projects/linux·blob master · refs

529 lines · 14.5 KiB

1// SPDX-License-Identifier: GPL-2.0
2/* Copyright(c) 2015 Intel Corporation. All rights reserved. */
3#include <linux/device.h>
4#include <linux/io.h>
5#include <linux/kasan.h>
6#include <linux/memory_hotplug.h>
7#include <linux/memremap.h>
8#include <linux/swap.h>
9#include <linux/mm.h>
10#include <linux/mmzone.h>
11#include <linux/swapops.h>
12#include <linux/types.h>
13#include <linux/wait_bit.h>
14#include <linux/xarray.h>
15#include "internal.h"
16
17static DEFINE_XARRAY(pgmap_array);
18
19/*
20 * The memremap() and memremap_pages() interfaces are alternately used
21 * to map persistent memory namespaces. These interfaces place different
22 * constraints on the alignment and size of the mapping (namespace).
23 * memremap() can map individual PAGE_SIZE pages. memremap_pages() can
24 * only map subsections (2MB), and at least one architecture (PowerPC)
25 * the minimum mapping granularity of memremap_pages() is 16MB.
26 *
27 * The role of memremap_compat_align() is to communicate the minimum
28 * arch supported alignment of a namespace such that it can freely
29 * switch modes without violating the arch constraint. Namely, do not
30 * allow a namespace to be PAGE_SIZE aligned since that namespace may be
31 * reconfigured into a mode that requires SUBSECTION_SIZE alignment.
32 */
33#ifndef CONFIG_ARCH_HAS_MEMREMAP_COMPAT_ALIGN
34unsigned long memremap_compat_align(void)
35{
36 return SUBSECTION_SIZE;
37}
38EXPORT_SYMBOL_GPL(memremap_compat_align);
39#endif
40
41static void pgmap_array_delete(struct range *range)
42{
43 xa_store_range(&pgmap_array, PHYS_PFN(range->start), PHYS_PFN(range->end),
44 NULL, GFP_KERNEL);
45 synchronize_rcu();
46}
47
48static unsigned long pfn_first(struct dev_pagemap *pgmap, int range_id)
49{
50 struct range *range = &pgmap->ranges[range_id];
51 unsigned long pfn = PHYS_PFN(range->start);
52
53 if (range_id)
54 return pfn;
55 return pfn + vmem_altmap_offset(pgmap_altmap(pgmap));
56}
57
58bool pgmap_pfn_valid(struct dev_pagemap *pgmap, unsigned long pfn)
59{
60 int i;
61
62 for (i = 0; i < pgmap->nr_range; i++) {
63 struct range *range = &pgmap->ranges[i];
64
65 if (pfn >= PHYS_PFN(range->start) &&
66 pfn <= PHYS_PFN(range->end))
67 return pfn >= pfn_first(pgmap, i);
68 }
69
70 return false;
71}
72
73static unsigned long pfn_end(struct dev_pagemap *pgmap, int range_id)
74{
75 const struct range *range = &pgmap->ranges[range_id];
76
77 return (range->start + range_len(range)) >> PAGE_SHIFT;
78}
79
80static unsigned long pfn_len(struct dev_pagemap *pgmap, unsigned long range_id)
81{
82 return (pfn_end(pgmap, range_id) -
83 pfn_first(pgmap, range_id)) >> pgmap->vmemmap_shift;
84}
85
86static void pageunmap_range(struct dev_pagemap *pgmap, int range_id)
87{
88 struct range *range = &pgmap->ranges[range_id];
89 struct page *first_page;
90
91 /* make sure to access a memmap that was actually initialized */
92 first_page = pfn_to_page(pfn_first(pgmap, range_id));
93
94 /* pages are dead and unused, undo the arch mapping */
95 mem_hotplug_begin();
96 remove_pfn_range_from_zone(page_zone(first_page), PHYS_PFN(range->start),
97 PHYS_PFN(range_len(range)));
98 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
99 __remove_pages(PHYS_PFN(range->start),
100 PHYS_PFN(range_len(range)), NULL, pgmap);
101 } else {
102 arch_remove_memory(range->start, range_len(range),
103 pgmap_altmap(pgmap), pgmap);
104 kasan_remove_zero_shadow(__va(range->start), range_len(range));
105 }
106 mem_hotplug_done();
107
108 pfnmap_untrack(PHYS_PFN(range->start), range_len(range));
109 pgmap_array_delete(range);
110}
111
112void memunmap_pages(struct dev_pagemap *pgmap)
113{
114 int i;
115
116 percpu_ref_kill(&pgmap->ref);
117 if (pgmap->type != MEMORY_DEVICE_PRIVATE &&
118 pgmap->type != MEMORY_DEVICE_COHERENT)
119 for (i = 0; i < pgmap->nr_range; i++)
120 percpu_ref_put_many(&pgmap->ref, pfn_len(pgmap, i));
121
122 wait_for_completion(&pgmap->done);
123
124 for (i = 0; i < pgmap->nr_range; i++)
125 pageunmap_range(pgmap, i);
126 percpu_ref_exit(&pgmap->ref);
127
128 WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
129}
130EXPORT_SYMBOL_GPL(memunmap_pages);
131
132static void devm_memremap_pages_release(void *data)
133{
134 memunmap_pages(data);
135}
136
137static void dev_pagemap_percpu_release(struct percpu_ref *ref)
138{
139 struct dev_pagemap *pgmap = container_of(ref, struct dev_pagemap, ref);
140
141 complete(&pgmap->done);
142}
143
144static int pagemap_range(struct dev_pagemap *pgmap, struct mhp_params *params,
145 int range_id, int nid)
146{
147 const bool is_private = pgmap->type == MEMORY_DEVICE_PRIVATE;
148 struct range *range = &pgmap->ranges[range_id];
149 struct dev_pagemap *conflict_pgmap;
150 int error, is_ram;
151
152 if (WARN_ONCE(pgmap_altmap(pgmap) && range_id > 0,
153 "altmap not supported for multiple ranges\n"))
154 return -EINVAL;
155
156 conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->start));
157 if (conflict_pgmap) {
158 WARN(1, "Conflicting mapping in same section\n");
159 put_dev_pagemap(conflict_pgmap);
160 return -ENOMEM;
161 }
162
163 conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->end));
164 if (conflict_pgmap) {
165 WARN(1, "Conflicting mapping in same section\n");
166 put_dev_pagemap(conflict_pgmap);
167 return -ENOMEM;
168 }
169
170 is_ram = region_intersects(range->start, range_len(range),
171 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
172
173 if (is_ram != REGION_DISJOINT) {
174 WARN_ONCE(1, "attempted on %s region %#llx-%#llx\n",
175 is_ram == REGION_MIXED ? "mixed" : "ram",
176 range->start, range->end);
177 return -ENXIO;
178 }
179
180 error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(range->start),
181 PHYS_PFN(range->end), pgmap, GFP_KERNEL));
182 if (error)
183 return error;
184
185 if (nid < 0)
186 nid = numa_mem_id();
187
188 error = pfnmap_track(PHYS_PFN(range->start), range_len(range),
189 &params->pgprot);
190 if (error)
191 goto err_pfn_remap;
192
193 if (!mhp_range_allowed(range->start, range_len(range), !is_private)) {
194 error = -EINVAL;
195 goto err_kasan;
196 }
197
198 mem_hotplug_begin();
199
200 /*
201 * For device private memory we call add_pages() as we only need to
202 * allocate and initialize struct page for the device memory. More-
203 * over the device memory is un-accessible thus we do not want to
204 * create a linear mapping for the memory like arch_add_memory()
205 * would do.
206 *
207 * For all other device memory types, which are accessible by
208 * the CPU, we do want the linear mapping and thus use
209 * arch_add_memory().
210 */
211 if (is_private) {
212 error = add_pages(nid, PHYS_PFN(range->start),
213 PHYS_PFN(range_len(range)), params);
214 } else {
215 error = kasan_add_zero_shadow(__va(range->start), range_len(range));
216 if (error) {
217 mem_hotplug_done();
218 goto err_kasan;
219 }
220
221 error = arch_add_memory(nid, range->start, range_len(range),
222 params);
223 }
224
225 if (!error) {
226 struct zone *zone;
227
228 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
229 move_pfn_range_to_zone(zone, PHYS_PFN(range->start),
230 PHYS_PFN(range_len(range)), params->altmap,
231 MIGRATE_MOVABLE, false);
232 }
233
234 mem_hotplug_done();
235 if (error)
236 goto err_add_memory;
237
238 /*
239 * Initialization of the pages has been deferred until now in order
240 * to allow us to do the work while not holding the hotplug lock.
241 */
242 memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
243 PHYS_PFN(range->start),
244 PHYS_PFN(range_len(range)), pgmap);
245 if (pgmap->type != MEMORY_DEVICE_PRIVATE &&
246 pgmap->type != MEMORY_DEVICE_COHERENT)
247 percpu_ref_get_many(&pgmap->ref, pfn_len(pgmap, range_id));
248 return 0;
249
250err_add_memory:
251 if (!is_private)
252 kasan_remove_zero_shadow(__va(range->start), range_len(range));
253err_kasan:
254 pfnmap_untrack(PHYS_PFN(range->start), range_len(range));
255err_pfn_remap:
256 pgmap_array_delete(range);
257 return error;
258}
259
260
261/*
262 * Not device managed version of devm_memremap_pages, undone by
263 * memunmap_pages(). Please use devm_memremap_pages if you have a struct
264 * device available.
265 */
266void *memremap_pages(struct dev_pagemap *pgmap, int nid)
267{
268 struct mhp_params params = {
269 .altmap = pgmap_altmap(pgmap),
270 .pgmap = pgmap,
271 .pgprot = PAGE_KERNEL,
272 };
273 const int nr_range = pgmap->nr_range;
274 int error, i;
275
276 if (WARN_ONCE(!nr_range, "nr_range must be specified\n"))
277 return ERR_PTR(-EINVAL);
278 if (WARN_ONCE(pgmap->vmemmap_shift > MAX_FOLIO_ORDER,
279 "requested folio size unsupported\n"))
280 return ERR_PTR(-EINVAL);
281
282 switch (pgmap->type) {
283 case MEMORY_DEVICE_PRIVATE:
284 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
285 WARN(1, "Device private memory not supported\n");
286 return ERR_PTR(-EINVAL);
287 }
288 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
289 WARN(1, "Missing migrate_to_ram method\n");
290 return ERR_PTR(-EINVAL);
291 }
292 if (!pgmap->ops->folio_free) {
293 WARN(1, "Missing folio_free method\n");
294 return ERR_PTR(-EINVAL);
295 }
296 if (!pgmap->owner) {
297 WARN(1, "Missing owner\n");
298 return ERR_PTR(-EINVAL);
299 }
300 break;
301 case MEMORY_DEVICE_COHERENT:
302 if (!pgmap->ops->folio_free) {
303 WARN(1, "Missing folio_free method\n");
304 return ERR_PTR(-EINVAL);
305 }
306 if (!pgmap->owner) {
307 WARN(1, "Missing owner\n");
308 return ERR_PTR(-EINVAL);
309 }
310 break;
311 case MEMORY_DEVICE_FS_DAX:
312 params.pgprot = pgprot_decrypted(params.pgprot);
313 break;
314 case MEMORY_DEVICE_GENERIC:
315 break;
316 case MEMORY_DEVICE_PCI_P2PDMA:
317 params.pgprot = pgprot_noncached(params.pgprot);
318 break;
319 default:
320 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
321 break;
322 }
323
324 init_completion(&pgmap->done);
325 error = percpu_ref_init(&pgmap->ref, dev_pagemap_percpu_release, 0,
326 GFP_KERNEL);
327 if (error)
328 return ERR_PTR(error);
329
330 /*
331 * Clear the pgmap nr_range as it will be incremented for each
332 * successfully processed range. This communicates how many
333 * regions to unwind in the abort case.
334 */
335 pgmap->nr_range = 0;
336 error = 0;
337 for (i = 0; i < nr_range; i++) {
338 error = pagemap_range(pgmap, &params, i, nid);
339 if (error)
340 break;
341 pgmap->nr_range++;
342 }
343
344 if (i < nr_range) {
345 memunmap_pages(pgmap);
346 pgmap->nr_range = nr_range;
347 return ERR_PTR(error);
348 }
349
350 return __va(pgmap->ranges[0].start);
351}
352EXPORT_SYMBOL_GPL(memremap_pages);
353
354/**
355 * devm_memremap_pages - remap and provide memmap backing for the given resource
356 * @dev: hosting device for @res
357 * @pgmap: pointer to a struct dev_pagemap
358 *
359 * Notes:
360 * 1/ At a minimum the range and type members of @pgmap must be initialized
361 * by the caller before passing it to this function
362 *
363 * 2/ The altmap field may optionally be initialized, in which case
364 * PGMAP_ALTMAP_VALID must be set in pgmap->flags.
365 *
366 * 3/ The ref field may optionally be provided, in which pgmap->ref must be
367 * 'live' on entry and will be killed and reaped at
368 * devm_memremap_pages_release() time, or if this routine fails.
369 *
370 * 4/ range is expected to be a host memory range that could feasibly be
371 * treated as a "System RAM" range, i.e. not a device mmio range, but
372 * this is not enforced.
373 */
374void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
375{
376 int error;
377 void *ret;
378
379 ret = memremap_pages(pgmap, dev_to_node(dev));
380 if (IS_ERR(ret))
381 return ret;
382
383 error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
384 pgmap);
385 if (error)
386 return ERR_PTR(error);
387 return ret;
388}
389EXPORT_SYMBOL_GPL(devm_memremap_pages);
390
391void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
392{
393 devm_release_action(dev, devm_memremap_pages_release, pgmap);
394}
395EXPORT_SYMBOL_GPL(devm_memunmap_pages);
396
397/**
398 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
399 * @pfn: page frame number to lookup page_map
400 */
401struct dev_pagemap *get_dev_pagemap(unsigned long pfn)
402{
403 struct dev_pagemap *pgmap;
404 resource_size_t phys = PFN_PHYS(pfn);
405
406 rcu_read_lock();
407 pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
408 if (pgmap && !percpu_ref_tryget_live_rcu(&pgmap->ref))
409 pgmap = NULL;
410 rcu_read_unlock();
411
412 return pgmap;
413}
414EXPORT_SYMBOL_GPL(get_dev_pagemap);
415
416void free_zone_device_folio(struct folio *folio)
417{
418 struct dev_pagemap *pgmap = folio->pgmap;
419 unsigned long nr = folio_nr_pages(folio);
420 int i;
421
422 if (WARN_ON_ONCE(!pgmap))
423 return;
424
425 mem_cgroup_uncharge(folio);
426
427 if (folio_test_anon(folio)) {
428 mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON, -1);
429 for (i = 0; i < nr; i++)
430 __ClearPageAnonExclusive(folio_page(folio, i));
431 }
432
433 /*
434 * When a device managed page is freed, the folio->mapping field
435 * may still contain a (stale) mapping value. For example, the
436 * lower bits of folio->mapping may still identify the folio as an
437 * anonymous folio. Ultimately, this entire field is just stale
438 * and wrong, and it will cause errors if not cleared.
439 *
440 * For other types of ZONE_DEVICE pages, migration is either
441 * handled differently or not done at all, so there is no need
442 * to clear folio->mapping.
443 *
444 * FS DAX pages clear the mapping when the folio->share count hits
445 * zero which indicating the page has been removed from the file
446 * system mapping.
447 */
448 if (pgmap->type != MEMORY_DEVICE_FS_DAX &&
449 pgmap->type != MEMORY_DEVICE_GENERIC)
450 folio->mapping = NULL;
451
452 switch (pgmap->type) {
453 case MEMORY_DEVICE_PRIVATE:
454 case MEMORY_DEVICE_COHERENT:
455 if (WARN_ON_ONCE(!pgmap->ops || !pgmap->ops->folio_free))
456 break;
457 pgmap->ops->folio_free(folio);
458 percpu_ref_put_many(&pgmap->ref, nr);
459 break;
460
461 case MEMORY_DEVICE_GENERIC:
462 /*
463 * Reset the refcount to 1 to prepare for handing out the page
464 * again.
465 */
466 folio_set_count(folio, 1);
467 break;
468
469 case MEMORY_DEVICE_FS_DAX:
470 wake_up_var(&folio->page);
471 break;
472
473 case MEMORY_DEVICE_PCI_P2PDMA:
474 if (WARN_ON_ONCE(!pgmap->ops || !pgmap->ops->folio_free))
475 break;
476 pgmap->ops->folio_free(folio);
477 break;
478 }
479}
480
481void zone_device_page_init(struct page *page, struct dev_pagemap *pgmap,
482 unsigned int order)
483{
484 struct page *new_page = page;
485 unsigned int i;
486
487 VM_WARN_ON_ONCE(order > MAX_ORDER_NR_PAGES);
488
489 for (i = 0; i < (1UL << order); ++i, ++new_page) {
490 struct folio *new_folio = (struct folio *)new_page;
491
492 /*
493 * new_page could have been part of previous higher order folio
494 * which encodes the order, in page + 1, in the flags bits. We
495 * blindly clear bits which could have set my order field here,
496 * including page head.
497 */
498 new_page->flags.f &= ~0xffUL; /* Clear possible order, page head */
499
500#ifdef NR_PAGES_IN_LARGE_FOLIO
501 /*
502 * This pointer math looks odd, but new_page could have been
503 * part of a previous higher order folio, which sets _nr_pages
504 * in page + 1 (new_page). Therefore, we use pointer casting to
505 * correctly locate the _nr_pages bits within new_page which
506 * could have modified by previous higher order folio.
507 */
508 ((struct folio *)(new_page - 1))->_nr_pages = 0;
509#endif
510
511 new_folio->mapping = NULL;
512 new_folio->pgmap = pgmap; /* Also clear compound head */
513 new_folio->share = 0; /* fsdax only, unused for device private */
514 VM_WARN_ON_FOLIO(folio_ref_count(new_folio), new_folio);
515 VM_WARN_ON_FOLIO(!folio_is_zone_device(new_folio), new_folio);
516 }
517
518 /*
519 * Drivers shouldn't be allocating pages after calling
520 * memunmap_pages().
521 */
522 WARN_ON_ONCE(!percpu_ref_tryget_many(&page_pgmap(page)->ref, 1 << order));
523 set_page_count(page, 1);
524 lock_page(page);
525
526 if (order)
527 prep_compound_page(page, order);
528}
529EXPORT_SYMBOL_GPL(zone_device_page_init);