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