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
show_mem_layout(void)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
setup_memory_node(int nid,void * kernel_end)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 unsigned long node_min_pfn, node_max_pfn;
68 int i;
69 unsigned long node_datasz = PFN_UP(sizeof(plat_pg_data_t));
70 int show_init = 0;
71
72 /* Find the bounds of current node */
73 node_pfn_start = (NODE_MEM_START(nid)) >> PAGE_SHIFT;
74 node_pfn_end = node_pfn_start + (NODE_MEM_SIZE(nid) >> PAGE_SHIFT);
75
76 /* Find free clusters, and init and free the bootmem accordingly. */
77 memdesc = (struct memdesc_struct *)
78 (hwrpb->mddt_offset + (unsigned long) hwrpb);
79
80 /* find the bounds of this node (node_min_pfn/node_max_pfn) */
81 node_min_pfn = ~0UL;
82 node_max_pfn = 0UL;
83 for_each_mem_cluster(memdesc, cluster, i) {
84 /* Bit 0 is console/PALcode reserved. Bit 1 is
85 non-volatile memory -- we might want to mark
86 this for later. */
87 if (cluster->usage & 3)
88 continue;
89
90 start = cluster->start_pfn;
91 end = start + cluster->numpages;
92
93 if (start >= node_pfn_end || end <= node_pfn_start)
94 continue;
95
96 if (!show_init) {
97 show_init = 1;
98 printk("Initialing bootmem allocator on Node ID %d\n", nid);
99 }
100 printk(" memcluster %2d, usage %1lx, start %8lu, end %8lu\n",
101 i, cluster->usage, cluster->start_pfn,
102 cluster->start_pfn + cluster->numpages);
103
104 if (start < node_pfn_start)
105 start = node_pfn_start;
106 if (end > node_pfn_end)
107 end = node_pfn_end;
108
109 if (start < node_min_pfn)
110 node_min_pfn = start;
111 if (end > node_max_pfn)
112 node_max_pfn = end;
113 }
114
115 if (mem_size_limit && node_max_pfn > mem_size_limit) {
116 static int msg_shown = 0;
117 if (!msg_shown) {
118 msg_shown = 1;
119 printk("setup: forcing memory size to %ldK (from %ldK).\n",
120 mem_size_limit << (PAGE_SHIFT - 10),
121 node_max_pfn << (PAGE_SHIFT - 10));
122 }
123 node_max_pfn = mem_size_limit;
124 }
125
126 if (node_min_pfn >= node_max_pfn)
127 return;
128
129 /* Update global {min,max}_low_pfn from node information. */
130 if (node_min_pfn < min_low_pfn)
131 min_low_pfn = node_min_pfn;
132 if (node_max_pfn > max_low_pfn)
133 max_low_pfn = node_max_pfn;
134
135 num_physpages += node_max_pfn - node_min_pfn;
136
137 /* Cute trick to make sure our local node data is on local memory */
138 PLAT_NODE_DATA(nid) = (plat_pg_data_t *)(__va(node_min_pfn << PAGE_SHIFT));
139 /* Quasi-mark the plat_pg_data_t as in-use */
140 node_min_pfn += node_datasz;
141 if (node_min_pfn >= node_max_pfn) {
142 printk(" not enough mem to reserve PLAT_NODE_DATA");
143 return;
144 }
145 NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
146
147 printk(" Detected node memory: start %8lu, end %8lu\n",
148 node_min_pfn, node_max_pfn);
149
150 DBGDCONT(" DISCONTIG: plat_node_data[%d] is at 0x%p\n", nid, PLAT_NODE_DATA(nid));
151 DBGDCONT(" DISCONTIG: NODE_DATA(%d)->bdata is at 0x%p\n", nid, NODE_DATA(nid)->bdata);
152
153 /* Find the bounds of kernel memory. */
154 start_kernel_pfn = PFN_DOWN(KERNEL_START_PHYS);
155 end_kernel_pfn = PFN_UP(virt_to_phys(kernel_end));
156 bootmap_start = -1;
157
158 if (!nid && (node_max_pfn < end_kernel_pfn || node_min_pfn > start_kernel_pfn))
159 panic("kernel loaded out of ram");
160
161 /* Zone start phys-addr must be 2^(MAX_ORDER-1) aligned */
162 node_min_pfn = (node_min_pfn + ((1UL << (MAX_ORDER-1))-1)) & ~((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. */
244 reserve_bootmem_node(NODE_DATA(nid), PFN_PHYS(bootmap_start), bootmap_size);
245 printk(" reserving pages %ld:%ld\n", bootmap_start, bootmap_start+PFN_UP(bootmap_size));
246
247 numnodes++;
248 }
249
250 void __init
setup_memory(void * kernel_end)251 setup_memory(void *kernel_end)
252 {
253 int nid;
254
255 show_mem_layout();
256
257 numnodes = 0;
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 reserve_bootmem_node(NODE_DATA(KVADDR_TO_NID(initrd_start)),
281 virt_to_phys((void *)initrd_start),
282 INITRD_SIZE);
283 }
284 }
285 #endif /* CONFIG_BLK_DEV_INITRD */
286 }
287
paging_init(void)288 void __init paging_init(void)
289 {
290 unsigned int nid;
291 unsigned long zones_size[MAX_NR_ZONES] = {0, };
292 unsigned long dma_local_pfn;
293
294 /*
295 * The old global MAX_DMA_ADDRESS per-arch API doesn't fit
296 * in the NUMA model, for now we convert it to a pfn and
297 * we interpret this pfn as a local per-node information.
298 * This issue isn't very important since none of these machines
299 * have legacy ISA slots anyways.
300 */
301 dma_local_pfn = virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
302
303 for (nid = 0; nid < numnodes; nid++) {
304 unsigned long start_pfn = plat_node_bdata[nid].node_boot_start >> PAGE_SHIFT;
305 unsigned long end_pfn = plat_node_bdata[nid].node_low_pfn;
306 unsigned long lmax_mapnr;
307
308 if (dma_local_pfn >= end_pfn - start_pfn)
309 zones_size[ZONE_DMA] = end_pfn - start_pfn;
310 else {
311 zones_size[ZONE_DMA] = dma_local_pfn;
312 zones_size[ZONE_NORMAL] = (end_pfn - start_pfn) - dma_local_pfn;
313 }
314 free_area_init_node(nid, NODE_DATA(nid), NULL, zones_size, start_pfn<<PAGE_SHIFT, NULL);
315 lmax_mapnr = PLAT_NODE_DATA_STARTNR(nid) + PLAT_NODE_DATA_SIZE(nid);
316 if (lmax_mapnr > max_mapnr) {
317 max_mapnr = lmax_mapnr;
318 DBGDCONT("Grow max_mapnr to %ld\n", max_mapnr);
319 }
320 }
321
322 /* Initialize the kernel's ZERO_PGE. */
323 memset((void *)ZERO_PGE, 0, PAGE_SIZE);
324 }
325
326 #define printkdot() \
327 do { \
328 if (!(i++ % ((100UL*1024*1024)>>PAGE_SHIFT))) \
329 printk("."); \
330 } while(0)
331
332 #define clobber(p, size) memset(page_address(p), 0xaa, (size))
333
mem_stress(void)334 void __init mem_stress(void)
335 {
336 LIST_HEAD(x);
337 LIST_HEAD(xx);
338 struct page * p;
339 unsigned long i = 0;
340
341 printk("starting memstress");
342 while ((p = alloc_pages(GFP_ATOMIC, 1))) {
343 clobber(p, PAGE_SIZE*2);
344 list_add(&p->list, &x);
345 printkdot();
346 }
347 while ((p = alloc_page(GFP_ATOMIC))) {
348 clobber(p, PAGE_SIZE);
349 list_add(&p->list, &xx);
350 printkdot();
351 }
352 while (!list_empty(&x)) {
353 p = list_entry(x.next, struct page, list);
354 clobber(p, PAGE_SIZE*2);
355 list_del(x.next);
356 __free_pages(p, 1);
357 printkdot();
358 }
359 while (!list_empty(&xx)) {
360 p = list_entry(xx.next, struct page, list);
361 clobber(p, PAGE_SIZE);
362 list_del(xx.next);
363 __free_pages(p, 0);
364 printkdot();
365 }
366 printk("I'm still alive duh!\n");
367 }
368
369 #undef printkdot
370 #undef clobber
371
mem_init(void)372 void __init mem_init(void)
373 {
374 unsigned long codesize, reservedpages, datasize, initsize, pfn;
375 extern int page_is_ram(unsigned long) __init;
376 extern char _text, _etext, _data, _edata;
377 extern char __init_begin, __init_end;
378 extern unsigned long totalram_pages;
379 unsigned long nid, i;
380 mem_map_t * lmem_map;
381
382 high_memory = (void *) __va(max_mapnr <<PAGE_SHIFT);
383
384 reservedpages = 0;
385 for (nid = 0; nid < numnodes; nid++) {
386 /*
387 * This will free up the bootmem, ie, slot 0 memory
388 */
389 totalram_pages += free_all_bootmem_node(NODE_DATA(nid));
390
391 lmem_map = NODE_MEM_MAP(nid);
392 pfn = NODE_DATA(nid)->node_start_paddr >> PAGE_SHIFT;
393 for (i = 0; i < PLAT_NODE_DATA_SIZE(nid); i++, pfn++)
394 if (page_is_ram(pfn) && PageReserved(lmem_map+i))
395 reservedpages++;
396 }
397
398 codesize = (unsigned long) &_etext - (unsigned long) &_text;
399 datasize = (unsigned long) &_edata - (unsigned long) &_data;
400 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin;
401
402 printk("Memory: %luk/%luk available (%luk kernel code, %luk reserved, "
403 "%luk data, %luk init)\n",
404 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
405 num_physpages << (PAGE_SHIFT-10),
406 codesize >> 10,
407 reservedpages << (PAGE_SHIFT-10),
408 datasize >> 10,
409 initsize >> 10);
410 #if 0
411 mem_stress();
412 #endif
413 }
414
415 void
show_mem(void)416 show_mem(void)
417 {
418 long i,free = 0,total = 0,reserved = 0;
419 long shared = 0, cached = 0;
420 int nid;
421
422 printk("\nMem-info:\n");
423 show_free_areas();
424 printk("Free swap: %6dkB\n",nr_swap_pages<<(PAGE_SHIFT-10));
425 for (nid = 0; nid < numnodes; nid++) {
426 mem_map_t * lmem_map = NODE_MEM_MAP(nid);
427 i = PLAT_NODE_DATA_SIZE(nid);
428 while (i-- > 0) {
429 total++;
430 if (PageReserved(lmem_map+i))
431 reserved++;
432 else if (PageSwapCache(lmem_map+i))
433 cached++;
434 else if (!page_count(lmem_map+i))
435 free++;
436 else
437 shared += atomic_read(&lmem_map[i].count) - 1;
438 }
439 }
440 printk("%ld pages of RAM\n",total);
441 printk("%ld free pages\n",free);
442 printk("%ld reserved pages\n",reserved);
443 printk("%ld pages shared\n",shared);
444 printk("%ld pages swap cached\n",cached);
445 printk("%ld pages in page table cache\n",pgtable_cache_size);
446 show_buffers();
447 }
448