xref: /linux/mm/sparse-vmemmap.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Virtual Memory Map support
4  *
5  * (C) 2007 sgi. Christoph Lameter.
6  *
7  * Virtual memory maps allow VM primitives pfn_to_page, page_to_pfn,
8  * virt_to_page, page_address() to be implemented as a base offset
9  * calculation without memory access.
10  *
11  * However, virtual mappings need a page table and TLBs. Many Linux
12  * architectures already map their physical space using 1-1 mappings
13  * via TLBs. For those arches the virtual memory map is essentially
14  * for free if we use the same page size as the 1-1 mappings. In that
15  * case the overhead consists of a few additional pages that are
16  * allocated to create a view of memory for vmemmap.
17  *
18  * The architecture is expected to provide a vmemmap_populate() function
19  * to instantiate the mapping.
20  */
21 #include <linux/mm.h>
22 #include <linux/mmzone.h>
23 #include <linux/bootmem.h>
24 #include <linux/memremap.h>
25 #include <linux/highmem.h>
26 #include <linux/slab.h>
27 #include <linux/spinlock.h>
28 #include <linux/vmalloc.h>
29 #include <linux/sched.h>
30 #include <asm/dma.h>
31 #include <asm/pgalloc.h>
32 #include <asm/pgtable.h>
33 
34 /*
35  * Allocate a block of memory to be used to back the virtual memory map
36  * or to back the page tables that are used to create the mapping.
37  * Uses the main allocators if they are available, else bootmem.
38  */
39 
40 static void * __ref __earlyonly_bootmem_alloc(int node,
41 				unsigned long size,
42 				unsigned long align,
43 				unsigned long goal)
44 {
45 	return memblock_virt_alloc_try_nid_raw(size, align, goal,
46 					    BOOTMEM_ALLOC_ACCESSIBLE, node);
47 }
48 
49 static void *vmemmap_buf;
50 static void *vmemmap_buf_end;
51 
52 void * __meminit vmemmap_alloc_block(unsigned long size, int node)
53 {
54 	/* If the main allocator is up use that, fallback to bootmem. */
55 	if (slab_is_available()) {
56 		gfp_t gfp_mask = GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
57 		int order = get_order(size);
58 		static bool warned;
59 		struct page *page;
60 
61 		page = alloc_pages_node(node, gfp_mask, order);
62 		if (page)
63 			return page_address(page);
64 
65 		if (!warned) {
66 			warn_alloc(gfp_mask & ~__GFP_NOWARN, NULL,
67 				   "vmemmap alloc failure: order:%u", order);
68 			warned = true;
69 		}
70 		return NULL;
71 	} else
72 		return __earlyonly_bootmem_alloc(node, size, size,
73 				__pa(MAX_DMA_ADDRESS));
74 }
75 
76 /* need to make sure size is all the same during early stage */
77 static void * __meminit alloc_block_buf(unsigned long size, int node)
78 {
79 	void *ptr;
80 
81 	if (!vmemmap_buf)
82 		return vmemmap_alloc_block(size, node);
83 
84 	/* take the from buf */
85 	ptr = (void *)ALIGN((unsigned long)vmemmap_buf, size);
86 	if (ptr + size > vmemmap_buf_end)
87 		return vmemmap_alloc_block(size, node);
88 
89 	vmemmap_buf = ptr + size;
90 
91 	return ptr;
92 }
93 
94 static unsigned long __meminit vmem_altmap_next_pfn(struct vmem_altmap *altmap)
95 {
96 	return altmap->base_pfn + altmap->reserve + altmap->alloc
97 		+ altmap->align;
98 }
99 
100 static unsigned long __meminit vmem_altmap_nr_free(struct vmem_altmap *altmap)
101 {
102 	unsigned long allocated = altmap->alloc + altmap->align;
103 
104 	if (altmap->free > allocated)
105 		return altmap->free - allocated;
106 	return 0;
107 }
108 
109 /**
110  * vmem_altmap_alloc - allocate pages from the vmem_altmap reservation
111  * @altmap - reserved page pool for the allocation
112  * @nr_pfns - size (in pages) of the allocation
113  *
114  * Allocations are aligned to the size of the request
115  */
116 static unsigned long __meminit vmem_altmap_alloc(struct vmem_altmap *altmap,
117 		unsigned long nr_pfns)
118 {
119 	unsigned long pfn = vmem_altmap_next_pfn(altmap);
120 	unsigned long nr_align;
121 
122 	nr_align = 1UL << find_first_bit(&nr_pfns, BITS_PER_LONG);
123 	nr_align = ALIGN(pfn, nr_align) - pfn;
124 
125 	if (nr_pfns + nr_align > vmem_altmap_nr_free(altmap))
126 		return ULONG_MAX;
127 	altmap->alloc += nr_pfns;
128 	altmap->align += nr_align;
129 	return pfn + nr_align;
130 }
131 
132 static void * __meminit altmap_alloc_block_buf(unsigned long size,
133 		struct vmem_altmap *altmap)
134 {
135 	unsigned long pfn, nr_pfns;
136 	void *ptr;
137 
138 	if (size & ~PAGE_MASK) {
139 		pr_warn_once("%s: allocations must be multiple of PAGE_SIZE (%ld)\n",
140 				__func__, size);
141 		return NULL;
142 	}
143 
144 	nr_pfns = size >> PAGE_SHIFT;
145 	pfn = vmem_altmap_alloc(altmap, nr_pfns);
146 	if (pfn < ULONG_MAX)
147 		ptr = __va(__pfn_to_phys(pfn));
148 	else
149 		ptr = NULL;
150 	pr_debug("%s: pfn: %#lx alloc: %ld align: %ld nr: %#lx\n",
151 			__func__, pfn, altmap->alloc, altmap->align, nr_pfns);
152 
153 	return ptr;
154 }
155 
156 /* need to make sure size is all the same during early stage */
157 void * __meminit __vmemmap_alloc_block_buf(unsigned long size, int node,
158 		struct vmem_altmap *altmap)
159 {
160 	if (altmap)
161 		return altmap_alloc_block_buf(size, altmap);
162 	return alloc_block_buf(size, node);
163 }
164 
165 void __meminit vmemmap_verify(pte_t *pte, int node,
166 				unsigned long start, unsigned long end)
167 {
168 	unsigned long pfn = pte_pfn(*pte);
169 	int actual_node = early_pfn_to_nid(pfn);
170 
171 	if (node_distance(actual_node, node) > LOCAL_DISTANCE)
172 		pr_warn("[%lx-%lx] potential offnode page_structs\n",
173 			start, end - 1);
174 }
175 
176 pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node)
177 {
178 	pte_t *pte = pte_offset_kernel(pmd, addr);
179 	if (pte_none(*pte)) {
180 		pte_t entry;
181 		void *p = alloc_block_buf(PAGE_SIZE, node);
182 		if (!p)
183 			return NULL;
184 		entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
185 		set_pte_at(&init_mm, addr, pte, entry);
186 	}
187 	return pte;
188 }
189 
190 static void * __meminit vmemmap_alloc_block_zero(unsigned long size, int node)
191 {
192 	void *p = vmemmap_alloc_block(size, node);
193 
194 	if (!p)
195 		return NULL;
196 	memset(p, 0, size);
197 
198 	return p;
199 }
200 
201 pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
202 {
203 	pmd_t *pmd = pmd_offset(pud, addr);
204 	if (pmd_none(*pmd)) {
205 		void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
206 		if (!p)
207 			return NULL;
208 		pmd_populate_kernel(&init_mm, pmd, p);
209 	}
210 	return pmd;
211 }
212 
213 pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, int node)
214 {
215 	pud_t *pud = pud_offset(p4d, addr);
216 	if (pud_none(*pud)) {
217 		void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
218 		if (!p)
219 			return NULL;
220 		pud_populate(&init_mm, pud, p);
221 	}
222 	return pud;
223 }
224 
225 p4d_t * __meminit vmemmap_p4d_populate(pgd_t *pgd, unsigned long addr, int node)
226 {
227 	p4d_t *p4d = p4d_offset(pgd, addr);
228 	if (p4d_none(*p4d)) {
229 		void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
230 		if (!p)
231 			return NULL;
232 		p4d_populate(&init_mm, p4d, p);
233 	}
234 	return p4d;
235 }
236 
237 pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
238 {
239 	pgd_t *pgd = pgd_offset_k(addr);
240 	if (pgd_none(*pgd)) {
241 		void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
242 		if (!p)
243 			return NULL;
244 		pgd_populate(&init_mm, pgd, p);
245 	}
246 	return pgd;
247 }
248 
249 int __meminit vmemmap_populate_basepages(unsigned long start,
250 					 unsigned long end, int node)
251 {
252 	unsigned long addr = start;
253 	pgd_t *pgd;
254 	p4d_t *p4d;
255 	pud_t *pud;
256 	pmd_t *pmd;
257 	pte_t *pte;
258 
259 	for (; addr < end; addr += PAGE_SIZE) {
260 		pgd = vmemmap_pgd_populate(addr, node);
261 		if (!pgd)
262 			return -ENOMEM;
263 		p4d = vmemmap_p4d_populate(pgd, addr, node);
264 		if (!p4d)
265 			return -ENOMEM;
266 		pud = vmemmap_pud_populate(p4d, addr, node);
267 		if (!pud)
268 			return -ENOMEM;
269 		pmd = vmemmap_pmd_populate(pud, addr, node);
270 		if (!pmd)
271 			return -ENOMEM;
272 		pte = vmemmap_pte_populate(pmd, addr, node);
273 		if (!pte)
274 			return -ENOMEM;
275 		vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
276 	}
277 
278 	return 0;
279 }
280 
281 struct page * __meminit sparse_mem_map_populate(unsigned long pnum, int nid)
282 {
283 	unsigned long start;
284 	unsigned long end;
285 	struct page *map;
286 
287 	map = pfn_to_page(pnum * PAGES_PER_SECTION);
288 	start = (unsigned long)map;
289 	end = (unsigned long)(map + PAGES_PER_SECTION);
290 
291 	if (vmemmap_populate(start, end, nid))
292 		return NULL;
293 
294 	return map;
295 }
296 
297 void __init sparse_mem_maps_populate_node(struct page **map_map,
298 					  unsigned long pnum_begin,
299 					  unsigned long pnum_end,
300 					  unsigned long map_count, int nodeid)
301 {
302 	unsigned long pnum;
303 	unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
304 	void *vmemmap_buf_start;
305 
306 	size = ALIGN(size, PMD_SIZE);
307 	vmemmap_buf_start = __earlyonly_bootmem_alloc(nodeid, size * map_count,
308 			 PMD_SIZE, __pa(MAX_DMA_ADDRESS));
309 
310 	if (vmemmap_buf_start) {
311 		vmemmap_buf = vmemmap_buf_start;
312 		vmemmap_buf_end = vmemmap_buf_start + size * map_count;
313 	}
314 
315 	for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
316 		struct mem_section *ms;
317 
318 		if (!present_section_nr(pnum))
319 			continue;
320 
321 		map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
322 		if (map_map[pnum])
323 			continue;
324 		ms = __nr_to_section(pnum);
325 		pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
326 		       __func__);
327 		ms->section_mem_map = 0;
328 	}
329 
330 	if (vmemmap_buf_start) {
331 		/* need to free left buf */
332 		memblock_free_early(__pa(vmemmap_buf),
333 				    vmemmap_buf_end - vmemmap_buf);
334 		vmemmap_buf = NULL;
335 		vmemmap_buf_end = NULL;
336 	}
337 }
338