Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/arch/alpha/mm/numa.c | |
3 | * | |
4 | * DISCONTIGMEM NUMA alpha support. | |
5 | * | |
6 | * Copyright (C) 2001 Andrea Arcangeli <andrea@suse.de> SuSE | |
7 | */ | |
8 | ||
1da177e4 LT |
9 | #include <linux/types.h> |
10 | #include <linux/kernel.h> | |
11 | #include <linux/mm.h> | |
12 | #include <linux/bootmem.h> | |
13 | #include <linux/swap.h> | |
14 | #include <linux/initrd.h> | |
22a9835c | 15 | #include <linux/pfn.h> |
cff52daf | 16 | #include <linux/module.h> |
1da177e4 LT |
17 | |
18 | #include <asm/hwrpb.h> | |
19 | #include <asm/pgalloc.h> | |
20 | ||
21 | pg_data_t node_data[MAX_NUMNODES]; | |
cff52daf | 22 | EXPORT_SYMBOL(node_data); |
1da177e4 LT |
23 | |
24 | #undef DEBUG_DISCONTIG | |
25 | #ifdef DEBUG_DISCONTIG | |
26 | #define DBGDCONT(args...) printk(args) | |
27 | #else | |
28 | #define DBGDCONT(args...) | |
29 | #endif | |
30 | ||
fb26b3e6 RK |
31 | #define for_each_mem_cluster(memdesc, _cluster, i) \ |
32 | for ((_cluster) = (memdesc)->cluster, (i) = 0; \ | |
33 | (i) < (memdesc)->numclusters; (i)++, (_cluster)++) | |
1da177e4 LT |
34 | |
35 | static void __init show_mem_layout(void) | |
36 | { | |
37 | struct memclust_struct * cluster; | |
38 | struct memdesc_struct * memdesc; | |
39 | int i; | |
40 | ||
41 | /* Find free clusters, and init and free the bootmem accordingly. */ | |
42 | memdesc = (struct memdesc_struct *) | |
43 | (hwrpb->mddt_offset + (unsigned long) hwrpb); | |
44 | ||
45 | printk("Raw memory layout:\n"); | |
46 | for_each_mem_cluster(memdesc, cluster, i) { | |
47 | printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n", | |
48 | i, cluster->usage, cluster->start_pfn, | |
49 | cluster->start_pfn + cluster->numpages); | |
50 | } | |
51 | } | |
52 | ||
53 | static void __init | |
54 | setup_memory_node(int nid, void *kernel_end) | |
55 | { | |
56 | extern unsigned long mem_size_limit; | |
57 | struct memclust_struct * cluster; | |
58 | struct memdesc_struct * memdesc; | |
59 | unsigned long start_kernel_pfn, end_kernel_pfn; | |
60 | unsigned long bootmap_size, bootmap_pages, bootmap_start; | |
61 | unsigned long start, end; | |
62 | unsigned long node_pfn_start, node_pfn_end; | |
63 | unsigned long node_min_pfn, node_max_pfn; | |
64 | int i; | |
65 | unsigned long node_datasz = PFN_UP(sizeof(pg_data_t)); | |
66 | int show_init = 0; | |
67 | ||
68 | /* Find the bounds of current node */ | |
69 | node_pfn_start = (node_mem_start(nid)) >> PAGE_SHIFT; | |
70 | node_pfn_end = node_pfn_start + (node_mem_size(nid) >> PAGE_SHIFT); | |
71 | ||
72 | /* Find free clusters, and init and free the bootmem accordingly. */ | |
73 | memdesc = (struct memdesc_struct *) | |
74 | (hwrpb->mddt_offset + (unsigned long) hwrpb); | |
75 | ||
76 | /* find the bounds of this node (node_min_pfn/node_max_pfn) */ | |
77 | node_min_pfn = ~0UL; | |
78 | node_max_pfn = 0UL; | |
79 | for_each_mem_cluster(memdesc, cluster, i) { | |
80 | /* Bit 0 is console/PALcode reserved. Bit 1 is | |
81 | non-volatile memory -- we might want to mark | |
82 | this for later. */ | |
83 | if (cluster->usage & 3) | |
84 | continue; | |
85 | ||
86 | start = cluster->start_pfn; | |
87 | end = start + cluster->numpages; | |
88 | ||
89 | if (start >= node_pfn_end || end <= node_pfn_start) | |
90 | continue; | |
91 | ||
92 | if (!show_init) { | |
93 | show_init = 1; | |
94 | printk("Initializing bootmem allocator on Node ID %d\n", nid); | |
95 | } | |
96 | printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n", | |
97 | i, cluster->usage, cluster->start_pfn, | |
98 | cluster->start_pfn + cluster->numpages); | |
99 | ||
100 | if (start < node_pfn_start) | |
101 | start = node_pfn_start; | |
102 | if (end > node_pfn_end) | |
103 | end = node_pfn_end; | |
104 | ||
105 | if (start < node_min_pfn) | |
106 | node_min_pfn = start; | |
107 | if (end > node_max_pfn) | |
108 | node_max_pfn = end; | |
109 | } | |
110 | ||
111 | if (mem_size_limit && node_max_pfn > mem_size_limit) { | |
112 | static int msg_shown = 0; | |
113 | if (!msg_shown) { | |
114 | msg_shown = 1; | |
115 | printk("setup: forcing memory size to %ldK (from %ldK).\n", | |
116 | mem_size_limit << (PAGE_SHIFT - 10), | |
117 | node_max_pfn << (PAGE_SHIFT - 10)); | |
118 | } | |
119 | node_max_pfn = mem_size_limit; | |
120 | } | |
121 | ||
122 | if (node_min_pfn >= node_max_pfn) | |
123 | return; | |
124 | ||
125 | /* Update global {min,max}_low_pfn from node information. */ | |
126 | if (node_min_pfn < min_low_pfn) | |
127 | min_low_pfn = node_min_pfn; | |
128 | if (node_max_pfn > max_low_pfn) | |
129 | max_pfn = max_low_pfn = node_max_pfn; | |
130 | ||
131 | num_physpages += node_max_pfn - node_min_pfn; | |
132 | ||
133 | #if 0 /* we'll try this one again in a little while */ | |
134 | /* Cute trick to make sure our local node data is on local memory */ | |
135 | node_data[nid] = (pg_data_t *)(__va(node_min_pfn << PAGE_SHIFT)); | |
136 | #endif | |
137 | /* Quasi-mark the pg_data_t as in-use */ | |
138 | node_min_pfn += node_datasz; | |
139 | if (node_min_pfn >= node_max_pfn) { | |
140 | printk(" not enough mem to reserve NODE_DATA"); | |
141 | return; | |
142 | } | |
b61bfa3c | 143 | NODE_DATA(nid)->bdata = &bootmem_node_data[nid]; |
1da177e4 LT |
144 | |
145 | printk(" Detected node memory: start %8lu, end %8lu\n", | |
146 | node_min_pfn, node_max_pfn); | |
147 | ||
148 | DBGDCONT(" DISCONTIG: node_data[%d] is at 0x%p\n", nid, NODE_DATA(nid)); | |
149 | DBGDCONT(" DISCONTIG: NODE_DATA(%d)->bdata is at 0x%p\n", nid, NODE_DATA(nid)->bdata); | |
150 | ||
151 | /* Find the bounds of kernel memory. */ | |
152 | start_kernel_pfn = PFN_DOWN(KERNEL_START_PHYS); | |
153 | end_kernel_pfn = PFN_UP(virt_to_phys(kernel_end)); | |
154 | bootmap_start = -1; | |
155 | ||
156 | if (!nid && (node_max_pfn < end_kernel_pfn || node_min_pfn > start_kernel_pfn)) | |
157 | panic("kernel loaded out of ram"); | |
158 | ||
159 | /* Zone start phys-addr must be 2^(MAX_ORDER-1) aligned. | |
160 | Note that we round this down, not up - node memory | |
161 | has much larger alignment than 8Mb, so it's safe. */ | |
162 | node_min_pfn &= ~((1UL << (MAX_ORDER-1))-1); | |
163 | ||
164 | /* We need to know how many physically contiguous pages | |
165 | we'll need for the bootmap. */ | |
166 | bootmap_pages = bootmem_bootmap_pages(node_max_pfn-node_min_pfn); | |
167 | ||
168 | /* Now find a good region where to allocate the bootmap. */ | |
169 | for_each_mem_cluster(memdesc, cluster, i) { | |
170 | if (cluster->usage & 3) | |
171 | continue; | |
172 | ||
173 | start = cluster->start_pfn; | |
174 | end = start + cluster->numpages; | |
175 | ||
176 | if (start >= node_max_pfn || end <= node_min_pfn) | |
177 | continue; | |
178 | ||
179 | if (end > node_max_pfn) | |
180 | end = node_max_pfn; | |
181 | if (start < node_min_pfn) | |
182 | start = node_min_pfn; | |
183 | ||
184 | if (start < start_kernel_pfn) { | |
185 | if (end > end_kernel_pfn | |
186 | && end - end_kernel_pfn >= bootmap_pages) { | |
187 | bootmap_start = end_kernel_pfn; | |
188 | break; | |
189 | } else if (end > start_kernel_pfn) | |
190 | end = start_kernel_pfn; | |
191 | } else if (start < end_kernel_pfn) | |
192 | start = end_kernel_pfn; | |
193 | if (end - start >= bootmap_pages) { | |
194 | bootmap_start = start; | |
195 | break; | |
196 | } | |
197 | } | |
198 | ||
199 | if (bootmap_start == -1) | |
200 | panic("couldn't find a contigous place for the bootmap"); | |
201 | ||
202 | /* Allocate the bootmap and mark the whole MM as reserved. */ | |
203 | bootmap_size = init_bootmem_node(NODE_DATA(nid), bootmap_start, | |
204 | node_min_pfn, node_max_pfn); | |
205 | DBGDCONT(" bootmap_start %lu, bootmap_size %lu, bootmap_pages %lu\n", | |
206 | bootmap_start, bootmap_size, bootmap_pages); | |
207 | ||
208 | /* Mark the free regions. */ | |
209 | for_each_mem_cluster(memdesc, cluster, i) { | |
210 | if (cluster->usage & 3) | |
211 | continue; | |
212 | ||
213 | start = cluster->start_pfn; | |
214 | end = cluster->start_pfn + cluster->numpages; | |
215 | ||
216 | if (start >= node_max_pfn || end <= node_min_pfn) | |
217 | continue; | |
218 | ||
219 | if (end > node_max_pfn) | |
220 | end = node_max_pfn; | |
221 | if (start < node_min_pfn) | |
222 | start = node_min_pfn; | |
223 | ||
224 | if (start < start_kernel_pfn) { | |
225 | if (end > end_kernel_pfn) { | |
226 | free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start), | |
227 | (PFN_PHYS(start_kernel_pfn) | |
228 | - PFN_PHYS(start))); | |
229 | printk(" freeing pages %ld:%ld\n", | |
230 | start, start_kernel_pfn); | |
231 | start = end_kernel_pfn; | |
232 | } else if (end > start_kernel_pfn) | |
233 | end = start_kernel_pfn; | |
234 | } else if (start < end_kernel_pfn) | |
235 | start = end_kernel_pfn; | |
236 | if (start >= end) | |
237 | continue; | |
238 | ||
239 | free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start), PFN_PHYS(end) - PFN_PHYS(start)); | |
240 | printk(" freeing pages %ld:%ld\n", start, end); | |
241 | } | |
242 | ||
243 | /* Reserve the bootmap memory. */ | |
72a7fe39 BW |
244 | reserve_bootmem_node(NODE_DATA(nid), PFN_PHYS(bootmap_start), |
245 | bootmap_size, BOOTMEM_DEFAULT); | |
1da177e4 LT |
246 | printk(" reserving pages %ld:%ld\n", bootmap_start, bootmap_start+PFN_UP(bootmap_size)); |
247 | ||
248 | node_set_online(nid); | |
249 | } | |
250 | ||
251 | void __init | |
252 | setup_memory(void *kernel_end) | |
253 | { | |
254 | int nid; | |
255 | ||
256 | show_mem_layout(); | |
257 | ||
258 | nodes_clear(node_online_map); | |
259 | ||
260 | min_low_pfn = ~0UL; | |
261 | max_low_pfn = 0UL; | |
262 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
263 | setup_memory_node(nid, kernel_end); | |
264 | ||
265 | #ifdef CONFIG_BLK_DEV_INITRD | |
266 | initrd_start = INITRD_START; | |
267 | if (initrd_start) { | |
268 | extern void *move_initrd(unsigned long); | |
269 | ||
270 | initrd_end = initrd_start+INITRD_SIZE; | |
271 | printk("Initial ramdisk at: 0x%p (%lu bytes)\n", | |
272 | (void *) initrd_start, INITRD_SIZE); | |
273 | ||
274 | if ((void *)initrd_end > phys_to_virt(PFN_PHYS(max_low_pfn))) { | |
275 | if (!move_initrd(PFN_PHYS(max_low_pfn))) | |
276 | printk("initrd extends beyond end of memory " | |
277 | "(0x%08lx > 0x%p)\ndisabling initrd\n", | |
278 | initrd_end, | |
279 | phys_to_virt(PFN_PHYS(max_low_pfn))); | |
280 | } else { | |
281 | nid = kvaddr_to_nid(initrd_start); | |
282 | reserve_bootmem_node(NODE_DATA(nid), | |
283 | virt_to_phys((void *)initrd_start), | |
72a7fe39 | 284 | INITRD_SIZE, BOOTMEM_DEFAULT); |
1da177e4 LT |
285 | } |
286 | } | |
287 | #endif /* CONFIG_BLK_DEV_INITRD */ | |
288 | } | |
289 | ||
290 | void __init paging_init(void) | |
291 | { | |
292 | unsigned int nid; | |
293 | unsigned long zones_size[MAX_NR_ZONES] = {0, }; | |
294 | unsigned long dma_local_pfn; | |
295 | ||
296 | /* | |
297 | * The old global MAX_DMA_ADDRESS per-arch API doesn't fit | |
298 | * in the NUMA model, for now we convert it to a pfn and | |
299 | * we interpret this pfn as a local per-node information. | |
300 | * This issue isn't very important since none of these machines | |
301 | * have legacy ISA slots anyways. | |
302 | */ | |
303 | dma_local_pfn = virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT; | |
304 | ||
305 | for_each_online_node(nid) { | |
b61bfa3c | 306 | bootmem_data_t *bdata = &bootmem_node_data[nid]; |
3560e249 | 307 | unsigned long start_pfn = bdata->node_min_pfn; |
b61bfa3c | 308 | unsigned long end_pfn = bdata->node_low_pfn; |
1da177e4 LT |
309 | |
310 | if (dma_local_pfn >= end_pfn - start_pfn) | |
311 | zones_size[ZONE_DMA] = end_pfn - start_pfn; | |
312 | else { | |
313 | zones_size[ZONE_DMA] = dma_local_pfn; | |
314 | zones_size[ZONE_NORMAL] = (end_pfn - start_pfn) - dma_local_pfn; | |
315 | } | |
9109fb7b | 316 | free_area_init_node(nid, zones_size, start_pfn, NULL); |
1da177e4 LT |
317 | } |
318 | ||
319 | /* Initialize the kernel's ZERO_PGE. */ | |
320 | memset((void *)ZERO_PGE, 0, PAGE_SIZE); | |
321 | } | |
322 | ||
323 | void __init mem_init(void) | |
324 | { | |
325 | unsigned long codesize, reservedpages, datasize, initsize, pfn; | |
326 | extern int page_is_ram(unsigned long) __init; | |
327 | extern char _text, _etext, _data, _edata; | |
328 | extern char __init_begin, __init_end; | |
329 | unsigned long nid, i; | |
1da177e4 LT |
330 | high_memory = (void *) __va(max_low_pfn << PAGE_SHIFT); |
331 | ||
332 | reservedpages = 0; | |
333 | for_each_online_node(nid) { | |
334 | /* | |
335 | * This will free up the bootmem, ie, slot 0 memory | |
336 | */ | |
337 | totalram_pages += free_all_bootmem_node(NODE_DATA(nid)); | |
338 | ||
1da177e4 LT |
339 | pfn = NODE_DATA(nid)->node_start_pfn; |
340 | for (i = 0; i < node_spanned_pages(nid); i++, pfn++) | |
408fde81 DH |
341 | if (page_is_ram(pfn) && |
342 | PageReserved(nid_page_nr(nid, i))) | |
1da177e4 LT |
343 | reservedpages++; |
344 | } | |
345 | ||
346 | codesize = (unsigned long) &_etext - (unsigned long) &_text; | |
347 | datasize = (unsigned long) &_edata - (unsigned long) &_data; | |
348 | initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin; | |
349 | ||
350 | printk("Memory: %luk/%luk available (%luk kernel code, %luk reserved, " | |
351 | "%luk data, %luk init)\n", | |
cc013a88 | 352 | nr_free_pages() << (PAGE_SHIFT-10), |
1da177e4 LT |
353 | num_physpages << (PAGE_SHIFT-10), |
354 | codesize >> 10, | |
355 | reservedpages << (PAGE_SHIFT-10), | |
356 | datasize >> 10, | |
357 | initsize >> 10); | |
358 | #if 0 | |
359 | mem_stress(); | |
360 | #endif | |
361 | } |