Merge master.kernel.org:/pub/scm/linux/kernel/git/jejb/scsi-rc-fixes-2.6
[linux-2.6] / 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/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>
15 #include <linux/pfn.h>
16 #include <linux/module.h>
17
18 #include <asm/hwrpb.h>
19 #include <asm/pgalloc.h>
20
21 pg_data_t node_data[MAX_NUMNODES];
22 bootmem_data_t node_bdata[MAX_NUMNODES];
23 EXPORT_SYMBOL(node_data);
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 for_each_mem_cluster(memdesc, cluster, i)               \
33         for ((cluster) = (memdesc)->cluster, (i) = 0;           \
34              (i) < (memdesc)->numclusters; (i)++, (cluster)++)
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
132         num_physpages += node_max_pfn - node_min_pfn;
133
134 #if 0 /* we'll try this one again in a little while */
135         /* Cute trick to make sure our local node data is on local memory */
136         node_data[nid] = (pg_data_t *)(__va(node_min_pfn << PAGE_SHIFT));
137 #endif
138         /* Quasi-mark the pg_data_t as in-use */
139         node_min_pfn += node_datasz;
140         if (node_min_pfn >= node_max_pfn) {
141                 printk(" not enough mem to reserve NODE_DATA");
142                 return;
143         }
144         NODE_DATA(nid)->bdata = &node_bdata[nid];
145
146         printk(" Detected node memory:   start %8lu, end %8lu\n",
147                node_min_pfn, node_max_pfn);
148
149         DBGDCONT(" DISCONTIG: node_data[%d]   is at 0x%p\n", nid, NODE_DATA(nid));
150         DBGDCONT(" DISCONTIG: NODE_DATA(%d)->bdata is at 0x%p\n", nid, NODE_DATA(nid)->bdata);
151
152         /* Find the bounds of kernel memory.  */
153         start_kernel_pfn = PFN_DOWN(KERNEL_START_PHYS);
154         end_kernel_pfn = PFN_UP(virt_to_phys(kernel_end));
155         bootmap_start = -1;
156
157         if (!nid && (node_max_pfn < end_kernel_pfn || node_min_pfn > start_kernel_pfn))
158                 panic("kernel loaded out of ram");
159
160         /* Zone start phys-addr must be 2^(MAX_ORDER-1) aligned.
161            Note that we round this down, not up - node memory
162            has much larger alignment than 8Mb, so it's safe. */
163         node_min_pfn &= ~((1UL << (MAX_ORDER-1))-1);
164
165         /* We need to know how many physically contiguous pages
166            we'll need for the bootmap.  */
167         bootmap_pages = bootmem_bootmap_pages(node_max_pfn-node_min_pfn);
168
169         /* Now find a good region where to allocate the bootmap.  */
170         for_each_mem_cluster(memdesc, cluster, i) {
171                 if (cluster->usage & 3)
172                         continue;
173
174                 start = cluster->start_pfn;
175                 end = start + cluster->numpages;
176
177                 if (start >= node_max_pfn || end <= node_min_pfn)
178                         continue;
179
180                 if (end > node_max_pfn)
181                         end = node_max_pfn;
182                 if (start < node_min_pfn)
183                         start = node_min_pfn;
184
185                 if (start < start_kernel_pfn) {
186                         if (end > end_kernel_pfn
187                             && end - end_kernel_pfn >= bootmap_pages) {
188                                 bootmap_start = end_kernel_pfn;
189                                 break;
190                         } else if (end > start_kernel_pfn)
191                                 end = start_kernel_pfn;
192                 } else if (start < end_kernel_pfn)
193                         start = end_kernel_pfn;
194                 if (end - start >= bootmap_pages) {
195                         bootmap_start = start;
196                         break;
197                 }
198         }
199
200         if (bootmap_start == -1)
201                 panic("couldn't find a contigous place for the bootmap");
202
203         /* Allocate the bootmap and mark the whole MM as reserved.  */
204         bootmap_size = init_bootmem_node(NODE_DATA(nid), bootmap_start,
205                                          node_min_pfn, node_max_pfn);
206         DBGDCONT(" bootmap_start %lu, bootmap_size %lu, bootmap_pages %lu\n",
207                  bootmap_start, bootmap_size, bootmap_pages);
208
209         /* Mark the free regions.  */
210         for_each_mem_cluster(memdesc, cluster, i) {
211                 if (cluster->usage & 3)
212                         continue;
213
214                 start = cluster->start_pfn;
215                 end = cluster->start_pfn + cluster->numpages;
216
217                 if (start >= node_max_pfn || end <= node_min_pfn)
218                         continue;
219
220                 if (end > node_max_pfn)
221                         end = node_max_pfn;
222                 if (start < node_min_pfn)
223                         start = node_min_pfn;
224
225                 if (start < start_kernel_pfn) {
226                         if (end > end_kernel_pfn) {
227                                 free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start),
228                                              (PFN_PHYS(start_kernel_pfn)
229                                               - PFN_PHYS(start)));
230                                 printk(" freeing pages %ld:%ld\n",
231                                        start, start_kernel_pfn);
232                                 start = end_kernel_pfn;
233                         } else if (end > start_kernel_pfn)
234                                 end = start_kernel_pfn;
235                 } else if (start < end_kernel_pfn)
236                         start = end_kernel_pfn;
237                 if (start >= end)
238                         continue;
239
240                 free_bootmem_node(NODE_DATA(nid), PFN_PHYS(start), PFN_PHYS(end) - PFN_PHYS(start));
241                 printk(" freeing pages %ld:%ld\n", start, end);
242         }
243
244         /* Reserve the bootmap memory.  */
245         reserve_bootmem_node(NODE_DATA(nid), PFN_PHYS(bootmap_start), bootmap_size);
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),
284                                              INITRD_SIZE);
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) {
306                 unsigned long start_pfn = node_bdata[nid].node_boot_start >> PAGE_SHIFT;
307                 unsigned long end_pfn = node_bdata[nid].node_low_pfn;
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                 }
315                 free_area_init_node(nid, NODE_DATA(nid), zones_size, start_pfn, NULL);
316         }
317
318         /* Initialize the kernel's ZERO_PGE. */
319         memset((void *)ZERO_PGE, 0, PAGE_SIZE);
320 }
321
322 void __init mem_init(void)
323 {
324         unsigned long codesize, reservedpages, datasize, initsize, pfn;
325         extern int page_is_ram(unsigned long) __init;
326         extern char _text, _etext, _data, _edata;
327         extern char __init_begin, __init_end;
328         unsigned long nid, i;
329         high_memory = (void *) __va(max_low_pfn << PAGE_SHIFT);
330
331         reservedpages = 0;
332         for_each_online_node(nid) {
333                 /*
334                  * This will free up the bootmem, ie, slot 0 memory
335                  */
336                 totalram_pages += free_all_bootmem_node(NODE_DATA(nid));
337
338                 pfn = NODE_DATA(nid)->node_start_pfn;
339                 for (i = 0; i < node_spanned_pages(nid); i++, pfn++)
340                         if (page_is_ram(pfn) &&
341                             PageReserved(nid_page_nr(nid, i)))
342                                 reservedpages++;
343         }
344
345         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
346         datasize =  (unsigned long) &_edata - (unsigned long) &_data;
347         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
348
349         printk("Memory: %luk/%luk available (%luk kernel code, %luk reserved, "
350                "%luk data, %luk init)\n",
351                (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
352                num_physpages << (PAGE_SHIFT-10),
353                codesize >> 10,
354                reservedpages << (PAGE_SHIFT-10),
355                datasize >> 10,
356                initsize >> 10);
357 #if 0
358         mem_stress();
359 #endif
360 }
361
362 void
363 show_mem(void)
364 {
365         long i,free = 0,total = 0,reserved = 0;
366         long shared = 0, cached = 0;
367         int nid;
368
369         printk("\nMem-info:\n");
370         show_free_areas();
371         printk("Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
372         for_each_online_node(nid) {
373                 unsigned long flags;
374                 pgdat_resize_lock(NODE_DATA(nid), &flags);
375                 i = node_spanned_pages(nid);
376                 while (i-- > 0) {
377                         struct page *page = nid_page_nr(nid, i);
378                         total++;
379                         if (PageReserved(page))
380                                 reserved++;
381                         else if (PageSwapCache(page))
382                                 cached++;
383                         else if (!page_count(page))
384                                 free++;
385                         else
386                                 shared += page_count(page) - 1;
387                 }
388                 pgdat_resize_unlock(NODE_DATA(nid), &flags);
389         }
390         printk("%ld pages of RAM\n",total);
391         printk("%ld free pages\n",free);
392         printk("%ld reserved pages\n",reserved);
393         printk("%ld pages shared\n",shared);
394         printk("%ld pages swap cached\n",cached);
395 }