1/*
2 * Copyright (c) 2005 Topspin Communications.  All rights reserved.
3 * Copyright (c) 2005 Cisco Systems.  All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5 *
6 * This software is available to you under a choice of one of two
7 * licenses.  You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
11 *
12 *     Redistribution and use in source and binary forms, with or
13 *     without modification, are permitted provided that the following
14 *     conditions are met:
15 *
16 *      - Redistributions of source code must retain the above
17 *        copyright notice, this list of conditions and the following
18 *        disclaimer.
19 *
20 *      - Redistributions in binary form must reproduce the above
21 *        copyright notice, this list of conditions and the following
22 *        disclaimer in the documentation and/or other materials
23 *        provided with the distribution.
24 *
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 * SOFTWARE.
33 */
34
35#include <linux/mm.h>
36#include <linux/dma-mapping.h>
37#include <linux/sched.h>
38#ifdef __linux__
39#include <linux/hugetlb.h>
40#endif
41#include <linux/dma-attrs.h>
42
43#include <sys/priv.h>
44#include <sys/resource.h>
45#include <sys/resourcevar.h>
46
47#include <vm/vm.h>
48#include <vm/vm_map.h>
49#include <vm/vm_object.h>
50#include <vm/vm_pageout.h>
51
52#include "uverbs.h"
53
54static int allow_weak_ordering;
55module_param(allow_weak_ordering, bool, 0444);
56MODULE_PARM_DESC(allow_weak_ordering,  "Allow weak ordering for data registered memory");
57
58#define IB_UMEM_MAX_PAGE_CHUNK						\
59	((PAGE_SIZE - offsetof(struct ib_umem_chunk, page_list)) /	\
60	 ((void *) &((struct ib_umem_chunk *) 0)->page_list[1] -	\
61	  (void *) &((struct ib_umem_chunk *) 0)->page_list[0]))
62
63#ifdef __ia64__
64extern int dma_map_sg_hp_wa;
65
66static int dma_map_sg_ia64(struct ib_device *ibdev,
67			   struct scatterlist *sg,
68			   int nents,
69			   enum dma_data_direction dir)
70{
71	int i, rc, j, lents = 0;
72	struct device *dev;
73
74	if (!dma_map_sg_hp_wa)
75		return ib_dma_map_sg(ibdev, sg, nents, dir);
76
77	dev = ibdev->dma_device;
78	for (i = 0; i < nents; ++i) {
79		rc = dma_map_sg(dev, sg + i, 1, dir);
80		if (rc <= 0) {
81			for (j = 0; j < i; ++j)
82				dma_unmap_sg(dev, sg + j, 1, dir);
83
84			return 0;
85		}
86		lents += rc;
87	}
88
89	return lents;
90}
91
92static void dma_unmap_sg_ia64(struct ib_device *ibdev,
93			      struct scatterlist *sg,
94			      int nents,
95			      enum dma_data_direction dir)
96{
97	int i;
98	struct device *dev;
99
100	if (!dma_map_sg_hp_wa)
101		return ib_dma_unmap_sg(ibdev, sg, nents, dir);
102
103	dev = ibdev->dma_device;
104	for (i = 0; i < nents; ++i)
105		dma_unmap_sg(dev, sg + i, 1, dir);
106}
107
108#define ib_dma_map_sg(dev, sg, nents, dir) dma_map_sg_ia64(dev, sg, nents, dir)
109#define ib_dma_unmap_sg(dev, sg, nents, dir) dma_unmap_sg_ia64(dev, sg, nents, dir)
110
111#endif
112
113static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
114{
115#ifdef __linux__
116	struct ib_umem_chunk *chunk, *tmp;
117	int i;
118
119	list_for_each_entry_safe(chunk, tmp, &umem->chunk_list, list) {
120		ib_dma_unmap_sg_attrs(dev, chunk->page_list,
121				      chunk->nents, DMA_BIDIRECTIONAL, &chunk->attrs);
122		for (i = 0; i < chunk->nents; ++i) {
123			struct page *page = sg_page(&chunk->page_list[i]);
124			if (umem->writable && dirty)
125				set_page_dirty_lock(page);
126			put_page(page);
127		}
128		kfree(chunk);
129	}
130#else
131	struct ib_umem_chunk *chunk, *tmp;
132	vm_object_t object;
133	int i;
134
135	object = NULL;
136	list_for_each_entry_safe(chunk, tmp, &umem->chunk_list, list) {
137		ib_dma_unmap_sg_attrs(dev, chunk->page_list,
138				      chunk->nents, DMA_BIDIRECTIONAL, &chunk->attrs);
139		for (i = 0; i < chunk->nents; ++i) {
140			struct page *page = sg_page(&chunk->page_list[i]);
141			if (umem->writable && dirty) {
142				if (object && object != page->object)
143					VM_OBJECT_WUNLOCK(object);
144				if (object != page->object) {
145					object = page->object;
146					VM_OBJECT_WLOCK(object);
147				}
148				vm_page_dirty(page);
149			}
150		}
151		kfree(chunk);
152	}
153	if (object)
154		VM_OBJECT_WUNLOCK(object);
155
156#endif
157}
158
159/**
160 * ib_umem_get - Pin and DMA map userspace memory.
161 * @context: userspace context to pin memory for
162 * @addr: userspace virtual address to start at
163 * @size: length of region to pin
164 * @access: IB_ACCESS_xxx flags for memory being pinned
165 * @dmasync: flush in-flight DMA when the memory region is written
166 */
167struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
168			    size_t size, int access, int dmasync)
169{
170#ifdef __linux__
171	struct ib_umem *umem;
172	struct page **page_list;
173	struct vm_area_struct **vma_list;
174	struct ib_umem_chunk *chunk;
175	unsigned long locked;
176	unsigned long lock_limit;
177	unsigned long cur_base;
178	unsigned long npages;
179	int ret;
180	int off;
181	int i;
182	DEFINE_DMA_ATTRS(attrs);
183
184	if (dmasync)
185		dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
186	else if (allow_weak_ordering)
187		dma_set_attr(DMA_ATTR_WEAK_ORDERING, &attrs);
188
189	if (!can_do_mlock())
190		return ERR_PTR(-EPERM);
191
192	umem = kmalloc(sizeof *umem, GFP_KERNEL);
193	if (!umem)
194		return ERR_PTR(-ENOMEM);
195
196	umem->context   = context;
197	umem->length    = size;
198	umem->offset    = addr & ~PAGE_MASK;
199	umem->page_size = PAGE_SIZE;
200	/*
201	 * We ask for writable memory if any access flags other than
202	 * "remote read" are set.  "Local write" and "remote write"
203	 * obviously require write access.  "Remote atomic" can do
204	 * things like fetch and add, which will modify memory, and
205	 * "MW bind" can change permissions by binding a window.
206	 */
207	umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
208
209	/* We assume the memory is from hugetlb until proved otherwise */
210	umem->hugetlb   = 1;
211
212	INIT_LIST_HEAD(&umem->chunk_list);
213
214	page_list = (struct page **) __get_free_page(GFP_KERNEL);
215	if (!page_list) {
216		kfree(umem);
217		return ERR_PTR(-ENOMEM);
218	}
219
220	/*
221	 * if we can't alloc the vma_list, it's not so bad;
222	 * just assume the memory is not hugetlb memory
223	 */
224	vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
225	if (!vma_list)
226		umem->hugetlb = 0;
227
228	npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
229
230	down_write(&current->mm->mmap_sem);
231
232	locked     = npages + current->mm->locked_vm;
233	lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
234
235	if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
236		ret = -ENOMEM;
237		goto out;
238	}
239
240	cur_base = addr & PAGE_MASK;
241
242	ret = 0;
243
244	while (npages) {
245		ret = get_user_pages(current, current->mm, cur_base,
246				     min_t(unsigned long, npages,
247					   PAGE_SIZE / sizeof (struct page *)),
248				     1, !umem->writable, page_list, vma_list);
249
250		if (ret < 0)
251			goto out;
252
253		cur_base += ret * PAGE_SIZE;
254		npages   -= ret;
255
256		off = 0;
257
258		while (ret) {
259			chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
260					min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
261					GFP_KERNEL);
262			if (!chunk) {
263				ret = -ENOMEM;
264				goto out;
265			}
266
267			chunk->attrs = attrs;
268			chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
269			sg_init_table(chunk->page_list, chunk->nents);
270			for (i = 0; i < chunk->nents; ++i) {
271				if (vma_list &&
272				    !is_vm_hugetlb_page(vma_list[i + off]))
273					umem->hugetlb = 0;
274				sg_set_page(&chunk->page_list[i], page_list[i + off], PAGE_SIZE, 0);
275			}
276
277			chunk->nmap = ib_dma_map_sg_attrs(context->device,
278							  &chunk->page_list[0],
279							  chunk->nents,
280							  DMA_BIDIRECTIONAL,
281							  &attrs);
282			if (chunk->nmap <= 0) {
283				for (i = 0; i < chunk->nents; ++i)
284					put_page(sg_page(&chunk->page_list[i]));
285				kfree(chunk);
286
287				ret = -ENOMEM;
288				goto out;
289			}
290
291			ret -= chunk->nents;
292			off += chunk->nents;
293			list_add_tail(&chunk->list, &umem->chunk_list);
294		}
295
296		ret = 0;
297	}
298
299out:
300	if (ret < 0) {
301		__ib_umem_release(context->device, umem, 0);
302		kfree(umem);
303	} else
304		current->mm->locked_vm = locked;
305
306	up_write(&current->mm->mmap_sem);
307	if (vma_list)
308		free_page((unsigned long) vma_list);
309	free_page((unsigned long) page_list);
310
311	return ret < 0 ? ERR_PTR(ret) : umem;
312#else
313	struct ib_umem *umem;
314	struct ib_umem_chunk *chunk;
315        struct proc *proc;
316	pmap_t pmap;
317        vm_offset_t end, last, start;
318        vm_size_t npages;
319        int error;
320	int ents;
321	int ret;
322	int i;
323	DEFINE_DMA_ATTRS(attrs);
324
325	error = priv_check(curthread, PRIV_VM_MLOCK);
326	if (error)
327		return ERR_PTR(-error);
328
329	last = addr + size;
330	start = addr & PAGE_MASK; /* Use the linux PAGE_MASK definition. */
331	end = roundup2(last, PAGE_SIZE); /* Use PAGE_MASK safe operation. */
332	if (last < addr || end < addr)
333		return ERR_PTR(-EINVAL);
334	npages = atop(end - start);
335	if (npages > vm_page_max_wired)
336		return ERR_PTR(-ENOMEM);
337	umem = kzalloc(sizeof *umem, GFP_KERNEL);
338	if (!umem)
339		return ERR_PTR(-ENOMEM);
340	proc = curthread->td_proc;
341	PROC_LOCK(proc);
342	if (ptoa(npages +
343	    pmap_wired_count(vm_map_pmap(&proc->p_vmspace->vm_map))) >
344	    lim_cur(proc, RLIMIT_MEMLOCK)) {
345		PROC_UNLOCK(proc);
346		kfree(umem);
347		return ERR_PTR(-ENOMEM);
348	}
349        PROC_UNLOCK(proc);
350	if (npages + cnt.v_wire_count > vm_page_max_wired) {
351		kfree(umem);
352		return ERR_PTR(-EAGAIN);
353	}
354	error = vm_map_wire(&proc->p_vmspace->vm_map, start, end,
355	    VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES |
356	    (umem->writable ? VM_MAP_WIRE_WRITE : 0));
357	if (error != KERN_SUCCESS) {
358		kfree(umem);
359		return ERR_PTR(-ENOMEM);
360	}
361
362	umem->context   = context;
363	umem->length    = size;
364	umem->offset    = addr & ~PAGE_MASK;
365	umem->page_size = PAGE_SIZE;
366	umem->start	= addr;
367	/*
368	 * We ask for writable memory if any access flags other than
369	 * "remote read" are set.  "Local write" and "remote write"
370	 * obviously require write access.  "Remote atomic" can do
371	 * things like fetch and add, which will modify memory, and
372	 * "MW bind" can change permissions by binding a window.
373	 */
374	umem->writable  = !!(access & ~IB_ACCESS_REMOTE_READ);
375	umem->hugetlb = 0;
376	INIT_LIST_HEAD(&umem->chunk_list);
377
378	pmap = vm_map_pmap(&proc->p_vmspace->vm_map);
379	ret = 0;
380	while (npages) {
381		ents = min_t(int, npages, IB_UMEM_MAX_PAGE_CHUNK);
382		chunk = kmalloc(sizeof(*chunk) +
383				(sizeof(struct scatterlist) * ents),
384				GFP_KERNEL);
385		if (!chunk) {
386			ret = -ENOMEM;
387			goto out;
388		}
389
390		chunk->attrs = attrs;
391		chunk->nents = ents;
392		sg_init_table(&chunk->page_list[0], ents);
393		for (i = 0; i < chunk->nents; ++i) {
394			vm_paddr_t pa;
395
396			pa = pmap_extract(pmap, start);
397			if (pa == 0) {
398				ret = -ENOMEM;
399				kfree(chunk);
400				goto out;
401			}
402			sg_set_page(&chunk->page_list[i], PHYS_TO_VM_PAGE(pa),
403			    PAGE_SIZE, 0);
404			npages--;
405			start += PAGE_SIZE;
406		}
407
408		chunk->nmap = ib_dma_map_sg_attrs(context->device,
409						  &chunk->page_list[0],
410						  chunk->nents,
411						  DMA_BIDIRECTIONAL,
412						  &attrs);
413		if (chunk->nmap != chunk->nents) {
414			kfree(chunk);
415			ret = -ENOMEM;
416			goto out;
417		}
418
419		list_add_tail(&chunk->list, &umem->chunk_list);
420	}
421
422out:
423	if (ret < 0) {
424		__ib_umem_release(context->device, umem, 0);
425		kfree(umem);
426	}
427
428	return ret < 0 ? ERR_PTR(ret) : umem;
429#endif
430}
431EXPORT_SYMBOL(ib_umem_get);
432
433#ifdef __linux__
434static void ib_umem_account(struct work_struct *work)
435{
436	struct ib_umem *umem = container_of(work, struct ib_umem, work);
437
438	down_write(&umem->mm->mmap_sem);
439	umem->mm->locked_vm -= umem->diff;
440	up_write(&umem->mm->mmap_sem);
441	mmput(umem->mm);
442	kfree(umem);
443}
444#endif
445
446/**
447 * ib_umem_release - release memory pinned with ib_umem_get
448 * @umem: umem struct to release
449 */
450void ib_umem_release(struct ib_umem *umem)
451{
452#ifdef __linux__
453	struct ib_ucontext *context = umem->context;
454	struct mm_struct *mm;
455	unsigned long diff;
456
457	__ib_umem_release(umem->context->device, umem, 1);
458
459	mm = get_task_mm(current);
460	if (!mm) {
461		kfree(umem);
462		return;
463	}
464
465	diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
466
467	/*
468	 * We may be called with the mm's mmap_sem already held.  This
469	 * can happen when a userspace munmap() is the call that drops
470	 * the last reference to our file and calls our release
471	 * method.  If there are memory regions to destroy, we'll end
472	 * up here and not be able to take the mmap_sem.  In that case
473	 * we defer the vm_locked accounting to the system workqueue.
474	 */
475	if (context->closing) {
476		if (!down_write_trylock(&mm->mmap_sem)) {
477			INIT_WORK(&umem->work, ib_umem_account);
478			umem->mm   = mm;
479			umem->diff = diff;
480
481			schedule_work(&umem->work);
482			return;
483		}
484	} else
485		down_write(&mm->mmap_sem);
486
487	current->mm->locked_vm -= diff;
488	up_write(&mm->mmap_sem);
489	mmput(mm);
490#else
491	vm_offset_t addr, end, last, start;
492	vm_size_t size;
493	int error;
494
495	__ib_umem_release(umem->context->device, umem, 1);
496	if (umem->context->closing) {
497		kfree(umem);
498		return;
499	}
500	error = priv_check(curthread, PRIV_VM_MUNLOCK);
501	if (error)
502		return;
503	addr = umem->start;
504	size = umem->length;
505	last = addr + size;
506        start = addr & PAGE_MASK; /* Use the linux PAGE_MASK definition. */
507	end = roundup2(last, PAGE_SIZE); /* Use PAGE_MASK safe operation. */
508	vm_map_unwire(&curthread->td_proc->p_vmspace->vm_map, start, end,
509	    VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES);
510
511#endif
512	kfree(umem);
513}
514EXPORT_SYMBOL(ib_umem_release);
515
516int ib_umem_page_count(struct ib_umem *umem)
517{
518	struct ib_umem_chunk *chunk;
519	int shift;
520	int i;
521	int n;
522
523	shift = ilog2(umem->page_size);
524
525	n = 0;
526	list_for_each_entry(chunk, &umem->chunk_list, list)
527		for (i = 0; i < chunk->nmap; ++i)
528			n += sg_dma_len(&chunk->page_list[i]) >> shift;
529
530	return n;
531}
532EXPORT_SYMBOL(ib_umem_page_count);
533
534/**********************************************/
535/*
536 * Stub functions for contiguous pages -
537 * We currently do not support this feature
538 */
539/**********************************************/
540
541/**
542 * ib_cmem_release_contiguous_pages - release memory allocated by
543 *                                              ib_cmem_alloc_contiguous_pages.
544 * @cmem: cmem struct to release
545 */
546void ib_cmem_release_contiguous_pages(struct ib_cmem *cmem)
547{
548}
549EXPORT_SYMBOL(ib_cmem_release_contiguous_pages);
550
551/**
552 *  * ib_cmem_alloc_contiguous_pages - allocate contiguous pages
553 *  *  @context: userspace context to allocate memory for
554 *   * @total_size: total required size for that allocation.
555 *    * @page_size_order: order of one contiguous page.
556 *     */
557struct ib_cmem *ib_cmem_alloc_contiguous_pages(struct ib_ucontext *context,
558                                unsigned long total_size,
559				                                unsigned long page_size_order)
560{
561	return NULL;
562}
563EXPORT_SYMBOL(ib_cmem_alloc_contiguous_pages);
564
565/**
566 *  * ib_cmem_map_contiguous_pages_to_vma - map contiguous pages into VMA
567 *   * @ib_cmem: cmem structure returned by ib_cmem_alloc_contiguous_pages
568 *    * @vma: VMA to inject pages into.
569 *     */
570int ib_cmem_map_contiguous_pages_to_vma(struct ib_cmem *ib_cmem,
571                                                struct vm_area_struct *vma)
572{
573	return 0;
574}
575EXPORT_SYMBOL(ib_cmem_map_contiguous_pages_to_vma);
576