1 /*
2  * Copyright(c) 2015-2018 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 #include <asm/page.h>
48 #include <linux/string.h>
49 
50 #include "mmu_rb.h"
51 #include "user_exp_rcv.h"
52 #include "trace.h"
53 
54 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
55 			    struct exp_tid_set *set,
56 			    struct hfi1_filedata *fd);
57 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages);
58 static int set_rcvarray_entry(struct hfi1_filedata *fd,
59 			      struct tid_user_buf *tbuf,
60 			      u32 rcventry, struct tid_group *grp,
61 			      u16 pageidx, unsigned int npages);
62 static int tid_rb_insert(void *arg, struct mmu_rb_node *node);
63 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
64 				    struct tid_rb_node *tnode);
65 static void tid_rb_remove(void *arg, struct mmu_rb_node *node);
66 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
67 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *,
68 			    struct tid_group *grp,
69 			    unsigned int start, u16 count,
70 			    u32 *tidlist, unsigned int *tididx,
71 			    unsigned int *pmapped);
72 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
73 			      struct tid_group **grp);
74 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
75 
76 static struct mmu_rb_ops tid_rb_ops = {
77 	.insert = tid_rb_insert,
78 	.remove = tid_rb_remove,
79 	.invalidate = tid_rb_invalidate
80 };
81 
82 /*
83  * Initialize context and file private data needed for Expected
84  * receive caching. This needs to be done after the context has
85  * been configured with the eager/expected RcvEntry counts.
86  */
87 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd,
88 			   struct hfi1_ctxtdata *uctxt)
89 {
90 	struct hfi1_devdata *dd = uctxt->dd;
91 	int ret = 0;
92 
93 	spin_lock_init(&fd->tid_lock);
94 	spin_lock_init(&fd->invalid_lock);
95 
96 	fd->entry_to_rb = kcalloc(uctxt->expected_count,
97 				  sizeof(struct rb_node *),
98 				  GFP_KERNEL);
99 	if (!fd->entry_to_rb)
100 		return -ENOMEM;
101 
102 	if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
103 		fd->invalid_tid_idx = 0;
104 		fd->invalid_tids = kcalloc(uctxt->expected_count,
105 					   sizeof(*fd->invalid_tids),
106 					   GFP_KERNEL);
107 		if (!fd->invalid_tids) {
108 			kfree(fd->entry_to_rb);
109 			fd->entry_to_rb = NULL;
110 			return -ENOMEM;
111 		}
112 
113 		/*
114 		 * Register MMU notifier callbacks. If the registration
115 		 * fails, continue without TID caching for this context.
116 		 */
117 		ret = hfi1_mmu_rb_register(fd, fd->mm, &tid_rb_ops,
118 					   dd->pport->hfi1_wq,
119 					   &fd->handler);
120 		if (ret) {
121 			dd_dev_info(dd,
122 				    "Failed MMU notifier registration %d\n",
123 				    ret);
124 			ret = 0;
125 		}
126 	}
127 
128 	/*
129 	 * PSM does not have a good way to separate, count, and
130 	 * effectively enforce a limit on RcvArray entries used by
131 	 * subctxts (when context sharing is used) when TID caching
132 	 * is enabled. To help with that, we calculate a per-process
133 	 * RcvArray entry share and enforce that.
134 	 * If TID caching is not in use, PSM deals with usage on its
135 	 * own. In that case, we allow any subctxt to take all of the
136 	 * entries.
137 	 *
138 	 * Make sure that we set the tid counts only after successful
139 	 * init.
140 	 */
141 	spin_lock(&fd->tid_lock);
142 	if (uctxt->subctxt_cnt && fd->handler) {
143 		u16 remainder;
144 
145 		fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
146 		remainder = uctxt->expected_count % uctxt->subctxt_cnt;
147 		if (remainder && fd->subctxt < remainder)
148 			fd->tid_limit++;
149 	} else {
150 		fd->tid_limit = uctxt->expected_count;
151 	}
152 	spin_unlock(&fd->tid_lock);
153 
154 	return ret;
155 }
156 
157 void hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
158 {
159 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
160 
161 	/*
162 	 * The notifier would have been removed when the process'es mm
163 	 * was freed.
164 	 */
165 	if (fd->handler) {
166 		hfi1_mmu_rb_unregister(fd->handler);
167 	} else {
168 		if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
169 			unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
170 		if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
171 			unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
172 	}
173 
174 	kfree(fd->invalid_tids);
175 	fd->invalid_tids = NULL;
176 
177 	kfree(fd->entry_to_rb);
178 	fd->entry_to_rb = NULL;
179 }
180 
181 /**
182  * Release pinned receive buffer pages.
183  *
184  * @mapped - true if the pages have been DMA mapped. false otherwise.
185  * @idx - Index of the first page to unpin.
186  * @npages - No of pages to unpin.
187  *
188  * If the pages have been DMA mapped (indicated by mapped parameter), their
189  * info will be passed via a struct tid_rb_node. If they haven't been mapped,
190  * their info will be passed via a struct tid_user_buf.
191  */
192 static void unpin_rcv_pages(struct hfi1_filedata *fd,
193 			    struct tid_user_buf *tidbuf,
194 			    struct tid_rb_node *node,
195 			    unsigned int idx,
196 			    unsigned int npages,
197 			    bool mapped)
198 {
199 	struct page **pages;
200 	struct hfi1_devdata *dd = fd->uctxt->dd;
201 
202 	if (mapped) {
203 		pci_unmap_single(dd->pcidev, node->dma_addr,
204 				 node->mmu.len, PCI_DMA_FROMDEVICE);
205 		pages = &node->pages[idx];
206 	} else {
207 		pages = &tidbuf->pages[idx];
208 	}
209 	hfi1_release_user_pages(fd->mm, pages, npages, mapped);
210 	fd->tid_n_pinned -= npages;
211 }
212 
213 /**
214  * Pin receive buffer pages.
215  */
216 static int pin_rcv_pages(struct hfi1_filedata *fd, struct tid_user_buf *tidbuf)
217 {
218 	int pinned;
219 	unsigned int npages;
220 	unsigned long vaddr = tidbuf->vaddr;
221 	struct page **pages = NULL;
222 	struct hfi1_devdata *dd = fd->uctxt->dd;
223 
224 	/* Get the number of pages the user buffer spans */
225 	npages = num_user_pages(vaddr, tidbuf->length);
226 	if (!npages)
227 		return -EINVAL;
228 
229 	if (npages > fd->uctxt->expected_count) {
230 		dd_dev_err(dd, "Expected buffer too big\n");
231 		return -EINVAL;
232 	}
233 
234 	/* Verify that access is OK for the user buffer */
235 	if (!access_ok(VERIFY_WRITE, (void __user *)vaddr,
236 		       npages * PAGE_SIZE)) {
237 		dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n",
238 			   (void *)vaddr, npages);
239 		return -EFAULT;
240 	}
241 	/* Allocate the array of struct page pointers needed for pinning */
242 	pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
243 	if (!pages)
244 		return -ENOMEM;
245 
246 	/*
247 	 * Pin all the pages of the user buffer. If we can't pin all the
248 	 * pages, accept the amount pinned so far and program only that.
249 	 * User space knows how to deal with partially programmed buffers.
250 	 */
251 	if (!hfi1_can_pin_pages(dd, fd->mm, fd->tid_n_pinned, npages)) {
252 		kfree(pages);
253 		return -ENOMEM;
254 	}
255 
256 	pinned = hfi1_acquire_user_pages(fd->mm, vaddr, npages, true, pages);
257 	if (pinned <= 0) {
258 		kfree(pages);
259 		return pinned;
260 	}
261 	tidbuf->pages = pages;
262 	tidbuf->npages = npages;
263 	fd->tid_n_pinned += pinned;
264 	return pinned;
265 }
266 
267 /*
268  * RcvArray entry allocation for Expected Receives is done by the
269  * following algorithm:
270  *
271  * The context keeps 3 lists of groups of RcvArray entries:
272  *   1. List of empty groups - tid_group_list
273  *      This list is created during user context creation and
274  *      contains elements which describe sets (of 8) of empty
275  *      RcvArray entries.
276  *   2. List of partially used groups - tid_used_list
277  *      This list contains sets of RcvArray entries which are
278  *      not completely used up. Another mapping request could
279  *      use some of all of the remaining entries.
280  *   3. List of full groups - tid_full_list
281  *      This is the list where sets that are completely used
282  *      up go.
283  *
284  * An attempt to optimize the usage of RcvArray entries is
285  * made by finding all sets of physically contiguous pages in a
286  * user's buffer.
287  * These physically contiguous sets are further split into
288  * sizes supported by the receive engine of the HFI. The
289  * resulting sets of pages are stored in struct tid_pageset,
290  * which describes the sets as:
291  *    * .count - number of pages in this set
292  *    * .idx - starting index into struct page ** array
293  *                    of this set
294  *
295  * From this point on, the algorithm deals with the page sets
296  * described above. The number of pagesets is divided by the
297  * RcvArray group size to produce the number of full groups
298  * needed.
299  *
300  * Groups from the 3 lists are manipulated using the following
301  * rules:
302  *   1. For each set of 8 pagesets, a complete group from
303  *      tid_group_list is taken, programmed, and moved to
304  *      the tid_full_list list.
305  *   2. For all remaining pagesets:
306  *      2.1 If the tid_used_list is empty and the tid_group_list
307  *          is empty, stop processing pageset and return only
308  *          what has been programmed up to this point.
309  *      2.2 If the tid_used_list is empty and the tid_group_list
310  *          is not empty, move a group from tid_group_list to
311  *          tid_used_list.
312  *      2.3 For each group is tid_used_group, program as much as
313  *          can fit into the group. If the group becomes fully
314  *          used, move it to tid_full_list.
315  */
316 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
317 			    struct hfi1_tid_info *tinfo)
318 {
319 	int ret = 0, need_group = 0, pinned;
320 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
321 	struct hfi1_devdata *dd = uctxt->dd;
322 	unsigned int ngroups, pageidx = 0, pageset_count,
323 		tididx = 0, mapped, mapped_pages = 0;
324 	u32 *tidlist = NULL;
325 	struct tid_user_buf *tidbuf;
326 
327 	tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL);
328 	if (!tidbuf)
329 		return -ENOMEM;
330 
331 	tidbuf->vaddr = tinfo->vaddr;
332 	tidbuf->length = tinfo->length;
333 	tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets),
334 				GFP_KERNEL);
335 	if (!tidbuf->psets) {
336 		kfree(tidbuf);
337 		return -ENOMEM;
338 	}
339 
340 	pinned = pin_rcv_pages(fd, tidbuf);
341 	if (pinned <= 0) {
342 		kfree(tidbuf->psets);
343 		kfree(tidbuf);
344 		return pinned;
345 	}
346 
347 	/* Find sets of physically contiguous pages */
348 	tidbuf->n_psets = find_phys_blocks(tidbuf, pinned);
349 
350 	/*
351 	 * We don't need to access this under a lock since tid_used is per
352 	 * process and the same process cannot be in hfi1_user_exp_rcv_clear()
353 	 * and hfi1_user_exp_rcv_setup() at the same time.
354 	 */
355 	spin_lock(&fd->tid_lock);
356 	if (fd->tid_used + tidbuf->n_psets > fd->tid_limit)
357 		pageset_count = fd->tid_limit - fd->tid_used;
358 	else
359 		pageset_count = tidbuf->n_psets;
360 	spin_unlock(&fd->tid_lock);
361 
362 	if (!pageset_count)
363 		goto bail;
364 
365 	ngroups = pageset_count / dd->rcv_entries.group_size;
366 	tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
367 	if (!tidlist) {
368 		ret = -ENOMEM;
369 		goto nomem;
370 	}
371 
372 	tididx = 0;
373 
374 	/*
375 	 * From this point on, we are going to be using shared (between master
376 	 * and subcontexts) context resources. We need to take the lock.
377 	 */
378 	mutex_lock(&uctxt->exp_mutex);
379 	/*
380 	 * The first step is to program the RcvArray entries which are complete
381 	 * groups.
382 	 */
383 	while (ngroups && uctxt->tid_group_list.count) {
384 		struct tid_group *grp =
385 			tid_group_pop(&uctxt->tid_group_list);
386 
387 		ret = program_rcvarray(fd, tidbuf, grp,
388 				       pageidx, dd->rcv_entries.group_size,
389 				       tidlist, &tididx, &mapped);
390 		/*
391 		 * If there was a failure to program the RcvArray
392 		 * entries for the entire group, reset the grp fields
393 		 * and add the grp back to the free group list.
394 		 */
395 		if (ret <= 0) {
396 			tid_group_add_tail(grp, &uctxt->tid_group_list);
397 			hfi1_cdbg(TID,
398 				  "Failed to program RcvArray group %d", ret);
399 			goto unlock;
400 		}
401 
402 		tid_group_add_tail(grp, &uctxt->tid_full_list);
403 		ngroups--;
404 		pageidx += ret;
405 		mapped_pages += mapped;
406 	}
407 
408 	while (pageidx < pageset_count) {
409 		struct tid_group *grp, *ptr;
410 		/*
411 		 * If we don't have any partially used tid groups, check
412 		 * if we have empty groups. If so, take one from there and
413 		 * put in the partially used list.
414 		 */
415 		if (!uctxt->tid_used_list.count || need_group) {
416 			if (!uctxt->tid_group_list.count)
417 				goto unlock;
418 
419 			grp = tid_group_pop(&uctxt->tid_group_list);
420 			tid_group_add_tail(grp, &uctxt->tid_used_list);
421 			need_group = 0;
422 		}
423 		/*
424 		 * There is an optimization opportunity here - instead of
425 		 * fitting as many page sets as we can, check for a group
426 		 * later on in the list that could fit all of them.
427 		 */
428 		list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
429 					 list) {
430 			unsigned use = min_t(unsigned, pageset_count - pageidx,
431 					     grp->size - grp->used);
432 
433 			ret = program_rcvarray(fd, tidbuf, grp,
434 					       pageidx, use, tidlist,
435 					       &tididx, &mapped);
436 			if (ret < 0) {
437 				hfi1_cdbg(TID,
438 					  "Failed to program RcvArray entries %d",
439 					  ret);
440 				goto unlock;
441 			} else if (ret > 0) {
442 				if (grp->used == grp->size)
443 					tid_group_move(grp,
444 						       &uctxt->tid_used_list,
445 						       &uctxt->tid_full_list);
446 				pageidx += ret;
447 				mapped_pages += mapped;
448 				need_group = 0;
449 				/* Check if we are done so we break out early */
450 				if (pageidx >= pageset_count)
451 					break;
452 			} else if (WARN_ON(ret == 0)) {
453 				/*
454 				 * If ret is 0, we did not program any entries
455 				 * into this group, which can only happen if
456 				 * we've screwed up the accounting somewhere.
457 				 * Warn and try to continue.
458 				 */
459 				need_group = 1;
460 			}
461 		}
462 	}
463 unlock:
464 	mutex_unlock(&uctxt->exp_mutex);
465 nomem:
466 	hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
467 		  mapped_pages, ret);
468 	if (tididx) {
469 		spin_lock(&fd->tid_lock);
470 		fd->tid_used += tididx;
471 		spin_unlock(&fd->tid_lock);
472 		tinfo->tidcnt = tididx;
473 		tinfo->length = mapped_pages * PAGE_SIZE;
474 
475 		if (copy_to_user(u64_to_user_ptr(tinfo->tidlist),
476 				 tidlist, sizeof(tidlist[0]) * tididx)) {
477 			/*
478 			 * On failure to copy to the user level, we need to undo
479 			 * everything done so far so we don't leak resources.
480 			 */
481 			tinfo->tidlist = (unsigned long)&tidlist;
482 			hfi1_user_exp_rcv_clear(fd, tinfo);
483 			tinfo->tidlist = 0;
484 			ret = -EFAULT;
485 			goto bail;
486 		}
487 	}
488 
489 	/*
490 	 * If not everything was mapped (due to insufficient RcvArray entries,
491 	 * for example), unpin all unmapped pages so we can pin them nex time.
492 	 */
493 	if (mapped_pages != pinned)
494 		unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages,
495 				(pinned - mapped_pages), false);
496 bail:
497 	kfree(tidbuf->psets);
498 	kfree(tidlist);
499 	kfree(tidbuf->pages);
500 	kfree(tidbuf);
501 	return ret > 0 ? 0 : ret;
502 }
503 
504 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
505 			    struct hfi1_tid_info *tinfo)
506 {
507 	int ret = 0;
508 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
509 	u32 *tidinfo;
510 	unsigned tididx;
511 
512 	if (unlikely(tinfo->tidcnt > fd->tid_used))
513 		return -EINVAL;
514 
515 	tidinfo = memdup_user(u64_to_user_ptr(tinfo->tidlist),
516 			      sizeof(tidinfo[0]) * tinfo->tidcnt);
517 	if (IS_ERR(tidinfo))
518 		return PTR_ERR(tidinfo);
519 
520 	mutex_lock(&uctxt->exp_mutex);
521 	for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
522 		ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
523 		if (ret) {
524 			hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
525 				  ret);
526 			break;
527 		}
528 	}
529 	spin_lock(&fd->tid_lock);
530 	fd->tid_used -= tididx;
531 	spin_unlock(&fd->tid_lock);
532 	tinfo->tidcnt = tididx;
533 	mutex_unlock(&uctxt->exp_mutex);
534 
535 	kfree(tidinfo);
536 	return ret;
537 }
538 
539 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
540 			      struct hfi1_tid_info *tinfo)
541 {
542 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
543 	unsigned long *ev = uctxt->dd->events +
544 		(uctxt_offset(uctxt) + fd->subctxt);
545 	u32 *array;
546 	int ret = 0;
547 
548 	/*
549 	 * copy_to_user() can sleep, which will leave the invalid_lock
550 	 * locked and cause the MMU notifier to be blocked on the lock
551 	 * for a long time.
552 	 * Copy the data to a local buffer so we can release the lock.
553 	 */
554 	array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
555 	if (!array)
556 		return -EFAULT;
557 
558 	spin_lock(&fd->invalid_lock);
559 	if (fd->invalid_tid_idx) {
560 		memcpy(array, fd->invalid_tids, sizeof(*array) *
561 		       fd->invalid_tid_idx);
562 		memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
563 		       fd->invalid_tid_idx);
564 		tinfo->tidcnt = fd->invalid_tid_idx;
565 		fd->invalid_tid_idx = 0;
566 		/*
567 		 * Reset the user flag while still holding the lock.
568 		 * Otherwise, PSM can miss events.
569 		 */
570 		clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
571 	} else {
572 		tinfo->tidcnt = 0;
573 	}
574 	spin_unlock(&fd->invalid_lock);
575 
576 	if (tinfo->tidcnt) {
577 		if (copy_to_user((void __user *)tinfo->tidlist,
578 				 array, sizeof(*array) * tinfo->tidcnt))
579 			ret = -EFAULT;
580 	}
581 	kfree(array);
582 
583 	return ret;
584 }
585 
586 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages)
587 {
588 	unsigned pagecount, pageidx, setcount = 0, i;
589 	unsigned long pfn, this_pfn;
590 	struct page **pages = tidbuf->pages;
591 	struct tid_pageset *list = tidbuf->psets;
592 
593 	if (!npages)
594 		return 0;
595 
596 	/*
597 	 * Look for sets of physically contiguous pages in the user buffer.
598 	 * This will allow us to optimize Expected RcvArray entry usage by
599 	 * using the bigger supported sizes.
600 	 */
601 	pfn = page_to_pfn(pages[0]);
602 	for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
603 		this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
604 
605 		/*
606 		 * If the pfn's are not sequential, pages are not physically
607 		 * contiguous.
608 		 */
609 		if (this_pfn != ++pfn) {
610 			/*
611 			 * At this point we have to loop over the set of
612 			 * physically contiguous pages and break them down it
613 			 * sizes supported by the HW.
614 			 * There are two main constraints:
615 			 *     1. The max buffer size is MAX_EXPECTED_BUFFER.
616 			 *        If the total set size is bigger than that
617 			 *        program only a MAX_EXPECTED_BUFFER chunk.
618 			 *     2. The buffer size has to be a power of two. If
619 			 *        it is not, round down to the closes power of
620 			 *        2 and program that size.
621 			 */
622 			while (pagecount) {
623 				int maxpages = pagecount;
624 				u32 bufsize = pagecount * PAGE_SIZE;
625 
626 				if (bufsize > MAX_EXPECTED_BUFFER)
627 					maxpages =
628 						MAX_EXPECTED_BUFFER >>
629 						PAGE_SHIFT;
630 				else if (!is_power_of_2(bufsize))
631 					maxpages =
632 						rounddown_pow_of_two(bufsize) >>
633 						PAGE_SHIFT;
634 
635 				list[setcount].idx = pageidx;
636 				list[setcount].count = maxpages;
637 				pagecount -= maxpages;
638 				pageidx += maxpages;
639 				setcount++;
640 			}
641 			pageidx = i;
642 			pagecount = 1;
643 			pfn = this_pfn;
644 		} else {
645 			pagecount++;
646 		}
647 	}
648 	return setcount;
649 }
650 
651 /**
652  * program_rcvarray() - program an RcvArray group with receive buffers
653  * @fd: filedata pointer
654  * @tbuf: pointer to struct tid_user_buf that has the user buffer starting
655  *	  virtual address, buffer length, page pointers, pagesets (array of
656  *	  struct tid_pageset holding information on physically contiguous
657  *	  chunks from the user buffer), and other fields.
658  * @grp: RcvArray group
659  * @start: starting index into sets array
660  * @count: number of struct tid_pageset's to program
661  * @tidlist: the array of u32 elements when the information about the
662  *           programmed RcvArray entries is to be encoded.
663  * @tididx: starting offset into tidlist
664  * @pmapped: (output parameter) number of pages programmed into the RcvArray
665  *           entries.
666  *
667  * This function will program up to 'count' number of RcvArray entries from the
668  * group 'grp'. To make best use of write-combining writes, the function will
669  * perform writes to the unused RcvArray entries which will be ignored by the
670  * HW. Each RcvArray entry will be programmed with a physically contiguous
671  * buffer chunk from the user's virtual buffer.
672  *
673  * Return:
674  * -EINVAL if the requested count is larger than the size of the group,
675  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
676  * number of RcvArray entries programmed.
677  */
678 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf,
679 			    struct tid_group *grp,
680 			    unsigned int start, u16 count,
681 			    u32 *tidlist, unsigned int *tididx,
682 			    unsigned int *pmapped)
683 {
684 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
685 	struct hfi1_devdata *dd = uctxt->dd;
686 	u16 idx;
687 	u32 tidinfo = 0, rcventry, useidx = 0;
688 	int mapped = 0;
689 
690 	/* Count should never be larger than the group size */
691 	if (count > grp->size)
692 		return -EINVAL;
693 
694 	/* Find the first unused entry in the group */
695 	for (idx = 0; idx < grp->size; idx++) {
696 		if (!(grp->map & (1 << idx))) {
697 			useidx = idx;
698 			break;
699 		}
700 		rcv_array_wc_fill(dd, grp->base + idx);
701 	}
702 
703 	idx = 0;
704 	while (idx < count) {
705 		u16 npages, pageidx, setidx = start + idx;
706 		int ret = 0;
707 
708 		/*
709 		 * If this entry in the group is used, move to the next one.
710 		 * If we go past the end of the group, exit the loop.
711 		 */
712 		if (useidx >= grp->size) {
713 			break;
714 		} else if (grp->map & (1 << useidx)) {
715 			rcv_array_wc_fill(dd, grp->base + useidx);
716 			useidx++;
717 			continue;
718 		}
719 
720 		rcventry = grp->base + useidx;
721 		npages = tbuf->psets[setidx].count;
722 		pageidx = tbuf->psets[setidx].idx;
723 
724 		ret = set_rcvarray_entry(fd, tbuf,
725 					 rcventry, grp, pageidx,
726 					 npages);
727 		if (ret)
728 			return ret;
729 		mapped += npages;
730 
731 		tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
732 			EXP_TID_SET(LEN, npages);
733 		tidlist[(*tididx)++] = tidinfo;
734 		grp->used++;
735 		grp->map |= 1 << useidx++;
736 		idx++;
737 	}
738 
739 	/* Fill the rest of the group with "blank" writes */
740 	for (; useidx < grp->size; useidx++)
741 		rcv_array_wc_fill(dd, grp->base + useidx);
742 	*pmapped = mapped;
743 	return idx;
744 }
745 
746 static int set_rcvarray_entry(struct hfi1_filedata *fd,
747 			      struct tid_user_buf *tbuf,
748 			      u32 rcventry, struct tid_group *grp,
749 			      u16 pageidx, unsigned int npages)
750 {
751 	int ret;
752 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
753 	struct tid_rb_node *node;
754 	struct hfi1_devdata *dd = uctxt->dd;
755 	dma_addr_t phys;
756 	struct page **pages = tbuf->pages + pageidx;
757 
758 	/*
759 	 * Allocate the node first so we can handle a potential
760 	 * failure before we've programmed anything.
761 	 */
762 	node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
763 		       GFP_KERNEL);
764 	if (!node)
765 		return -ENOMEM;
766 
767 	phys = pci_map_single(dd->pcidev,
768 			      __va(page_to_phys(pages[0])),
769 			      npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
770 	if (dma_mapping_error(&dd->pcidev->dev, phys)) {
771 		dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
772 			   phys);
773 		kfree(node);
774 		return -EFAULT;
775 	}
776 
777 	node->mmu.addr = tbuf->vaddr + (pageidx * PAGE_SIZE);
778 	node->mmu.len = npages * PAGE_SIZE;
779 	node->phys = page_to_phys(pages[0]);
780 	node->npages = npages;
781 	node->rcventry = rcventry;
782 	node->dma_addr = phys;
783 	node->grp = grp;
784 	node->freed = false;
785 	memcpy(node->pages, pages, sizeof(struct page *) * npages);
786 
787 	if (!fd->handler)
788 		ret = tid_rb_insert(fd, &node->mmu);
789 	else
790 		ret = hfi1_mmu_rb_insert(fd->handler, &node->mmu);
791 
792 	if (ret) {
793 		hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
794 			  node->rcventry, node->mmu.addr, node->phys, ret);
795 		pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
796 				 PCI_DMA_FROMDEVICE);
797 		kfree(node);
798 		return -EFAULT;
799 	}
800 	hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
801 	trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
802 			       node->mmu.addr, node->phys, phys);
803 	return 0;
804 }
805 
806 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
807 			      struct tid_group **grp)
808 {
809 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
810 	struct hfi1_devdata *dd = uctxt->dd;
811 	struct tid_rb_node *node;
812 	u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
813 	u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
814 
815 	if (tididx >= uctxt->expected_count) {
816 		dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
817 			   tididx, uctxt->ctxt);
818 		return -EINVAL;
819 	}
820 
821 	if (tidctrl == 0x3)
822 		return -EINVAL;
823 
824 	rcventry = tididx + (tidctrl - 1);
825 
826 	node = fd->entry_to_rb[rcventry];
827 	if (!node || node->rcventry != (uctxt->expected_base + rcventry))
828 		return -EBADF;
829 
830 	if (grp)
831 		*grp = node->grp;
832 
833 	if (!fd->handler)
834 		cacheless_tid_rb_remove(fd, node);
835 	else
836 		hfi1_mmu_rb_remove(fd->handler, &node->mmu);
837 
838 	return 0;
839 }
840 
841 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
842 {
843 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
844 	struct hfi1_devdata *dd = uctxt->dd;
845 
846 	trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
847 				 node->npages, node->mmu.addr, node->phys,
848 				 node->dma_addr);
849 
850 	/*
851 	 * Make sure device has seen the write before we unpin the
852 	 * pages.
853 	 */
854 	hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0);
855 
856 	unpin_rcv_pages(fd, NULL, node, 0, node->npages, true);
857 
858 	node->grp->used--;
859 	node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
860 
861 	if (node->grp->used == node->grp->size - 1)
862 		tid_group_move(node->grp, &uctxt->tid_full_list,
863 			       &uctxt->tid_used_list);
864 	else if (!node->grp->used)
865 		tid_group_move(node->grp, &uctxt->tid_used_list,
866 			       &uctxt->tid_group_list);
867 	kfree(node);
868 }
869 
870 /*
871  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
872  * clearing nodes in the non-cached case.
873  */
874 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
875 			    struct exp_tid_set *set,
876 			    struct hfi1_filedata *fd)
877 {
878 	struct tid_group *grp, *ptr;
879 	int i;
880 
881 	list_for_each_entry_safe(grp, ptr, &set->list, list) {
882 		list_del_init(&grp->list);
883 
884 		for (i = 0; i < grp->size; i++) {
885 			if (grp->map & (1 << i)) {
886 				u16 rcventry = grp->base + i;
887 				struct tid_rb_node *node;
888 
889 				node = fd->entry_to_rb[rcventry -
890 							  uctxt->expected_base];
891 				if (!node || node->rcventry != rcventry)
892 					continue;
893 
894 				cacheless_tid_rb_remove(fd, node);
895 			}
896 		}
897 	}
898 }
899 
900 /*
901  * Always return 0 from this function.  A non-zero return indicates that the
902  * remove operation will be called and that memory should be unpinned.
903  * However, the driver cannot unpin out from under PSM.  Instead, retain the
904  * memory (by returning 0) and inform PSM that the memory is going away.  PSM
905  * will call back later when it has removed the memory from its list.
906  */
907 static int tid_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
908 {
909 	struct hfi1_filedata *fdata = arg;
910 	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
911 	struct tid_rb_node *node =
912 		container_of(mnode, struct tid_rb_node, mmu);
913 
914 	if (node->freed)
915 		return 0;
916 
917 	trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt, node->mmu.addr,
918 				 node->rcventry, node->npages, node->dma_addr);
919 	node->freed = true;
920 
921 	spin_lock(&fdata->invalid_lock);
922 	if (fdata->invalid_tid_idx < uctxt->expected_count) {
923 		fdata->invalid_tids[fdata->invalid_tid_idx] =
924 			rcventry2tidinfo(node->rcventry - uctxt->expected_base);
925 		fdata->invalid_tids[fdata->invalid_tid_idx] |=
926 			EXP_TID_SET(LEN, node->npages);
927 		if (!fdata->invalid_tid_idx) {
928 			unsigned long *ev;
929 
930 			/*
931 			 * hfi1_set_uevent_bits() sets a user event flag
932 			 * for all processes. Because calling into the
933 			 * driver to process TID cache invalidations is
934 			 * expensive and TID cache invalidations are
935 			 * handled on a per-process basis, we can
936 			 * optimize this to set the flag only for the
937 			 * process in question.
938 			 */
939 			ev = uctxt->dd->events +
940 				(uctxt_offset(uctxt) + fdata->subctxt);
941 			set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
942 		}
943 		fdata->invalid_tid_idx++;
944 	}
945 	spin_unlock(&fdata->invalid_lock);
946 	return 0;
947 }
948 
949 static int tid_rb_insert(void *arg, struct mmu_rb_node *node)
950 {
951 	struct hfi1_filedata *fdata = arg;
952 	struct tid_rb_node *tnode =
953 		container_of(node, struct tid_rb_node, mmu);
954 	u32 base = fdata->uctxt->expected_base;
955 
956 	fdata->entry_to_rb[tnode->rcventry - base] = tnode;
957 	return 0;
958 }
959 
960 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
961 				    struct tid_rb_node *tnode)
962 {
963 	u32 base = fdata->uctxt->expected_base;
964 
965 	fdata->entry_to_rb[tnode->rcventry - base] = NULL;
966 	clear_tid_node(fdata, tnode);
967 }
968 
969 static void tid_rb_remove(void *arg, struct mmu_rb_node *node)
970 {
971 	struct hfi1_filedata *fdata = arg;
972 	struct tid_rb_node *tnode =
973 		container_of(node, struct tid_rb_node, mmu);
974 
975 	cacheless_tid_rb_remove(fdata, tnode);
976 }
977