xref: /openbmc/linux/mm/cma.c (revision 8a0a6c9af053fb93c0edf4581518c77fd131803e)
1  // SPDX-License-Identifier: GPL-2.0-or-later
2  /*
3   * Contiguous Memory Allocator
4   *
5   * Copyright (c) 2010-2011 by Samsung Electronics.
6   * Copyright IBM Corporation, 2013
7   * Copyright LG Electronics Inc., 2014
8   * Written by:
9   *	Marek Szyprowski <m.szyprowski@samsung.com>
10   *	Michal Nazarewicz <mina86@mina86.com>
11   *	Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
12   *	Joonsoo Kim <iamjoonsoo.kim@lge.com>
13   */
14  
15  #define pr_fmt(fmt) "cma: " fmt
16  
17  #ifdef CONFIG_CMA_DEBUG
18  #ifndef DEBUG
19  #  define DEBUG
20  #endif
21  #endif
22  #define CREATE_TRACE_POINTS
23  
24  #include <linux/memblock.h>
25  #include <linux/err.h>
26  #include <linux/mm.h>
27  #include <linux/sizes.h>
28  #include <linux/slab.h>
29  #include <linux/log2.h>
30  #include <linux/cma.h>
31  #include <linux/highmem.h>
32  #include <linux/io.h>
33  #include <linux/kmemleak.h>
34  #include <trace/events/cma.h>
35  
36  #include "cma.h"
37  
38  struct cma cma_areas[MAX_CMA_AREAS];
39  unsigned cma_area_count;
40  static DEFINE_MUTEX(cma_mutex);
41  
42  phys_addr_t cma_get_base(const struct cma *cma)
43  {
44  	return PFN_PHYS(cma->base_pfn);
45  }
46  
47  unsigned long cma_get_size(const struct cma *cma)
48  {
49  	return cma->count << PAGE_SHIFT;
50  }
51  
52  const char *cma_get_name(const struct cma *cma)
53  {
54  	return cma->name;
55  }
56  
57  static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
58  					     unsigned int align_order)
59  {
60  	if (align_order <= cma->order_per_bit)
61  		return 0;
62  	return (1UL << (align_order - cma->order_per_bit)) - 1;
63  }
64  
65  /*
66   * Find the offset of the base PFN from the specified align_order.
67   * The value returned is represented in order_per_bits.
68   */
69  static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
70  					       unsigned int align_order)
71  {
72  	return (cma->base_pfn & ((1UL << align_order) - 1))
73  		>> cma->order_per_bit;
74  }
75  
76  static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
77  					      unsigned long pages)
78  {
79  	return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
80  }
81  
82  static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
83  			     unsigned long count)
84  {
85  	unsigned long bitmap_no, bitmap_count;
86  	unsigned long flags;
87  
88  	bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
89  	bitmap_count = cma_bitmap_pages_to_bits(cma, count);
90  
91  	spin_lock_irqsave(&cma->lock, flags);
92  	bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
93  	spin_unlock_irqrestore(&cma->lock, flags);
94  }
95  
96  static void __init cma_activate_area(struct cma *cma)
97  {
98  	unsigned long base_pfn = cma->base_pfn, pfn;
99  	struct zone *zone;
100  
101  	cma->bitmap = bitmap_zalloc(cma_bitmap_maxno(cma), GFP_KERNEL);
102  	if (!cma->bitmap)
103  		goto out_error;
104  
105  	/*
106  	 * alloc_contig_range() requires the pfn range specified to be in the
107  	 * same zone. Simplify by forcing the entire CMA resv range to be in the
108  	 * same zone.
109  	 */
110  	WARN_ON_ONCE(!pfn_valid(base_pfn));
111  	zone = page_zone(pfn_to_page(base_pfn));
112  	for (pfn = base_pfn + 1; pfn < base_pfn + cma->count; pfn++) {
113  		WARN_ON_ONCE(!pfn_valid(pfn));
114  		if (page_zone(pfn_to_page(pfn)) != zone)
115  			goto not_in_zone;
116  	}
117  
118  	for (pfn = base_pfn; pfn < base_pfn + cma->count;
119  	     pfn += pageblock_nr_pages)
120  		init_cma_reserved_pageblock(pfn_to_page(pfn));
121  
122  	spin_lock_init(&cma->lock);
123  
124  #ifdef CONFIG_CMA_DEBUGFS
125  	INIT_HLIST_HEAD(&cma->mem_head);
126  	spin_lock_init(&cma->mem_head_lock);
127  #endif
128  
129  	return;
130  
131  not_in_zone:
132  	bitmap_free(cma->bitmap);
133  out_error:
134  	/* Expose all pages to the buddy, they are useless for CMA. */
135  	if (!cma->reserve_pages_on_error) {
136  		for (pfn = base_pfn; pfn < base_pfn + cma->count; pfn++)
137  			free_reserved_page(pfn_to_page(pfn));
138  	}
139  	totalcma_pages -= cma->count;
140  	cma->count = 0;
141  	pr_err("CMA area %s could not be activated\n", cma->name);
142  	return;
143  }
144  
145  static int __init cma_init_reserved_areas(void)
146  {
147  	int i;
148  
149  	for (i = 0; i < cma_area_count; i++)
150  		cma_activate_area(&cma_areas[i]);
151  
152  	return 0;
153  }
154  core_initcall(cma_init_reserved_areas);
155  
156  void __init cma_reserve_pages_on_error(struct cma *cma)
157  {
158  	cma->reserve_pages_on_error = true;
159  }
160  
161  /**
162   * cma_init_reserved_mem() - create custom contiguous area from reserved memory
163   * @base: Base address of the reserved area
164   * @size: Size of the reserved area (in bytes),
165   * @order_per_bit: Order of pages represented by one bit on bitmap.
166   * @name: The name of the area. If this parameter is NULL, the name of
167   *        the area will be set to "cmaN", where N is a running counter of
168   *        used areas.
169   * @res_cma: Pointer to store the created cma region.
170   *
171   * This function creates custom contiguous area from already reserved memory.
172   */
173  int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
174  				 unsigned int order_per_bit,
175  				 const char *name,
176  				 struct cma **res_cma)
177  {
178  	struct cma *cma;
179  
180  	/* Sanity checks */
181  	if (cma_area_count == ARRAY_SIZE(cma_areas)) {
182  		pr_err("Not enough slots for CMA reserved regions!\n");
183  		return -ENOSPC;
184  	}
185  
186  	if (!size || !memblock_is_region_reserved(base, size))
187  		return -EINVAL;
188  
189  	/* alignment should be aligned with order_per_bit */
190  	if (!IS_ALIGNED(CMA_MIN_ALIGNMENT_PAGES, 1 << order_per_bit))
191  		return -EINVAL;
192  
193  	/* ensure minimal alignment required by mm core */
194  	if (!IS_ALIGNED(base | size, CMA_MIN_ALIGNMENT_BYTES))
195  		return -EINVAL;
196  
197  	/*
198  	 * Each reserved area must be initialised later, when more kernel
199  	 * subsystems (like slab allocator) are available.
200  	 */
201  	cma = &cma_areas[cma_area_count];
202  
203  	if (name)
204  		snprintf(cma->name, CMA_MAX_NAME, name);
205  	else
206  		snprintf(cma->name, CMA_MAX_NAME,  "cma%d\n", cma_area_count);
207  
208  	cma->base_pfn = PFN_DOWN(base);
209  	cma->count = size >> PAGE_SHIFT;
210  	cma->order_per_bit = order_per_bit;
211  	*res_cma = cma;
212  	cma_area_count++;
213  	totalcma_pages += (size / PAGE_SIZE);
214  
215  	return 0;
216  }
217  
218  /**
219   * cma_declare_contiguous_nid() - reserve custom contiguous area
220   * @base: Base address of the reserved area optional, use 0 for any
221   * @size: Size of the reserved area (in bytes),
222   * @limit: End address of the reserved memory (optional, 0 for any).
223   * @alignment: Alignment for the CMA area, should be power of 2 or zero
224   * @order_per_bit: Order of pages represented by one bit on bitmap.
225   * @fixed: hint about where to place the reserved area
226   * @name: The name of the area. See function cma_init_reserved_mem()
227   * @res_cma: Pointer to store the created cma region.
228   * @nid: nid of the free area to find, %NUMA_NO_NODE for any node
229   *
230   * This function reserves memory from early allocator. It should be
231   * called by arch specific code once the early allocator (memblock or bootmem)
232   * has been activated and all other subsystems have already allocated/reserved
233   * memory. This function allows to create custom reserved areas.
234   *
235   * If @fixed is true, reserve contiguous area at exactly @base.  If false,
236   * reserve in range from @base to @limit.
237   */
238  int __init cma_declare_contiguous_nid(phys_addr_t base,
239  			phys_addr_t size, phys_addr_t limit,
240  			phys_addr_t alignment, unsigned int order_per_bit,
241  			bool fixed, const char *name, struct cma **res_cma,
242  			int nid)
243  {
244  	phys_addr_t memblock_end = memblock_end_of_DRAM();
245  	phys_addr_t highmem_start;
246  	int ret = 0;
247  
248  	/*
249  	 * We can't use __pa(high_memory) directly, since high_memory
250  	 * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
251  	 * complain. Find the boundary by adding one to the last valid
252  	 * address.
253  	 */
254  	highmem_start = __pa(high_memory - 1) + 1;
255  	pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
256  		__func__, &size, &base, &limit, &alignment);
257  
258  	if (cma_area_count == ARRAY_SIZE(cma_areas)) {
259  		pr_err("Not enough slots for CMA reserved regions!\n");
260  		return -ENOSPC;
261  	}
262  
263  	if (!size)
264  		return -EINVAL;
265  
266  	if (alignment && !is_power_of_2(alignment))
267  		return -EINVAL;
268  
269  	/* Sanitise input arguments. */
270  	alignment = max_t(phys_addr_t, alignment, CMA_MIN_ALIGNMENT_BYTES);
271  	if (fixed && base & (alignment - 1)) {
272  		ret = -EINVAL;
273  		pr_err("Region at %pa must be aligned to %pa bytes\n",
274  			&base, &alignment);
275  		goto err;
276  	}
277  	base = ALIGN(base, alignment);
278  	size = ALIGN(size, alignment);
279  	limit &= ~(alignment - 1);
280  
281  	if (!base)
282  		fixed = false;
283  
284  	/* size should be aligned with order_per_bit */
285  	if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
286  		return -EINVAL;
287  
288  	/*
289  	 * If allocating at a fixed base the request region must not cross the
290  	 * low/high memory boundary.
291  	 */
292  	if (fixed && base < highmem_start && base + size > highmem_start) {
293  		ret = -EINVAL;
294  		pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
295  			&base, &highmem_start);
296  		goto err;
297  	}
298  
299  	/*
300  	 * If the limit is unspecified or above the memblock end, its effective
301  	 * value will be the memblock end. Set it explicitly to simplify further
302  	 * checks.
303  	 */
304  	if (limit == 0 || limit > memblock_end)
305  		limit = memblock_end;
306  
307  	if (base + size > limit) {
308  		ret = -EINVAL;
309  		pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
310  			&size, &base, &limit);
311  		goto err;
312  	}
313  
314  	/* Reserve memory */
315  	if (fixed) {
316  		if (memblock_is_region_reserved(base, size) ||
317  		    memblock_reserve(base, size) < 0) {
318  			ret = -EBUSY;
319  			goto err;
320  		}
321  	} else {
322  		phys_addr_t addr = 0;
323  
324  		/*
325  		 * If there is enough memory, try a bottom-up allocation first.
326  		 * It will place the new cma area close to the start of the node
327  		 * and guarantee that the compaction is moving pages out of the
328  		 * cma area and not into it.
329  		 * Avoid using first 4GB to not interfere with constrained zones
330  		 * like DMA/DMA32.
331  		 */
332  #ifdef CONFIG_PHYS_ADDR_T_64BIT
333  		if (!memblock_bottom_up() && memblock_end >= SZ_4G + size) {
334  			memblock_set_bottom_up(true);
335  			addr = memblock_alloc_range_nid(size, alignment, SZ_4G,
336  							limit, nid, true);
337  			memblock_set_bottom_up(false);
338  		}
339  #endif
340  
341  		/*
342  		 * All pages in the reserved area must come from the same zone.
343  		 * If the requested region crosses the low/high memory boundary,
344  		 * try allocating from high memory first and fall back to low
345  		 * memory in case of failure.
346  		 */
347  		if (!addr && base < highmem_start && limit > highmem_start) {
348  			addr = memblock_alloc_range_nid(size, alignment,
349  					highmem_start, limit, nid, true);
350  			limit = highmem_start;
351  		}
352  
353  		if (!addr) {
354  			addr = memblock_alloc_range_nid(size, alignment, base,
355  					limit, nid, true);
356  			if (!addr) {
357  				ret = -ENOMEM;
358  				goto err;
359  			}
360  		}
361  
362  		/*
363  		 * kmemleak scans/reads tracked objects for pointers to other
364  		 * objects but this address isn't mapped and accessible
365  		 */
366  		kmemleak_ignore_phys(addr);
367  		base = addr;
368  	}
369  
370  	ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
371  	if (ret)
372  		goto free_mem;
373  
374  	pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
375  		&base);
376  	return 0;
377  
378  free_mem:
379  	memblock_phys_free(base, size);
380  err:
381  	pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
382  	return ret;
383  }
384  
385  #ifdef CONFIG_CMA_DEBUG
386  static void cma_debug_show_areas(struct cma *cma)
387  {
388  	unsigned long next_zero_bit, next_set_bit, nr_zero;
389  	unsigned long start = 0;
390  	unsigned long nr_part, nr_total = 0;
391  	unsigned long nbits = cma_bitmap_maxno(cma);
392  
393  	spin_lock_irq(&cma->lock);
394  	pr_info("number of available pages: ");
395  	for (;;) {
396  		next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
397  		if (next_zero_bit >= nbits)
398  			break;
399  		next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
400  		nr_zero = next_set_bit - next_zero_bit;
401  		nr_part = nr_zero << cma->order_per_bit;
402  		pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
403  			next_zero_bit);
404  		nr_total += nr_part;
405  		start = next_zero_bit + nr_zero;
406  	}
407  	pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
408  	spin_unlock_irq(&cma->lock);
409  }
410  #else
411  static inline void cma_debug_show_areas(struct cma *cma) { }
412  #endif
413  
414  /**
415   * cma_alloc() - allocate pages from contiguous area
416   * @cma:   Contiguous memory region for which the allocation is performed.
417   * @count: Requested number of pages.
418   * @align: Requested alignment of pages (in PAGE_SIZE order).
419   * @no_warn: Avoid printing message about failed allocation
420   *
421   * This function allocates part of contiguous memory on specific
422   * contiguous memory area.
423   */
424  struct page *cma_alloc(struct cma *cma, unsigned long count,
425  		       unsigned int align, bool no_warn)
426  {
427  	unsigned long mask, offset;
428  	unsigned long pfn = -1;
429  	unsigned long start = 0;
430  	unsigned long bitmap_maxno, bitmap_no, bitmap_count;
431  	unsigned long i;
432  	struct page *page = NULL;
433  	int ret = -ENOMEM;
434  
435  	if (!cma || !cma->count || !cma->bitmap)
436  		goto out;
437  
438  	pr_debug("%s(cma %p, count %lu, align %d)\n", __func__, (void *)cma,
439  		 count, align);
440  
441  	if (!count)
442  		goto out;
443  
444  	trace_cma_alloc_start(cma->name, count, align);
445  
446  	mask = cma_bitmap_aligned_mask(cma, align);
447  	offset = cma_bitmap_aligned_offset(cma, align);
448  	bitmap_maxno = cma_bitmap_maxno(cma);
449  	bitmap_count = cma_bitmap_pages_to_bits(cma, count);
450  
451  	if (bitmap_count > bitmap_maxno)
452  		goto out;
453  
454  	for (;;) {
455  		spin_lock_irq(&cma->lock);
456  		bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
457  				bitmap_maxno, start, bitmap_count, mask,
458  				offset);
459  		if (bitmap_no >= bitmap_maxno) {
460  			spin_unlock_irq(&cma->lock);
461  			break;
462  		}
463  		bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
464  		/*
465  		 * It's safe to drop the lock here. We've marked this region for
466  		 * our exclusive use. If the migration fails we will take the
467  		 * lock again and unmark it.
468  		 */
469  		spin_unlock_irq(&cma->lock);
470  
471  		pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
472  		mutex_lock(&cma_mutex);
473  		ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
474  				     GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
475  		mutex_unlock(&cma_mutex);
476  		if (ret == 0) {
477  			page = pfn_to_page(pfn);
478  			break;
479  		}
480  
481  		cma_clear_bitmap(cma, pfn, count);
482  		if (ret != -EBUSY)
483  			break;
484  
485  		pr_debug("%s(): memory range at %p is busy, retrying\n",
486  			 __func__, pfn_to_page(pfn));
487  
488  		trace_cma_alloc_busy_retry(cma->name, pfn, pfn_to_page(pfn),
489  					   count, align);
490  		/* try again with a bit different memory target */
491  		start = bitmap_no + mask + 1;
492  	}
493  
494  	trace_cma_alloc_finish(cma->name, pfn, page, count, align, ret);
495  
496  	/*
497  	 * CMA can allocate multiple page blocks, which results in different
498  	 * blocks being marked with different tags. Reset the tags to ignore
499  	 * those page blocks.
500  	 */
501  	if (page) {
502  		for (i = 0; i < count; i++)
503  			page_kasan_tag_reset(page + i);
504  	}
505  
506  	if (ret && !no_warn) {
507  		pr_err_ratelimited("%s: %s: alloc failed, req-size: %lu pages, ret: %d\n",
508  				   __func__, cma->name, count, ret);
509  		cma_debug_show_areas(cma);
510  	}
511  
512  	pr_debug("%s(): returned %p\n", __func__, page);
513  out:
514  	if (page) {
515  		count_vm_event(CMA_ALLOC_SUCCESS);
516  		cma_sysfs_account_success_pages(cma, count);
517  	} else {
518  		count_vm_event(CMA_ALLOC_FAIL);
519  		if (cma)
520  			cma_sysfs_account_fail_pages(cma, count);
521  	}
522  
523  	return page;
524  }
525  
526  bool cma_pages_valid(struct cma *cma, const struct page *pages,
527  		     unsigned long count)
528  {
529  	unsigned long pfn;
530  
531  	if (!cma || !pages)
532  		return false;
533  
534  	pfn = page_to_pfn(pages);
535  
536  	if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count) {
537  		pr_debug("%s(page %p, count %lu)\n", __func__,
538  						(void *)pages, count);
539  		return false;
540  	}
541  
542  	return true;
543  }
544  
545  /**
546   * cma_release() - release allocated pages
547   * @cma:   Contiguous memory region for which the allocation is performed.
548   * @pages: Allocated pages.
549   * @count: Number of allocated pages.
550   *
551   * This function releases memory allocated by cma_alloc().
552   * It returns false when provided pages do not belong to contiguous area and
553   * true otherwise.
554   */
555  bool cma_release(struct cma *cma, const struct page *pages,
556  		 unsigned long count)
557  {
558  	unsigned long pfn;
559  
560  	if (!cma_pages_valid(cma, pages, count))
561  		return false;
562  
563  	pr_debug("%s(page %p, count %lu)\n", __func__, (void *)pages, count);
564  
565  	pfn = page_to_pfn(pages);
566  
567  	VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
568  
569  	free_contig_range(pfn, count);
570  	cma_clear_bitmap(cma, pfn, count);
571  	trace_cma_release(cma->name, pfn, pages, count);
572  
573  	return true;
574  }
575  
576  int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
577  {
578  	int i;
579  
580  	for (i = 0; i < cma_area_count; i++) {
581  		int ret = it(&cma_areas[i], data);
582  
583  		if (ret)
584  			return ret;
585  	}
586  
587  	return 0;
588  }
589