umem_odp.c

  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/hmm-dma.h>
 45#include <linux/pagemap.h>
 46
 47#include <rdma/ib_umem_odp.h>
 48
 49#include "uverbs.h"
 50
 51static void ib_init_umem_implicit_odp(struct ib_umem_odp *umem_odp)
 52{
 53	umem_odp->is_implicit_odp = 1;
 54	umem_odp->umem.is_odp = 1;
 55	mutex_init(&umem_odp->umem_mutex);
 56}
 57
 58static int ib_init_umem_odp(struct ib_umem_odp *umem_odp,
 59			    const struct mmu_interval_notifier_ops *ops)
 60{
 61	struct ib_device *dev = umem_odp->umem.ibdev;
 62	size_t page_size = 1UL << umem_odp->page_shift;
 63	struct hmm_dma_map *map;
 64	unsigned long start;
 65	unsigned long end;
 66	size_t nr_entries;
 67	int ret = 0;
 68
 69	umem_odp->umem.is_odp = 1;
 70	mutex_init(&umem_odp->umem_mutex);
 71
 72	start = ALIGN_DOWN(umem_odp->umem.address, page_size);
 73	if (check_add_overflow(umem_odp->umem.address,
 74			       (unsigned long)umem_odp->umem.length, &end))
 75		return -EOVERFLOW;
 76	end = ALIGN(end, page_size);
 77	if (unlikely(end < page_size))
 78		return -EOVERFLOW;
 79	/*
 80	 * The mmu notifier can be called within reclaim contexts and takes the
 81	 * umem_mutex. This is rare to trigger in testing, teach lockdep about
 82	 * it.
 83	 */
 84	if (IS_ENABLED(CONFIG_LOCKDEP)) {
 85		fs_reclaim_acquire(GFP_KERNEL);
 86		mutex_lock(&umem_odp->umem_mutex);
 87		mutex_unlock(&umem_odp->umem_mutex);
 88		fs_reclaim_release(GFP_KERNEL);
 89	}
 90
 91	nr_entries = (end - start) >> PAGE_SHIFT;
 92	if (!(nr_entries * PAGE_SIZE / page_size))
 93		return -EINVAL;
 94
 95	map = &umem_odp->map;
 96	if (ib_uses_virt_dma(dev)) {
 97		map->pfn_list = kvcalloc(nr_entries, sizeof(*map->pfn_list),
 98					 GFP_KERNEL | __GFP_NOWARN);
 99		if (!map->pfn_list)
100			ret = -ENOMEM;
101	} else
102		ret = hmm_dma_map_alloc(dev->dma_device, map,
103					(end - start) >> PAGE_SHIFT,
104					1 << umem_odp->page_shift);
105	if (ret)
106		return ret;
107
108	ret = mmu_interval_notifier_insert(&umem_odp->notifier,
109					   umem_odp->umem.owning_mm, start,
110					   end - start, ops);
111	if (ret)
112		goto out_free_map;
113
114	return 0;
115
116out_free_map:
117	if (ib_uses_virt_dma(dev))
118		kvfree(map->pfn_list);
119	else
120		hmm_dma_map_free(dev->dma_device, map);
121	return ret;
122}
123
124/**
125 * ib_umem_odp_alloc_implicit - Allocate a parent implicit ODP umem
126 *
127 * Implicit ODP umems do not have a VA range and do not have any page lists.
128 * They exist only to hold the per_mm reference to help the driver create
129 * children umems.
130 *
131 * @device: IB device to create UMEM
132 * @access: ib_reg_mr access flags
133 */
134struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
135					       int access)
136{
137	struct ib_umem *umem;
138	struct ib_umem_odp *umem_odp;
139
140	if (access & IB_ACCESS_HUGETLB)
141		return ERR_PTR(-EINVAL);
142
143	umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL);
144	if (!umem_odp)
145		return ERR_PTR(-ENOMEM);
146	umem = &umem_odp->umem;
147	umem->ibdev = device;
148	umem->writable = ib_access_writable(access);
149	umem->owning_mm = current->mm;
150	umem_odp->page_shift = PAGE_SHIFT;
151
152	umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
153	ib_init_umem_implicit_odp(umem_odp);
154	return umem_odp;
155}
156EXPORT_SYMBOL(ib_umem_odp_alloc_implicit);
157
158/**
159 * ib_umem_odp_alloc_child - Allocate a child ODP umem under an implicit
160 *                           parent ODP umem
161 *
162 * @root: The parent umem enclosing the child. This must be allocated using
163 *        ib_alloc_implicit_odp_umem()
164 * @addr: The starting userspace VA
165 * @size: The length of the userspace VA
166 * @ops: MMU interval ops, currently only @invalidate
167 */
168struct ib_umem_odp *
169ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
170			size_t size,
171			const struct mmu_interval_notifier_ops *ops)
172{
173	/*
174	 * Caller must ensure that root cannot be freed during the call to
175	 * ib_alloc_odp_umem.
176	 */
177	struct ib_umem_odp *odp_data;
178	struct ib_umem *umem;
179	int ret;
180
181	if (WARN_ON(!root->is_implicit_odp))
182		return ERR_PTR(-EINVAL);
183
184	odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
185	if (!odp_data)
186		return ERR_PTR(-ENOMEM);
187	umem = &odp_data->umem;
188	umem->ibdev = root->umem.ibdev;
189	umem->length     = size;
190	umem->address    = addr;
191	umem->writable   = root->umem.writable;
192	umem->owning_mm  = root->umem.owning_mm;
193	odp_data->page_shift = PAGE_SHIFT;
194	odp_data->notifier.ops = ops;
195
196	/*
197	 * A mmget must be held when registering a notifier, the owming_mm only
198	 * has a mm_grab at this point.
199	 */
200	if (!mmget_not_zero(umem->owning_mm)) {
201		ret = -EFAULT;
202		goto out_free;
203	}
204
205	odp_data->tgid = get_pid(root->tgid);
206	ret = ib_init_umem_odp(odp_data, ops);
207	if (ret)
208		goto out_tgid;
209	mmput(umem->owning_mm);
210	return odp_data;
211
212out_tgid:
213	put_pid(odp_data->tgid);
214	mmput(umem->owning_mm);
215out_free:
216	kfree(odp_data);
217	return ERR_PTR(ret);
218}
219EXPORT_SYMBOL(ib_umem_odp_alloc_child);
220
221/**
222 * ib_umem_odp_get - Create a umem_odp for a userspace va
223 *
224 * @device: IB device struct to get UMEM
225 * @addr: userspace virtual address to start at
226 * @size: length of region to pin
227 * @access: IB_ACCESS_xxx flags for memory being pinned
228 * @ops: MMU interval ops, currently only @invalidate
229 *
230 * The driver should use when the access flags indicate ODP memory. It avoids
231 * pinning, instead, stores the mm for future page fault handling in
232 * conjunction with MMU notifiers.
233 */
234struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
235				    unsigned long addr, size_t size, int access,
236				    const struct mmu_interval_notifier_ops *ops)
237{
238	struct ib_umem_odp *umem_odp;
239	int ret;
240
241	if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)))
242		return ERR_PTR(-EINVAL);
243
244	umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
245	if (!umem_odp)
246		return ERR_PTR(-ENOMEM);
247
248	umem_odp->umem.ibdev = device;
249	umem_odp->umem.length = size;
250	umem_odp->umem.address = addr;
251	umem_odp->umem.writable = ib_access_writable(access);
252	umem_odp->umem.owning_mm = current->mm;
253	umem_odp->notifier.ops = ops;
254
255	umem_odp->page_shift = PAGE_SHIFT;
256#ifdef CONFIG_HUGETLB_PAGE
257	if (access & IB_ACCESS_HUGETLB)
258		umem_odp->page_shift = HPAGE_SHIFT;
259#endif
260
261	umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
262	ret = ib_init_umem_odp(umem_odp, ops);
263	if (ret)
264		goto err_put_pid;
265	return umem_odp;
266
267err_put_pid:
268	put_pid(umem_odp->tgid);
269	kfree(umem_odp);
270	return ERR_PTR(ret);
271}
272EXPORT_SYMBOL(ib_umem_odp_get);
273
274static void ib_umem_odp_free(struct ib_umem_odp *umem_odp)
275{
276	struct ib_device *dev = umem_odp->umem.ibdev;
277
278	/*
279	 * Ensure that no more pages are mapped in the umem.
280	 *
281	 * It is the driver's responsibility to ensure, before calling us,
282	 * that the hardware will not attempt to access the MR any more.
283	 */
284	mutex_lock(&umem_odp->umem_mutex);
285	ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
286				    ib_umem_end(umem_odp));
287	mutex_unlock(&umem_odp->umem_mutex);
288	mmu_interval_notifier_remove(&umem_odp->notifier);
289	if (ib_uses_virt_dma(dev))
290		kvfree(umem_odp->map.pfn_list);
291	else
292		hmm_dma_map_free(dev->dma_device, &umem_odp->map);
293}
294
295void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
296{
297	if (!umem_odp->is_implicit_odp)
298		ib_umem_odp_free(umem_odp);
299
300	put_pid(umem_odp->tgid);
301	kfree(umem_odp);
302}
303EXPORT_SYMBOL(ib_umem_odp_release);
304
305/**
306 * ib_umem_odp_map_dma_and_lock - DMA map userspace memory in an ODP MR and lock it.
307 *
308 * Maps the range passed in the argument to DMA addresses.
309 * Upon success the ODP MR will be locked to let caller complete its device
310 * page table update.
311 *
312 * Returns the number of pages mapped in success, negative error code
313 * for failure.
314 * @umem_odp: the umem to map and pin
315 * @user_virt: the address from which we need to map.
316 * @bcnt: the minimal number of bytes to pin and map. The mapping might be
317 *        bigger due to alignment, and may also be smaller in case of an error
318 *        pinning or mapping a page. The actual pages mapped is returned in
319 *        the return value.
320 * @access_mask: bit mask of the requested access permissions for the given
321 *               range.
322 * @fault: is faulting required for the given range
323 */
324int ib_umem_odp_map_dma_and_lock(struct ib_umem_odp *umem_odp, u64 user_virt,
325				 u64 bcnt, u64 access_mask, bool fault)
326			__acquires(&umem_odp->umem_mutex)
327{
328	struct task_struct *owning_process  = NULL;
329	struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
330	int pfn_index, dma_index, ret = 0, start_idx;
331	unsigned int page_shift, hmm_order, pfn_start_idx;
332	unsigned long num_pfns, current_seq;
333	struct hmm_range range = {};
334	unsigned long timeout;
335
336	if (user_virt < ib_umem_start(umem_odp) ||
337	    user_virt + bcnt > ib_umem_end(umem_odp))
338		return -EFAULT;
339
340	page_shift = umem_odp->page_shift;
341
342	/*
343	 * owning_process is allowed to be NULL, this means somehow the mm is
344	 * existing beyond the lifetime of the originating process.. Presumably
345	 * mmget_not_zero will fail in this case.
346	 */
347	owning_process = get_pid_task(umem_odp->tgid, PIDTYPE_PID);
348	if (!owning_process || !mmget_not_zero(owning_mm)) {
349		ret = -EINVAL;
350		goto out_put_task;
351	}
352
353	range.notifier = &umem_odp->notifier;
354	range.start = ALIGN_DOWN(user_virt, 1UL << page_shift);
355	range.end = ALIGN(user_virt + bcnt, 1UL << page_shift);
356	pfn_start_idx = (range.start - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
357	num_pfns = (range.end - range.start) >> PAGE_SHIFT;
358	if (fault) {
359		range.default_flags = HMM_PFN_REQ_FAULT;
360
361		if (access_mask & HMM_PFN_WRITE)
362			range.default_flags |= HMM_PFN_REQ_WRITE;
363	}
364
365	range.hmm_pfns = &(umem_odp->map.pfn_list[pfn_start_idx]);
366	timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
367
368retry:
369	current_seq = range.notifier_seq =
370		mmu_interval_read_begin(&umem_odp->notifier);
371
372	mmap_read_lock(owning_mm);
373	ret = hmm_range_fault(&range);
374	mmap_read_unlock(owning_mm);
375	if (unlikely(ret)) {
376		if (ret == -EBUSY && !time_after(jiffies, timeout))
377			goto retry;
378		goto out_put_mm;
379	}
380
381	start_idx = (range.start - ib_umem_start(umem_odp)) >> page_shift;
382	dma_index = start_idx;
383
384	mutex_lock(&umem_odp->umem_mutex);
385	if (mmu_interval_read_retry(&umem_odp->notifier, current_seq)) {
386		mutex_unlock(&umem_odp->umem_mutex);
387		goto retry;
388	}
389
390	for (pfn_index = 0; pfn_index < num_pfns;
391		pfn_index += 1 << (page_shift - PAGE_SHIFT), dma_index++) {
392
393		/*
394		 * Since we asked for hmm_range_fault() to populate
395		 * pages it shouldn't return an error entry on success.
396		 */
397		WARN_ON(fault && range.hmm_pfns[pfn_index] & HMM_PFN_ERROR);
398		WARN_ON(fault && !(range.hmm_pfns[pfn_index] & HMM_PFN_VALID));
399		if (!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID))
400			continue;
401
402		if (range.hmm_pfns[pfn_index] & HMM_PFN_DMA_MAPPED)
403			continue;
404
405		hmm_order = hmm_pfn_to_map_order(range.hmm_pfns[pfn_index]);
406		/* If a hugepage was detected and ODP wasn't set for, the umem
407		 * page_shift will be used, the opposite case is an error.
408		 */
409		if (hmm_order + PAGE_SHIFT < page_shift) {
410			ret = -EINVAL;
411			ibdev_dbg(umem_odp->umem.ibdev,
412				  "%s: un-expected hmm_order %u, page_shift %u\n",
413				  __func__, hmm_order, page_shift);
414			break;
415		}
416	}
417	/* upon success lock should stay on hold for the callee */
418	if (!ret)
419		ret = dma_index - start_idx;
420	else
421		mutex_unlock(&umem_odp->umem_mutex);
422
423out_put_mm:
424	mmput_async(owning_mm);
425out_put_task:
426	if (owning_process)
427		put_task_struct(owning_process);
428	return ret;
429}
430EXPORT_SYMBOL(ib_umem_odp_map_dma_and_lock);
431
432void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
433				 u64 bound)
434{
435	struct ib_device *dev = umem_odp->umem.ibdev;
436	u64 addr;
437
438	lockdep_assert_held(&umem_odp->umem_mutex);
439
440	virt = max_t(u64, virt, ib_umem_start(umem_odp));
441	bound = min_t(u64, bound, ib_umem_end(umem_odp));
442	for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
443		u64 offset = addr - ib_umem_start(umem_odp);
444		size_t idx = offset >> umem_odp->page_shift;
445		unsigned long pfn = umem_odp->map.pfn_list[idx];
446
447		if (!hmm_dma_unmap_pfn(dev->dma_device, &umem_odp->map, idx))
448			goto clear;
449
450		if (pfn & HMM_PFN_WRITE) {
451			struct page *page = hmm_pfn_to_page(pfn);
452			struct page *head_page = compound_head(page);
453			/*
454			 * set_page_dirty prefers being called with
455			 * the page lock. However, MMU notifiers are
456			 * called sometimes with and sometimes without
457			 * the lock. We rely on the umem_mutex instead
458			 * to prevent other mmu notifiers from
459			 * continuing and allowing the page mapping to
460			 * be removed.
461			 */
462			set_page_dirty(head_page);
463		}
464		umem_odp->npages--;
465clear:
466		umem_odp->map.pfn_list[idx] &= ~HMM_PFN_FLAGS;
467	}
468}
469EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);