summaryrefslogtreecommitdiff
path: root/mm/cma.c (plain)
blob: 8a01df35f42eddfdf7eb9be5858f8d888cc84a92
1/*
2 * Contiguous Memory Allocator
3 *
4 * Copyright (c) 2010-2011 by Samsung Electronics.
5 * Copyright IBM Corporation, 2013
6 * Copyright LG Electronics Inc., 2014
7 * Written by:
8 * Marek Szyprowski <m.szyprowski@samsung.com>
9 * Michal Nazarewicz <mina86@mina86.com>
10 * Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
11 * Joonsoo Kim <iamjoonsoo.kim@lge.com>
12 *
13 * This program is free software; you can redistribute it and/or
14 * modify it under the terms of the GNU General Public License as
15 * published by the Free Software Foundation; either version 2 of the
16 * License or (at your optional) any later version of the license.
17 */
18
19#define pr_fmt(fmt) "cma: " fmt
20
21#ifdef CONFIG_CMA_DEBUG
22#ifndef DEBUG
23# define DEBUG
24#endif
25#endif
26#define CREATE_TRACE_POINTS
27
28#include <linux/memblock.h>
29#include <linux/err.h>
30#include <linux/mm.h>
31#include <linux/mutex.h>
32#include <linux/sizes.h>
33#include <linux/slab.h>
34#include <linux/log2.h>
35#include <linux/cma.h>
36#include <linux/highmem.h>
37#include <linux/io.h>
38#include <trace/events/cma.h>
39#ifdef CONFIG_AMLOGIC_CMA
40#include <asm/pgtable.h>
41#include <linux/amlogic/aml_cma.h>
42#include <linux/delay.h>
43#include <linux/amlogic/secmon.h>
44#endif /* CONFIG_AMLOGIC_CMA */
45
46#include "cma.h"
47
48struct cma cma_areas[MAX_CMA_AREAS];
49unsigned cma_area_count;
50static DEFINE_MUTEX(cma_mutex);
51
52#ifdef CONFIG_AMLOGIC_CMA
53void cma_init_clear(struct cma *cma, bool clear)
54{
55 cma->clear_map = clear;
56}
57
58#ifdef CONFIG_ARM64
59static int clear_cma_pagemap2(struct cma *cma)
60{
61 pgd_t *pgd;
62 pud_t *pud;
63 pmd_t *pmd;
64 unsigned long addr, end;
65 struct mm_struct *mm;
66
67 addr = (unsigned long)pfn_to_kaddr(cma->base_pfn);
68 end = addr + cma->count * PAGE_SIZE;
69 mm = &init_mm;
70 for (; addr < end; addr += SECTION_SIZE) {
71 pgd = pgd_offset(mm, addr);
72 if (pgd_none(*pgd) || pgd_bad(*pgd))
73 break;
74
75 pud = pud_offset(pgd, addr);
76 if (pud_none(*pud) || pud_bad(*pud))
77 break;
78
79 pmd = pmd_offset(pud, addr);
80 if (pmd_none(*pmd))
81 break;
82
83 pr_debug("%s, addr:%lx, pgd:%p %llx, pmd:%p %llx\n",
84 __func__, addr, pgd, pgd_val(*pgd), pmd, pmd_val(*pmd));
85 pmd_clear(pmd);
86 }
87
88 return 0;
89}
90#endif
91
92int setup_cma_full_pagemap(struct cma *cma)
93{
94#ifdef CONFIG_ARM
95 /*
96 * arm already create level 3 mmu mapping for lowmem cma.
97 * And if high mem cma, there is no mapping. So nothing to
98 * do for arch arm.
99 */
100 return 0;
101#elif defined(CONFIG_ARM64)
102 struct vm_area_struct vma = {};
103 unsigned long addr, size;
104 int ret;
105
106 clear_cma_pagemap2(cma);
107 addr = (unsigned long)pfn_to_kaddr(cma->base_pfn);
108 size = cma->count * PAGE_SIZE;
109 vma.vm_mm = &init_mm;
110 vma.vm_start = addr;
111 vma.vm_end = addr + size;
112 vma.vm_page_prot = PAGE_KERNEL;
113 ret = remap_pfn_range(&vma, addr, cma->base_pfn,
114 size, vma.vm_page_prot);
115 if (ret < 0)
116 pr_info("%s, remap pte failed:%d, cma:%lx\n",
117 __func__, ret, cma->base_pfn);
118 return 0;
119#else
120 #error "NOT supported ARCH"
121#endif
122}
123
124int cma_mmu_op(struct page *page, int count, bool set)
125{
126 pgd_t *pgd;
127 pud_t *pud;
128 pmd_t *pmd;
129 pte_t *pte;
130 unsigned long addr, end;
131 struct mm_struct *mm;
132
133 if (!page || PageHighMem(page))
134 return -EINVAL;
135
136 addr = (unsigned long)page_address(page);
137 end = addr + count * PAGE_SIZE;
138 mm = &init_mm;
139 for (; addr < end; addr += PAGE_SIZE) {
140 pgd = pgd_offset(mm, addr);
141 if (pgd_none(*pgd) || pgd_bad(*pgd))
142 break;
143
144 pud = pud_offset(pgd, addr);
145 if (pud_none(*pud) || pud_bad(*pud))
146 break;
147
148 pmd = pmd_offset(pud, addr);
149 if (pmd_none(*pmd))
150 break;
151
152 pte = pte_offset_map(pmd, addr);
153 if (set)
154 set_pte_at(mm, addr, pte, mk_pte(page, PAGE_KERNEL));
155 else
156 pte_clear(mm, addr, pte);
157 pte_unmap(pte);
158 #ifdef CONFIG_ARM
159 pr_debug("%s, add:%lx, pgd:%p %x, pmd:%p %x, pte:%p %x\n",
160 __func__, addr, pgd, (int)pgd_val(*pgd),
161 pmd, (int)pmd_val(*pmd), pte, (int)pte_val(*pte));
162 #elif defined(CONFIG_ARM64)
163 pr_debug("%s, add:%lx, pgd:%p %llx, pmd:%p %llx, pte:%p %llx\n",
164 __func__, addr, pgd, pgd_val(*pgd),
165 pmd, pmd_val(*pmd), pte, pte_val(*pte));
166 #endif
167 page++;
168 }
169 return 0;
170}
171#endif
172
173phys_addr_t cma_get_base(const struct cma *cma)
174{
175 return PFN_PHYS(cma->base_pfn);
176}
177
178unsigned long cma_get_size(const struct cma *cma)
179{
180 return cma->count << PAGE_SHIFT;
181}
182
183static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
184 unsigned int align_order)
185{
186 if (align_order <= cma->order_per_bit)
187 return 0;
188 return (1UL << (align_order - cma->order_per_bit)) - 1;
189}
190
191/*
192 * Find the offset of the base PFN from the specified align_order.
193 * The value returned is represented in order_per_bits.
194 */
195static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
196 unsigned int align_order)
197{
198 return (cma->base_pfn & ((1UL << align_order) - 1))
199 >> cma->order_per_bit;
200}
201
202static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
203 unsigned long pages)
204{
205 return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
206}
207
208static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
209 unsigned int count)
210{
211 unsigned long bitmap_no, bitmap_count;
212
213 bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
214 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
215
216 mutex_lock(&cma->lock);
217 bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
218 mutex_unlock(&cma->lock);
219}
220
221static int __init cma_activate_area(struct cma *cma)
222{
223 int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
224 unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
225 unsigned i = cma->count >> pageblock_order;
226 struct zone *zone;
227
228 cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
229
230 if (!cma->bitmap)
231 return -ENOMEM;
232
233 WARN_ON_ONCE(!pfn_valid(pfn));
234 zone = page_zone(pfn_to_page(pfn));
235
236 do {
237 unsigned j;
238
239 base_pfn = pfn;
240 for (j = pageblock_nr_pages; j; --j, pfn++) {
241 WARN_ON_ONCE(!pfn_valid(pfn));
242 /*
243 * alloc_contig_range requires the pfn range
244 * specified to be in the same zone. Make this
245 * simple by forcing the entire CMA resv range
246 * to be in the same zone.
247 */
248 if (page_zone(pfn_to_page(pfn)) != zone)
249 goto err;
250 }
251 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
252 } while (--i);
253
254 mutex_init(&cma->lock);
255
256#ifdef CONFIG_AMLOGIC_CMA
257 if (cma->clear_map)
258 setup_cma_full_pagemap(cma);
259#endif
260
261#ifdef CONFIG_CMA_DEBUGFS
262 INIT_HLIST_HEAD(&cma->mem_head);
263 spin_lock_init(&cma->mem_head_lock);
264#endif
265
266 return 0;
267
268err:
269 kfree(cma->bitmap);
270 cma->count = 0;
271 return -EINVAL;
272}
273
274static int __init cma_init_reserved_areas(void)
275{
276 int i;
277
278 for (i = 0; i < cma_area_count; i++) {
279 int ret = cma_activate_area(&cma_areas[i]);
280
281 if (ret)
282 return ret;
283 }
284#ifdef CONFIG_AMLOGIC_SEC
285 /*
286 * A73 cache speculate prefetch may cause SError when boot.
287 * because it may prefetch cache line in secure memory range
288 * which have already reserved by bootloader. So we must
289 * clear mmu of secmon range before A73 core boot up
290 */
291 secmon_clear_cma_mmu();
292#endif
293 return 0;
294}
295#ifdef CONFIG_AMLOGIC_CMA
296early_initcall(cma_init_reserved_areas);
297#else
298core_initcall(cma_init_reserved_areas);
299#endif
300
301/**
302 * cma_init_reserved_mem() - create custom contiguous area from reserved memory
303 * @base: Base address of the reserved area
304 * @size: Size of the reserved area (in bytes),
305 * @order_per_bit: Order of pages represented by one bit on bitmap.
306 * @res_cma: Pointer to store the created cma region.
307 *
308 * This function creates custom contiguous area from already reserved memory.
309 */
310int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
311 unsigned int order_per_bit,
312 struct cma **res_cma)
313{
314 struct cma *cma;
315 phys_addr_t alignment;
316
317 /* Sanity checks */
318 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
319 pr_err("Not enough slots for CMA reserved regions!\n");
320 return -ENOSPC;
321 }
322
323 if (!size || !memblock_is_region_reserved(base, size))
324 return -EINVAL;
325
326 /* ensure minimal alignment required by mm core */
327 alignment = PAGE_SIZE <<
328 max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
329
330 /* alignment should be aligned with order_per_bit */
331 if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
332 return -EINVAL;
333
334 if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
335 return -EINVAL;
336
337 /*
338 * Each reserved area must be initialised later, when more kernel
339 * subsystems (like slab allocator) are available.
340 */
341 cma = &cma_areas[cma_area_count];
342 cma->base_pfn = PFN_DOWN(base);
343 cma->count = size >> PAGE_SHIFT;
344 cma->order_per_bit = order_per_bit;
345 *res_cma = cma;
346 cma_area_count++;
347 totalcma_pages += (size / PAGE_SIZE);
348
349 return 0;
350}
351
352/**
353 * cma_declare_contiguous() - reserve custom contiguous area
354 * @base: Base address of the reserved area optional, use 0 for any
355 * @size: Size of the reserved area (in bytes),
356 * @limit: End address of the reserved memory (optional, 0 for any).
357 * @alignment: Alignment for the CMA area, should be power of 2 or zero
358 * @order_per_bit: Order of pages represented by one bit on bitmap.
359 * @fixed: hint about where to place the reserved area
360 * @res_cma: Pointer to store the created cma region.
361 *
362 * This function reserves memory from early allocator. It should be
363 * called by arch specific code once the early allocator (memblock or bootmem)
364 * has been activated and all other subsystems have already allocated/reserved
365 * memory. This function allows to create custom reserved areas.
366 *
367 * If @fixed is true, reserve contiguous area at exactly @base. If false,
368 * reserve in range from @base to @limit.
369 */
370int __init cma_declare_contiguous(phys_addr_t base,
371 phys_addr_t size, phys_addr_t limit,
372 phys_addr_t alignment, unsigned int order_per_bit,
373 bool fixed, struct cma **res_cma)
374{
375 phys_addr_t memblock_end = memblock_end_of_DRAM();
376 phys_addr_t highmem_start;
377 int ret = 0;
378
379#ifdef CONFIG_X86
380 /*
381 * high_memory isn't direct mapped memory so retrieving its physical
382 * address isn't appropriate. But it would be useful to check the
383 * physical address of the highmem boundary so it's justifiable to get
384 * the physical address from it. On x86 there is a validation check for
385 * this case, so the following workaround is needed to avoid it.
386 */
387 highmem_start = __pa_nodebug(high_memory);
388#else
389 highmem_start = __pa(high_memory);
390#endif
391 pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
392 __func__, &size, &base, &limit, &alignment);
393
394 if (cma_area_count == ARRAY_SIZE(cma_areas)) {
395 pr_err("Not enough slots for CMA reserved regions!\n");
396 return -ENOSPC;
397 }
398
399 if (!size)
400 return -EINVAL;
401
402 if (alignment && !is_power_of_2(alignment))
403 return -EINVAL;
404
405 /*
406 * Sanitise input arguments.
407 * Pages both ends in CMA area could be merged into adjacent unmovable
408 * migratetype page by page allocator's buddy algorithm. In the case,
409 * you couldn't get a contiguous memory, which is not what we want.
410 */
411 alignment = max(alignment, (phys_addr_t)PAGE_SIZE <<
412 max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
413 base = ALIGN(base, alignment);
414 size = ALIGN(size, alignment);
415 limit &= ~(alignment - 1);
416
417 if (!base)
418 fixed = false;
419
420 /* size should be aligned with order_per_bit */
421 if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
422 return -EINVAL;
423
424 /*
425 * If allocating at a fixed base the request region must not cross the
426 * low/high memory boundary.
427 */
428 if (fixed && base < highmem_start && base + size > highmem_start) {
429 ret = -EINVAL;
430 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
431 &base, &highmem_start);
432 goto err;
433 }
434
435 /*
436 * If the limit is unspecified or above the memblock end, its effective
437 * value will be the memblock end. Set it explicitly to simplify further
438 * checks.
439 */
440 if (limit == 0 || limit > memblock_end)
441 limit = memblock_end;
442
443 /* Reserve memory */
444 if (fixed) {
445 if (memblock_is_region_reserved(base, size) ||
446 memblock_reserve(base, size) < 0) {
447 ret = -EBUSY;
448 goto err;
449 }
450 } else {
451 phys_addr_t addr = 0;
452
453 /*
454 * All pages in the reserved area must come from the same zone.
455 * If the requested region crosses the low/high memory boundary,
456 * try allocating from high memory first and fall back to low
457 * memory in case of failure.
458 */
459 if (base < highmem_start && limit > highmem_start) {
460 addr = memblock_alloc_range(size, alignment,
461 highmem_start, limit,
462 MEMBLOCK_NONE);
463 limit = highmem_start;
464 }
465
466 if (!addr) {
467 addr = memblock_alloc_range(size, alignment, base,
468 limit,
469 MEMBLOCK_NONE);
470 if (!addr) {
471 ret = -ENOMEM;
472 goto err;
473 }
474 }
475
476 /*
477 * kmemleak scans/reads tracked objects for pointers to other
478 * objects but this address isn't mapped and accessible
479 */
480 kmemleak_ignore_phys(addr);
481 base = addr;
482 }
483
484 ret = cma_init_reserved_mem(base, size, order_per_bit, res_cma);
485 if (ret)
486 goto free_mem;
487
488 pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
489 &base);
490 return 0;
491
492free_mem:
493 memblock_free(base, size);
494err:
495 pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
496 return ret;
497}
498
499/**
500 * cma_alloc() - allocate pages from contiguous area
501 * @cma: Contiguous memory region for which the allocation is performed.
502 * @count: Requested number of pages.
503 * @align: Requested alignment of pages (in PAGE_SIZE order).
504 *
505 * This function allocates part of contiguous memory on specific
506 * contiguous memory area.
507 */
508struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align)
509{
510 unsigned long mask, offset;
511 unsigned long pfn = -1;
512 unsigned long start = 0;
513 unsigned long bitmap_maxno, bitmap_no, bitmap_count;
514 struct page *page = NULL;
515 int ret;
516#ifdef CONFIG_AMLOGIC_CMA
517 int dummy;
518 unsigned long long tick;
519 unsigned long long in_tick, timeout;
520
521 in_tick = sched_clock();
522#endif /* CONFIG_AMLOGIC_CMA */
523
524 if (!cma || !cma->count)
525 return NULL;
526
527 pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
528 count, align);
529
530#ifdef CONFIG_AMLOGIC_CMA
531 tick = sched_clock();
532 cma_debug(0, NULL, "(cma %p, count %zu, align %d)\n",
533 (void *)cma, count, align);
534 in_tick = sched_clock();
535 timeout = 2ULL * 1000000 * (1 + ((count * PAGE_SIZE) >> 20));
536#endif
537 if (!count)
538 return NULL;
539
540 mask = cma_bitmap_aligned_mask(cma, align);
541 offset = cma_bitmap_aligned_offset(cma, align);
542 bitmap_maxno = cma_bitmap_maxno(cma);
543 bitmap_count = cma_bitmap_pages_to_bits(cma, count);
544
545 if (bitmap_count > bitmap_maxno)
546 return NULL;
547
548#ifdef CONFIG_AMLOGIC_CMA
549 aml_cma_alloc_pre_hook(&dummy, count);
550#endif /* CONFIG_AMLOGIC_CMA */
551
552 for (;;) {
553 mutex_lock(&cma->lock);
554 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
555 bitmap_maxno, start, bitmap_count, mask,
556 offset);
557 if (bitmap_no >= bitmap_maxno) {
558 mutex_unlock(&cma->lock);
559 break;
560 }
561 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
562 /*
563 * It's safe to drop the lock here. We've marked this region for
564 * our exclusive use. If the migration fails we will take the
565 * lock again and unmark it.
566 */
567 mutex_unlock(&cma->lock);
568
569 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
570 mutex_lock(&cma_mutex);
571 #ifdef CONFIG_AMLOGIC_CMA
572 ret = aml_cma_alloc_range(pfn, pfn + count);
573 #else
574 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA);
575 #endif /* CONFIG_AMLOGIC_CMA */
576 mutex_unlock(&cma_mutex);
577 if (ret == 0) {
578 page = pfn_to_page(pfn);
579 break;
580 }
581
582 cma_clear_bitmap(cma, pfn, count);
583 if (ret != -EBUSY)
584 break;
585
586 pr_debug("%s(): memory range at %p is busy, retrying\n",
587 __func__, pfn_to_page(pfn));
588 #ifndef CONFIG_AMLOGIC_CMA
589 /* try again with a bit different memory target */
590 start = bitmap_no + mask + 1;
591 #else
592 /*
593 * CMA allocation time out, may blocked on some pages
594 * relax CPU and try later
595 */
596 if ((sched_clock() - in_tick) >= timeout)
597 usleep_range(1000, 2000);
598 #endif /* CONFIG_AMLOGIC_CMA */
599 }
600
601 trace_cma_alloc(pfn, page, count, align);
602
603#ifdef CONFIG_AMLOGIC_CMA
604 aml_cma_alloc_post_hook(&dummy, count, page);
605 cma_debug(0, NULL, "return page:%lx, tick:%lld\n",
606 page ? page_to_pfn(page) : 0, sched_clock() - tick);
607#endif /* CONFIG_AMLOGIC_CMA */
608 pr_debug("%s(): returned %p\n", __func__, page);
609 return page;
610}
611
612/**
613 * cma_release() - release allocated pages
614 * @cma: Contiguous memory region for which the allocation is performed.
615 * @pages: Allocated pages.
616 * @count: Number of allocated pages.
617 *
618 * This function releases memory allocated by alloc_cma().
619 * It returns false when provided pages do not belong to contiguous area and
620 * true otherwise.
621 */
622bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
623{
624 unsigned long pfn;
625
626 if (!cma || !pages)
627 return false;
628
629 pr_debug("%s(page %p)\n", __func__, (void *)pages);
630
631 pfn = page_to_pfn(pages);
632
633 if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
634 return false;
635
636 VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
637
638#ifdef CONFIG_AMLOGIC_CMA
639 aml_cma_release_hook(count, (struct page *)pages);
640 aml_cma_free(pfn, count);
641#else
642 free_contig_range(pfn, count);
643#endif /* CONFIG_AMLOGIC_CMA */
644 cma_clear_bitmap(cma, pfn, count);
645 trace_cma_release(pfn, pages, count);
646
647 return true;
648}
649
650
generated by cgit at 2024-05-18 17:55:04 (GMT)