1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * Copyright (C) 2012 ARM Ltd. |
4 | * Copyright (C) 2020 Google LLC |
5 | */ |
6 | #include <linux/cma.h> |
7 | #include <linux/debugfs.h> |
8 | #include <linux/dma-map-ops.h> |
9 | #include <linux/dma-direct.h> |
10 | #include <linux/init.h> |
11 | #include <linux/genalloc.h> |
12 | #include <linux/set_memory.h> |
13 | #include <linux/slab.h> |
14 | #include <linux/workqueue.h> |
15 | |
16 | static struct gen_pool *atomic_pool_dma __ro_after_init; |
17 | static unsigned long pool_size_dma; |
18 | static struct gen_pool *atomic_pool_dma32 __ro_after_init; |
19 | static unsigned long pool_size_dma32; |
20 | static struct gen_pool *atomic_pool_kernel __ro_after_init; |
21 | static unsigned long pool_size_kernel; |
22 | |
23 | /* Size can be defined by the coherent_pool command line */ |
24 | static size_t atomic_pool_size; |
25 | |
26 | /* Dynamic background expansion when the atomic pool is near capacity */ |
27 | static struct work_struct atomic_pool_work; |
28 | |
29 | static int __init early_coherent_pool(char *p) |
30 | { |
31 | atomic_pool_size = memparse(ptr: p, retptr: &p); |
32 | return 0; |
33 | } |
34 | early_param("coherent_pool" , early_coherent_pool); |
35 | |
36 | static void __init dma_atomic_pool_debugfs_init(void) |
37 | { |
38 | struct dentry *root; |
39 | |
40 | root = debugfs_create_dir(name: "dma_pools" , NULL); |
41 | debugfs_create_ulong(name: "pool_size_dma" , mode: 0400, parent: root, value: &pool_size_dma); |
42 | debugfs_create_ulong(name: "pool_size_dma32" , mode: 0400, parent: root, value: &pool_size_dma32); |
43 | debugfs_create_ulong(name: "pool_size_kernel" , mode: 0400, parent: root, value: &pool_size_kernel); |
44 | } |
45 | |
46 | static void dma_atomic_pool_size_add(gfp_t gfp, size_t size) |
47 | { |
48 | if (gfp & __GFP_DMA) |
49 | pool_size_dma += size; |
50 | else if (gfp & __GFP_DMA32) |
51 | pool_size_dma32 += size; |
52 | else |
53 | pool_size_kernel += size; |
54 | } |
55 | |
56 | static bool cma_in_zone(gfp_t gfp) |
57 | { |
58 | unsigned long size; |
59 | phys_addr_t end; |
60 | struct cma *cma; |
61 | |
62 | cma = dev_get_cma_area(NULL); |
63 | if (!cma) |
64 | return false; |
65 | |
66 | size = cma_get_size(cma); |
67 | if (!size) |
68 | return false; |
69 | |
70 | /* CMA can't cross zone boundaries, see cma_activate_area() */ |
71 | end = cma_get_base(cma) + size - 1; |
72 | if (IS_ENABLED(CONFIG_ZONE_DMA) && (gfp & GFP_DMA)) |
73 | return end <= DMA_BIT_MASK(zone_dma_bits); |
74 | if (IS_ENABLED(CONFIG_ZONE_DMA32) && (gfp & GFP_DMA32)) |
75 | return end <= DMA_BIT_MASK(32); |
76 | return true; |
77 | } |
78 | |
79 | static int atomic_pool_expand(struct gen_pool *pool, size_t pool_size, |
80 | gfp_t gfp) |
81 | { |
82 | unsigned int order; |
83 | struct page *page = NULL; |
84 | void *addr; |
85 | int ret = -ENOMEM; |
86 | |
87 | /* Cannot allocate larger than MAX_PAGE_ORDER */ |
88 | order = min(get_order(pool_size), MAX_PAGE_ORDER); |
89 | |
90 | do { |
91 | pool_size = 1 << (PAGE_SHIFT + order); |
92 | if (cma_in_zone(gfp)) |
93 | page = dma_alloc_from_contiguous(NULL, count: 1 << order, |
94 | order, no_warn: false); |
95 | if (!page) |
96 | page = alloc_pages(gfp, order); |
97 | } while (!page && order-- > 0); |
98 | if (!page) |
99 | goto out; |
100 | |
101 | arch_dma_prep_coherent(page, size: pool_size); |
102 | |
103 | #ifdef CONFIG_DMA_DIRECT_REMAP |
104 | addr = dma_common_contiguous_remap(page, pool_size, |
105 | pgprot_dmacoherent(PAGE_KERNEL), |
106 | __builtin_return_address(0)); |
107 | if (!addr) |
108 | goto free_page; |
109 | #else |
110 | addr = page_to_virt(page); |
111 | #endif |
112 | /* |
113 | * Memory in the atomic DMA pools must be unencrypted, the pools do not |
114 | * shrink so no re-encryption occurs in dma_direct_free(). |
115 | */ |
116 | ret = set_memory_decrypted(addr: (unsigned long)page_to_virt(page), |
117 | numpages: 1 << order); |
118 | if (ret) |
119 | goto remove_mapping; |
120 | ret = gen_pool_add_virt(pool, addr: (unsigned long)addr, page_to_phys(page), |
121 | size: pool_size, NUMA_NO_NODE); |
122 | if (ret) |
123 | goto encrypt_mapping; |
124 | |
125 | dma_atomic_pool_size_add(gfp, size: pool_size); |
126 | return 0; |
127 | |
128 | encrypt_mapping: |
129 | ret = set_memory_encrypted(addr: (unsigned long)page_to_virt(page), |
130 | numpages: 1 << order); |
131 | if (WARN_ON_ONCE(ret)) { |
132 | /* Decrypt succeeded but encrypt failed, purposely leak */ |
133 | goto out; |
134 | } |
135 | remove_mapping: |
136 | #ifdef CONFIG_DMA_DIRECT_REMAP |
137 | dma_common_free_remap(addr, pool_size); |
138 | free_page: |
139 | __free_pages(page, order); |
140 | #endif |
141 | out: |
142 | return ret; |
143 | } |
144 | |
145 | static void atomic_pool_resize(struct gen_pool *pool, gfp_t gfp) |
146 | { |
147 | if (pool && gen_pool_avail(pool) < atomic_pool_size) |
148 | atomic_pool_expand(pool, pool_size: gen_pool_size(pool), gfp); |
149 | } |
150 | |
151 | static void atomic_pool_work_fn(struct work_struct *work) |
152 | { |
153 | if (IS_ENABLED(CONFIG_ZONE_DMA)) |
154 | atomic_pool_resize(pool: atomic_pool_dma, |
155 | GFP_KERNEL | GFP_DMA); |
156 | if (IS_ENABLED(CONFIG_ZONE_DMA32)) |
157 | atomic_pool_resize(pool: atomic_pool_dma32, |
158 | GFP_KERNEL | GFP_DMA32); |
159 | atomic_pool_resize(pool: atomic_pool_kernel, GFP_KERNEL); |
160 | } |
161 | |
162 | static __init struct gen_pool *__dma_atomic_pool_init(size_t pool_size, |
163 | gfp_t gfp) |
164 | { |
165 | struct gen_pool *pool; |
166 | int ret; |
167 | |
168 | pool = gen_pool_create(PAGE_SHIFT, NUMA_NO_NODE); |
169 | if (!pool) |
170 | return NULL; |
171 | |
172 | gen_pool_set_algo(pool, algo: gen_pool_first_fit_order_align, NULL); |
173 | |
174 | ret = atomic_pool_expand(pool, pool_size, gfp); |
175 | if (ret) { |
176 | gen_pool_destroy(pool); |
177 | pr_err("DMA: failed to allocate %zu KiB %pGg pool for atomic allocation\n" , |
178 | pool_size >> 10, &gfp); |
179 | return NULL; |
180 | } |
181 | |
182 | pr_info("DMA: preallocated %zu KiB %pGg pool for atomic allocations\n" , |
183 | gen_pool_size(pool) >> 10, &gfp); |
184 | return pool; |
185 | } |
186 | |
187 | static int __init dma_atomic_pool_init(void) |
188 | { |
189 | int ret = 0; |
190 | |
191 | /* |
192 | * If coherent_pool was not used on the command line, default the pool |
193 | * sizes to 128KB per 1GB of memory, min 128KB, max MAX_PAGE_ORDER. |
194 | */ |
195 | if (!atomic_pool_size) { |
196 | unsigned long pages = totalram_pages() / (SZ_1G / SZ_128K); |
197 | pages = min_t(unsigned long, pages, MAX_ORDER_NR_PAGES); |
198 | atomic_pool_size = max_t(size_t, pages << PAGE_SHIFT, SZ_128K); |
199 | } |
200 | INIT_WORK(&atomic_pool_work, atomic_pool_work_fn); |
201 | |
202 | atomic_pool_kernel = __dma_atomic_pool_init(pool_size: atomic_pool_size, |
203 | GFP_KERNEL); |
204 | if (!atomic_pool_kernel) |
205 | ret = -ENOMEM; |
206 | if (has_managed_dma()) { |
207 | atomic_pool_dma = __dma_atomic_pool_init(pool_size: atomic_pool_size, |
208 | GFP_KERNEL | GFP_DMA); |
209 | if (!atomic_pool_dma) |
210 | ret = -ENOMEM; |
211 | } |
212 | if (IS_ENABLED(CONFIG_ZONE_DMA32)) { |
213 | atomic_pool_dma32 = __dma_atomic_pool_init(pool_size: atomic_pool_size, |
214 | GFP_KERNEL | GFP_DMA32); |
215 | if (!atomic_pool_dma32) |
216 | ret = -ENOMEM; |
217 | } |
218 | |
219 | dma_atomic_pool_debugfs_init(); |
220 | return ret; |
221 | } |
222 | postcore_initcall(dma_atomic_pool_init); |
223 | |
224 | static inline struct gen_pool *dma_guess_pool(struct gen_pool *prev, gfp_t gfp) |
225 | { |
226 | if (prev == NULL) { |
227 | if (IS_ENABLED(CONFIG_ZONE_DMA32) && (gfp & GFP_DMA32)) |
228 | return atomic_pool_dma32; |
229 | if (atomic_pool_dma && (gfp & GFP_DMA)) |
230 | return atomic_pool_dma; |
231 | return atomic_pool_kernel; |
232 | } |
233 | if (prev == atomic_pool_kernel) |
234 | return atomic_pool_dma32 ? atomic_pool_dma32 : atomic_pool_dma; |
235 | if (prev == atomic_pool_dma32) |
236 | return atomic_pool_dma; |
237 | return NULL; |
238 | } |
239 | |
240 | static struct page *__dma_alloc_from_pool(struct device *dev, size_t size, |
241 | struct gen_pool *pool, void **cpu_addr, |
242 | bool (*phys_addr_ok)(struct device *, phys_addr_t, size_t)) |
243 | { |
244 | unsigned long addr; |
245 | phys_addr_t phys; |
246 | |
247 | addr = gen_pool_alloc(pool, size); |
248 | if (!addr) |
249 | return NULL; |
250 | |
251 | phys = gen_pool_virt_to_phys(pool, addr); |
252 | if (phys_addr_ok && !phys_addr_ok(dev, phys, size)) { |
253 | gen_pool_free(pool, addr, size); |
254 | return NULL; |
255 | } |
256 | |
257 | if (gen_pool_avail(pool) < atomic_pool_size) |
258 | schedule_work(work: &atomic_pool_work); |
259 | |
260 | *cpu_addr = (void *)addr; |
261 | memset(*cpu_addr, 0, size); |
262 | return pfn_to_page(__phys_to_pfn(phys)); |
263 | } |
264 | |
265 | struct page *dma_alloc_from_pool(struct device *dev, size_t size, |
266 | void **cpu_addr, gfp_t gfp, |
267 | bool (*phys_addr_ok)(struct device *, phys_addr_t, size_t)) |
268 | { |
269 | struct gen_pool *pool = NULL; |
270 | struct page *page; |
271 | |
272 | while ((pool = dma_guess_pool(prev: pool, gfp))) { |
273 | page = __dma_alloc_from_pool(dev, size, pool, cpu_addr, |
274 | phys_addr_ok); |
275 | if (page) |
276 | return page; |
277 | } |
278 | |
279 | WARN(1, "Failed to get suitable pool for %s\n" , dev_name(dev)); |
280 | return NULL; |
281 | } |
282 | |
283 | bool dma_free_from_pool(struct device *dev, void *start, size_t size) |
284 | { |
285 | struct gen_pool *pool = NULL; |
286 | |
287 | while ((pool = dma_guess_pool(prev: pool, gfp: 0))) { |
288 | if (!gen_pool_has_addr(pool, start: (unsigned long)start, size)) |
289 | continue; |
290 | gen_pool_free(pool, addr: (unsigned long)start, size); |
291 | return true; |
292 | } |
293 | |
294 | return false; |
295 | } |
296 | |