Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * arch/sparc64/mm/init.c
4 *
5 * Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
6 * Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
7 */
8
9#include <linux/extable.h>
10#include <linux/kernel.h>
11#include <linux/sched.h>
12#include <linux/string.h>
13#include <linux/init.h>
14#include <linux/memblock.h>
15#include <linux/mm.h>
16#include <linux/hugetlb.h>
17#include <linux/initrd.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/poison.h>
21#include <linux/fs.h>
22#include <linux/seq_file.h>
23#include <linux/kprobes.h>
24#include <linux/cache.h>
25#include <linux/sort.h>
26#include <linux/ioport.h>
27#include <linux/percpu.h>
28#include <linux/mmzone.h>
29#include <linux/gfp.h>
30#include <linux/bootmem_info.h>
31
32#include <asm/head.h>
33#include <asm/page.h>
34#include <asm/pgalloc.h>
35#include <asm/oplib.h>
36#include <asm/iommu.h>
37#include <asm/io.h>
38#include <linux/uaccess.h>
39#include <asm/mmu_context.h>
40#include <asm/tlbflush.h>
41#include <asm/dma.h>
42#include <asm/starfire.h>
43#include <asm/tlb.h>
44#include <asm/spitfire.h>
45#include <asm/sections.h>
46#include <asm/tsb.h>
47#include <asm/hypervisor.h>
48#include <asm/prom.h>
49#include <asm/mdesc.h>
50#include <asm/cpudata.h>
51#include <asm/setup.h>
52#include <asm/irq.h>
53
54#include "init_64.h"
55
56unsigned long kern_linear_pte_xor[4] __read_mostly;
57static unsigned long page_cache4v_flag;
58
59/* A bitmap, two bits for every 256MB of physical memory. These two
60 * bits determine what page size we use for kernel linear
61 * translations. They form an index into kern_linear_pte_xor[]. The
62 * value in the indexed slot is XOR'd with the TLB miss virtual
63 * address to form the resulting TTE. The mapping is:
64 *
65 * 0 ==> 4MB
66 * 1 ==> 256MB
67 * 2 ==> 2GB
68 * 3 ==> 16GB
69 *
70 * All sun4v chips support 256MB pages. Only SPARC-T4 and later
71 * support 2GB pages, and hopefully future cpus will support the 16GB
72 * pages as well. For slots 2 and 3, we encode a 256MB TTE xor there
73 * if these larger page sizes are not supported by the cpu.
74 *
75 * It would be nice to determine this from the machine description
76 * 'cpu' properties, but we need to have this table setup before the
77 * MDESC is initialized.
78 */
79
80#ifndef CONFIG_DEBUG_PAGEALLOC
81/* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
82 * Space is allocated for this right after the trap table in
83 * arch/sparc64/kernel/head.S
84 */
85extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
86#endif
87extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
88
89static unsigned long cpu_pgsz_mask;
90
91#define MAX_BANKS 1024
92
93static struct linux_prom64_registers pavail[MAX_BANKS];
94static int pavail_ents;
95
96u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
97
98static int cmp_p64(const void *a, const void *b)
99{
100 const struct linux_prom64_registers *x = a, *y = b;
101
102 if (x->phys_addr > y->phys_addr)
103 return 1;
104 if (x->phys_addr < y->phys_addr)
105 return -1;
106 return 0;
107}
108
109static void __init read_obp_memory(const char *property,
110 struct linux_prom64_registers *regs,
111 int *num_ents)
112{
113 phandle node = prom_finddevice("/memory");
114 int prop_size = prom_getproplen(node, property);
115 int ents, ret, i;
116
117 ents = prop_size / sizeof(struct linux_prom64_registers);
118 if (ents > MAX_BANKS) {
119 prom_printf("The machine has more %s property entries than "
120 "this kernel can support (%d).\n",
121 property, MAX_BANKS);
122 prom_halt();
123 }
124
125 ret = prom_getproperty(node, property, (char *) regs, prop_size);
126 if (ret == -1) {
127 prom_printf("Couldn't get %s property from /memory.\n",
128 property);
129 prom_halt();
130 }
131
132 /* Sanitize what we got from the firmware, by page aligning
133 * everything.
134 */
135 for (i = 0; i < ents; i++) {
136 unsigned long base, size;
137
138 base = regs[i].phys_addr;
139 size = regs[i].reg_size;
140
141 size &= PAGE_MASK;
142 if (base & ~PAGE_MASK) {
143 unsigned long new_base = PAGE_ALIGN(base);
144
145 size -= new_base - base;
146 if ((long) size < 0L)
147 size = 0UL;
148 base = new_base;
149 }
150 if (size == 0UL) {
151 /* If it is empty, simply get rid of it.
152 * This simplifies the logic of the other
153 * functions that process these arrays.
154 */
155 memmove(®s[i], ®s[i + 1],
156 (ents - i - 1) * sizeof(regs[0]));
157 i--;
158 ents--;
159 continue;
160 }
161 regs[i].phys_addr = base;
162 regs[i].reg_size = size;
163 }
164
165 *num_ents = ents;
166
167 sort(regs, ents, sizeof(struct linux_prom64_registers),
168 cmp_p64, NULL);
169}
170
171/* Kernel physical address base and size in bytes. */
172unsigned long kern_base __read_mostly;
173unsigned long kern_size __read_mostly;
174
175/* Initial ramdisk setup */
176extern unsigned long sparc_ramdisk_image64;
177extern unsigned int sparc_ramdisk_image;
178extern unsigned int sparc_ramdisk_size;
179
180struct page *mem_map_zero __read_mostly;
181EXPORT_SYMBOL(mem_map_zero);
182
183unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
184
185unsigned long sparc64_kern_pri_context __read_mostly;
186unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
187unsigned long sparc64_kern_sec_context __read_mostly;
188
189int num_kernel_image_mappings;
190
191#ifdef CONFIG_DEBUG_DCFLUSH
192atomic_t dcpage_flushes = ATOMIC_INIT(0);
193#ifdef CONFIG_SMP
194atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
195#endif
196#endif
197
198inline void flush_dcache_page_impl(struct page *page)
199{
200 BUG_ON(tlb_type == hypervisor);
201#ifdef CONFIG_DEBUG_DCFLUSH
202 atomic_inc(&dcpage_flushes);
203#endif
204
205#ifdef DCACHE_ALIASING_POSSIBLE
206 __flush_dcache_page(page_address(page),
207 ((tlb_type == spitfire) &&
208 page_mapping_file(page) != NULL));
209#else
210 if (page_mapping_file(page) != NULL &&
211 tlb_type == spitfire)
212 __flush_icache_page(__pa(page_address(page)));
213#endif
214}
215
216#define PG_dcache_dirty PG_arch_1
217#define PG_dcache_cpu_shift 32UL
218#define PG_dcache_cpu_mask \
219 ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
220
221#define dcache_dirty_cpu(page) \
222 (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
223
224static inline void set_dcache_dirty(struct page *page, int this_cpu)
225{
226 unsigned long mask = this_cpu;
227 unsigned long non_cpu_bits;
228
229 non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
230 mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
231
232 __asm__ __volatile__("1:\n\t"
233 "ldx [%2], %%g7\n\t"
234 "and %%g7, %1, %%g1\n\t"
235 "or %%g1, %0, %%g1\n\t"
236 "casx [%2], %%g7, %%g1\n\t"
237 "cmp %%g7, %%g1\n\t"
238 "bne,pn %%xcc, 1b\n\t"
239 " nop"
240 : /* no outputs */
241 : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
242 : "g1", "g7");
243}
244
245static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
246{
247 unsigned long mask = (1UL << PG_dcache_dirty);
248
249 __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
250 "1:\n\t"
251 "ldx [%2], %%g7\n\t"
252 "srlx %%g7, %4, %%g1\n\t"
253 "and %%g1, %3, %%g1\n\t"
254 "cmp %%g1, %0\n\t"
255 "bne,pn %%icc, 2f\n\t"
256 " andn %%g7, %1, %%g1\n\t"
257 "casx [%2], %%g7, %%g1\n\t"
258 "cmp %%g7, %%g1\n\t"
259 "bne,pn %%xcc, 1b\n\t"
260 " nop\n"
261 "2:"
262 : /* no outputs */
263 : "r" (cpu), "r" (mask), "r" (&page->flags),
264 "i" (PG_dcache_cpu_mask),
265 "i" (PG_dcache_cpu_shift)
266 : "g1", "g7");
267}
268
269static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
270{
271 unsigned long tsb_addr = (unsigned long) ent;
272
273 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
274 tsb_addr = __pa(tsb_addr);
275
276 __tsb_insert(tsb_addr, tag, pte);
277}
278
279unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
280
281static void flush_dcache(unsigned long pfn)
282{
283 struct page *page;
284
285 page = pfn_to_page(pfn);
286 if (page) {
287 unsigned long pg_flags;
288
289 pg_flags = page->flags;
290 if (pg_flags & (1UL << PG_dcache_dirty)) {
291 int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
292 PG_dcache_cpu_mask);
293 int this_cpu = get_cpu();
294
295 /* This is just to optimize away some function calls
296 * in the SMP case.
297 */
298 if (cpu == this_cpu)
299 flush_dcache_page_impl(page);
300 else
301 smp_flush_dcache_page_impl(page, cpu);
302
303 clear_dcache_dirty_cpu(page, cpu);
304
305 put_cpu();
306 }
307 }
308}
309
310/* mm->context.lock must be held */
311static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
312 unsigned long tsb_hash_shift, unsigned long address,
313 unsigned long tte)
314{
315 struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
316 unsigned long tag;
317
318 if (unlikely(!tsb))
319 return;
320
321 tsb += ((address >> tsb_hash_shift) &
322 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
323 tag = (address >> 22UL);
324 tsb_insert(tsb, tag, tte);
325}
326
327#ifdef CONFIG_HUGETLB_PAGE
328static int __init hugetlbpage_init(void)
329{
330 hugetlb_add_hstate(HPAGE_64K_SHIFT - PAGE_SHIFT);
331 hugetlb_add_hstate(HPAGE_SHIFT - PAGE_SHIFT);
332 hugetlb_add_hstate(HPAGE_256MB_SHIFT - PAGE_SHIFT);
333 hugetlb_add_hstate(HPAGE_2GB_SHIFT - PAGE_SHIFT);
334
335 return 0;
336}
337
338arch_initcall(hugetlbpage_init);
339
340static void __init pud_huge_patch(void)
341{
342 struct pud_huge_patch_entry *p;
343 unsigned long addr;
344
345 p = &__pud_huge_patch;
346 addr = p->addr;
347 *(unsigned int *)addr = p->insn;
348
349 __asm__ __volatile__("flush %0" : : "r" (addr));
350}
351
352bool __init arch_hugetlb_valid_size(unsigned long size)
353{
354 unsigned int hugepage_shift = ilog2(size);
355 unsigned short hv_pgsz_idx;
356 unsigned int hv_pgsz_mask;
357
358 switch (hugepage_shift) {
359 case HPAGE_16GB_SHIFT:
360 hv_pgsz_mask = HV_PGSZ_MASK_16GB;
361 hv_pgsz_idx = HV_PGSZ_IDX_16GB;
362 pud_huge_patch();
363 break;
364 case HPAGE_2GB_SHIFT:
365 hv_pgsz_mask = HV_PGSZ_MASK_2GB;
366 hv_pgsz_idx = HV_PGSZ_IDX_2GB;
367 break;
368 case HPAGE_256MB_SHIFT:
369 hv_pgsz_mask = HV_PGSZ_MASK_256MB;
370 hv_pgsz_idx = HV_PGSZ_IDX_256MB;
371 break;
372 case HPAGE_SHIFT:
373 hv_pgsz_mask = HV_PGSZ_MASK_4MB;
374 hv_pgsz_idx = HV_PGSZ_IDX_4MB;
375 break;
376 case HPAGE_64K_SHIFT:
377 hv_pgsz_mask = HV_PGSZ_MASK_64K;
378 hv_pgsz_idx = HV_PGSZ_IDX_64K;
379 break;
380 default:
381 hv_pgsz_mask = 0;
382 }
383
384 if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U)
385 return false;
386
387 return true;
388}
389#endif /* CONFIG_HUGETLB_PAGE */
390
391void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
392{
393 struct mm_struct *mm;
394 unsigned long flags;
395 bool is_huge_tsb;
396 pte_t pte = *ptep;
397
398 if (tlb_type != hypervisor) {
399 unsigned long pfn = pte_pfn(pte);
400
401 if (pfn_valid(pfn))
402 flush_dcache(pfn);
403 }
404
405 mm = vma->vm_mm;
406
407 /* Don't insert a non-valid PTE into the TSB, we'll deadlock. */
408 if (!pte_accessible(mm, pte))
409 return;
410
411 spin_lock_irqsave(&mm->context.lock, flags);
412
413 is_huge_tsb = false;
414#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
415 if (mm->context.hugetlb_pte_count || mm->context.thp_pte_count) {
416 unsigned long hugepage_size = PAGE_SIZE;
417
418 if (is_vm_hugetlb_page(vma))
419 hugepage_size = huge_page_size(hstate_vma(vma));
420
421 if (hugepage_size >= PUD_SIZE) {
422 unsigned long mask = 0x1ffc00000UL;
423
424 /* Transfer bits [32:22] from address to resolve
425 * at 4M granularity.
426 */
427 pte_val(pte) &= ~mask;
428 pte_val(pte) |= (address & mask);
429 } else if (hugepage_size >= PMD_SIZE) {
430 /* We are fabricating 8MB pages using 4MB
431 * real hw pages.
432 */
433 pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT));
434 }
435
436 if (hugepage_size >= PMD_SIZE) {
437 __update_mmu_tsb_insert(mm, MM_TSB_HUGE,
438 REAL_HPAGE_SHIFT, address, pte_val(pte));
439 is_huge_tsb = true;
440 }
441 }
442#endif
443 if (!is_huge_tsb)
444 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
445 address, pte_val(pte));
446
447 spin_unlock_irqrestore(&mm->context.lock, flags);
448}
449
450void flush_dcache_page(struct page *page)
451{
452 struct address_space *mapping;
453 int this_cpu;
454
455 if (tlb_type == hypervisor)
456 return;
457
458 /* Do not bother with the expensive D-cache flush if it
459 * is merely the zero page. The 'bigcore' testcase in GDB
460 * causes this case to run millions of times.
461 */
462 if (page == ZERO_PAGE(0))
463 return;
464
465 this_cpu = get_cpu();
466
467 mapping = page_mapping_file(page);
468 if (mapping && !mapping_mapped(mapping)) {
469 int dirty = test_bit(PG_dcache_dirty, &page->flags);
470 if (dirty) {
471 int dirty_cpu = dcache_dirty_cpu(page);
472
473 if (dirty_cpu == this_cpu)
474 goto out;
475 smp_flush_dcache_page_impl(page, dirty_cpu);
476 }
477 set_dcache_dirty(page, this_cpu);
478 } else {
479 /* We could delay the flush for the !page_mapping
480 * case too. But that case is for exec env/arg
481 * pages and those are %99 certainly going to get
482 * faulted into the tlb (and thus flushed) anyways.
483 */
484 flush_dcache_page_impl(page);
485 }
486
487out:
488 put_cpu();
489}
490EXPORT_SYMBOL(flush_dcache_page);
491
492void __kprobes flush_icache_range(unsigned long start, unsigned long end)
493{
494 /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
495 if (tlb_type == spitfire) {
496 unsigned long kaddr;
497
498 /* This code only runs on Spitfire cpus so this is
499 * why we can assume _PAGE_PADDR_4U.
500 */
501 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
502 unsigned long paddr, mask = _PAGE_PADDR_4U;
503
504 if (kaddr >= PAGE_OFFSET)
505 paddr = kaddr & mask;
506 else {
507 pte_t *ptep = virt_to_kpte(kaddr);
508
509 paddr = pte_val(*ptep) & mask;
510 }
511 __flush_icache_page(paddr);
512 }
513 }
514}
515EXPORT_SYMBOL(flush_icache_range);
516
517void mmu_info(struct seq_file *m)
518{
519 static const char *pgsz_strings[] = {
520 "8K", "64K", "512K", "4MB", "32MB",
521 "256MB", "2GB", "16GB",
522 };
523 int i, printed;
524
525 if (tlb_type == cheetah)
526 seq_printf(m, "MMU Type\t: Cheetah\n");
527 else if (tlb_type == cheetah_plus)
528 seq_printf(m, "MMU Type\t: Cheetah+\n");
529 else if (tlb_type == spitfire)
530 seq_printf(m, "MMU Type\t: Spitfire\n");
531 else if (tlb_type == hypervisor)
532 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
533 else
534 seq_printf(m, "MMU Type\t: ???\n");
535
536 seq_printf(m, "MMU PGSZs\t: ");
537 printed = 0;
538 for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
539 if (cpu_pgsz_mask & (1UL << i)) {
540 seq_printf(m, "%s%s",
541 printed ? "," : "", pgsz_strings[i]);
542 printed++;
543 }
544 }
545 seq_putc(m, '\n');
546
547#ifdef CONFIG_DEBUG_DCFLUSH
548 seq_printf(m, "DCPageFlushes\t: %d\n",
549 atomic_read(&dcpage_flushes));
550#ifdef CONFIG_SMP
551 seq_printf(m, "DCPageFlushesXC\t: %d\n",
552 atomic_read(&dcpage_flushes_xcall));
553#endif /* CONFIG_SMP */
554#endif /* CONFIG_DEBUG_DCFLUSH */
555}
556
557struct linux_prom_translation prom_trans[512] __read_mostly;
558unsigned int prom_trans_ents __read_mostly;
559
560unsigned long kern_locked_tte_data;
561
562/* The obp translations are saved based on 8k pagesize, since obp can
563 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
564 * HI_OBP_ADDRESS range are handled in ktlb.S.
565 */
566static inline int in_obp_range(unsigned long vaddr)
567{
568 return (vaddr >= LOW_OBP_ADDRESS &&
569 vaddr < HI_OBP_ADDRESS);
570}
571
572static int cmp_ptrans(const void *a, const void *b)
573{
574 const struct linux_prom_translation *x = a, *y = b;
575
576 if (x->virt > y->virt)
577 return 1;
578 if (x->virt < y->virt)
579 return -1;
580 return 0;
581}
582
583/* Read OBP translations property into 'prom_trans[]'. */
584static void __init read_obp_translations(void)
585{
586 int n, node, ents, first, last, i;
587
588 node = prom_finddevice("/virtual-memory");
589 n = prom_getproplen(node, "translations");
590 if (unlikely(n == 0 || n == -1)) {
591 prom_printf("prom_mappings: Couldn't get size.\n");
592 prom_halt();
593 }
594 if (unlikely(n > sizeof(prom_trans))) {
595 prom_printf("prom_mappings: Size %d is too big.\n", n);
596 prom_halt();
597 }
598
599 if ((n = prom_getproperty(node, "translations",
600 (char *)&prom_trans[0],
601 sizeof(prom_trans))) == -1) {
602 prom_printf("prom_mappings: Couldn't get property.\n");
603 prom_halt();
604 }
605
606 n = n / sizeof(struct linux_prom_translation);
607
608 ents = n;
609
610 sort(prom_trans, ents, sizeof(struct linux_prom_translation),
611 cmp_ptrans, NULL);
612
613 /* Now kick out all the non-OBP entries. */
614 for (i = 0; i < ents; i++) {
615 if (in_obp_range(prom_trans[i].virt))
616 break;
617 }
618 first = i;
619 for (; i < ents; i++) {
620 if (!in_obp_range(prom_trans[i].virt))
621 break;
622 }
623 last = i;
624
625 for (i = 0; i < (last - first); i++) {
626 struct linux_prom_translation *src = &prom_trans[i + first];
627 struct linux_prom_translation *dest = &prom_trans[i];
628
629 *dest = *src;
630 }
631 for (; i < ents; i++) {
632 struct linux_prom_translation *dest = &prom_trans[i];
633 dest->virt = dest->size = dest->data = 0x0UL;
634 }
635
636 prom_trans_ents = last - first;
637
638 if (tlb_type == spitfire) {
639 /* Clear diag TTE bits. */
640 for (i = 0; i < prom_trans_ents; i++)
641 prom_trans[i].data &= ~0x0003fe0000000000UL;
642 }
643
644 /* Force execute bit on. */
645 for (i = 0; i < prom_trans_ents; i++)
646 prom_trans[i].data |= (tlb_type == hypervisor ?
647 _PAGE_EXEC_4V : _PAGE_EXEC_4U);
648}
649
650static void __init hypervisor_tlb_lock(unsigned long vaddr,
651 unsigned long pte,
652 unsigned long mmu)
653{
654 unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
655
656 if (ret != 0) {
657 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
658 "errors with %lx\n", vaddr, 0, pte, mmu, ret);
659 prom_halt();
660 }
661}
662
663static unsigned long kern_large_tte(unsigned long paddr);
664
665static void __init remap_kernel(void)
666{
667 unsigned long phys_page, tte_vaddr, tte_data;
668 int i, tlb_ent = sparc64_highest_locked_tlbent();
669
670 tte_vaddr = (unsigned long) KERNBASE;
671 phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
672 tte_data = kern_large_tte(phys_page);
673
674 kern_locked_tte_data = tte_data;
675
676 /* Now lock us into the TLBs via Hypervisor or OBP. */
677 if (tlb_type == hypervisor) {
678 for (i = 0; i < num_kernel_image_mappings; i++) {
679 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
680 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
681 tte_vaddr += 0x400000;
682 tte_data += 0x400000;
683 }
684 } else {
685 for (i = 0; i < num_kernel_image_mappings; i++) {
686 prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
687 prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
688 tte_vaddr += 0x400000;
689 tte_data += 0x400000;
690 }
691 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
692 }
693 if (tlb_type == cheetah_plus) {
694 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
695 CTX_CHEETAH_PLUS_NUC);
696 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
697 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
698 }
699}
700
701
702static void __init inherit_prom_mappings(void)
703{
704 /* Now fixup OBP's idea about where we really are mapped. */
705 printk("Remapping the kernel... ");
706 remap_kernel();
707 printk("done.\n");
708}
709
710void prom_world(int enter)
711{
712 /*
713 * No need to change the address space any more, just flush
714 * the register windows
715 */
716 __asm__ __volatile__("flushw");
717}
718
719void __flush_dcache_range(unsigned long start, unsigned long end)
720{
721 unsigned long va;
722
723 if (tlb_type == spitfire) {
724 int n = 0;
725
726 for (va = start; va < end; va += 32) {
727 spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
728 if (++n >= 512)
729 break;
730 }
731 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
732 start = __pa(start);
733 end = __pa(end);
734 for (va = start; va < end; va += 32)
735 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
736 "membar #Sync"
737 : /* no outputs */
738 : "r" (va),
739 "i" (ASI_DCACHE_INVALIDATE));
740 }
741}
742EXPORT_SYMBOL(__flush_dcache_range);
743
744/* get_new_mmu_context() uses "cache + 1". */
745DEFINE_SPINLOCK(ctx_alloc_lock);
746unsigned long tlb_context_cache = CTX_FIRST_VERSION;
747#define MAX_CTX_NR (1UL << CTX_NR_BITS)
748#define CTX_BMAP_SLOTS BITS_TO_LONGS(MAX_CTX_NR)
749DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
750DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0};
751
752static void mmu_context_wrap(void)
753{
754 unsigned long old_ver = tlb_context_cache & CTX_VERSION_MASK;
755 unsigned long new_ver, new_ctx, old_ctx;
756 struct mm_struct *mm;
757 int cpu;
758
759 bitmap_zero(mmu_context_bmap, 1 << CTX_NR_BITS);
760
761 /* Reserve kernel context */
762 set_bit(0, mmu_context_bmap);
763
764 new_ver = (tlb_context_cache & CTX_VERSION_MASK) + CTX_FIRST_VERSION;
765 if (unlikely(new_ver == 0))
766 new_ver = CTX_FIRST_VERSION;
767 tlb_context_cache = new_ver;
768
769 /*
770 * Make sure that any new mm that are added into per_cpu_secondary_mm,
771 * are going to go through get_new_mmu_context() path.
772 */
773 mb();
774
775 /*
776 * Updated versions to current on those CPUs that had valid secondary
777 * contexts
778 */
779 for_each_online_cpu(cpu) {
780 /*
781 * If a new mm is stored after we took this mm from the array,
782 * it will go into get_new_mmu_context() path, because we
783 * already bumped the version in tlb_context_cache.
784 */
785 mm = per_cpu(per_cpu_secondary_mm, cpu);
786
787 if (unlikely(!mm || mm == &init_mm))
788 continue;
789
790 old_ctx = mm->context.sparc64_ctx_val;
791 if (likely((old_ctx & CTX_VERSION_MASK) == old_ver)) {
792 new_ctx = (old_ctx & ~CTX_VERSION_MASK) | new_ver;
793 set_bit(new_ctx & CTX_NR_MASK, mmu_context_bmap);
794 mm->context.sparc64_ctx_val = new_ctx;
795 }
796 }
797}
798
799/* Caller does TLB context flushing on local CPU if necessary.
800 * The caller also ensures that CTX_VALID(mm->context) is false.
801 *
802 * We must be careful about boundary cases so that we never
803 * let the user have CTX 0 (nucleus) or we ever use a CTX
804 * version of zero (and thus NO_CONTEXT would not be caught
805 * by version mis-match tests in mmu_context.h).
806 *
807 * Always invoked with interrupts disabled.
808 */
809void get_new_mmu_context(struct mm_struct *mm)
810{
811 unsigned long ctx, new_ctx;
812 unsigned long orig_pgsz_bits;
813
814 spin_lock(&ctx_alloc_lock);
815retry:
816 /* wrap might have happened, test again if our context became valid */
817 if (unlikely(CTX_VALID(mm->context)))
818 goto out;
819 orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
820 ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
821 new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
822 if (new_ctx >= (1 << CTX_NR_BITS)) {
823 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
824 if (new_ctx >= ctx) {
825 mmu_context_wrap();
826 goto retry;
827 }
828 }
829 if (mm->context.sparc64_ctx_val)
830 cpumask_clear(mm_cpumask(mm));
831 mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
832 new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
833 tlb_context_cache = new_ctx;
834 mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
835out:
836 spin_unlock(&ctx_alloc_lock);
837}
838
839static int numa_enabled = 1;
840static int numa_debug;
841
842static int __init early_numa(char *p)
843{
844 if (!p)
845 return 0;
846
847 if (strstr(p, "off"))
848 numa_enabled = 0;
849
850 if (strstr(p, "debug"))
851 numa_debug = 1;
852
853 return 0;
854}
855early_param("numa", early_numa);
856
857#define numadbg(f, a...) \
858do { if (numa_debug) \
859 printk(KERN_INFO f, ## a); \
860} while (0)
861
862static void __init find_ramdisk(unsigned long phys_base)
863{
864#ifdef CONFIG_BLK_DEV_INITRD
865 if (sparc_ramdisk_image || sparc_ramdisk_image64) {
866 unsigned long ramdisk_image;
867
868 /* Older versions of the bootloader only supported a
869 * 32-bit physical address for the ramdisk image
870 * location, stored at sparc_ramdisk_image. Newer
871 * SILO versions set sparc_ramdisk_image to zero and
872 * provide a full 64-bit physical address at
873 * sparc_ramdisk_image64.
874 */
875 ramdisk_image = sparc_ramdisk_image;
876 if (!ramdisk_image)
877 ramdisk_image = sparc_ramdisk_image64;
878
879 /* Another bootloader quirk. The bootloader normalizes
880 * the physical address to KERNBASE, so we have to
881 * factor that back out and add in the lowest valid
882 * physical page address to get the true physical address.
883 */
884 ramdisk_image -= KERNBASE;
885 ramdisk_image += phys_base;
886
887 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
888 ramdisk_image, sparc_ramdisk_size);
889
890 initrd_start = ramdisk_image;
891 initrd_end = ramdisk_image + sparc_ramdisk_size;
892
893 memblock_reserve(initrd_start, sparc_ramdisk_size);
894
895 initrd_start += PAGE_OFFSET;
896 initrd_end += PAGE_OFFSET;
897 }
898#endif
899}
900
901struct node_mem_mask {
902 unsigned long mask;
903 unsigned long match;
904};
905static struct node_mem_mask node_masks[MAX_NUMNODES];
906static int num_node_masks;
907
908#ifdef CONFIG_NUMA
909
910struct mdesc_mlgroup {
911 u64 node;
912 u64 latency;
913 u64 match;
914 u64 mask;
915};
916
917static struct mdesc_mlgroup *mlgroups;
918static int num_mlgroups;
919
920int numa_cpu_lookup_table[NR_CPUS];
921cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
922
923struct mdesc_mblock {
924 u64 base;
925 u64 size;
926 u64 offset; /* RA-to-PA */
927};
928static struct mdesc_mblock *mblocks;
929static int num_mblocks;
930
931static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr)
932{
933 struct mdesc_mblock *m = NULL;
934 int i;
935
936 for (i = 0; i < num_mblocks; i++) {
937 m = &mblocks[i];
938
939 if (addr >= m->base &&
940 addr < (m->base + m->size)) {
941 break;
942 }
943 }
944
945 return m;
946}
947
948static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid)
949{
950 int prev_nid, new_nid;
951
952 prev_nid = NUMA_NO_NODE;
953 for ( ; start < end; start += PAGE_SIZE) {
954 for (new_nid = 0; new_nid < num_node_masks; new_nid++) {
955 struct node_mem_mask *p = &node_masks[new_nid];
956
957 if ((start & p->mask) == p->match) {
958 if (prev_nid == NUMA_NO_NODE)
959 prev_nid = new_nid;
960 break;
961 }
962 }
963
964 if (new_nid == num_node_masks) {
965 prev_nid = 0;
966 WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.",
967 start);
968 break;
969 }
970
971 if (prev_nid != new_nid)
972 break;
973 }
974 *nid = prev_nid;
975
976 return start > end ? end : start;
977}
978
979static u64 __init memblock_nid_range(u64 start, u64 end, int *nid)
980{
981 u64 ret_end, pa_start, m_mask, m_match, m_end;
982 struct mdesc_mblock *mblock;
983 int _nid, i;
984
985 if (tlb_type != hypervisor)
986 return memblock_nid_range_sun4u(start, end, nid);
987
988 mblock = addr_to_mblock(start);
989 if (!mblock) {
990 WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]",
991 start);
992
993 _nid = 0;
994 ret_end = end;
995 goto done;
996 }
997
998 pa_start = start + mblock->offset;
999 m_match = 0;
1000 m_mask = 0;
1001
1002 for (_nid = 0; _nid < num_node_masks; _nid++) {
1003 struct node_mem_mask *const m = &node_masks[_nid];
1004
1005 if ((pa_start & m->mask) == m->match) {
1006 m_match = m->match;
1007 m_mask = m->mask;
1008 break;
1009 }
1010 }
1011
1012 if (num_node_masks == _nid) {
1013 /* We could not find NUMA group, so default to 0, but lets
1014 * search for latency group, so we could calculate the correct
1015 * end address that we return
1016 */
1017 _nid = 0;
1018
1019 for (i = 0; i < num_mlgroups; i++) {
1020 struct mdesc_mlgroup *const m = &mlgroups[i];
1021
1022 if ((pa_start & m->mask) == m->match) {
1023 m_match = m->match;
1024 m_mask = m->mask;
1025 break;
1026 }
1027 }
1028
1029 if (i == num_mlgroups) {
1030 WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]",
1031 start);
1032
1033 ret_end = end;
1034 goto done;
1035 }
1036 }
1037
1038 /*
1039 * Each latency group has match and mask, and each memory block has an
1040 * offset. An address belongs to a latency group if its address matches
1041 * the following formula: ((addr + offset) & mask) == match
1042 * It is, however, slow to check every single page if it matches a
1043 * particular latency group. As optimization we calculate end value by
1044 * using bit arithmetics.
1045 */
1046 m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset;
1047 m_end += pa_start & ~((1ul << fls64(m_mask)) - 1);
1048 ret_end = m_end > end ? end : m_end;
1049
1050done:
1051 *nid = _nid;
1052 return ret_end;
1053}
1054#endif
1055
1056/* This must be invoked after performing all of the necessary
1057 * memblock_set_node() calls for 'nid'. We need to be able to get
1058 * correct data from get_pfn_range_for_nid().
1059 */
1060static void __init allocate_node_data(int nid)
1061{
1062 struct pglist_data *p;
1063 unsigned long start_pfn, end_pfn;
1064#ifdef CONFIG_NUMA
1065
1066 NODE_DATA(nid) = memblock_alloc_node(sizeof(struct pglist_data),
1067 SMP_CACHE_BYTES, nid);
1068 if (!NODE_DATA(nid)) {
1069 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
1070 prom_halt();
1071 }
1072
1073 NODE_DATA(nid)->node_id = nid;
1074#endif
1075
1076 p = NODE_DATA(nid);
1077
1078 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1079 p->node_start_pfn = start_pfn;
1080 p->node_spanned_pages = end_pfn - start_pfn;
1081}
1082
1083static void init_node_masks_nonnuma(void)
1084{
1085#ifdef CONFIG_NUMA
1086 int i;
1087#endif
1088
1089 numadbg("Initializing tables for non-numa.\n");
1090
1091 node_masks[0].mask = 0;
1092 node_masks[0].match = 0;
1093 num_node_masks = 1;
1094
1095#ifdef CONFIG_NUMA
1096 for (i = 0; i < NR_CPUS; i++)
1097 numa_cpu_lookup_table[i] = 0;
1098
1099 cpumask_setall(&numa_cpumask_lookup_table[0]);
1100#endif
1101}
1102
1103#ifdef CONFIG_NUMA
1104struct pglist_data *node_data[MAX_NUMNODES];
1105
1106EXPORT_SYMBOL(numa_cpu_lookup_table);
1107EXPORT_SYMBOL(numa_cpumask_lookup_table);
1108EXPORT_SYMBOL(node_data);
1109
1110static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
1111 u32 cfg_handle)
1112{
1113 u64 arc;
1114
1115 mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
1116 u64 target = mdesc_arc_target(md, arc);
1117 const u64 *val;
1118
1119 val = mdesc_get_property(md, target,
1120 "cfg-handle", NULL);
1121 if (val && *val == cfg_handle)
1122 return 0;
1123 }
1124 return -ENODEV;
1125}
1126
1127static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
1128 u32 cfg_handle)
1129{
1130 u64 arc, candidate, best_latency = ~(u64)0;
1131
1132 candidate = MDESC_NODE_NULL;
1133 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1134 u64 target = mdesc_arc_target(md, arc);
1135 const char *name = mdesc_node_name(md, target);
1136 const u64 *val;
1137
1138 if (strcmp(name, "pio-latency-group"))
1139 continue;
1140
1141 val = mdesc_get_property(md, target, "latency", NULL);
1142 if (!val)
1143 continue;
1144
1145 if (*val < best_latency) {
1146 candidate = target;
1147 best_latency = *val;
1148 }
1149 }
1150
1151 if (candidate == MDESC_NODE_NULL)
1152 return -ENODEV;
1153
1154 return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
1155}
1156
1157int of_node_to_nid(struct device_node *dp)
1158{
1159 const struct linux_prom64_registers *regs;
1160 struct mdesc_handle *md;
1161 u32 cfg_handle;
1162 int count, nid;
1163 u64 grp;
1164
1165 /* This is the right thing to do on currently supported
1166 * SUN4U NUMA platforms as well, as the PCI controller does
1167 * not sit behind any particular memory controller.
1168 */
1169 if (!mlgroups)
1170 return -1;
1171
1172 regs = of_get_property(dp, "reg", NULL);
1173 if (!regs)
1174 return -1;
1175
1176 cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1177
1178 md = mdesc_grab();
1179
1180 count = 0;
1181 nid = NUMA_NO_NODE;
1182 mdesc_for_each_node_by_name(md, grp, "group") {
1183 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1184 nid = count;
1185 break;
1186 }
1187 count++;
1188 }
1189
1190 mdesc_release(md);
1191
1192 return nid;
1193}
1194
1195static void __init add_node_ranges(void)
1196{
1197 phys_addr_t start, end;
1198 unsigned long prev_max;
1199 u64 i;
1200
1201memblock_resized:
1202 prev_max = memblock.memory.max;
1203
1204 for_each_mem_range(i, &start, &end) {
1205 while (start < end) {
1206 unsigned long this_end;
1207 int nid;
1208
1209 this_end = memblock_nid_range(start, end, &nid);
1210
1211 numadbg("Setting memblock NUMA node nid[%d] "
1212 "start[%llx] end[%lx]\n",
1213 nid, start, this_end);
1214
1215 memblock_set_node(start, this_end - start,
1216 &memblock.memory, nid);
1217 if (memblock.memory.max != prev_max)
1218 goto memblock_resized;
1219 start = this_end;
1220 }
1221 }
1222}
1223
1224static int __init grab_mlgroups(struct mdesc_handle *md)
1225{
1226 unsigned long paddr;
1227 int count = 0;
1228 u64 node;
1229
1230 mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1231 count++;
1232 if (!count)
1233 return -ENOENT;
1234
1235 paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mlgroup),
1236 SMP_CACHE_BYTES);
1237 if (!paddr)
1238 return -ENOMEM;
1239
1240 mlgroups = __va(paddr);
1241 num_mlgroups = count;
1242
1243 count = 0;
1244 mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1245 struct mdesc_mlgroup *m = &mlgroups[count++];
1246 const u64 *val;
1247
1248 m->node = node;
1249
1250 val = mdesc_get_property(md, node, "latency", NULL);
1251 m->latency = *val;
1252 val = mdesc_get_property(md, node, "address-match", NULL);
1253 m->match = *val;
1254 val = mdesc_get_property(md, node, "address-mask", NULL);
1255 m->mask = *val;
1256
1257 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1258 "match[%llx] mask[%llx]\n",
1259 count - 1, m->node, m->latency, m->match, m->mask);
1260 }
1261
1262 return 0;
1263}
1264
1265static int __init grab_mblocks(struct mdesc_handle *md)
1266{
1267 unsigned long paddr;
1268 int count = 0;
1269 u64 node;
1270
1271 mdesc_for_each_node_by_name(md, node, "mblock")
1272 count++;
1273 if (!count)
1274 return -ENOENT;
1275
1276 paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mblock),
1277 SMP_CACHE_BYTES);
1278 if (!paddr)
1279 return -ENOMEM;
1280
1281 mblocks = __va(paddr);
1282 num_mblocks = count;
1283
1284 count = 0;
1285 mdesc_for_each_node_by_name(md, node, "mblock") {
1286 struct mdesc_mblock *m = &mblocks[count++];
1287 const u64 *val;
1288
1289 val = mdesc_get_property(md, node, "base", NULL);
1290 m->base = *val;
1291 val = mdesc_get_property(md, node, "size", NULL);
1292 m->size = *val;
1293 val = mdesc_get_property(md, node,
1294 "address-congruence-offset", NULL);
1295
1296 /* The address-congruence-offset property is optional.
1297 * Explicity zero it be identifty this.
1298 */
1299 if (val)
1300 m->offset = *val;
1301 else
1302 m->offset = 0UL;
1303
1304 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1305 count - 1, m->base, m->size, m->offset);
1306 }
1307
1308 return 0;
1309}
1310
1311static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1312 u64 grp, cpumask_t *mask)
1313{
1314 u64 arc;
1315
1316 cpumask_clear(mask);
1317
1318 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1319 u64 target = mdesc_arc_target(md, arc);
1320 const char *name = mdesc_node_name(md, target);
1321 const u64 *id;
1322
1323 if (strcmp(name, "cpu"))
1324 continue;
1325 id = mdesc_get_property(md, target, "id", NULL);
1326 if (*id < nr_cpu_ids)
1327 cpumask_set_cpu(*id, mask);
1328 }
1329}
1330
1331static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1332{
1333 int i;
1334
1335 for (i = 0; i < num_mlgroups; i++) {
1336 struct mdesc_mlgroup *m = &mlgroups[i];
1337 if (m->node == node)
1338 return m;
1339 }
1340 return NULL;
1341}
1342
1343int __node_distance(int from, int to)
1344{
1345 if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1346 pr_warn("Returning default NUMA distance value for %d->%d\n",
1347 from, to);
1348 return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1349 }
1350 return numa_latency[from][to];
1351}
1352EXPORT_SYMBOL(__node_distance);
1353
1354static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1355{
1356 int i;
1357
1358 for (i = 0; i < MAX_NUMNODES; i++) {
1359 struct node_mem_mask *n = &node_masks[i];
1360
1361 if ((grp->mask == n->mask) && (grp->match == n->match))
1362 break;
1363 }
1364 return i;
1365}
1366
1367static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1368 u64 grp, int index)
1369{
1370 u64 arc;
1371
1372 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1373 int tnode;
1374 u64 target = mdesc_arc_target(md, arc);
1375 struct mdesc_mlgroup *m = find_mlgroup(target);
1376
1377 if (!m)
1378 continue;
1379 tnode = find_best_numa_node_for_mlgroup(m);
1380 if (tnode == MAX_NUMNODES)
1381 continue;
1382 numa_latency[index][tnode] = m->latency;
1383 }
1384}
1385
1386static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1387 int index)
1388{
1389 struct mdesc_mlgroup *candidate = NULL;
1390 u64 arc, best_latency = ~(u64)0;
1391 struct node_mem_mask *n;
1392
1393 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1394 u64 target = mdesc_arc_target(md, arc);
1395 struct mdesc_mlgroup *m = find_mlgroup(target);
1396 if (!m)
1397 continue;
1398 if (m->latency < best_latency) {
1399 candidate = m;
1400 best_latency = m->latency;
1401 }
1402 }
1403 if (!candidate)
1404 return -ENOENT;
1405
1406 if (num_node_masks != index) {
1407 printk(KERN_ERR "Inconsistent NUMA state, "
1408 "index[%d] != num_node_masks[%d]\n",
1409 index, num_node_masks);
1410 return -EINVAL;
1411 }
1412
1413 n = &node_masks[num_node_masks++];
1414
1415 n->mask = candidate->mask;
1416 n->match = candidate->match;
1417
1418 numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n",
1419 index, n->mask, n->match, candidate->latency);
1420
1421 return 0;
1422}
1423
1424static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1425 int index)
1426{
1427 cpumask_t mask;
1428 int cpu;
1429
1430 numa_parse_mdesc_group_cpus(md, grp, &mask);
1431
1432 for_each_cpu(cpu, &mask)
1433 numa_cpu_lookup_table[cpu] = index;
1434 cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1435
1436 if (numa_debug) {
1437 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1438 for_each_cpu(cpu, &mask)
1439 printk("%d ", cpu);
1440 printk("]\n");
1441 }
1442
1443 return numa_attach_mlgroup(md, grp, index);
1444}
1445
1446static int __init numa_parse_mdesc(void)
1447{
1448 struct mdesc_handle *md = mdesc_grab();
1449 int i, j, err, count;
1450 u64 node;
1451
1452 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1453 if (node == MDESC_NODE_NULL) {
1454 mdesc_release(md);
1455 return -ENOENT;
1456 }
1457
1458 err = grab_mblocks(md);
1459 if (err < 0)
1460 goto out;
1461
1462 err = grab_mlgroups(md);
1463 if (err < 0)
1464 goto out;
1465
1466 count = 0;
1467 mdesc_for_each_node_by_name(md, node, "group") {
1468 err = numa_parse_mdesc_group(md, node, count);
1469 if (err < 0)
1470 break;
1471 count++;
1472 }
1473
1474 count = 0;
1475 mdesc_for_each_node_by_name(md, node, "group") {
1476 find_numa_latencies_for_group(md, node, count);
1477 count++;
1478 }
1479
1480 /* Normalize numa latency matrix according to ACPI SLIT spec. */
1481 for (i = 0; i < MAX_NUMNODES; i++) {
1482 u64 self_latency = numa_latency[i][i];
1483
1484 for (j = 0; j < MAX_NUMNODES; j++) {
1485 numa_latency[i][j] =
1486 (numa_latency[i][j] * LOCAL_DISTANCE) /
1487 self_latency;
1488 }
1489 }
1490
1491 add_node_ranges();
1492
1493 for (i = 0; i < num_node_masks; i++) {
1494 allocate_node_data(i);
1495 node_set_online(i);
1496 }
1497
1498 err = 0;
1499out:
1500 mdesc_release(md);
1501 return err;
1502}
1503
1504static int __init numa_parse_jbus(void)
1505{
1506 unsigned long cpu, index;
1507
1508 /* NUMA node id is encoded in bits 36 and higher, and there is
1509 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1510 */
1511 index = 0;
1512 for_each_present_cpu(cpu) {
1513 numa_cpu_lookup_table[cpu] = index;
1514 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1515 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1516 node_masks[index].match = cpu << 36UL;
1517
1518 index++;
1519 }
1520 num_node_masks = index;
1521
1522 add_node_ranges();
1523
1524 for (index = 0; index < num_node_masks; index++) {
1525 allocate_node_data(index);
1526 node_set_online(index);
1527 }
1528
1529 return 0;
1530}
1531
1532static int __init numa_parse_sun4u(void)
1533{
1534 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1535 unsigned long ver;
1536
1537 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1538 if ((ver >> 32UL) == __JALAPENO_ID ||
1539 (ver >> 32UL) == __SERRANO_ID)
1540 return numa_parse_jbus();
1541 }
1542 return -1;
1543}
1544
1545static int __init bootmem_init_numa(void)
1546{
1547 int i, j;
1548 int err = -1;
1549
1550 numadbg("bootmem_init_numa()\n");
1551
1552 /* Some sane defaults for numa latency values */
1553 for (i = 0; i < MAX_NUMNODES; i++) {
1554 for (j = 0; j < MAX_NUMNODES; j++)
1555 numa_latency[i][j] = (i == j) ?
1556 LOCAL_DISTANCE : REMOTE_DISTANCE;
1557 }
1558
1559 if (numa_enabled) {
1560 if (tlb_type == hypervisor)
1561 err = numa_parse_mdesc();
1562 else
1563 err = numa_parse_sun4u();
1564 }
1565 return err;
1566}
1567
1568#else
1569
1570static int bootmem_init_numa(void)
1571{
1572 return -1;
1573}
1574
1575#endif
1576
1577static void __init bootmem_init_nonnuma(void)
1578{
1579 unsigned long top_of_ram = memblock_end_of_DRAM();
1580 unsigned long total_ram = memblock_phys_mem_size();
1581
1582 numadbg("bootmem_init_nonnuma()\n");
1583
1584 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1585 top_of_ram, total_ram);
1586 printk(KERN_INFO "Memory hole size: %ldMB\n",
1587 (top_of_ram - total_ram) >> 20);
1588
1589 init_node_masks_nonnuma();
1590 memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
1591 allocate_node_data(0);
1592 node_set_online(0);
1593}
1594
1595static unsigned long __init bootmem_init(unsigned long phys_base)
1596{
1597 unsigned long end_pfn;
1598
1599 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1600 max_pfn = max_low_pfn = end_pfn;
1601 min_low_pfn = (phys_base >> PAGE_SHIFT);
1602
1603 if (bootmem_init_numa() < 0)
1604 bootmem_init_nonnuma();
1605
1606 /* Dump memblock with node info. */
1607 memblock_dump_all();
1608
1609 /* XXX cpu notifier XXX */
1610
1611 sparse_init();
1612
1613 return end_pfn;
1614}
1615
1616static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1617static int pall_ents __initdata;
1618
1619static unsigned long max_phys_bits = 40;
1620
1621bool kern_addr_valid(unsigned long addr)
1622{
1623 pgd_t *pgd;
1624 p4d_t *p4d;
1625 pud_t *pud;
1626 pmd_t *pmd;
1627 pte_t *pte;
1628
1629 if ((long)addr < 0L) {
1630 unsigned long pa = __pa(addr);
1631
1632 if ((pa >> max_phys_bits) != 0UL)
1633 return false;
1634
1635 return pfn_valid(pa >> PAGE_SHIFT);
1636 }
1637
1638 if (addr >= (unsigned long) KERNBASE &&
1639 addr < (unsigned long)&_end)
1640 return true;
1641
1642 pgd = pgd_offset_k(addr);
1643 if (pgd_none(*pgd))
1644 return false;
1645
1646 p4d = p4d_offset(pgd, addr);
1647 if (p4d_none(*p4d))
1648 return false;
1649
1650 pud = pud_offset(p4d, addr);
1651 if (pud_none(*pud))
1652 return false;
1653
1654 if (pud_large(*pud))
1655 return pfn_valid(pud_pfn(*pud));
1656
1657 pmd = pmd_offset(pud, addr);
1658 if (pmd_none(*pmd))
1659 return false;
1660
1661 if (pmd_large(*pmd))
1662 return pfn_valid(pmd_pfn(*pmd));
1663
1664 pte = pte_offset_kernel(pmd, addr);
1665 if (pte_none(*pte))
1666 return false;
1667
1668 return pfn_valid(pte_pfn(*pte));
1669}
1670
1671static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1672 unsigned long vend,
1673 pud_t *pud)
1674{
1675 const unsigned long mask16gb = (1UL << 34) - 1UL;
1676 u64 pte_val = vstart;
1677
1678 /* Each PUD is 8GB */
1679 if ((vstart & mask16gb) ||
1680 (vend - vstart <= mask16gb)) {
1681 pte_val ^= kern_linear_pte_xor[2];
1682 pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1683
1684 return vstart + PUD_SIZE;
1685 }
1686
1687 pte_val ^= kern_linear_pte_xor[3];
1688 pte_val |= _PAGE_PUD_HUGE;
1689
1690 vend = vstart + mask16gb + 1UL;
1691 while (vstart < vend) {
1692 pud_val(*pud) = pte_val;
1693
1694 pte_val += PUD_SIZE;
1695 vstart += PUD_SIZE;
1696 pud++;
1697 }
1698 return vstart;
1699}
1700
1701static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1702 bool guard)
1703{
1704 if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1705 return true;
1706
1707 return false;
1708}
1709
1710static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1711 unsigned long vend,
1712 pmd_t *pmd)
1713{
1714 const unsigned long mask256mb = (1UL << 28) - 1UL;
1715 const unsigned long mask2gb = (1UL << 31) - 1UL;
1716 u64 pte_val = vstart;
1717
1718 /* Each PMD is 8MB */
1719 if ((vstart & mask256mb) ||
1720 (vend - vstart <= mask256mb)) {
1721 pte_val ^= kern_linear_pte_xor[0];
1722 pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1723
1724 return vstart + PMD_SIZE;
1725 }
1726
1727 if ((vstart & mask2gb) ||
1728 (vend - vstart <= mask2gb)) {
1729 pte_val ^= kern_linear_pte_xor[1];
1730 pte_val |= _PAGE_PMD_HUGE;
1731 vend = vstart + mask256mb + 1UL;
1732 } else {
1733 pte_val ^= kern_linear_pte_xor[2];
1734 pte_val |= _PAGE_PMD_HUGE;
1735 vend = vstart + mask2gb + 1UL;
1736 }
1737
1738 while (vstart < vend) {
1739 pmd_val(*pmd) = pte_val;
1740
1741 pte_val += PMD_SIZE;
1742 vstart += PMD_SIZE;
1743 pmd++;
1744 }
1745
1746 return vstart;
1747}
1748
1749static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1750 bool guard)
1751{
1752 if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1753 return true;
1754
1755 return false;
1756}
1757
1758static unsigned long __ref kernel_map_range(unsigned long pstart,
1759 unsigned long pend, pgprot_t prot,
1760 bool use_huge)
1761{
1762 unsigned long vstart = PAGE_OFFSET + pstart;
1763 unsigned long vend = PAGE_OFFSET + pend;
1764 unsigned long alloc_bytes = 0UL;
1765
1766 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1767 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1768 vstart, vend);
1769 prom_halt();
1770 }
1771
1772 while (vstart < vend) {
1773 unsigned long this_end, paddr = __pa(vstart);
1774 pgd_t *pgd = pgd_offset_k(vstart);
1775 p4d_t *p4d;
1776 pud_t *pud;
1777 pmd_t *pmd;
1778 pte_t *pte;
1779
1780 if (pgd_none(*pgd)) {
1781 pud_t *new;
1782
1783 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1784 PAGE_SIZE);
1785 if (!new)
1786 goto err_alloc;
1787 alloc_bytes += PAGE_SIZE;
1788 pgd_populate(&init_mm, pgd, new);
1789 }
1790
1791 p4d = p4d_offset(pgd, vstart);
1792 if (p4d_none(*p4d)) {
1793 pud_t *new;
1794
1795 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1796 PAGE_SIZE);
1797 if (!new)
1798 goto err_alloc;
1799 alloc_bytes += PAGE_SIZE;
1800 p4d_populate(&init_mm, p4d, new);
1801 }
1802
1803 pud = pud_offset(p4d, vstart);
1804 if (pud_none(*pud)) {
1805 pmd_t *new;
1806
1807 if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1808 vstart = kernel_map_hugepud(vstart, vend, pud);
1809 continue;
1810 }
1811 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1812 PAGE_SIZE);
1813 if (!new)
1814 goto err_alloc;
1815 alloc_bytes += PAGE_SIZE;
1816 pud_populate(&init_mm, pud, new);
1817 }
1818
1819 pmd = pmd_offset(pud, vstart);
1820 if (pmd_none(*pmd)) {
1821 pte_t *new;
1822
1823 if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1824 vstart = kernel_map_hugepmd(vstart, vend, pmd);
1825 continue;
1826 }
1827 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1828 PAGE_SIZE);
1829 if (!new)
1830 goto err_alloc;
1831 alloc_bytes += PAGE_SIZE;
1832 pmd_populate_kernel(&init_mm, pmd, new);
1833 }
1834
1835 pte = pte_offset_kernel(pmd, vstart);
1836 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1837 if (this_end > vend)
1838 this_end = vend;
1839
1840 while (vstart < this_end) {
1841 pte_val(*pte) = (paddr | pgprot_val(prot));
1842
1843 vstart += PAGE_SIZE;
1844 paddr += PAGE_SIZE;
1845 pte++;
1846 }
1847 }
1848
1849 return alloc_bytes;
1850
1851err_alloc:
1852 panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n",
1853 __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1854 return -ENOMEM;
1855}
1856
1857static void __init flush_all_kernel_tsbs(void)
1858{
1859 int i;
1860
1861 for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1862 struct tsb *ent = &swapper_tsb[i];
1863
1864 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1865 }
1866#ifndef CONFIG_DEBUG_PAGEALLOC
1867 for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1868 struct tsb *ent = &swapper_4m_tsb[i];
1869
1870 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1871 }
1872#endif
1873}
1874
1875extern unsigned int kvmap_linear_patch[1];
1876
1877static void __init kernel_physical_mapping_init(void)
1878{
1879 unsigned long i, mem_alloced = 0UL;
1880 bool use_huge = true;
1881
1882#ifdef CONFIG_DEBUG_PAGEALLOC
1883 use_huge = false;
1884#endif
1885 for (i = 0; i < pall_ents; i++) {
1886 unsigned long phys_start, phys_end;
1887
1888 phys_start = pall[i].phys_addr;
1889 phys_end = phys_start + pall[i].reg_size;
1890
1891 mem_alloced += kernel_map_range(phys_start, phys_end,
1892 PAGE_KERNEL, use_huge);
1893 }
1894
1895 printk("Allocated %ld bytes for kernel page tables.\n",
1896 mem_alloced);
1897
1898 kvmap_linear_patch[0] = 0x01000000; /* nop */
1899 flushi(&kvmap_linear_patch[0]);
1900
1901 flush_all_kernel_tsbs();
1902
1903 __flush_tlb_all();
1904}
1905
1906#ifdef CONFIG_DEBUG_PAGEALLOC
1907void __kernel_map_pages(struct page *page, int numpages, int enable)
1908{
1909 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1910 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1911
1912 kernel_map_range(phys_start, phys_end,
1913 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1914
1915 flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1916 PAGE_OFFSET + phys_end);
1917
1918 /* we should perform an IPI and flush all tlbs,
1919 * but that can deadlock->flush only current cpu.
1920 */
1921 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1922 PAGE_OFFSET + phys_end);
1923}
1924#endif
1925
1926unsigned long __init find_ecache_flush_span(unsigned long size)
1927{
1928 int i;
1929
1930 for (i = 0; i < pavail_ents; i++) {
1931 if (pavail[i].reg_size >= size)
1932 return pavail[i].phys_addr;
1933 }
1934
1935 return ~0UL;
1936}
1937
1938unsigned long PAGE_OFFSET;
1939EXPORT_SYMBOL(PAGE_OFFSET);
1940
1941unsigned long VMALLOC_END = 0x0000010000000000UL;
1942EXPORT_SYMBOL(VMALLOC_END);
1943
1944unsigned long sparc64_va_hole_top = 0xfffff80000000000UL;
1945unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1946
1947static void __init setup_page_offset(void)
1948{
1949 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1950 /* Cheetah/Panther support a full 64-bit virtual
1951 * address, so we can use all that our page tables
1952 * support.
1953 */
1954 sparc64_va_hole_top = 0xfff0000000000000UL;
1955 sparc64_va_hole_bottom = 0x0010000000000000UL;
1956
1957 max_phys_bits = 42;
1958 } else if (tlb_type == hypervisor) {
1959 switch (sun4v_chip_type) {
1960 case SUN4V_CHIP_NIAGARA1:
1961 case SUN4V_CHIP_NIAGARA2:
1962 /* T1 and T2 support 48-bit virtual addresses. */
1963 sparc64_va_hole_top = 0xffff800000000000UL;
1964 sparc64_va_hole_bottom = 0x0000800000000000UL;
1965
1966 max_phys_bits = 39;
1967 break;
1968 case SUN4V_CHIP_NIAGARA3:
1969 /* T3 supports 48-bit virtual addresses. */
1970 sparc64_va_hole_top = 0xffff800000000000UL;
1971 sparc64_va_hole_bottom = 0x0000800000000000UL;
1972
1973 max_phys_bits = 43;
1974 break;
1975 case SUN4V_CHIP_NIAGARA4:
1976 case SUN4V_CHIP_NIAGARA5:
1977 case SUN4V_CHIP_SPARC64X:
1978 case SUN4V_CHIP_SPARC_M6:
1979 /* T4 and later support 52-bit virtual addresses. */
1980 sparc64_va_hole_top = 0xfff8000000000000UL;
1981 sparc64_va_hole_bottom = 0x0008000000000000UL;
1982 max_phys_bits = 47;
1983 break;
1984 case SUN4V_CHIP_SPARC_M7:
1985 case SUN4V_CHIP_SPARC_SN:
1986 /* M7 and later support 52-bit virtual addresses. */
1987 sparc64_va_hole_top = 0xfff8000000000000UL;
1988 sparc64_va_hole_bottom = 0x0008000000000000UL;
1989 max_phys_bits = 49;
1990 break;
1991 case SUN4V_CHIP_SPARC_M8:
1992 default:
1993 /* M8 and later support 54-bit virtual addresses.
1994 * However, restricting M8 and above VA bits to 53
1995 * as 4-level page table cannot support more than
1996 * 53 VA bits.
1997 */
1998 sparc64_va_hole_top = 0xfff0000000000000UL;
1999 sparc64_va_hole_bottom = 0x0010000000000000UL;
2000 max_phys_bits = 51;
2001 break;
2002 }
2003 }
2004
2005 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
2006 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
2007 max_phys_bits);
2008 prom_halt();
2009 }
2010
2011 PAGE_OFFSET = sparc64_va_hole_top;
2012 VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
2013 (sparc64_va_hole_bottom >> 2));
2014
2015 pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
2016 PAGE_OFFSET, max_phys_bits);
2017 pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
2018 VMALLOC_START, VMALLOC_END);
2019 pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
2020 VMEMMAP_BASE, VMEMMAP_BASE << 1);
2021}
2022
2023static void __init tsb_phys_patch(void)
2024{
2025 struct tsb_ldquad_phys_patch_entry *pquad;
2026 struct tsb_phys_patch_entry *p;
2027
2028 pquad = &__tsb_ldquad_phys_patch;
2029 while (pquad < &__tsb_ldquad_phys_patch_end) {
2030 unsigned long addr = pquad->addr;
2031
2032 if (tlb_type == hypervisor)
2033 *(unsigned int *) addr = pquad->sun4v_insn;
2034 else
2035 *(unsigned int *) addr = pquad->sun4u_insn;
2036 wmb();
2037 __asm__ __volatile__("flush %0"
2038 : /* no outputs */
2039 : "r" (addr));
2040
2041 pquad++;
2042 }
2043
2044 p = &__tsb_phys_patch;
2045 while (p < &__tsb_phys_patch_end) {
2046 unsigned long addr = p->addr;
2047
2048 *(unsigned int *) addr = p->insn;
2049 wmb();
2050 __asm__ __volatile__("flush %0"
2051 : /* no outputs */
2052 : "r" (addr));
2053
2054 p++;
2055 }
2056}
2057
2058/* Don't mark as init, we give this to the Hypervisor. */
2059#ifndef CONFIG_DEBUG_PAGEALLOC
2060#define NUM_KTSB_DESCR 2
2061#else
2062#define NUM_KTSB_DESCR 1
2063#endif
2064static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
2065
2066/* The swapper TSBs are loaded with a base sequence of:
2067 *
2068 * sethi %uhi(SYMBOL), REG1
2069 * sethi %hi(SYMBOL), REG2
2070 * or REG1, %ulo(SYMBOL), REG1
2071 * or REG2, %lo(SYMBOL), REG2
2072 * sllx REG1, 32, REG1
2073 * or REG1, REG2, REG1
2074 *
2075 * When we use physical addressing for the TSB accesses, we patch the
2076 * first four instructions in the above sequence.
2077 */
2078
2079static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
2080{
2081 unsigned long high_bits, low_bits;
2082
2083 high_bits = (pa >> 32) & 0xffffffff;
2084 low_bits = (pa >> 0) & 0xffffffff;
2085
2086 while (start < end) {
2087 unsigned int *ia = (unsigned int *)(unsigned long)*start;
2088
2089 ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
2090 __asm__ __volatile__("flush %0" : : "r" (ia));
2091
2092 ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
2093 __asm__ __volatile__("flush %0" : : "r" (ia + 1));
2094
2095 ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
2096 __asm__ __volatile__("flush %0" : : "r" (ia + 2));
2097
2098 ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
2099 __asm__ __volatile__("flush %0" : : "r" (ia + 3));
2100
2101 start++;
2102 }
2103}
2104
2105static void ktsb_phys_patch(void)
2106{
2107 extern unsigned int __swapper_tsb_phys_patch;
2108 extern unsigned int __swapper_tsb_phys_patch_end;
2109 unsigned long ktsb_pa;
2110
2111 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2112 patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
2113 &__swapper_tsb_phys_patch_end, ktsb_pa);
2114#ifndef CONFIG_DEBUG_PAGEALLOC
2115 {
2116 extern unsigned int __swapper_4m_tsb_phys_patch;
2117 extern unsigned int __swapper_4m_tsb_phys_patch_end;
2118 ktsb_pa = (kern_base +
2119 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2120 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
2121 &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
2122 }
2123#endif
2124}
2125
2126static void __init sun4v_ktsb_init(void)
2127{
2128 unsigned long ktsb_pa;
2129
2130 /* First KTSB for PAGE_SIZE mappings. */
2131 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2132
2133 switch (PAGE_SIZE) {
2134 case 8 * 1024:
2135 default:
2136 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
2137 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
2138 break;
2139
2140 case 64 * 1024:
2141 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
2142 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
2143 break;
2144
2145 case 512 * 1024:
2146 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
2147 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
2148 break;
2149
2150 case 4 * 1024 * 1024:
2151 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
2152 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
2153 break;
2154 }
2155
2156 ktsb_descr[0].assoc = 1;
2157 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
2158 ktsb_descr[0].ctx_idx = 0;
2159 ktsb_descr[0].tsb_base = ktsb_pa;
2160 ktsb_descr[0].resv = 0;
2161
2162#ifndef CONFIG_DEBUG_PAGEALLOC
2163 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */
2164 ktsb_pa = (kern_base +
2165 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2166
2167 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
2168 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
2169 HV_PGSZ_MASK_256MB |
2170 HV_PGSZ_MASK_2GB |
2171 HV_PGSZ_MASK_16GB) &
2172 cpu_pgsz_mask);
2173 ktsb_descr[1].assoc = 1;
2174 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
2175 ktsb_descr[1].ctx_idx = 0;
2176 ktsb_descr[1].tsb_base = ktsb_pa;
2177 ktsb_descr[1].resv = 0;
2178#endif
2179}
2180
2181void sun4v_ktsb_register(void)
2182{
2183 unsigned long pa, ret;
2184
2185 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
2186
2187 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
2188 if (ret != 0) {
2189 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
2190 "errors with %lx\n", pa, ret);
2191 prom_halt();
2192 }
2193}
2194
2195static void __init sun4u_linear_pte_xor_finalize(void)
2196{
2197#ifndef CONFIG_DEBUG_PAGEALLOC
2198 /* This is where we would add Panther support for
2199 * 32MB and 256MB pages.
2200 */
2201#endif
2202}
2203
2204static void __init sun4v_linear_pte_xor_finalize(void)
2205{
2206 unsigned long pagecv_flag;
2207
2208 /* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
2209 * enables MCD error. Do not set bit 9 on M7 processor.
2210 */
2211 switch (sun4v_chip_type) {
2212 case SUN4V_CHIP_SPARC_M7:
2213 case SUN4V_CHIP_SPARC_M8:
2214 case SUN4V_CHIP_SPARC_SN:
2215 pagecv_flag = 0x00;
2216 break;
2217 default:
2218 pagecv_flag = _PAGE_CV_4V;
2219 break;
2220 }
2221#ifndef CONFIG_DEBUG_PAGEALLOC
2222 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
2223 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2224 PAGE_OFFSET;
2225 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2226 _PAGE_P_4V | _PAGE_W_4V);
2227 } else {
2228 kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2229 }
2230
2231 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2232 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2233 PAGE_OFFSET;
2234 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2235 _PAGE_P_4V | _PAGE_W_4V);
2236 } else {
2237 kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2238 }
2239
2240 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2241 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2242 PAGE_OFFSET;
2243 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2244 _PAGE_P_4V | _PAGE_W_4V);
2245 } else {
2246 kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2247 }
2248#endif
2249}
2250
2251/* paging_init() sets up the page tables */
2252
2253static unsigned long last_valid_pfn;
2254
2255static void sun4u_pgprot_init(void);
2256static void sun4v_pgprot_init(void);
2257
2258#define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U)
2259#define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V)
2260#define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2261#define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2262#define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2263#define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2264
2265/* We need to exclude reserved regions. This exclusion will include
2266 * vmlinux and initrd. To be more precise the initrd size could be used to
2267 * compute a new lower limit because it is freed later during initialization.
2268 */
2269static void __init reduce_memory(phys_addr_t limit_ram)
2270{
2271 limit_ram += memblock_reserved_size();
2272 memblock_enforce_memory_limit(limit_ram);
2273}
2274
2275void __init paging_init(void)
2276{
2277 unsigned long end_pfn, shift, phys_base;
2278 unsigned long real_end, i;
2279
2280 setup_page_offset();
2281
2282 /* These build time checkes make sure that the dcache_dirty_cpu()
2283 * page->flags usage will work.
2284 *
2285 * When a page gets marked as dcache-dirty, we store the
2286 * cpu number starting at bit 32 in the page->flags. Also,
2287 * functions like clear_dcache_dirty_cpu use the cpu mask
2288 * in 13-bit signed-immediate instruction fields.
2289 */
2290
2291 /*
2292 * Page flags must not reach into upper 32 bits that are used
2293 * for the cpu number
2294 */
2295 BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2296
2297 /*
2298 * The bit fields placed in the high range must not reach below
2299 * the 32 bit boundary. Otherwise we cannot place the cpu field
2300 * at the 32 bit boundary.
2301 */
2302 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2303 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2304
2305 BUILD_BUG_ON(NR_CPUS > 4096);
2306
2307 kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2308 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2309
2310 /* Invalidate both kernel TSBs. */
2311 memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2312#ifndef CONFIG_DEBUG_PAGEALLOC
2313 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2314#endif
2315
2316 /* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2317 * bit on M7 processor. This is a conflicting usage of the same
2318 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2319 * Detection error on all pages and this will lead to problems
2320 * later. Kernel does not run with MCD enabled and hence rest
2321 * of the required steps to fully configure memory corruption
2322 * detection are not taken. We need to ensure TTE.mcde is not
2323 * set on M7 processor. Compute the value of cacheability
2324 * flag for use later taking this into consideration.
2325 */
2326 switch (sun4v_chip_type) {
2327 case SUN4V_CHIP_SPARC_M7:
2328 case SUN4V_CHIP_SPARC_M8:
2329 case SUN4V_CHIP_SPARC_SN:
2330 page_cache4v_flag = _PAGE_CP_4V;
2331 break;
2332 default:
2333 page_cache4v_flag = _PAGE_CACHE_4V;
2334 break;
2335 }
2336
2337 if (tlb_type == hypervisor)
2338 sun4v_pgprot_init();
2339 else
2340 sun4u_pgprot_init();
2341
2342 if (tlb_type == cheetah_plus ||
2343 tlb_type == hypervisor) {
2344 tsb_phys_patch();
2345 ktsb_phys_patch();
2346 }
2347
2348 if (tlb_type == hypervisor)
2349 sun4v_patch_tlb_handlers();
2350
2351 /* Find available physical memory...
2352 *
2353 * Read it twice in order to work around a bug in openfirmware.
2354 * The call to grab this table itself can cause openfirmware to
2355 * allocate memory, which in turn can take away some space from
2356 * the list of available memory. Reading it twice makes sure
2357 * we really do get the final value.
2358 */
2359 read_obp_translations();
2360 read_obp_memory("reg", &pall[0], &pall_ents);
2361 read_obp_memory("available", &pavail[0], &pavail_ents);
2362 read_obp_memory("available", &pavail[0], &pavail_ents);
2363
2364 phys_base = 0xffffffffffffffffUL;
2365 for (i = 0; i < pavail_ents; i++) {
2366 phys_base = min(phys_base, pavail[i].phys_addr);
2367 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2368 }
2369
2370 memblock_reserve(kern_base, kern_size);
2371
2372 find_ramdisk(phys_base);
2373
2374 if (cmdline_memory_size)
2375 reduce_memory(cmdline_memory_size);
2376
2377 memblock_allow_resize();
2378 memblock_dump_all();
2379
2380 set_bit(0, mmu_context_bmap);
2381
2382 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2383
2384 real_end = (unsigned long)_end;
2385 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2386 printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2387 num_kernel_image_mappings);
2388
2389 /* Set kernel pgd to upper alias so physical page computations
2390 * work.
2391 */
2392 init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2393
2394 memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2395
2396 inherit_prom_mappings();
2397
2398 /* Ok, we can use our TLB miss and window trap handlers safely. */
2399 setup_tba();
2400
2401 __flush_tlb_all();
2402
2403 prom_build_devicetree();
2404 of_populate_present_mask();
2405#ifndef CONFIG_SMP
2406 of_fill_in_cpu_data();
2407#endif
2408
2409 if (tlb_type == hypervisor) {
2410 sun4v_mdesc_init();
2411 mdesc_populate_present_mask(cpu_all_mask);
2412#ifndef CONFIG_SMP
2413 mdesc_fill_in_cpu_data(cpu_all_mask);
2414#endif
2415 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2416
2417 sun4v_linear_pte_xor_finalize();
2418
2419 sun4v_ktsb_init();
2420 sun4v_ktsb_register();
2421 } else {
2422 unsigned long impl, ver;
2423
2424 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2425 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2426
2427 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2428 impl = ((ver >> 32) & 0xffff);
2429 if (impl == PANTHER_IMPL)
2430 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2431 HV_PGSZ_MASK_256MB);
2432
2433 sun4u_linear_pte_xor_finalize();
2434 }
2435
2436 /* Flush the TLBs and the 4M TSB so that the updated linear
2437 * pte XOR settings are realized for all mappings.
2438 */
2439 __flush_tlb_all();
2440#ifndef CONFIG_DEBUG_PAGEALLOC
2441 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2442#endif
2443 __flush_tlb_all();
2444
2445 /* Setup bootmem... */
2446 last_valid_pfn = end_pfn = bootmem_init(phys_base);
2447
2448 kernel_physical_mapping_init();
2449
2450 {
2451 unsigned long max_zone_pfns[MAX_NR_ZONES];
2452
2453 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2454
2455 max_zone_pfns[ZONE_NORMAL] = end_pfn;
2456
2457 free_area_init(max_zone_pfns);
2458 }
2459
2460 printk("Booting Linux...\n");
2461}
2462
2463int page_in_phys_avail(unsigned long paddr)
2464{
2465 int i;
2466
2467 paddr &= PAGE_MASK;
2468
2469 for (i = 0; i < pavail_ents; i++) {
2470 unsigned long start, end;
2471
2472 start = pavail[i].phys_addr;
2473 end = start + pavail[i].reg_size;
2474
2475 if (paddr >= start && paddr < end)
2476 return 1;
2477 }
2478 if (paddr >= kern_base && paddr < (kern_base + kern_size))
2479 return 1;
2480#ifdef CONFIG_BLK_DEV_INITRD
2481 if (paddr >= __pa(initrd_start) &&
2482 paddr < __pa(PAGE_ALIGN(initrd_end)))
2483 return 1;
2484#endif
2485
2486 return 0;
2487}
2488
2489static void __init register_page_bootmem_info(void)
2490{
2491#ifdef CONFIG_NUMA
2492 int i;
2493
2494 for_each_online_node(i)
2495 if (NODE_DATA(i)->node_spanned_pages)
2496 register_page_bootmem_info_node(NODE_DATA(i));
2497#endif
2498}
2499void __init mem_init(void)
2500{
2501 high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2502
2503 memblock_free_all();
2504
2505 /*
2506 * Must be done after boot memory is put on freelist, because here we
2507 * might set fields in deferred struct pages that have not yet been
2508 * initialized, and memblock_free_all() initializes all the reserved
2509 * deferred pages for us.
2510 */
2511 register_page_bootmem_info();
2512
2513 /*
2514 * Set up the zero page, mark it reserved, so that page count
2515 * is not manipulated when freeing the page from user ptes.
2516 */
2517 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2518 if (mem_map_zero == NULL) {
2519 prom_printf("paging_init: Cannot alloc zero page.\n");
2520 prom_halt();
2521 }
2522 mark_page_reserved(mem_map_zero);
2523
2524
2525 if (tlb_type == cheetah || tlb_type == cheetah_plus)
2526 cheetah_ecache_flush_init();
2527}
2528
2529void free_initmem(void)
2530{
2531 unsigned long addr, initend;
2532 int do_free = 1;
2533
2534 /* If the physical memory maps were trimmed by kernel command
2535 * line options, don't even try freeing this initmem stuff up.
2536 * The kernel image could have been in the trimmed out region
2537 * and if so the freeing below will free invalid page structs.
2538 */
2539 if (cmdline_memory_size)
2540 do_free = 0;
2541
2542 /*
2543 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2544 */
2545 addr = PAGE_ALIGN((unsigned long)(__init_begin));
2546 initend = (unsigned long)(__init_end) & PAGE_MASK;
2547 for (; addr < initend; addr += PAGE_SIZE) {
2548 unsigned long page;
2549
2550 page = (addr +
2551 ((unsigned long) __va(kern_base)) -
2552 ((unsigned long) KERNBASE));
2553 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2554
2555 if (do_free)
2556 free_reserved_page(virt_to_page(page));
2557 }
2558}
2559
2560pgprot_t PAGE_KERNEL __read_mostly;
2561EXPORT_SYMBOL(PAGE_KERNEL);
2562
2563pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2564pgprot_t PAGE_COPY __read_mostly;
2565
2566pgprot_t PAGE_SHARED __read_mostly;
2567EXPORT_SYMBOL(PAGE_SHARED);
2568
2569unsigned long pg_iobits __read_mostly;
2570
2571unsigned long _PAGE_IE __read_mostly;
2572EXPORT_SYMBOL(_PAGE_IE);
2573
2574unsigned long _PAGE_E __read_mostly;
2575EXPORT_SYMBOL(_PAGE_E);
2576
2577unsigned long _PAGE_CACHE __read_mostly;
2578EXPORT_SYMBOL(_PAGE_CACHE);
2579
2580#ifdef CONFIG_SPARSEMEM_VMEMMAP
2581int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2582 int node, struct vmem_altmap *altmap)
2583{
2584 unsigned long pte_base;
2585
2586 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2587 _PAGE_CP_4U | _PAGE_CV_4U |
2588 _PAGE_P_4U | _PAGE_W_4U);
2589 if (tlb_type == hypervisor)
2590 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2591 page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2592
2593 pte_base |= _PAGE_PMD_HUGE;
2594
2595 vstart = vstart & PMD_MASK;
2596 vend = ALIGN(vend, PMD_SIZE);
2597 for (; vstart < vend; vstart += PMD_SIZE) {
2598 pgd_t *pgd = vmemmap_pgd_populate(vstart, node);
2599 unsigned long pte;
2600 p4d_t *p4d;
2601 pud_t *pud;
2602 pmd_t *pmd;
2603
2604 if (!pgd)
2605 return -ENOMEM;
2606
2607 p4d = vmemmap_p4d_populate(pgd, vstart, node);
2608 if (!p4d)
2609 return -ENOMEM;
2610
2611 pud = vmemmap_pud_populate(p4d, vstart, node);
2612 if (!pud)
2613 return -ENOMEM;
2614
2615 pmd = pmd_offset(pud, vstart);
2616 pte = pmd_val(*pmd);
2617 if (!(pte & _PAGE_VALID)) {
2618 void *block = vmemmap_alloc_block(PMD_SIZE, node);
2619
2620 if (!block)
2621 return -ENOMEM;
2622
2623 pmd_val(*pmd) = pte_base | __pa(block);
2624 }
2625 }
2626
2627 return 0;
2628}
2629
2630void vmemmap_free(unsigned long start, unsigned long end,
2631 struct vmem_altmap *altmap)
2632{
2633}
2634#endif /* CONFIG_SPARSEMEM_VMEMMAP */
2635
2636/* These are actually filled in at boot time by sun4{u,v}_pgprot_init() */
2637static pgprot_t protection_map[16] __ro_after_init;
2638
2639static void prot_init_common(unsigned long page_none,
2640 unsigned long page_shared,
2641 unsigned long page_copy,
2642 unsigned long page_readonly,
2643 unsigned long page_exec_bit)
2644{
2645 PAGE_COPY = __pgprot(page_copy);
2646 PAGE_SHARED = __pgprot(page_shared);
2647
2648 protection_map[0x0] = __pgprot(page_none);
2649 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2650 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2651 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2652 protection_map[0x4] = __pgprot(page_readonly);
2653 protection_map[0x5] = __pgprot(page_readonly);
2654 protection_map[0x6] = __pgprot(page_copy);
2655 protection_map[0x7] = __pgprot(page_copy);
2656 protection_map[0x8] = __pgprot(page_none);
2657 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2658 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2659 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2660 protection_map[0xc] = __pgprot(page_readonly);
2661 protection_map[0xd] = __pgprot(page_readonly);
2662 protection_map[0xe] = __pgprot(page_shared);
2663 protection_map[0xf] = __pgprot(page_shared);
2664}
2665
2666static void __init sun4u_pgprot_init(void)
2667{
2668 unsigned long page_none, page_shared, page_copy, page_readonly;
2669 unsigned long page_exec_bit;
2670 int i;
2671
2672 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2673 _PAGE_CACHE_4U | _PAGE_P_4U |
2674 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2675 _PAGE_EXEC_4U);
2676 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2677 _PAGE_CACHE_4U | _PAGE_P_4U |
2678 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2679 _PAGE_EXEC_4U | _PAGE_L_4U);
2680
2681 _PAGE_IE = _PAGE_IE_4U;
2682 _PAGE_E = _PAGE_E_4U;
2683 _PAGE_CACHE = _PAGE_CACHE_4U;
2684
2685 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2686 __ACCESS_BITS_4U | _PAGE_E_4U);
2687
2688#ifdef CONFIG_DEBUG_PAGEALLOC
2689 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2690#else
2691 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2692 PAGE_OFFSET;
2693#endif
2694 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2695 _PAGE_P_4U | _PAGE_W_4U);
2696
2697 for (i = 1; i < 4; i++)
2698 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2699
2700 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2701 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2702 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2703
2704
2705 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2706 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2707 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2708 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2709 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2710 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2711 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2712
2713 page_exec_bit = _PAGE_EXEC_4U;
2714
2715 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2716 page_exec_bit);
2717}
2718
2719static void __init sun4v_pgprot_init(void)
2720{
2721 unsigned long page_none, page_shared, page_copy, page_readonly;
2722 unsigned long page_exec_bit;
2723 int i;
2724
2725 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2726 page_cache4v_flag | _PAGE_P_4V |
2727 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2728 _PAGE_EXEC_4V);
2729 PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2730
2731 _PAGE_IE = _PAGE_IE_4V;
2732 _PAGE_E = _PAGE_E_4V;
2733 _PAGE_CACHE = page_cache4v_flag;
2734
2735#ifdef CONFIG_DEBUG_PAGEALLOC
2736 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2737#else
2738 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2739 PAGE_OFFSET;
2740#endif
2741 kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2742 _PAGE_W_4V);
2743
2744 for (i = 1; i < 4; i++)
2745 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2746
2747 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2748 __ACCESS_BITS_4V | _PAGE_E_4V);
2749
2750 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2751 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2752 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2753 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2754
2755 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2756 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2757 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2758 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2759 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2760 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2761 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2762
2763 page_exec_bit = _PAGE_EXEC_4V;
2764
2765 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2766 page_exec_bit);
2767}
2768
2769unsigned long pte_sz_bits(unsigned long sz)
2770{
2771 if (tlb_type == hypervisor) {
2772 switch (sz) {
2773 case 8 * 1024:
2774 default:
2775 return _PAGE_SZ8K_4V;
2776 case 64 * 1024:
2777 return _PAGE_SZ64K_4V;
2778 case 512 * 1024:
2779 return _PAGE_SZ512K_4V;
2780 case 4 * 1024 * 1024:
2781 return _PAGE_SZ4MB_4V;
2782 }
2783 } else {
2784 switch (sz) {
2785 case 8 * 1024:
2786 default:
2787 return _PAGE_SZ8K_4U;
2788 case 64 * 1024:
2789 return _PAGE_SZ64K_4U;
2790 case 512 * 1024:
2791 return _PAGE_SZ512K_4U;
2792 case 4 * 1024 * 1024:
2793 return _PAGE_SZ4MB_4U;
2794 }
2795 }
2796}
2797
2798pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2799{
2800 pte_t pte;
2801
2802 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot));
2803 pte_val(pte) |= (((unsigned long)space) << 32);
2804 pte_val(pte) |= pte_sz_bits(page_size);
2805
2806 return pte;
2807}
2808
2809static unsigned long kern_large_tte(unsigned long paddr)
2810{
2811 unsigned long val;
2812
2813 val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2814 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2815 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2816 if (tlb_type == hypervisor)
2817 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2818 page_cache4v_flag | _PAGE_P_4V |
2819 _PAGE_EXEC_4V | _PAGE_W_4V);
2820
2821 return val | paddr;
2822}
2823
2824/* If not locked, zap it. */
2825void __flush_tlb_all(void)
2826{
2827 unsigned long pstate;
2828 int i;
2829
2830 __asm__ __volatile__("flushw\n\t"
2831 "rdpr %%pstate, %0\n\t"
2832 "wrpr %0, %1, %%pstate"
2833 : "=r" (pstate)
2834 : "i" (PSTATE_IE));
2835 if (tlb_type == hypervisor) {
2836 sun4v_mmu_demap_all();
2837 } else if (tlb_type == spitfire) {
2838 for (i = 0; i < 64; i++) {
2839 /* Spitfire Errata #32 workaround */
2840 /* NOTE: Always runs on spitfire, so no
2841 * cheetah+ page size encodings.
2842 */
2843 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2844 "flush %%g6"
2845 : /* No outputs */
2846 : "r" (0),
2847 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2848
2849 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2850 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2851 "membar #Sync"
2852 : /* no outputs */
2853 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2854 spitfire_put_dtlb_data(i, 0x0UL);
2855 }
2856
2857 /* Spitfire Errata #32 workaround */
2858 /* NOTE: Always runs on spitfire, so no
2859 * cheetah+ page size encodings.
2860 */
2861 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2862 "flush %%g6"
2863 : /* No outputs */
2864 : "r" (0),
2865 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2866
2867 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2868 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2869 "membar #Sync"
2870 : /* no outputs */
2871 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2872 spitfire_put_itlb_data(i, 0x0UL);
2873 }
2874 }
2875 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2876 cheetah_flush_dtlb_all();
2877 cheetah_flush_itlb_all();
2878 }
2879 __asm__ __volatile__("wrpr %0, 0, %%pstate"
2880 : : "r" (pstate));
2881}
2882
2883pte_t *pte_alloc_one_kernel(struct mm_struct *mm)
2884{
2885 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2886 pte_t *pte = NULL;
2887
2888 if (page)
2889 pte = (pte_t *) page_address(page);
2890
2891 return pte;
2892}
2893
2894pgtable_t pte_alloc_one(struct mm_struct *mm)
2895{
2896 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2897 if (!page)
2898 return NULL;
2899 if (!pgtable_pte_page_ctor(page)) {
2900 __free_page(page);
2901 return NULL;
2902 }
2903 return (pte_t *) page_address(page);
2904}
2905
2906void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2907{
2908 free_page((unsigned long)pte);
2909}
2910
2911static void __pte_free(pgtable_t pte)
2912{
2913 struct page *page = virt_to_page(pte);
2914
2915 pgtable_pte_page_dtor(page);
2916 __free_page(page);
2917}
2918
2919void pte_free(struct mm_struct *mm, pgtable_t pte)
2920{
2921 __pte_free(pte);
2922}
2923
2924void pgtable_free(void *table, bool is_page)
2925{
2926 if (is_page)
2927 __pte_free(table);
2928 else
2929 kmem_cache_free(pgtable_cache, table);
2930}
2931
2932#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2933void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2934 pmd_t *pmd)
2935{
2936 unsigned long pte, flags;
2937 struct mm_struct *mm;
2938 pmd_t entry = *pmd;
2939
2940 if (!pmd_large(entry) || !pmd_young(entry))
2941 return;
2942
2943 pte = pmd_val(entry);
2944
2945 /* Don't insert a non-valid PMD into the TSB, we'll deadlock. */
2946 if (!(pte & _PAGE_VALID))
2947 return;
2948
2949 /* We are fabricating 8MB pages using 4MB real hw pages. */
2950 pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2951
2952 mm = vma->vm_mm;
2953
2954 spin_lock_irqsave(&mm->context.lock, flags);
2955
2956 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2957 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2958 addr, pte);
2959
2960 spin_unlock_irqrestore(&mm->context.lock, flags);
2961}
2962#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2963
2964#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2965static void context_reload(void *__data)
2966{
2967 struct mm_struct *mm = __data;
2968
2969 if (mm == current->mm)
2970 load_secondary_context(mm);
2971}
2972
2973void hugetlb_setup(struct pt_regs *regs)
2974{
2975 struct mm_struct *mm = current->mm;
2976 struct tsb_config *tp;
2977
2978 if (faulthandler_disabled() || !mm) {
2979 const struct exception_table_entry *entry;
2980
2981 entry = search_exception_tables(regs->tpc);
2982 if (entry) {
2983 regs->tpc = entry->fixup;
2984 regs->tnpc = regs->tpc + 4;
2985 return;
2986 }
2987 pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2988 die_if_kernel("HugeTSB in atomic", regs);
2989 }
2990
2991 tp = &mm->context.tsb_block[MM_TSB_HUGE];
2992 if (likely(tp->tsb == NULL))
2993 tsb_grow(mm, MM_TSB_HUGE, 0);
2994
2995 tsb_context_switch(mm);
2996 smp_tsb_sync(mm);
2997
2998 /* On UltraSPARC-III+ and later, configure the second half of
2999 * the Data-TLB for huge pages.
3000 */
3001 if (tlb_type == cheetah_plus) {
3002 bool need_context_reload = false;
3003 unsigned long ctx;
3004
3005 spin_lock_irq(&ctx_alloc_lock);
3006 ctx = mm->context.sparc64_ctx_val;
3007 ctx &= ~CTX_PGSZ_MASK;
3008 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
3009 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
3010
3011 if (ctx != mm->context.sparc64_ctx_val) {
3012 /* When changing the page size fields, we
3013 * must perform a context flush so that no
3014 * stale entries match. This flush must
3015 * occur with the original context register
3016 * settings.
3017 */
3018 do_flush_tlb_mm(mm);
3019
3020 /* Reload the context register of all processors
3021 * also executing in this address space.
3022 */
3023 mm->context.sparc64_ctx_val = ctx;
3024 need_context_reload = true;
3025 }
3026 spin_unlock_irq(&ctx_alloc_lock);
3027
3028 if (need_context_reload)
3029 on_each_cpu(context_reload, mm, 0);
3030 }
3031}
3032#endif
3033
3034static struct resource code_resource = {
3035 .name = "Kernel code",
3036 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3037};
3038
3039static struct resource data_resource = {
3040 .name = "Kernel data",
3041 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3042};
3043
3044static struct resource bss_resource = {
3045 .name = "Kernel bss",
3046 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3047};
3048
3049static inline resource_size_t compute_kern_paddr(void *addr)
3050{
3051 return (resource_size_t) (addr - KERNBASE + kern_base);
3052}
3053
3054static void __init kernel_lds_init(void)
3055{
3056 code_resource.start = compute_kern_paddr(_text);
3057 code_resource.end = compute_kern_paddr(_etext - 1);
3058 data_resource.start = compute_kern_paddr(_etext);
3059 data_resource.end = compute_kern_paddr(_edata - 1);
3060 bss_resource.start = compute_kern_paddr(__bss_start);
3061 bss_resource.end = compute_kern_paddr(_end - 1);
3062}
3063
3064static int __init report_memory(void)
3065{
3066 int i;
3067 struct resource *res;
3068
3069 kernel_lds_init();
3070
3071 for (i = 0; i < pavail_ents; i++) {
3072 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
3073
3074 if (!res) {
3075 pr_warn("Failed to allocate source.\n");
3076 break;
3077 }
3078
3079 res->name = "System RAM";
3080 res->start = pavail[i].phys_addr;
3081 res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
3082 res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
3083
3084 if (insert_resource(&iomem_resource, res) < 0) {
3085 pr_warn("Resource insertion failed.\n");
3086 break;
3087 }
3088
3089 insert_resource(res, &code_resource);
3090 insert_resource(res, &data_resource);
3091 insert_resource(res, &bss_resource);
3092 }
3093
3094 return 0;
3095}
3096arch_initcall(report_memory);
3097
3098#ifdef CONFIG_SMP
3099#define do_flush_tlb_kernel_range smp_flush_tlb_kernel_range
3100#else
3101#define do_flush_tlb_kernel_range __flush_tlb_kernel_range
3102#endif
3103
3104void flush_tlb_kernel_range(unsigned long start, unsigned long end)
3105{
3106 if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
3107 if (start < LOW_OBP_ADDRESS) {
3108 flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
3109 do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
3110 }
3111 if (end > HI_OBP_ADDRESS) {
3112 flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
3113 do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
3114 }
3115 } else {
3116 flush_tsb_kernel_range(start, end);
3117 do_flush_tlb_kernel_range(start, end);
3118 }
3119}
3120
3121void copy_user_highpage(struct page *to, struct page *from,
3122 unsigned long vaddr, struct vm_area_struct *vma)
3123{
3124 char *vfrom, *vto;
3125
3126 vfrom = kmap_atomic(from);
3127 vto = kmap_atomic(to);
3128 copy_user_page(vto, vfrom, vaddr, to);
3129 kunmap_atomic(vto);
3130 kunmap_atomic(vfrom);
3131
3132 /* If this page has ADI enabled, copy over any ADI tags
3133 * as well
3134 */
3135 if (vma->vm_flags & VM_SPARC_ADI) {
3136 unsigned long pfrom, pto, i, adi_tag;
3137
3138 pfrom = page_to_phys(from);
3139 pto = page_to_phys(to);
3140
3141 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3142 asm volatile("ldxa [%1] %2, %0\n\t"
3143 : "=r" (adi_tag)
3144 : "r" (i), "i" (ASI_MCD_REAL));
3145 asm volatile("stxa %0, [%1] %2\n\t"
3146 :
3147 : "r" (adi_tag), "r" (pto),
3148 "i" (ASI_MCD_REAL));
3149 pto += adi_blksize();
3150 }
3151 asm volatile("membar #Sync\n\t");
3152 }
3153}
3154EXPORT_SYMBOL(copy_user_highpage);
3155
3156void copy_highpage(struct page *to, struct page *from)
3157{
3158 char *vfrom, *vto;
3159
3160 vfrom = kmap_atomic(from);
3161 vto = kmap_atomic(to);
3162 copy_page(vto, vfrom);
3163 kunmap_atomic(vto);
3164 kunmap_atomic(vfrom);
3165
3166 /* If this platform is ADI enabled, copy any ADI tags
3167 * as well
3168 */
3169 if (adi_capable()) {
3170 unsigned long pfrom, pto, i, adi_tag;
3171
3172 pfrom = page_to_phys(from);
3173 pto = page_to_phys(to);
3174
3175 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3176 asm volatile("ldxa [%1] %2, %0\n\t"
3177 : "=r" (adi_tag)
3178 : "r" (i), "i" (ASI_MCD_REAL));
3179 asm volatile("stxa %0, [%1] %2\n\t"
3180 :
3181 : "r" (adi_tag), "r" (pto),
3182 "i" (ASI_MCD_REAL));
3183 pto += adi_blksize();
3184 }
3185 asm volatile("membar #Sync\n\t");
3186 }
3187}
3188EXPORT_SYMBOL(copy_highpage);
3189
3190pgprot_t vm_get_page_prot(unsigned long vm_flags)
3191{
3192 unsigned long prot = pgprot_val(protection_map[vm_flags &
3193 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]);
3194
3195 if (vm_flags & VM_SPARC_ADI)
3196 prot |= _PAGE_MCD_4V;
3197
3198 return __pgprot(prot);
3199}
3200EXPORT_SYMBOL(vm_get_page_prot);
1/*
2 * arch/sparc64/mm/init.c
3 *
4 * Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
5 * Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
6 */
7
8#include <linux/module.h>
9#include <linux/kernel.h>
10#include <linux/sched.h>
11#include <linux/string.h>
12#include <linux/init.h>
13#include <linux/bootmem.h>
14#include <linux/mm.h>
15#include <linux/hugetlb.h>
16#include <linux/initrd.h>
17#include <linux/swap.h>
18#include <linux/pagemap.h>
19#include <linux/poison.h>
20#include <linux/fs.h>
21#include <linux/seq_file.h>
22#include <linux/kprobes.h>
23#include <linux/cache.h>
24#include <linux/sort.h>
25#include <linux/percpu.h>
26#include <linux/memblock.h>
27#include <linux/mmzone.h>
28#include <linux/gfp.h>
29
30#include <asm/head.h>
31#include <asm/page.h>
32#include <asm/pgalloc.h>
33#include <asm/pgtable.h>
34#include <asm/oplib.h>
35#include <asm/iommu.h>
36#include <asm/io.h>
37#include <asm/uaccess.h>
38#include <asm/mmu_context.h>
39#include <asm/tlbflush.h>
40#include <asm/dma.h>
41#include <asm/starfire.h>
42#include <asm/tlb.h>
43#include <asm/spitfire.h>
44#include <asm/sections.h>
45#include <asm/tsb.h>
46#include <asm/hypervisor.h>
47#include <asm/prom.h>
48#include <asm/mdesc.h>
49#include <asm/cpudata.h>
50#include <asm/irq.h>
51
52#include "init_64.h"
53
54unsigned long kern_linear_pte_xor[4] __read_mostly;
55
56/* A bitmap, two bits for every 256MB of physical memory. These two
57 * bits determine what page size we use for kernel linear
58 * translations. They form an index into kern_linear_pte_xor[]. The
59 * value in the indexed slot is XOR'd with the TLB miss virtual
60 * address to form the resulting TTE. The mapping is:
61 *
62 * 0 ==> 4MB
63 * 1 ==> 256MB
64 * 2 ==> 2GB
65 * 3 ==> 16GB
66 *
67 * All sun4v chips support 256MB pages. Only SPARC-T4 and later
68 * support 2GB pages, and hopefully future cpus will support the 16GB
69 * pages as well. For slots 2 and 3, we encode a 256MB TTE xor there
70 * if these larger page sizes are not supported by the cpu.
71 *
72 * It would be nice to determine this from the machine description
73 * 'cpu' properties, but we need to have this table setup before the
74 * MDESC is initialized.
75 */
76unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
77
78#ifndef CONFIG_DEBUG_PAGEALLOC
79/* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
80 * Space is allocated for this right after the trap table in
81 * arch/sparc64/kernel/head.S
82 */
83extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
84#endif
85
86static unsigned long cpu_pgsz_mask;
87
88#define MAX_BANKS 32
89
90static struct linux_prom64_registers pavail[MAX_BANKS];
91static int pavail_ents;
92
93static int cmp_p64(const void *a, const void *b)
94{
95 const struct linux_prom64_registers *x = a, *y = b;
96
97 if (x->phys_addr > y->phys_addr)
98 return 1;
99 if (x->phys_addr < y->phys_addr)
100 return -1;
101 return 0;
102}
103
104static void __init read_obp_memory(const char *property,
105 struct linux_prom64_registers *regs,
106 int *num_ents)
107{
108 phandle node = prom_finddevice("/memory");
109 int prop_size = prom_getproplen(node, property);
110 int ents, ret, i;
111
112 ents = prop_size / sizeof(struct linux_prom64_registers);
113 if (ents > MAX_BANKS) {
114 prom_printf("The machine has more %s property entries than "
115 "this kernel can support (%d).\n",
116 property, MAX_BANKS);
117 prom_halt();
118 }
119
120 ret = prom_getproperty(node, property, (char *) regs, prop_size);
121 if (ret == -1) {
122 prom_printf("Couldn't get %s property from /memory.\n",
123 property);
124 prom_halt();
125 }
126
127 /* Sanitize what we got from the firmware, by page aligning
128 * everything.
129 */
130 for (i = 0; i < ents; i++) {
131 unsigned long base, size;
132
133 base = regs[i].phys_addr;
134 size = regs[i].reg_size;
135
136 size &= PAGE_MASK;
137 if (base & ~PAGE_MASK) {
138 unsigned long new_base = PAGE_ALIGN(base);
139
140 size -= new_base - base;
141 if ((long) size < 0L)
142 size = 0UL;
143 base = new_base;
144 }
145 if (size == 0UL) {
146 /* If it is empty, simply get rid of it.
147 * This simplifies the logic of the other
148 * functions that process these arrays.
149 */
150 memmove(®s[i], ®s[i + 1],
151 (ents - i - 1) * sizeof(regs[0]));
152 i--;
153 ents--;
154 continue;
155 }
156 regs[i].phys_addr = base;
157 regs[i].reg_size = size;
158 }
159
160 *num_ents = ents;
161
162 sort(regs, ents, sizeof(struct linux_prom64_registers),
163 cmp_p64, NULL);
164}
165
166unsigned long sparc64_valid_addr_bitmap[VALID_ADDR_BITMAP_BYTES /
167 sizeof(unsigned long)];
168EXPORT_SYMBOL(sparc64_valid_addr_bitmap);
169
170/* Kernel physical address base and size in bytes. */
171unsigned long kern_base __read_mostly;
172unsigned long kern_size __read_mostly;
173
174/* Initial ramdisk setup */
175extern unsigned long sparc_ramdisk_image64;
176extern unsigned int sparc_ramdisk_image;
177extern unsigned int sparc_ramdisk_size;
178
179struct page *mem_map_zero __read_mostly;
180EXPORT_SYMBOL(mem_map_zero);
181
182unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
183
184unsigned long sparc64_kern_pri_context __read_mostly;
185unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
186unsigned long sparc64_kern_sec_context __read_mostly;
187
188int num_kernel_image_mappings;
189
190#ifdef CONFIG_DEBUG_DCFLUSH
191atomic_t dcpage_flushes = ATOMIC_INIT(0);
192#ifdef CONFIG_SMP
193atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
194#endif
195#endif
196
197inline void flush_dcache_page_impl(struct page *page)
198{
199 BUG_ON(tlb_type == hypervisor);
200#ifdef CONFIG_DEBUG_DCFLUSH
201 atomic_inc(&dcpage_flushes);
202#endif
203
204#ifdef DCACHE_ALIASING_POSSIBLE
205 __flush_dcache_page(page_address(page),
206 ((tlb_type == spitfire) &&
207 page_mapping(page) != NULL));
208#else
209 if (page_mapping(page) != NULL &&
210 tlb_type == spitfire)
211 __flush_icache_page(__pa(page_address(page)));
212#endif
213}
214
215#define PG_dcache_dirty PG_arch_1
216#define PG_dcache_cpu_shift 32UL
217#define PG_dcache_cpu_mask \
218 ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
219
220#define dcache_dirty_cpu(page) \
221 (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
222
223static inline void set_dcache_dirty(struct page *page, int this_cpu)
224{
225 unsigned long mask = this_cpu;
226 unsigned long non_cpu_bits;
227
228 non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
229 mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
230
231 __asm__ __volatile__("1:\n\t"
232 "ldx [%2], %%g7\n\t"
233 "and %%g7, %1, %%g1\n\t"
234 "or %%g1, %0, %%g1\n\t"
235 "casx [%2], %%g7, %%g1\n\t"
236 "cmp %%g7, %%g1\n\t"
237 "bne,pn %%xcc, 1b\n\t"
238 " nop"
239 : /* no outputs */
240 : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
241 : "g1", "g7");
242}
243
244static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
245{
246 unsigned long mask = (1UL << PG_dcache_dirty);
247
248 __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
249 "1:\n\t"
250 "ldx [%2], %%g7\n\t"
251 "srlx %%g7, %4, %%g1\n\t"
252 "and %%g1, %3, %%g1\n\t"
253 "cmp %%g1, %0\n\t"
254 "bne,pn %%icc, 2f\n\t"
255 " andn %%g7, %1, %%g1\n\t"
256 "casx [%2], %%g7, %%g1\n\t"
257 "cmp %%g7, %%g1\n\t"
258 "bne,pn %%xcc, 1b\n\t"
259 " nop\n"
260 "2:"
261 : /* no outputs */
262 : "r" (cpu), "r" (mask), "r" (&page->flags),
263 "i" (PG_dcache_cpu_mask),
264 "i" (PG_dcache_cpu_shift)
265 : "g1", "g7");
266}
267
268static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
269{
270 unsigned long tsb_addr = (unsigned long) ent;
271
272 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
273 tsb_addr = __pa(tsb_addr);
274
275 __tsb_insert(tsb_addr, tag, pte);
276}
277
278unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
279
280static void flush_dcache(unsigned long pfn)
281{
282 struct page *page;
283
284 page = pfn_to_page(pfn);
285 if (page) {
286 unsigned long pg_flags;
287
288 pg_flags = page->flags;
289 if (pg_flags & (1UL << PG_dcache_dirty)) {
290 int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
291 PG_dcache_cpu_mask);
292 int this_cpu = get_cpu();
293
294 /* This is just to optimize away some function calls
295 * in the SMP case.
296 */
297 if (cpu == this_cpu)
298 flush_dcache_page_impl(page);
299 else
300 smp_flush_dcache_page_impl(page, cpu);
301
302 clear_dcache_dirty_cpu(page, cpu);
303
304 put_cpu();
305 }
306 }
307}
308
309/* mm->context.lock must be held */
310static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
311 unsigned long tsb_hash_shift, unsigned long address,
312 unsigned long tte)
313{
314 struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
315 unsigned long tag;
316
317 if (unlikely(!tsb))
318 return;
319
320 tsb += ((address >> tsb_hash_shift) &
321 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
322 tag = (address >> 22UL);
323 tsb_insert(tsb, tag, tte);
324}
325
326#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
327static inline bool is_hugetlb_pte(pte_t pte)
328{
329 if ((tlb_type == hypervisor &&
330 (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
331 (tlb_type != hypervisor &&
332 (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U))
333 return true;
334 return false;
335}
336#endif
337
338void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
339{
340 struct mm_struct *mm;
341 unsigned long flags;
342 pte_t pte = *ptep;
343
344 if (tlb_type != hypervisor) {
345 unsigned long pfn = pte_pfn(pte);
346
347 if (pfn_valid(pfn))
348 flush_dcache(pfn);
349 }
350
351 mm = vma->vm_mm;
352
353 spin_lock_irqsave(&mm->context.lock, flags);
354
355#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
356 if (mm->context.huge_pte_count && is_hugetlb_pte(pte))
357 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
358 address, pte_val(pte));
359 else
360#endif
361 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
362 address, pte_val(pte));
363
364 spin_unlock_irqrestore(&mm->context.lock, flags);
365}
366
367void flush_dcache_page(struct page *page)
368{
369 struct address_space *mapping;
370 int this_cpu;
371
372 if (tlb_type == hypervisor)
373 return;
374
375 /* Do not bother with the expensive D-cache flush if it
376 * is merely the zero page. The 'bigcore' testcase in GDB
377 * causes this case to run millions of times.
378 */
379 if (page == ZERO_PAGE(0))
380 return;
381
382 this_cpu = get_cpu();
383
384 mapping = page_mapping(page);
385 if (mapping && !mapping_mapped(mapping)) {
386 int dirty = test_bit(PG_dcache_dirty, &page->flags);
387 if (dirty) {
388 int dirty_cpu = dcache_dirty_cpu(page);
389
390 if (dirty_cpu == this_cpu)
391 goto out;
392 smp_flush_dcache_page_impl(page, dirty_cpu);
393 }
394 set_dcache_dirty(page, this_cpu);
395 } else {
396 /* We could delay the flush for the !page_mapping
397 * case too. But that case is for exec env/arg
398 * pages and those are %99 certainly going to get
399 * faulted into the tlb (and thus flushed) anyways.
400 */
401 flush_dcache_page_impl(page);
402 }
403
404out:
405 put_cpu();
406}
407EXPORT_SYMBOL(flush_dcache_page);
408
409void __kprobes flush_icache_range(unsigned long start, unsigned long end)
410{
411 /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
412 if (tlb_type == spitfire) {
413 unsigned long kaddr;
414
415 /* This code only runs on Spitfire cpus so this is
416 * why we can assume _PAGE_PADDR_4U.
417 */
418 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
419 unsigned long paddr, mask = _PAGE_PADDR_4U;
420
421 if (kaddr >= PAGE_OFFSET)
422 paddr = kaddr & mask;
423 else {
424 pgd_t *pgdp = pgd_offset_k(kaddr);
425 pud_t *pudp = pud_offset(pgdp, kaddr);
426 pmd_t *pmdp = pmd_offset(pudp, kaddr);
427 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
428
429 paddr = pte_val(*ptep) & mask;
430 }
431 __flush_icache_page(paddr);
432 }
433 }
434}
435EXPORT_SYMBOL(flush_icache_range);
436
437void mmu_info(struct seq_file *m)
438{
439 static const char *pgsz_strings[] = {
440 "8K", "64K", "512K", "4MB", "32MB",
441 "256MB", "2GB", "16GB",
442 };
443 int i, printed;
444
445 if (tlb_type == cheetah)
446 seq_printf(m, "MMU Type\t: Cheetah\n");
447 else if (tlb_type == cheetah_plus)
448 seq_printf(m, "MMU Type\t: Cheetah+\n");
449 else if (tlb_type == spitfire)
450 seq_printf(m, "MMU Type\t: Spitfire\n");
451 else if (tlb_type == hypervisor)
452 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
453 else
454 seq_printf(m, "MMU Type\t: ???\n");
455
456 seq_printf(m, "MMU PGSZs\t: ");
457 printed = 0;
458 for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
459 if (cpu_pgsz_mask & (1UL << i)) {
460 seq_printf(m, "%s%s",
461 printed ? "," : "", pgsz_strings[i]);
462 printed++;
463 }
464 }
465 seq_putc(m, '\n');
466
467#ifdef CONFIG_DEBUG_DCFLUSH
468 seq_printf(m, "DCPageFlushes\t: %d\n",
469 atomic_read(&dcpage_flushes));
470#ifdef CONFIG_SMP
471 seq_printf(m, "DCPageFlushesXC\t: %d\n",
472 atomic_read(&dcpage_flushes_xcall));
473#endif /* CONFIG_SMP */
474#endif /* CONFIG_DEBUG_DCFLUSH */
475}
476
477struct linux_prom_translation prom_trans[512] __read_mostly;
478unsigned int prom_trans_ents __read_mostly;
479
480unsigned long kern_locked_tte_data;
481
482/* The obp translations are saved based on 8k pagesize, since obp can
483 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
484 * HI_OBP_ADDRESS range are handled in ktlb.S.
485 */
486static inline int in_obp_range(unsigned long vaddr)
487{
488 return (vaddr >= LOW_OBP_ADDRESS &&
489 vaddr < HI_OBP_ADDRESS);
490}
491
492static int cmp_ptrans(const void *a, const void *b)
493{
494 const struct linux_prom_translation *x = a, *y = b;
495
496 if (x->virt > y->virt)
497 return 1;
498 if (x->virt < y->virt)
499 return -1;
500 return 0;
501}
502
503/* Read OBP translations property into 'prom_trans[]'. */
504static void __init read_obp_translations(void)
505{
506 int n, node, ents, first, last, i;
507
508 node = prom_finddevice("/virtual-memory");
509 n = prom_getproplen(node, "translations");
510 if (unlikely(n == 0 || n == -1)) {
511 prom_printf("prom_mappings: Couldn't get size.\n");
512 prom_halt();
513 }
514 if (unlikely(n > sizeof(prom_trans))) {
515 prom_printf("prom_mappings: Size %d is too big.\n", n);
516 prom_halt();
517 }
518
519 if ((n = prom_getproperty(node, "translations",
520 (char *)&prom_trans[0],
521 sizeof(prom_trans))) == -1) {
522 prom_printf("prom_mappings: Couldn't get property.\n");
523 prom_halt();
524 }
525
526 n = n / sizeof(struct linux_prom_translation);
527
528 ents = n;
529
530 sort(prom_trans, ents, sizeof(struct linux_prom_translation),
531 cmp_ptrans, NULL);
532
533 /* Now kick out all the non-OBP entries. */
534 for (i = 0; i < ents; i++) {
535 if (in_obp_range(prom_trans[i].virt))
536 break;
537 }
538 first = i;
539 for (; i < ents; i++) {
540 if (!in_obp_range(prom_trans[i].virt))
541 break;
542 }
543 last = i;
544
545 for (i = 0; i < (last - first); i++) {
546 struct linux_prom_translation *src = &prom_trans[i + first];
547 struct linux_prom_translation *dest = &prom_trans[i];
548
549 *dest = *src;
550 }
551 for (; i < ents; i++) {
552 struct linux_prom_translation *dest = &prom_trans[i];
553 dest->virt = dest->size = dest->data = 0x0UL;
554 }
555
556 prom_trans_ents = last - first;
557
558 if (tlb_type == spitfire) {
559 /* Clear diag TTE bits. */
560 for (i = 0; i < prom_trans_ents; i++)
561 prom_trans[i].data &= ~0x0003fe0000000000UL;
562 }
563
564 /* Force execute bit on. */
565 for (i = 0; i < prom_trans_ents; i++)
566 prom_trans[i].data |= (tlb_type == hypervisor ?
567 _PAGE_EXEC_4V : _PAGE_EXEC_4U);
568}
569
570static void __init hypervisor_tlb_lock(unsigned long vaddr,
571 unsigned long pte,
572 unsigned long mmu)
573{
574 unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
575
576 if (ret != 0) {
577 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
578 "errors with %lx\n", vaddr, 0, pte, mmu, ret);
579 prom_halt();
580 }
581}
582
583static unsigned long kern_large_tte(unsigned long paddr);
584
585static void __init remap_kernel(void)
586{
587 unsigned long phys_page, tte_vaddr, tte_data;
588 int i, tlb_ent = sparc64_highest_locked_tlbent();
589
590 tte_vaddr = (unsigned long) KERNBASE;
591 phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
592 tte_data = kern_large_tte(phys_page);
593
594 kern_locked_tte_data = tte_data;
595
596 /* Now lock us into the TLBs via Hypervisor or OBP. */
597 if (tlb_type == hypervisor) {
598 for (i = 0; i < num_kernel_image_mappings; i++) {
599 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
600 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
601 tte_vaddr += 0x400000;
602 tte_data += 0x400000;
603 }
604 } else {
605 for (i = 0; i < num_kernel_image_mappings; i++) {
606 prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
607 prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
608 tte_vaddr += 0x400000;
609 tte_data += 0x400000;
610 }
611 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
612 }
613 if (tlb_type == cheetah_plus) {
614 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
615 CTX_CHEETAH_PLUS_NUC);
616 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
617 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
618 }
619}
620
621
622static void __init inherit_prom_mappings(void)
623{
624 /* Now fixup OBP's idea about where we really are mapped. */
625 printk("Remapping the kernel... ");
626 remap_kernel();
627 printk("done.\n");
628}
629
630void prom_world(int enter)
631{
632 if (!enter)
633 set_fs(get_fs());
634
635 __asm__ __volatile__("flushw");
636}
637
638void __flush_dcache_range(unsigned long start, unsigned long end)
639{
640 unsigned long va;
641
642 if (tlb_type == spitfire) {
643 int n = 0;
644
645 for (va = start; va < end; va += 32) {
646 spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
647 if (++n >= 512)
648 break;
649 }
650 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
651 start = __pa(start);
652 end = __pa(end);
653 for (va = start; va < end; va += 32)
654 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
655 "membar #Sync"
656 : /* no outputs */
657 : "r" (va),
658 "i" (ASI_DCACHE_INVALIDATE));
659 }
660}
661EXPORT_SYMBOL(__flush_dcache_range);
662
663/* get_new_mmu_context() uses "cache + 1". */
664DEFINE_SPINLOCK(ctx_alloc_lock);
665unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
666#define MAX_CTX_NR (1UL << CTX_NR_BITS)
667#define CTX_BMAP_SLOTS BITS_TO_LONGS(MAX_CTX_NR)
668DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
669
670/* Caller does TLB context flushing on local CPU if necessary.
671 * The caller also ensures that CTX_VALID(mm->context) is false.
672 *
673 * We must be careful about boundary cases so that we never
674 * let the user have CTX 0 (nucleus) or we ever use a CTX
675 * version of zero (and thus NO_CONTEXT would not be caught
676 * by version mis-match tests in mmu_context.h).
677 *
678 * Always invoked with interrupts disabled.
679 */
680void get_new_mmu_context(struct mm_struct *mm)
681{
682 unsigned long ctx, new_ctx;
683 unsigned long orig_pgsz_bits;
684 int new_version;
685
686 spin_lock(&ctx_alloc_lock);
687 orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
688 ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
689 new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
690 new_version = 0;
691 if (new_ctx >= (1 << CTX_NR_BITS)) {
692 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
693 if (new_ctx >= ctx) {
694 int i;
695 new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
696 CTX_FIRST_VERSION;
697 if (new_ctx == 1)
698 new_ctx = CTX_FIRST_VERSION;
699
700 /* Don't call memset, for 16 entries that's just
701 * plain silly...
702 */
703 mmu_context_bmap[0] = 3;
704 mmu_context_bmap[1] = 0;
705 mmu_context_bmap[2] = 0;
706 mmu_context_bmap[3] = 0;
707 for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
708 mmu_context_bmap[i + 0] = 0;
709 mmu_context_bmap[i + 1] = 0;
710 mmu_context_bmap[i + 2] = 0;
711 mmu_context_bmap[i + 3] = 0;
712 }
713 new_version = 1;
714 goto out;
715 }
716 }
717 mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
718 new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
719out:
720 tlb_context_cache = new_ctx;
721 mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
722 spin_unlock(&ctx_alloc_lock);
723
724 if (unlikely(new_version))
725 smp_new_mmu_context_version();
726}
727
728static int numa_enabled = 1;
729static int numa_debug;
730
731static int __init early_numa(char *p)
732{
733 if (!p)
734 return 0;
735
736 if (strstr(p, "off"))
737 numa_enabled = 0;
738
739 if (strstr(p, "debug"))
740 numa_debug = 1;
741
742 return 0;
743}
744early_param("numa", early_numa);
745
746#define numadbg(f, a...) \
747do { if (numa_debug) \
748 printk(KERN_INFO f, ## a); \
749} while (0)
750
751static void __init find_ramdisk(unsigned long phys_base)
752{
753#ifdef CONFIG_BLK_DEV_INITRD
754 if (sparc_ramdisk_image || sparc_ramdisk_image64) {
755 unsigned long ramdisk_image;
756
757 /* Older versions of the bootloader only supported a
758 * 32-bit physical address for the ramdisk image
759 * location, stored at sparc_ramdisk_image. Newer
760 * SILO versions set sparc_ramdisk_image to zero and
761 * provide a full 64-bit physical address at
762 * sparc_ramdisk_image64.
763 */
764 ramdisk_image = sparc_ramdisk_image;
765 if (!ramdisk_image)
766 ramdisk_image = sparc_ramdisk_image64;
767
768 /* Another bootloader quirk. The bootloader normalizes
769 * the physical address to KERNBASE, so we have to
770 * factor that back out and add in the lowest valid
771 * physical page address to get the true physical address.
772 */
773 ramdisk_image -= KERNBASE;
774 ramdisk_image += phys_base;
775
776 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
777 ramdisk_image, sparc_ramdisk_size);
778
779 initrd_start = ramdisk_image;
780 initrd_end = ramdisk_image + sparc_ramdisk_size;
781
782 memblock_reserve(initrd_start, sparc_ramdisk_size);
783
784 initrd_start += PAGE_OFFSET;
785 initrd_end += PAGE_OFFSET;
786 }
787#endif
788}
789
790struct node_mem_mask {
791 unsigned long mask;
792 unsigned long val;
793};
794static struct node_mem_mask node_masks[MAX_NUMNODES];
795static int num_node_masks;
796
797int numa_cpu_lookup_table[NR_CPUS];
798cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
799
800#ifdef CONFIG_NEED_MULTIPLE_NODES
801
802struct mdesc_mblock {
803 u64 base;
804 u64 size;
805 u64 offset; /* RA-to-PA */
806};
807static struct mdesc_mblock *mblocks;
808static int num_mblocks;
809
810static unsigned long ra_to_pa(unsigned long addr)
811{
812 int i;
813
814 for (i = 0; i < num_mblocks; i++) {
815 struct mdesc_mblock *m = &mblocks[i];
816
817 if (addr >= m->base &&
818 addr < (m->base + m->size)) {
819 addr += m->offset;
820 break;
821 }
822 }
823 return addr;
824}
825
826static int find_node(unsigned long addr)
827{
828 int i;
829
830 addr = ra_to_pa(addr);
831 for (i = 0; i < num_node_masks; i++) {
832 struct node_mem_mask *p = &node_masks[i];
833
834 if ((addr & p->mask) == p->val)
835 return i;
836 }
837 return -1;
838}
839
840static u64 memblock_nid_range(u64 start, u64 end, int *nid)
841{
842 *nid = find_node(start);
843 start += PAGE_SIZE;
844 while (start < end) {
845 int n = find_node(start);
846
847 if (n != *nid)
848 break;
849 start += PAGE_SIZE;
850 }
851
852 if (start > end)
853 start = end;
854
855 return start;
856}
857#endif
858
859/* This must be invoked after performing all of the necessary
860 * memblock_set_node() calls for 'nid'. We need to be able to get
861 * correct data from get_pfn_range_for_nid().
862 */
863static void __init allocate_node_data(int nid)
864{
865 struct pglist_data *p;
866 unsigned long start_pfn, end_pfn;
867#ifdef CONFIG_NEED_MULTIPLE_NODES
868 unsigned long paddr;
869
870 paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
871 if (!paddr) {
872 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
873 prom_halt();
874 }
875 NODE_DATA(nid) = __va(paddr);
876 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
877
878 NODE_DATA(nid)->node_id = nid;
879#endif
880
881 p = NODE_DATA(nid);
882
883 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
884 p->node_start_pfn = start_pfn;
885 p->node_spanned_pages = end_pfn - start_pfn;
886}
887
888static void init_node_masks_nonnuma(void)
889{
890 int i;
891
892 numadbg("Initializing tables for non-numa.\n");
893
894 node_masks[0].mask = node_masks[0].val = 0;
895 num_node_masks = 1;
896
897 for (i = 0; i < NR_CPUS; i++)
898 numa_cpu_lookup_table[i] = 0;
899
900 cpumask_setall(&numa_cpumask_lookup_table[0]);
901}
902
903#ifdef CONFIG_NEED_MULTIPLE_NODES
904struct pglist_data *node_data[MAX_NUMNODES];
905
906EXPORT_SYMBOL(numa_cpu_lookup_table);
907EXPORT_SYMBOL(numa_cpumask_lookup_table);
908EXPORT_SYMBOL(node_data);
909
910struct mdesc_mlgroup {
911 u64 node;
912 u64 latency;
913 u64 match;
914 u64 mask;
915};
916static struct mdesc_mlgroup *mlgroups;
917static int num_mlgroups;
918
919static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
920 u32 cfg_handle)
921{
922 u64 arc;
923
924 mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
925 u64 target = mdesc_arc_target(md, arc);
926 const u64 *val;
927
928 val = mdesc_get_property(md, target,
929 "cfg-handle", NULL);
930 if (val && *val == cfg_handle)
931 return 0;
932 }
933 return -ENODEV;
934}
935
936static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
937 u32 cfg_handle)
938{
939 u64 arc, candidate, best_latency = ~(u64)0;
940
941 candidate = MDESC_NODE_NULL;
942 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
943 u64 target = mdesc_arc_target(md, arc);
944 const char *name = mdesc_node_name(md, target);
945 const u64 *val;
946
947 if (strcmp(name, "pio-latency-group"))
948 continue;
949
950 val = mdesc_get_property(md, target, "latency", NULL);
951 if (!val)
952 continue;
953
954 if (*val < best_latency) {
955 candidate = target;
956 best_latency = *val;
957 }
958 }
959
960 if (candidate == MDESC_NODE_NULL)
961 return -ENODEV;
962
963 return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
964}
965
966int of_node_to_nid(struct device_node *dp)
967{
968 const struct linux_prom64_registers *regs;
969 struct mdesc_handle *md;
970 u32 cfg_handle;
971 int count, nid;
972 u64 grp;
973
974 /* This is the right thing to do on currently supported
975 * SUN4U NUMA platforms as well, as the PCI controller does
976 * not sit behind any particular memory controller.
977 */
978 if (!mlgroups)
979 return -1;
980
981 regs = of_get_property(dp, "reg", NULL);
982 if (!regs)
983 return -1;
984
985 cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
986
987 md = mdesc_grab();
988
989 count = 0;
990 nid = -1;
991 mdesc_for_each_node_by_name(md, grp, "group") {
992 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
993 nid = count;
994 break;
995 }
996 count++;
997 }
998
999 mdesc_release(md);
1000
1001 return nid;
1002}
1003
1004static void __init add_node_ranges(void)
1005{
1006 struct memblock_region *reg;
1007
1008 for_each_memblock(memory, reg) {
1009 unsigned long size = reg->size;
1010 unsigned long start, end;
1011
1012 start = reg->base;
1013 end = start + size;
1014 while (start < end) {
1015 unsigned long this_end;
1016 int nid;
1017
1018 this_end = memblock_nid_range(start, end, &nid);
1019
1020 numadbg("Setting memblock NUMA node nid[%d] "
1021 "start[%lx] end[%lx]\n",
1022 nid, start, this_end);
1023
1024 memblock_set_node(start, this_end - start,
1025 &memblock.memory, nid);
1026 start = this_end;
1027 }
1028 }
1029}
1030
1031static int __init grab_mlgroups(struct mdesc_handle *md)
1032{
1033 unsigned long paddr;
1034 int count = 0;
1035 u64 node;
1036
1037 mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1038 count++;
1039 if (!count)
1040 return -ENOENT;
1041
1042 paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1043 SMP_CACHE_BYTES);
1044 if (!paddr)
1045 return -ENOMEM;
1046
1047 mlgroups = __va(paddr);
1048 num_mlgroups = count;
1049
1050 count = 0;
1051 mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1052 struct mdesc_mlgroup *m = &mlgroups[count++];
1053 const u64 *val;
1054
1055 m->node = node;
1056
1057 val = mdesc_get_property(md, node, "latency", NULL);
1058 m->latency = *val;
1059 val = mdesc_get_property(md, node, "address-match", NULL);
1060 m->match = *val;
1061 val = mdesc_get_property(md, node, "address-mask", NULL);
1062 m->mask = *val;
1063
1064 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1065 "match[%llx] mask[%llx]\n",
1066 count - 1, m->node, m->latency, m->match, m->mask);
1067 }
1068
1069 return 0;
1070}
1071
1072static int __init grab_mblocks(struct mdesc_handle *md)
1073{
1074 unsigned long paddr;
1075 int count = 0;
1076 u64 node;
1077
1078 mdesc_for_each_node_by_name(md, node, "mblock")
1079 count++;
1080 if (!count)
1081 return -ENOENT;
1082
1083 paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1084 SMP_CACHE_BYTES);
1085 if (!paddr)
1086 return -ENOMEM;
1087
1088 mblocks = __va(paddr);
1089 num_mblocks = count;
1090
1091 count = 0;
1092 mdesc_for_each_node_by_name(md, node, "mblock") {
1093 struct mdesc_mblock *m = &mblocks[count++];
1094 const u64 *val;
1095
1096 val = mdesc_get_property(md, node, "base", NULL);
1097 m->base = *val;
1098 val = mdesc_get_property(md, node, "size", NULL);
1099 m->size = *val;
1100 val = mdesc_get_property(md, node,
1101 "address-congruence-offset", NULL);
1102
1103 /* The address-congruence-offset property is optional.
1104 * Explicity zero it be identifty this.
1105 */
1106 if (val)
1107 m->offset = *val;
1108 else
1109 m->offset = 0UL;
1110
1111 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1112 count - 1, m->base, m->size, m->offset);
1113 }
1114
1115 return 0;
1116}
1117
1118static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1119 u64 grp, cpumask_t *mask)
1120{
1121 u64 arc;
1122
1123 cpumask_clear(mask);
1124
1125 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1126 u64 target = mdesc_arc_target(md, arc);
1127 const char *name = mdesc_node_name(md, target);
1128 const u64 *id;
1129
1130 if (strcmp(name, "cpu"))
1131 continue;
1132 id = mdesc_get_property(md, target, "id", NULL);
1133 if (*id < nr_cpu_ids)
1134 cpumask_set_cpu(*id, mask);
1135 }
1136}
1137
1138static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1139{
1140 int i;
1141
1142 for (i = 0; i < num_mlgroups; i++) {
1143 struct mdesc_mlgroup *m = &mlgroups[i];
1144 if (m->node == node)
1145 return m;
1146 }
1147 return NULL;
1148}
1149
1150static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1151 int index)
1152{
1153 struct mdesc_mlgroup *candidate = NULL;
1154 u64 arc, best_latency = ~(u64)0;
1155 struct node_mem_mask *n;
1156
1157 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1158 u64 target = mdesc_arc_target(md, arc);
1159 struct mdesc_mlgroup *m = find_mlgroup(target);
1160 if (!m)
1161 continue;
1162 if (m->latency < best_latency) {
1163 candidate = m;
1164 best_latency = m->latency;
1165 }
1166 }
1167 if (!candidate)
1168 return -ENOENT;
1169
1170 if (num_node_masks != index) {
1171 printk(KERN_ERR "Inconsistent NUMA state, "
1172 "index[%d] != num_node_masks[%d]\n",
1173 index, num_node_masks);
1174 return -EINVAL;
1175 }
1176
1177 n = &node_masks[num_node_masks++];
1178
1179 n->mask = candidate->mask;
1180 n->val = candidate->match;
1181
1182 numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1183 index, n->mask, n->val, candidate->latency);
1184
1185 return 0;
1186}
1187
1188static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1189 int index)
1190{
1191 cpumask_t mask;
1192 int cpu;
1193
1194 numa_parse_mdesc_group_cpus(md, grp, &mask);
1195
1196 for_each_cpu(cpu, &mask)
1197 numa_cpu_lookup_table[cpu] = index;
1198 cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1199
1200 if (numa_debug) {
1201 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1202 for_each_cpu(cpu, &mask)
1203 printk("%d ", cpu);
1204 printk("]\n");
1205 }
1206
1207 return numa_attach_mlgroup(md, grp, index);
1208}
1209
1210static int __init numa_parse_mdesc(void)
1211{
1212 struct mdesc_handle *md = mdesc_grab();
1213 int i, err, count;
1214 u64 node;
1215
1216 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1217 if (node == MDESC_NODE_NULL) {
1218 mdesc_release(md);
1219 return -ENOENT;
1220 }
1221
1222 err = grab_mblocks(md);
1223 if (err < 0)
1224 goto out;
1225
1226 err = grab_mlgroups(md);
1227 if (err < 0)
1228 goto out;
1229
1230 count = 0;
1231 mdesc_for_each_node_by_name(md, node, "group") {
1232 err = numa_parse_mdesc_group(md, node, count);
1233 if (err < 0)
1234 break;
1235 count++;
1236 }
1237
1238 add_node_ranges();
1239
1240 for (i = 0; i < num_node_masks; i++) {
1241 allocate_node_data(i);
1242 node_set_online(i);
1243 }
1244
1245 err = 0;
1246out:
1247 mdesc_release(md);
1248 return err;
1249}
1250
1251static int __init numa_parse_jbus(void)
1252{
1253 unsigned long cpu, index;
1254
1255 /* NUMA node id is encoded in bits 36 and higher, and there is
1256 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1257 */
1258 index = 0;
1259 for_each_present_cpu(cpu) {
1260 numa_cpu_lookup_table[cpu] = index;
1261 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1262 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1263 node_masks[index].val = cpu << 36UL;
1264
1265 index++;
1266 }
1267 num_node_masks = index;
1268
1269 add_node_ranges();
1270
1271 for (index = 0; index < num_node_masks; index++) {
1272 allocate_node_data(index);
1273 node_set_online(index);
1274 }
1275
1276 return 0;
1277}
1278
1279static int __init numa_parse_sun4u(void)
1280{
1281 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1282 unsigned long ver;
1283
1284 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1285 if ((ver >> 32UL) == __JALAPENO_ID ||
1286 (ver >> 32UL) == __SERRANO_ID)
1287 return numa_parse_jbus();
1288 }
1289 return -1;
1290}
1291
1292static int __init bootmem_init_numa(void)
1293{
1294 int err = -1;
1295
1296 numadbg("bootmem_init_numa()\n");
1297
1298 if (numa_enabled) {
1299 if (tlb_type == hypervisor)
1300 err = numa_parse_mdesc();
1301 else
1302 err = numa_parse_sun4u();
1303 }
1304 return err;
1305}
1306
1307#else
1308
1309static int bootmem_init_numa(void)
1310{
1311 return -1;
1312}
1313
1314#endif
1315
1316static void __init bootmem_init_nonnuma(void)
1317{
1318 unsigned long top_of_ram = memblock_end_of_DRAM();
1319 unsigned long total_ram = memblock_phys_mem_size();
1320
1321 numadbg("bootmem_init_nonnuma()\n");
1322
1323 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1324 top_of_ram, total_ram);
1325 printk(KERN_INFO "Memory hole size: %ldMB\n",
1326 (top_of_ram - total_ram) >> 20);
1327
1328 init_node_masks_nonnuma();
1329 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1330 allocate_node_data(0);
1331 node_set_online(0);
1332}
1333
1334static unsigned long __init bootmem_init(unsigned long phys_base)
1335{
1336 unsigned long end_pfn;
1337
1338 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1339 max_pfn = max_low_pfn = end_pfn;
1340 min_low_pfn = (phys_base >> PAGE_SHIFT);
1341
1342 if (bootmem_init_numa() < 0)
1343 bootmem_init_nonnuma();
1344
1345 /* Dump memblock with node info. */
1346 memblock_dump_all();
1347
1348 /* XXX cpu notifier XXX */
1349
1350 sparse_memory_present_with_active_regions(MAX_NUMNODES);
1351 sparse_init();
1352
1353 return end_pfn;
1354}
1355
1356static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1357static int pall_ents __initdata;
1358
1359#ifdef CONFIG_DEBUG_PAGEALLOC
1360static unsigned long __ref kernel_map_range(unsigned long pstart,
1361 unsigned long pend, pgprot_t prot)
1362{
1363 unsigned long vstart = PAGE_OFFSET + pstart;
1364 unsigned long vend = PAGE_OFFSET + pend;
1365 unsigned long alloc_bytes = 0UL;
1366
1367 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1368 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1369 vstart, vend);
1370 prom_halt();
1371 }
1372
1373 while (vstart < vend) {
1374 unsigned long this_end, paddr = __pa(vstart);
1375 pgd_t *pgd = pgd_offset_k(vstart);
1376 pud_t *pud;
1377 pmd_t *pmd;
1378 pte_t *pte;
1379
1380 pud = pud_offset(pgd, vstart);
1381 if (pud_none(*pud)) {
1382 pmd_t *new;
1383
1384 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1385 alloc_bytes += PAGE_SIZE;
1386 pud_populate(&init_mm, pud, new);
1387 }
1388
1389 pmd = pmd_offset(pud, vstart);
1390 if (!pmd_present(*pmd)) {
1391 pte_t *new;
1392
1393 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1394 alloc_bytes += PAGE_SIZE;
1395 pmd_populate_kernel(&init_mm, pmd, new);
1396 }
1397
1398 pte = pte_offset_kernel(pmd, vstart);
1399 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1400 if (this_end > vend)
1401 this_end = vend;
1402
1403 while (vstart < this_end) {
1404 pte_val(*pte) = (paddr | pgprot_val(prot));
1405
1406 vstart += PAGE_SIZE;
1407 paddr += PAGE_SIZE;
1408 pte++;
1409 }
1410 }
1411
1412 return alloc_bytes;
1413}
1414
1415extern unsigned int kvmap_linear_patch[1];
1416#endif /* CONFIG_DEBUG_PAGEALLOC */
1417
1418static void __init kpte_set_val(unsigned long index, unsigned long val)
1419{
1420 unsigned long *ptr = kpte_linear_bitmap;
1421
1422 val <<= ((index % (BITS_PER_LONG / 2)) * 2);
1423 ptr += (index / (BITS_PER_LONG / 2));
1424
1425 *ptr |= val;
1426}
1427
1428static const unsigned long kpte_shift_min = 28; /* 256MB */
1429static const unsigned long kpte_shift_max = 34; /* 16GB */
1430static const unsigned long kpte_shift_incr = 3;
1431
1432static unsigned long kpte_mark_using_shift(unsigned long start, unsigned long end,
1433 unsigned long shift)
1434{
1435 unsigned long size = (1UL << shift);
1436 unsigned long mask = (size - 1UL);
1437 unsigned long remains = end - start;
1438 unsigned long val;
1439
1440 if (remains < size || (start & mask))
1441 return start;
1442
1443 /* VAL maps:
1444 *
1445 * shift 28 --> kern_linear_pte_xor index 1
1446 * shift 31 --> kern_linear_pte_xor index 2
1447 * shift 34 --> kern_linear_pte_xor index 3
1448 */
1449 val = ((shift - kpte_shift_min) / kpte_shift_incr) + 1;
1450
1451 remains &= ~mask;
1452 if (shift != kpte_shift_max)
1453 remains = size;
1454
1455 while (remains) {
1456 unsigned long index = start >> kpte_shift_min;
1457
1458 kpte_set_val(index, val);
1459
1460 start += 1UL << kpte_shift_min;
1461 remains -= 1UL << kpte_shift_min;
1462 }
1463
1464 return start;
1465}
1466
1467static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1468{
1469 unsigned long smallest_size, smallest_mask;
1470 unsigned long s;
1471
1472 smallest_size = (1UL << kpte_shift_min);
1473 smallest_mask = (smallest_size - 1UL);
1474
1475 while (start < end) {
1476 unsigned long orig_start = start;
1477
1478 for (s = kpte_shift_max; s >= kpte_shift_min; s -= kpte_shift_incr) {
1479 start = kpte_mark_using_shift(start, end, s);
1480
1481 if (start != orig_start)
1482 break;
1483 }
1484
1485 if (start == orig_start)
1486 start = (start + smallest_size) & ~smallest_mask;
1487 }
1488}
1489
1490static void __init init_kpte_bitmap(void)
1491{
1492 unsigned long i;
1493
1494 for (i = 0; i < pall_ents; i++) {
1495 unsigned long phys_start, phys_end;
1496
1497 phys_start = pall[i].phys_addr;
1498 phys_end = phys_start + pall[i].reg_size;
1499
1500 mark_kpte_bitmap(phys_start, phys_end);
1501 }
1502}
1503
1504static void __init kernel_physical_mapping_init(void)
1505{
1506#ifdef CONFIG_DEBUG_PAGEALLOC
1507 unsigned long i, mem_alloced = 0UL;
1508
1509 for (i = 0; i < pall_ents; i++) {
1510 unsigned long phys_start, phys_end;
1511
1512 phys_start = pall[i].phys_addr;
1513 phys_end = phys_start + pall[i].reg_size;
1514
1515 mem_alloced += kernel_map_range(phys_start, phys_end,
1516 PAGE_KERNEL);
1517 }
1518
1519 printk("Allocated %ld bytes for kernel page tables.\n",
1520 mem_alloced);
1521
1522 kvmap_linear_patch[0] = 0x01000000; /* nop */
1523 flushi(&kvmap_linear_patch[0]);
1524
1525 __flush_tlb_all();
1526#endif
1527}
1528
1529#ifdef CONFIG_DEBUG_PAGEALLOC
1530void kernel_map_pages(struct page *page, int numpages, int enable)
1531{
1532 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1533 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1534
1535 kernel_map_range(phys_start, phys_end,
1536 (enable ? PAGE_KERNEL : __pgprot(0)));
1537
1538 flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1539 PAGE_OFFSET + phys_end);
1540
1541 /* we should perform an IPI and flush all tlbs,
1542 * but that can deadlock->flush only current cpu.
1543 */
1544 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1545 PAGE_OFFSET + phys_end);
1546}
1547#endif
1548
1549unsigned long __init find_ecache_flush_span(unsigned long size)
1550{
1551 int i;
1552
1553 for (i = 0; i < pavail_ents; i++) {
1554 if (pavail[i].reg_size >= size)
1555 return pavail[i].phys_addr;
1556 }
1557
1558 return ~0UL;
1559}
1560
1561unsigned long PAGE_OFFSET;
1562EXPORT_SYMBOL(PAGE_OFFSET);
1563
1564static void __init page_offset_shift_patch_one(unsigned int *insn, unsigned long phys_bits)
1565{
1566 unsigned long final_shift;
1567 unsigned int val = *insn;
1568 unsigned int cnt;
1569
1570 /* We are patching in ilog2(max_supported_phys_address), and
1571 * we are doing so in a manner similar to a relocation addend.
1572 * That is, we are adding the shift value to whatever value
1573 * is in the shift instruction count field already.
1574 */
1575 cnt = (val & 0x3f);
1576 val &= ~0x3f;
1577
1578 /* If we are trying to shift >= 64 bits, clear the destination
1579 * register. This can happen when phys_bits ends up being equal
1580 * to MAX_PHYS_ADDRESS_BITS.
1581 */
1582 final_shift = (cnt + (64 - phys_bits));
1583 if (final_shift >= 64) {
1584 unsigned int rd = (val >> 25) & 0x1f;
1585
1586 val = 0x80100000 | (rd << 25);
1587 } else {
1588 val |= final_shift;
1589 }
1590 *insn = val;
1591
1592 __asm__ __volatile__("flush %0"
1593 : /* no outputs */
1594 : "r" (insn));
1595}
1596
1597static void __init page_offset_shift_patch(unsigned long phys_bits)
1598{
1599 extern unsigned int __page_offset_shift_patch;
1600 extern unsigned int __page_offset_shift_patch_end;
1601 unsigned int *p;
1602
1603 p = &__page_offset_shift_patch;
1604 while (p < &__page_offset_shift_patch_end) {
1605 unsigned int *insn = (unsigned int *)(unsigned long)*p;
1606
1607 page_offset_shift_patch_one(insn, phys_bits);
1608
1609 p++;
1610 }
1611}
1612
1613static void __init setup_page_offset(void)
1614{
1615 unsigned long max_phys_bits = 40;
1616
1617 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1618 max_phys_bits = 42;
1619 } else if (tlb_type == hypervisor) {
1620 switch (sun4v_chip_type) {
1621 case SUN4V_CHIP_NIAGARA1:
1622 case SUN4V_CHIP_NIAGARA2:
1623 max_phys_bits = 39;
1624 break;
1625 case SUN4V_CHIP_NIAGARA3:
1626 max_phys_bits = 43;
1627 break;
1628 case SUN4V_CHIP_NIAGARA4:
1629 case SUN4V_CHIP_NIAGARA5:
1630 case SUN4V_CHIP_SPARC64X:
1631 default:
1632 max_phys_bits = 47;
1633 break;
1634 }
1635 }
1636
1637 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1638 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1639 max_phys_bits);
1640 prom_halt();
1641 }
1642
1643 PAGE_OFFSET = PAGE_OFFSET_BY_BITS(max_phys_bits);
1644
1645 pr_info("PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1646 PAGE_OFFSET, max_phys_bits);
1647
1648 page_offset_shift_patch(max_phys_bits);
1649}
1650
1651static void __init tsb_phys_patch(void)
1652{
1653 struct tsb_ldquad_phys_patch_entry *pquad;
1654 struct tsb_phys_patch_entry *p;
1655
1656 pquad = &__tsb_ldquad_phys_patch;
1657 while (pquad < &__tsb_ldquad_phys_patch_end) {
1658 unsigned long addr = pquad->addr;
1659
1660 if (tlb_type == hypervisor)
1661 *(unsigned int *) addr = pquad->sun4v_insn;
1662 else
1663 *(unsigned int *) addr = pquad->sun4u_insn;
1664 wmb();
1665 __asm__ __volatile__("flush %0"
1666 : /* no outputs */
1667 : "r" (addr));
1668
1669 pquad++;
1670 }
1671
1672 p = &__tsb_phys_patch;
1673 while (p < &__tsb_phys_patch_end) {
1674 unsigned long addr = p->addr;
1675
1676 *(unsigned int *) addr = p->insn;
1677 wmb();
1678 __asm__ __volatile__("flush %0"
1679 : /* no outputs */
1680 : "r" (addr));
1681
1682 p++;
1683 }
1684}
1685
1686/* Don't mark as init, we give this to the Hypervisor. */
1687#ifndef CONFIG_DEBUG_PAGEALLOC
1688#define NUM_KTSB_DESCR 2
1689#else
1690#define NUM_KTSB_DESCR 1
1691#endif
1692static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1693extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1694
1695static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1696{
1697 pa >>= KTSB_PHYS_SHIFT;
1698
1699 while (start < end) {
1700 unsigned int *ia = (unsigned int *)(unsigned long)*start;
1701
1702 ia[0] = (ia[0] & ~0x3fffff) | (pa >> 10);
1703 __asm__ __volatile__("flush %0" : : "r" (ia));
1704
1705 ia[1] = (ia[1] & ~0x3ff) | (pa & 0x3ff);
1706 __asm__ __volatile__("flush %0" : : "r" (ia + 1));
1707
1708 start++;
1709 }
1710}
1711
1712static void ktsb_phys_patch(void)
1713{
1714 extern unsigned int __swapper_tsb_phys_patch;
1715 extern unsigned int __swapper_tsb_phys_patch_end;
1716 unsigned long ktsb_pa;
1717
1718 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1719 patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1720 &__swapper_tsb_phys_patch_end, ktsb_pa);
1721#ifndef CONFIG_DEBUG_PAGEALLOC
1722 {
1723 extern unsigned int __swapper_4m_tsb_phys_patch;
1724 extern unsigned int __swapper_4m_tsb_phys_patch_end;
1725 ktsb_pa = (kern_base +
1726 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1727 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1728 &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1729 }
1730#endif
1731}
1732
1733static void __init sun4v_ktsb_init(void)
1734{
1735 unsigned long ktsb_pa;
1736
1737 /* First KTSB for PAGE_SIZE mappings. */
1738 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1739
1740 switch (PAGE_SIZE) {
1741 case 8 * 1024:
1742 default:
1743 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1744 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1745 break;
1746
1747 case 64 * 1024:
1748 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1749 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1750 break;
1751
1752 case 512 * 1024:
1753 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1754 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1755 break;
1756
1757 case 4 * 1024 * 1024:
1758 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1759 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1760 break;
1761 }
1762
1763 ktsb_descr[0].assoc = 1;
1764 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1765 ktsb_descr[0].ctx_idx = 0;
1766 ktsb_descr[0].tsb_base = ktsb_pa;
1767 ktsb_descr[0].resv = 0;
1768
1769#ifndef CONFIG_DEBUG_PAGEALLOC
1770 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */
1771 ktsb_pa = (kern_base +
1772 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1773
1774 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1775 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
1776 HV_PGSZ_MASK_256MB |
1777 HV_PGSZ_MASK_2GB |
1778 HV_PGSZ_MASK_16GB) &
1779 cpu_pgsz_mask);
1780 ktsb_descr[1].assoc = 1;
1781 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1782 ktsb_descr[1].ctx_idx = 0;
1783 ktsb_descr[1].tsb_base = ktsb_pa;
1784 ktsb_descr[1].resv = 0;
1785#endif
1786}
1787
1788void sun4v_ktsb_register(void)
1789{
1790 unsigned long pa, ret;
1791
1792 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1793
1794 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1795 if (ret != 0) {
1796 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1797 "errors with %lx\n", pa, ret);
1798 prom_halt();
1799 }
1800}
1801
1802static void __init sun4u_linear_pte_xor_finalize(void)
1803{
1804#ifndef CONFIG_DEBUG_PAGEALLOC
1805 /* This is where we would add Panther support for
1806 * 32MB and 256MB pages.
1807 */
1808#endif
1809}
1810
1811static void __init sun4v_linear_pte_xor_finalize(void)
1812{
1813#ifndef CONFIG_DEBUG_PAGEALLOC
1814 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
1815 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
1816 PAGE_OFFSET;
1817 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1818 _PAGE_P_4V | _PAGE_W_4V);
1819 } else {
1820 kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
1821 }
1822
1823 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
1824 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
1825 PAGE_OFFSET;
1826 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1827 _PAGE_P_4V | _PAGE_W_4V);
1828 } else {
1829 kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
1830 }
1831
1832 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
1833 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
1834 PAGE_OFFSET;
1835 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1836 _PAGE_P_4V | _PAGE_W_4V);
1837 } else {
1838 kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
1839 }
1840#endif
1841}
1842
1843/* paging_init() sets up the page tables */
1844
1845static unsigned long last_valid_pfn;
1846pgd_t swapper_pg_dir[PTRS_PER_PGD];
1847
1848static void sun4u_pgprot_init(void);
1849static void sun4v_pgprot_init(void);
1850
1851void __init paging_init(void)
1852{
1853 unsigned long end_pfn, shift, phys_base;
1854 unsigned long real_end, i;
1855 int node;
1856
1857 setup_page_offset();
1858
1859 /* These build time checkes make sure that the dcache_dirty_cpu()
1860 * page->flags usage will work.
1861 *
1862 * When a page gets marked as dcache-dirty, we store the
1863 * cpu number starting at bit 32 in the page->flags. Also,
1864 * functions like clear_dcache_dirty_cpu use the cpu mask
1865 * in 13-bit signed-immediate instruction fields.
1866 */
1867
1868 /*
1869 * Page flags must not reach into upper 32 bits that are used
1870 * for the cpu number
1871 */
1872 BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1873
1874 /*
1875 * The bit fields placed in the high range must not reach below
1876 * the 32 bit boundary. Otherwise we cannot place the cpu field
1877 * at the 32 bit boundary.
1878 */
1879 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1880 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1881
1882 BUILD_BUG_ON(NR_CPUS > 4096);
1883
1884 kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
1885 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1886
1887 /* Invalidate both kernel TSBs. */
1888 memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1889#ifndef CONFIG_DEBUG_PAGEALLOC
1890 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1891#endif
1892
1893 if (tlb_type == hypervisor)
1894 sun4v_pgprot_init();
1895 else
1896 sun4u_pgprot_init();
1897
1898 if (tlb_type == cheetah_plus ||
1899 tlb_type == hypervisor) {
1900 tsb_phys_patch();
1901 ktsb_phys_patch();
1902 }
1903
1904 if (tlb_type == hypervisor)
1905 sun4v_patch_tlb_handlers();
1906
1907 /* Find available physical memory...
1908 *
1909 * Read it twice in order to work around a bug in openfirmware.
1910 * The call to grab this table itself can cause openfirmware to
1911 * allocate memory, which in turn can take away some space from
1912 * the list of available memory. Reading it twice makes sure
1913 * we really do get the final value.
1914 */
1915 read_obp_translations();
1916 read_obp_memory("reg", &pall[0], &pall_ents);
1917 read_obp_memory("available", &pavail[0], &pavail_ents);
1918 read_obp_memory("available", &pavail[0], &pavail_ents);
1919
1920 phys_base = 0xffffffffffffffffUL;
1921 for (i = 0; i < pavail_ents; i++) {
1922 phys_base = min(phys_base, pavail[i].phys_addr);
1923 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
1924 }
1925
1926 memblock_reserve(kern_base, kern_size);
1927
1928 find_ramdisk(phys_base);
1929
1930 memblock_enforce_memory_limit(cmdline_memory_size);
1931
1932 memblock_allow_resize();
1933 memblock_dump_all();
1934
1935 set_bit(0, mmu_context_bmap);
1936
1937 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1938
1939 real_end = (unsigned long)_end;
1940 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
1941 printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1942 num_kernel_image_mappings);
1943
1944 /* Set kernel pgd to upper alias so physical page computations
1945 * work.
1946 */
1947 init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1948
1949 memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1950
1951 /* Now can init the kernel/bad page tables. */
1952 pud_set(pud_offset(&swapper_pg_dir[0], 0),
1953 swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1954
1955 inherit_prom_mappings();
1956
1957 init_kpte_bitmap();
1958
1959 /* Ok, we can use our TLB miss and window trap handlers safely. */
1960 setup_tba();
1961
1962 __flush_tlb_all();
1963
1964 prom_build_devicetree();
1965 of_populate_present_mask();
1966#ifndef CONFIG_SMP
1967 of_fill_in_cpu_data();
1968#endif
1969
1970 if (tlb_type == hypervisor) {
1971 sun4v_mdesc_init();
1972 mdesc_populate_present_mask(cpu_all_mask);
1973#ifndef CONFIG_SMP
1974 mdesc_fill_in_cpu_data(cpu_all_mask);
1975#endif
1976 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
1977
1978 sun4v_linear_pte_xor_finalize();
1979
1980 sun4v_ktsb_init();
1981 sun4v_ktsb_register();
1982 } else {
1983 unsigned long impl, ver;
1984
1985 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
1986 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
1987
1988 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
1989 impl = ((ver >> 32) & 0xffff);
1990 if (impl == PANTHER_IMPL)
1991 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
1992 HV_PGSZ_MASK_256MB);
1993
1994 sun4u_linear_pte_xor_finalize();
1995 }
1996
1997 /* Flush the TLBs and the 4M TSB so that the updated linear
1998 * pte XOR settings are realized for all mappings.
1999 */
2000 __flush_tlb_all();
2001#ifndef CONFIG_DEBUG_PAGEALLOC
2002 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2003#endif
2004 __flush_tlb_all();
2005
2006 /* Setup bootmem... */
2007 last_valid_pfn = end_pfn = bootmem_init(phys_base);
2008
2009 /* Once the OF device tree and MDESC have been setup, we know
2010 * the list of possible cpus. Therefore we can allocate the
2011 * IRQ stacks.
2012 */
2013 for_each_possible_cpu(i) {
2014 node = cpu_to_node(i);
2015
2016 softirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2017 THREAD_SIZE,
2018 THREAD_SIZE, 0);
2019 hardirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2020 THREAD_SIZE,
2021 THREAD_SIZE, 0);
2022 }
2023
2024 kernel_physical_mapping_init();
2025
2026 {
2027 unsigned long max_zone_pfns[MAX_NR_ZONES];
2028
2029 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2030
2031 max_zone_pfns[ZONE_NORMAL] = end_pfn;
2032
2033 free_area_init_nodes(max_zone_pfns);
2034 }
2035
2036 printk("Booting Linux...\n");
2037}
2038
2039int page_in_phys_avail(unsigned long paddr)
2040{
2041 int i;
2042
2043 paddr &= PAGE_MASK;
2044
2045 for (i = 0; i < pavail_ents; i++) {
2046 unsigned long start, end;
2047
2048 start = pavail[i].phys_addr;
2049 end = start + pavail[i].reg_size;
2050
2051 if (paddr >= start && paddr < end)
2052 return 1;
2053 }
2054 if (paddr >= kern_base && paddr < (kern_base + kern_size))
2055 return 1;
2056#ifdef CONFIG_BLK_DEV_INITRD
2057 if (paddr >= __pa(initrd_start) &&
2058 paddr < __pa(PAGE_ALIGN(initrd_end)))
2059 return 1;
2060#endif
2061
2062 return 0;
2063}
2064
2065static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
2066static int pavail_rescan_ents __initdata;
2067
2068/* Certain OBP calls, such as fetching "available" properties, can
2069 * claim physical memory. So, along with initializing the valid
2070 * address bitmap, what we do here is refetch the physical available
2071 * memory list again, and make sure it provides at least as much
2072 * memory as 'pavail' does.
2073 */
2074static void __init setup_valid_addr_bitmap_from_pavail(unsigned long *bitmap)
2075{
2076 int i;
2077
2078 read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
2079
2080 for (i = 0; i < pavail_ents; i++) {
2081 unsigned long old_start, old_end;
2082
2083 old_start = pavail[i].phys_addr;
2084 old_end = old_start + pavail[i].reg_size;
2085 while (old_start < old_end) {
2086 int n;
2087
2088 for (n = 0; n < pavail_rescan_ents; n++) {
2089 unsigned long new_start, new_end;
2090
2091 new_start = pavail_rescan[n].phys_addr;
2092 new_end = new_start +
2093 pavail_rescan[n].reg_size;
2094
2095 if (new_start <= old_start &&
2096 new_end >= (old_start + PAGE_SIZE)) {
2097 set_bit(old_start >> ILOG2_4MB, bitmap);
2098 goto do_next_page;
2099 }
2100 }
2101
2102 prom_printf("mem_init: Lost memory in pavail\n");
2103 prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
2104 pavail[i].phys_addr,
2105 pavail[i].reg_size);
2106 prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
2107 pavail_rescan[i].phys_addr,
2108 pavail_rescan[i].reg_size);
2109 prom_printf("mem_init: Cannot continue, aborting.\n");
2110 prom_halt();
2111
2112 do_next_page:
2113 old_start += PAGE_SIZE;
2114 }
2115 }
2116}
2117
2118static void __init patch_tlb_miss_handler_bitmap(void)
2119{
2120 extern unsigned int valid_addr_bitmap_insn[];
2121 extern unsigned int valid_addr_bitmap_patch[];
2122
2123 valid_addr_bitmap_insn[1] = valid_addr_bitmap_patch[1];
2124 mb();
2125 valid_addr_bitmap_insn[0] = valid_addr_bitmap_patch[0];
2126 flushi(&valid_addr_bitmap_insn[0]);
2127}
2128
2129static void __init register_page_bootmem_info(void)
2130{
2131#ifdef CONFIG_NEED_MULTIPLE_NODES
2132 int i;
2133
2134 for_each_online_node(i)
2135 if (NODE_DATA(i)->node_spanned_pages)
2136 register_page_bootmem_info_node(NODE_DATA(i));
2137#endif
2138}
2139void __init mem_init(void)
2140{
2141 unsigned long addr, last;
2142
2143 addr = PAGE_OFFSET + kern_base;
2144 last = PAGE_ALIGN(kern_size) + addr;
2145 while (addr < last) {
2146 set_bit(__pa(addr) >> ILOG2_4MB, sparc64_valid_addr_bitmap);
2147 addr += PAGE_SIZE;
2148 }
2149
2150 setup_valid_addr_bitmap_from_pavail(sparc64_valid_addr_bitmap);
2151 patch_tlb_miss_handler_bitmap();
2152
2153 high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2154
2155 register_page_bootmem_info();
2156 free_all_bootmem();
2157
2158 /*
2159 * Set up the zero page, mark it reserved, so that page count
2160 * is not manipulated when freeing the page from user ptes.
2161 */
2162 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2163 if (mem_map_zero == NULL) {
2164 prom_printf("paging_init: Cannot alloc zero page.\n");
2165 prom_halt();
2166 }
2167 mark_page_reserved(mem_map_zero);
2168
2169 mem_init_print_info(NULL);
2170
2171 if (tlb_type == cheetah || tlb_type == cheetah_plus)
2172 cheetah_ecache_flush_init();
2173}
2174
2175void free_initmem(void)
2176{
2177 unsigned long addr, initend;
2178 int do_free = 1;
2179
2180 /* If the physical memory maps were trimmed by kernel command
2181 * line options, don't even try freeing this initmem stuff up.
2182 * The kernel image could have been in the trimmed out region
2183 * and if so the freeing below will free invalid page structs.
2184 */
2185 if (cmdline_memory_size)
2186 do_free = 0;
2187
2188 /*
2189 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2190 */
2191 addr = PAGE_ALIGN((unsigned long)(__init_begin));
2192 initend = (unsigned long)(__init_end) & PAGE_MASK;
2193 for (; addr < initend; addr += PAGE_SIZE) {
2194 unsigned long page;
2195
2196 page = (addr +
2197 ((unsigned long) __va(kern_base)) -
2198 ((unsigned long) KERNBASE));
2199 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2200
2201 if (do_free)
2202 free_reserved_page(virt_to_page(page));
2203 }
2204}
2205
2206#ifdef CONFIG_BLK_DEV_INITRD
2207void free_initrd_mem(unsigned long start, unsigned long end)
2208{
2209 free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2210 "initrd");
2211}
2212#endif
2213
2214#define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U)
2215#define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V)
2216#define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2217#define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2218#define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2219#define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2220
2221pgprot_t PAGE_KERNEL __read_mostly;
2222EXPORT_SYMBOL(PAGE_KERNEL);
2223
2224pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2225pgprot_t PAGE_COPY __read_mostly;
2226
2227pgprot_t PAGE_SHARED __read_mostly;
2228EXPORT_SYMBOL(PAGE_SHARED);
2229
2230unsigned long pg_iobits __read_mostly;
2231
2232unsigned long _PAGE_IE __read_mostly;
2233EXPORT_SYMBOL(_PAGE_IE);
2234
2235unsigned long _PAGE_E __read_mostly;
2236EXPORT_SYMBOL(_PAGE_E);
2237
2238unsigned long _PAGE_CACHE __read_mostly;
2239EXPORT_SYMBOL(_PAGE_CACHE);
2240
2241#ifdef CONFIG_SPARSEMEM_VMEMMAP
2242unsigned long vmemmap_table[VMEMMAP_SIZE];
2243
2244static long __meminitdata addr_start, addr_end;
2245static int __meminitdata node_start;
2246
2247int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2248 int node)
2249{
2250 unsigned long phys_start = (vstart - VMEMMAP_BASE);
2251 unsigned long phys_end = (vend - VMEMMAP_BASE);
2252 unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2253 unsigned long end = VMEMMAP_ALIGN(phys_end);
2254 unsigned long pte_base;
2255
2256 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2257 _PAGE_CP_4U | _PAGE_CV_4U |
2258 _PAGE_P_4U | _PAGE_W_4U);
2259 if (tlb_type == hypervisor)
2260 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2261 _PAGE_CP_4V | _PAGE_CV_4V |
2262 _PAGE_P_4V | _PAGE_W_4V);
2263
2264 for (; addr < end; addr += VMEMMAP_CHUNK) {
2265 unsigned long *vmem_pp =
2266 vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2267 void *block;
2268
2269 if (!(*vmem_pp & _PAGE_VALID)) {
2270 block = vmemmap_alloc_block(1UL << ILOG2_4MB, node);
2271 if (!block)
2272 return -ENOMEM;
2273
2274 *vmem_pp = pte_base | __pa(block);
2275
2276 /* check to see if we have contiguous blocks */
2277 if (addr_end != addr || node_start != node) {
2278 if (addr_start)
2279 printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2280 addr_start, addr_end-1, node_start);
2281 addr_start = addr;
2282 node_start = node;
2283 }
2284 addr_end = addr + VMEMMAP_CHUNK;
2285 }
2286 }
2287 return 0;
2288}
2289
2290void __meminit vmemmap_populate_print_last(void)
2291{
2292 if (addr_start) {
2293 printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2294 addr_start, addr_end-1, node_start);
2295 addr_start = 0;
2296 addr_end = 0;
2297 node_start = 0;
2298 }
2299}
2300
2301void vmemmap_free(unsigned long start, unsigned long end)
2302{
2303}
2304
2305#endif /* CONFIG_SPARSEMEM_VMEMMAP */
2306
2307static void prot_init_common(unsigned long page_none,
2308 unsigned long page_shared,
2309 unsigned long page_copy,
2310 unsigned long page_readonly,
2311 unsigned long page_exec_bit)
2312{
2313 PAGE_COPY = __pgprot(page_copy);
2314 PAGE_SHARED = __pgprot(page_shared);
2315
2316 protection_map[0x0] = __pgprot(page_none);
2317 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2318 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2319 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2320 protection_map[0x4] = __pgprot(page_readonly);
2321 protection_map[0x5] = __pgprot(page_readonly);
2322 protection_map[0x6] = __pgprot(page_copy);
2323 protection_map[0x7] = __pgprot(page_copy);
2324 protection_map[0x8] = __pgprot(page_none);
2325 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2326 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2327 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2328 protection_map[0xc] = __pgprot(page_readonly);
2329 protection_map[0xd] = __pgprot(page_readonly);
2330 protection_map[0xe] = __pgprot(page_shared);
2331 protection_map[0xf] = __pgprot(page_shared);
2332}
2333
2334static void __init sun4u_pgprot_init(void)
2335{
2336 unsigned long page_none, page_shared, page_copy, page_readonly;
2337 unsigned long page_exec_bit;
2338 int i;
2339
2340 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2341 _PAGE_CACHE_4U | _PAGE_P_4U |
2342 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2343 _PAGE_EXEC_4U);
2344 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2345 _PAGE_CACHE_4U | _PAGE_P_4U |
2346 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2347 _PAGE_EXEC_4U | _PAGE_L_4U);
2348
2349 _PAGE_IE = _PAGE_IE_4U;
2350 _PAGE_E = _PAGE_E_4U;
2351 _PAGE_CACHE = _PAGE_CACHE_4U;
2352
2353 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2354 __ACCESS_BITS_4U | _PAGE_E_4U);
2355
2356#ifdef CONFIG_DEBUG_PAGEALLOC
2357 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2358#else
2359 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2360 PAGE_OFFSET;
2361#endif
2362 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2363 _PAGE_P_4U | _PAGE_W_4U);
2364
2365 for (i = 1; i < 4; i++)
2366 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2367
2368 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2369 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2370 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2371
2372
2373 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2374 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2375 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2376 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2377 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2378 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2379 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2380
2381 page_exec_bit = _PAGE_EXEC_4U;
2382
2383 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2384 page_exec_bit);
2385}
2386
2387static void __init sun4v_pgprot_init(void)
2388{
2389 unsigned long page_none, page_shared, page_copy, page_readonly;
2390 unsigned long page_exec_bit;
2391 int i;
2392
2393 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2394 _PAGE_CACHE_4V | _PAGE_P_4V |
2395 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2396 _PAGE_EXEC_4V);
2397 PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2398
2399 _PAGE_IE = _PAGE_IE_4V;
2400 _PAGE_E = _PAGE_E_4V;
2401 _PAGE_CACHE = _PAGE_CACHE_4V;
2402
2403#ifdef CONFIG_DEBUG_PAGEALLOC
2404 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2405#else
2406 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2407 PAGE_OFFSET;
2408#endif
2409 kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2410 _PAGE_P_4V | _PAGE_W_4V);
2411
2412 for (i = 1; i < 4; i++)
2413 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2414
2415 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2416 __ACCESS_BITS_4V | _PAGE_E_4V);
2417
2418 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2419 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2420 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2421 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2422
2423 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2424 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2425 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2426 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2427 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2428 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2429 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2430
2431 page_exec_bit = _PAGE_EXEC_4V;
2432
2433 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2434 page_exec_bit);
2435}
2436
2437unsigned long pte_sz_bits(unsigned long sz)
2438{
2439 if (tlb_type == hypervisor) {
2440 switch (sz) {
2441 case 8 * 1024:
2442 default:
2443 return _PAGE_SZ8K_4V;
2444 case 64 * 1024:
2445 return _PAGE_SZ64K_4V;
2446 case 512 * 1024:
2447 return _PAGE_SZ512K_4V;
2448 case 4 * 1024 * 1024:
2449 return _PAGE_SZ4MB_4V;
2450 }
2451 } else {
2452 switch (sz) {
2453 case 8 * 1024:
2454 default:
2455 return _PAGE_SZ8K_4U;
2456 case 64 * 1024:
2457 return _PAGE_SZ64K_4U;
2458 case 512 * 1024:
2459 return _PAGE_SZ512K_4U;
2460 case 4 * 1024 * 1024:
2461 return _PAGE_SZ4MB_4U;
2462 }
2463 }
2464}
2465
2466pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2467{
2468 pte_t pte;
2469
2470 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot));
2471 pte_val(pte) |= (((unsigned long)space) << 32);
2472 pte_val(pte) |= pte_sz_bits(page_size);
2473
2474 return pte;
2475}
2476
2477static unsigned long kern_large_tte(unsigned long paddr)
2478{
2479 unsigned long val;
2480
2481 val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2482 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2483 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2484 if (tlb_type == hypervisor)
2485 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2486 _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2487 _PAGE_EXEC_4V | _PAGE_W_4V);
2488
2489 return val | paddr;
2490}
2491
2492/* If not locked, zap it. */
2493void __flush_tlb_all(void)
2494{
2495 unsigned long pstate;
2496 int i;
2497
2498 __asm__ __volatile__("flushw\n\t"
2499 "rdpr %%pstate, %0\n\t"
2500 "wrpr %0, %1, %%pstate"
2501 : "=r" (pstate)
2502 : "i" (PSTATE_IE));
2503 if (tlb_type == hypervisor) {
2504 sun4v_mmu_demap_all();
2505 } else if (tlb_type == spitfire) {
2506 for (i = 0; i < 64; i++) {
2507 /* Spitfire Errata #32 workaround */
2508 /* NOTE: Always runs on spitfire, so no
2509 * cheetah+ page size encodings.
2510 */
2511 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2512 "flush %%g6"
2513 : /* No outputs */
2514 : "r" (0),
2515 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2516
2517 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2518 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2519 "membar #Sync"
2520 : /* no outputs */
2521 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2522 spitfire_put_dtlb_data(i, 0x0UL);
2523 }
2524
2525 /* Spitfire Errata #32 workaround */
2526 /* NOTE: Always runs on spitfire, so no
2527 * cheetah+ page size encodings.
2528 */
2529 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2530 "flush %%g6"
2531 : /* No outputs */
2532 : "r" (0),
2533 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2534
2535 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2536 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2537 "membar #Sync"
2538 : /* no outputs */
2539 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2540 spitfire_put_itlb_data(i, 0x0UL);
2541 }
2542 }
2543 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2544 cheetah_flush_dtlb_all();
2545 cheetah_flush_itlb_all();
2546 }
2547 __asm__ __volatile__("wrpr %0, 0, %%pstate"
2548 : : "r" (pstate));
2549}
2550
2551pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2552 unsigned long address)
2553{
2554 struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2555 __GFP_REPEAT | __GFP_ZERO);
2556 pte_t *pte = NULL;
2557
2558 if (page)
2559 pte = (pte_t *) page_address(page);
2560
2561 return pte;
2562}
2563
2564pgtable_t pte_alloc_one(struct mm_struct *mm,
2565 unsigned long address)
2566{
2567 struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2568 __GFP_REPEAT | __GFP_ZERO);
2569 if (!page)
2570 return NULL;
2571 if (!pgtable_page_ctor(page)) {
2572 free_hot_cold_page(page, 0);
2573 return NULL;
2574 }
2575 return (pte_t *) page_address(page);
2576}
2577
2578void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2579{
2580 free_page((unsigned long)pte);
2581}
2582
2583static void __pte_free(pgtable_t pte)
2584{
2585 struct page *page = virt_to_page(pte);
2586
2587 pgtable_page_dtor(page);
2588 __free_page(page);
2589}
2590
2591void pte_free(struct mm_struct *mm, pgtable_t pte)
2592{
2593 __pte_free(pte);
2594}
2595
2596void pgtable_free(void *table, bool is_page)
2597{
2598 if (is_page)
2599 __pte_free(table);
2600 else
2601 kmem_cache_free(pgtable_cache, table);
2602}
2603
2604#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2605void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2606 pmd_t *pmd)
2607{
2608 unsigned long pte, flags;
2609 struct mm_struct *mm;
2610 pmd_t entry = *pmd;
2611
2612 if (!pmd_large(entry) || !pmd_young(entry))
2613 return;
2614
2615 pte = pmd_val(entry);
2616
2617 /* We are fabricating 8MB pages using 4MB real hw pages. */
2618 pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2619
2620 mm = vma->vm_mm;
2621
2622 spin_lock_irqsave(&mm->context.lock, flags);
2623
2624 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2625 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2626 addr, pte);
2627
2628 spin_unlock_irqrestore(&mm->context.lock, flags);
2629}
2630#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2631
2632#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2633static void context_reload(void *__data)
2634{
2635 struct mm_struct *mm = __data;
2636
2637 if (mm == current->mm)
2638 load_secondary_context(mm);
2639}
2640
2641void hugetlb_setup(struct pt_regs *regs)
2642{
2643 struct mm_struct *mm = current->mm;
2644 struct tsb_config *tp;
2645
2646 if (in_atomic() || !mm) {
2647 const struct exception_table_entry *entry;
2648
2649 entry = search_exception_tables(regs->tpc);
2650 if (entry) {
2651 regs->tpc = entry->fixup;
2652 regs->tnpc = regs->tpc + 4;
2653 return;
2654 }
2655 pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2656 die_if_kernel("HugeTSB in atomic", regs);
2657 }
2658
2659 tp = &mm->context.tsb_block[MM_TSB_HUGE];
2660 if (likely(tp->tsb == NULL))
2661 tsb_grow(mm, MM_TSB_HUGE, 0);
2662
2663 tsb_context_switch(mm);
2664 smp_tsb_sync(mm);
2665
2666 /* On UltraSPARC-III+ and later, configure the second half of
2667 * the Data-TLB for huge pages.
2668 */
2669 if (tlb_type == cheetah_plus) {
2670 unsigned long ctx;
2671
2672 spin_lock(&ctx_alloc_lock);
2673 ctx = mm->context.sparc64_ctx_val;
2674 ctx &= ~CTX_PGSZ_MASK;
2675 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2676 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2677
2678 if (ctx != mm->context.sparc64_ctx_val) {
2679 /* When changing the page size fields, we
2680 * must perform a context flush so that no
2681 * stale entries match. This flush must
2682 * occur with the original context register
2683 * settings.
2684 */
2685 do_flush_tlb_mm(mm);
2686
2687 /* Reload the context register of all processors
2688 * also executing in this address space.
2689 */
2690 mm->context.sparc64_ctx_val = ctx;
2691 on_each_cpu(context_reload, mm, 0);
2692 }
2693 spin_unlock(&ctx_alloc_lock);
2694 }
2695}
2696#endif