[IA64] Percpu quicklist for combined allocator for pgd/pmd/pte.
[linux-2.6.git] / arch / ia64 / mm / init.c
1 /*
2  * Initialize MMU support.
3  *
4  * Copyright (C) 1998-2003 Hewlett-Packard Co
5  *      David Mosberger-Tang <davidm@hpl.hp.com>
6  */
7 #include <linux/config.h>
8 #include <linux/kernel.h>
9 #include <linux/init.h>
10
11 #include <linux/bootmem.h>
12 #include <linux/efi.h>
13 #include <linux/elf.h>
14 #include <linux/mm.h>
15 #include <linux/mmzone.h>
16 #include <linux/module.h>
17 #include <linux/personality.h>
18 #include <linux/reboot.h>
19 #include <linux/slab.h>
20 #include <linux/swap.h>
21 #include <linux/proc_fs.h>
22 #include <linux/bitops.h>
23
24 #include <asm/a.out.h>
25 #include <asm/dma.h>
26 #include <asm/ia32.h>
27 #include <asm/io.h>
28 #include <asm/machvec.h>
29 #include <asm/numa.h>
30 #include <asm/patch.h>
31 #include <asm/pgalloc.h>
32 #include <asm/sal.h>
33 #include <asm/sections.h>
34 #include <asm/system.h>
35 #include <asm/tlb.h>
36 #include <asm/uaccess.h>
37 #include <asm/unistd.h>
38 #include <asm/mca.h>
39
40 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
41
42 DEFINE_PER_CPU(unsigned long *, __pgtable_quicklist);
43 DEFINE_PER_CPU(long, __pgtable_quicklist_size);
44
45 extern void ia64_tlb_init (void);
46
47 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL;
48
49 #ifdef CONFIG_VIRTUAL_MEM_MAP
50 unsigned long vmalloc_end = VMALLOC_END_INIT;
51 EXPORT_SYMBOL(vmalloc_end);
52 struct page *vmem_map;
53 EXPORT_SYMBOL(vmem_map);
54 #endif
55
56 struct page *zero_page_memmap_ptr;      /* map entry for zero page */
57 EXPORT_SYMBOL(zero_page_memmap_ptr);
58
59 #define MIN_PGT_PAGES                   25UL
60 #define MAX_PGT_FREES_PER_PASS          16
61 #define PGT_FRACTION_OF_NODE_MEM        16
62
63 static inline long
64 max_pgt_pages(void)
65 {
66         u64 node_free_pages, max_pgt_pages;
67
68 #ifndef CONFIG_NUMA
69         node_free_pages = nr_free_pages();
70 #else
71         node_free_pages = nr_free_pages_pgdat(NODE_DATA(numa_node_id()));
72 #endif
73         max_pgt_pages = node_free_pages / PGT_FRACTION_OF_NODE_MEM;
74         max_pgt_pages = max(max_pgt_pages, MIN_PGT_PAGES);
75         return max_pgt_pages;
76 }
77
78 static inline long
79 min_pages_to_free(void)
80 {
81         long pages_to_free;
82
83         pages_to_free = pgtable_quicklist_size - max_pgt_pages();
84         pages_to_free = min(pages_to_free, MAX_PGT_FREES_PER_PASS);
85         return pages_to_free;
86 }
87
88 void
89 check_pgt_cache(void)
90 {
91         long pages_to_free;
92
93         if (unlikely(pgtable_quicklist_size <= MIN_PGT_PAGES))
94                 return;
95
96         preempt_disable();
97         while (unlikely((pages_to_free = min_pages_to_free()) > 0)) {
98                 while (pages_to_free--) {
99                         free_page((unsigned long)pgtable_quicklist_alloc());
100                 }
101                 preempt_enable();
102                 preempt_disable();
103         }
104         preempt_enable();
105 }
106
107 void
108 lazy_mmu_prot_update (pte_t pte)
109 {
110         unsigned long addr;
111         struct page *page;
112
113         if (!pte_exec(pte))
114                 return;                         /* not an executable page... */
115
116         page = pte_page(pte);
117         addr = (unsigned long) page_address(page);
118
119         if (test_bit(PG_arch_1, &page->flags))
120                 return;                         /* i-cache is already coherent with d-cache */
121
122         flush_icache_range(addr, addr + PAGE_SIZE);
123         set_bit(PG_arch_1, &page->flags);       /* mark page as clean */
124 }
125
126 inline void
127 ia64_set_rbs_bot (void)
128 {
129         unsigned long stack_size = current->signal->rlim[RLIMIT_STACK].rlim_max & -16;
130
131         if (stack_size > MAX_USER_STACK_SIZE)
132                 stack_size = MAX_USER_STACK_SIZE;
133         current->thread.rbs_bot = STACK_TOP - stack_size;
134 }
135
136 /*
137  * This performs some platform-dependent address space initialization.
138  * On IA-64, we want to setup the VM area for the register backing
139  * store (which grows upwards) and install the gateway page which is
140  * used for signal trampolines, etc.
141  */
142 void
143 ia64_init_addr_space (void)
144 {
145         struct vm_area_struct *vma;
146
147         ia64_set_rbs_bot();
148
149         /*
150          * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
151          * the problem.  When the process attempts to write to the register backing store
152          * for the first time, it will get a SEGFAULT in this case.
153          */
154         vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
155         if (vma) {
156                 memset(vma, 0, sizeof(*vma));
157                 vma->vm_mm = current->mm;
158                 vma->vm_start = current->thread.rbs_bot & PAGE_MASK;
159                 vma->vm_end = vma->vm_start + PAGE_SIZE;
160                 vma->vm_page_prot = protection_map[VM_DATA_DEFAULT_FLAGS & 0x7];
161                 vma->vm_flags = VM_DATA_DEFAULT_FLAGS | VM_GROWSUP;
162                 down_write(&current->mm->mmap_sem);
163                 if (insert_vm_struct(current->mm, vma)) {
164                         up_write(&current->mm->mmap_sem);
165                         kmem_cache_free(vm_area_cachep, vma);
166                         return;
167                 }
168                 up_write(&current->mm->mmap_sem);
169         }
170
171         /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
172         if (!(current->personality & MMAP_PAGE_ZERO)) {
173                 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
174                 if (vma) {
175                         memset(vma, 0, sizeof(*vma));
176                         vma->vm_mm = current->mm;
177                         vma->vm_end = PAGE_SIZE;
178                         vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
179                         vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO | VM_RESERVED;
180                         down_write(&current->mm->mmap_sem);
181                         if (insert_vm_struct(current->mm, vma)) {
182                                 up_write(&current->mm->mmap_sem);
183                                 kmem_cache_free(vm_area_cachep, vma);
184                                 return;
185                         }
186                         up_write(&current->mm->mmap_sem);
187                 }
188         }
189 }
190
191 void
192 free_initmem (void)
193 {
194         unsigned long addr, eaddr;
195
196         addr = (unsigned long) ia64_imva(__init_begin);
197         eaddr = (unsigned long) ia64_imva(__init_end);
198         while (addr < eaddr) {
199                 ClearPageReserved(virt_to_page(addr));
200                 set_page_count(virt_to_page(addr), 1);
201                 free_page(addr);
202                 ++totalram_pages;
203                 addr += PAGE_SIZE;
204         }
205         printk(KERN_INFO "Freeing unused kernel memory: %ldkB freed\n",
206                (__init_end - __init_begin) >> 10);
207 }
208
209 void
210 free_initrd_mem (unsigned long start, unsigned long end)
211 {
212         struct page *page;
213         /*
214          * EFI uses 4KB pages while the kernel can use 4KB or bigger.
215          * Thus EFI and the kernel may have different page sizes. It is
216          * therefore possible to have the initrd share the same page as
217          * the end of the kernel (given current setup).
218          *
219          * To avoid freeing/using the wrong page (kernel sized) we:
220          *      - align up the beginning of initrd
221          *      - align down the end of initrd
222          *
223          *  |             |
224          *  |=============| a000
225          *  |             |
226          *  |             |
227          *  |             | 9000
228          *  |/////////////|
229          *  |/////////////|
230          *  |=============| 8000
231          *  |///INITRD////|
232          *  |/////////////|
233          *  |/////////////| 7000
234          *  |             |
235          *  |KKKKKKKKKKKKK|
236          *  |=============| 6000
237          *  |KKKKKKKKKKKKK|
238          *  |KKKKKKKKKKKKK|
239          *  K=kernel using 8KB pages
240          *
241          * In this example, we must free page 8000 ONLY. So we must align up
242          * initrd_start and keep initrd_end as is.
243          */
244         start = PAGE_ALIGN(start);
245         end = end & PAGE_MASK;
246
247         if (start < end)
248                 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
249
250         for (; start < end; start += PAGE_SIZE) {
251                 if (!virt_addr_valid(start))
252                         continue;
253                 page = virt_to_page(start);
254                 ClearPageReserved(page);
255                 set_page_count(page, 1);
256                 free_page(start);
257                 ++totalram_pages;
258         }
259 }
260
261 /*
262  * This installs a clean page in the kernel's page table.
263  */
264 struct page *
265 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot)
266 {
267         pgd_t *pgd;
268         pud_t *pud;
269         pmd_t *pmd;
270         pte_t *pte;
271
272         if (!PageReserved(page))
273                 printk(KERN_ERR "put_kernel_page: page at 0x%p not in reserved memory\n",
274                        page_address(page));
275
276         pgd = pgd_offset_k(address);            /* note: this is NOT pgd_offset()! */
277
278         spin_lock(&init_mm.page_table_lock);
279         {
280                 pud = pud_alloc(&init_mm, pgd, address);
281                 if (!pud)
282                         goto out;
283
284                 pmd = pmd_alloc(&init_mm, pud, address);
285                 if (!pmd)
286                         goto out;
287                 pte = pte_alloc_map(&init_mm, pmd, address);
288                 if (!pte)
289                         goto out;
290                 if (!pte_none(*pte)) {
291                         pte_unmap(pte);
292                         goto out;
293                 }
294                 set_pte(pte, mk_pte(page, pgprot));
295                 pte_unmap(pte);
296         }
297   out:  spin_unlock(&init_mm.page_table_lock);
298         /* no need for flush_tlb */
299         return page;
300 }
301
302 static void
303 setup_gate (void)
304 {
305         struct page *page;
306
307         /*
308          * Map the gate page twice: once read-only to export the ELF headers etc. and once
309          * execute-only page to enable privilege-promotion via "epc":
310          */
311         page = virt_to_page(ia64_imva(__start_gate_section));
312         put_kernel_page(page, GATE_ADDR, PAGE_READONLY);
313 #ifdef HAVE_BUGGY_SEGREL
314         page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE));
315         put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE);
316 #else
317         put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE);
318 #endif
319         ia64_patch_gate();
320 }
321
322 void __devinit
323 ia64_mmu_init (void *my_cpu_data)
324 {
325         unsigned long psr, pta, impl_va_bits;
326         extern void __devinit tlb_init (void);
327
328 #ifdef CONFIG_DISABLE_VHPT
329 #       define VHPT_ENABLE_BIT  0
330 #else
331 #       define VHPT_ENABLE_BIT  1
332 #endif
333
334         /* Pin mapping for percpu area into TLB */
335         psr = ia64_clear_ic();
336         ia64_itr(0x2, IA64_TR_PERCPU_DATA, PERCPU_ADDR,
337                  pte_val(pfn_pte(__pa(my_cpu_data) >> PAGE_SHIFT, PAGE_KERNEL)),
338                  PERCPU_PAGE_SHIFT);
339
340         ia64_set_psr(psr);
341         ia64_srlz_i();
342
343         /*
344          * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
345          * address space.  The IA-64 architecture guarantees that at least 50 bits of
346          * virtual address space are implemented but if we pick a large enough page size
347          * (e.g., 64KB), the mapped address space is big enough that it will overlap with
348          * VMLPT.  I assume that once we run on machines big enough to warrant 64KB pages,
349          * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
350          * problem in practice.  Alternatively, we could truncate the top of the mapped
351          * address space to not permit mappings that would overlap with the VMLPT.
352          * --davidm 00/12/06
353          */
354 #       define pte_bits                 3
355 #       define mapped_space_bits        (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
356         /*
357          * The virtual page table has to cover the entire implemented address space within
358          * a region even though not all of this space may be mappable.  The reason for
359          * this is that the Access bit and Dirty bit fault handlers perform
360          * non-speculative accesses to the virtual page table, so the address range of the
361          * virtual page table itself needs to be covered by virtual page table.
362          */
363 #       define vmlpt_bits               (impl_va_bits - PAGE_SHIFT + pte_bits)
364 #       define POW2(n)                  (1ULL << (n))
365
366         impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
367
368         if (impl_va_bits < 51 || impl_va_bits > 61)
369                 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
370
371         /* place the VMLPT at the end of each page-table mapped region: */
372         pta = POW2(61) - POW2(vmlpt_bits);
373
374         if (POW2(mapped_space_bits) >= pta)
375                 panic("mm/init: overlap between virtually mapped linear page table and "
376                       "mapped kernel space!");
377         /*
378          * Set the (virtually mapped linear) page table address.  Bit
379          * 8 selects between the short and long format, bits 2-7 the
380          * size of the table, and bit 0 whether the VHPT walker is
381          * enabled.
382          */
383         ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
384
385         ia64_tlb_init();
386
387 #ifdef  CONFIG_HUGETLB_PAGE
388         ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2);
389         ia64_srlz_d();
390 #endif
391 }
392
393 #ifdef CONFIG_VIRTUAL_MEM_MAP
394
395 int
396 create_mem_map_page_table (u64 start, u64 end, void *arg)
397 {
398         unsigned long address, start_page, end_page;
399         struct page *map_start, *map_end;
400         int node;
401         pgd_t *pgd;
402         pud_t *pud;
403         pmd_t *pmd;
404         pte_t *pte;
405
406         map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
407         map_end   = vmem_map + (__pa(end) >> PAGE_SHIFT);
408
409         start_page = (unsigned long) map_start & PAGE_MASK;
410         end_page = PAGE_ALIGN((unsigned long) map_end);
411         node = paddr_to_nid(__pa(start));
412
413         for (address = start_page; address < end_page; address += PAGE_SIZE) {
414                 pgd = pgd_offset_k(address);
415                 if (pgd_none(*pgd))
416                         pgd_populate(&init_mm, pgd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
417                 pud = pud_offset(pgd, address);
418
419                 if (pud_none(*pud))
420                         pud_populate(&init_mm, pud, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
421                 pmd = pmd_offset(pud, address);
422
423                 if (pmd_none(*pmd))
424                         pmd_populate_kernel(&init_mm, pmd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
425                 pte = pte_offset_kernel(pmd, address);
426
427                 if (pte_none(*pte))
428                         set_pte(pte, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE)) >> PAGE_SHIFT,
429                                              PAGE_KERNEL));
430         }
431         return 0;
432 }
433
434 struct memmap_init_callback_data {
435         struct page *start;
436         struct page *end;
437         int nid;
438         unsigned long zone;
439 };
440
441 static int
442 virtual_memmap_init (u64 start, u64 end, void *arg)
443 {
444         struct memmap_init_callback_data *args;
445         struct page *map_start, *map_end;
446
447         args = (struct memmap_init_callback_data *) arg;
448         map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
449         map_end   = vmem_map + (__pa(end) >> PAGE_SHIFT);
450
451         if (map_start < args->start)
452                 map_start = args->start;
453         if (map_end > args->end)
454                 map_end = args->end;
455
456         /*
457          * We have to initialize "out of bounds" struct page elements that fit completely
458          * on the same pages that were allocated for the "in bounds" elements because they
459          * may be referenced later (and found to be "reserved").
460          */
461         map_start -= ((unsigned long) map_start & (PAGE_SIZE - 1)) / sizeof(struct page);
462         map_end += ((PAGE_ALIGN((unsigned long) map_end) - (unsigned long) map_end)
463                     / sizeof(struct page));
464
465         if (map_start < map_end)
466                 memmap_init_zone((unsigned long)(map_end - map_start),
467                                  args->nid, args->zone, page_to_pfn(map_start));
468         return 0;
469 }
470
471 void
472 memmap_init (unsigned long size, int nid, unsigned long zone,
473              unsigned long start_pfn)
474 {
475         if (!vmem_map)
476                 memmap_init_zone(size, nid, zone, start_pfn);
477         else {
478                 struct page *start;
479                 struct memmap_init_callback_data args;
480
481                 start = pfn_to_page(start_pfn);
482                 args.start = start;
483                 args.end = start + size;
484                 args.nid = nid;
485                 args.zone = zone;
486
487                 efi_memmap_walk(virtual_memmap_init, &args);
488         }
489 }
490
491 int
492 ia64_pfn_valid (unsigned long pfn)
493 {
494         char byte;
495         struct page *pg = pfn_to_page(pfn);
496
497         return     (__get_user(byte, (char __user *) pg) == 0)
498                 && ((((u64)pg & PAGE_MASK) == (((u64)(pg + 1) - 1) & PAGE_MASK))
499                         || (__get_user(byte, (char __user *) (pg + 1) - 1) == 0));
500 }
501 EXPORT_SYMBOL(ia64_pfn_valid);
502
503 int
504 find_largest_hole (u64 start, u64 end, void *arg)
505 {
506         u64 *max_gap = arg;
507
508         static u64 last_end = PAGE_OFFSET;
509
510         /* NOTE: this algorithm assumes efi memmap table is ordered */
511
512         if (*max_gap < (start - last_end))
513                 *max_gap = start - last_end;
514         last_end = end;
515         return 0;
516 }
517 #endif /* CONFIG_VIRTUAL_MEM_MAP */
518
519 static int
520 count_reserved_pages (u64 start, u64 end, void *arg)
521 {
522         unsigned long num_reserved = 0;
523         unsigned long *count = arg;
524
525         for (; start < end; start += PAGE_SIZE)
526                 if (PageReserved(virt_to_page(start)))
527                         ++num_reserved;
528         *count += num_reserved;
529         return 0;
530 }
531
532 /*
533  * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
534  * system call handler.  When this option is in effect, all fsyscalls will end up bubbling
535  * down into the kernel and calling the normal (heavy-weight) syscall handler.  This is
536  * useful for performance testing, but conceivably could also come in handy for debugging
537  * purposes.
538  */
539
540 static int nolwsys;
541
542 static int __init
543 nolwsys_setup (char *s)
544 {
545         nolwsys = 1;
546         return 1;
547 }
548
549 __setup("nolwsys", nolwsys_setup);
550
551 void
552 mem_init (void)
553 {
554         long reserved_pages, codesize, datasize, initsize;
555         pg_data_t *pgdat;
556         int i;
557         static struct kcore_list kcore_mem, kcore_vmem, kcore_kernel;
558
559         BUG_ON(PTRS_PER_PGD * sizeof(pgd_t) != PAGE_SIZE);
560         BUG_ON(PTRS_PER_PMD * sizeof(pmd_t) != PAGE_SIZE);
561         BUG_ON(PTRS_PER_PTE * sizeof(pte_t) != PAGE_SIZE);
562
563 #ifdef CONFIG_PCI
564         /*
565          * This needs to be called _after_ the command line has been parsed but _before_
566          * any drivers that may need the PCI DMA interface are initialized or bootmem has
567          * been freed.
568          */
569         platform_dma_init();
570 #endif
571
572 #ifndef CONFIG_DISCONTIGMEM
573         if (!mem_map)
574                 BUG();
575         max_mapnr = max_low_pfn;
576 #endif
577
578         high_memory = __va(max_low_pfn * PAGE_SIZE);
579
580         kclist_add(&kcore_mem, __va(0), max_low_pfn * PAGE_SIZE);
581         kclist_add(&kcore_vmem, (void *)VMALLOC_START, VMALLOC_END-VMALLOC_START);
582         kclist_add(&kcore_kernel, _stext, _end - _stext);
583
584         for_each_pgdat(pgdat)
585                 totalram_pages += free_all_bootmem_node(pgdat);
586
587         reserved_pages = 0;
588         efi_memmap_walk(count_reserved_pages, &reserved_pages);
589
590         codesize =  (unsigned long) _etext - (unsigned long) _stext;
591         datasize =  (unsigned long) _edata - (unsigned long) _etext;
592         initsize =  (unsigned long) __init_end - (unsigned long) __init_begin;
593
594         printk(KERN_INFO "Memory: %luk/%luk available (%luk code, %luk reserved, "
595                "%luk data, %luk init)\n", (unsigned long) nr_free_pages() << (PAGE_SHIFT - 10),
596                num_physpages << (PAGE_SHIFT - 10), codesize >> 10,
597                reserved_pages << (PAGE_SHIFT - 10), datasize >> 10, initsize >> 10);
598
599
600         /*
601          * For fsyscall entrpoints with no light-weight handler, use the ordinary
602          * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
603          * code can tell them apart.
604          */
605         for (i = 0; i < NR_syscalls; ++i) {
606                 extern unsigned long fsyscall_table[NR_syscalls];
607                 extern unsigned long sys_call_table[NR_syscalls];
608
609                 if (!fsyscall_table[i] || nolwsys)
610                         fsyscall_table[i] = sys_call_table[i] | 1;
611         }
612         setup_gate();
613
614 #ifdef CONFIG_IA32_SUPPORT
615         ia32_mem_init();
616 #endif
617 }