1 /*
2  * Copyright(c) 2020 Cornelis Networks, Inc.
3  * Copyright(c) 2015-2018 Intel Corporation.
4  *
5  * This file is provided under a dual BSD/GPLv2 license.  When using or
6  * redistributing this file, you may do so under either license.
7  *
8  * GPL LICENSE SUMMARY
9  *
10  * This program is free software; you can redistribute it and/or modify
11  * it under the terms of version 2 of the GNU General Public License as
12  * published by the Free Software Foundation.
13  *
14  * This program is distributed in the hope that it will be useful, but
15  * WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17  * General Public License for more details.
18  *
19  * BSD LICENSE
20  *
21  * Redistribution and use in source and binary forms, with or without
22  * modification, are permitted provided that the following conditions
23  * are met:
24  *
25  *  - Redistributions of source code must retain the above copyright
26  *    notice, this list of conditions and the following disclaimer.
27  *  - Redistributions in binary form must reproduce the above copyright
28  *    notice, this list of conditions and the following disclaimer in
29  *    the documentation and/or other materials provided with the
30  *    distribution.
31  *  - Neither the name of Intel Corporation nor the names of its
32  *    contributors may be used to endorse or promote products derived
33  *    from this software without specific prior written permission.
34  *
35  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
36  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
37  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
38  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
39  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
40  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
41  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
42  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
43  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
44  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
45  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
46  *
47  */
48 #include <asm/page.h>
49 #include <linux/string.h>
50 
51 #include "mmu_rb.h"
52 #include "user_exp_rcv.h"
53 #include "trace.h"
54 
55 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
56 			    struct exp_tid_set *set,
57 			    struct hfi1_filedata *fd);
58 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages);
59 static int set_rcvarray_entry(struct hfi1_filedata *fd,
60 			      struct tid_user_buf *tbuf,
61 			      u32 rcventry, struct tid_group *grp,
62 			      u16 pageidx, unsigned int npages);
63 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
64 				    struct tid_rb_node *tnode);
65 static bool tid_rb_invalidate(struct mmu_interval_notifier *mni,
66 			      const struct mmu_notifier_range *range,
67 			      unsigned long cur_seq);
68 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *,
69 			    struct tid_group *grp,
70 			    unsigned int start, u16 count,
71 			    u32 *tidlist, unsigned int *tididx,
72 			    unsigned int *pmapped);
73 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
74 			      struct tid_group **grp);
75 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
76 
77 static const struct mmu_interval_notifier_ops tid_mn_ops = {
78 	.invalidate = tid_rb_invalidate,
79 };
80 
81 /*
82  * Initialize context and file private data needed for Expected
83  * receive caching. This needs to be done after the context has
84  * been configured with the eager/expected RcvEntry counts.
85  */
86 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd,
87 			   struct hfi1_ctxtdata *uctxt)
88 {
89 	int ret = 0;
90 
91 	fd->entry_to_rb = kcalloc(uctxt->expected_count,
92 				  sizeof(struct rb_node *),
93 				  GFP_KERNEL);
94 	if (!fd->entry_to_rb)
95 		return -ENOMEM;
96 
97 	if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
98 		fd->invalid_tid_idx = 0;
99 		fd->invalid_tids = kcalloc(uctxt->expected_count,
100 					   sizeof(*fd->invalid_tids),
101 					   GFP_KERNEL);
102 		if (!fd->invalid_tids) {
103 			kfree(fd->entry_to_rb);
104 			fd->entry_to_rb = NULL;
105 			return -ENOMEM;
106 		}
107 		fd->use_mn = true;
108 	}
109 
110 	/*
111 	 * PSM does not have a good way to separate, count, and
112 	 * effectively enforce a limit on RcvArray entries used by
113 	 * subctxts (when context sharing is used) when TID caching
114 	 * is enabled. To help with that, we calculate a per-process
115 	 * RcvArray entry share and enforce that.
116 	 * If TID caching is not in use, PSM deals with usage on its
117 	 * own. In that case, we allow any subctxt to take all of the
118 	 * entries.
119 	 *
120 	 * Make sure that we set the tid counts only after successful
121 	 * init.
122 	 */
123 	spin_lock(&fd->tid_lock);
124 	if (uctxt->subctxt_cnt && fd->use_mn) {
125 		u16 remainder;
126 
127 		fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
128 		remainder = uctxt->expected_count % uctxt->subctxt_cnt;
129 		if (remainder && fd->subctxt < remainder)
130 			fd->tid_limit++;
131 	} else {
132 		fd->tid_limit = uctxt->expected_count;
133 	}
134 	spin_unlock(&fd->tid_lock);
135 
136 	return ret;
137 }
138 
139 void hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
140 {
141 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
142 
143 	mutex_lock(&uctxt->exp_mutex);
144 	if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
145 		unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
146 	if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
147 		unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
148 	mutex_unlock(&uctxt->exp_mutex);
149 
150 	kfree(fd->invalid_tids);
151 	fd->invalid_tids = NULL;
152 
153 	kfree(fd->entry_to_rb);
154 	fd->entry_to_rb = NULL;
155 }
156 
157 /*
158  * Release pinned receive buffer pages.
159  *
160  * @mapped: true if the pages have been DMA mapped. false otherwise.
161  * @idx: Index of the first page to unpin.
162  * @npages: No of pages to unpin.
163  *
164  * If the pages have been DMA mapped (indicated by mapped parameter), their
165  * info will be passed via a struct tid_rb_node. If they haven't been mapped,
166  * their info will be passed via a struct tid_user_buf.
167  */
168 static void unpin_rcv_pages(struct hfi1_filedata *fd,
169 			    struct tid_user_buf *tidbuf,
170 			    struct tid_rb_node *node,
171 			    unsigned int idx,
172 			    unsigned int npages,
173 			    bool mapped)
174 {
175 	struct page **pages;
176 	struct hfi1_devdata *dd = fd->uctxt->dd;
177 	struct mm_struct *mm;
178 
179 	if (mapped) {
180 		pci_unmap_single(dd->pcidev, node->dma_addr,
181 				 node->npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
182 		pages = &node->pages[idx];
183 		mm = mm_from_tid_node(node);
184 	} else {
185 		pages = &tidbuf->pages[idx];
186 		mm = current->mm;
187 	}
188 	hfi1_release_user_pages(mm, pages, npages, mapped);
189 	fd->tid_n_pinned -= npages;
190 }
191 
192 /*
193  * Pin receive buffer pages.
194  */
195 static int pin_rcv_pages(struct hfi1_filedata *fd, struct tid_user_buf *tidbuf)
196 {
197 	int pinned;
198 	unsigned int npages;
199 	unsigned long vaddr = tidbuf->vaddr;
200 	struct page **pages = NULL;
201 	struct hfi1_devdata *dd = fd->uctxt->dd;
202 
203 	/* Get the number of pages the user buffer spans */
204 	npages = num_user_pages(vaddr, tidbuf->length);
205 	if (!npages)
206 		return -EINVAL;
207 
208 	if (npages > fd->uctxt->expected_count) {
209 		dd_dev_err(dd, "Expected buffer too big\n");
210 		return -EINVAL;
211 	}
212 
213 	/* Allocate the array of struct page pointers needed for pinning */
214 	pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
215 	if (!pages)
216 		return -ENOMEM;
217 
218 	/*
219 	 * Pin all the pages of the user buffer. If we can't pin all the
220 	 * pages, accept the amount pinned so far and program only that.
221 	 * User space knows how to deal with partially programmed buffers.
222 	 */
223 	if (!hfi1_can_pin_pages(dd, current->mm, fd->tid_n_pinned, npages)) {
224 		kfree(pages);
225 		return -ENOMEM;
226 	}
227 
228 	pinned = hfi1_acquire_user_pages(current->mm, vaddr, npages, true, pages);
229 	if (pinned <= 0) {
230 		kfree(pages);
231 		return pinned;
232 	}
233 	tidbuf->pages = pages;
234 	tidbuf->npages = npages;
235 	fd->tid_n_pinned += pinned;
236 	return pinned;
237 }
238 
239 /*
240  * RcvArray entry allocation for Expected Receives is done by the
241  * following algorithm:
242  *
243  * The context keeps 3 lists of groups of RcvArray entries:
244  *   1. List of empty groups - tid_group_list
245  *      This list is created during user context creation and
246  *      contains elements which describe sets (of 8) of empty
247  *      RcvArray entries.
248  *   2. List of partially used groups - tid_used_list
249  *      This list contains sets of RcvArray entries which are
250  *      not completely used up. Another mapping request could
251  *      use some of all of the remaining entries.
252  *   3. List of full groups - tid_full_list
253  *      This is the list where sets that are completely used
254  *      up go.
255  *
256  * An attempt to optimize the usage of RcvArray entries is
257  * made by finding all sets of physically contiguous pages in a
258  * user's buffer.
259  * These physically contiguous sets are further split into
260  * sizes supported by the receive engine of the HFI. The
261  * resulting sets of pages are stored in struct tid_pageset,
262  * which describes the sets as:
263  *    * .count - number of pages in this set
264  *    * .idx - starting index into struct page ** array
265  *                    of this set
266  *
267  * From this point on, the algorithm deals with the page sets
268  * described above. The number of pagesets is divided by the
269  * RcvArray group size to produce the number of full groups
270  * needed.
271  *
272  * Groups from the 3 lists are manipulated using the following
273  * rules:
274  *   1. For each set of 8 pagesets, a complete group from
275  *      tid_group_list is taken, programmed, and moved to
276  *      the tid_full_list list.
277  *   2. For all remaining pagesets:
278  *      2.1 If the tid_used_list is empty and the tid_group_list
279  *          is empty, stop processing pageset and return only
280  *          what has been programmed up to this point.
281  *      2.2 If the tid_used_list is empty and the tid_group_list
282  *          is not empty, move a group from tid_group_list to
283  *          tid_used_list.
284  *      2.3 For each group is tid_used_group, program as much as
285  *          can fit into the group. If the group becomes fully
286  *          used, move it to tid_full_list.
287  */
288 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
289 			    struct hfi1_tid_info *tinfo)
290 {
291 	int ret = 0, need_group = 0, pinned;
292 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
293 	struct hfi1_devdata *dd = uctxt->dd;
294 	unsigned int ngroups, pageidx = 0, pageset_count,
295 		tididx = 0, mapped, mapped_pages = 0;
296 	u32 *tidlist = NULL;
297 	struct tid_user_buf *tidbuf;
298 
299 	if (!PAGE_ALIGNED(tinfo->vaddr))
300 		return -EINVAL;
301 
302 	tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL);
303 	if (!tidbuf)
304 		return -ENOMEM;
305 
306 	tidbuf->vaddr = tinfo->vaddr;
307 	tidbuf->length = tinfo->length;
308 	tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets),
309 				GFP_KERNEL);
310 	if (!tidbuf->psets) {
311 		kfree(tidbuf);
312 		return -ENOMEM;
313 	}
314 
315 	pinned = pin_rcv_pages(fd, tidbuf);
316 	if (pinned <= 0) {
317 		kfree(tidbuf->psets);
318 		kfree(tidbuf);
319 		return pinned;
320 	}
321 
322 	/* Find sets of physically contiguous pages */
323 	tidbuf->n_psets = find_phys_blocks(tidbuf, pinned);
324 
325 	/*
326 	 * We don't need to access this under a lock since tid_used is per
327 	 * process and the same process cannot be in hfi1_user_exp_rcv_clear()
328 	 * and hfi1_user_exp_rcv_setup() at the same time.
329 	 */
330 	spin_lock(&fd->tid_lock);
331 	if (fd->tid_used + tidbuf->n_psets > fd->tid_limit)
332 		pageset_count = fd->tid_limit - fd->tid_used;
333 	else
334 		pageset_count = tidbuf->n_psets;
335 	spin_unlock(&fd->tid_lock);
336 
337 	if (!pageset_count)
338 		goto bail;
339 
340 	ngroups = pageset_count / dd->rcv_entries.group_size;
341 	tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
342 	if (!tidlist) {
343 		ret = -ENOMEM;
344 		goto nomem;
345 	}
346 
347 	tididx = 0;
348 
349 	/*
350 	 * From this point on, we are going to be using shared (between master
351 	 * and subcontexts) context resources. We need to take the lock.
352 	 */
353 	mutex_lock(&uctxt->exp_mutex);
354 	/*
355 	 * The first step is to program the RcvArray entries which are complete
356 	 * groups.
357 	 */
358 	while (ngroups && uctxt->tid_group_list.count) {
359 		struct tid_group *grp =
360 			tid_group_pop(&uctxt->tid_group_list);
361 
362 		ret = program_rcvarray(fd, tidbuf, grp,
363 				       pageidx, dd->rcv_entries.group_size,
364 				       tidlist, &tididx, &mapped);
365 		/*
366 		 * If there was a failure to program the RcvArray
367 		 * entries for the entire group, reset the grp fields
368 		 * and add the grp back to the free group list.
369 		 */
370 		if (ret <= 0) {
371 			tid_group_add_tail(grp, &uctxt->tid_group_list);
372 			hfi1_cdbg(TID,
373 				  "Failed to program RcvArray group %d", ret);
374 			goto unlock;
375 		}
376 
377 		tid_group_add_tail(grp, &uctxt->tid_full_list);
378 		ngroups--;
379 		pageidx += ret;
380 		mapped_pages += mapped;
381 	}
382 
383 	while (pageidx < pageset_count) {
384 		struct tid_group *grp, *ptr;
385 		/*
386 		 * If we don't have any partially used tid groups, check
387 		 * if we have empty groups. If so, take one from there and
388 		 * put in the partially used list.
389 		 */
390 		if (!uctxt->tid_used_list.count || need_group) {
391 			if (!uctxt->tid_group_list.count)
392 				goto unlock;
393 
394 			grp = tid_group_pop(&uctxt->tid_group_list);
395 			tid_group_add_tail(grp, &uctxt->tid_used_list);
396 			need_group = 0;
397 		}
398 		/*
399 		 * There is an optimization opportunity here - instead of
400 		 * fitting as many page sets as we can, check for a group
401 		 * later on in the list that could fit all of them.
402 		 */
403 		list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
404 					 list) {
405 			unsigned use = min_t(unsigned, pageset_count - pageidx,
406 					     grp->size - grp->used);
407 
408 			ret = program_rcvarray(fd, tidbuf, grp,
409 					       pageidx, use, tidlist,
410 					       &tididx, &mapped);
411 			if (ret < 0) {
412 				hfi1_cdbg(TID,
413 					  "Failed to program RcvArray entries %d",
414 					  ret);
415 				goto unlock;
416 			} else if (ret > 0) {
417 				if (grp->used == grp->size)
418 					tid_group_move(grp,
419 						       &uctxt->tid_used_list,
420 						       &uctxt->tid_full_list);
421 				pageidx += ret;
422 				mapped_pages += mapped;
423 				need_group = 0;
424 				/* Check if we are done so we break out early */
425 				if (pageidx >= pageset_count)
426 					break;
427 			} else if (WARN_ON(ret == 0)) {
428 				/*
429 				 * If ret is 0, we did not program any entries
430 				 * into this group, which can only happen if
431 				 * we've screwed up the accounting somewhere.
432 				 * Warn and try to continue.
433 				 */
434 				need_group = 1;
435 			}
436 		}
437 	}
438 unlock:
439 	mutex_unlock(&uctxt->exp_mutex);
440 nomem:
441 	hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
442 		  mapped_pages, ret);
443 	if (tididx) {
444 		spin_lock(&fd->tid_lock);
445 		fd->tid_used += tididx;
446 		spin_unlock(&fd->tid_lock);
447 		tinfo->tidcnt = tididx;
448 		tinfo->length = mapped_pages * PAGE_SIZE;
449 
450 		if (copy_to_user(u64_to_user_ptr(tinfo->tidlist),
451 				 tidlist, sizeof(tidlist[0]) * tididx)) {
452 			/*
453 			 * On failure to copy to the user level, we need to undo
454 			 * everything done so far so we don't leak resources.
455 			 */
456 			tinfo->tidlist = (unsigned long)&tidlist;
457 			hfi1_user_exp_rcv_clear(fd, tinfo);
458 			tinfo->tidlist = 0;
459 			ret = -EFAULT;
460 			goto bail;
461 		}
462 	}
463 
464 	/*
465 	 * If not everything was mapped (due to insufficient RcvArray entries,
466 	 * for example), unpin all unmapped pages so we can pin them nex time.
467 	 */
468 	if (mapped_pages != pinned)
469 		unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages,
470 				(pinned - mapped_pages), false);
471 bail:
472 	kfree(tidbuf->psets);
473 	kfree(tidlist);
474 	kfree(tidbuf->pages);
475 	kfree(tidbuf);
476 	return ret > 0 ? 0 : ret;
477 }
478 
479 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
480 			    struct hfi1_tid_info *tinfo)
481 {
482 	int ret = 0;
483 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
484 	u32 *tidinfo;
485 	unsigned tididx;
486 
487 	if (unlikely(tinfo->tidcnt > fd->tid_used))
488 		return -EINVAL;
489 
490 	tidinfo = memdup_user(u64_to_user_ptr(tinfo->tidlist),
491 			      sizeof(tidinfo[0]) * tinfo->tidcnt);
492 	if (IS_ERR(tidinfo))
493 		return PTR_ERR(tidinfo);
494 
495 	mutex_lock(&uctxt->exp_mutex);
496 	for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
497 		ret = unprogram_rcvarray(fd, tidinfo[tididx], NULL);
498 		if (ret) {
499 			hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
500 				  ret);
501 			break;
502 		}
503 	}
504 	spin_lock(&fd->tid_lock);
505 	fd->tid_used -= tididx;
506 	spin_unlock(&fd->tid_lock);
507 	tinfo->tidcnt = tididx;
508 	mutex_unlock(&uctxt->exp_mutex);
509 
510 	kfree(tidinfo);
511 	return ret;
512 }
513 
514 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
515 			      struct hfi1_tid_info *tinfo)
516 {
517 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
518 	unsigned long *ev = uctxt->dd->events +
519 		(uctxt_offset(uctxt) + fd->subctxt);
520 	u32 *array;
521 	int ret = 0;
522 
523 	/*
524 	 * copy_to_user() can sleep, which will leave the invalid_lock
525 	 * locked and cause the MMU notifier to be blocked on the lock
526 	 * for a long time.
527 	 * Copy the data to a local buffer so we can release the lock.
528 	 */
529 	array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
530 	if (!array)
531 		return -EFAULT;
532 
533 	spin_lock(&fd->invalid_lock);
534 	if (fd->invalid_tid_idx) {
535 		memcpy(array, fd->invalid_tids, sizeof(*array) *
536 		       fd->invalid_tid_idx);
537 		memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
538 		       fd->invalid_tid_idx);
539 		tinfo->tidcnt = fd->invalid_tid_idx;
540 		fd->invalid_tid_idx = 0;
541 		/*
542 		 * Reset the user flag while still holding the lock.
543 		 * Otherwise, PSM can miss events.
544 		 */
545 		clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
546 	} else {
547 		tinfo->tidcnt = 0;
548 	}
549 	spin_unlock(&fd->invalid_lock);
550 
551 	if (tinfo->tidcnt) {
552 		if (copy_to_user((void __user *)tinfo->tidlist,
553 				 array, sizeof(*array) * tinfo->tidcnt))
554 			ret = -EFAULT;
555 	}
556 	kfree(array);
557 
558 	return ret;
559 }
560 
561 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages)
562 {
563 	unsigned pagecount, pageidx, setcount = 0, i;
564 	unsigned long pfn, this_pfn;
565 	struct page **pages = tidbuf->pages;
566 	struct tid_pageset *list = tidbuf->psets;
567 
568 	if (!npages)
569 		return 0;
570 
571 	/*
572 	 * Look for sets of physically contiguous pages in the user buffer.
573 	 * This will allow us to optimize Expected RcvArray entry usage by
574 	 * using the bigger supported sizes.
575 	 */
576 	pfn = page_to_pfn(pages[0]);
577 	for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
578 		this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
579 
580 		/*
581 		 * If the pfn's are not sequential, pages are not physically
582 		 * contiguous.
583 		 */
584 		if (this_pfn != ++pfn) {
585 			/*
586 			 * At this point we have to loop over the set of
587 			 * physically contiguous pages and break them down it
588 			 * sizes supported by the HW.
589 			 * There are two main constraints:
590 			 *     1. The max buffer size is MAX_EXPECTED_BUFFER.
591 			 *        If the total set size is bigger than that
592 			 *        program only a MAX_EXPECTED_BUFFER chunk.
593 			 *     2. The buffer size has to be a power of two. If
594 			 *        it is not, round down to the closes power of
595 			 *        2 and program that size.
596 			 */
597 			while (pagecount) {
598 				int maxpages = pagecount;
599 				u32 bufsize = pagecount * PAGE_SIZE;
600 
601 				if (bufsize > MAX_EXPECTED_BUFFER)
602 					maxpages =
603 						MAX_EXPECTED_BUFFER >>
604 						PAGE_SHIFT;
605 				else if (!is_power_of_2(bufsize))
606 					maxpages =
607 						rounddown_pow_of_two(bufsize) >>
608 						PAGE_SHIFT;
609 
610 				list[setcount].idx = pageidx;
611 				list[setcount].count = maxpages;
612 				pagecount -= maxpages;
613 				pageidx += maxpages;
614 				setcount++;
615 			}
616 			pageidx = i;
617 			pagecount = 1;
618 			pfn = this_pfn;
619 		} else {
620 			pagecount++;
621 		}
622 	}
623 	return setcount;
624 }
625 
626 /**
627  * program_rcvarray() - program an RcvArray group with receive buffers
628  * @fd: filedata pointer
629  * @tbuf: pointer to struct tid_user_buf that has the user buffer starting
630  *	  virtual address, buffer length, page pointers, pagesets (array of
631  *	  struct tid_pageset holding information on physically contiguous
632  *	  chunks from the user buffer), and other fields.
633  * @grp: RcvArray group
634  * @start: starting index into sets array
635  * @count: number of struct tid_pageset's to program
636  * @tidlist: the array of u32 elements when the information about the
637  *           programmed RcvArray entries is to be encoded.
638  * @tididx: starting offset into tidlist
639  * @pmapped: (output parameter) number of pages programmed into the RcvArray
640  *           entries.
641  *
642  * This function will program up to 'count' number of RcvArray entries from the
643  * group 'grp'. To make best use of write-combining writes, the function will
644  * perform writes to the unused RcvArray entries which will be ignored by the
645  * HW. Each RcvArray entry will be programmed with a physically contiguous
646  * buffer chunk from the user's virtual buffer.
647  *
648  * Return:
649  * -EINVAL if the requested count is larger than the size of the group,
650  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
651  * number of RcvArray entries programmed.
652  */
653 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf,
654 			    struct tid_group *grp,
655 			    unsigned int start, u16 count,
656 			    u32 *tidlist, unsigned int *tididx,
657 			    unsigned int *pmapped)
658 {
659 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
660 	struct hfi1_devdata *dd = uctxt->dd;
661 	u16 idx;
662 	u32 tidinfo = 0, rcventry, useidx = 0;
663 	int mapped = 0;
664 
665 	/* Count should never be larger than the group size */
666 	if (count > grp->size)
667 		return -EINVAL;
668 
669 	/* Find the first unused entry in the group */
670 	for (idx = 0; idx < grp->size; idx++) {
671 		if (!(grp->map & (1 << idx))) {
672 			useidx = idx;
673 			break;
674 		}
675 		rcv_array_wc_fill(dd, grp->base + idx);
676 	}
677 
678 	idx = 0;
679 	while (idx < count) {
680 		u16 npages, pageidx, setidx = start + idx;
681 		int ret = 0;
682 
683 		/*
684 		 * If this entry in the group is used, move to the next one.
685 		 * If we go past the end of the group, exit the loop.
686 		 */
687 		if (useidx >= grp->size) {
688 			break;
689 		} else if (grp->map & (1 << useidx)) {
690 			rcv_array_wc_fill(dd, grp->base + useidx);
691 			useidx++;
692 			continue;
693 		}
694 
695 		rcventry = grp->base + useidx;
696 		npages = tbuf->psets[setidx].count;
697 		pageidx = tbuf->psets[setidx].idx;
698 
699 		ret = set_rcvarray_entry(fd, tbuf,
700 					 rcventry, grp, pageidx,
701 					 npages);
702 		if (ret)
703 			return ret;
704 		mapped += npages;
705 
706 		tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) |
707 			EXP_TID_SET(LEN, npages);
708 		tidlist[(*tididx)++] = tidinfo;
709 		grp->used++;
710 		grp->map |= 1 << useidx++;
711 		idx++;
712 	}
713 
714 	/* Fill the rest of the group with "blank" writes */
715 	for (; useidx < grp->size; useidx++)
716 		rcv_array_wc_fill(dd, grp->base + useidx);
717 	*pmapped = mapped;
718 	return idx;
719 }
720 
721 static int set_rcvarray_entry(struct hfi1_filedata *fd,
722 			      struct tid_user_buf *tbuf,
723 			      u32 rcventry, struct tid_group *grp,
724 			      u16 pageidx, unsigned int npages)
725 {
726 	int ret;
727 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
728 	struct tid_rb_node *node;
729 	struct hfi1_devdata *dd = uctxt->dd;
730 	dma_addr_t phys;
731 	struct page **pages = tbuf->pages + pageidx;
732 
733 	/*
734 	 * Allocate the node first so we can handle a potential
735 	 * failure before we've programmed anything.
736 	 */
737 	node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages),
738 		       GFP_KERNEL);
739 	if (!node)
740 		return -ENOMEM;
741 
742 	phys = pci_map_single(dd->pcidev,
743 			      __va(page_to_phys(pages[0])),
744 			      npages * PAGE_SIZE, PCI_DMA_FROMDEVICE);
745 	if (dma_mapping_error(&dd->pcidev->dev, phys)) {
746 		dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n",
747 			   phys);
748 		kfree(node);
749 		return -EFAULT;
750 	}
751 
752 	node->fdata = fd;
753 	node->phys = page_to_phys(pages[0]);
754 	node->npages = npages;
755 	node->rcventry = rcventry;
756 	node->dma_addr = phys;
757 	node->grp = grp;
758 	node->freed = false;
759 	memcpy(node->pages, pages, sizeof(struct page *) * npages);
760 
761 	if (fd->use_mn) {
762 		ret = mmu_interval_notifier_insert(
763 			&node->notifier, current->mm,
764 			tbuf->vaddr + (pageidx * PAGE_SIZE), npages * PAGE_SIZE,
765 			&tid_mn_ops);
766 		if (ret)
767 			goto out_unmap;
768 		/*
769 		 * FIXME: This is in the wrong order, the notifier should be
770 		 * established before the pages are pinned by pin_rcv_pages.
771 		 */
772 		mmu_interval_read_begin(&node->notifier);
773 	}
774 	fd->entry_to_rb[node->rcventry - uctxt->expected_base] = node;
775 
776 	hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
777 	trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
778 			       node->notifier.interval_tree.start, node->phys,
779 			       phys);
780 	return 0;
781 
782 out_unmap:
783 	hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
784 		  node->rcventry, node->notifier.interval_tree.start,
785 		  node->phys, ret);
786 	pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE,
787 			 PCI_DMA_FROMDEVICE);
788 	kfree(node);
789 	return -EFAULT;
790 }
791 
792 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo,
793 			      struct tid_group **grp)
794 {
795 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
796 	struct hfi1_devdata *dd = uctxt->dd;
797 	struct tid_rb_node *node;
798 	u8 tidctrl = EXP_TID_GET(tidinfo, CTRL);
799 	u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
800 
801 	if (tididx >= uctxt->expected_count) {
802 		dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
803 			   tididx, uctxt->ctxt);
804 		return -EINVAL;
805 	}
806 
807 	if (tidctrl == 0x3)
808 		return -EINVAL;
809 
810 	rcventry = tididx + (tidctrl - 1);
811 
812 	node = fd->entry_to_rb[rcventry];
813 	if (!node || node->rcventry != (uctxt->expected_base + rcventry))
814 		return -EBADF;
815 
816 	if (grp)
817 		*grp = node->grp;
818 
819 	if (fd->use_mn)
820 		mmu_interval_notifier_remove(&node->notifier);
821 	cacheless_tid_rb_remove(fd, node);
822 
823 	return 0;
824 }
825 
826 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
827 {
828 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
829 	struct hfi1_devdata *dd = uctxt->dd;
830 
831 	trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
832 				 node->npages,
833 				 node->notifier.interval_tree.start, node->phys,
834 				 node->dma_addr);
835 
836 	/*
837 	 * Make sure device has seen the write before we unpin the
838 	 * pages.
839 	 */
840 	hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0);
841 
842 	unpin_rcv_pages(fd, NULL, node, 0, node->npages, true);
843 
844 	node->grp->used--;
845 	node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
846 
847 	if (node->grp->used == node->grp->size - 1)
848 		tid_group_move(node->grp, &uctxt->tid_full_list,
849 			       &uctxt->tid_used_list);
850 	else if (!node->grp->used)
851 		tid_group_move(node->grp, &uctxt->tid_used_list,
852 			       &uctxt->tid_group_list);
853 	kfree(node);
854 }
855 
856 /*
857  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
858  * clearing nodes in the non-cached case.
859  */
860 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
861 			    struct exp_tid_set *set,
862 			    struct hfi1_filedata *fd)
863 {
864 	struct tid_group *grp, *ptr;
865 	int i;
866 
867 	list_for_each_entry_safe(grp, ptr, &set->list, list) {
868 		list_del_init(&grp->list);
869 
870 		for (i = 0; i < grp->size; i++) {
871 			if (grp->map & (1 << i)) {
872 				u16 rcventry = grp->base + i;
873 				struct tid_rb_node *node;
874 
875 				node = fd->entry_to_rb[rcventry -
876 							  uctxt->expected_base];
877 				if (!node || node->rcventry != rcventry)
878 					continue;
879 
880 				if (fd->use_mn)
881 					mmu_interval_notifier_remove(
882 						&node->notifier);
883 				cacheless_tid_rb_remove(fd, node);
884 			}
885 		}
886 	}
887 }
888 
889 static bool tid_rb_invalidate(struct mmu_interval_notifier *mni,
890 			      const struct mmu_notifier_range *range,
891 			      unsigned long cur_seq)
892 {
893 	struct tid_rb_node *node =
894 		container_of(mni, struct tid_rb_node, notifier);
895 	struct hfi1_filedata *fdata = node->fdata;
896 	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
897 
898 	if (node->freed)
899 		return true;
900 
901 	trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt,
902 				 node->notifier.interval_tree.start,
903 				 node->rcventry, node->npages, node->dma_addr);
904 	node->freed = true;
905 
906 	spin_lock(&fdata->invalid_lock);
907 	if (fdata->invalid_tid_idx < uctxt->expected_count) {
908 		fdata->invalid_tids[fdata->invalid_tid_idx] =
909 			rcventry2tidinfo(node->rcventry - uctxt->expected_base);
910 		fdata->invalid_tids[fdata->invalid_tid_idx] |=
911 			EXP_TID_SET(LEN, node->npages);
912 		if (!fdata->invalid_tid_idx) {
913 			unsigned long *ev;
914 
915 			/*
916 			 * hfi1_set_uevent_bits() sets a user event flag
917 			 * for all processes. Because calling into the
918 			 * driver to process TID cache invalidations is
919 			 * expensive and TID cache invalidations are
920 			 * handled on a per-process basis, we can
921 			 * optimize this to set the flag only for the
922 			 * process in question.
923 			 */
924 			ev = uctxt->dd->events +
925 				(uctxt_offset(uctxt) + fdata->subctxt);
926 			set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
927 		}
928 		fdata->invalid_tid_idx++;
929 	}
930 	spin_unlock(&fdata->invalid_lock);
931 	return true;
932 }
933 
934 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
935 				    struct tid_rb_node *tnode)
936 {
937 	u32 base = fdata->uctxt->expected_base;
938 
939 	fdata->entry_to_rb[tnode->rcventry - base] = NULL;
940 	clear_tid_node(fdata, tnode);
941 }
942