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> | |
f3beeb4a | 20 | #include <asm/sections.h> |
1da177e4 LT |
21 | |
22 | pg_data_t node_data[MAX_NUMNODES]; | |
cff52daf | 23 | EXPORT_SYMBOL(node_data); |
1da177e4 LT |
24 | |
25 | #undef DEBUG_DISCONTIG | |
26 | #ifdef DEBUG_DISCONTIG | |
27 | #define DBGDCONT(args...) printk(args) | |
28 | #else | |
29 | #define DBGDCONT(args...) | |
30 | #endif | |
31 | ||
fb26b3e6 RK |
32 | #define for_each_mem_cluster(memdesc, _cluster, i) \ |
33 | for ((_cluster) = (memdesc)->cluster, (i) = 0; \ | |
34 | (i) < (memdesc)->numclusters; (i)++, (_cluster)++) | |
1da177e4 LT |
35 | |
36 | static void __init show_mem_layout(void) | |
37 | { | |
38 | struct memclust_struct * cluster; | |
39 | struct memdesc_struct * memdesc; | |
40 | int i; | |
41 | ||
42 | /* Find free clusters, and init and free the bootmem accordingly. */ | |
43 | memdesc = (struct memdesc_struct *) | |
44 | (hwrpb->mddt_offset + (unsigned long) hwrpb); | |
45 | ||
46 | printk("Raw memory layout:\n"); | |
47 | for_each_mem_cluster(memdesc, cluster, i) { | |
48 | printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n", | |
49 | i, cluster->usage, cluster->start_pfn, | |
50 | cluster->start_pfn + cluster->numpages); | |
51 | } | |
52 | } | |
53 | ||
54 | static void __init | |
55 | setup_memory_node(int nid, void *kernel_end) | |
56 | { | |
57 | extern unsigned long mem_size_limit; | |
58 | struct memclust_struct * cluster; | |
59 | struct memdesc_struct * memdesc; | |
60 | unsigned long start_kernel_pfn, end_kernel_pfn; | |
61 | unsigned long bootmap_size, bootmap_pages, bootmap_start; | |
62 | unsigned long start, end; | |
63 | unsigned long node_pfn_start, node_pfn_end; | |
64 | unsigned long node_min_pfn, node_max_pfn; | |
65 | int i; | |
66 | unsigned long node_datasz = PFN_UP(sizeof(pg_data_t)); | |
67 | int show_init = 0; | |
68 | ||
69 | /* Find the bounds of current node */ | |
70 | node_pfn_start = (node_mem_start(nid)) >> PAGE_SHIFT; | |
71 | node_pfn_end = node_pfn_start + (node_mem_size(nid) >> PAGE_SHIFT); | |
72 | ||
73 | /* Find free clusters, and init and free the bootmem accordingly. */ | |
74 | memdesc = (struct memdesc_struct *) | |
75 | (hwrpb->mddt_offset + (unsigned long) hwrpb); | |
76 | ||
77 | /* find the bounds of this node (node_min_pfn/node_max_pfn) */ | |
78 | node_min_pfn = ~0UL; | |
79 | node_max_pfn = 0UL; | |
80 | for_each_mem_cluster(memdesc, cluster, i) { | |
81 | /* Bit 0 is console/PALcode reserved. Bit 1 is | |
82 | non-volatile memory -- we might want to mark | |
83 | this for later. */ | |
84 | if (cluster->usage & 3) | |
85 | continue; | |
86 | ||
87 | start = cluster->start_pfn; | |
88 | end = start + cluster->numpages; | |
89 | ||
90 | if (start >= node_pfn_end || end <= node_pfn_start) | |
91 | continue; | |
92 | ||
93 | if (!show_init) { | |
94 | show_init = 1; | |
95 | printk("Initializing bootmem allocator on Node ID %d\n", nid); | |
96 | } | |
97 | printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n", | |
98 | i, cluster->usage, cluster->start_pfn, | |
99 | cluster->start_pfn + cluster->numpages); | |
100 | ||
101 | if (start < node_pfn_start) | |
102 | start = node_pfn_start; | |
103 | if (end > node_pfn_end) | |
104 | end = node_pfn_end; | |
105 | ||
106 | if (start < node_min_pfn) | |
107 | node_min_pfn = start; | |
108 | if (end > node_max_pfn) | |
109 | node_max_pfn = end; | |
110 | } | |
111 | ||
112 | if (mem_size_limit && node_max_pfn > mem_size_limit) { | |
113 | static int msg_shown = 0; | |
114 | if (!msg_shown) { | |
115 | msg_shown = 1; | |
116 | printk("setup: forcing memory size to %ldK (from %ldK).\n", | |
117 | mem_size_limit << (PAGE_SHIFT - 10), | |
118 | node_max_pfn << (PAGE_SHIFT - 10)); | |
119 | } | |
120 | node_max_pfn = mem_size_limit; | |
121 | } | |
122 | ||
123 | if (node_min_pfn >= node_max_pfn) | |
124 | return; | |
125 | ||
126 | /* Update global {min,max}_low_pfn from node information. */ | |
127 | if (node_min_pfn < min_low_pfn) | |
128 | min_low_pfn = node_min_pfn; | |
129 | if (node_max_pfn > max_low_pfn) | |
130 | max_pfn = max_low_pfn = node_max_pfn; | |
131 | ||
1da177e4 LT |
132 | #if 0 /* we'll try this one again in a little while */ |
133 | /* Cute trick to make sure our local node data is on local memory */ | |
134 | node_data[nid] = (pg_data_t *)(__va(node_min_pfn << PAGE_SHIFT)); | |
135 | #endif | |
136 | /* Quasi-mark the pg_data_t as in-use */ | |
137 | node_min_pfn += node_datasz; | |
138 | if (node_min_pfn >= node_max_pfn) { | |
139 | printk(" not enough mem to reserve NODE_DATA"); | |
140 | return; | |
141 | } | |
b61bfa3c | 142 | NODE_DATA(nid)->bdata = &bootmem_node_data[nid]; |
1da177e4 LT |
143 | |
144 | printk(" Detected node memory: start %8lu, end %8lu\n", | |
145 | node_min_pfn, node_max_pfn); | |
146 | ||
147 | DBGDCONT(" DISCONTIG: node_data[%d] is at 0x%p\n", nid, NODE_DATA(nid)); | |
148 | DBGDCONT(" DISCONTIG: NODE_DATA(%d)->bdata is at 0x%p\n", nid, NODE_DATA(nid)->bdata); | |
149 | ||
150 | /* Find the bounds of kernel memory. */ | |
151 | start_kernel_pfn = PFN_DOWN(KERNEL_START_PHYS); | |
152 | end_kernel_pfn = PFN_UP(virt_to_phys(kernel_end)); | |
153 | bootmap_start = -1; | |
154 | ||
155 | if (!nid && (node_max_pfn < end_kernel_pfn || node_min_pfn > start_kernel_pfn)) | |
156 | panic("kernel loaded out of ram"); | |
157 | ||
158 | /* Zone start phys-addr must be 2^(MAX_ORDER-1) aligned. | |
159 | Note that we round this down, not up - node memory | |
160 | has much larger alignment than 8Mb, so it's safe. */ | |
161 | node_min_pfn &= ~((1UL << (MAX_ORDER-1))-1); | |
162 | ||
163 | /* We need to know how many physically contiguous pages | |
164 | we'll need for the bootmap. */ | |
165 | bootmap_pages = bootmem_bootmap_pages(node_max_pfn-node_min_pfn); | |
166 | ||
167 | /* Now find a good region where to allocate the bootmap. */ | |
168 | for_each_mem_cluster(memdesc, cluster, i) { | |
169 | if (cluster->usage & 3) | |
170 | continue; | |
171 | ||
172 | start = cluster->start_pfn; | |
173 | end = start + cluster->numpages; | |
174 | ||
175 | if (start >= node_max_pfn || end <= node_min_pfn) | |
176 | continue; | |
177 | ||
178 | if (end > node_max_pfn) | |
179 | end = node_max_pfn; | |
180 | if (start < node_min_pfn) | |
181 | start = node_min_pfn; | |
182 | ||
183 | if (start < start_kernel_pfn) { | |
184 | if (end > end_kernel_pfn | |
185 | && end - end_kernel_pfn >= bootmap_pages) { | |
186 | bootmap_start = end_kernel_pfn; | |
187 | break; | |
188 | } else if (end > start_kernel_pfn) | |
189 | end = start_kernel_pfn; | |
190 | } else if (start < end_kernel_pfn) | |
191 | start = end_kernel_pfn; | |
192 | if (end - start >= bootmap_pages) { | |
193 | bootmap_start = start; | |
194 | break; | |
195 | } | |
196 | } | |
197 | ||
198 | if (bootmap_start == -1) | |
af901ca1 | 199 | panic("couldn't find a contiguous place for the bootmap"); |
1da177e4 LT |
200 | |
201 | /* Allocate the bootmap and mark the whole MM as reserved. */ | |
202 | bootmap_size = init_bootmem_node(NODE_DATA(nid), bootmap_start, | |
203 | node_min_pfn, node_max_pfn); | |
204 | DBGDCONT(" bootmap_start %lu, bootmap_size %lu, bootmap_pages %lu\n", | |
205 | bootmap_start, bootmap_size, bootmap_pages); | |
206 | ||
207 | /* Mark the free regions. */ | |
208 | for_each_mem_cluster(memdesc, cluster, i) { | |
209 | if (cluster->usage & 3) | |
210 | continue; | |
211 | ||
212 | start = cluster->start_pfn; | |
213 | end = cluster->start_pfn + cluster->numpages; | |
214 | ||
215 | if (start >= node_max_pfn || end <= node_min_pfn) | |
216 | continue; | |
217 | ||
218 | if (end > node_max_pfn) | |
219 | end = node_max_pfn; | |
220 | if (start < node_min_pfn) | |
221 | start = node_min_pfn; | |
222 | ||
223 | if (start < start_kernel_pfn) { | |
224 | if (end > end_kernel_pfn) { | |
225 | free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start), | |
226 | (PFN_PHYS(start_kernel_pfn) | |
227 | - PFN_PHYS(start))); | |
228 | printk(" freeing pages %ld:%ld\n", | |
229 | start, start_kernel_pfn); | |
230 | start = end_kernel_pfn; | |
231 | } else if (end > start_kernel_pfn) | |
232 | end = start_kernel_pfn; | |
233 | } else if (start < end_kernel_pfn) | |
234 | start = end_kernel_pfn; | |
235 | if (start >= end) | |
236 | continue; | |
237 | ||
238 | free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start), PFN_PHYS(end) - PFN_PHYS(start)); | |
239 | printk(" freeing pages %ld:%ld\n", start, end); | |
240 | } | |
241 | ||
242 | /* Reserve the bootmap memory. */ | |
72a7fe39 BW |
243 | reserve_bootmem_node(NODE_DATA(nid), PFN_PHYS(bootmap_start), |
244 | bootmap_size, BOOTMEM_DEFAULT); | |
1da177e4 LT |
245 | printk(" reserving pages %ld:%ld\n", bootmap_start, bootmap_start+PFN_UP(bootmap_size)); |
246 | ||
247 | node_set_online(nid); | |
248 | } | |
249 | ||
250 | void __init | |
251 | setup_memory(void *kernel_end) | |
252 | { | |
253 | int nid; | |
254 | ||
255 | show_mem_layout(); | |
256 | ||
257 | nodes_clear(node_online_map); | |
258 | ||
259 | min_low_pfn = ~0UL; | |
260 | max_low_pfn = 0UL; | |
261 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
262 | setup_memory_node(nid, kernel_end); | |
263 | ||
264 | #ifdef CONFIG_BLK_DEV_INITRD | |
265 | initrd_start = INITRD_START; | |
266 | if (initrd_start) { | |
267 | extern void *move_initrd(unsigned long); | |
268 | ||
269 | initrd_end = initrd_start+INITRD_SIZE; | |
270 | printk("Initial ramdisk at: 0x%p (%lu bytes)\n", | |
271 | (void *) initrd_start, INITRD_SIZE); | |
272 | ||
273 | if ((void *)initrd_end > phys_to_virt(PFN_PHYS(max_low_pfn))) { | |
274 | if (!move_initrd(PFN_PHYS(max_low_pfn))) | |
275 | printk("initrd extends beyond end of memory " | |
276 | "(0x%08lx > 0x%p)\ndisabling initrd\n", | |
277 | initrd_end, | |
278 | phys_to_virt(PFN_PHYS(max_low_pfn))); | |
279 | } else { | |
280 | nid = kvaddr_to_nid(initrd_start); | |
281 | reserve_bootmem_node(NODE_DATA(nid), | |
282 | virt_to_phys((void *)initrd_start), | |
72a7fe39 | 283 | INITRD_SIZE, BOOTMEM_DEFAULT); |
1da177e4 LT |
284 | } |
285 | } | |
286 | #endif /* CONFIG_BLK_DEV_INITRD */ | |
287 | } | |
288 | ||
289 | void __init paging_init(void) | |
290 | { | |
291 | unsigned int nid; | |
292 | unsigned long zones_size[MAX_NR_ZONES] = {0, }; | |
293 | unsigned long dma_local_pfn; | |
294 | ||
295 | /* | |
296 | * The old global MAX_DMA_ADDRESS per-arch API doesn't fit | |
297 | * in the NUMA model, for now we convert it to a pfn and | |
298 | * we interpret this pfn as a local per-node information. | |
299 | * This issue isn't very important since none of these machines | |
300 | * have legacy ISA slots anyways. | |
301 | */ | |
302 | dma_local_pfn = virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT; | |
303 | ||
304 | for_each_online_node(nid) { | |
b61bfa3c | 305 | bootmem_data_t *bdata = &bootmem_node_data[nid]; |
3560e249 | 306 | unsigned long start_pfn = bdata->node_min_pfn; |
b61bfa3c | 307 | unsigned long end_pfn = bdata->node_low_pfn; |
1da177e4 LT |
308 | |
309 | if (dma_local_pfn >= end_pfn - start_pfn) | |
310 | zones_size[ZONE_DMA] = end_pfn - start_pfn; | |
311 | else { | |
312 | zones_size[ZONE_DMA] = dma_local_pfn; | |
313 | zones_size[ZONE_NORMAL] = (end_pfn - start_pfn) - dma_local_pfn; | |
314 | } | |
ad8b4b28 | 315 | node_set_state(nid, N_NORMAL_MEMORY); |
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 | } |