xref: /linux/drivers/infiniband/hw/hfi1/user_exp_rcv.c (revision 06d07429858317ded2db7986113a9e0129cd599b)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright(c) 2020 Cornelis Networks, Inc.
4  * Copyright(c) 2015-2018 Intel Corporation.
5  */
6 #include <asm/page.h>
7 #include <linux/string.h>
8 
9 #include "mmu_rb.h"
10 #include "user_exp_rcv.h"
11 #include "trace.h"
12 
13 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
14 			    struct exp_tid_set *set,
15 			    struct hfi1_filedata *fd);
16 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages);
17 static int set_rcvarray_entry(struct hfi1_filedata *fd,
18 			      struct tid_user_buf *tbuf,
19 			      u32 rcventry, struct tid_group *grp,
20 			      u16 pageidx, unsigned int npages);
21 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
22 				    struct tid_rb_node *tnode);
23 static bool tid_rb_invalidate(struct mmu_interval_notifier *mni,
24 			      const struct mmu_notifier_range *range,
25 			      unsigned long cur_seq);
26 static bool tid_cover_invalidate(struct mmu_interval_notifier *mni,
27 			         const struct mmu_notifier_range *range,
28 			         unsigned long cur_seq);
29 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *,
30 			    struct tid_group *grp, u16 count,
31 			    u32 *tidlist, unsigned int *tididx,
32 			    unsigned int *pmapped);
33 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo);
34 static void __clear_tid_node(struct hfi1_filedata *fd,
35 			     struct tid_rb_node *node);
36 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node);
37 
38 static const struct mmu_interval_notifier_ops tid_mn_ops = {
39 	.invalidate = tid_rb_invalidate,
40 };
41 static const struct mmu_interval_notifier_ops tid_cover_ops = {
42 	.invalidate = tid_cover_invalidate,
43 };
44 
45 /*
46  * Initialize context and file private data needed for Expected
47  * receive caching. This needs to be done after the context has
48  * been configured with the eager/expected RcvEntry counts.
49  */
hfi1_user_exp_rcv_init(struct hfi1_filedata * fd,struct hfi1_ctxtdata * uctxt)50 int hfi1_user_exp_rcv_init(struct hfi1_filedata *fd,
51 			   struct hfi1_ctxtdata *uctxt)
52 {
53 	int ret = 0;
54 
55 	fd->entry_to_rb = kcalloc(uctxt->expected_count,
56 				  sizeof(struct rb_node *),
57 				  GFP_KERNEL);
58 	if (!fd->entry_to_rb)
59 		return -ENOMEM;
60 
61 	if (!HFI1_CAP_UGET_MASK(uctxt->flags, TID_UNMAP)) {
62 		fd->invalid_tid_idx = 0;
63 		fd->invalid_tids = kcalloc(uctxt->expected_count,
64 					   sizeof(*fd->invalid_tids),
65 					   GFP_KERNEL);
66 		if (!fd->invalid_tids) {
67 			kfree(fd->entry_to_rb);
68 			fd->entry_to_rb = NULL;
69 			return -ENOMEM;
70 		}
71 		fd->use_mn = true;
72 	}
73 
74 	/*
75 	 * PSM does not have a good way to separate, count, and
76 	 * effectively enforce a limit on RcvArray entries used by
77 	 * subctxts (when context sharing is used) when TID caching
78 	 * is enabled. To help with that, we calculate a per-process
79 	 * RcvArray entry share and enforce that.
80 	 * If TID caching is not in use, PSM deals with usage on its
81 	 * own. In that case, we allow any subctxt to take all of the
82 	 * entries.
83 	 *
84 	 * Make sure that we set the tid counts only after successful
85 	 * init.
86 	 */
87 	spin_lock(&fd->tid_lock);
88 	if (uctxt->subctxt_cnt && fd->use_mn) {
89 		u16 remainder;
90 
91 		fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt;
92 		remainder = uctxt->expected_count % uctxt->subctxt_cnt;
93 		if (remainder && fd->subctxt < remainder)
94 			fd->tid_limit++;
95 	} else {
96 		fd->tid_limit = uctxt->expected_count;
97 	}
98 	spin_unlock(&fd->tid_lock);
99 
100 	return ret;
101 }
102 
hfi1_user_exp_rcv_free(struct hfi1_filedata * fd)103 void hfi1_user_exp_rcv_free(struct hfi1_filedata *fd)
104 {
105 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
106 
107 	mutex_lock(&uctxt->exp_mutex);
108 	if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list))
109 		unlock_exp_tids(uctxt, &uctxt->tid_full_list, fd);
110 	if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list))
111 		unlock_exp_tids(uctxt, &uctxt->tid_used_list, fd);
112 	mutex_unlock(&uctxt->exp_mutex);
113 
114 	kfree(fd->invalid_tids);
115 	fd->invalid_tids = NULL;
116 
117 	kfree(fd->entry_to_rb);
118 	fd->entry_to_rb = NULL;
119 }
120 
121 /*
122  * Release pinned receive buffer pages.
123  *
124  * @mapped: true if the pages have been DMA mapped. false otherwise.
125  * @idx: Index of the first page to unpin.
126  * @npages: No of pages to unpin.
127  *
128  * If the pages have been DMA mapped (indicated by mapped parameter), their
129  * info will be passed via a struct tid_rb_node. If they haven't been mapped,
130  * their info will be passed via a struct tid_user_buf.
131  */
unpin_rcv_pages(struct hfi1_filedata * fd,struct tid_user_buf * tidbuf,struct tid_rb_node * node,unsigned int idx,unsigned int npages,bool mapped)132 static void unpin_rcv_pages(struct hfi1_filedata *fd,
133 			    struct tid_user_buf *tidbuf,
134 			    struct tid_rb_node *node,
135 			    unsigned int idx,
136 			    unsigned int npages,
137 			    bool mapped)
138 {
139 	struct page **pages;
140 	struct hfi1_devdata *dd = fd->uctxt->dd;
141 	struct mm_struct *mm;
142 
143 	if (mapped) {
144 		dma_unmap_single(&dd->pcidev->dev, node->dma_addr,
145 				 node->npages * PAGE_SIZE, DMA_FROM_DEVICE);
146 		pages = &node->pages[idx];
147 		mm = mm_from_tid_node(node);
148 	} else {
149 		pages = &tidbuf->pages[idx];
150 		mm = current->mm;
151 	}
152 	hfi1_release_user_pages(mm, pages, npages, mapped);
153 	fd->tid_n_pinned -= npages;
154 }
155 
156 /*
157  * Pin receive buffer pages.
158  */
pin_rcv_pages(struct hfi1_filedata * fd,struct tid_user_buf * tidbuf)159 static int pin_rcv_pages(struct hfi1_filedata *fd, struct tid_user_buf *tidbuf)
160 {
161 	int pinned;
162 	unsigned int npages = tidbuf->npages;
163 	unsigned long vaddr = tidbuf->vaddr;
164 	struct page **pages = NULL;
165 	struct hfi1_devdata *dd = fd->uctxt->dd;
166 
167 	if (npages > fd->uctxt->expected_count) {
168 		dd_dev_err(dd, "Expected buffer too big\n");
169 		return -EINVAL;
170 	}
171 
172 	/* Allocate the array of struct page pointers needed for pinning */
173 	pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
174 	if (!pages)
175 		return -ENOMEM;
176 
177 	/*
178 	 * Pin all the pages of the user buffer. If we can't pin all the
179 	 * pages, accept the amount pinned so far and program only that.
180 	 * User space knows how to deal with partially programmed buffers.
181 	 */
182 	if (!hfi1_can_pin_pages(dd, current->mm, fd->tid_n_pinned, npages)) {
183 		kfree(pages);
184 		return -ENOMEM;
185 	}
186 
187 	pinned = hfi1_acquire_user_pages(current->mm, vaddr, npages, true, pages);
188 	if (pinned <= 0) {
189 		kfree(pages);
190 		return pinned;
191 	}
192 	tidbuf->pages = pages;
193 	fd->tid_n_pinned += pinned;
194 	return pinned;
195 }
196 
197 /*
198  * RcvArray entry allocation for Expected Receives is done by the
199  * following algorithm:
200  *
201  * The context keeps 3 lists of groups of RcvArray entries:
202  *   1. List of empty groups - tid_group_list
203  *      This list is created during user context creation and
204  *      contains elements which describe sets (of 8) of empty
205  *      RcvArray entries.
206  *   2. List of partially used groups - tid_used_list
207  *      This list contains sets of RcvArray entries which are
208  *      not completely used up. Another mapping request could
209  *      use some of all of the remaining entries.
210  *   3. List of full groups - tid_full_list
211  *      This is the list where sets that are completely used
212  *      up go.
213  *
214  * An attempt to optimize the usage of RcvArray entries is
215  * made by finding all sets of physically contiguous pages in a
216  * user's buffer.
217  * These physically contiguous sets are further split into
218  * sizes supported by the receive engine of the HFI. The
219  * resulting sets of pages are stored in struct tid_pageset,
220  * which describes the sets as:
221  *    * .count - number of pages in this set
222  *    * .idx - starting index into struct page ** array
223  *                    of this set
224  *
225  * From this point on, the algorithm deals with the page sets
226  * described above. The number of pagesets is divided by the
227  * RcvArray group size to produce the number of full groups
228  * needed.
229  *
230  * Groups from the 3 lists are manipulated using the following
231  * rules:
232  *   1. For each set of 8 pagesets, a complete group from
233  *      tid_group_list is taken, programmed, and moved to
234  *      the tid_full_list list.
235  *   2. For all remaining pagesets:
236  *      2.1 If the tid_used_list is empty and the tid_group_list
237  *          is empty, stop processing pageset and return only
238  *          what has been programmed up to this point.
239  *      2.2 If the tid_used_list is empty and the tid_group_list
240  *          is not empty, move a group from tid_group_list to
241  *          tid_used_list.
242  *      2.3 For each group is tid_used_group, program as much as
243  *          can fit into the group. If the group becomes fully
244  *          used, move it to tid_full_list.
245  */
hfi1_user_exp_rcv_setup(struct hfi1_filedata * fd,struct hfi1_tid_info * tinfo)246 int hfi1_user_exp_rcv_setup(struct hfi1_filedata *fd,
247 			    struct hfi1_tid_info *tinfo)
248 {
249 	int ret = 0, need_group = 0, pinned;
250 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
251 	struct hfi1_devdata *dd = uctxt->dd;
252 	unsigned int ngroups, pageset_count,
253 		tididx = 0, mapped, mapped_pages = 0;
254 	u32 *tidlist = NULL;
255 	struct tid_user_buf *tidbuf;
256 	unsigned long mmu_seq = 0;
257 
258 	if (!PAGE_ALIGNED(tinfo->vaddr))
259 		return -EINVAL;
260 	if (tinfo->length == 0)
261 		return -EINVAL;
262 
263 	tidbuf = kzalloc(sizeof(*tidbuf), GFP_KERNEL);
264 	if (!tidbuf)
265 		return -ENOMEM;
266 
267 	mutex_init(&tidbuf->cover_mutex);
268 	tidbuf->vaddr = tinfo->vaddr;
269 	tidbuf->length = tinfo->length;
270 	tidbuf->npages = num_user_pages(tidbuf->vaddr, tidbuf->length);
271 	tidbuf->psets = kcalloc(uctxt->expected_count, sizeof(*tidbuf->psets),
272 				GFP_KERNEL);
273 	if (!tidbuf->psets) {
274 		ret = -ENOMEM;
275 		goto fail_release_mem;
276 	}
277 
278 	if (fd->use_mn) {
279 		ret = mmu_interval_notifier_insert(
280 			&tidbuf->notifier, current->mm,
281 			tidbuf->vaddr, tidbuf->npages * PAGE_SIZE,
282 			&tid_cover_ops);
283 		if (ret)
284 			goto fail_release_mem;
285 		mmu_seq = mmu_interval_read_begin(&tidbuf->notifier);
286 	}
287 
288 	pinned = pin_rcv_pages(fd, tidbuf);
289 	if (pinned <= 0) {
290 		ret = (pinned < 0) ? pinned : -ENOSPC;
291 		goto fail_unpin;
292 	}
293 
294 	/* Find sets of physically contiguous pages */
295 	tidbuf->n_psets = find_phys_blocks(tidbuf, pinned);
296 
297 	/* Reserve the number of expected tids to be used. */
298 	spin_lock(&fd->tid_lock);
299 	if (fd->tid_used + tidbuf->n_psets > fd->tid_limit)
300 		pageset_count = fd->tid_limit - fd->tid_used;
301 	else
302 		pageset_count = tidbuf->n_psets;
303 	fd->tid_used += pageset_count;
304 	spin_unlock(&fd->tid_lock);
305 
306 	if (!pageset_count) {
307 		ret = -ENOSPC;
308 		goto fail_unreserve;
309 	}
310 
311 	ngroups = pageset_count / dd->rcv_entries.group_size;
312 	tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL);
313 	if (!tidlist) {
314 		ret = -ENOMEM;
315 		goto fail_unreserve;
316 	}
317 
318 	tididx = 0;
319 
320 	/*
321 	 * From this point on, we are going to be using shared (between master
322 	 * and subcontexts) context resources. We need to take the lock.
323 	 */
324 	mutex_lock(&uctxt->exp_mutex);
325 	/*
326 	 * The first step is to program the RcvArray entries which are complete
327 	 * groups.
328 	 */
329 	while (ngroups && uctxt->tid_group_list.count) {
330 		struct tid_group *grp =
331 			tid_group_pop(&uctxt->tid_group_list);
332 
333 		ret = program_rcvarray(fd, tidbuf, grp,
334 				       dd->rcv_entries.group_size,
335 				       tidlist, &tididx, &mapped);
336 		/*
337 		 * If there was a failure to program the RcvArray
338 		 * entries for the entire group, reset the grp fields
339 		 * and add the grp back to the free group list.
340 		 */
341 		if (ret <= 0) {
342 			tid_group_add_tail(grp, &uctxt->tid_group_list);
343 			hfi1_cdbg(TID,
344 				  "Failed to program RcvArray group %d", ret);
345 			goto unlock;
346 		}
347 
348 		tid_group_add_tail(grp, &uctxt->tid_full_list);
349 		ngroups--;
350 		mapped_pages += mapped;
351 	}
352 
353 	while (tididx < pageset_count) {
354 		struct tid_group *grp, *ptr;
355 		/*
356 		 * If we don't have any partially used tid groups, check
357 		 * if we have empty groups. If so, take one from there and
358 		 * put in the partially used list.
359 		 */
360 		if (!uctxt->tid_used_list.count || need_group) {
361 			if (!uctxt->tid_group_list.count)
362 				goto unlock;
363 
364 			grp = tid_group_pop(&uctxt->tid_group_list);
365 			tid_group_add_tail(grp, &uctxt->tid_used_list);
366 			need_group = 0;
367 		}
368 		/*
369 		 * There is an optimization opportunity here - instead of
370 		 * fitting as many page sets as we can, check for a group
371 		 * later on in the list that could fit all of them.
372 		 */
373 		list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list,
374 					 list) {
375 			unsigned use = min_t(unsigned, pageset_count - tididx,
376 					     grp->size - grp->used);
377 
378 			ret = program_rcvarray(fd, tidbuf, grp,
379 					       use, tidlist,
380 					       &tididx, &mapped);
381 			if (ret < 0) {
382 				hfi1_cdbg(TID,
383 					  "Failed to program RcvArray entries %d",
384 					  ret);
385 				goto unlock;
386 			} else if (ret > 0) {
387 				if (grp->used == grp->size)
388 					tid_group_move(grp,
389 						       &uctxt->tid_used_list,
390 						       &uctxt->tid_full_list);
391 				mapped_pages += mapped;
392 				need_group = 0;
393 				/* Check if we are done so we break out early */
394 				if (tididx >= pageset_count)
395 					break;
396 			} else if (WARN_ON(ret == 0)) {
397 				/*
398 				 * If ret is 0, we did not program any entries
399 				 * into this group, which can only happen if
400 				 * we've screwed up the accounting somewhere.
401 				 * Warn and try to continue.
402 				 */
403 				need_group = 1;
404 			}
405 		}
406 	}
407 unlock:
408 	mutex_unlock(&uctxt->exp_mutex);
409 	hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx,
410 		  mapped_pages, ret);
411 
412 	/* fail if nothing was programmed, set error if none provided */
413 	if (tididx == 0) {
414 		if (ret >= 0)
415 			ret = -ENOSPC;
416 		goto fail_unreserve;
417 	}
418 
419 	/* adjust reserved tid_used to actual count */
420 	spin_lock(&fd->tid_lock);
421 	fd->tid_used -= pageset_count - tididx;
422 	spin_unlock(&fd->tid_lock);
423 
424 	/* unpin all pages not covered by a TID */
425 	unpin_rcv_pages(fd, tidbuf, NULL, mapped_pages, pinned - mapped_pages,
426 			false);
427 
428 	if (fd->use_mn) {
429 		/* check for an invalidate during setup */
430 		bool fail = false;
431 
432 		mutex_lock(&tidbuf->cover_mutex);
433 		fail = mmu_interval_read_retry(&tidbuf->notifier, mmu_seq);
434 		mutex_unlock(&tidbuf->cover_mutex);
435 
436 		if (fail) {
437 			ret = -EBUSY;
438 			goto fail_unprogram;
439 		}
440 	}
441 
442 	tinfo->tidcnt = tididx;
443 	tinfo->length = mapped_pages * PAGE_SIZE;
444 
445 	if (copy_to_user(u64_to_user_ptr(tinfo->tidlist),
446 			 tidlist, sizeof(tidlist[0]) * tididx)) {
447 		ret = -EFAULT;
448 		goto fail_unprogram;
449 	}
450 
451 	if (fd->use_mn)
452 		mmu_interval_notifier_remove(&tidbuf->notifier);
453 	kfree(tidbuf->pages);
454 	kfree(tidbuf->psets);
455 	kfree(tidbuf);
456 	kfree(tidlist);
457 	return 0;
458 
459 fail_unprogram:
460 	/* unprogram, unmap, and unpin all allocated TIDs */
461 	tinfo->tidlist = (unsigned long)tidlist;
462 	hfi1_user_exp_rcv_clear(fd, tinfo);
463 	tinfo->tidlist = 0;
464 	pinned = 0;		/* nothing left to unpin */
465 	pageset_count = 0;	/* nothing left reserved */
466 fail_unreserve:
467 	spin_lock(&fd->tid_lock);
468 	fd->tid_used -= pageset_count;
469 	spin_unlock(&fd->tid_lock);
470 fail_unpin:
471 	if (fd->use_mn)
472 		mmu_interval_notifier_remove(&tidbuf->notifier);
473 	if (pinned > 0)
474 		unpin_rcv_pages(fd, tidbuf, NULL, 0, pinned, false);
475 fail_release_mem:
476 	kfree(tidbuf->pages);
477 	kfree(tidbuf->psets);
478 	kfree(tidbuf);
479 	kfree(tidlist);
480 	return ret;
481 }
482 
hfi1_user_exp_rcv_clear(struct hfi1_filedata * fd,struct hfi1_tid_info * tinfo)483 int hfi1_user_exp_rcv_clear(struct hfi1_filedata *fd,
484 			    struct hfi1_tid_info *tinfo)
485 {
486 	int ret = 0;
487 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
488 	u32 *tidinfo;
489 	unsigned tididx;
490 
491 	if (unlikely(tinfo->tidcnt > fd->tid_used))
492 		return -EINVAL;
493 
494 	tidinfo = memdup_array_user(u64_to_user_ptr(tinfo->tidlist),
495 				    tinfo->tidcnt, sizeof(tidinfo[0]));
496 	if (IS_ERR(tidinfo))
497 		return PTR_ERR(tidinfo);
498 
499 	mutex_lock(&uctxt->exp_mutex);
500 	for (tididx = 0; tididx < tinfo->tidcnt; tididx++) {
501 		ret = unprogram_rcvarray(fd, tidinfo[tididx]);
502 		if (ret) {
503 			hfi1_cdbg(TID, "Failed to unprogram rcv array %d",
504 				  ret);
505 			break;
506 		}
507 	}
508 	spin_lock(&fd->tid_lock);
509 	fd->tid_used -= tididx;
510 	spin_unlock(&fd->tid_lock);
511 	tinfo->tidcnt = tididx;
512 	mutex_unlock(&uctxt->exp_mutex);
513 
514 	kfree(tidinfo);
515 	return ret;
516 }
517 
hfi1_user_exp_rcv_invalid(struct hfi1_filedata * fd,struct hfi1_tid_info * tinfo)518 int hfi1_user_exp_rcv_invalid(struct hfi1_filedata *fd,
519 			      struct hfi1_tid_info *tinfo)
520 {
521 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
522 	unsigned long *ev = uctxt->dd->events +
523 		(uctxt_offset(uctxt) + fd->subctxt);
524 	u32 *array;
525 	int ret = 0;
526 
527 	/*
528 	 * copy_to_user() can sleep, which will leave the invalid_lock
529 	 * locked and cause the MMU notifier to be blocked on the lock
530 	 * for a long time.
531 	 * Copy the data to a local buffer so we can release the lock.
532 	 */
533 	array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL);
534 	if (!array)
535 		return -EFAULT;
536 
537 	spin_lock(&fd->invalid_lock);
538 	if (fd->invalid_tid_idx) {
539 		memcpy(array, fd->invalid_tids, sizeof(*array) *
540 		       fd->invalid_tid_idx);
541 		memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) *
542 		       fd->invalid_tid_idx);
543 		tinfo->tidcnt = fd->invalid_tid_idx;
544 		fd->invalid_tid_idx = 0;
545 		/*
546 		 * Reset the user flag while still holding the lock.
547 		 * Otherwise, PSM can miss events.
548 		 */
549 		clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
550 	} else {
551 		tinfo->tidcnt = 0;
552 	}
553 	spin_unlock(&fd->invalid_lock);
554 
555 	if (tinfo->tidcnt) {
556 		if (copy_to_user((void __user *)tinfo->tidlist,
557 				 array, sizeof(*array) * tinfo->tidcnt))
558 			ret = -EFAULT;
559 	}
560 	kfree(array);
561 
562 	return ret;
563 }
564 
find_phys_blocks(struct tid_user_buf * tidbuf,unsigned int npages)565 static u32 find_phys_blocks(struct tid_user_buf *tidbuf, unsigned int npages)
566 {
567 	unsigned pagecount, pageidx, setcount = 0, i;
568 	unsigned long pfn, this_pfn;
569 	struct page **pages = tidbuf->pages;
570 	struct tid_pageset *list = tidbuf->psets;
571 
572 	if (!npages)
573 		return 0;
574 
575 	/*
576 	 * Look for sets of physically contiguous pages in the user buffer.
577 	 * This will allow us to optimize Expected RcvArray entry usage by
578 	 * using the bigger supported sizes.
579 	 */
580 	pfn = page_to_pfn(pages[0]);
581 	for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) {
582 		this_pfn = i < npages ? page_to_pfn(pages[i]) : 0;
583 
584 		/*
585 		 * If the pfn's are not sequential, pages are not physically
586 		 * contiguous.
587 		 */
588 		if (this_pfn != ++pfn) {
589 			/*
590 			 * At this point we have to loop over the set of
591 			 * physically contiguous pages and break them down it
592 			 * sizes supported by the HW.
593 			 * There are two main constraints:
594 			 *     1. The max buffer size is MAX_EXPECTED_BUFFER.
595 			 *        If the total set size is bigger than that
596 			 *        program only a MAX_EXPECTED_BUFFER chunk.
597 			 *     2. The buffer size has to be a power of two. If
598 			 *        it is not, round down to the closes power of
599 			 *        2 and program that size.
600 			 */
601 			while (pagecount) {
602 				int maxpages = pagecount;
603 				u32 bufsize = pagecount * PAGE_SIZE;
604 
605 				if (bufsize > MAX_EXPECTED_BUFFER)
606 					maxpages =
607 						MAX_EXPECTED_BUFFER >>
608 						PAGE_SHIFT;
609 				else if (!is_power_of_2(bufsize))
610 					maxpages =
611 						rounddown_pow_of_two(bufsize) >>
612 						PAGE_SHIFT;
613 
614 				list[setcount].idx = pageidx;
615 				list[setcount].count = maxpages;
616 				pagecount -= maxpages;
617 				pageidx += maxpages;
618 				setcount++;
619 			}
620 			pageidx = i;
621 			pagecount = 1;
622 			pfn = this_pfn;
623 		} else {
624 			pagecount++;
625 		}
626 	}
627 	return setcount;
628 }
629 
630 /**
631  * program_rcvarray() - program an RcvArray group with receive buffers
632  * @fd: filedata pointer
633  * @tbuf: pointer to struct tid_user_buf that has the user buffer starting
634  *	  virtual address, buffer length, page pointers, pagesets (array of
635  *	  struct tid_pageset holding information on physically contiguous
636  *	  chunks from the user buffer), and other fields.
637  * @grp: RcvArray group
638  * @count: number of struct tid_pageset's to program
639  * @tidlist: the array of u32 elements when the information about the
640  *           programmed RcvArray entries is to be encoded.
641  * @tididx: starting offset into tidlist
642  * @pmapped: (output parameter) number of pages programmed into the RcvArray
643  *           entries.
644  *
645  * This function will program up to 'count' number of RcvArray entries from the
646  * group 'grp'. To make best use of write-combining writes, the function will
647  * perform writes to the unused RcvArray entries which will be ignored by the
648  * HW. Each RcvArray entry will be programmed with a physically contiguous
649  * buffer chunk from the user's virtual buffer.
650  *
651  * Return:
652  * -EINVAL if the requested count is larger than the size of the group,
653  * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or
654  * number of RcvArray entries programmed.
655  */
program_rcvarray(struct hfi1_filedata * fd,struct tid_user_buf * tbuf,struct tid_group * grp,u16 count,u32 * tidlist,unsigned int * tididx,unsigned int * pmapped)656 static int program_rcvarray(struct hfi1_filedata *fd, struct tid_user_buf *tbuf,
657 			    struct tid_group *grp, u16 count,
658 			    u32 *tidlist, unsigned int *tididx,
659 			    unsigned int *pmapped)
660 {
661 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
662 	struct hfi1_devdata *dd = uctxt->dd;
663 	u16 idx;
664 	unsigned int start = *tididx;
665 	u32 tidinfo = 0, rcventry, useidx = 0;
666 	int mapped = 0;
667 
668 	/* Count should never be larger than the group size */
669 	if (count > grp->size)
670 		return -EINVAL;
671 
672 	/* Find the first unused entry in the group */
673 	for (idx = 0; idx < grp->size; idx++) {
674 		if (!(grp->map & (1 << idx))) {
675 			useidx = idx;
676 			break;
677 		}
678 		rcv_array_wc_fill(dd, grp->base + idx);
679 	}
680 
681 	idx = 0;
682 	while (idx < count) {
683 		u16 npages, pageidx, setidx = start + idx;
684 		int ret = 0;
685 
686 		/*
687 		 * If this entry in the group is used, move to the next one.
688 		 * If we go past the end of the group, exit the loop.
689 		 */
690 		if (useidx >= grp->size) {
691 			break;
692 		} else if (grp->map & (1 << useidx)) {
693 			rcv_array_wc_fill(dd, grp->base + useidx);
694 			useidx++;
695 			continue;
696 		}
697 
698 		rcventry = grp->base + useidx;
699 		npages = tbuf->psets[setidx].count;
700 		pageidx = tbuf->psets[setidx].idx;
701 
702 		ret = set_rcvarray_entry(fd, tbuf,
703 					 rcventry, grp, pageidx,
704 					 npages);
705 		if (ret)
706 			return ret;
707 		mapped += npages;
708 
709 		tidinfo = create_tid(rcventry - uctxt->expected_base, npages);
710 		tidlist[(*tididx)++] = tidinfo;
711 		grp->used++;
712 		grp->map |= 1 << useidx++;
713 		idx++;
714 	}
715 
716 	/* Fill the rest of the group with "blank" writes */
717 	for (; useidx < grp->size; useidx++)
718 		rcv_array_wc_fill(dd, grp->base + useidx);
719 	*pmapped = mapped;
720 	return idx;
721 }
722 
set_rcvarray_entry(struct hfi1_filedata * fd,struct tid_user_buf * tbuf,u32 rcventry,struct tid_group * grp,u16 pageidx,unsigned int npages)723 static int set_rcvarray_entry(struct hfi1_filedata *fd,
724 			      struct tid_user_buf *tbuf,
725 			      u32 rcventry, struct tid_group *grp,
726 			      u16 pageidx, unsigned int npages)
727 {
728 	int ret;
729 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
730 	struct tid_rb_node *node;
731 	struct hfi1_devdata *dd = uctxt->dd;
732 	dma_addr_t phys;
733 	struct page **pages = tbuf->pages + pageidx;
734 
735 	/*
736 	 * Allocate the node first so we can handle a potential
737 	 * failure before we've programmed anything.
738 	 */
739 	node = kzalloc(struct_size(node, pages, npages), GFP_KERNEL);
740 	if (!node)
741 		return -ENOMEM;
742 
743 	phys = dma_map_single(&dd->pcidev->dev, __va(page_to_phys(pages[0])),
744 			      npages * PAGE_SIZE, DMA_FROM_DEVICE);
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 	mutex_init(&node->invalidate_mutex);
754 	node->phys = page_to_phys(pages[0]);
755 	node->npages = npages;
756 	node->rcventry = rcventry;
757 	node->dma_addr = phys;
758 	node->grp = grp;
759 	node->freed = false;
760 	memcpy(node->pages, pages, flex_array_size(node, pages, npages));
761 
762 	if (fd->use_mn) {
763 		ret = mmu_interval_notifier_insert(
764 			&node->notifier, current->mm,
765 			tbuf->vaddr + (pageidx * PAGE_SIZE), npages * PAGE_SIZE,
766 			&tid_mn_ops);
767 		if (ret)
768 			goto out_unmap;
769 	}
770 	fd->entry_to_rb[node->rcventry - uctxt->expected_base] = node;
771 
772 	hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1);
773 	trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, npages,
774 			       node->notifier.interval_tree.start, node->phys,
775 			       phys);
776 	return 0;
777 
778 out_unmap:
779 	hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d",
780 		  node->rcventry, node->notifier.interval_tree.start,
781 		  node->phys, ret);
782 	dma_unmap_single(&dd->pcidev->dev, phys, npages * PAGE_SIZE,
783 			 DMA_FROM_DEVICE);
784 	kfree(node);
785 	return -EFAULT;
786 }
787 
unprogram_rcvarray(struct hfi1_filedata * fd,u32 tidinfo)788 static int unprogram_rcvarray(struct hfi1_filedata *fd, u32 tidinfo)
789 {
790 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
791 	struct hfi1_devdata *dd = uctxt->dd;
792 	struct tid_rb_node *node;
793 	u32 tidctrl = EXP_TID_GET(tidinfo, CTRL);
794 	u32 tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry;
795 
796 	if (tidctrl == 0x3 || tidctrl == 0x0)
797 		return -EINVAL;
798 
799 	rcventry = tididx + (tidctrl - 1);
800 
801 	if (rcventry >= uctxt->expected_count) {
802 		dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n",
803 			   rcventry, uctxt->ctxt);
804 		return -EINVAL;
805 	}
806 
807 	node = fd->entry_to_rb[rcventry];
808 	if (!node || node->rcventry != (uctxt->expected_base + rcventry))
809 		return -EBADF;
810 
811 	if (fd->use_mn)
812 		mmu_interval_notifier_remove(&node->notifier);
813 	cacheless_tid_rb_remove(fd, node);
814 
815 	return 0;
816 }
817 
__clear_tid_node(struct hfi1_filedata * fd,struct tid_rb_node * node)818 static void __clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
819 {
820 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
821 	struct hfi1_devdata *dd = uctxt->dd;
822 
823 	mutex_lock(&node->invalidate_mutex);
824 	if (node->freed)
825 		goto done;
826 	node->freed = true;
827 
828 	trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry,
829 				 node->npages,
830 				 node->notifier.interval_tree.start, node->phys,
831 				 node->dma_addr);
832 
833 	/* Make sure device has seen the write before pages are unpinned */
834 	hfi1_put_tid(dd, node->rcventry, PT_INVALID_FLUSH, 0, 0);
835 
836 	unpin_rcv_pages(fd, NULL, node, 0, node->npages, true);
837 done:
838 	mutex_unlock(&node->invalidate_mutex);
839 }
840 
clear_tid_node(struct hfi1_filedata * fd,struct tid_rb_node * node)841 static void clear_tid_node(struct hfi1_filedata *fd, struct tid_rb_node *node)
842 {
843 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
844 
845 	__clear_tid_node(fd, node);
846 
847 	node->grp->used--;
848 	node->grp->map &= ~(1 << (node->rcventry - node->grp->base));
849 
850 	if (node->grp->used == node->grp->size - 1)
851 		tid_group_move(node->grp, &uctxt->tid_full_list,
852 			       &uctxt->tid_used_list);
853 	else if (!node->grp->used)
854 		tid_group_move(node->grp, &uctxt->tid_used_list,
855 			       &uctxt->tid_group_list);
856 	kfree(node);
857 }
858 
859 /*
860  * As a simple helper for hfi1_user_exp_rcv_free, this function deals with
861  * clearing nodes in the non-cached case.
862  */
unlock_exp_tids(struct hfi1_ctxtdata * uctxt,struct exp_tid_set * set,struct hfi1_filedata * fd)863 static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt,
864 			    struct exp_tid_set *set,
865 			    struct hfi1_filedata *fd)
866 {
867 	struct tid_group *grp, *ptr;
868 	int i;
869 
870 	list_for_each_entry_safe(grp, ptr, &set->list, list) {
871 		list_del_init(&grp->list);
872 
873 		for (i = 0; i < grp->size; i++) {
874 			if (grp->map & (1 << i)) {
875 				u16 rcventry = grp->base + i;
876 				struct tid_rb_node *node;
877 
878 				node = fd->entry_to_rb[rcventry -
879 							  uctxt->expected_base];
880 				if (!node || node->rcventry != rcventry)
881 					continue;
882 
883 				if (fd->use_mn)
884 					mmu_interval_notifier_remove(
885 						&node->notifier);
886 				cacheless_tid_rb_remove(fd, node);
887 			}
888 		}
889 	}
890 }
891 
tid_rb_invalidate(struct mmu_interval_notifier * mni,const struct mmu_notifier_range * range,unsigned long cur_seq)892 static bool tid_rb_invalidate(struct mmu_interval_notifier *mni,
893 			      const struct mmu_notifier_range *range,
894 			      unsigned long cur_seq)
895 {
896 	struct tid_rb_node *node =
897 		container_of(mni, struct tid_rb_node, notifier);
898 	struct hfi1_filedata *fdata = node->fdata;
899 	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
900 
901 	if (node->freed)
902 		return true;
903 
904 	/* take action only if unmapping */
905 	if (range->event != MMU_NOTIFY_UNMAP)
906 		return true;
907 
908 	trace_hfi1_exp_tid_inval(uctxt->ctxt, fdata->subctxt,
909 				 node->notifier.interval_tree.start,
910 				 node->rcventry, node->npages, node->dma_addr);
911 
912 	/* clear the hardware rcvarray entry */
913 	__clear_tid_node(fdata, node);
914 
915 	spin_lock(&fdata->invalid_lock);
916 	if (fdata->invalid_tid_idx < uctxt->expected_count) {
917 		fdata->invalid_tids[fdata->invalid_tid_idx] =
918 			create_tid(node->rcventry - uctxt->expected_base,
919 				   node->npages);
920 		if (!fdata->invalid_tid_idx) {
921 			unsigned long *ev;
922 
923 			/*
924 			 * hfi1_set_uevent_bits() sets a user event flag
925 			 * for all processes. Because calling into the
926 			 * driver to process TID cache invalidations is
927 			 * expensive and TID cache invalidations are
928 			 * handled on a per-process basis, we can
929 			 * optimize this to set the flag only for the
930 			 * process in question.
931 			 */
932 			ev = uctxt->dd->events +
933 				(uctxt_offset(uctxt) + fdata->subctxt);
934 			set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev);
935 		}
936 		fdata->invalid_tid_idx++;
937 	}
938 	spin_unlock(&fdata->invalid_lock);
939 	return true;
940 }
941 
tid_cover_invalidate(struct mmu_interval_notifier * mni,const struct mmu_notifier_range * range,unsigned long cur_seq)942 static bool tid_cover_invalidate(struct mmu_interval_notifier *mni,
943 			         const struct mmu_notifier_range *range,
944 			         unsigned long cur_seq)
945 {
946 	struct tid_user_buf *tidbuf =
947 		container_of(mni, struct tid_user_buf, notifier);
948 
949 	/* take action only if unmapping */
950 	if (range->event == MMU_NOTIFY_UNMAP) {
951 		mutex_lock(&tidbuf->cover_mutex);
952 		mmu_interval_set_seq(mni, cur_seq);
953 		mutex_unlock(&tidbuf->cover_mutex);
954 	}
955 
956 	return true;
957 }
958 
cacheless_tid_rb_remove(struct hfi1_filedata * fdata,struct tid_rb_node * tnode)959 static void cacheless_tid_rb_remove(struct hfi1_filedata *fdata,
960 				    struct tid_rb_node *tnode)
961 {
962 	u32 base = fdata->uctxt->expected_base;
963 
964 	fdata->entry_to_rb[tnode->rcventry - base] = NULL;
965 	clear_tid_node(fdata, tnode);
966 }
967