File: /usr/src/linux/arch/alpha/mm/numa.c
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
9 #include <linux/config.h>
10 #include <linux/types.h>
11 #include <linux/kernel.h>
12 #include <linux/mm.h>
13 #include <linux/bootmem.h>
14 #include <linux/swap.h>
15 #ifdef CONFIG_BLK_DEV_INITRD
16 #include <linux/blk.h>
17 #endif
18
19 #include <asm/hwrpb.h>
20 #include <asm/pgalloc.h>
21
22 plat_pg_data_t *plat_node_data[MAX_NUMNODES];
23 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
24
25 #undef DEBUG_DISCONTIG
26 #ifdef DEBUG_DISCONTIG
27 #define DBGDCONT(args...) printk(args)
28 #else
29 #define DBGDCONT(args...)
30 #endif
31
32 #define PFN_UP(x) (((x) + PAGE_SIZE-1) >> PAGE_SHIFT)
33 #define PFN_DOWN(x) ((x) >> PAGE_SHIFT)
34 #define PFN_PHYS(x) ((x) << PAGE_SHIFT)
35 #define for_each_mem_cluster(memdesc, cluster, i) \
36 for ((cluster) = (memdesc)->cluster, (i) = 0; \
37 (i) < (memdesc)->numclusters; (i)++, (cluster)++)
38
39 static void __init show_mem_layout(void)
40 {
41 struct memclust_struct * cluster;
42 struct memdesc_struct * memdesc;
43 int i;
44
45 /* Find free clusters, and init and free the bootmem accordingly. */
46 memdesc = (struct memdesc_struct *)
47 (hwrpb->mddt_offset + (unsigned long) hwrpb);
48
49 printk("Raw memory layout:\n");
50 for_each_mem_cluster(memdesc, cluster, i) {
51 printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n",
52 i, cluster->usage, cluster->start_pfn,
53 cluster->start_pfn + cluster->numpages);
54 }
55 }
56
57 static void __init
58 setup_memory_node(int nid, void *kernel_end)
59 {
60 extern unsigned long mem_size_limit;
61 struct memclust_struct * cluster;
62 struct memdesc_struct * memdesc;
63 unsigned long start_kernel_pfn, end_kernel_pfn;
64 unsigned long bootmap_size, bootmap_pages, bootmap_start;
65 unsigned long start, end;
66 unsigned long node_pfn_start, node_pfn_end;
67 int i;
68 unsigned long node_datasz = PFN_UP(sizeof(plat_pg_data_t));
69 int show_init = 0;
70
71 /* Find the bounds of current node */
72 node_pfn_start = (nid * NODE_MAX_MEM_SIZE) >> PAGE_SHIFT;
73 node_pfn_end = node_pfn_start + (NODE_MAX_MEM_SIZE >> PAGE_SHIFT);
74
75 /* Find free clusters, and init and free the bootmem accordingly. */
76 memdesc = (struct memdesc_struct *)
77 (hwrpb->mddt_offset + (unsigned long) hwrpb);
78
79 /* find the bounds of this node (min_low_pfn/max_low_pfn) */
80 min_low_pfn = ~0UL;
81 for_each_mem_cluster(memdesc, cluster, i) {
82 /* Bit 0 is console/PALcode reserved. Bit 1 is
83 non-volatile memory -- we might want to mark
84 this for later. */
85 if (cluster->usage & 3)
86 continue;
87
88 start = cluster->start_pfn;
89 end = start + cluster->numpages;
90
91 if (start >= node_pfn_end || end <= node_pfn_start)
92 continue;
93
94 if (!show_init) {
95 show_init = 1;
96 printk("Initialing bootmem allocator on Node ID %d\n", nid);
97 }
98 printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n",
99 i, cluster->usage, cluster->start_pfn,
100 cluster->start_pfn + cluster->numpages);
101
102 if (start < node_pfn_start)
103 start = node_pfn_start;
104 if (end > node_pfn_end)
105 end = node_pfn_end;
106
107 if (start < min_low_pfn)
108 min_low_pfn = start;
109 if (end > max_low_pfn)
110 max_low_pfn = end;
111 }
112
113 if (mem_size_limit && max_low_pfn >= mem_size_limit) {
114 printk("setup: forcing memory size to %ldK (from %ldK).\n",
115 mem_size_limit << (PAGE_SHIFT - 10),
116 max_low_pfn << (PAGE_SHIFT - 10));
117 max_low_pfn = mem_size_limit;
118 }
119
120 if (min_low_pfn >= max_low_pfn)
121 return;
122
123 num_physpages += max_low_pfn - min_low_pfn;
124
125 /* Cute trick to make sure our local node data is on local memory */
126 PLAT_NODE_DATA(nid) = (plat_pg_data_t *)(__va(min_low_pfn << PAGE_SHIFT));
127 /* Quasi-mark the plat_pg_data_t as in-use */
128 min_low_pfn += node_datasz;
129 if (min_low_pfn >= max_low_pfn) {
130 printk(" not enough mem to reserve PLAT_NODE_DATA");
131 return;
132 }
133 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
134
135 printk(" Detected node memory: start %8lu, end %8lu\n",
136 min_low_pfn, max_low_pfn);
137
138 DBGDCONT(" DISCONTIG: plat_node_data[%d] is at 0x%p\n", nid, PLAT_NODE_DATA(nid));
139 DBGDCONT(" DISCONTIG: NODE_DATA(%d)->bdata is at 0x%p\n", nid, NODE_DATA(nid)->bdata);
140
141 /* Find the bounds of kernel memory. */
142 start_kernel_pfn = PFN_DOWN(KERNEL_START_PHYS);
143 end_kernel_pfn = PFN_UP(virt_to_phys(kernel_end));
144 bootmap_start = -1;
145
146 if (!nid && (max_low_pfn < end_kernel_pfn || min_low_pfn > start_kernel_pfn))
147 panic("kernel loaded out of ram");
148
149 /* Zone start phys-addr must be 2^(MAX_ORDER-1) aligned */
150 min_low_pfn = (min_low_pfn + ((1UL << (MAX_ORDER-1))-1)) & ~((1UL << (MAX_ORDER-1))-1);
151
152 /* We need to know how many physically contiguous pages
153 we'll need for the bootmap. */
154 bootmap_pages = bootmem_bootmap_pages(max_low_pfn-min_low_pfn);
155
156 /* Now find a good region where to allocate the bootmap. */
157 for_each_mem_cluster(memdesc, cluster, i) {
158 if (cluster->usage & 3)
159 continue;
160
161 start = cluster->start_pfn;
162 end = start + cluster->numpages;
163
164 if (start >= max_low_pfn || end <= min_low_pfn)
165 continue;
166
167 if (end > max_low_pfn)
168 end = max_low_pfn;
169 if (start < min_low_pfn)
170 start = min_low_pfn;
171
172 if (start < start_kernel_pfn) {
173 if (end > end_kernel_pfn
174 && end - end_kernel_pfn >= bootmap_pages) {
175 bootmap_start = end_kernel_pfn;
176 break;
177 } else if (end > start_kernel_pfn)
178 end = start_kernel_pfn;
179 } else if (start < end_kernel_pfn)
180 start = end_kernel_pfn;
181 if (end - start >= bootmap_pages) {
182 bootmap_start = start;
183 break;
184 }
185 }
186
187 if (bootmap_start == -1)
188 panic("couldn't find a contigous place for the bootmap");
189
190 /* Allocate the bootmap and mark the whole MM as reserved. */
191 bootmap_size = init_bootmem_node(NODE_DATA(nid), bootmap_start,
192 min_low_pfn, max_low_pfn);
193 DBGDCONT(" bootmap_start %lu, bootmap_size %lu, bootmap_pages %lu\n",
194 bootmap_start, bootmap_size, bootmap_pages);
195
196 /* Mark the free regions. */
197 for_each_mem_cluster(memdesc, cluster, i) {
198 if (cluster->usage & 3)
199 continue;
200
201 start = cluster->start_pfn;
202 end = cluster->start_pfn + cluster->numpages;
203
204 if (start >= max_low_pfn || end <= min_low_pfn)
205 continue;
206
207 if (end > max_low_pfn)
208 end = max_low_pfn;
209 if (start < min_low_pfn)
210 start = min_low_pfn;
211
212 if (start < start_kernel_pfn) {
213 if (end > end_kernel_pfn) {
214 free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start),
215 (PFN_PHYS(start_kernel_pfn)
216 - PFN_PHYS(start)));
217 printk(" freeing pages %ld:%ld\n",
218 start, start_kernel_pfn);
219 start = end_kernel_pfn;
220 } else if (end > start_kernel_pfn)
221 end = start_kernel_pfn;
222 } else if (start < end_kernel_pfn)
223 start = end_kernel_pfn;
224 if (start >= end)
225 continue;
226
227 free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start), PFN_PHYS(end) - PFN_PHYS(start));
228 printk(" freeing pages %ld:%ld\n", start, end);
229 }
230
231 /* Reserve the bootmap memory. */
232 reserve_bootmem_node(NODE_DATA(nid), PFN_PHYS(bootmap_start), bootmap_size);
233 printk(" reserving pages %ld:%ld\n", bootmap_start, bootmap_start+PFN_UP(bootmap_size));
234
235 numnodes++;
236 }
237
238 void __init
239 setup_memory(void *kernel_end)
240 {
241 int nid;
242
243 show_mem_layout();
244
245 numnodes = 0;
246 for (nid = 0; nid < MAX_NUMNODES; nid++)
247 setup_memory_node(nid, kernel_end);
248
249 #ifdef CONFIG_BLK_DEV_INITRD
250 initrd_start = INITRD_START;
251 if (initrd_start) {
252 initrd_end = initrd_start+INITRD_SIZE;
253 printk("Initial ramdisk at: 0x%p (%lu bytes)\n",
254 (void *) initrd_start, INITRD_SIZE);
255
256 if ((void *)initrd_end > phys_to_virt(PFN_PHYS(max_low_pfn))) {
257 printk("initrd extends beyond end of memory "
258 "(0x%08lx > 0x%p)\ndisabling initrd\n",
259 initrd_end,
260 phys_to_virt(PFN_PHYS(max_low_pfn)));
261 initrd_start = initrd_end = 0;
262 } else {
263 reserve_bootmem_node(NODE_DATA(KVADDR_TO_NID(initrd_start)),
264 virt_to_phys((void *)initrd_start),
265 INITRD_SIZE);
266 }
267 }
268 #endif /* CONFIG_BLK_DEV_INITRD */
269 }
270
271 void __init paging_init(void)
272 {
273 unsigned int nid;
274 unsigned long zones_size[MAX_NR_ZONES] = {0, };
275 unsigned long dma_local_pfn;
276
277 /*
278 * The old global MAX_DMA_ADDRESS per-arch API doesn't fit
279 * in the NUMA model, for now we convert it to a pfn and
280 * we interpret this pfn as a local per-node information.
281 * This issue isn't very important since none of these machines
282 * have legacy ISA slots anyways.
283 */
284 dma_local_pfn = virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
285
286 for (nid = 0; nid < numnodes; nid++) {
287 unsigned long start_pfn = plat_node_bdata[nid].node_boot_start >> PAGE_SHIFT;
288 unsigned long end_pfn = plat_node_bdata[nid].node_low_pfn;
289 unsigned long lmax_mapnr;
290
291 if (dma_local_pfn >= end_pfn - start_pfn)
292 zones_size[ZONE_DMA] = end_pfn - start_pfn;
293 else {
294 zones_size[ZONE_DMA] = dma_local_pfn;
295 zones_size[ZONE_NORMAL] = (end_pfn - start_pfn) - dma_local_pfn;
296 }
297 free_area_init_node(nid, NODE_DATA(nid), NULL, zones_size, start_pfn<<PAGE_SHIFT, NULL);
298 lmax_mapnr = PLAT_NODE_DATA_STARTNR(nid) + PLAT_NODE_DATA_SIZE(nid);
299 if (lmax_mapnr > max_mapnr) {
300 max_mapnr = lmax_mapnr;
301 DBGDCONT("Grow max_mapnr to %ld\n", max_mapnr);
302 }
303 }
304
305 /* Initialize the kernel's ZERO_PGE. */
306 memset((void *)ZERO_PGE, 0, PAGE_SIZE);
307 }
308
309 #define printkdot() \
310 do { \
311 if (!(i++ % ((100UL*1024*1024)>>PAGE_SHIFT))) \
312 printk("."); \
313 } while(0)
314
315 #define clobber(p, size) memset((p)->virtual, 0xaa, (size))
316
317 void __init mem_stress(void)
318 {
319 LIST_HEAD(x);
320 LIST_HEAD(xx);
321 struct page * p;
322 unsigned long i = 0;
323
324 printk("starting memstress");
325 while ((p = alloc_pages(GFP_ATOMIC, 1))) {
326 clobber(p, PAGE_SIZE*2);
327 list_add(&p->list, &x);
328 printkdot();
329 }
330 while ((p = alloc_page(GFP_ATOMIC))) {
331 clobber(p, PAGE_SIZE);
332 list_add(&p->list, &xx);
333 printkdot();
334 }
335 while (!list_empty(&x)) {
336 p = list_entry(x.next, struct page, list);
337 clobber(p, PAGE_SIZE*2);
338 list_del(x.next);
339 __free_pages(p, 1);
340 printkdot();
341 }
342 while (!list_empty(&xx)) {
343 p = list_entry(xx.next, struct page, list);
344 clobber(p, PAGE_SIZE);
345 list_del(xx.next);
346 __free_pages(p, 0);
347 printkdot();
348 }
349 printk("I'm still alive duh!\n");
350 }
351
352 #undef printkdot
353 #undef clobber
354
355 void __init mem_init(void)
356 {
357 unsigned long codesize, reservedpages, datasize, initsize, pfn;
358 extern int page_is_ram(unsigned long) __init;
359 extern char _text, _etext, _data, _edata;
360 extern char __init_begin, __init_end;
361 extern unsigned long totalram_pages;
362 unsigned long nid, i;
363 mem_map_t * lmem_map;
364
365 high_memory = (void *) __va(max_mapnr <<PAGE_SHIFT);
366
367 reservedpages = 0;
368 for (nid = 0; nid < numnodes; nid++) {
369 /*
370 * This will free up the bootmem, ie, slot 0 memory
371 */
372 totalram_pages += free_all_bootmem_node(NODE_DATA(nid));
373
374 lmem_map = NODE_MEM_MAP(nid);
375 pfn = NODE_DATA(nid)->node_start_paddr >> PAGE_SHIFT;
376 for (i = 0; i < PLAT_NODE_DATA_SIZE(nid); i++, pfn++)
377 if (page_is_ram(pfn) && PageReserved(lmem_map+i))
378 reservedpages++;
379 }
380
381 codesize = (unsigned long) &_etext - (unsigned long) &_text;
382 datasize = (unsigned long) &_edata - (unsigned long) &_data;
383 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin;
384
385 printk("Memory: %luk/%luk available (%luk kernel code, %luk reserved, "
386 "%luk data, %luk init)\n",
387 nr_free_pages() << (PAGE_SHIFT-10),
388 num_physpages << (PAGE_SHIFT-10),
389 codesize >> 10,
390 reservedpages << (PAGE_SHIFT-10),
391 datasize >> 10,
392 initsize >> 10);
393 #if 0
394 mem_stress();
395 #endif
396 }
397
398 void
399 show_mem(void)
400 {
401 long i,free = 0,total = 0,reserved = 0;
402 long shared = 0, cached = 0;
403 int nid;
404
405 printk("\nMem-info:\n");
406 show_free_areas();
407 printk("Free swap: %6dkB\n",nr_swap_pages<<(PAGE_SHIFT-10));
408 for (nid = 0; nid < numnodes; nid++) {
409 mem_map_t * lmem_map = NODE_MEM_MAP(nid);
410 i = PLAT_NODE_DATA_SIZE(nid);
411 while (i-- > 0) {
412 total++;
413 if (PageReserved(lmem_map+i))
414 reserved++;
415 else if (PageSwapCache(lmem_map+i))
416 cached++;
417 else if (!page_count(lmem_map+i))
418 free++;
419 else
420 shared += atomic_read(&lmem_map[i].count) - 1;
421 }
422 }
423 printk("%ld pages of RAM\n",total);
424 printk("%ld free pages\n",free);
425 printk("%ld reserved pages\n",reserved);
426 printk("%ld pages shared\n",shared);
427 printk("%ld pages swap cached\n",cached);
428 printk("%ld pages in page table cache\n",pgtable_cache_size);
429 show_buffers();
430 }
431