1 /*
2 * Intel MIC Platform Software Stack (MPSS)
3 *
4 * Copyright(c) 2015 Intel Corporation.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2, as
8 * published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License for more details.
14 *
15 * Intel SCIF driver.
16 *
17 */
18 #include <linux/dma_remapping.h>
19 #include <linux/pagemap.h>
20 #include <linux/sched/mm.h>
21 #include <linux/sched/signal.h>
22
23 #include "scif_main.h"
24 #include "scif_map.h"
25
26 /* Used to skip ulimit checks for registrations with SCIF_MAP_KERNEL flag */
27 #define SCIF_MAP_ULIMIT 0x40
28
29 bool scif_ulimit_check = 1;
30
31 /**
32 * scif_rma_ep_init:
33 * @ep: end point
34 *
35 * Initialize RMA per EP data structures.
36 */
scif_rma_ep_init(struct scif_endpt * ep)37 void scif_rma_ep_init(struct scif_endpt *ep)
38 {
39 struct scif_endpt_rma_info *rma = &ep->rma_info;
40
41 mutex_init(&rma->rma_lock);
42 init_iova_domain(&rma->iovad, PAGE_SIZE, SCIF_IOVA_START_PFN,
43 SCIF_DMA_64BIT_PFN);
44 spin_lock_init(&rma->tc_lock);
45 mutex_init(&rma->mmn_lock);
46 INIT_LIST_HEAD(&rma->reg_list);
47 INIT_LIST_HEAD(&rma->remote_reg_list);
48 atomic_set(&rma->tw_refcount, 0);
49 atomic_set(&rma->tcw_refcount, 0);
50 atomic_set(&rma->tcw_total_pages, 0);
51 atomic_set(&rma->fence_refcount, 0);
52
53 rma->async_list_del = 0;
54 rma->dma_chan = NULL;
55 INIT_LIST_HEAD(&rma->mmn_list);
56 INIT_LIST_HEAD(&rma->vma_list);
57 init_waitqueue_head(&rma->markwq);
58 }
59
60 /**
61 * scif_rma_ep_can_uninit:
62 * @ep: end point
63 *
64 * Returns 1 if an endpoint can be uninitialized and 0 otherwise.
65 */
scif_rma_ep_can_uninit(struct scif_endpt * ep)66 int scif_rma_ep_can_uninit(struct scif_endpt *ep)
67 {
68 int ret = 0;
69
70 mutex_lock(&ep->rma_info.rma_lock);
71 /* Destroy RMA Info only if both lists are empty */
72 if (list_empty(&ep->rma_info.reg_list) &&
73 list_empty(&ep->rma_info.remote_reg_list) &&
74 list_empty(&ep->rma_info.mmn_list) &&
75 !atomic_read(&ep->rma_info.tw_refcount) &&
76 !atomic_read(&ep->rma_info.tcw_refcount) &&
77 !atomic_read(&ep->rma_info.fence_refcount))
78 ret = 1;
79 mutex_unlock(&ep->rma_info.rma_lock);
80 return ret;
81 }
82
83 /**
84 * scif_create_pinned_pages:
85 * @nr_pages: number of pages in window
86 * @prot: read/write protection
87 *
88 * Allocate and prepare a set of pinned pages.
89 */
90 static struct scif_pinned_pages *
scif_create_pinned_pages(int nr_pages,int prot)91 scif_create_pinned_pages(int nr_pages, int prot)
92 {
93 struct scif_pinned_pages *pin;
94
95 might_sleep();
96 pin = scif_zalloc(sizeof(*pin));
97 if (!pin)
98 goto error;
99
100 pin->pages = scif_zalloc(nr_pages * sizeof(*pin->pages));
101 if (!pin->pages)
102 goto error_free_pinned_pages;
103
104 pin->prot = prot;
105 pin->magic = SCIFEP_MAGIC;
106 return pin;
107
108 error_free_pinned_pages:
109 scif_free(pin, sizeof(*pin));
110 error:
111 return NULL;
112 }
113
114 /**
115 * scif_destroy_pinned_pages:
116 * @pin: A set of pinned pages.
117 *
118 * Deallocate resources for pinned pages.
119 */
scif_destroy_pinned_pages(struct scif_pinned_pages * pin)120 static int scif_destroy_pinned_pages(struct scif_pinned_pages *pin)
121 {
122 int j;
123 int writeable = pin->prot & SCIF_PROT_WRITE;
124 int kernel = SCIF_MAP_KERNEL & pin->map_flags;
125
126 for (j = 0; j < pin->nr_pages; j++) {
127 if (pin->pages[j] && !kernel) {
128 if (writeable)
129 SetPageDirty(pin->pages[j]);
130 put_page(pin->pages[j]);
131 }
132 }
133
134 scif_free(pin->pages,
135 pin->nr_pages * sizeof(*pin->pages));
136 scif_free(pin, sizeof(*pin));
137 return 0;
138 }
139
140 /*
141 * scif_create_window:
142 * @ep: end point
143 * @nr_pages: number of pages
144 * @offset: registration offset
145 * @temp: true if a temporary window is being created
146 *
147 * Allocate and prepare a self registration window.
148 */
scif_create_window(struct scif_endpt * ep,int nr_pages,s64 offset,bool temp)149 struct scif_window *scif_create_window(struct scif_endpt *ep, int nr_pages,
150 s64 offset, bool temp)
151 {
152 struct scif_window *window;
153
154 might_sleep();
155 window = scif_zalloc(sizeof(*window));
156 if (!window)
157 goto error;
158
159 window->dma_addr = scif_zalloc(nr_pages * sizeof(*window->dma_addr));
160 if (!window->dma_addr)
161 goto error_free_window;
162
163 window->num_pages = scif_zalloc(nr_pages * sizeof(*window->num_pages));
164 if (!window->num_pages)
165 goto error_free_window;
166
167 window->offset = offset;
168 window->ep = (u64)ep;
169 window->magic = SCIFEP_MAGIC;
170 window->reg_state = OP_IDLE;
171 init_waitqueue_head(&window->regwq);
172 window->unreg_state = OP_IDLE;
173 init_waitqueue_head(&window->unregwq);
174 INIT_LIST_HEAD(&window->list);
175 window->type = SCIF_WINDOW_SELF;
176 window->temp = temp;
177 return window;
178
179 error_free_window:
180 scif_free(window->dma_addr,
181 nr_pages * sizeof(*window->dma_addr));
182 scif_free(window, sizeof(*window));
183 error:
184 return NULL;
185 }
186
187 /**
188 * scif_destroy_incomplete_window:
189 * @ep: end point
190 * @window: registration window
191 *
192 * Deallocate resources for self window.
193 */
scif_destroy_incomplete_window(struct scif_endpt * ep,struct scif_window * window)194 static void scif_destroy_incomplete_window(struct scif_endpt *ep,
195 struct scif_window *window)
196 {
197 int err;
198 int nr_pages = window->nr_pages;
199 struct scif_allocmsg *alloc = &window->alloc_handle;
200 struct scifmsg msg;
201
202 retry:
203 /* Wait for a SCIF_ALLOC_GNT/REJ message */
204 err = wait_event_timeout(alloc->allocwq,
205 alloc->state != OP_IN_PROGRESS,
206 SCIF_NODE_ALIVE_TIMEOUT);
207 if (!err && scifdev_alive(ep))
208 goto retry;
209
210 mutex_lock(&ep->rma_info.rma_lock);
211 if (alloc->state == OP_COMPLETED) {
212 msg.uop = SCIF_FREE_VIRT;
213 msg.src = ep->port;
214 msg.payload[0] = ep->remote_ep;
215 msg.payload[1] = window->alloc_handle.vaddr;
216 msg.payload[2] = (u64)window;
217 msg.payload[3] = SCIF_REGISTER;
218 _scif_nodeqp_send(ep->remote_dev, &msg);
219 }
220 mutex_unlock(&ep->rma_info.rma_lock);
221
222 scif_free_window_offset(ep, window, window->offset);
223 scif_free(window->dma_addr, nr_pages * sizeof(*window->dma_addr));
224 scif_free(window->num_pages, nr_pages * sizeof(*window->num_pages));
225 scif_free(window, sizeof(*window));
226 }
227
228 /**
229 * scif_unmap_window:
230 * @remote_dev: SCIF remote device
231 * @window: registration window
232 *
233 * Delete any DMA mappings created for a registered self window
234 */
scif_unmap_window(struct scif_dev * remote_dev,struct scif_window * window)235 void scif_unmap_window(struct scif_dev *remote_dev, struct scif_window *window)
236 {
237 int j;
238
239 if (scif_is_iommu_enabled() && !scifdev_self(remote_dev)) {
240 if (window->st) {
241 dma_unmap_sg(&remote_dev->sdev->dev,
242 window->st->sgl, window->st->nents,
243 DMA_BIDIRECTIONAL);
244 sg_free_table(window->st);
245 kfree(window->st);
246 window->st = NULL;
247 }
248 } else {
249 for (j = 0; j < window->nr_contig_chunks; j++) {
250 if (window->dma_addr[j]) {
251 scif_unmap_single(window->dma_addr[j],
252 remote_dev,
253 window->num_pages[j] <<
254 PAGE_SHIFT);
255 window->dma_addr[j] = 0x0;
256 }
257 }
258 }
259 }
260
__scif_acquire_mm(void)261 static inline struct mm_struct *__scif_acquire_mm(void)
262 {
263 if (scif_ulimit_check)
264 return get_task_mm(current);
265 return NULL;
266 }
267
__scif_release_mm(struct mm_struct * mm)268 static inline void __scif_release_mm(struct mm_struct *mm)
269 {
270 if (mm)
271 mmput(mm);
272 }
273
274 static inline int
__scif_dec_pinned_vm_lock(struct mm_struct * mm,int nr_pages,bool try_lock)275 __scif_dec_pinned_vm_lock(struct mm_struct *mm,
276 int nr_pages, bool try_lock)
277 {
278 if (!mm || !nr_pages || !scif_ulimit_check)
279 return 0;
280 if (try_lock) {
281 if (!down_write_trylock(&mm->mmap_sem)) {
282 dev_err(scif_info.mdev.this_device,
283 "%s %d err\n", __func__, __LINE__);
284 return -1;
285 }
286 } else {
287 down_write(&mm->mmap_sem);
288 }
289 mm->pinned_vm -= nr_pages;
290 up_write(&mm->mmap_sem);
291 return 0;
292 }
293
__scif_check_inc_pinned_vm(struct mm_struct * mm,int nr_pages)294 static inline int __scif_check_inc_pinned_vm(struct mm_struct *mm,
295 int nr_pages)
296 {
297 unsigned long locked, lock_limit;
298
299 if (!mm || !nr_pages || !scif_ulimit_check)
300 return 0;
301
302 locked = nr_pages;
303 locked += mm->pinned_vm;
304 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
305 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
306 dev_err(scif_info.mdev.this_device,
307 "locked(%lu) > lock_limit(%lu)\n",
308 locked, lock_limit);
309 return -ENOMEM;
310 }
311 mm->pinned_vm = locked;
312 return 0;
313 }
314
315 /**
316 * scif_destroy_window:
317 * @ep: end point
318 * @window: registration window
319 *
320 * Deallocate resources for self window.
321 */
scif_destroy_window(struct scif_endpt * ep,struct scif_window * window)322 int scif_destroy_window(struct scif_endpt *ep, struct scif_window *window)
323 {
324 int j;
325 struct scif_pinned_pages *pinned_pages = window->pinned_pages;
326 int nr_pages = window->nr_pages;
327
328 might_sleep();
329 if (!window->temp && window->mm) {
330 __scif_dec_pinned_vm_lock(window->mm, window->nr_pages, 0);
331 __scif_release_mm(window->mm);
332 window->mm = NULL;
333 }
334
335 scif_free_window_offset(ep, window, window->offset);
336 scif_unmap_window(ep->remote_dev, window);
337 /*
338 * Decrement references for this set of pinned pages from
339 * this window.
340 */
341 j = atomic_sub_return(1, &pinned_pages->ref_count);
342 if (j < 0)
343 dev_err(scif_info.mdev.this_device,
344 "%s %d incorrect ref count %d\n",
345 __func__, __LINE__, j);
346 /*
347 * If the ref count for pinned_pages is zero then someone
348 * has already called scif_unpin_pages() for it and we should
349 * destroy the page cache.
350 */
351 if (!j)
352 scif_destroy_pinned_pages(window->pinned_pages);
353 scif_free(window->dma_addr, nr_pages * sizeof(*window->dma_addr));
354 scif_free(window->num_pages, nr_pages * sizeof(*window->num_pages));
355 window->magic = 0;
356 scif_free(window, sizeof(*window));
357 return 0;
358 }
359
360 /**
361 * scif_create_remote_lookup:
362 * @remote_dev: SCIF remote device
363 * @window: remote window
364 *
365 * Allocate and prepare lookup entries for the remote
366 * end to copy over the physical addresses.
367 * Returns 0 on success and appropriate errno on failure.
368 */
scif_create_remote_lookup(struct scif_dev * remote_dev,struct scif_window * window)369 static int scif_create_remote_lookup(struct scif_dev *remote_dev,
370 struct scif_window *window)
371 {
372 int i, j, err = 0;
373 int nr_pages = window->nr_pages;
374 bool vmalloc_dma_phys, vmalloc_num_pages;
375
376 might_sleep();
377 /* Map window */
378 err = scif_map_single(&window->mapped_offset,
379 window, remote_dev, sizeof(*window));
380 if (err)
381 goto error_window;
382
383 /* Compute the number of lookup entries. 21 == 2MB Shift */
384 window->nr_lookup = ALIGN(nr_pages * PAGE_SIZE,
385 ((2) * 1024 * 1024)) >> 21;
386
387 window->dma_addr_lookup.lookup =
388 scif_alloc_coherent(&window->dma_addr_lookup.offset,
389 remote_dev, window->nr_lookup *
390 sizeof(*window->dma_addr_lookup.lookup),
391 GFP_KERNEL | __GFP_ZERO);
392 if (!window->dma_addr_lookup.lookup) {
393 err = -ENOMEM;
394 goto error_window;
395 }
396
397 window->num_pages_lookup.lookup =
398 scif_alloc_coherent(&window->num_pages_lookup.offset,
399 remote_dev, window->nr_lookup *
400 sizeof(*window->num_pages_lookup.lookup),
401 GFP_KERNEL | __GFP_ZERO);
402 if (!window->num_pages_lookup.lookup) {
403 err = -ENOMEM;
404 goto error_window;
405 }
406
407 vmalloc_dma_phys = is_vmalloc_addr(&window->dma_addr[0]);
408 vmalloc_num_pages = is_vmalloc_addr(&window->num_pages[0]);
409
410 /* Now map each of the pages containing physical addresses */
411 for (i = 0, j = 0; i < nr_pages; i += SCIF_NR_ADDR_IN_PAGE, j++) {
412 err = scif_map_page(&window->dma_addr_lookup.lookup[j],
413 vmalloc_dma_phys ?
414 vmalloc_to_page(&window->dma_addr[i]) :
415 virt_to_page(&window->dma_addr[i]),
416 remote_dev);
417 if (err)
418 goto error_window;
419 err = scif_map_page(&window->num_pages_lookup.lookup[j],
420 vmalloc_num_pages ?
421 vmalloc_to_page(&window->num_pages[i]) :
422 virt_to_page(&window->num_pages[i]),
423 remote_dev);
424 if (err)
425 goto error_window;
426 }
427 return 0;
428 error_window:
429 return err;
430 }
431
432 /**
433 * scif_destroy_remote_lookup:
434 * @remote_dev: SCIF remote device
435 * @window: remote window
436 *
437 * Destroy lookup entries used for the remote
438 * end to copy over the physical addresses.
439 */
scif_destroy_remote_lookup(struct scif_dev * remote_dev,struct scif_window * window)440 static void scif_destroy_remote_lookup(struct scif_dev *remote_dev,
441 struct scif_window *window)
442 {
443 int i, j;
444
445 if (window->nr_lookup) {
446 struct scif_rma_lookup *lup = &window->dma_addr_lookup;
447 struct scif_rma_lookup *npup = &window->num_pages_lookup;
448
449 for (i = 0, j = 0; i < window->nr_pages;
450 i += SCIF_NR_ADDR_IN_PAGE, j++) {
451 if (lup->lookup && lup->lookup[j])
452 scif_unmap_single(lup->lookup[j],
453 remote_dev,
454 PAGE_SIZE);
455 if (npup->lookup && npup->lookup[j])
456 scif_unmap_single(npup->lookup[j],
457 remote_dev,
458 PAGE_SIZE);
459 }
460 if (lup->lookup)
461 scif_free_coherent(lup->lookup, lup->offset,
462 remote_dev, window->nr_lookup *
463 sizeof(*lup->lookup));
464 if (npup->lookup)
465 scif_free_coherent(npup->lookup, npup->offset,
466 remote_dev, window->nr_lookup *
467 sizeof(*npup->lookup));
468 if (window->mapped_offset)
469 scif_unmap_single(window->mapped_offset,
470 remote_dev, sizeof(*window));
471 window->nr_lookup = 0;
472 }
473 }
474
475 /**
476 * scif_create_remote_window:
477 * @ep: end point
478 * @nr_pages: number of pages in window
479 *
480 * Allocate and prepare a remote registration window.
481 */
482 static struct scif_window *
scif_create_remote_window(struct scif_dev * scifdev,int nr_pages)483 scif_create_remote_window(struct scif_dev *scifdev, int nr_pages)
484 {
485 struct scif_window *window;
486
487 might_sleep();
488 window = scif_zalloc(sizeof(*window));
489 if (!window)
490 goto error_ret;
491
492 window->magic = SCIFEP_MAGIC;
493 window->nr_pages = nr_pages;
494
495 window->dma_addr = scif_zalloc(nr_pages * sizeof(*window->dma_addr));
496 if (!window->dma_addr)
497 goto error_window;
498
499 window->num_pages = scif_zalloc(nr_pages *
500 sizeof(*window->num_pages));
501 if (!window->num_pages)
502 goto error_window;
503
504 if (scif_create_remote_lookup(scifdev, window))
505 goto error_window;
506
507 window->type = SCIF_WINDOW_PEER;
508 window->unreg_state = OP_IDLE;
509 INIT_LIST_HEAD(&window->list);
510 return window;
511 error_window:
512 scif_destroy_remote_window(window);
513 error_ret:
514 return NULL;
515 }
516
517 /**
518 * scif_destroy_remote_window:
519 * @ep: end point
520 * @window: remote registration window
521 *
522 * Deallocate resources for remote window.
523 */
524 void
scif_destroy_remote_window(struct scif_window * window)525 scif_destroy_remote_window(struct scif_window *window)
526 {
527 scif_free(window->dma_addr, window->nr_pages *
528 sizeof(*window->dma_addr));
529 scif_free(window->num_pages, window->nr_pages *
530 sizeof(*window->num_pages));
531 window->magic = 0;
532 scif_free(window, sizeof(*window));
533 }
534
535 /**
536 * scif_iommu_map: create DMA mappings if the IOMMU is enabled
537 * @remote_dev: SCIF remote device
538 * @window: remote registration window
539 *
540 * Map the physical pages using dma_map_sg(..) and then detect the number
541 * of contiguous DMA mappings allocated
542 */
scif_iommu_map(struct scif_dev * remote_dev,struct scif_window * window)543 static int scif_iommu_map(struct scif_dev *remote_dev,
544 struct scif_window *window)
545 {
546 struct scatterlist *sg;
547 int i, err;
548 scif_pinned_pages_t pin = window->pinned_pages;
549
550 window->st = kzalloc(sizeof(*window->st), GFP_KERNEL);
551 if (!window->st)
552 return -ENOMEM;
553
554 err = sg_alloc_table(window->st, window->nr_pages, GFP_KERNEL);
555 if (err)
556 return err;
557
558 for_each_sg(window->st->sgl, sg, window->st->nents, i)
559 sg_set_page(sg, pin->pages[i], PAGE_SIZE, 0x0);
560
561 err = dma_map_sg(&remote_dev->sdev->dev, window->st->sgl,
562 window->st->nents, DMA_BIDIRECTIONAL);
563 if (!err)
564 return -ENOMEM;
565 /* Detect contiguous ranges of DMA mappings */
566 sg = window->st->sgl;
567 for (i = 0; sg; i++) {
568 dma_addr_t last_da;
569
570 window->dma_addr[i] = sg_dma_address(sg);
571 window->num_pages[i] = sg_dma_len(sg) >> PAGE_SHIFT;
572 last_da = sg_dma_address(sg) + sg_dma_len(sg);
573 while ((sg = sg_next(sg)) && sg_dma_address(sg) == last_da) {
574 window->num_pages[i] +=
575 (sg_dma_len(sg) >> PAGE_SHIFT);
576 last_da = window->dma_addr[i] +
577 sg_dma_len(sg);
578 }
579 window->nr_contig_chunks++;
580 }
581 return 0;
582 }
583
584 /**
585 * scif_map_window:
586 * @remote_dev: SCIF remote device
587 * @window: self registration window
588 *
589 * Map pages of a window into the aperture/PCI.
590 * Also determine addresses required for DMA.
591 */
592 int
scif_map_window(struct scif_dev * remote_dev,struct scif_window * window)593 scif_map_window(struct scif_dev *remote_dev, struct scif_window *window)
594 {
595 int i, j, k, err = 0, nr_contig_pages;
596 scif_pinned_pages_t pin;
597 phys_addr_t phys_prev, phys_curr;
598
599 might_sleep();
600
601 pin = window->pinned_pages;
602
603 if (intel_iommu_enabled && !scifdev_self(remote_dev))
604 return scif_iommu_map(remote_dev, window);
605
606 for (i = 0, j = 0; i < window->nr_pages; i += nr_contig_pages, j++) {
607 phys_prev = page_to_phys(pin->pages[i]);
608 nr_contig_pages = 1;
609
610 /* Detect physically contiguous chunks */
611 for (k = i + 1; k < window->nr_pages; k++) {
612 phys_curr = page_to_phys(pin->pages[k]);
613 if (phys_curr != (phys_prev + PAGE_SIZE))
614 break;
615 phys_prev = phys_curr;
616 nr_contig_pages++;
617 }
618 window->num_pages[j] = nr_contig_pages;
619 window->nr_contig_chunks++;
620 if (scif_is_mgmt_node()) {
621 /*
622 * Management node has to deal with SMPT on X100 and
623 * hence the DMA mapping is required
624 */
625 err = scif_map_single(&window->dma_addr[j],
626 phys_to_virt(page_to_phys(
627 pin->pages[i])),
628 remote_dev,
629 nr_contig_pages << PAGE_SHIFT);
630 if (err)
631 return err;
632 } else {
633 window->dma_addr[j] = page_to_phys(pin->pages[i]);
634 }
635 }
636 return err;
637 }
638
639 /**
640 * scif_send_scif_unregister:
641 * @ep: end point
642 * @window: self registration window
643 *
644 * Send a SCIF_UNREGISTER message.
645 */
scif_send_scif_unregister(struct scif_endpt * ep,struct scif_window * window)646 static int scif_send_scif_unregister(struct scif_endpt *ep,
647 struct scif_window *window)
648 {
649 struct scifmsg msg;
650
651 msg.uop = SCIF_UNREGISTER;
652 msg.src = ep->port;
653 msg.payload[0] = window->alloc_handle.vaddr;
654 msg.payload[1] = (u64)window;
655 return scif_nodeqp_send(ep->remote_dev, &msg);
656 }
657
658 /**
659 * scif_unregister_window:
660 * @window: self registration window
661 *
662 * Send an unregistration request and wait for a response.
663 */
scif_unregister_window(struct scif_window * window)664 int scif_unregister_window(struct scif_window *window)
665 {
666 int err = 0;
667 struct scif_endpt *ep = (struct scif_endpt *)window->ep;
668 bool send_msg = false;
669
670 might_sleep();
671 switch (window->unreg_state) {
672 case OP_IDLE:
673 {
674 window->unreg_state = OP_IN_PROGRESS;
675 send_msg = true;
676 /* fall through */
677 }
678 case OP_IN_PROGRESS:
679 {
680 scif_get_window(window, 1);
681 mutex_unlock(&ep->rma_info.rma_lock);
682 if (send_msg) {
683 err = scif_send_scif_unregister(ep, window);
684 if (err) {
685 window->unreg_state = OP_COMPLETED;
686 goto done;
687 }
688 } else {
689 /* Return ENXIO since unregistration is in progress */
690 mutex_lock(&ep->rma_info.rma_lock);
691 return -ENXIO;
692 }
693 retry:
694 /* Wait for a SCIF_UNREGISTER_(N)ACK message */
695 err = wait_event_timeout(window->unregwq,
696 window->unreg_state != OP_IN_PROGRESS,
697 SCIF_NODE_ALIVE_TIMEOUT);
698 if (!err && scifdev_alive(ep))
699 goto retry;
700 if (!err) {
701 err = -ENODEV;
702 window->unreg_state = OP_COMPLETED;
703 dev_err(scif_info.mdev.this_device,
704 "%s %d err %d\n", __func__, __LINE__, err);
705 }
706 if (err > 0)
707 err = 0;
708 done:
709 mutex_lock(&ep->rma_info.rma_lock);
710 scif_put_window(window, 1);
711 break;
712 }
713 case OP_FAILED:
714 {
715 if (!scifdev_alive(ep)) {
716 err = -ENODEV;
717 window->unreg_state = OP_COMPLETED;
718 }
719 break;
720 }
721 case OP_COMPLETED:
722 break;
723 default:
724 err = -ENODEV;
725 }
726
727 if (window->unreg_state == OP_COMPLETED && window->ref_count)
728 scif_put_window(window, window->nr_pages);
729
730 if (!window->ref_count) {
731 atomic_inc(&ep->rma_info.tw_refcount);
732 list_del_init(&window->list);
733 scif_free_window_offset(ep, window, window->offset);
734 mutex_unlock(&ep->rma_info.rma_lock);
735 if ((!!(window->pinned_pages->map_flags & SCIF_MAP_KERNEL)) &&
736 scifdev_alive(ep)) {
737 scif_drain_dma_intr(ep->remote_dev->sdev,
738 ep->rma_info.dma_chan);
739 } else {
740 if (!__scif_dec_pinned_vm_lock(window->mm,
741 window->nr_pages, 1)) {
742 __scif_release_mm(window->mm);
743 window->mm = NULL;
744 }
745 }
746 scif_queue_for_cleanup(window, &scif_info.rma);
747 mutex_lock(&ep->rma_info.rma_lock);
748 }
749 return err;
750 }
751
752 /**
753 * scif_send_alloc_request:
754 * @ep: end point
755 * @window: self registration window
756 *
757 * Send a remote window allocation request
758 */
scif_send_alloc_request(struct scif_endpt * ep,struct scif_window * window)759 static int scif_send_alloc_request(struct scif_endpt *ep,
760 struct scif_window *window)
761 {
762 struct scifmsg msg;
763 struct scif_allocmsg *alloc = &window->alloc_handle;
764
765 /* Set up the Alloc Handle */
766 alloc->state = OP_IN_PROGRESS;
767 init_waitqueue_head(&alloc->allocwq);
768
769 /* Send out an allocation request */
770 msg.uop = SCIF_ALLOC_REQ;
771 msg.payload[1] = window->nr_pages;
772 msg.payload[2] = (u64)&window->alloc_handle;
773 return _scif_nodeqp_send(ep->remote_dev, &msg);
774 }
775
776 /**
777 * scif_prep_remote_window:
778 * @ep: end point
779 * @window: self registration window
780 *
781 * Send a remote window allocation request, wait for an allocation response,
782 * and prepares the remote window by copying over the page lists
783 */
scif_prep_remote_window(struct scif_endpt * ep,struct scif_window * window)784 static int scif_prep_remote_window(struct scif_endpt *ep,
785 struct scif_window *window)
786 {
787 struct scifmsg msg;
788 struct scif_window *remote_window;
789 struct scif_allocmsg *alloc = &window->alloc_handle;
790 dma_addr_t *dma_phys_lookup, *tmp, *num_pages_lookup, *tmp1;
791 int i = 0, j = 0;
792 int nr_contig_chunks, loop_nr_contig_chunks;
793 int remaining_nr_contig_chunks, nr_lookup;
794 int err, map_err;
795
796 map_err = scif_map_window(ep->remote_dev, window);
797 if (map_err)
798 dev_err(&ep->remote_dev->sdev->dev,
799 "%s %d map_err %d\n", __func__, __LINE__, map_err);
800 remaining_nr_contig_chunks = window->nr_contig_chunks;
801 nr_contig_chunks = window->nr_contig_chunks;
802 retry:
803 /* Wait for a SCIF_ALLOC_GNT/REJ message */
804 err = wait_event_timeout(alloc->allocwq,
805 alloc->state != OP_IN_PROGRESS,
806 SCIF_NODE_ALIVE_TIMEOUT);
807 mutex_lock(&ep->rma_info.rma_lock);
808 /* Synchronize with the thread waking up allocwq */
809 mutex_unlock(&ep->rma_info.rma_lock);
810 if (!err && scifdev_alive(ep))
811 goto retry;
812
813 if (!err)
814 err = -ENODEV;
815
816 if (err > 0)
817 err = 0;
818 else
819 return err;
820
821 /* Bail out. The remote end rejected this request */
822 if (alloc->state == OP_FAILED)
823 return -ENOMEM;
824
825 if (map_err) {
826 dev_err(&ep->remote_dev->sdev->dev,
827 "%s %d err %d\n", __func__, __LINE__, map_err);
828 msg.uop = SCIF_FREE_VIRT;
829 msg.src = ep->port;
830 msg.payload[0] = ep->remote_ep;
831 msg.payload[1] = window->alloc_handle.vaddr;
832 msg.payload[2] = (u64)window;
833 msg.payload[3] = SCIF_REGISTER;
834 spin_lock(&ep->lock);
835 if (ep->state == SCIFEP_CONNECTED)
836 err = _scif_nodeqp_send(ep->remote_dev, &msg);
837 else
838 err = -ENOTCONN;
839 spin_unlock(&ep->lock);
840 return err;
841 }
842
843 remote_window = scif_ioremap(alloc->phys_addr, sizeof(*window),
844 ep->remote_dev);
845
846 /* Compute the number of lookup entries. 21 == 2MB Shift */
847 nr_lookup = ALIGN(nr_contig_chunks, SCIF_NR_ADDR_IN_PAGE)
848 >> ilog2(SCIF_NR_ADDR_IN_PAGE);
849
850 dma_phys_lookup =
851 scif_ioremap(remote_window->dma_addr_lookup.offset,
852 nr_lookup *
853 sizeof(*remote_window->dma_addr_lookup.lookup),
854 ep->remote_dev);
855 num_pages_lookup =
856 scif_ioremap(remote_window->num_pages_lookup.offset,
857 nr_lookup *
858 sizeof(*remote_window->num_pages_lookup.lookup),
859 ep->remote_dev);
860
861 while (remaining_nr_contig_chunks) {
862 loop_nr_contig_chunks = min_t(int, remaining_nr_contig_chunks,
863 (int)SCIF_NR_ADDR_IN_PAGE);
864 /* #1/2 - Copy physical addresses over to the remote side */
865
866 /* #2/2 - Copy DMA addresses (addresses that are fed into the
867 * DMA engine) We transfer bus addresses which are then
868 * converted into a MIC physical address on the remote
869 * side if it is a MIC, if the remote node is a mgmt node we
870 * transfer the MIC physical address
871 */
872 tmp = scif_ioremap(dma_phys_lookup[j],
873 loop_nr_contig_chunks *
874 sizeof(*window->dma_addr),
875 ep->remote_dev);
876 tmp1 = scif_ioremap(num_pages_lookup[j],
877 loop_nr_contig_chunks *
878 sizeof(*window->num_pages),
879 ep->remote_dev);
880 if (scif_is_mgmt_node()) {
881 memcpy_toio((void __force __iomem *)tmp,
882 &window->dma_addr[i], loop_nr_contig_chunks
883 * sizeof(*window->dma_addr));
884 memcpy_toio((void __force __iomem *)tmp1,
885 &window->num_pages[i], loop_nr_contig_chunks
886 * sizeof(*window->num_pages));
887 } else {
888 if (scifdev_is_p2p(ep->remote_dev)) {
889 /*
890 * add remote node's base address for this node
891 * to convert it into a MIC address
892 */
893 int m;
894 dma_addr_t dma_addr;
895
896 for (m = 0; m < loop_nr_contig_chunks; m++) {
897 dma_addr = window->dma_addr[i + m] +
898 ep->remote_dev->base_addr;
899 writeq(dma_addr,
900 (void __force __iomem *)&tmp[m]);
901 }
902 memcpy_toio((void __force __iomem *)tmp1,
903 &window->num_pages[i],
904 loop_nr_contig_chunks
905 * sizeof(*window->num_pages));
906 } else {
907 /* Mgmt node or loopback - transfer DMA
908 * addresses as is, this is the same as a
909 * MIC physical address (we use the dma_addr
910 * and not the phys_addr array since the
911 * phys_addr is only setup if there is a mmap()
912 * request from the mgmt node)
913 */
914 memcpy_toio((void __force __iomem *)tmp,
915 &window->dma_addr[i],
916 loop_nr_contig_chunks *
917 sizeof(*window->dma_addr));
918 memcpy_toio((void __force __iomem *)tmp1,
919 &window->num_pages[i],
920 loop_nr_contig_chunks *
921 sizeof(*window->num_pages));
922 }
923 }
924 remaining_nr_contig_chunks -= loop_nr_contig_chunks;
925 i += loop_nr_contig_chunks;
926 j++;
927 scif_iounmap(tmp, loop_nr_contig_chunks *
928 sizeof(*window->dma_addr), ep->remote_dev);
929 scif_iounmap(tmp1, loop_nr_contig_chunks *
930 sizeof(*window->num_pages), ep->remote_dev);
931 }
932
933 /* Prepare the remote window for the peer */
934 remote_window->peer_window = (u64)window;
935 remote_window->offset = window->offset;
936 remote_window->prot = window->prot;
937 remote_window->nr_contig_chunks = nr_contig_chunks;
938 remote_window->ep = ep->remote_ep;
939 scif_iounmap(num_pages_lookup,
940 nr_lookup *
941 sizeof(*remote_window->num_pages_lookup.lookup),
942 ep->remote_dev);
943 scif_iounmap(dma_phys_lookup,
944 nr_lookup *
945 sizeof(*remote_window->dma_addr_lookup.lookup),
946 ep->remote_dev);
947 scif_iounmap(remote_window, sizeof(*remote_window), ep->remote_dev);
948 window->peer_window = alloc->vaddr;
949 return err;
950 }
951
952 /**
953 * scif_send_scif_register:
954 * @ep: end point
955 * @window: self registration window
956 *
957 * Send a SCIF_REGISTER message if EP is connected and wait for a
958 * SCIF_REGISTER_(N)ACK message else send a SCIF_FREE_VIRT
959 * message so that the peer can free its remote window allocated earlier.
960 */
scif_send_scif_register(struct scif_endpt * ep,struct scif_window * window)961 static int scif_send_scif_register(struct scif_endpt *ep,
962 struct scif_window *window)
963 {
964 int err = 0;
965 struct scifmsg msg;
966
967 msg.src = ep->port;
968 msg.payload[0] = ep->remote_ep;
969 msg.payload[1] = window->alloc_handle.vaddr;
970 msg.payload[2] = (u64)window;
971 spin_lock(&ep->lock);
972 if (ep->state == SCIFEP_CONNECTED) {
973 msg.uop = SCIF_REGISTER;
974 window->reg_state = OP_IN_PROGRESS;
975 err = _scif_nodeqp_send(ep->remote_dev, &msg);
976 spin_unlock(&ep->lock);
977 if (!err) {
978 retry:
979 /* Wait for a SCIF_REGISTER_(N)ACK message */
980 err = wait_event_timeout(window->regwq,
981 window->reg_state !=
982 OP_IN_PROGRESS,
983 SCIF_NODE_ALIVE_TIMEOUT);
984 if (!err && scifdev_alive(ep))
985 goto retry;
986 err = !err ? -ENODEV : 0;
987 if (window->reg_state == OP_FAILED)
988 err = -ENOTCONN;
989 }
990 } else {
991 msg.uop = SCIF_FREE_VIRT;
992 msg.payload[3] = SCIF_REGISTER;
993 err = _scif_nodeqp_send(ep->remote_dev, &msg);
994 spin_unlock(&ep->lock);
995 if (!err)
996 err = -ENOTCONN;
997 }
998 return err;
999 }
1000
1001 /**
1002 * scif_get_window_offset:
1003 * @ep: end point descriptor
1004 * @flags: flags
1005 * @offset: offset hint
1006 * @num_pages: number of pages
1007 * @out_offset: computed offset returned by reference.
1008 *
1009 * Compute/Claim a new offset for this EP.
1010 */
scif_get_window_offset(struct scif_endpt * ep,int flags,s64 offset,int num_pages,s64 * out_offset)1011 int scif_get_window_offset(struct scif_endpt *ep, int flags, s64 offset,
1012 int num_pages, s64 *out_offset)
1013 {
1014 s64 page_index;
1015 struct iova *iova_ptr;
1016 int err = 0;
1017
1018 if (flags & SCIF_MAP_FIXED) {
1019 page_index = SCIF_IOVA_PFN(offset);
1020 iova_ptr = reserve_iova(&ep->rma_info.iovad, page_index,
1021 page_index + num_pages - 1);
1022 if (!iova_ptr)
1023 err = -EADDRINUSE;
1024 } else {
1025 iova_ptr = alloc_iova(&ep->rma_info.iovad, num_pages,
1026 SCIF_DMA_63BIT_PFN - 1, 0);
1027 if (!iova_ptr)
1028 err = -ENOMEM;
1029 }
1030 if (!err)
1031 *out_offset = (iova_ptr->pfn_lo) << PAGE_SHIFT;
1032 return err;
1033 }
1034
1035 /**
1036 * scif_free_window_offset:
1037 * @ep: end point descriptor
1038 * @window: registration window
1039 * @offset: Offset to be freed
1040 *
1041 * Free offset for this EP. The callee is supposed to grab
1042 * the RMA mutex before calling this API.
1043 */
scif_free_window_offset(struct scif_endpt * ep,struct scif_window * window,s64 offset)1044 void scif_free_window_offset(struct scif_endpt *ep,
1045 struct scif_window *window, s64 offset)
1046 {
1047 if ((window && !window->offset_freed) || !window) {
1048 free_iova(&ep->rma_info.iovad, offset >> PAGE_SHIFT);
1049 if (window)
1050 window->offset_freed = true;
1051 }
1052 }
1053
1054 /**
1055 * scif_alloc_req: Respond to SCIF_ALLOC_REQ interrupt message
1056 * @msg: Interrupt message
1057 *
1058 * Remote side is requesting a memory allocation.
1059 */
scif_alloc_req(struct scif_dev * scifdev,struct scifmsg * msg)1060 void scif_alloc_req(struct scif_dev *scifdev, struct scifmsg *msg)
1061 {
1062 int err;
1063 struct scif_window *window = NULL;
1064 int nr_pages = msg->payload[1];
1065
1066 window = scif_create_remote_window(scifdev, nr_pages);
1067 if (!window) {
1068 err = -ENOMEM;
1069 goto error;
1070 }
1071
1072 /* The peer's allocation request is granted */
1073 msg->uop = SCIF_ALLOC_GNT;
1074 msg->payload[0] = (u64)window;
1075 msg->payload[1] = window->mapped_offset;
1076 err = scif_nodeqp_send(scifdev, msg);
1077 if (err)
1078 scif_destroy_remote_window(window);
1079 return;
1080 error:
1081 /* The peer's allocation request is rejected */
1082 dev_err(&scifdev->sdev->dev,
1083 "%s %d error %d alloc_ptr %p nr_pages 0x%x\n",
1084 __func__, __LINE__, err, window, nr_pages);
1085 msg->uop = SCIF_ALLOC_REJ;
1086 scif_nodeqp_send(scifdev, msg);
1087 }
1088
1089 /**
1090 * scif_alloc_gnt_rej: Respond to SCIF_ALLOC_GNT/REJ interrupt message
1091 * @msg: Interrupt message
1092 *
1093 * Remote side responded to a memory allocation.
1094 */
scif_alloc_gnt_rej(struct scif_dev * scifdev,struct scifmsg * msg)1095 void scif_alloc_gnt_rej(struct scif_dev *scifdev, struct scifmsg *msg)
1096 {
1097 struct scif_allocmsg *handle = (struct scif_allocmsg *)msg->payload[2];
1098 struct scif_window *window = container_of(handle, struct scif_window,
1099 alloc_handle);
1100 struct scif_endpt *ep = (struct scif_endpt *)window->ep;
1101
1102 mutex_lock(&ep->rma_info.rma_lock);
1103 handle->vaddr = msg->payload[0];
1104 handle->phys_addr = msg->payload[1];
1105 if (msg->uop == SCIF_ALLOC_GNT)
1106 handle->state = OP_COMPLETED;
1107 else
1108 handle->state = OP_FAILED;
1109 wake_up(&handle->allocwq);
1110 mutex_unlock(&ep->rma_info.rma_lock);
1111 }
1112
1113 /**
1114 * scif_free_virt: Respond to SCIF_FREE_VIRT interrupt message
1115 * @msg: Interrupt message
1116 *
1117 * Free up memory kmalloc'd earlier.
1118 */
scif_free_virt(struct scif_dev * scifdev,struct scifmsg * msg)1119 void scif_free_virt(struct scif_dev *scifdev, struct scifmsg *msg)
1120 {
1121 struct scif_window *window = (struct scif_window *)msg->payload[1];
1122
1123 scif_destroy_remote_window(window);
1124 }
1125
1126 static void
scif_fixup_aper_base(struct scif_dev * dev,struct scif_window * window)1127 scif_fixup_aper_base(struct scif_dev *dev, struct scif_window *window)
1128 {
1129 int j;
1130 struct scif_hw_dev *sdev = dev->sdev;
1131 phys_addr_t apt_base = 0;
1132
1133 /*
1134 * Add the aperture base if the DMA address is not card relative
1135 * since the DMA addresses need to be an offset into the bar
1136 */
1137 if (!scifdev_self(dev) && window->type == SCIF_WINDOW_PEER &&
1138 sdev->aper && !sdev->card_rel_da)
1139 apt_base = sdev->aper->pa;
1140 else
1141 return;
1142
1143 for (j = 0; j < window->nr_contig_chunks; j++) {
1144 if (window->num_pages[j])
1145 window->dma_addr[j] += apt_base;
1146 else
1147 break;
1148 }
1149 }
1150
1151 /**
1152 * scif_recv_reg: Respond to SCIF_REGISTER interrupt message
1153 * @msg: Interrupt message
1154 *
1155 * Update remote window list with a new registered window.
1156 */
scif_recv_reg(struct scif_dev * scifdev,struct scifmsg * msg)1157 void scif_recv_reg(struct scif_dev *scifdev, struct scifmsg *msg)
1158 {
1159 struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
1160 struct scif_window *window =
1161 (struct scif_window *)msg->payload[1];
1162
1163 mutex_lock(&ep->rma_info.rma_lock);
1164 spin_lock(&ep->lock);
1165 if (ep->state == SCIFEP_CONNECTED) {
1166 msg->uop = SCIF_REGISTER_ACK;
1167 scif_nodeqp_send(ep->remote_dev, msg);
1168 scif_fixup_aper_base(ep->remote_dev, window);
1169 /* No further failures expected. Insert new window */
1170 scif_insert_window(window, &ep->rma_info.remote_reg_list);
1171 } else {
1172 msg->uop = SCIF_REGISTER_NACK;
1173 scif_nodeqp_send(ep->remote_dev, msg);
1174 }
1175 spin_unlock(&ep->lock);
1176 mutex_unlock(&ep->rma_info.rma_lock);
1177 /* free up any lookup resources now that page lists are transferred */
1178 scif_destroy_remote_lookup(ep->remote_dev, window);
1179 /*
1180 * We could not insert the window but we need to
1181 * destroy the window.
1182 */
1183 if (msg->uop == SCIF_REGISTER_NACK)
1184 scif_destroy_remote_window(window);
1185 }
1186
1187 /**
1188 * scif_recv_unreg: Respond to SCIF_UNREGISTER interrupt message
1189 * @msg: Interrupt message
1190 *
1191 * Remove window from remote registration list;
1192 */
scif_recv_unreg(struct scif_dev * scifdev,struct scifmsg * msg)1193 void scif_recv_unreg(struct scif_dev *scifdev, struct scifmsg *msg)
1194 {
1195 struct scif_rma_req req;
1196 struct scif_window *window = NULL;
1197 struct scif_window *recv_window =
1198 (struct scif_window *)msg->payload[0];
1199 struct scif_endpt *ep;
1200 int del_window = 0;
1201
1202 ep = (struct scif_endpt *)recv_window->ep;
1203 req.out_window = &window;
1204 req.offset = recv_window->offset;
1205 req.prot = 0;
1206 req.nr_bytes = recv_window->nr_pages << PAGE_SHIFT;
1207 req.type = SCIF_WINDOW_FULL;
1208 req.head = &ep->rma_info.remote_reg_list;
1209 msg->payload[0] = ep->remote_ep;
1210
1211 mutex_lock(&ep->rma_info.rma_lock);
1212 /* Does a valid window exist? */
1213 if (scif_query_window(&req)) {
1214 dev_err(&scifdev->sdev->dev,
1215 "%s %d -ENXIO\n", __func__, __LINE__);
1216 msg->uop = SCIF_UNREGISTER_ACK;
1217 goto error;
1218 }
1219 if (window) {
1220 if (window->ref_count)
1221 scif_put_window(window, window->nr_pages);
1222 else
1223 dev_err(&scifdev->sdev->dev,
1224 "%s %d ref count should be +ve\n",
1225 __func__, __LINE__);
1226 window->unreg_state = OP_COMPLETED;
1227 if (!window->ref_count) {
1228 msg->uop = SCIF_UNREGISTER_ACK;
1229 atomic_inc(&ep->rma_info.tw_refcount);
1230 ep->rma_info.async_list_del = 1;
1231 list_del_init(&window->list);
1232 del_window = 1;
1233 } else {
1234 /* NACK! There are valid references to this window */
1235 msg->uop = SCIF_UNREGISTER_NACK;
1236 }
1237 } else {
1238 /* The window did not make its way to the list at all. ACK */
1239 msg->uop = SCIF_UNREGISTER_ACK;
1240 scif_destroy_remote_window(recv_window);
1241 }
1242 error:
1243 mutex_unlock(&ep->rma_info.rma_lock);
1244 if (del_window)
1245 scif_drain_dma_intr(ep->remote_dev->sdev,
1246 ep->rma_info.dma_chan);
1247 scif_nodeqp_send(ep->remote_dev, msg);
1248 if (del_window)
1249 scif_queue_for_cleanup(window, &scif_info.rma);
1250 }
1251
1252 /**
1253 * scif_recv_reg_ack: Respond to SCIF_REGISTER_ACK interrupt message
1254 * @msg: Interrupt message
1255 *
1256 * Wake up the window waiting to complete registration.
1257 */
scif_recv_reg_ack(struct scif_dev * scifdev,struct scifmsg * msg)1258 void scif_recv_reg_ack(struct scif_dev *scifdev, struct scifmsg *msg)
1259 {
1260 struct scif_window *window =
1261 (struct scif_window *)msg->payload[2];
1262 struct scif_endpt *ep = (struct scif_endpt *)window->ep;
1263
1264 mutex_lock(&ep->rma_info.rma_lock);
1265 window->reg_state = OP_COMPLETED;
1266 wake_up(&window->regwq);
1267 mutex_unlock(&ep->rma_info.rma_lock);
1268 }
1269
1270 /**
1271 * scif_recv_reg_nack: Respond to SCIF_REGISTER_NACK interrupt message
1272 * @msg: Interrupt message
1273 *
1274 * Wake up the window waiting to inform it that registration
1275 * cannot be completed.
1276 */
scif_recv_reg_nack(struct scif_dev * scifdev,struct scifmsg * msg)1277 void scif_recv_reg_nack(struct scif_dev *scifdev, struct scifmsg *msg)
1278 {
1279 struct scif_window *window =
1280 (struct scif_window *)msg->payload[2];
1281 struct scif_endpt *ep = (struct scif_endpt *)window->ep;
1282
1283 mutex_lock(&ep->rma_info.rma_lock);
1284 window->reg_state = OP_FAILED;
1285 wake_up(&window->regwq);
1286 mutex_unlock(&ep->rma_info.rma_lock);
1287 }
1288
1289 /**
1290 * scif_recv_unreg_ack: Respond to SCIF_UNREGISTER_ACK interrupt message
1291 * @msg: Interrupt message
1292 *
1293 * Wake up the window waiting to complete unregistration.
1294 */
scif_recv_unreg_ack(struct scif_dev * scifdev,struct scifmsg * msg)1295 void scif_recv_unreg_ack(struct scif_dev *scifdev, struct scifmsg *msg)
1296 {
1297 struct scif_window *window =
1298 (struct scif_window *)msg->payload[1];
1299 struct scif_endpt *ep = (struct scif_endpt *)window->ep;
1300
1301 mutex_lock(&ep->rma_info.rma_lock);
1302 window->unreg_state = OP_COMPLETED;
1303 wake_up(&window->unregwq);
1304 mutex_unlock(&ep->rma_info.rma_lock);
1305 }
1306
1307 /**
1308 * scif_recv_unreg_nack: Respond to SCIF_UNREGISTER_NACK interrupt message
1309 * @msg: Interrupt message
1310 *
1311 * Wake up the window waiting to inform it that unregistration
1312 * cannot be completed immediately.
1313 */
scif_recv_unreg_nack(struct scif_dev * scifdev,struct scifmsg * msg)1314 void scif_recv_unreg_nack(struct scif_dev *scifdev, struct scifmsg *msg)
1315 {
1316 struct scif_window *window =
1317 (struct scif_window *)msg->payload[1];
1318 struct scif_endpt *ep = (struct scif_endpt *)window->ep;
1319
1320 mutex_lock(&ep->rma_info.rma_lock);
1321 window->unreg_state = OP_FAILED;
1322 wake_up(&window->unregwq);
1323 mutex_unlock(&ep->rma_info.rma_lock);
1324 }
1325
__scif_pin_pages(void * addr,size_t len,int * out_prot,int map_flags,scif_pinned_pages_t * pages)1326 int __scif_pin_pages(void *addr, size_t len, int *out_prot,
1327 int map_flags, scif_pinned_pages_t *pages)
1328 {
1329 struct scif_pinned_pages *pinned_pages;
1330 int nr_pages, err = 0, i;
1331 bool vmalloc_addr = false;
1332 bool try_upgrade = false;
1333 int prot = *out_prot;
1334 int ulimit = 0;
1335 struct mm_struct *mm = NULL;
1336
1337 /* Unsupported flags */
1338 if (map_flags & ~(SCIF_MAP_KERNEL | SCIF_MAP_ULIMIT))
1339 return -EINVAL;
1340 ulimit = !!(map_flags & SCIF_MAP_ULIMIT);
1341
1342 /* Unsupported protection requested */
1343 if (prot & ~(SCIF_PROT_READ | SCIF_PROT_WRITE))
1344 return -EINVAL;
1345
1346 /* addr/len must be page aligned. len should be non zero */
1347 if (!len ||
1348 (ALIGN((u64)addr, PAGE_SIZE) != (u64)addr) ||
1349 (ALIGN((u64)len, PAGE_SIZE) != (u64)len))
1350 return -EINVAL;
1351
1352 might_sleep();
1353
1354 nr_pages = len >> PAGE_SHIFT;
1355
1356 /* Allocate a set of pinned pages */
1357 pinned_pages = scif_create_pinned_pages(nr_pages, prot);
1358 if (!pinned_pages)
1359 return -ENOMEM;
1360
1361 if (map_flags & SCIF_MAP_KERNEL) {
1362 if (is_vmalloc_addr(addr))
1363 vmalloc_addr = true;
1364
1365 for (i = 0; i < nr_pages; i++) {
1366 if (vmalloc_addr)
1367 pinned_pages->pages[i] =
1368 vmalloc_to_page(addr + (i * PAGE_SIZE));
1369 else
1370 pinned_pages->pages[i] =
1371 virt_to_page(addr + (i * PAGE_SIZE));
1372 }
1373 pinned_pages->nr_pages = nr_pages;
1374 pinned_pages->map_flags = SCIF_MAP_KERNEL;
1375 } else {
1376 /*
1377 * SCIF supports registration caching. If a registration has
1378 * been requested with read only permissions, then we try
1379 * to pin the pages with RW permissions so that a subsequent
1380 * transfer with RW permission can hit the cache instead of
1381 * invalidating it. If the upgrade fails with RW then we
1382 * revert back to R permission and retry
1383 */
1384 if (prot == SCIF_PROT_READ)
1385 try_upgrade = true;
1386 prot |= SCIF_PROT_WRITE;
1387 retry:
1388 mm = current->mm;
1389 down_write(&mm->mmap_sem);
1390 if (ulimit) {
1391 err = __scif_check_inc_pinned_vm(mm, nr_pages);
1392 if (err) {
1393 up_write(&mm->mmap_sem);
1394 pinned_pages->nr_pages = 0;
1395 goto error_unmap;
1396 }
1397 }
1398
1399 pinned_pages->nr_pages = get_user_pages(
1400 (u64)addr,
1401 nr_pages,
1402 (prot & SCIF_PROT_WRITE) ? FOLL_WRITE : 0,
1403 pinned_pages->pages,
1404 NULL);
1405 up_write(&mm->mmap_sem);
1406 if (nr_pages != pinned_pages->nr_pages) {
1407 if (try_upgrade) {
1408 if (ulimit)
1409 __scif_dec_pinned_vm_lock(mm,
1410 nr_pages, 0);
1411 /* Roll back any pinned pages */
1412 for (i = 0; i < pinned_pages->nr_pages; i++) {
1413 if (pinned_pages->pages[i])
1414 put_page(
1415 pinned_pages->pages[i]);
1416 }
1417 prot &= ~SCIF_PROT_WRITE;
1418 try_upgrade = false;
1419 goto retry;
1420 }
1421 }
1422 pinned_pages->map_flags = 0;
1423 }
1424
1425 if (pinned_pages->nr_pages < nr_pages) {
1426 err = -EFAULT;
1427 pinned_pages->nr_pages = nr_pages;
1428 goto dec_pinned;
1429 }
1430
1431 *out_prot = prot;
1432 atomic_set(&pinned_pages->ref_count, 1);
1433 *pages = pinned_pages;
1434 return err;
1435 dec_pinned:
1436 if (ulimit)
1437 __scif_dec_pinned_vm_lock(mm, nr_pages, 0);
1438 /* Something went wrong! Rollback */
1439 error_unmap:
1440 pinned_pages->nr_pages = nr_pages;
1441 scif_destroy_pinned_pages(pinned_pages);
1442 *pages = NULL;
1443 dev_dbg(scif_info.mdev.this_device,
1444 "%s %d err %d len 0x%lx\n", __func__, __LINE__, err, len);
1445 return err;
1446 }
1447
scif_pin_pages(void * addr,size_t len,int prot,int map_flags,scif_pinned_pages_t * pages)1448 int scif_pin_pages(void *addr, size_t len, int prot,
1449 int map_flags, scif_pinned_pages_t *pages)
1450 {
1451 return __scif_pin_pages(addr, len, &prot, map_flags, pages);
1452 }
1453 EXPORT_SYMBOL_GPL(scif_pin_pages);
1454
scif_unpin_pages(scif_pinned_pages_t pinned_pages)1455 int scif_unpin_pages(scif_pinned_pages_t pinned_pages)
1456 {
1457 int err = 0, ret;
1458
1459 if (!pinned_pages || SCIFEP_MAGIC != pinned_pages->magic)
1460 return -EINVAL;
1461
1462 ret = atomic_sub_return(1, &pinned_pages->ref_count);
1463 if (ret < 0) {
1464 dev_err(scif_info.mdev.this_device,
1465 "%s %d scif_unpin_pages called without pinning? rc %d\n",
1466 __func__, __LINE__, ret);
1467 return -EINVAL;
1468 }
1469 /*
1470 * Destroy the window if the ref count for this set of pinned
1471 * pages has dropped to zero. If it is positive then there is
1472 * a valid registered window which is backed by these pages and
1473 * it will be destroyed once all such windows are unregistered.
1474 */
1475 if (!ret)
1476 err = scif_destroy_pinned_pages(pinned_pages);
1477
1478 return err;
1479 }
1480 EXPORT_SYMBOL_GPL(scif_unpin_pages);
1481
1482 static inline void
scif_insert_local_window(struct scif_window * window,struct scif_endpt * ep)1483 scif_insert_local_window(struct scif_window *window, struct scif_endpt *ep)
1484 {
1485 mutex_lock(&ep->rma_info.rma_lock);
1486 scif_insert_window(window, &ep->rma_info.reg_list);
1487 mutex_unlock(&ep->rma_info.rma_lock);
1488 }
1489
scif_register_pinned_pages(scif_epd_t epd,scif_pinned_pages_t pinned_pages,off_t offset,int map_flags)1490 off_t scif_register_pinned_pages(scif_epd_t epd,
1491 scif_pinned_pages_t pinned_pages,
1492 off_t offset, int map_flags)
1493 {
1494 struct scif_endpt *ep = (struct scif_endpt *)epd;
1495 s64 computed_offset;
1496 struct scif_window *window;
1497 int err;
1498 size_t len;
1499 struct device *spdev;
1500
1501 /* Unsupported flags */
1502 if (map_flags & ~SCIF_MAP_FIXED)
1503 return -EINVAL;
1504
1505 len = pinned_pages->nr_pages << PAGE_SHIFT;
1506
1507 /*
1508 * Offset is not page aligned/negative or offset+len
1509 * wraps around with SCIF_MAP_FIXED.
1510 */
1511 if ((map_flags & SCIF_MAP_FIXED) &&
1512 ((ALIGN(offset, PAGE_SIZE) != offset) ||
1513 (offset < 0) ||
1514 (len > LONG_MAX - offset)))
1515 return -EINVAL;
1516
1517 might_sleep();
1518
1519 err = scif_verify_epd(ep);
1520 if (err)
1521 return err;
1522 /*
1523 * It is an error to pass pinned_pages to scif_register_pinned_pages()
1524 * after calling scif_unpin_pages().
1525 */
1526 if (!atomic_add_unless(&pinned_pages->ref_count, 1, 0))
1527 return -EINVAL;
1528
1529 /* Compute the offset for this registration */
1530 err = scif_get_window_offset(ep, map_flags, offset,
1531 len, &computed_offset);
1532 if (err) {
1533 atomic_sub(1, &pinned_pages->ref_count);
1534 return err;
1535 }
1536
1537 /* Allocate and prepare self registration window */
1538 window = scif_create_window(ep, pinned_pages->nr_pages,
1539 computed_offset, false);
1540 if (!window) {
1541 atomic_sub(1, &pinned_pages->ref_count);
1542 scif_free_window_offset(ep, NULL, computed_offset);
1543 return -ENOMEM;
1544 }
1545
1546 window->pinned_pages = pinned_pages;
1547 window->nr_pages = pinned_pages->nr_pages;
1548 window->prot = pinned_pages->prot;
1549
1550 spdev = scif_get_peer_dev(ep->remote_dev);
1551 if (IS_ERR(spdev)) {
1552 err = PTR_ERR(spdev);
1553 scif_destroy_window(ep, window);
1554 return err;
1555 }
1556 err = scif_send_alloc_request(ep, window);
1557 if (err) {
1558 dev_err(&ep->remote_dev->sdev->dev,
1559 "%s %d err %d\n", __func__, __LINE__, err);
1560 goto error_unmap;
1561 }
1562
1563 /* Prepare the remote registration window */
1564 err = scif_prep_remote_window(ep, window);
1565 if (err) {
1566 dev_err(&ep->remote_dev->sdev->dev,
1567 "%s %d err %d\n", __func__, __LINE__, err);
1568 goto error_unmap;
1569 }
1570
1571 /* Tell the peer about the new window */
1572 err = scif_send_scif_register(ep, window);
1573 if (err) {
1574 dev_err(&ep->remote_dev->sdev->dev,
1575 "%s %d err %d\n", __func__, __LINE__, err);
1576 goto error_unmap;
1577 }
1578
1579 scif_put_peer_dev(spdev);
1580 /* No further failures expected. Insert new window */
1581 scif_insert_local_window(window, ep);
1582 return computed_offset;
1583 error_unmap:
1584 scif_destroy_window(ep, window);
1585 scif_put_peer_dev(spdev);
1586 dev_err(&ep->remote_dev->sdev->dev,
1587 "%s %d err %d\n", __func__, __LINE__, err);
1588 return err;
1589 }
1590 EXPORT_SYMBOL_GPL(scif_register_pinned_pages);
1591
scif_register(scif_epd_t epd,void * addr,size_t len,off_t offset,int prot,int map_flags)1592 off_t scif_register(scif_epd_t epd, void *addr, size_t len, off_t offset,
1593 int prot, int map_flags)
1594 {
1595 scif_pinned_pages_t pinned_pages;
1596 off_t err;
1597 struct scif_endpt *ep = (struct scif_endpt *)epd;
1598 s64 computed_offset;
1599 struct scif_window *window;
1600 struct mm_struct *mm = NULL;
1601 struct device *spdev;
1602
1603 dev_dbg(scif_info.mdev.this_device,
1604 "SCIFAPI register: ep %p addr %p len 0x%lx offset 0x%lx prot 0x%x map_flags 0x%x\n",
1605 epd, addr, len, offset, prot, map_flags);
1606 /* Unsupported flags */
1607 if (map_flags & ~(SCIF_MAP_FIXED | SCIF_MAP_KERNEL))
1608 return -EINVAL;
1609
1610 /*
1611 * Offset is not page aligned/negative or offset+len
1612 * wraps around with SCIF_MAP_FIXED.
1613 */
1614 if ((map_flags & SCIF_MAP_FIXED) &&
1615 ((ALIGN(offset, PAGE_SIZE) != offset) ||
1616 (offset < 0) ||
1617 (len > LONG_MAX - offset)))
1618 return -EINVAL;
1619
1620 /* Unsupported protection requested */
1621 if (prot & ~(SCIF_PROT_READ | SCIF_PROT_WRITE))
1622 return -EINVAL;
1623
1624 /* addr/len must be page aligned. len should be non zero */
1625 if (!len || (ALIGN((u64)addr, PAGE_SIZE) != (u64)addr) ||
1626 (ALIGN(len, PAGE_SIZE) != len))
1627 return -EINVAL;
1628
1629 might_sleep();
1630
1631 err = scif_verify_epd(ep);
1632 if (err)
1633 return err;
1634
1635 /* Compute the offset for this registration */
1636 err = scif_get_window_offset(ep, map_flags, offset,
1637 len >> PAGE_SHIFT, &computed_offset);
1638 if (err)
1639 return err;
1640
1641 spdev = scif_get_peer_dev(ep->remote_dev);
1642 if (IS_ERR(spdev)) {
1643 err = PTR_ERR(spdev);
1644 scif_free_window_offset(ep, NULL, computed_offset);
1645 return err;
1646 }
1647 /* Allocate and prepare self registration window */
1648 window = scif_create_window(ep, len >> PAGE_SHIFT,
1649 computed_offset, false);
1650 if (!window) {
1651 scif_free_window_offset(ep, NULL, computed_offset);
1652 scif_put_peer_dev(spdev);
1653 return -ENOMEM;
1654 }
1655
1656 window->nr_pages = len >> PAGE_SHIFT;
1657
1658 err = scif_send_alloc_request(ep, window);
1659 if (err) {
1660 scif_destroy_incomplete_window(ep, window);
1661 scif_put_peer_dev(spdev);
1662 return err;
1663 }
1664
1665 if (!(map_flags & SCIF_MAP_KERNEL)) {
1666 mm = __scif_acquire_mm();
1667 map_flags |= SCIF_MAP_ULIMIT;
1668 }
1669 /* Pin down the pages */
1670 err = __scif_pin_pages(addr, len, &prot,
1671 map_flags & (SCIF_MAP_KERNEL | SCIF_MAP_ULIMIT),
1672 &pinned_pages);
1673 if (err) {
1674 scif_destroy_incomplete_window(ep, window);
1675 __scif_release_mm(mm);
1676 goto error;
1677 }
1678
1679 window->pinned_pages = pinned_pages;
1680 window->prot = pinned_pages->prot;
1681 window->mm = mm;
1682
1683 /* Prepare the remote registration window */
1684 err = scif_prep_remote_window(ep, window);
1685 if (err) {
1686 dev_err(&ep->remote_dev->sdev->dev,
1687 "%s %d err %ld\n", __func__, __LINE__, err);
1688 goto error_unmap;
1689 }
1690
1691 /* Tell the peer about the new window */
1692 err = scif_send_scif_register(ep, window);
1693 if (err) {
1694 dev_err(&ep->remote_dev->sdev->dev,
1695 "%s %d err %ld\n", __func__, __LINE__, err);
1696 goto error_unmap;
1697 }
1698
1699 scif_put_peer_dev(spdev);
1700 /* No further failures expected. Insert new window */
1701 scif_insert_local_window(window, ep);
1702 dev_dbg(&ep->remote_dev->sdev->dev,
1703 "SCIFAPI register: ep %p addr %p len 0x%lx computed_offset 0x%llx\n",
1704 epd, addr, len, computed_offset);
1705 return computed_offset;
1706 error_unmap:
1707 scif_destroy_window(ep, window);
1708 error:
1709 scif_put_peer_dev(spdev);
1710 dev_err(&ep->remote_dev->sdev->dev,
1711 "%s %d err %ld\n", __func__, __LINE__, err);
1712 return err;
1713 }
1714 EXPORT_SYMBOL_GPL(scif_register);
1715
1716 int
scif_unregister(scif_epd_t epd,off_t offset,size_t len)1717 scif_unregister(scif_epd_t epd, off_t offset, size_t len)
1718 {
1719 struct scif_endpt *ep = (struct scif_endpt *)epd;
1720 struct scif_window *window = NULL;
1721 struct scif_rma_req req;
1722 int nr_pages, err;
1723 struct device *spdev;
1724
1725 dev_dbg(scif_info.mdev.this_device,
1726 "SCIFAPI unregister: ep %p offset 0x%lx len 0x%lx\n",
1727 ep, offset, len);
1728 /* len must be page aligned. len should be non zero */
1729 if (!len ||
1730 (ALIGN((u64)len, PAGE_SIZE) != (u64)len))
1731 return -EINVAL;
1732
1733 /* Offset is not page aligned or offset+len wraps around */
1734 if ((ALIGN(offset, PAGE_SIZE) != offset) ||
1735 (offset < 0) ||
1736 (len > LONG_MAX - offset))
1737 return -EINVAL;
1738
1739 err = scif_verify_epd(ep);
1740 if (err)
1741 return err;
1742
1743 might_sleep();
1744 nr_pages = len >> PAGE_SHIFT;
1745
1746 req.out_window = &window;
1747 req.offset = offset;
1748 req.prot = 0;
1749 req.nr_bytes = len;
1750 req.type = SCIF_WINDOW_FULL;
1751 req.head = &ep->rma_info.reg_list;
1752
1753 spdev = scif_get_peer_dev(ep->remote_dev);
1754 if (IS_ERR(spdev)) {
1755 err = PTR_ERR(spdev);
1756 return err;
1757 }
1758 mutex_lock(&ep->rma_info.rma_lock);
1759 /* Does a valid window exist? */
1760 err = scif_query_window(&req);
1761 if (err) {
1762 dev_err(&ep->remote_dev->sdev->dev,
1763 "%s %d err %d\n", __func__, __LINE__, err);
1764 goto error;
1765 }
1766 /* Unregister all the windows in this range */
1767 err = scif_rma_list_unregister(window, offset, nr_pages);
1768 if (err)
1769 dev_err(&ep->remote_dev->sdev->dev,
1770 "%s %d err %d\n", __func__, __LINE__, err);
1771 error:
1772 mutex_unlock(&ep->rma_info.rma_lock);
1773 scif_put_peer_dev(spdev);
1774 return err;
1775 }
1776 EXPORT_SYMBOL_GPL(scif_unregister);
1777