
1451 | comp_cb->dst_window, comp_cb->temp_buf + comp_cb->header_padding, | |
1452 | comp_cb->len, false); | |
1453 | } | |
1454 | #ifndef _MIC_SCIF_ | |
1455 | micscif_pci_dev(comp_cb->remote_node, &pdev); | |
1456 | mic_ctx_unmap_single(get_per_dev_ctx(comp_cb->remote_node - 1), | |
1457 | comp_cb->temp_phys, KMEM_UNALIGNED_BUF_SIZE); | |
1458 | #endif | |
1459 | if (comp_cb->is_cache) | |
1460 | micscif_kmem_cache_free(comp_cb->temp_buf_to_free); | |
1461 | else | |
1462 | kfree(comp_cb->temp_buf_to_free); | |
1463 | } | |
1464 | kfree(comp_cb); | |
1465 | } | |
1466 | ||
1467 | static void __micscif_rma_destroy_tcw_ep(struct endpt *ep); | |
1468 | static | |
1469 | bool micscif_rma_tc_can_cache(struct endpt *ep, size_t cur_bytes) | |
1470 | { | |
1471 | if ((cur_bytes >> PAGE_SHIFT) > ms_info.mi_rma_tc_limit) | |
1472 | return false; | |
1473 | if ((atomic_read(&ep->rma_info.tcw_total_pages) | |
1474 | + (cur_bytes >> PAGE_SHIFT)) > | |
1475 | ms_info.mi_rma_tc_limit) { | |
1476 | printk(KERN_ALERT "%s %d total=%d, current=%zu reached max\n", | |
1477 | __func__, __LINE__, | |
1478 | atomic_read(&ep->rma_info.tcw_total_pages), | |
1479 | (1 + (cur_bytes >> PAGE_SHIFT))); | |
1480 | micscif_rma_destroy_tcw_invalid(&ms_info.mi_rma_tc); | |
1481 | __micscif_rma_destroy_tcw_ep(ep); | |
1482 | } | |
1483 | return true; | |
1484 | } | |
1485 | ||
1486 | /** | |
1487 | * micscif_rma_copy: | |
1488 | * @epd: end point descriptor. | |
1489 | * @loffset: offset in local registered address space to/from which to copy | |
1490 | * @addr: user virtual address to/from which to copy | |
1491 | * @len: length of range to copy | |
1492 | * @roffset: offset in remote registered address space to/from which to copy | |
1493 | * @flags: flags | |
1494 | * @dir: LOCAL->REMOTE or vice versa. | |
1495 | * | |
1496 | * Validate parameters, check if src/dst registered ranges requested for copy | |
1497 | * are valid and initiate either CPU or DMA copy. | |
1498 | */ | |
1499 | int micscif_rma_copy(scif_epd_t epd, off_t loffset, void *addr, size_t len, | |
1500 | off_t roffset, int flags, enum rma_direction dir, bool last_chunk) | |
1501 | { | |
1502 | struct endpt *ep = (struct endpt *)epd; | |
1503 | struct micscif_rma_req remote_req; | |
1504 | struct micscif_rma_req req; | |
1505 | struct reg_range_t *window = NULL; | |
1506 | struct reg_range_t *remote_window = NULL; | |
1507 | struct mic_copy_work copy_work; | |
1508 | bool loopback; | |
1509 | int err = 0; | |
1510 | struct dma_channel *chan; | |
1511 | struct rma_mmu_notifier *mmn = NULL; | |
1512 | bool insert_window = false; | |
1513 | bool cache = false; | |
1514 | ||
1515 | if ((err = verify_epd(ep))) | |
1516 | return err; | |
1517 | ||
1518 | if (flags && !(flags & (SCIF_RMA_USECPU | SCIF_RMA_USECACHE | SCIF_RMA_SYNC | SCIF_RMA_ORDERED))) | |
1519 | return -EINVAL; | |
1520 | ||
1521 | if (!len) | |
1522 | return -EINVAL; | |
1523 | loopback = is_self_scifdev(ep->remote_dev) ? true : false; | |
1524 | copy_work.fence_type = ((flags & SCIF_RMA_SYNC) && last_chunk) ? DO_DMA_POLLING : 0; | |
1525 | copy_work.ordered = !!((flags & SCIF_RMA_ORDERED) && last_chunk); | |
1526 | ||
1527 | #ifdef CONFIG_MMU_NOTIFIER | |
1528 | if (!mic_reg_cache_enable) | |
1529 | flags &= ~SCIF_RMA_USECACHE; | |
1530 | #else | |
1531 | flags &= ~SCIF_RMA_USECACHE; | |
1532 | #endif | |
1533 | #ifndef _MIC_SCIF_ | |
1534 | #ifdef CONFIG_ML1OM | |
1535 | /* Use DMA Copies even if CPU copy is requested on KNF MIC from Host */ | |
1536 | if (flags & SCIF_RMA_USECPU) { | |
1537 | flags &= ~SCIF_RMA_USECPU; | |
1538 | if (last_chunk) | |
1539 | copy_work.fence_type = DO_DMA_POLLING; | |
1540 | } | |
1541 | #endif | |
1542 | /* Use CPU for Host<->Host Copies */ | |
1543 | if (loopback) { | |
1544 | flags |= SCIF_RMA_USECPU; | |
1545 | copy_work.fence_type = 0x0; | |
1546 | } | |
1547 | #endif | |
1548 | ||
1549 | cache = flags & SCIF_RMA_USECACHE; | |
1550 | ||
1551 | /* Trying to wrap around */ | |
1552 | if ((loffset && (loffset + (off_t)len < loffset)) || | |
1553 | (roffset + (off_t)len < roffset)) | |
1554 | return -EINVAL; | |
1555 | ||
1556 | remote_req.out_window = &remote_window; | |
1557 | remote_req.offset = roffset; | |
1558 | remote_req.nr_bytes = len; | |
1559 | /* | |
1560 | * If transfer is from local to remote then the remote window | |
1561 | * must be writeable and vice versa. | |
1562 | */ | |
1563 | remote_req.prot = LOCAL_TO_REMOTE == dir ? VM_WRITE : VM_READ; | |
1564 | remote_req.type = WINDOW_PARTIAL; | |
1565 | remote_req.head = &ep->rma_info.remote_reg_list; | |
1566 | ||
1567 | #ifdef CONFIG_MMU_NOTIFIER | |
1568 | if (addr && cache) { | |
1569 | mutex_lock(&ep->rma_info.mmn_lock); | |
1570 | mmn = find_mmu_notifier(current->mm, &ep->rma_info); | |
1571 | if (!mmn) { | |
1572 | mmn = kzalloc(sizeof(*mmn), GFP_KERNEL); | |
1573 | if (!mmn) { | |
1574 | mutex_unlock(&ep->rma_info.mmn_lock); | |
1575 | return -ENOMEM; | |
1576 | } | |
1577 | init_mmu_notifier(mmn, current->mm, ep); | |
1578 | if (mmu_notifier_register(&mmn->ep_mmu_notifier, current->mm)) { | |
1579 | mutex_unlock(&ep->rma_info.mmn_lock); | |
1580 | kfree(mmn); | |
1581 | return -EBUSY; | |
1582 | } | |
1583 | #ifdef RMA_DEBUG | |
1584 | atomic_long_add_return(1, &ms_info.mmu_notif_cnt); | |
1585 | #endif | |
1586 | list_add(&mmn->list_member, &ep->rma_info.mmn_list); | |
1587 | } | |
1588 | mutex_unlock(&ep->rma_info.mmn_lock); | |
1589 | } | |
1590 | #endif | |
1591 | ||
1592 | micscif_inc_node_refcnt(ep->remote_dev, 1); | |
1593 | #ifdef _MIC_SCIF_ | |
1594 | if (!(flags & SCIF_RMA_USECPU)) { | |
1595 | /* | |
1596 | * Proxy the DMA only for P2P reads with transfer size | |
1597 | * greater than proxy DMA threshold. scif_vreadfrom(..) | |
1598 | * and scif_vwriteto(..) is not supported since the peer | |
1599 | * does not have the page lists required to perform the | |
1600 | * proxy DMA. | |
1601 | */ | |
1602 | if (ep->remote_dev->sd_proxy_dma_reads && | |
1603 | !addr && dir == REMOTE_TO_LOCAL && | |
1604 | ep->rma_info.proxy_dma_va && | |
1605 | len >= ms_info.mi_proxy_dma_threshold) { | |
1606 | copy_work.len = len; | |
1607 | copy_work.src_offset = roffset; | |
1608 | copy_work.dst_offset = loffset; | |
1609 | /* Fall through if there were errors */ | |
1610 | if (!(err = micscif_proxy_dma(epd, ©_work))) | |
1611 | goto error; | |
1612 | } | |
1613 | } | |
1614 | #endif | |
1615 | mutex_lock(&ep->rma_info.rma_lock); | |
1616 | if (addr) { | |
1617 | req.out_window = &window; | |
1618 | req.nr_bytes = ALIGN(len + ((uint64_t)addr & ~PAGE_MASK), PAGE_SIZE); | |
1619 | if (mmn) | |
1620 | req.head = &mmn->tc_reg_list; | |
1621 | req.va_for_temp = (void*)((uint64_t)addr & PAGE_MASK); | |
1622 | req.prot = (LOCAL_TO_REMOTE == dir ? VM_READ : VM_WRITE | VM_READ); | |
1623 | /* Does a valid local window exist? */ | |
1624 | ||
1625 | pr_debug("%s %d req.va_for_temp %p addr %p req.nr_bytes 0x%lx len 0x%lx\n", | |
1626 | __func__, __LINE__, req.va_for_temp, addr, req.nr_bytes, len); | |
1627 | spin_lock(&ep->rma_info.tc_lock); | |
1628 | if (!mmn || (err = micscif_query_tcw(ep, &req))) { | |
1629 | pr_debug("%s %d err %d req.va_for_temp %p addr %p req.nr_bytes 0x%lx len 0x%lx\n", | |
1630 | __func__, __LINE__, err, req.va_for_temp, addr, req.nr_bytes, len); | |
1631 | spin_unlock(&ep->rma_info.tc_lock); | |
1632 | mutex_unlock(&ep->rma_info.rma_lock); | |
1633 | if (cache) | |
1634 | if (!micscif_rma_tc_can_cache(ep, req.nr_bytes)) | |
1635 | cache = false; | |
1636 | if ((err = micscif_register_temp(epd, req.va_for_temp, req.nr_bytes, | |
1637 | req.prot, | |
1638 | &loffset, &window))) { | |
1639 | goto error; | |
1640 | } | |
1641 | mutex_lock(&ep->rma_info.rma_lock); | |
1642 | pr_debug("New temp window created addr %p\n", addr); | |
1643 | if (cache) { | |
1644 | atomic_inc(&ep->rma_info.tcw_refcount); | |
1645 | atomic_add_return((int32_t)window->nr_pages, &ep->rma_info.tcw_total_pages); | |
1646 | if (mmn) { | |
1647 | spin_lock(&ep->rma_info.tc_lock); | |
1648 | micscif_insert_tcw(window, &mmn->tc_reg_list); | |
1649 | spin_unlock(&ep->rma_info.tc_lock); | |
1650 | } | |
1651 | } | |
1652 | insert_window = true; | |
1653 | } else { | |
1654 | spin_unlock(&ep->rma_info.tc_lock); | |
1655 | pr_debug("window found for addr %p\n", addr); | |
1656 | BUG_ON(window->va_for_temp > addr); | |
1657 | } | |
1658 | loffset = window->offset + ((uint64_t)addr - (uint64_t)window->va_for_temp); | |
1659 | pr_debug("%s %d addr %p loffset 0x%lx window->nr_pages 0x%llx" | |
1660 | " window->va_for_temp %p\n", __func__, __LINE__, | |
1661 | addr, loffset, window->nr_pages, window->va_for_temp); | |
1662 | RMA_MAGIC(window); | |
1663 | } | |
1664 | ||
1665 | /* Does a valid remote window exist? */ | |
1666 | if ((err = micscif_query_window(&remote_req))) { | |
1667 | pr_debug("%s %d err %d roffset 0x%lx len 0x%lx\n", | |
1668 | __func__, __LINE__, err, roffset, len); | |
1669 | mutex_unlock(&ep->rma_info.rma_lock); | |
1670 | goto error; | |
1671 | } | |
1672 | RMA_MAGIC(remote_window); | |
1673 | if (!addr) { | |
1674 | req.out_window = &window; | |
1675 | req.offset = loffset; | |
1676 | /* | |
1677 | * If transfer is from local to remote then the self window | |
1678 | * must be readable and vice versa. | |
1679 | */ | |
1680 | req.prot = LOCAL_TO_REMOTE == dir ? VM_READ : VM_WRITE; | |
1681 | req.nr_bytes = len; | |
1682 | req.type = WINDOW_PARTIAL; | |
1683 | req.head = &ep->rma_info.reg_list; | |
1684 | /* Does a valid local window exist? */ | |
1685 | if ((err = micscif_query_window(&req))) { | |
1686 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); | |
1687 | mutex_unlock(&ep->rma_info.rma_lock); | |
1688 | goto error; | |
1689 | } | |
1690 | RMA_MAGIC(window); | |
1691 | } | |
1692 | ||
1693 | /* | |
1694 | * Preprare copy_work for submitting work to the DMA kernel thread | |
1695 | * or CPU copy routine. | |
1696 | */ | |
1697 | copy_work.len = len; | |
1698 | copy_work.loopback = loopback; | |
1699 | copy_work.remote_dev = ep->remote_dev; | |
1700 | copy_work.dma_chan_released = false; | |
1701 | if (LOCAL_TO_REMOTE == dir) { | |
1702 | copy_work.src_offset = loffset; | |
1703 | copy_work.src_window = window; | |
1704 | copy_work.dst_offset = roffset; | |
1705 | copy_work.dst_window = remote_window; | |
1706 | } else { | |
1707 | copy_work.src_offset = roffset; | |
1708 | copy_work.src_window = remote_window; | |
1709 | copy_work.dst_offset = loffset; | |
1710 | copy_work.dst_window = window; | |
1711 | } | |
1712 | ||
1713 | if (!(flags & SCIF_RMA_USECPU)) { | |
1714 | chan = ep->rma_info.dma_chan; | |
1715 | if ((err = request_dma_channel(chan))) { | |
1716 | mutex_unlock(&ep->rma_info.rma_lock); | |
1717 | goto error; | |
1718 | } | |
1719 | err = micscif_rma_list_dma_copy_wrapper(epd, ©_work, | |
1720 | chan, loffset); | |
1721 | if (!copy_work.dma_chan_released) | |
1722 | free_dma_channel(chan); | |
1723 | } | |
1724 | if (flags & SCIF_RMA_USECPU) { | |
1725 | /* Initiate synchronous CPU copy */ | |
1726 | micscif_rma_list_cpu_copy(©_work); | |
1727 | } | |
1728 | if (insert_window && !cache) { | |
1729 | atomic_inc(&ep->rma_info.tw_refcount); | |
1730 | atomic_add_return((int32_t)window->nr_pages, &ep->rma_info.tw_total_pages); | |
1731 | } | |
1732 | ||
1733 | mutex_unlock(&ep->rma_info.rma_lock); | |
1734 | ||
1735 | if (last_chunk) { | |
1736 | if (DO_DMA_POLLING == copy_work.fence_type) | |
1737 | err = drain_dma_poll(ep->rma_info.dma_chan); | |
1738 | else if (DO_DMA_INTR == copy_work.fence_type) | |
1739 | err = drain_dma_intr(ep->rma_info.dma_chan); | |
1740 | } | |
1741 | ||
1742 | micscif_dec_node_refcnt(ep->remote_dev, 1); | |
1743 | if (insert_window && !cache) | |
1744 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); | |
1745 | return err; | |
1746 | error: | |
1747 | if (err) { | |
1748 | if (addr && window && !cache) | |
1749 | micscif_destroy_window(ep, window); | |
1750 | printk(KERN_ERR "%s %d err %d len 0x%lx\n", __func__, __LINE__, err, len); | |
1751 | } | |
1752 | micscif_dec_node_refcnt(ep->remote_dev, 1); | |
1753 | return err; | |
1754 | } | |
1755 | ||
1756 | /** | |
1757 | * micscif_send_fence_mark: | |
1758 | * @epd: end point descriptor. | |
1759 | * @out_mark: Output DMA mark reported by peer. | |
1760 | * | |
1761 | * Send a remote fence mark request. | |
1762 | */ | |
1763 | int micscif_send_fence_mark(scif_epd_t epd, int *out_mark) | |
1764 | { | |
1765 | int err; | |
1766 | struct nodemsg msg; | |
1767 | struct fence_info *fence_req; | |
1768 | struct endpt *ep = (struct endpt *)epd; | |
1769 | ||
1770 | if (!(fence_req = kmalloc(sizeof(*fence_req), GFP_KERNEL))) { | |
1771 | err = -ENOMEM; | |
1772 | goto error; | |
1773 | } | |
1774 | ||
1775 | fence_req->state = OP_IN_PROGRESS; | |
1776 | init_waitqueue_head(&fence_req->wq); | |
1777 | ||
1778 | msg.src = ep->port; | |
1779 | msg.uop = SCIF_MARK; | |
1780 | msg.payload[0] = ep->remote_ep; | |
1781 | msg.payload[1] = (uint64_t)fence_req; | |
1782 | ||
1783 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) | |
1784 | goto error; | |
1785 | ||
1786 | retry: | |
1787 | err = wait_event_timeout(fence_req->wq, | |
1788 | (OP_IN_PROGRESS != fence_req->state), NODE_ALIVE_TIMEOUT); | |
1789 | if (!err && scifdev_alive(ep)) | |
1790 | goto retry; | |
1791 | if (!err) | |
1792 | err = -ENODEV; | |
1793 | if (err > 0) | |
1794 | err = 0; | |
1795 | if (err < 0) { | |
1796 | mutex_lock(&ep->rma_info.rma_lock); | |
1797 | if (OP_IN_PROGRESS == fence_req->state) | |
1798 | fence_req->state = OP_FAILED; | |
1799 | mutex_unlock(&ep->rma_info.rma_lock); | |
1800 | } | |
1801 | if (OP_COMPLETED == fence_req->state) | |
1802 | *out_mark = SCIF_REMOTE_FENCE | fence_req->dma_mark; | |
1803 | ||
1804 | if (OP_FAILED == fence_req->state && !err) | |
1805 | err = -ENOMEM; | |
1806 | mutex_lock(&ep->rma_info.rma_lock); | |
1807 | mutex_unlock(&ep->rma_info.rma_lock); | |
1808 | kfree(fence_req); | |
1809 | error: | |
1810 | return err; | |
1811 | } | |
1812 | ||
1813 | /** | |
1814 | * micscif_send_fence_wait: | |
1815 | * @epd: end point descriptor. | |
1816 | * @mark: DMA mark to wait for. | |
1817 | * | |
1818 | * Send a remote fence wait request. | |
1819 | */ | |
1820 | int micscif_send_fence_wait(scif_epd_t epd, int mark) | |
1821 | { | |
1822 | int err; | |
1823 | struct nodemsg msg; | |
1824 | struct fence_info *fence_req; | |
1825 | struct endpt *ep = (struct endpt *)epd; | |
1826 | ||
1827 | if (!(fence_req = kmalloc(sizeof(*fence_req), GFP_KERNEL))) { | |
1828 | err = -ENOMEM; | |
1829 | goto error; | |
1830 | } | |
1831 | ||
1832 | fence_req->state = OP_IN_PROGRESS; | |
1833 | init_waitqueue_head(&fence_req->wq); | |
1834 | ||
1835 | msg.src = ep->port; | |
1836 | msg.uop = SCIF_WAIT; | |
1837 | msg.payload[0] = ep->remote_ep; | |
1838 | msg.payload[1] = (uint64_t)fence_req; | |
1839 | msg.payload[2] = mark; | |
1840 | ||
1841 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) | |
1842 | goto error; | |
1843 | retry: | |
1844 | err = wait_event_timeout(fence_req->wq, | |
1845 | (OP_IN_PROGRESS != fence_req->state), NODE_ALIVE_TIMEOUT); | |
1846 | if (!err && scifdev_alive(ep)) | |
1847 | goto retry; | |
1848 | if (!err) | |
1849 | err = -ENODEV; | |
1850 | if (err > 0) | |
1851 | err = 0; | |
1852 | if (err < 0) { | |
1853 | mutex_lock(&ep->rma_info.rma_lock); | |
1854 | if (OP_IN_PROGRESS == fence_req->state) | |
1855 | fence_req->state = OP_FAILED; | |
1856 | mutex_unlock(&ep->rma_info.rma_lock); | |
1857 | } | |
1858 | if (OP_FAILED == fence_req->state && !err) | |
1859 | err = -ENOMEM; | |
1860 | mutex_lock(&ep->rma_info.rma_lock); | |
1861 | mutex_unlock(&ep->rma_info.rma_lock); | |
1862 | kfree(fence_req); | |
1863 | error: | |
1864 | return err; | |
1865 | } | |
1866 | ||
1867 | /** | |
1868 | * micscif_send_fence_signal: | |
1869 | * @epd - endpoint descriptor | |
1870 | * @loff - local offset | |
1871 | * @lval - local value to write to loffset | |
1872 | * @roff - remote offset | |
1873 | * @rval - remote value to write to roffset | |
1874 | * @flags - flags | |
1875 | * | |
1876 | * Sends a remote fence signal request | |
1877 | */ | |
1878 | int micscif_send_fence_signal(scif_epd_t epd, off_t roff, uint64_t rval, | |
1879 | off_t loff, uint64_t lval, int flags) | |
1880 | { | |
1881 | int err = 0; | |
1882 | struct nodemsg msg; | |
1883 | struct fence_info *fence_req; | |
1884 | struct endpt *ep = (struct endpt *)epd; | |
1885 | ||
1886 | if (!(fence_req = kmalloc(sizeof(*fence_req), GFP_KERNEL))) { | |
1887 | err = -ENOMEM; | |
1888 | goto error; | |
1889 | } | |
1890 | ||
1891 | fence_req->state = OP_IN_PROGRESS; | |
1892 | init_waitqueue_head(&fence_req->wq); | |
1893 | ||
1894 | msg.src = ep->port; | |
1895 | if (flags & SCIF_SIGNAL_LOCAL) { | |
1896 | msg.uop = SCIF_SIG_LOCAL; | |
1897 | msg.payload[0] = ep->remote_ep; | |
1898 | msg.payload[1] = roff; | |
1899 | msg.payload[2] = rval; | |
1900 | msg.payload[3] = (uint64_t)fence_req; | |
1901 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) | |
1902 | goto error_free; | |
1903 | retry1: | |
1904 | err = wait_event_timeout(fence_req->wq, | |
1905 | (OP_IN_PROGRESS != fence_req->state), NODE_ALIVE_TIMEOUT); | |
1906 | if (!err && scifdev_alive(ep)) | |
1907 | goto retry1; | |
1908 | if (!err) | |
1909 | err = -ENODEV; | |
1910 | if (err > 0) | |
1911 | err = 0; | |
1912 | if (err < 0) { | |
1913 | mutex_lock(&ep->rma_info.rma_lock); | |
1914 | if (OP_IN_PROGRESS == fence_req->state) | |
1915 | fence_req->state = OP_FAILED; | |
1916 | mutex_unlock(&ep->rma_info.rma_lock); | |
1917 | } | |
1918 | if (OP_FAILED == fence_req->state && !err) { | |
1919 | err = -ENXIO; | |
1920 | goto error_free; | |
1921 | } | |
1922 | } | |
1923 | fence_req->state = OP_IN_PROGRESS; | |
1924 | ||
1925 | if (flags & SCIF_SIGNAL_REMOTE) { | |
1926 | msg.uop = SCIF_SIG_REMOTE; | |
1927 | msg.payload[0] = ep->remote_ep; | |
1928 | msg.payload[1] = loff; | |
1929 | msg.payload[2] = lval; | |
1930 | msg.payload[3] = (uint64_t)fence_req; | |
1931 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) | |
1932 | goto error_free; | |
1933 | retry2: | |
1934 | err = wait_event_timeout(fence_req->wq, | |
1935 | (OP_IN_PROGRESS != fence_req->state), NODE_ALIVE_TIMEOUT); | |
1936 | if (!err && scifdev_alive(ep)) | |
1937 | goto retry2; | |
1938 | if (!err) | |
1939 | err = -ENODEV; | |
1940 | if (err > 0) | |
1941 | err = 0; | |
1942 | if (err < 0) { | |
1943 | mutex_lock(&ep->rma_info.rma_lock); | |
1944 | if (OP_IN_PROGRESS == fence_req->state) | |
1945 | fence_req->state = OP_FAILED; | |
1946 | mutex_unlock(&ep->rma_info.rma_lock); | |
1947 | } | |
1948 | if (OP_FAILED == fence_req->state && !err) { | |
1949 | err = -ENXIO; | |
1950 | goto error_free; | |
1951 | } | |
1952 | } | |
1953 | error_free: | |
1954 | mutex_lock(&ep->rma_info.rma_lock); | |
1955 | mutex_unlock(&ep->rma_info.rma_lock); | |
1956 | kfree(fence_req); | |
1957 | error: | |
1958 | return err; | |
1959 | } | |
1960 | ||
1961 | /* | |
1962 | * micscif_fence_mark: | |
1963 | * | |
1964 | * @epd - endpoint descriptor | |
1965 | * Set up a mark for this endpoint and return the value of the mark. | |
1966 | */ | |
1967 | int micscif_fence_mark(scif_epd_t epd) | |
1968 | { | |
1969 | int mark = 0; | |
1970 | struct endpt *ep = (struct endpt *)epd; | |
1971 | struct dma_channel *chan = ep->rma_info.dma_chan; | |
1972 | ||
1973 | if ((mark = request_dma_channel(chan))) | |
1974 | goto error; | |
1975 | ||
1976 | mark = program_dma_mark(chan); | |
1977 | ||
1978 | free_dma_channel(chan); | |
1979 | error: | |
1980 | return mark; | |
1981 | } | |
1982 | ||
1983 | /** | |
1984 | * micscif_rma_destroy_temp_windows: | |
1985 | * | |
1986 | * This routine destroys temporary registered windows created | |
1987 | * by scif_vreadfrom() and scif_vwriteto(). | |
1988 | */ | |
1989 | void micscif_rma_destroy_temp_windows(void) | |
1990 | { | |
1991 | struct list_head *item, *tmp; | |
1992 | struct reg_range_t *window; | |
1993 | struct endpt *ep; | |
1994 | struct dma_channel *chan; | |
1995 | might_sleep(); | |
1996 | restart: | |
1997 | spin_lock(&ms_info.mi_rmalock); | |
1998 | list_for_each_safe(item, tmp, &ms_info.mi_rma) { | |
1999 | window = list_entry(item, | |
2000 | struct reg_range_t, list_member); | |
2001 | ep = (struct endpt *)window->ep; | |
2002 | chan = ep->rma_info.dma_chan; | |
2003 | ||
2004 | list_del(&window->list_member); | |
2005 | spin_unlock(&ms_info.mi_rmalock); | |
2006 | micscif_inc_node_refcnt(ep->remote_dev, 1); | |
2007 | if (!chan || | |
2008 | !scifdev_alive(ep) || | |
2009 | (!is_current_dma_mark(chan, window->dma_mark) && | |
2010 | is_dma_mark_processed(chan, window->dma_mark)) || | |
2011 | !drain_dma_intr(chan)) { | |
2012 | micscif_dec_node_refcnt(ep->remote_dev, 1); | |
2013 | /* Remove window from global list */ | |
2014 | window->unreg_state = OP_COMPLETED; | |
2015 | } else { | |
2016 | micscif_dec_node_refcnt(ep->remote_dev, 1); | |
2017 | /* DMA engine hung ?? */ | |
2018 | printk(KERN_ERR "%s %d DMA channel %d hung ep->state %d " | |
2019 | "window->dma_mark 0x%x channel_mark 0x%x\n", | |
2020 | __func__, __LINE__, get_chan_num(chan), | |
2021 | ep->sd_state, window->dma_mark, get_dma_mark(chan)); | |
2022 | WARN_ON(1); | |
2023 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); | |
2024 | goto restart; | |
2025 | } | |
2026 | ||
2027 | if (OP_COMPLETED == window->unreg_state) { | |
2028 | BUG_ON(atomic_sub_return((int32_t)window->nr_pages, | |
2029 | &ep->rma_info.tw_total_pages) < 0); | |
2030 | if (RMA_WINDOW_SELF == window->type) | |
2031 | micscif_destroy_window(ep, window); | |
2032 | else | |
2033 | micscif_destroy_remote_window(ep, window); | |
2034 | BUG_ON(atomic_dec_return( | |
2035 | &ep->rma_info.tw_refcount) < 0); | |
2036 | } | |
2037 | goto restart; | |
2038 | } | |
2039 | spin_unlock(&ms_info.mi_rmalock); | |
2040 | } | |
2041 | ||
2042 | /** | |
2043 | * micscif_rma_destroy_tcw: | |
2044 | * | |
2045 | * This routine destroys temporary registered windows created | |
2046 | * by scif_vreadfrom() and scif_vwriteto(). | |
2047 | */ | |
2048 | static | |
2049 | void __micscif_rma_destroy_tcw(struct rma_mmu_notifier *mmn, | |
2050 | struct endpt *ep, bool inrange, | |
2051 | uint64_t start, uint64_t len) | |
2052 | { | |
2053 | struct list_head *item, *tmp; | |
2054 | struct reg_range_t *window; | |
2055 | uint64_t start_va, end_va; | |
2056 | uint64_t end = start + len; | |
2057 | list_for_each_safe(item, tmp, &mmn->tc_reg_list) { | |
2058 | window = list_entry(item, | |
2059 | struct reg_range_t, list_member); | |
2060 | ep = (struct endpt *)window->ep; | |
2061 | if (inrange) { | |
2062 | if (0 == len) | |
2063 | break; | |
2064 | start_va = (uint64_t)window->va_for_temp; | |
2065 | end_va = start_va+ (window->nr_pages << PAGE_SHIFT); | |
2066 | if (start < start_va) { | |
2067 | if (end <= start_va) { | |
2068 | break; | |
2069 | } else { | |
2070 | } | |
2071 | ||
2072 | } else { | |
2073 | if (start >= end_va) { | |
2074 | continue; | |
2075 | } else { | |
2076 | } | |
2077 | } | |
2078 | } | |
2079 | __micscif_rma_destroy_tcw_helper(window); | |
2080 | } | |
2081 | } | |
2082 | ||
2083 | static inline | |
2084 | void micscif_rma_destroy_tcw(struct rma_mmu_notifier *mmn, | |
2085 | struct endpt *ep, bool inrange, | |
2086 | uint64_t start, uint64_t len) | |
2087 | { | |
2088 | unsigned long sflags; | |
2089 | ||
2090 | spin_lock_irqsave(&ep->rma_info.tc_lock, sflags); | |
2091 | __micscif_rma_destroy_tcw(mmn, ep, inrange, start, len); | |
2092 | spin_unlock_irqrestore(&ep->rma_info.tc_lock, sflags); | |
2093 | } | |
2094 | ||
2095 | static void __micscif_rma_destroy_tcw_ep(struct endpt *ep) | |
2096 | { | |
2097 | struct list_head *item, *tmp; | |
2098 | struct rma_mmu_notifier *mmn; | |
2099 | spin_lock(&ep->rma_info.tc_lock); | |
2100 | list_for_each_safe(item, tmp, &ep->rma_info.mmn_list) { | |
2101 | mmn = list_entry(item, | |
2102 | struct rma_mmu_notifier, list_member); | |
2103 | __micscif_rma_destroy_tcw(mmn, ep, false, 0, 0); | |
2104 | } | |
2105 | spin_unlock(&ep->rma_info.tc_lock); | |
2106 | } | |
2107 | ||
2108 | void micscif_rma_destroy_tcw_ep(struct endpt *ep) | |
2109 | { | |
2110 | struct list_head *item, *tmp; | |
2111 | struct rma_mmu_notifier *mmn; | |
2112 | list_for_each_safe(item, tmp, &ep->rma_info.mmn_list) { | |
2113 | mmn = list_entry(item, | |
2114 | struct rma_mmu_notifier, list_member); | |
2115 | micscif_rma_destroy_tcw(mmn, ep, false, 0, 0); | |
2116 | } | |
2117 | } | |
2118 | ||
2119 | /** | |
2120 | * micscif_rma_destroy_tcw: | |
2121 | * | |
2122 | * This routine destroys temporary registered windows created | |
2123 | * by scif_vreadfrom() and scif_vwriteto(). | |
2124 | */ | |
2125 | void micscif_rma_destroy_tcw_invalid(struct list_head *list) | |
2126 | { | |
2127 | struct list_head *item, *tmp; | |
2128 | struct reg_range_t *window; | |
2129 | struct endpt *ep; | |
2130 | struct dma_channel *chan; | |
2131 | might_sleep(); | |
2132 | restart: | |
2133 | spin_lock(&ms_info.mi_rmalock); | |
2134 | list_for_each_safe(item, tmp, list) { | |
2135 | window = list_entry(item, | |
2136 | struct reg_range_t, list_member); | |
2137 | ep = (struct endpt *)window->ep; | |
2138 | chan = ep->rma_info.dma_chan; | |
2139 | list_del(&window->list_member); | |
2140 | spin_unlock(&ms_info.mi_rmalock); | |
2141 | micscif_inc_node_refcnt(ep->remote_dev, 1); | |
2142 | mutex_lock(&ep->rma_info.rma_lock); | |
2143 | if (!chan || | |
2144 | !scifdev_alive(ep) || | |
2145 | (!is_current_dma_mark(chan, window->dma_mark) && | |
2146 | is_dma_mark_processed(chan, window->dma_mark)) || | |
2147 | !drain_dma_intr(chan)) { | |
2148 | micscif_dec_node_refcnt(ep->remote_dev, 1); | |
2149 | BUG_ON(atomic_sub_return((int32_t)window->nr_pages, | |
2150 | &ep->rma_info.tcw_total_pages) < 0); | |
2151 | micscif_destroy_window(ep, window); | |
2152 | BUG_ON(atomic_dec_return( | |
2153 | &ep->rma_info.tcw_refcount) < 0); | |
2154 | } else { | |
2155 | /* DMA engine hung ?? */ | |
2156 | printk(KERN_ERR "%s %d DMA channel %d hung ep->state %d " | |
2157 | "window->dma_mark 0x%x channel_mark 0x%x\n", | |
2158 | __func__, __LINE__, get_chan_num(chan), | |
2159 | ep->sd_state, window->dma_mark, get_dma_mark(chan)); | |
2160 | WARN_ON(1); | |
2161 | mutex_unlock(&ep->rma_info.rma_lock); | |
2162 | micscif_dec_node_refcnt(ep->remote_dev, 1); | |
2163 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); | |
2164 | goto restart; | |
2165 | } | |
2166 | mutex_unlock(&ep->rma_info.rma_lock); | |
2167 | goto restart; | |
2168 | } | |
2169 | spin_unlock(&ms_info.mi_rmalock); | |
2170 | } | |
2171 | ||
2172 | /** | |
2173 | * micscif_rma_handle_remote_fences: | |
2174 | * | |
2175 | * This routine services remote fence requests. | |
2176 | */ | |
2177 | void micscif_rma_handle_remote_fences(void) | |
2178 | { | |
2179 | struct list_head *item, *tmp; | |
2180 | struct remote_fence_info *fence; | |
2181 | struct endpt *ep; | |
2182 | int mark; | |
2183 | ||
2184 | might_sleep(); | |
2185 | mutex_lock(&ms_info.mi_fencelock); | |
2186 | list_for_each_safe(item, tmp, &ms_info.mi_fence) { | |
2187 | fence = list_entry(item, | |
2188 | struct remote_fence_info, list_member); | |
2189 | /* Remove fence from global list */ | |
2190 | list_del(&fence->list_member); | |
2191 | ||
2192 | /* Initiate the fence operation */ | |
2193 | ep = (struct endpt *)fence->msg.payload[0]; | |
2194 | mark = (int)fence->msg.payload[2]; | |
2195 | BUG_ON(!(mark & SCIF_REMOTE_FENCE)); | |
2196 | if (dma_mark_wait(ep->rma_info.dma_chan, | |
2197 | mark & ~SCIF_REMOTE_FENCE, false)) { | |
2198 | printk(KERN_ERR "%s %d err\n", __func__, __LINE__); | |
2199 | fence->msg.uop = SCIF_WAIT_NACK; | |
2200 | } else { | |
2201 | fence->msg.uop = SCIF_WAIT_ACK; | |
2202 | } | |
2203 | micscif_inc_node_refcnt(ep->remote_dev, 1); | |
2204 | fence->msg.payload[0] = ep->remote_ep; | |
2205 | /* No error handling for Notification messages. */ | |
2206 | micscif_nodeqp_send(ep->remote_dev, &fence->msg, ep); | |
2207 | micscif_dec_node_refcnt(ep->remote_dev, 1); | |
2208 | kfree(fence); | |
2209 | /* | |
2210 | * Decrement ref count and wake up | |
2211 | * any thread blocked in the EP close routine waiting | |
2212 | * for all such remote fence requests to complete. | |
2213 | */ | |
2214 | ep->rma_info.fence_refcount--; | |
2215 | wake_up(&ep->rma_info.fence_wq); | |
2216 | } | |
2217 | mutex_unlock(&ms_info.mi_fencelock); | |
2218 | } | |
2219 | ||
2220 | #ifdef CONFIG_MMU_NOTIFIER | |
2221 | void micscif_mmu_notif_handler(struct work_struct *work) | |
2222 | { | |
2223 | struct list_head *pos, *tmpq; | |
2224 | struct endpt *ep; | |
2225 | restart: | |
2226 | micscif_rma_destroy_tcw_invalid(&ms_info.mi_rma_tc); | |
2227 | spin_lock(&ms_info.mi_rmalock); | |
2228 | list_for_each_safe(pos, tmpq, &ms_info.mi_mmu_notif_cleanup) { | |
2229 | ep = list_entry(pos, struct endpt, mmu_list); | |
2230 | list_del(&ep->mmu_list); | |
2231 | spin_unlock(&ms_info.mi_rmalock); | |
2232 | BUG_ON(list_empty(&ep->rma_info.mmn_list)); | |
2233 | ||
2234 | micscif_rma_destroy_tcw_ep(ep); | |
2235 | ep_unregister_mmu_notifier(ep); | |
2236 | queue_work(ms_info.mi_misc_wq, &ms_info.mi_misc_work); | |
2237 | goto restart; | |
2238 | } | |
2239 | spin_unlock(&ms_info.mi_rmalock); | |
2240 | } | |
2241 | #endif | |
2242 | ||
2243 | /** | |
2244 | * micscif_reserve_dma_chan: | |
2245 | * @ep: Endpoint Descriptor. | |
2246 | * | |
2247 | * This routine reserves a DMA channel for a particular | |
2248 | * endpoint. All DMA transfers for an endpoint are always | |
2249 | * programmed on the same DMA channel. | |
2250 | */ | |
2251 | int micscif_reserve_dma_chan(struct endpt *ep) | |
2252 | { | |
2253 | int err = 0; | |
2254 | #ifndef _MIC_SCIF_ | |
2255 | /* | |
2256 | * Host Loopback cannot use DMA by design and hence | |
2257 | * reserving DMA channels is a nop. | |
2258 | */ | |
2259 | if (is_self_scifdev(ep->remote_dev)) | |
2260 | return 0; | |
2261 | #endif | |
2262 | mutex_lock(&ep->rma_info.rma_lock); | |
2263 | if (!ep->rma_info.dma_chan) { | |
2264 | struct dma_channel **chan = &ep->rma_info.dma_chan; | |
2265 | unsigned long ts = jiffies; | |
2266 | #ifndef _MIC_SCIF_ | |
2267 | mic_ctx_t *mic_ctx = | |
2268 | get_per_dev_ctx(ep->remote_dev->sd_node - 1); | |
2269 | BUG_ON(!ep->remote_dev->sd_node); | |
2270 | #endif | |
2271 | while (true) { | |
2272 | if (!(err = allocate_dma_channel((struct mic_dma_ctx_t *) | |
2273 | #ifdef _MIC_SCIF_ | |
2274 | mic_dma_handle, | |
2275 | #else | |
2276 | mic_ctx->dma_handle, | |
2277 | #endif | |
2278 | chan))) | |
2279 | break; | |
2280 | schedule(); | |
2281 | if (time_after(jiffies, | |
2282 | ts + NODE_ALIVE_TIMEOUT)) { | |
2283 | err = -EBUSY; | |
2284 | goto error; | |
2285 | } | |
2286 | } | |
2287 | mic_dma_thread_free_chan(*chan); | |
2288 | } | |
2289 | error: | |
2290 | mutex_unlock(&ep->rma_info.rma_lock); | |
2291 | return err; | |
2292 | } | |
2293 | ||
2294 | /* | |
2295 | * micscif_prog_signal: | |
2296 | * @epd - Endpoint Descriptor | |
2297 | * @offset - registered address | |
2298 | * @val - Value to be programmed in SUD. | |
2299 | * @type - Type of the window. | |
2300 | * | |
2301 | * Program a status update descriptor adter ensuring that the offset | |
2302 | * provided is indeed valid. | |
2303 | */ | |
2304 | int micscif_prog_signal(scif_epd_t epd, off_t offset, uint64_t val, | |
2305 | enum rma_window_type type) | |
2306 | { | |
2307 | struct endpt *ep = (struct endpt *)epd; | |
2308 | struct dma_channel *chan = ep->rma_info.dma_chan; | |
2309 | struct reg_range_t *window = NULL; | |
2310 | struct micscif_rma_req req; | |
2311 | int err; | |
2312 | dma_addr_t phys; | |
2313 | ||
2314 | mutex_lock(&ep->rma_info.rma_lock); | |
2315 | req.out_window = &window; | |
2316 | req.offset = offset; | |
2317 | req.nr_bytes = sizeof(uint64_t); | |
2318 | req.prot = SCIF_PROT_WRITE; | |
2319 | req.type = WINDOW_SINGLE; | |
2320 | if (RMA_WINDOW_SELF == type) | |
2321 | req.head = &ep->rma_info.reg_list; | |
2322 | else | |
2323 | req.head = &ep->rma_info.remote_reg_list; | |
2324 | /* Does a valid window exist? */ | |
2325 | if ((err = micscif_query_window(&req))) { | |
2326 | printk(KERN_ERR "%s %d err %d\n", | |
2327 | __func__, __LINE__, err); | |
2328 | goto unlock_ret; | |
2329 | } | |
2330 | RMA_MAGIC(window); | |
2331 | ||
2332 | #ifndef _MIC_SCIF_ | |
2333 | if (unlikely(is_self_scifdev(ep->remote_dev))) { | |
2334 | void *dst_virt; | |
2335 | if (RMA_WINDOW_SELF == type) | |
2336 | dst_virt = get_local_va(offset, window, | |
2337 | sizeof(uint32_t)); | |
2338 | else { | |
2339 | struct page **pages = ((struct reg_range_t *) | |
2340 | (window->peer_window))->pinned_pages->pages; | |
2341 | int page_nr = (int) ( (offset - window->offset) >> PAGE_SHIFT ); | |
2342 | off_t page_off = offset & ~PAGE_MASK; | |
2343 | dst_virt = (void *)((uint64_t)phys_to_virt(page_to_phys( | |
2344 | pages[page_nr])) | page_off); | |
2345 | } | |
2346 | *(uint64_t*)dst_virt = val; | |
2347 | goto unlock_ret; | |
2348 | } | |
2349 | #endif | |
2350 | phys = micscif_get_dma_addr(window, offset, NULL, NULL, NULL); | |
2351 | if ((err = request_dma_channel(chan))) | |
2352 | goto unlock_ret; | |
2353 | err = do_status_update(chan, phys, val); | |
2354 | free_dma_channel(chan); | |
2355 | unlock_ret: | |
2356 | mutex_unlock(&ep->rma_info.rma_lock); | |
2357 | return err; | |
2358 | } | |
2359 | ||
2360 | /* | |
2361 | * __micscif_kill_apps_with_mmaps: | |
2362 | * @ep - The SCIF endpoint | |
2363 | * | |
2364 | * Kill the applications which have valid remote memory mappings | |
2365 | * created via scif_mmap(..). | |
2366 | */ | |
2367 | static void __micscif_kill_apps_with_mmaps(struct endpt *ep) | |
2368 | { | |
2369 | struct list_head *item; | |
2370 | struct rma_task_info *info; | |
2371 | ||
2372 | spin_lock(&ep->lock); | |
2373 | list_for_each(item, &ep->rma_info.task_list) { | |
2374 | info = list_entry(item, struct rma_task_info, list_member); | |
2375 | kill_pid(info->pid, SIGKILL, 1); | |
2376 | pr_debug("%s ep %p pid %p ref %d\n", | |
2377 | __func__, ep, info->pid, info->ref_count); | |
2378 | } | |
2379 | spin_unlock(&ep->lock); | |
2380 | } | |
2381 | ||
2382 | /* | |
2383 | * _micscif_kill_apps_with_mmaps: | |
2384 | * @node - remote node id. | |
2385 | * @head - head of the list of endpoints to kill. | |
2386 | * | |
2387 | * Traverse the list of endpoints for a particular remote node and | |
2388 | * kill applications with valid remote memory mappings. | |
2389 | */ | |
2390 | static void _micscif_kill_apps_with_mmaps(int node, struct list_head *head) | |
2391 | { | |
2392 | struct endpt *ep; | |
2393 | unsigned long sflags; | |
2394 | struct list_head *item; | |
2395 | ||
2396 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); | |
2397 | list_for_each(item, head) { | |
2398 | ep = list_entry(item, struct endpt, list); | |
2399 | if (ep->remote_dev->sd_node == node) | |
2400 | __micscif_kill_apps_with_mmaps(ep); | |
2401 | } | |
2402 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); | |
2403 | } | |
2404 | ||
2405 | /* | |
2406 | * micscif_kill_apps_with_mmaps: | |
2407 | * @node - remote node id. | |
2408 | * | |
2409 | * Wrapper for killing applications with valid remote memory mappings | |
2410 | * for a particular node. This API is called by peer nodes as part of | |
2411 | * handling a lost node. | |
2412 | */ | |
2413 | void micscif_kill_apps_with_mmaps(int node) | |
2414 | { | |
2415 | _micscif_kill_apps_with_mmaps(node, &ms_info.mi_connected); | |
2416 | _micscif_kill_apps_with_mmaps(node, &ms_info.mi_disconnected); | |
2417 | } | |
2418 | ||
2419 | /* | |
2420 | * micscif_query_apps_with_mmaps: | |
2421 | * @node - remote node id. | |
2422 | * @head - head of the list of endpoints to query. | |
2423 | * | |
2424 | * Query if any applications for a remote node have valid remote memory | |
2425 | * mappings. | |
2426 | */ | |
2427 | static bool micscif_query_apps_with_mmaps(int node, struct list_head *head) | |
2428 | { | |
2429 | struct endpt *ep; | |
2430 | unsigned long sflags; | |
2431 | struct list_head *item; | |
2432 | bool ret = false; | |
2433 | ||
2434 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); | |
2435 | list_for_each(item, head) { | |
2436 | ep = list_entry(item, struct endpt, list); | |
2437 | if (ep->remote_dev->sd_node == node && | |
2438 | !list_empty(&ep->rma_info.task_list)) { | |
2439 | ret = true; | |
2440 | break; | |
2441 | } | |
2442 | } | |
2443 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); | |
2444 | return ret; | |
2445 | } | |
2446 | ||
2447 | /* | |
2448 | * micscif_rma_do_apps_have_mmaps: | |
2449 | * @node - remote node id. | |
2450 | * | |
2451 | * Wrapper for querying if any applications have remote memory mappings | |
2452 | * for a particular node. | |
2453 | */ | |
2454 | bool micscif_rma_do_apps_have_mmaps(int node) | |
2455 | { | |
2456 | return (micscif_query_apps_with_mmaps(node, &ms_info.mi_connected) || | |
2457 | micscif_query_apps_with_mmaps(node, &ms_info.mi_disconnected)); | |
2458 | } | |
2459 | ||
2460 | /* | |
2461 | * __micscif_cleanup_rma_for_zombies: | |
2462 | * @ep - The SCIF endpoint | |
2463 | * | |
2464 | * This API is only called while handling a lost node: | |
2465 | * a) Remote node is dead. | |
2466 | * b) All endpoints with remote memory mappings have been killed. | |
2467 | * So we can traverse the remote_reg_list without any locks. Since | |
2468 | * the window has not yet been unregistered we can drop the ref count | |
2469 | * and queue it to the cleanup thread. | |
2470 | */ | |
2471 | static void __micscif_cleanup_rma_for_zombies(struct endpt *ep) | |
2472 | { | |
2473 | struct list_head *pos, *tmp; | |
2474 | struct reg_range_t *window; | |
2475 | ||
2476 | list_for_each_safe(pos, tmp, &ep->rma_info.remote_reg_list) { | |
2477 | window = list_entry(pos, struct reg_range_t, list_member); | |
2478 | /* If unregistration is complete then why is it on the list? */ | |
2479 | WARN_ON(window->unreg_state == OP_COMPLETED); | |
2480 | if (window->ref_count) | |
2481 | put_window_ref_count(window, window->nr_pages); | |
2482 | if (!window->ref_count) { | |
2483 | atomic_inc(&ep->rma_info.tw_refcount); | |
2484 | atomic_add_return((int32_t)window->nr_pages, | |
2485 | &ep->rma_info.tw_total_pages); | |
2486 | list_del(&window->list_member); | |
2487 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); | |
2488 | } | |
2489 | } | |
2490 | } | |
2491 | ||
2492 | /* | |
2493 | * micscif_cleanup_rma_for_zombies: | |
2494 | * @node - remote node id. | |
2495 | * | |
2496 | * Cleanup remote registration lists for zombie endpoints. | |
2497 | */ | |
2498 | void micscif_cleanup_rma_for_zombies(int node) | |
2499 | { | |
2500 | struct endpt *ep; | |
2501 | unsigned long sflags; | |
2502 | struct list_head *item; | |
2503 | ||
2504 | spin_lock_irqsave(&ms_info.mi_eplock, sflags); | |
2505 | list_for_each(item, &ms_info.mi_zombie) { | |
2506 | ep = list_entry(item, struct endpt, list); | |
2507 | if (ep->remote_dev && ep->remote_dev->sd_node == node) { | |
2508 | /* | |
2509 | * If the zombie endpoint remote node matches the lost | |
2510 | * node then the scifdev should not be alive. | |
2511 | */ | |
2512 | WARN_ON(scifdev_alive(ep)); | |
2513 | __micscif_cleanup_rma_for_zombies(ep); | |
2514 | } | |
2515 | } | |
2516 | spin_unlock_irqrestore(&ms_info.mi_eplock, sflags); | |
2517 | } | |
2518 | ||
2519 | /* | |
2520 | * micscif_rma_get_task: | |
2521 | * | |
2522 | * Store the parent task struct and bump up the number of remote mappings. | |
2523 | * If this is the first remote memory mapping for this endpoint then | |
2524 | * create a new rma_task_info entry in the epd task list. | |
2525 | */ | |
2526 | int micscif_rma_get_task(struct endpt *ep, int nr_pages) | |
2527 | { | |
2528 | struct list_head *item; | |
2529 | struct rma_task_info *info; | |
2530 | int err = 0; | |
2531 | ||
2532 | spin_lock(&ep->lock); | |
2533 | list_for_each(item, &ep->rma_info.task_list) { | |
2534 | info = list_entry(item, struct rma_task_info, list_member); | |
2535 | if (info->pid == task_tgid(current)) { | |
2536 | info->ref_count += nr_pages; | |
2537 | pr_debug("%s ep %p existing pid %p ref %d\n", | |
2538 | __func__, ep, info->pid, info->ref_count); | |
2539 | goto unlock; | |
2540 | } | |
2541 | } | |
2542 | spin_unlock(&ep->lock); | |
2543 | ||
2544 | /* A new task is mapping this window. Create a new entry */ | |
2545 | if (!(info = kzalloc(sizeof(*info), GFP_KERNEL))) { | |
2546 | err = -ENOMEM; | |
2547 | goto done; | |
2548 | } | |
2549 | ||
2550 | info->pid = get_pid(task_tgid(current)); | |
2551 | info->ref_count = nr_pages; | |
2552 | pr_debug("%s ep %p new pid %p ref %d\n", | |
2553 | __func__, ep, info->pid, info->ref_count); | |
2554 | spin_lock(&ep->lock); | |
2555 | list_add_tail(&info->list_member, &ep->rma_info.task_list); | |
2556 | unlock: | |
2557 | spin_unlock(&ep->lock); | |
2558 | done: | |
2559 | return err; | |
2560 | } | |
2561 | ||
2562 | /* | |
2563 | * micscif_rma_put_task: | |
2564 | * | |
2565 | * Bump down the number of remote mappings. if the ref count for this | |
2566 | * particular task drops to zero then remove the rma_task_info from | |
2567 | * the epd task list. | |
2568 | */ | |
2569 | void micscif_rma_put_task(struct endpt *ep, int nr_pages) | |
2570 | { | |
2571 | struct list_head *item; | |
2572 | struct rma_task_info *info; | |
2573 | ||
2574 | spin_lock(&ep->lock); | |
2575 | list_for_each(item, &ep->rma_info.task_list) { | |
2576 | info = list_entry(item, struct rma_task_info, list_member); | |
2577 | if (info->pid == task_tgid(current)) { | |
2578 | info->ref_count -= nr_pages; | |
2579 | pr_debug("%s ep %p pid %p ref %d\n", | |
2580 | __func__, ep, info->pid, info->ref_count); | |
2581 | if (!info->ref_count) { | |
2582 | list_del(&info->list_member); | |
2583 | put_pid(info->pid); | |
2584 | kfree(info); | |
2585 | } | |
2586 | goto done; | |
2587 | } | |
2588 | } | |
2589 | /* Why was the task not found? This is a bug. */ | |
2590 | WARN_ON(1); | |
2591 | done: | |
2592 | spin_unlock(&ep->lock); | |
2593 | return; | |
2594 | } | |
2595 | ||
2596 | /* Only debug API's below */ | |
2597 | void micscif_display_window(struct reg_range_t *window, const char *s, int line) | |
2598 | { | |
2599 | int j; | |
2600 | ||
2601 | printk("%s %d window %p type %d temp %d offset 0x%llx" | |
2602 | " nr_pages 0x%llx nr_contig_chunks 0x%llx" | |
2603 | " prot %d ref_count %d magic 0x%llx peer_window 0x%llx" | |
2604 | " unreg_state 0x%x va_for_temp %p\n", | |
2605 | s, line, window, window->type, window->temp, | |
2606 | window->offset, window->nr_pages, window->nr_contig_chunks, | |
2607 | window->prot, window->ref_count, window->magic, | |
2608 | window->peer_window, window->unreg_state, window->va_for_temp); | |
2609 | ||
2610 | for (j = 0; j < window->nr_contig_chunks; j++) | |
2611 | pr_debug("page[%d] = dma_addr 0x%llx num_pages 0x%x\n", | |
2612 | j, | |
2613 | window->dma_addr[j], | |
2614 | window->num_pages[j]); | |
2615 | ||
2616 | if (RMA_WINDOW_SELF == window->type && window->pinned_pages) | |
2617 | for (j = 0; j < window->nr_pages; j++) | |
2618 | pr_debug("page[%d] = pinned_pages %p address %p\n", | |
2619 | j, window->pinned_pages->pages[j], | |
2620 | page_address(window->pinned_pages->pages[j])); | |
2621 | ||
2622 | #ifdef CONFIG_ML1OM | |
2623 | if (window->temp_phys_addr) | |
2624 | for (j = 0; j < window->nr_contig_chunks; j++) | |
2625 | pr_debug("page[%d] = temp_phys_addr 0x%llx\n", | |
2626 | j, window->temp_phys_addr[j]); | |
2627 | if (window->phys_addr) | |
2628 | for (j = 0; j < window->nr_pages; j++) | |
2629 | pr_debug("page[%d] = phys_addr 0x%llx\n", | |
2630 | j, window->phys_addr[j]); | |
2631 | #endif | |
2632 | RMA_MAGIC(window); | |
2633 | } |