xref: /openbmc/linux/mm/percpu-vm.c (revision 93274f1d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * mm/percpu-vm.c - vmalloc area based chunk allocation
4  *
5  * Copyright (C) 2010		SUSE Linux Products GmbH
6  * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
7  *
8  * Chunks are mapped into vmalloc areas and populated page by page.
9  * This is the default chunk allocator.
10  */
11 #include "internal.h"
12 
pcpu_chunk_page(struct pcpu_chunk * chunk,unsigned int cpu,int page_idx)13 static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
14 				    unsigned int cpu, int page_idx)
15 {
16 	/* must not be used on pre-mapped chunk */
17 	WARN_ON(chunk->immutable);
18 
19 	return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
20 }
21 
22 /**
23  * pcpu_get_pages - get temp pages array
24  *
25  * Returns pointer to array of pointers to struct page which can be indexed
26  * with pcpu_page_idx().  Note that there is only one array and accesses
27  * should be serialized by pcpu_alloc_mutex.
28  *
29  * RETURNS:
30  * Pointer to temp pages array on success.
31  */
pcpu_get_pages(void)32 static struct page **pcpu_get_pages(void)
33 {
34 	static struct page **pages;
35 	size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
36 
37 	lockdep_assert_held(&pcpu_alloc_mutex);
38 
39 	if (!pages)
40 		pages = pcpu_mem_zalloc(pages_size, GFP_KERNEL);
41 	return pages;
42 }
43 
44 /**
45  * pcpu_free_pages - free pages which were allocated for @chunk
46  * @chunk: chunk pages were allocated for
47  * @pages: array of pages to be freed, indexed by pcpu_page_idx()
48  * @page_start: page index of the first page to be freed
49  * @page_end: page index of the last page to be freed + 1
50  *
51  * Free pages [@page_start and @page_end) in @pages for all units.
52  * The pages were allocated for @chunk.
53  */
pcpu_free_pages(struct pcpu_chunk * chunk,struct page ** pages,int page_start,int page_end)54 static void pcpu_free_pages(struct pcpu_chunk *chunk,
55 			    struct page **pages, int page_start, int page_end)
56 {
57 	unsigned int cpu;
58 	int i;
59 
60 	for_each_possible_cpu(cpu) {
61 		for (i = page_start; i < page_end; i++) {
62 			struct page *page = pages[pcpu_page_idx(cpu, i)];
63 
64 			if (page)
65 				__free_page(page);
66 		}
67 	}
68 }
69 
70 /**
71  * pcpu_alloc_pages - allocates pages for @chunk
72  * @chunk: target chunk
73  * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
74  * @page_start: page index of the first page to be allocated
75  * @page_end: page index of the last page to be allocated + 1
76  * @gfp: allocation flags passed to the underlying allocator
77  *
78  * Allocate pages [@page_start,@page_end) into @pages for all units.
79  * The allocation is for @chunk.  Percpu core doesn't care about the
80  * content of @pages and will pass it verbatim to pcpu_map_pages().
81  */
pcpu_alloc_pages(struct pcpu_chunk * chunk,struct page ** pages,int page_start,int page_end,gfp_t gfp)82 static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
83 			    struct page **pages, int page_start, int page_end,
84 			    gfp_t gfp)
85 {
86 	unsigned int cpu, tcpu;
87 	int i;
88 
89 	gfp |= __GFP_HIGHMEM;
90 
91 	for_each_possible_cpu(cpu) {
92 		for (i = page_start; i < page_end; i++) {
93 			struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
94 
95 			*pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
96 			if (!*pagep)
97 				goto err;
98 		}
99 	}
100 	return 0;
101 
102 err:
103 	while (--i >= page_start)
104 		__free_page(pages[pcpu_page_idx(cpu, i)]);
105 
106 	for_each_possible_cpu(tcpu) {
107 		if (tcpu == cpu)
108 			break;
109 		for (i = page_start; i < page_end; i++)
110 			__free_page(pages[pcpu_page_idx(tcpu, i)]);
111 	}
112 	return -ENOMEM;
113 }
114 
115 /**
116  * pcpu_pre_unmap_flush - flush cache prior to unmapping
117  * @chunk: chunk the regions to be flushed belongs to
118  * @page_start: page index of the first page to be flushed
119  * @page_end: page index of the last page to be flushed + 1
120  *
121  * Pages in [@page_start,@page_end) of @chunk are about to be
122  * unmapped.  Flush cache.  As each flushing trial can be very
123  * expensive, issue flush on the whole region at once rather than
124  * doing it for each cpu.  This could be an overkill but is more
125  * scalable.
126  */
pcpu_pre_unmap_flush(struct pcpu_chunk * chunk,int page_start,int page_end)127 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
128 				 int page_start, int page_end)
129 {
130 	flush_cache_vunmap(
131 		pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
132 		pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
133 }
134 
__pcpu_unmap_pages(unsigned long addr,int nr_pages)135 static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
136 {
137 	vunmap_range_noflush(addr, addr + (nr_pages << PAGE_SHIFT));
138 }
139 
140 /**
141  * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
142  * @chunk: chunk of interest
143  * @pages: pages array which can be used to pass information to free
144  * @page_start: page index of the first page to unmap
145  * @page_end: page index of the last page to unmap + 1
146  *
147  * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
148  * Corresponding elements in @pages were cleared by the caller and can
149  * be used to carry information to pcpu_free_pages() which will be
150  * called after all unmaps are finished.  The caller should call
151  * proper pre/post flush functions.
152  */
pcpu_unmap_pages(struct pcpu_chunk * chunk,struct page ** pages,int page_start,int page_end)153 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
154 			     struct page **pages, int page_start, int page_end)
155 {
156 	unsigned int cpu;
157 	int i;
158 
159 	for_each_possible_cpu(cpu) {
160 		for (i = page_start; i < page_end; i++) {
161 			struct page *page;
162 
163 			page = pcpu_chunk_page(chunk, cpu, i);
164 			WARN_ON(!page);
165 			pages[pcpu_page_idx(cpu, i)] = page;
166 		}
167 		__pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
168 				   page_end - page_start);
169 	}
170 }
171 
172 /**
173  * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
174  * @chunk: pcpu_chunk the regions to be flushed belong to
175  * @page_start: page index of the first page to be flushed
176  * @page_end: page index of the last page to be flushed + 1
177  *
178  * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
179  * TLB for the regions.  This can be skipped if the area is to be
180  * returned to vmalloc as vmalloc will handle TLB flushing lazily.
181  *
182  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
183  * for the whole region.
184  */
pcpu_post_unmap_tlb_flush(struct pcpu_chunk * chunk,int page_start,int page_end)185 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
186 				      int page_start, int page_end)
187 {
188 	flush_tlb_kernel_range(
189 		pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
190 		pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
191 }
192 
__pcpu_map_pages(unsigned long addr,struct page ** pages,int nr_pages)193 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
194 			    int nr_pages)
195 {
196 	return vmap_pages_range_noflush(addr, addr + (nr_pages << PAGE_SHIFT),
197 					PAGE_KERNEL, pages, PAGE_SHIFT);
198 }
199 
200 /**
201  * pcpu_map_pages - map pages into a pcpu_chunk
202  * @chunk: chunk of interest
203  * @pages: pages array containing pages to be mapped
204  * @page_start: page index of the first page to map
205  * @page_end: page index of the last page to map + 1
206  *
207  * For each cpu, map pages [@page_start,@page_end) into @chunk.  The
208  * caller is responsible for calling pcpu_post_map_flush() after all
209  * mappings are complete.
210  *
211  * This function is responsible for setting up whatever is necessary for
212  * reverse lookup (addr -> chunk).
213  */
pcpu_map_pages(struct pcpu_chunk * chunk,struct page ** pages,int page_start,int page_end)214 static int pcpu_map_pages(struct pcpu_chunk *chunk,
215 			  struct page **pages, int page_start, int page_end)
216 {
217 	unsigned int cpu, tcpu;
218 	int i, err;
219 
220 	for_each_possible_cpu(cpu) {
221 		err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
222 				       &pages[pcpu_page_idx(cpu, page_start)],
223 				       page_end - page_start);
224 		if (err < 0)
225 			goto err;
226 
227 		for (i = page_start; i < page_end; i++)
228 			pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
229 					    chunk);
230 	}
231 	return 0;
232 err:
233 	for_each_possible_cpu(tcpu) {
234 		if (tcpu == cpu)
235 			break;
236 		__pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
237 				   page_end - page_start);
238 	}
239 	pcpu_post_unmap_tlb_flush(chunk, page_start, page_end);
240 	return err;
241 }
242 
243 /**
244  * pcpu_post_map_flush - flush cache after mapping
245  * @chunk: pcpu_chunk the regions to be flushed belong to
246  * @page_start: page index of the first page to be flushed
247  * @page_end: page index of the last page to be flushed + 1
248  *
249  * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
250  * cache.
251  *
252  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
253  * for the whole region.
254  */
pcpu_post_map_flush(struct pcpu_chunk * chunk,int page_start,int page_end)255 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
256 				int page_start, int page_end)
257 {
258 	flush_cache_vmap(
259 		pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
260 		pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
261 }
262 
263 /**
264  * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
265  * @chunk: chunk of interest
266  * @page_start: the start page
267  * @page_end: the end page
268  * @gfp: allocation flags passed to the underlying memory allocator
269  *
270  * For each cpu, populate and map pages [@page_start,@page_end) into
271  * @chunk.
272  *
273  * CONTEXT:
274  * pcpu_alloc_mutex, does GFP_KERNEL allocation.
275  */
pcpu_populate_chunk(struct pcpu_chunk * chunk,int page_start,int page_end,gfp_t gfp)276 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
277 			       int page_start, int page_end, gfp_t gfp)
278 {
279 	struct page **pages;
280 
281 	pages = pcpu_get_pages();
282 	if (!pages)
283 		return -ENOMEM;
284 
285 	if (pcpu_alloc_pages(chunk, pages, page_start, page_end, gfp))
286 		return -ENOMEM;
287 
288 	if (pcpu_map_pages(chunk, pages, page_start, page_end)) {
289 		pcpu_free_pages(chunk, pages, page_start, page_end);
290 		return -ENOMEM;
291 	}
292 	pcpu_post_map_flush(chunk, page_start, page_end);
293 
294 	return 0;
295 }
296 
297 /**
298  * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
299  * @chunk: chunk to depopulate
300  * @page_start: the start page
301  * @page_end: the end page
302  *
303  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
304  * from @chunk.
305  *
306  * Caller is required to call pcpu_post_unmap_tlb_flush() if not returning the
307  * region back to vmalloc() which will lazily flush the tlb.
308  *
309  * CONTEXT:
310  * pcpu_alloc_mutex.
311  */
pcpu_depopulate_chunk(struct pcpu_chunk * chunk,int page_start,int page_end)312 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
313 				  int page_start, int page_end)
314 {
315 	struct page **pages;
316 
317 	/*
318 	 * If control reaches here, there must have been at least one
319 	 * successful population attempt so the temp pages array must
320 	 * be available now.
321 	 */
322 	pages = pcpu_get_pages();
323 	BUG_ON(!pages);
324 
325 	/* unmap and free */
326 	pcpu_pre_unmap_flush(chunk, page_start, page_end);
327 
328 	pcpu_unmap_pages(chunk, pages, page_start, page_end);
329 
330 	pcpu_free_pages(chunk, pages, page_start, page_end);
331 }
332 
pcpu_create_chunk(gfp_t gfp)333 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp)
334 {
335 	struct pcpu_chunk *chunk;
336 	struct vm_struct **vms;
337 
338 	chunk = pcpu_alloc_chunk(gfp);
339 	if (!chunk)
340 		return NULL;
341 
342 	vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
343 				pcpu_nr_groups, pcpu_atom_size);
344 	if (!vms) {
345 		pcpu_free_chunk(chunk);
346 		return NULL;
347 	}
348 
349 	chunk->data = vms;
350 	chunk->base_addr = vms[0]->addr - pcpu_group_offsets[0];
351 
352 	pcpu_stats_chunk_alloc();
353 	trace_percpu_create_chunk(chunk->base_addr);
354 
355 	return chunk;
356 }
357 
pcpu_destroy_chunk(struct pcpu_chunk * chunk)358 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
359 {
360 	if (!chunk)
361 		return;
362 
363 	pcpu_stats_chunk_dealloc();
364 	trace_percpu_destroy_chunk(chunk->base_addr);
365 
366 	if (chunk->data)
367 		pcpu_free_vm_areas(chunk->data, pcpu_nr_groups);
368 	pcpu_free_chunk(chunk);
369 }
370 
pcpu_addr_to_page(void * addr)371 static struct page *pcpu_addr_to_page(void *addr)
372 {
373 	return vmalloc_to_page(addr);
374 }
375 
pcpu_verify_alloc_info(const struct pcpu_alloc_info * ai)376 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
377 {
378 	/* no extra restriction */
379 	return 0;
380 }
381 
382 /**
383  * pcpu_should_reclaim_chunk - determine if a chunk should go into reclaim
384  * @chunk: chunk of interest
385  *
386  * This is the entry point for percpu reclaim.  If a chunk qualifies, it is then
387  * isolated and managed in separate lists at the back of pcpu_slot: sidelined
388  * and to_depopulate respectively.  The to_depopulate list holds chunks slated
389  * for depopulation.  They no longer contribute to pcpu_nr_empty_pop_pages once
390  * they are on this list.  Once depopulated, they are moved onto the sidelined
391  * list which enables them to be pulled back in for allocation if no other chunk
392  * can suffice the allocation.
393  */
pcpu_should_reclaim_chunk(struct pcpu_chunk * chunk)394 static bool pcpu_should_reclaim_chunk(struct pcpu_chunk *chunk)
395 {
396 	/* do not reclaim either the first chunk or reserved chunk */
397 	if (chunk == pcpu_first_chunk || chunk == pcpu_reserved_chunk)
398 		return false;
399 
400 	/*
401 	 * If it is isolated, it may be on the sidelined list so move it back to
402 	 * the to_depopulate list.  If we hit at least 1/4 pages empty pages AND
403 	 * there is no system-wide shortage of empty pages aside from this
404 	 * chunk, move it to the to_depopulate list.
405 	 */
406 	return ((chunk->isolated && chunk->nr_empty_pop_pages) ||
407 		(pcpu_nr_empty_pop_pages >
408 		 (PCPU_EMPTY_POP_PAGES_HIGH + chunk->nr_empty_pop_pages) &&
409 		 chunk->nr_empty_pop_pages >= chunk->nr_pages / 4));
410 }
411