Linux Audio

Check our new training course

Loading...
v6.8
  1/*
  2 * Copyright (c) 2014 Mellanox Technologies. All rights reserved.
  3 *
  4 * This software is available to you under a choice of one of two
  5 * licenses.  You may choose to be licensed under the terms of the GNU
  6 * General Public License (GPL) Version 2, available from the file
  7 * COPYING in the main directory of this source tree, or the
  8 * OpenIB.org BSD license below:
  9 *
 10 *     Redistribution and use in source and binary forms, with or
 11 *     without modification, are permitted provided that the following
 12 *     conditions are met:
 13 *
 14 *      - Redistributions of source code must retain the above
 15 *        copyright notice, this list of conditions and the following
 16 *        disclaimer.
 17 *
 18 *      - Redistributions in binary form must reproduce the above
 19 *        copyright notice, this list of conditions and the following
 20 *        disclaimer in the documentation and/or other materials
 21 *        provided with the distribution.
 22 *
 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 30 * SOFTWARE.
 31 */
 32
 33#include <linux/types.h>
 34#include <linux/sched.h>
 35#include <linux/sched/mm.h>
 36#include <linux/sched/task.h>
 37#include <linux/pid.h>
 38#include <linux/slab.h>
 39#include <linux/export.h>
 40#include <linux/vmalloc.h>
 41#include <linux/hugetlb.h>
 42#include <linux/interval_tree.h>
 43#include <linux/hmm.h>
 44#include <linux/pagemap.h>
 45
 
 
 46#include <rdma/ib_umem_odp.h>
 47
 48#include "uverbs.h"
 49
 50static inline int ib_init_umem_odp(struct ib_umem_odp *umem_odp,
 51				   const struct mmu_interval_notifier_ops *ops)
 52{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 53	int ret;
 54
 55	umem_odp->umem.is_odp = 1;
 56	mutex_init(&umem_odp->umem_mutex);
 57
 58	if (!umem_odp->is_implicit_odp) {
 59		size_t page_size = 1UL << umem_odp->page_shift;
 60		unsigned long start;
 61		unsigned long end;
 62		size_t ndmas, npfns;
 63
 64		start = ALIGN_DOWN(umem_odp->umem.address, page_size);
 
 65		if (check_add_overflow(umem_odp->umem.address,
 66				       (unsigned long)umem_odp->umem.length,
 67				       &end))
 68			return -EOVERFLOW;
 69		end = ALIGN(end, page_size);
 70		if (unlikely(end < page_size))
 
 71			return -EOVERFLOW;
 72
 73		ndmas = (end - start) >> umem_odp->page_shift;
 74		if (!ndmas)
 
 
 75			return -EINVAL;
 76
 77		npfns = (end - start) >> PAGE_SHIFT;
 78		umem_odp->pfn_list = kvcalloc(
 79			npfns, sizeof(*umem_odp->pfn_list), GFP_KERNEL);
 80		if (!umem_odp->pfn_list)
 
 
 
 
 
 
 81			return -ENOMEM;
 82
 83		umem_odp->dma_list = kvcalloc(
 84			ndmas, sizeof(*umem_odp->dma_list), GFP_KERNEL);
 85		if (!umem_odp->dma_list) {
 86			ret = -ENOMEM;
 87			goto out_pfn_list;
 88		}
 
 
 
 
 
 
 
 
 
 
 
 
 89
 90		ret = mmu_interval_notifier_insert(&umem_odp->notifier,
 91						   umem_odp->umem.owning_mm,
 92						   start, end - start, ops);
 93		if (ret)
 94			goto out_dma_list;
 95	}
 
 96
 97	return 0;
 98
 99out_dma_list:
100	kvfree(umem_odp->dma_list);
101out_pfn_list:
102	kvfree(umem_odp->pfn_list);
103	return ret;
104}
105
106/**
107 * ib_umem_odp_alloc_implicit - Allocate a parent implicit ODP umem
108 *
109 * Implicit ODP umems do not have a VA range and do not have any page lists.
110 * They exist only to hold the per_mm reference to help the driver create
111 * children umems.
112 *
113 * @device: IB device to create UMEM
114 * @access: ib_reg_mr access flags
115 */
116struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
117					       int access)
118{
 
 
 
119	struct ib_umem *umem;
120	struct ib_umem_odp *umem_odp;
121	int ret;
122
123	if (access & IB_ACCESS_HUGETLB)
124		return ERR_PTR(-EINVAL);
125
 
 
 
 
 
126	umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL);
127	if (!umem_odp)
128		return ERR_PTR(-ENOMEM);
129	umem = &umem_odp->umem;
130	umem->ibdev = device;
131	umem->writable = ib_access_writable(access);
132	umem->owning_mm = current->mm;
133	umem_odp->is_implicit_odp = 1;
134	umem_odp->page_shift = PAGE_SHIFT;
135
136	umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
137	ret = ib_init_umem_odp(umem_odp, NULL);
138	if (ret) {
139		put_pid(umem_odp->tgid);
140		kfree(umem_odp);
141		return ERR_PTR(ret);
142	}
143	return umem_odp;
144}
145EXPORT_SYMBOL(ib_umem_odp_alloc_implicit);
146
147/**
148 * ib_umem_odp_alloc_child - Allocate a child ODP umem under an implicit
149 *                           parent ODP umem
150 *
151 * @root: The parent umem enclosing the child. This must be allocated using
152 *        ib_alloc_implicit_odp_umem()
153 * @addr: The starting userspace VA
154 * @size: The length of the userspace VA
155 * @ops: MMU interval ops, currently only @invalidate
156 */
157struct ib_umem_odp *
158ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
159			size_t size,
160			const struct mmu_interval_notifier_ops *ops)
161{
162	/*
163	 * Caller must ensure that root cannot be freed during the call to
164	 * ib_alloc_odp_umem.
165	 */
166	struct ib_umem_odp *odp_data;
167	struct ib_umem *umem;
168	int ret;
169
170	if (WARN_ON(!root->is_implicit_odp))
171		return ERR_PTR(-EINVAL);
172
173	odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
174	if (!odp_data)
175		return ERR_PTR(-ENOMEM);
176	umem = &odp_data->umem;
177	umem->ibdev = root->umem.ibdev;
178	umem->length     = size;
179	umem->address    = addr;
180	umem->writable   = root->umem.writable;
181	umem->owning_mm  = root->umem.owning_mm;
182	odp_data->page_shift = PAGE_SHIFT;
183	odp_data->notifier.ops = ops;
184
185	/*
186	 * A mmget must be held when registering a notifier, the owming_mm only
187	 * has a mm_grab at this point.
188	 */
189	if (!mmget_not_zero(umem->owning_mm)) {
190		ret = -EFAULT;
191		goto out_free;
192	}
193
194	odp_data->tgid = get_pid(root->tgid);
195	ret = ib_init_umem_odp(odp_data, ops);
196	if (ret)
197		goto out_tgid;
198	mmput(umem->owning_mm);
199	return odp_data;
200
201out_tgid:
202	put_pid(odp_data->tgid);
203	mmput(umem->owning_mm);
204out_free:
205	kfree(odp_data);
206	return ERR_PTR(ret);
207}
208EXPORT_SYMBOL(ib_umem_odp_alloc_child);
209
210/**
211 * ib_umem_odp_get - Create a umem_odp for a userspace va
212 *
213 * @device: IB device struct to get UMEM
214 * @addr: userspace virtual address to start at
215 * @size: length of region to pin
216 * @access: IB_ACCESS_xxx flags for memory being pinned
217 * @ops: MMU interval ops, currently only @invalidate
218 *
219 * The driver should use when the access flags indicate ODP memory. It avoids
220 * pinning, instead, stores the mm for future page fault handling in
221 * conjunction with MMU notifiers.
222 */
223struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
224				    unsigned long addr, size_t size, int access,
225				    const struct mmu_interval_notifier_ops *ops)
226{
227	struct ib_umem_odp *umem_odp;
 
 
228	int ret;
229
230	if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)))
 
 
 
 
 
 
 
 
 
231		return ERR_PTR(-EINVAL);
232
233	umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
234	if (!umem_odp)
235		return ERR_PTR(-ENOMEM);
236
237	umem_odp->umem.ibdev = device;
238	umem_odp->umem.length = size;
239	umem_odp->umem.address = addr;
240	umem_odp->umem.writable = ib_access_writable(access);
241	umem_odp->umem.owning_mm = current->mm;
242	umem_odp->notifier.ops = ops;
243
244	umem_odp->page_shift = PAGE_SHIFT;
245#ifdef CONFIG_HUGETLB_PAGE
246	if (access & IB_ACCESS_HUGETLB)
247		umem_odp->page_shift = HPAGE_SHIFT;
248#endif
 
 
 
 
 
 
 
 
 
 
 
249
250	umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
251	ret = ib_init_umem_odp(umem_odp, ops);
252	if (ret)
253		goto err_put_pid;
254	return umem_odp;
255
256err_put_pid:
257	put_pid(umem_odp->tgid);
258	kfree(umem_odp);
259	return ERR_PTR(ret);
260}
261EXPORT_SYMBOL(ib_umem_odp_get);
262
263void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
264{
 
 
265	/*
266	 * Ensure that no more pages are mapped in the umem.
267	 *
268	 * It is the driver's responsibility to ensure, before calling us,
269	 * that the hardware will not attempt to access the MR any more.
270	 */
271	if (!umem_odp->is_implicit_odp) {
272		mutex_lock(&umem_odp->umem_mutex);
273		ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
274					    ib_umem_end(umem_odp));
275		mutex_unlock(&umem_odp->umem_mutex);
276		mmu_interval_notifier_remove(&umem_odp->notifier);
277		kvfree(umem_odp->dma_list);
278		kvfree(umem_odp->pfn_list);
279	}
280	put_pid(umem_odp->tgid);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
281	kfree(umem_odp);
282}
283EXPORT_SYMBOL(ib_umem_odp_release);
284
285/*
286 * Map for DMA and insert a single page into the on-demand paging page tables.
287 *
288 * @umem: the umem to insert the page to.
289 * @dma_index: index in the umem to add the dma to.
290 * @page: the page struct to map and add.
291 * @access_mask: access permissions needed for this page.
 
 
 
292 *
293 * The function returns -EFAULT if the DMA mapping operation fails.
 
294 *
 
 
 
295 */
296static int ib_umem_odp_map_dma_single_page(
297		struct ib_umem_odp *umem_odp,
298		unsigned int dma_index,
299		struct page *page,
300		u64 access_mask)
 
301{
302	struct ib_device *dev = umem_odp->umem.ibdev;
303	dma_addr_t *dma_addr = &umem_odp->dma_list[dma_index];
 
 
304
305	if (*dma_addr) {
306		/*
307		 * If the page is already dma mapped it means it went through
308		 * a non-invalidating trasition, like read-only to writable.
309		 * Resync the flags.
310		 */
311		*dma_addr = (*dma_addr & ODP_DMA_ADDR_MASK) | access_mask;
312		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
313	}
314
315	*dma_addr = ib_dma_map_page(dev, page, 0, 1 << umem_odp->page_shift,
316				    DMA_BIDIRECTIONAL);
317	if (ib_dma_mapping_error(dev, *dma_addr)) {
318		*dma_addr = 0;
319		return -EFAULT;
320	}
321	umem_odp->npages++;
322	*dma_addr |= access_mask;
323	return 0;
324}
325
326/**
327 * ib_umem_odp_map_dma_and_lock - DMA map userspace memory in an ODP MR and lock it.
328 *
329 * Maps the range passed in the argument to DMA addresses.
330 * The DMA addresses of the mapped pages is updated in umem_odp->dma_list.
331 * Upon success the ODP MR will be locked to let caller complete its device
332 * page table update.
333 *
334 * Returns the number of pages mapped in success, negative error code
335 * for failure.
 
 
 
 
336 * @umem_odp: the umem to map and pin
337 * @user_virt: the address from which we need to map.
338 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
339 *        bigger due to alignment, and may also be smaller in case of an error
340 *        pinning or mapping a page. The actual pages mapped is returned in
341 *        the return value.
342 * @access_mask: bit mask of the requested access permissions for the given
343 *               range.
344 * @fault: is faulting required for the given range
 
 
345 */
346int ib_umem_odp_map_dma_and_lock(struct ib_umem_odp *umem_odp, u64 user_virt,
347				 u64 bcnt, u64 access_mask, bool fault)
348			__acquires(&umem_odp->umem_mutex)
349{
350	struct task_struct *owning_process  = NULL;
351	struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
352	int pfn_index, dma_index, ret = 0, start_idx;
353	unsigned int page_shift, hmm_order, pfn_start_idx;
354	unsigned long num_pfns, current_seq;
355	struct hmm_range range = {};
356	unsigned long timeout;
357
358	if (access_mask == 0)
359		return -EINVAL;
360
361	if (user_virt < ib_umem_start(umem_odp) ||
362	    user_virt + bcnt > ib_umem_end(umem_odp))
363		return -EFAULT;
364
 
 
 
 
365	page_shift = umem_odp->page_shift;
 
 
 
 
366
367	/*
368	 * owning_process is allowed to be NULL, this means somehow the mm is
369	 * existing beyond the lifetime of the originating process.. Presumably
370	 * mmget_not_zero will fail in this case.
371	 */
372	owning_process = get_pid_task(umem_odp->tgid, PIDTYPE_PID);
373	if (!owning_process || !mmget_not_zero(owning_mm)) {
374		ret = -EINVAL;
375		goto out_put_task;
376	}
377
378	range.notifier = &umem_odp->notifier;
379	range.start = ALIGN_DOWN(user_virt, 1UL << page_shift);
380	range.end = ALIGN(user_virt + bcnt, 1UL << page_shift);
381	pfn_start_idx = (range.start - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
382	num_pfns = (range.end - range.start) >> PAGE_SHIFT;
383	if (fault) {
384		range.default_flags = HMM_PFN_REQ_FAULT;
385
386		if (access_mask & ODP_WRITE_ALLOWED_BIT)
387			range.default_flags |= HMM_PFN_REQ_WRITE;
388	}
389
390	range.hmm_pfns = &(umem_odp->pfn_list[pfn_start_idx]);
391	timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
392
393retry:
394	current_seq = range.notifier_seq =
395		mmu_interval_read_begin(&umem_odp->notifier);
396
397	mmap_read_lock(owning_mm);
398	ret = hmm_range_fault(&range);
399	mmap_read_unlock(owning_mm);
400	if (unlikely(ret)) {
401		if (ret == -EBUSY && !time_after(jiffies, timeout))
402			goto retry;
403		goto out_put_mm;
404	}
405
406	start_idx = (range.start - ib_umem_start(umem_odp)) >> page_shift;
407	dma_index = start_idx;
408
409	mutex_lock(&umem_odp->umem_mutex);
410	if (mmu_interval_read_retry(&umem_odp->notifier, current_seq)) {
411		mutex_unlock(&umem_odp->umem_mutex);
412		goto retry;
413	}
414
415	for (pfn_index = 0; pfn_index < num_pfns;
416		pfn_index += 1 << (page_shift - PAGE_SHIFT), dma_index++) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
417
418		if (fault) {
419			/*
420			 * Since we asked for hmm_range_fault() to populate
421			 * pages it shouldn't return an error entry on success.
422			 */
423			WARN_ON(range.hmm_pfns[pfn_index] & HMM_PFN_ERROR);
424			WARN_ON(!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID));
425		} else {
426			if (!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID)) {
427				WARN_ON(umem_odp->dma_list[dma_index]);
428				continue;
429			}
430			access_mask = ODP_READ_ALLOWED_BIT;
431			if (range.hmm_pfns[pfn_index] & HMM_PFN_WRITE)
432				access_mask |= ODP_WRITE_ALLOWED_BIT;
433		}
434
435		hmm_order = hmm_pfn_to_map_order(range.hmm_pfns[pfn_index]);
436		/* If a hugepage was detected and ODP wasn't set for, the umem
437		 * page_shift will be used, the opposite case is an error.
438		 */
439		if (hmm_order + PAGE_SHIFT < page_shift) {
440			ret = -EINVAL;
441			ibdev_dbg(umem_odp->umem.ibdev,
442				  "%s: un-expected hmm_order %u, page_shift %u\n",
443				  __func__, hmm_order, page_shift);
444			break;
 
 
 
445		}
 
446
447		ret = ib_umem_odp_map_dma_single_page(
448				umem_odp, dma_index, hmm_pfn_to_page(range.hmm_pfns[pfn_index]),
449				access_mask);
450		if (ret < 0) {
451			ibdev_dbg(umem_odp->umem.ibdev,
452				  "ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
 
 
 
 
 
 
453			break;
454		}
455	}
456	/* upon success lock should stay on hold for the callee */
457	if (!ret)
458		ret = dma_index - start_idx;
459	else
460		mutex_unlock(&umem_odp->umem_mutex);
461
462out_put_mm:
463	mmput_async(owning_mm);
 
 
 
 
 
 
464out_put_task:
465	if (owning_process)
466		put_task_struct(owning_process);
 
467	return ret;
468}
469EXPORT_SYMBOL(ib_umem_odp_map_dma_and_lock);
470
471void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
472				 u64 bound)
473{
474	dma_addr_t dma_addr;
475	dma_addr_t dma;
476	int idx;
477	u64 addr;
478	struct ib_device *dev = umem_odp->umem.ibdev;
479
480	lockdep_assert_held(&umem_odp->umem_mutex);
481
482	virt = max_t(u64, virt, ib_umem_start(umem_odp));
483	bound = min_t(u64, bound, ib_umem_end(umem_odp));
 
 
 
 
 
484	for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
485		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
486		dma = umem_odp->dma_list[idx];
 
 
 
487
488		/* The access flags guaranteed a valid DMA address in case was NULL */
489		if (dma) {
490			unsigned long pfn_idx = (addr - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
491			struct page *page = hmm_pfn_to_page(umem_odp->pfn_list[pfn_idx]);
492
493			dma_addr = dma & ODP_DMA_ADDR_MASK;
494			ib_dma_unmap_page(dev, dma_addr,
495					  BIT(umem_odp->page_shift),
496					  DMA_BIDIRECTIONAL);
497			if (dma & ODP_WRITE_ALLOWED_BIT) {
498				struct page *head_page = compound_head(page);
499				/*
500				 * set_page_dirty prefers being called with
501				 * the page lock. However, MMU notifiers are
502				 * called sometimes with and sometimes without
503				 * the lock. We rely on the umem_mutex instead
504				 * to prevent other mmu notifiers from
505				 * continuing and allowing the page mapping to
506				 * be removed.
507				 */
508				set_page_dirty(head_page);
509			}
 
510			umem_odp->dma_list[idx] = 0;
511			umem_odp->npages--;
512		}
513	}
514}
515EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
v5.4
  1/*
  2 * Copyright (c) 2014 Mellanox Technologies. All rights reserved.
  3 *
  4 * This software is available to you under a choice of one of two
  5 * licenses.  You may choose to be licensed under the terms of the GNU
  6 * General Public License (GPL) Version 2, available from the file
  7 * COPYING in the main directory of this source tree, or the
  8 * OpenIB.org BSD license below:
  9 *
 10 *     Redistribution and use in source and binary forms, with or
 11 *     without modification, are permitted provided that the following
 12 *     conditions are met:
 13 *
 14 *      - Redistributions of source code must retain the above
 15 *        copyright notice, this list of conditions and the following
 16 *        disclaimer.
 17 *
 18 *      - Redistributions in binary form must reproduce the above
 19 *        copyright notice, this list of conditions and the following
 20 *        disclaimer in the documentation and/or other materials
 21 *        provided with the distribution.
 22 *
 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 30 * SOFTWARE.
 31 */
 32
 33#include <linux/types.h>
 34#include <linux/sched.h>
 35#include <linux/sched/mm.h>
 36#include <linux/sched/task.h>
 37#include <linux/pid.h>
 38#include <linux/slab.h>
 39#include <linux/export.h>
 40#include <linux/vmalloc.h>
 41#include <linux/hugetlb.h>
 42#include <linux/interval_tree.h>
 
 43#include <linux/pagemap.h>
 44
 45#include <rdma/ib_verbs.h>
 46#include <rdma/ib_umem.h>
 47#include <rdma/ib_umem_odp.h>
 48
 49#include "uverbs.h"
 50
 51static void ib_umem_notifier_start_account(struct ib_umem_odp *umem_odp)
 
 52{
 53	mutex_lock(&umem_odp->umem_mutex);
 54	if (umem_odp->notifiers_count++ == 0)
 55		/*
 56		 * Initialize the completion object for waiting on
 57		 * notifiers. Since notifier_count is zero, no one should be
 58		 * waiting right now.
 59		 */
 60		reinit_completion(&umem_odp->notifier_completion);
 61	mutex_unlock(&umem_odp->umem_mutex);
 62}
 63
 64static void ib_umem_notifier_end_account(struct ib_umem_odp *umem_odp)
 65{
 66	mutex_lock(&umem_odp->umem_mutex);
 67	/*
 68	 * This sequence increase will notify the QP page fault that the page
 69	 * that is going to be mapped in the spte could have been freed.
 70	 */
 71	++umem_odp->notifiers_seq;
 72	if (--umem_odp->notifiers_count == 0)
 73		complete_all(&umem_odp->notifier_completion);
 74	mutex_unlock(&umem_odp->umem_mutex);
 75}
 76
 77static void ib_umem_notifier_release(struct mmu_notifier *mn,
 78				     struct mm_struct *mm)
 79{
 80	struct ib_ucontext_per_mm *per_mm =
 81		container_of(mn, struct ib_ucontext_per_mm, mn);
 82	struct rb_node *node;
 83
 84	down_read(&per_mm->umem_rwsem);
 85	if (!per_mm->mn.users)
 86		goto out;
 87
 88	for (node = rb_first_cached(&per_mm->umem_tree); node;
 89	     node = rb_next(node)) {
 90		struct ib_umem_odp *umem_odp =
 91			rb_entry(node, struct ib_umem_odp, interval_tree.rb);
 92
 93		/*
 94		 * Increase the number of notifiers running, to prevent any
 95		 * further fault handling on this MR.
 96		 */
 97		ib_umem_notifier_start_account(umem_odp);
 98		complete_all(&umem_odp->notifier_completion);
 99		umem_odp->umem.ibdev->ops.invalidate_range(
100			umem_odp, ib_umem_start(umem_odp),
101			ib_umem_end(umem_odp));
102	}
103
104out:
105	up_read(&per_mm->umem_rwsem);
106}
107
108static int invalidate_range_start_trampoline(struct ib_umem_odp *item,
109					     u64 start, u64 end, void *cookie)
110{
111	ib_umem_notifier_start_account(item);
112	item->umem.ibdev->ops.invalidate_range(item, start, end);
113	return 0;
114}
115
116static int ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn,
117				const struct mmu_notifier_range *range)
118{
119	struct ib_ucontext_per_mm *per_mm =
120		container_of(mn, struct ib_ucontext_per_mm, mn);
121	int rc;
122
123	if (mmu_notifier_range_blockable(range))
124		down_read(&per_mm->umem_rwsem);
125	else if (!down_read_trylock(&per_mm->umem_rwsem))
126		return -EAGAIN;
127
128	if (!per_mm->mn.users) {
129		up_read(&per_mm->umem_rwsem);
130		/*
131		 * At this point users is permanently zero and visible to this
132		 * CPU without a lock, that fact is relied on to skip the unlock
133		 * in range_end.
134		 */
135		return 0;
136	}
137
138	rc = rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
139					   range->end,
140					   invalidate_range_start_trampoline,
141					   mmu_notifier_range_blockable(range),
142					   NULL);
143	if (rc)
144		up_read(&per_mm->umem_rwsem);
145	return rc;
146}
147
148static int invalidate_range_end_trampoline(struct ib_umem_odp *item, u64 start,
149					   u64 end, void *cookie)
150{
151	ib_umem_notifier_end_account(item);
152	return 0;
153}
154
155static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn,
156				const struct mmu_notifier_range *range)
157{
158	struct ib_ucontext_per_mm *per_mm =
159		container_of(mn, struct ib_ucontext_per_mm, mn);
160
161	if (unlikely(!per_mm->mn.users))
162		return;
163
164	rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start,
165				      range->end,
166				      invalidate_range_end_trampoline, true, NULL);
167	up_read(&per_mm->umem_rwsem);
168}
169
170static struct mmu_notifier *ib_umem_alloc_notifier(struct mm_struct *mm)
171{
172	struct ib_ucontext_per_mm *per_mm;
173
174	per_mm = kzalloc(sizeof(*per_mm), GFP_KERNEL);
175	if (!per_mm)
176		return ERR_PTR(-ENOMEM);
177
178	per_mm->umem_tree = RB_ROOT_CACHED;
179	init_rwsem(&per_mm->umem_rwsem);
180
181	WARN_ON(mm != current->mm);
182	rcu_read_lock();
183	per_mm->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
184	rcu_read_unlock();
185	return &per_mm->mn;
186}
187
188static void ib_umem_free_notifier(struct mmu_notifier *mn)
189{
190	struct ib_ucontext_per_mm *per_mm =
191		container_of(mn, struct ib_ucontext_per_mm, mn);
192
193	WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root));
194
195	put_pid(per_mm->tgid);
196	kfree(per_mm);
197}
198
199static const struct mmu_notifier_ops ib_umem_notifiers = {
200	.release                    = ib_umem_notifier_release,
201	.invalidate_range_start     = ib_umem_notifier_invalidate_range_start,
202	.invalidate_range_end       = ib_umem_notifier_invalidate_range_end,
203	.alloc_notifier		    = ib_umem_alloc_notifier,
204	.free_notifier		    = ib_umem_free_notifier,
205};
206
207static inline int ib_init_umem_odp(struct ib_umem_odp *umem_odp)
208{
209	struct ib_ucontext_per_mm *per_mm;
210	struct mmu_notifier *mn;
211	int ret;
212
213	umem_odp->umem.is_odp = 1;
 
 
214	if (!umem_odp->is_implicit_odp) {
215		size_t page_size = 1UL << umem_odp->page_shift;
216		size_t pages;
 
 
217
218		umem_odp->interval_tree.start =
219			ALIGN_DOWN(umem_odp->umem.address, page_size);
220		if (check_add_overflow(umem_odp->umem.address,
221				       (unsigned long)umem_odp->umem.length,
222				       &umem_odp->interval_tree.last))
223			return -EOVERFLOW;
224		umem_odp->interval_tree.last =
225			ALIGN(umem_odp->interval_tree.last, page_size);
226		if (unlikely(umem_odp->interval_tree.last < page_size))
227			return -EOVERFLOW;
228
229		pages = (umem_odp->interval_tree.last -
230			 umem_odp->interval_tree.start) >>
231			umem_odp->page_shift;
232		if (!pages)
233			return -EINVAL;
234
235		/*
236		 * Note that the representation of the intervals in the
237		 * interval tree considers the ending point as contained in
238		 * the interval.
239		 */
240		umem_odp->interval_tree.last--;
241
242		umem_odp->page_list = kvcalloc(
243			pages, sizeof(*umem_odp->page_list), GFP_KERNEL);
244		if (!umem_odp->page_list)
245			return -ENOMEM;
246
247		umem_odp->dma_list = kvcalloc(
248			pages, sizeof(*umem_odp->dma_list), GFP_KERNEL);
249		if (!umem_odp->dma_list) {
250			ret = -ENOMEM;
251			goto out_page_list;
252		}
253	}
254
255	mn = mmu_notifier_get(&ib_umem_notifiers, umem_odp->umem.owning_mm);
256	if (IS_ERR(mn)) {
257		ret = PTR_ERR(mn);
258		goto out_dma_list;
259	}
260	umem_odp->per_mm = per_mm =
261		container_of(mn, struct ib_ucontext_per_mm, mn);
262
263	mutex_init(&umem_odp->umem_mutex);
264	init_completion(&umem_odp->notifier_completion);
265
266	if (!umem_odp->is_implicit_odp) {
267		down_write(&per_mm->umem_rwsem);
268		interval_tree_insert(&umem_odp->interval_tree,
269				     &per_mm->umem_tree);
270		up_write(&per_mm->umem_rwsem);
271	}
272	mmgrab(umem_odp->umem.owning_mm);
273
274	return 0;
275
276out_dma_list:
277	kvfree(umem_odp->dma_list);
278out_page_list:
279	kvfree(umem_odp->page_list);
280	return ret;
281}
282
283/**
284 * ib_umem_odp_alloc_implicit - Allocate a parent implicit ODP umem
285 *
286 * Implicit ODP umems do not have a VA range and do not have any page lists.
287 * They exist only to hold the per_mm reference to help the driver create
288 * children umems.
289 *
290 * @udata: udata from the syscall being used to create the umem
291 * @access: ib_reg_mr access flags
292 */
293struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_udata *udata,
294					       int access)
295{
296	struct ib_ucontext *context =
297		container_of(udata, struct uverbs_attr_bundle, driver_udata)
298			->context;
299	struct ib_umem *umem;
300	struct ib_umem_odp *umem_odp;
301	int ret;
302
303	if (access & IB_ACCESS_HUGETLB)
304		return ERR_PTR(-EINVAL);
305
306	if (!context)
307		return ERR_PTR(-EIO);
308	if (WARN_ON_ONCE(!context->device->ops.invalidate_range))
309		return ERR_PTR(-EINVAL);
310
311	umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL);
312	if (!umem_odp)
313		return ERR_PTR(-ENOMEM);
314	umem = &umem_odp->umem;
315	umem->ibdev = context->device;
316	umem->writable = ib_access_writable(access);
317	umem->owning_mm = current->mm;
318	umem_odp->is_implicit_odp = 1;
319	umem_odp->page_shift = PAGE_SHIFT;
320
321	ret = ib_init_umem_odp(umem_odp);
 
322	if (ret) {
 
323		kfree(umem_odp);
324		return ERR_PTR(ret);
325	}
326	return umem_odp;
327}
328EXPORT_SYMBOL(ib_umem_odp_alloc_implicit);
329
330/**
331 * ib_umem_odp_alloc_child - Allocate a child ODP umem under an implicit
332 *                           parent ODP umem
333 *
334 * @root: The parent umem enclosing the child. This must be allocated using
335 *        ib_alloc_implicit_odp_umem()
336 * @addr: The starting userspace VA
337 * @size: The length of the userspace VA
 
338 */
339struct ib_umem_odp *ib_umem_odp_alloc_child(struct ib_umem_odp *root,
340					    unsigned long addr, size_t size)
 
 
341{
342	/*
343	 * Caller must ensure that root cannot be freed during the call to
344	 * ib_alloc_odp_umem.
345	 */
346	struct ib_umem_odp *odp_data;
347	struct ib_umem *umem;
348	int ret;
349
350	if (WARN_ON(!root->is_implicit_odp))
351		return ERR_PTR(-EINVAL);
352
353	odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
354	if (!odp_data)
355		return ERR_PTR(-ENOMEM);
356	umem = &odp_data->umem;
357	umem->ibdev = root->umem.ibdev;
358	umem->length     = size;
359	umem->address    = addr;
360	umem->writable   = root->umem.writable;
361	umem->owning_mm  = root->umem.owning_mm;
362	odp_data->page_shift = PAGE_SHIFT;
 
363
364	ret = ib_init_umem_odp(odp_data);
365	if (ret) {
366		kfree(odp_data);
367		return ERR_PTR(ret);
 
 
 
368	}
 
 
 
 
 
 
369	return odp_data;
 
 
 
 
 
 
 
370}
371EXPORT_SYMBOL(ib_umem_odp_alloc_child);
372
373/**
374 * ib_umem_odp_get - Create a umem_odp for a userspace va
375 *
376 * @udata: userspace context to pin memory for
377 * @addr: userspace virtual address to start at
378 * @size: length of region to pin
379 * @access: IB_ACCESS_xxx flags for memory being pinned
 
380 *
381 * The driver should use when the access flags indicate ODP memory. It avoids
382 * pinning, instead, stores the mm for future page fault handling in
383 * conjunction with MMU notifiers.
384 */
385struct ib_umem_odp *ib_umem_odp_get(struct ib_udata *udata, unsigned long addr,
386				    size_t size, int access)
 
387{
388	struct ib_umem_odp *umem_odp;
389	struct ib_ucontext *context;
390	struct mm_struct *mm;
391	int ret;
392
393	if (!udata)
394		return ERR_PTR(-EIO);
395
396	context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
397			  ->context;
398	if (!context)
399		return ERR_PTR(-EIO);
400
401	if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)) ||
402	    WARN_ON_ONCE(!context->device->ops.invalidate_range))
403		return ERR_PTR(-EINVAL);
404
405	umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
406	if (!umem_odp)
407		return ERR_PTR(-ENOMEM);
408
409	umem_odp->umem.ibdev = context->device;
410	umem_odp->umem.length = size;
411	umem_odp->umem.address = addr;
412	umem_odp->umem.writable = ib_access_writable(access);
413	umem_odp->umem.owning_mm = mm = current->mm;
 
414
415	umem_odp->page_shift = PAGE_SHIFT;
416	if (access & IB_ACCESS_HUGETLB) {
417		struct vm_area_struct *vma;
418		struct hstate *h;
419
420		down_read(&mm->mmap_sem);
421		vma = find_vma(mm, ib_umem_start(umem_odp));
422		if (!vma || !is_vm_hugetlb_page(vma)) {
423			up_read(&mm->mmap_sem);
424			ret = -EINVAL;
425			goto err_free;
426		}
427		h = hstate_vma(vma);
428		umem_odp->page_shift = huge_page_shift(h);
429		up_read(&mm->mmap_sem);
430	}
431
432	ret = ib_init_umem_odp(umem_odp);
 
433	if (ret)
434		goto err_free;
435	return umem_odp;
436
437err_free:
 
438	kfree(umem_odp);
439	return ERR_PTR(ret);
440}
441EXPORT_SYMBOL(ib_umem_odp_get);
442
443void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
444{
445	struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
446
447	/*
448	 * Ensure that no more pages are mapped in the umem.
449	 *
450	 * It is the driver's responsibility to ensure, before calling us,
451	 * that the hardware will not attempt to access the MR any more.
452	 */
453	if (!umem_odp->is_implicit_odp) {
454		mutex_lock(&umem_odp->umem_mutex);
455		ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
456					    ib_umem_end(umem_odp));
457		mutex_unlock(&umem_odp->umem_mutex);
 
458		kvfree(umem_odp->dma_list);
459		kvfree(umem_odp->page_list);
460	}
461
462	down_write(&per_mm->umem_rwsem);
463	if (!umem_odp->is_implicit_odp) {
464		interval_tree_remove(&umem_odp->interval_tree,
465				     &per_mm->umem_tree);
466		complete_all(&umem_odp->notifier_completion);
467	}
468	/*
469	 * NOTE! mmu_notifier_unregister() can happen between a start/end
470	 * callback, resulting in a missing end, and thus an unbalanced
471	 * lock. This doesn't really matter to us since we are about to kfree
472	 * the memory that holds the lock, however LOCKDEP doesn't like this.
473	 * Thus we call the mmu_notifier_put under the rwsem and test the
474	 * internal users count to reliably see if we are past this point.
475	 */
476	mmu_notifier_put(&per_mm->mn);
477	up_write(&per_mm->umem_rwsem);
478
479	mmdrop(umem_odp->umem.owning_mm);
480	kfree(umem_odp);
481}
482EXPORT_SYMBOL(ib_umem_odp_release);
483
484/*
485 * Map for DMA and insert a single page into the on-demand paging page tables.
486 *
487 * @umem: the umem to insert the page to.
488 * @page_index: index in the umem to add the page to.
489 * @page: the page struct to map and add.
490 * @access_mask: access permissions needed for this page.
491 * @current_seq: sequence number for synchronization with invalidations.
492 *               the sequence number is taken from
493 *               umem_odp->notifiers_seq.
494 *
495 * The function returns -EFAULT if the DMA mapping operation fails. It returns
496 * -EAGAIN if a concurrent invalidation prevents us from updating the page.
497 *
498 * The page is released via put_user_page even if the operation failed. For
499 * on-demand pinning, the page is released whenever it isn't stored in the
500 * umem.
501 */
502static int ib_umem_odp_map_dma_single_page(
503		struct ib_umem_odp *umem_odp,
504		int page_index,
505		struct page *page,
506		u64 access_mask,
507		unsigned long current_seq)
508{
509	struct ib_device *dev = umem_odp->umem.ibdev;
510	dma_addr_t dma_addr;
511	int remove_existing_mapping = 0;
512	int ret = 0;
513
514	/*
515	 * Note: we avoid writing if seq is different from the initial seq, to
516	 * handle case of a racing notifier. This check also allows us to bail
517	 * early if we have a notifier running in parallel with us.
518	 */
519	if (ib_umem_mmu_notifier_retry(umem_odp, current_seq)) {
520		ret = -EAGAIN;
521		goto out;
522	}
523	if (!(umem_odp->dma_list[page_index])) {
524		dma_addr =
525			ib_dma_map_page(dev, page, 0, BIT(umem_odp->page_shift),
526					DMA_BIDIRECTIONAL);
527		if (ib_dma_mapping_error(dev, dma_addr)) {
528			ret = -EFAULT;
529			goto out;
530		}
531		umem_odp->dma_list[page_index] = dma_addr | access_mask;
532		umem_odp->page_list[page_index] = page;
533		umem_odp->npages++;
534	} else if (umem_odp->page_list[page_index] == page) {
535		umem_odp->dma_list[page_index] |= access_mask;
536	} else {
537		pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n",
538		       umem_odp->page_list[page_index], page);
539		/* Better remove the mapping now, to prevent any further
540		 * damage. */
541		remove_existing_mapping = 1;
542	}
543
544out:
545	put_user_page(page);
546
547	if (remove_existing_mapping) {
548		ib_umem_notifier_start_account(umem_odp);
549		dev->ops.invalidate_range(
550			umem_odp,
551			ib_umem_start(umem_odp) +
552				(page_index << umem_odp->page_shift),
553			ib_umem_start(umem_odp) +
554				((page_index + 1) << umem_odp->page_shift));
555		ib_umem_notifier_end_account(umem_odp);
556		ret = -EAGAIN;
557	}
558
559	return ret;
 
 
 
 
 
 
 
 
560}
561
562/**
563 * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR.
564 *
565 * Pins the range of pages passed in the argument, and maps them to
566 * DMA addresses. The DMA addresses of the mapped pages is updated in
567 * umem_odp->dma_list.
 
568 *
569 * Returns the number of pages mapped in success, negative error code
570 * for failure.
571 * An -EAGAIN error code is returned when a concurrent mmu notifier prevents
572 * the function from completing its task.
573 * An -ENOENT error code indicates that userspace process is being terminated
574 * and mm was already destroyed.
575 * @umem_odp: the umem to map and pin
576 * @user_virt: the address from which we need to map.
577 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
578 *        bigger due to alignment, and may also be smaller in case of an error
579 *        pinning or mapping a page. The actual pages mapped is returned in
580 *        the return value.
581 * @access_mask: bit mask of the requested access permissions for the given
582 *               range.
583 * @current_seq: the MMU notifiers sequance value for synchronization with
584 *               invalidations. the sequance number is read from
585 *               umem_odp->notifiers_seq before calling this function
586 */
587int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt,
588			      u64 bcnt, u64 access_mask,
589			      unsigned long current_seq)
590{
591	struct task_struct *owning_process  = NULL;
592	struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
593	struct page       **local_page_list = NULL;
594	u64 page_mask, off;
595	int j, k, ret = 0, start_idx, npages = 0;
596	unsigned int flags = 0, page_shift;
597	phys_addr_t p = 0;
598
599	if (access_mask == 0)
600		return -EINVAL;
601
602	if (user_virt < ib_umem_start(umem_odp) ||
603	    user_virt + bcnt > ib_umem_end(umem_odp))
604		return -EFAULT;
605
606	local_page_list = (struct page **)__get_free_page(GFP_KERNEL);
607	if (!local_page_list)
608		return -ENOMEM;
609
610	page_shift = umem_odp->page_shift;
611	page_mask = ~(BIT(page_shift) - 1);
612	off = user_virt & (~page_mask);
613	user_virt = user_virt & page_mask;
614	bcnt += off; /* Charge for the first page offset as well. */
615
616	/*
617	 * owning_process is allowed to be NULL, this means somehow the mm is
618	 * existing beyond the lifetime of the originating process.. Presumably
619	 * mmget_not_zero will fail in this case.
620	 */
621	owning_process = get_pid_task(umem_odp->per_mm->tgid, PIDTYPE_PID);
622	if (!owning_process || !mmget_not_zero(owning_mm)) {
623		ret = -EINVAL;
624		goto out_put_task;
625	}
626
627	if (access_mask & ODP_WRITE_ALLOWED_BIT)
628		flags |= FOLL_WRITE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
629
630	start_idx = (user_virt - ib_umem_start(umem_odp)) >> page_shift;
631	k = start_idx;
632
633	while (bcnt > 0) {
634		const size_t gup_num_pages = min_t(size_t,
635				(bcnt + BIT(page_shift) - 1) >> page_shift,
636				PAGE_SIZE / sizeof(struct page *));
 
637
638		down_read(&owning_mm->mmap_sem);
639		/*
640		 * Note: this might result in redundent page getting. We can
641		 * avoid this by checking dma_list to be 0 before calling
642		 * get_user_pages. However, this make the code much more
643		 * complex (and doesn't gain us much performance in most use
644		 * cases).
645		 */
646		npages = get_user_pages_remote(owning_process, owning_mm,
647				user_virt, gup_num_pages,
648				flags, local_page_list, NULL, NULL);
649		up_read(&owning_mm->mmap_sem);
650
651		if (npages < 0) {
652			if (npages != -EAGAIN)
653				pr_warn("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
654			else
655				pr_debug("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
656			break;
657		}
658
659		bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
660		mutex_lock(&umem_odp->umem_mutex);
661		for (j = 0; j < npages; j++, user_virt += PAGE_SIZE) {
662			if (user_virt & ~page_mask) {
663				p += PAGE_SIZE;
664				if (page_to_phys(local_page_list[j]) != p) {
665					ret = -EFAULT;
666					break;
667				}
668				put_user_page(local_page_list[j]);
669				continue;
670			}
 
 
 
 
671
672			ret = ib_umem_odp_map_dma_single_page(
673					umem_odp, k, local_page_list[j],
674					access_mask, current_seq);
675			if (ret < 0) {
676				if (ret != -EAGAIN)
677					pr_warn("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
678				else
679					pr_debug("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
680				break;
681			}
682
683			p = page_to_phys(local_page_list[j]);
684			k++;
685		}
686		mutex_unlock(&umem_odp->umem_mutex);
687
 
 
 
688		if (ret < 0) {
689			/*
690			 * Release pages, remembering that the first page
691			 * to hit an error was already released by
692			 * ib_umem_odp_map_dma_single_page().
693			 */
694			if (npages - (j + 1) > 0)
695				put_user_pages(&local_page_list[j+1],
696					       npages - (j + 1));
697			break;
698		}
699	}
 
 
 
 
 
700
701	if (ret >= 0) {
702		if (npages < 0 && k == start_idx)
703			ret = npages;
704		else
705			ret = k - start_idx;
706	}
707
708	mmput(owning_mm);
709out_put_task:
710	if (owning_process)
711		put_task_struct(owning_process);
712	free_page((unsigned long)local_page_list);
713	return ret;
714}
715EXPORT_SYMBOL(ib_umem_odp_map_dma_pages);
716
717void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
718				 u64 bound)
719{
 
 
720	int idx;
721	u64 addr;
722	struct ib_device *dev = umem_odp->umem.ibdev;
723
724	lockdep_assert_held(&umem_odp->umem_mutex);
725
726	virt = max_t(u64, virt, ib_umem_start(umem_odp));
727	bound = min_t(u64, bound, ib_umem_end(umem_odp));
728	/* Note that during the run of this function, the
729	 * notifiers_count of the MR is > 0, preventing any racing
730	 * faults from completion. We might be racing with other
731	 * invalidations, so we must make sure we free each page only
732	 * once. */
733	for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
734		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
735		if (umem_odp->page_list[idx]) {
736			struct page *page = umem_odp->page_list[idx];
737			dma_addr_t dma = umem_odp->dma_list[idx];
738			dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK;
739
740			WARN_ON(!dma_addr);
 
 
 
741
 
742			ib_dma_unmap_page(dev, dma_addr,
743					  BIT(umem_odp->page_shift),
744					  DMA_BIDIRECTIONAL);
745			if (dma & ODP_WRITE_ALLOWED_BIT) {
746				struct page *head_page = compound_head(page);
747				/*
748				 * set_page_dirty prefers being called with
749				 * the page lock. However, MMU notifiers are
750				 * called sometimes with and sometimes without
751				 * the lock. We rely on the umem_mutex instead
752				 * to prevent other mmu notifiers from
753				 * continuing and allowing the page mapping to
754				 * be removed.
755				 */
756				set_page_dirty(head_page);
757			}
758			umem_odp->page_list[idx] = NULL;
759			umem_odp->dma_list[idx] = 0;
760			umem_odp->npages--;
761		}
762	}
763}
764EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
765
766/* @last is not a part of the interval. See comment for function
767 * node_last.
768 */
769int rbt_ib_umem_for_each_in_range(struct rb_root_cached *root,
770				  u64 start, u64 last,
771				  umem_call_back cb,
772				  bool blockable,
773				  void *cookie)
774{
775	int ret_val = 0;
776	struct interval_tree_node *node, *next;
777	struct ib_umem_odp *umem;
778
779	if (unlikely(start == last))
780		return ret_val;
781
782	for (node = interval_tree_iter_first(root, start, last - 1);
783			node; node = next) {
784		/* TODO move the blockable decision up to the callback */
785		if (!blockable)
786			return -EAGAIN;
787		next = interval_tree_iter_next(node, start, last - 1);
788		umem = container_of(node, struct ib_umem_odp, interval_tree);
789		ret_val = cb(umem, start, last, cookie) || ret_val;
790	}
791
792	return ret_val;
793}