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/bootmem.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/memblock.h>
29#include <linux/mmzone.h>
30#include <linux/gfp.h>
31
32#include <asm/head.h>
33#include <asm/page.h>
34#include <asm/pgalloc.h>
35#include <asm/pgtable.h>
36#include <asm/oplib.h>
37#include <asm/iommu.h>
38#include <asm/io.h>
39#include <linux/uaccess.h>
40#include <asm/mmu_context.h>
41#include <asm/tlbflush.h>
42#include <asm/dma.h>
43#include <asm/starfire.h>
44#include <asm/tlb.h>
45#include <asm/spitfire.h>
46#include <asm/sections.h>
47#include <asm/tsb.h>
48#include <asm/hypervisor.h>
49#include <asm/prom.h>
50#include <asm/mdesc.h>
51#include <asm/cpudata.h>
52#include <asm/setup.h>
53#include <asm/irq.h>
54
55#include "init_64.h"
56
57unsigned long kern_linear_pte_xor[4] __read_mostly;
58static unsigned long page_cache4v_flag;
59
60/* A bitmap, two bits for every 256MB of physical memory. These two
61 * bits determine what page size we use for kernel linear
62 * translations. They form an index into kern_linear_pte_xor[]. The
63 * value in the indexed slot is XOR'd with the TLB miss virtual
64 * address to form the resulting TTE. The mapping is:
65 *
66 * 0 ==> 4MB
67 * 1 ==> 256MB
68 * 2 ==> 2GB
69 * 3 ==> 16GB
70 *
71 * All sun4v chips support 256MB pages. Only SPARC-T4 and later
72 * support 2GB pages, and hopefully future cpus will support the 16GB
73 * pages as well. For slots 2 and 3, we encode a 256MB TTE xor there
74 * if these larger page sizes are not supported by the cpu.
75 *
76 * It would be nice to determine this from the machine description
77 * 'cpu' properties, but we need to have this table setup before the
78 * MDESC is initialized.
79 */
80
81#ifndef CONFIG_DEBUG_PAGEALLOC
82/* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
83 * Space is allocated for this right after the trap table in
84 * arch/sparc64/kernel/head.S
85 */
86extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
87#endif
88extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
89
90static unsigned long cpu_pgsz_mask;
91
92#define MAX_BANKS 1024
93
94static struct linux_prom64_registers pavail[MAX_BANKS];
95static int pavail_ents;
96
97u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
98
99static int cmp_p64(const void *a, const void *b)
100{
101 const struct linux_prom64_registers *x = a, *y = b;
102
103 if (x->phys_addr > y->phys_addr)
104 return 1;
105 if (x->phys_addr < y->phys_addr)
106 return -1;
107 return 0;
108}
109
110static void __init read_obp_memory(const char *property,
111 struct linux_prom64_registers *regs,
112 int *num_ents)
113{
114 phandle node = prom_finddevice("/memory");
115 int prop_size = prom_getproplen(node, property);
116 int ents, ret, i;
117
118 ents = prop_size / sizeof(struct linux_prom64_registers);
119 if (ents > MAX_BANKS) {
120 prom_printf("The machine has more %s property entries than "
121 "this kernel can support (%d).\n",
122 property, MAX_BANKS);
123 prom_halt();
124 }
125
126 ret = prom_getproperty(node, property, (char *) regs, prop_size);
127 if (ret == -1) {
128 prom_printf("Couldn't get %s property from /memory.\n",
129 property);
130 prom_halt();
131 }
132
133 /* Sanitize what we got from the firmware, by page aligning
134 * everything.
135 */
136 for (i = 0; i < ents; i++) {
137 unsigned long base, size;
138
139 base = regs[i].phys_addr;
140 size = regs[i].reg_size;
141
142 size &= PAGE_MASK;
143 if (base & ~PAGE_MASK) {
144 unsigned long new_base = PAGE_ALIGN(base);
145
146 size -= new_base - base;
147 if ((long) size < 0L)
148 size = 0UL;
149 base = new_base;
150 }
151 if (size == 0UL) {
152 /* If it is empty, simply get rid of it.
153 * This simplifies the logic of the other
154 * functions that process these arrays.
155 */
156 memmove(®s[i], ®s[i + 1],
157 (ents - i - 1) * sizeof(regs[0]));
158 i--;
159 ents--;
160 continue;
161 }
162 regs[i].phys_addr = base;
163 regs[i].reg_size = size;
164 }
165
166 *num_ents = ents;
167
168 sort(regs, ents, sizeof(struct linux_prom64_registers),
169 cmp_p64, NULL);
170}
171
172/* Kernel physical address base and size in bytes. */
173unsigned long kern_base __read_mostly;
174unsigned long kern_size __read_mostly;
175
176/* Initial ramdisk setup */
177extern unsigned long sparc_ramdisk_image64;
178extern unsigned int sparc_ramdisk_image;
179extern unsigned int sparc_ramdisk_size;
180
181struct page *mem_map_zero __read_mostly;
182EXPORT_SYMBOL(mem_map_zero);
183
184unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
185
186unsigned long sparc64_kern_pri_context __read_mostly;
187unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
188unsigned long sparc64_kern_sec_context __read_mostly;
189
190int num_kernel_image_mappings;
191
192#ifdef CONFIG_DEBUG_DCFLUSH
193atomic_t dcpage_flushes = ATOMIC_INIT(0);
194#ifdef CONFIG_SMP
195atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
196#endif
197#endif
198
199inline void flush_dcache_page_impl(struct page *page)
200{
201 BUG_ON(tlb_type == hypervisor);
202#ifdef CONFIG_DEBUG_DCFLUSH
203 atomic_inc(&dcpage_flushes);
204#endif
205
206#ifdef DCACHE_ALIASING_POSSIBLE
207 __flush_dcache_page(page_address(page),
208 ((tlb_type == spitfire) &&
209 page_mapping_file(page) != NULL));
210#else
211 if (page_mapping_file(page) != NULL &&
212 tlb_type == spitfire)
213 __flush_icache_page(__pa(page_address(page)));
214#endif
215}
216
217#define PG_dcache_dirty PG_arch_1
218#define PG_dcache_cpu_shift 32UL
219#define PG_dcache_cpu_mask \
220 ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
221
222#define dcache_dirty_cpu(page) \
223 (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
224
225static inline void set_dcache_dirty(struct page *page, int this_cpu)
226{
227 unsigned long mask = this_cpu;
228 unsigned long non_cpu_bits;
229
230 non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
231 mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
232
233 __asm__ __volatile__("1:\n\t"
234 "ldx [%2], %%g7\n\t"
235 "and %%g7, %1, %%g1\n\t"
236 "or %%g1, %0, %%g1\n\t"
237 "casx [%2], %%g7, %%g1\n\t"
238 "cmp %%g7, %%g1\n\t"
239 "bne,pn %%xcc, 1b\n\t"
240 " nop"
241 : /* no outputs */
242 : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
243 : "g1", "g7");
244}
245
246static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
247{
248 unsigned long mask = (1UL << PG_dcache_dirty);
249
250 __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
251 "1:\n\t"
252 "ldx [%2], %%g7\n\t"
253 "srlx %%g7, %4, %%g1\n\t"
254 "and %%g1, %3, %%g1\n\t"
255 "cmp %%g1, %0\n\t"
256 "bne,pn %%icc, 2f\n\t"
257 " andn %%g7, %1, %%g1\n\t"
258 "casx [%2], %%g7, %%g1\n\t"
259 "cmp %%g7, %%g1\n\t"
260 "bne,pn %%xcc, 1b\n\t"
261 " nop\n"
262 "2:"
263 : /* no outputs */
264 : "r" (cpu), "r" (mask), "r" (&page->flags),
265 "i" (PG_dcache_cpu_mask),
266 "i" (PG_dcache_cpu_shift)
267 : "g1", "g7");
268}
269
270static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
271{
272 unsigned long tsb_addr = (unsigned long) ent;
273
274 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
275 tsb_addr = __pa(tsb_addr);
276
277 __tsb_insert(tsb_addr, tag, pte);
278}
279
280unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
281
282static void flush_dcache(unsigned long pfn)
283{
284 struct page *page;
285
286 page = pfn_to_page(pfn);
287 if (page) {
288 unsigned long pg_flags;
289
290 pg_flags = page->flags;
291 if (pg_flags & (1UL << PG_dcache_dirty)) {
292 int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
293 PG_dcache_cpu_mask);
294 int this_cpu = get_cpu();
295
296 /* This is just to optimize away some function calls
297 * in the SMP case.
298 */
299 if (cpu == this_cpu)
300 flush_dcache_page_impl(page);
301 else
302 smp_flush_dcache_page_impl(page, cpu);
303
304 clear_dcache_dirty_cpu(page, cpu);
305
306 put_cpu();
307 }
308 }
309}
310
311/* mm->context.lock must be held */
312static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
313 unsigned long tsb_hash_shift, unsigned long address,
314 unsigned long tte)
315{
316 struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
317 unsigned long tag;
318
319 if (unlikely(!tsb))
320 return;
321
322 tsb += ((address >> tsb_hash_shift) &
323 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
324 tag = (address >> 22UL);
325 tsb_insert(tsb, tag, tte);
326}
327
328#ifdef CONFIG_HUGETLB_PAGE
329static void __init add_huge_page_size(unsigned long size)
330{
331 unsigned int order;
332
333 if (size_to_hstate(size))
334 return;
335
336 order = ilog2(size) - PAGE_SHIFT;
337 hugetlb_add_hstate(order);
338}
339
340static int __init hugetlbpage_init(void)
341{
342 add_huge_page_size(1UL << HPAGE_64K_SHIFT);
343 add_huge_page_size(1UL << HPAGE_SHIFT);
344 add_huge_page_size(1UL << HPAGE_256MB_SHIFT);
345 add_huge_page_size(1UL << HPAGE_2GB_SHIFT);
346
347 return 0;
348}
349
350arch_initcall(hugetlbpage_init);
351
352static void __init pud_huge_patch(void)
353{
354 struct pud_huge_patch_entry *p;
355 unsigned long addr;
356
357 p = &__pud_huge_patch;
358 addr = p->addr;
359 *(unsigned int *)addr = p->insn;
360
361 __asm__ __volatile__("flush %0" : : "r" (addr));
362}
363
364static int __init setup_hugepagesz(char *string)
365{
366 unsigned long long hugepage_size;
367 unsigned int hugepage_shift;
368 unsigned short hv_pgsz_idx;
369 unsigned int hv_pgsz_mask;
370 int rc = 0;
371
372 hugepage_size = memparse(string, &string);
373 hugepage_shift = ilog2(hugepage_size);
374
375 switch (hugepage_shift) {
376 case HPAGE_16GB_SHIFT:
377 hv_pgsz_mask = HV_PGSZ_MASK_16GB;
378 hv_pgsz_idx = HV_PGSZ_IDX_16GB;
379 pud_huge_patch();
380 break;
381 case HPAGE_2GB_SHIFT:
382 hv_pgsz_mask = HV_PGSZ_MASK_2GB;
383 hv_pgsz_idx = HV_PGSZ_IDX_2GB;
384 break;
385 case HPAGE_256MB_SHIFT:
386 hv_pgsz_mask = HV_PGSZ_MASK_256MB;
387 hv_pgsz_idx = HV_PGSZ_IDX_256MB;
388 break;
389 case HPAGE_SHIFT:
390 hv_pgsz_mask = HV_PGSZ_MASK_4MB;
391 hv_pgsz_idx = HV_PGSZ_IDX_4MB;
392 break;
393 case HPAGE_64K_SHIFT:
394 hv_pgsz_mask = HV_PGSZ_MASK_64K;
395 hv_pgsz_idx = HV_PGSZ_IDX_64K;
396 break;
397 default:
398 hv_pgsz_mask = 0;
399 }
400
401 if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U) {
402 hugetlb_bad_size();
403 pr_err("hugepagesz=%llu not supported by MMU.\n",
404 hugepage_size);
405 goto out;
406 }
407
408 add_huge_page_size(hugepage_size);
409 rc = 1;
410
411out:
412 return rc;
413}
414__setup("hugepagesz=", setup_hugepagesz);
415#endif /* CONFIG_HUGETLB_PAGE */
416
417void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
418{
419 struct mm_struct *mm;
420 unsigned long flags;
421 bool is_huge_tsb;
422 pte_t pte = *ptep;
423
424 if (tlb_type != hypervisor) {
425 unsigned long pfn = pte_pfn(pte);
426
427 if (pfn_valid(pfn))
428 flush_dcache(pfn);
429 }
430
431 mm = vma->vm_mm;
432
433 /* Don't insert a non-valid PTE into the TSB, we'll deadlock. */
434 if (!pte_accessible(mm, pte))
435 return;
436
437 spin_lock_irqsave(&mm->context.lock, flags);
438
439 is_huge_tsb = false;
440#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
441 if (mm->context.hugetlb_pte_count || mm->context.thp_pte_count) {
442 unsigned long hugepage_size = PAGE_SIZE;
443
444 if (is_vm_hugetlb_page(vma))
445 hugepage_size = huge_page_size(hstate_vma(vma));
446
447 if (hugepage_size >= PUD_SIZE) {
448 unsigned long mask = 0x1ffc00000UL;
449
450 /* Transfer bits [32:22] from address to resolve
451 * at 4M granularity.
452 */
453 pte_val(pte) &= ~mask;
454 pte_val(pte) |= (address & mask);
455 } else if (hugepage_size >= PMD_SIZE) {
456 /* We are fabricating 8MB pages using 4MB
457 * real hw pages.
458 */
459 pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT));
460 }
461
462 if (hugepage_size >= PMD_SIZE) {
463 __update_mmu_tsb_insert(mm, MM_TSB_HUGE,
464 REAL_HPAGE_SHIFT, address, pte_val(pte));
465 is_huge_tsb = true;
466 }
467 }
468#endif
469 if (!is_huge_tsb)
470 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
471 address, pte_val(pte));
472
473 spin_unlock_irqrestore(&mm->context.lock, flags);
474}
475
476void flush_dcache_page(struct page *page)
477{
478 struct address_space *mapping;
479 int this_cpu;
480
481 if (tlb_type == hypervisor)
482 return;
483
484 /* Do not bother with the expensive D-cache flush if it
485 * is merely the zero page. The 'bigcore' testcase in GDB
486 * causes this case to run millions of times.
487 */
488 if (page == ZERO_PAGE(0))
489 return;
490
491 this_cpu = get_cpu();
492
493 mapping = page_mapping_file(page);
494 if (mapping && !mapping_mapped(mapping)) {
495 int dirty = test_bit(PG_dcache_dirty, &page->flags);
496 if (dirty) {
497 int dirty_cpu = dcache_dirty_cpu(page);
498
499 if (dirty_cpu == this_cpu)
500 goto out;
501 smp_flush_dcache_page_impl(page, dirty_cpu);
502 }
503 set_dcache_dirty(page, this_cpu);
504 } else {
505 /* We could delay the flush for the !page_mapping
506 * case too. But that case is for exec env/arg
507 * pages and those are %99 certainly going to get
508 * faulted into the tlb (and thus flushed) anyways.
509 */
510 flush_dcache_page_impl(page);
511 }
512
513out:
514 put_cpu();
515}
516EXPORT_SYMBOL(flush_dcache_page);
517
518void __kprobes flush_icache_range(unsigned long start, unsigned long end)
519{
520 /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
521 if (tlb_type == spitfire) {
522 unsigned long kaddr;
523
524 /* This code only runs on Spitfire cpus so this is
525 * why we can assume _PAGE_PADDR_4U.
526 */
527 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
528 unsigned long paddr, mask = _PAGE_PADDR_4U;
529
530 if (kaddr >= PAGE_OFFSET)
531 paddr = kaddr & mask;
532 else {
533 pgd_t *pgdp = pgd_offset_k(kaddr);
534 pud_t *pudp = pud_offset(pgdp, kaddr);
535 pmd_t *pmdp = pmd_offset(pudp, kaddr);
536 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
537
538 paddr = pte_val(*ptep) & mask;
539 }
540 __flush_icache_page(paddr);
541 }
542 }
543}
544EXPORT_SYMBOL(flush_icache_range);
545
546void mmu_info(struct seq_file *m)
547{
548 static const char *pgsz_strings[] = {
549 "8K", "64K", "512K", "4MB", "32MB",
550 "256MB", "2GB", "16GB",
551 };
552 int i, printed;
553
554 if (tlb_type == cheetah)
555 seq_printf(m, "MMU Type\t: Cheetah\n");
556 else if (tlb_type == cheetah_plus)
557 seq_printf(m, "MMU Type\t: Cheetah+\n");
558 else if (tlb_type == spitfire)
559 seq_printf(m, "MMU Type\t: Spitfire\n");
560 else if (tlb_type == hypervisor)
561 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
562 else
563 seq_printf(m, "MMU Type\t: ???\n");
564
565 seq_printf(m, "MMU PGSZs\t: ");
566 printed = 0;
567 for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
568 if (cpu_pgsz_mask & (1UL << i)) {
569 seq_printf(m, "%s%s",
570 printed ? "," : "", pgsz_strings[i]);
571 printed++;
572 }
573 }
574 seq_putc(m, '\n');
575
576#ifdef CONFIG_DEBUG_DCFLUSH
577 seq_printf(m, "DCPageFlushes\t: %d\n",
578 atomic_read(&dcpage_flushes));
579#ifdef CONFIG_SMP
580 seq_printf(m, "DCPageFlushesXC\t: %d\n",
581 atomic_read(&dcpage_flushes_xcall));
582#endif /* CONFIG_SMP */
583#endif /* CONFIG_DEBUG_DCFLUSH */
584}
585
586struct linux_prom_translation prom_trans[512] __read_mostly;
587unsigned int prom_trans_ents __read_mostly;
588
589unsigned long kern_locked_tte_data;
590
591/* The obp translations are saved based on 8k pagesize, since obp can
592 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
593 * HI_OBP_ADDRESS range are handled in ktlb.S.
594 */
595static inline int in_obp_range(unsigned long vaddr)
596{
597 return (vaddr >= LOW_OBP_ADDRESS &&
598 vaddr < HI_OBP_ADDRESS);
599}
600
601static int cmp_ptrans(const void *a, const void *b)
602{
603 const struct linux_prom_translation *x = a, *y = b;
604
605 if (x->virt > y->virt)
606 return 1;
607 if (x->virt < y->virt)
608 return -1;
609 return 0;
610}
611
612/* Read OBP translations property into 'prom_trans[]'. */
613static void __init read_obp_translations(void)
614{
615 int n, node, ents, first, last, i;
616
617 node = prom_finddevice("/virtual-memory");
618 n = prom_getproplen(node, "translations");
619 if (unlikely(n == 0 || n == -1)) {
620 prom_printf("prom_mappings: Couldn't get size.\n");
621 prom_halt();
622 }
623 if (unlikely(n > sizeof(prom_trans))) {
624 prom_printf("prom_mappings: Size %d is too big.\n", n);
625 prom_halt();
626 }
627
628 if ((n = prom_getproperty(node, "translations",
629 (char *)&prom_trans[0],
630 sizeof(prom_trans))) == -1) {
631 prom_printf("prom_mappings: Couldn't get property.\n");
632 prom_halt();
633 }
634
635 n = n / sizeof(struct linux_prom_translation);
636
637 ents = n;
638
639 sort(prom_trans, ents, sizeof(struct linux_prom_translation),
640 cmp_ptrans, NULL);
641
642 /* Now kick out all the non-OBP entries. */
643 for (i = 0; i < ents; i++) {
644 if (in_obp_range(prom_trans[i].virt))
645 break;
646 }
647 first = i;
648 for (; i < ents; i++) {
649 if (!in_obp_range(prom_trans[i].virt))
650 break;
651 }
652 last = i;
653
654 for (i = 0; i < (last - first); i++) {
655 struct linux_prom_translation *src = &prom_trans[i + first];
656 struct linux_prom_translation *dest = &prom_trans[i];
657
658 *dest = *src;
659 }
660 for (; i < ents; i++) {
661 struct linux_prom_translation *dest = &prom_trans[i];
662 dest->virt = dest->size = dest->data = 0x0UL;
663 }
664
665 prom_trans_ents = last - first;
666
667 if (tlb_type == spitfire) {
668 /* Clear diag TTE bits. */
669 for (i = 0; i < prom_trans_ents; i++)
670 prom_trans[i].data &= ~0x0003fe0000000000UL;
671 }
672
673 /* Force execute bit on. */
674 for (i = 0; i < prom_trans_ents; i++)
675 prom_trans[i].data |= (tlb_type == hypervisor ?
676 _PAGE_EXEC_4V : _PAGE_EXEC_4U);
677}
678
679static void __init hypervisor_tlb_lock(unsigned long vaddr,
680 unsigned long pte,
681 unsigned long mmu)
682{
683 unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
684
685 if (ret != 0) {
686 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
687 "errors with %lx\n", vaddr, 0, pte, mmu, ret);
688 prom_halt();
689 }
690}
691
692static unsigned long kern_large_tte(unsigned long paddr);
693
694static void __init remap_kernel(void)
695{
696 unsigned long phys_page, tte_vaddr, tte_data;
697 int i, tlb_ent = sparc64_highest_locked_tlbent();
698
699 tte_vaddr = (unsigned long) KERNBASE;
700 phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
701 tte_data = kern_large_tte(phys_page);
702
703 kern_locked_tte_data = tte_data;
704
705 /* Now lock us into the TLBs via Hypervisor or OBP. */
706 if (tlb_type == hypervisor) {
707 for (i = 0; i < num_kernel_image_mappings; i++) {
708 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
709 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
710 tte_vaddr += 0x400000;
711 tte_data += 0x400000;
712 }
713 } else {
714 for (i = 0; i < num_kernel_image_mappings; i++) {
715 prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
716 prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
717 tte_vaddr += 0x400000;
718 tte_data += 0x400000;
719 }
720 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
721 }
722 if (tlb_type == cheetah_plus) {
723 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
724 CTX_CHEETAH_PLUS_NUC);
725 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
726 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
727 }
728}
729
730
731static void __init inherit_prom_mappings(void)
732{
733 /* Now fixup OBP's idea about where we really are mapped. */
734 printk("Remapping the kernel... ");
735 remap_kernel();
736 printk("done.\n");
737}
738
739void prom_world(int enter)
740{
741 if (!enter)
742 set_fs(get_fs());
743
744 __asm__ __volatile__("flushw");
745}
746
747void __flush_dcache_range(unsigned long start, unsigned long end)
748{
749 unsigned long va;
750
751 if (tlb_type == spitfire) {
752 int n = 0;
753
754 for (va = start; va < end; va += 32) {
755 spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
756 if (++n >= 512)
757 break;
758 }
759 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
760 start = __pa(start);
761 end = __pa(end);
762 for (va = start; va < end; va += 32)
763 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
764 "membar #Sync"
765 : /* no outputs */
766 : "r" (va),
767 "i" (ASI_DCACHE_INVALIDATE));
768 }
769}
770EXPORT_SYMBOL(__flush_dcache_range);
771
772/* get_new_mmu_context() uses "cache + 1". */
773DEFINE_SPINLOCK(ctx_alloc_lock);
774unsigned long tlb_context_cache = CTX_FIRST_VERSION;
775#define MAX_CTX_NR (1UL << CTX_NR_BITS)
776#define CTX_BMAP_SLOTS BITS_TO_LONGS(MAX_CTX_NR)
777DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
778DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0};
779
780static void mmu_context_wrap(void)
781{
782 unsigned long old_ver = tlb_context_cache & CTX_VERSION_MASK;
783 unsigned long new_ver, new_ctx, old_ctx;
784 struct mm_struct *mm;
785 int cpu;
786
787 bitmap_zero(mmu_context_bmap, 1 << CTX_NR_BITS);
788
789 /* Reserve kernel context */
790 set_bit(0, mmu_context_bmap);
791
792 new_ver = (tlb_context_cache & CTX_VERSION_MASK) + CTX_FIRST_VERSION;
793 if (unlikely(new_ver == 0))
794 new_ver = CTX_FIRST_VERSION;
795 tlb_context_cache = new_ver;
796
797 /*
798 * Make sure that any new mm that are added into per_cpu_secondary_mm,
799 * are going to go through get_new_mmu_context() path.
800 */
801 mb();
802
803 /*
804 * Updated versions to current on those CPUs that had valid secondary
805 * contexts
806 */
807 for_each_online_cpu(cpu) {
808 /*
809 * If a new mm is stored after we took this mm from the array,
810 * it will go into get_new_mmu_context() path, because we
811 * already bumped the version in tlb_context_cache.
812 */
813 mm = per_cpu(per_cpu_secondary_mm, cpu);
814
815 if (unlikely(!mm || mm == &init_mm))
816 continue;
817
818 old_ctx = mm->context.sparc64_ctx_val;
819 if (likely((old_ctx & CTX_VERSION_MASK) == old_ver)) {
820 new_ctx = (old_ctx & ~CTX_VERSION_MASK) | new_ver;
821 set_bit(new_ctx & CTX_NR_MASK, mmu_context_bmap);
822 mm->context.sparc64_ctx_val = new_ctx;
823 }
824 }
825}
826
827/* Caller does TLB context flushing on local CPU if necessary.
828 * The caller also ensures that CTX_VALID(mm->context) is false.
829 *
830 * We must be careful about boundary cases so that we never
831 * let the user have CTX 0 (nucleus) or we ever use a CTX
832 * version of zero (and thus NO_CONTEXT would not be caught
833 * by version mis-match tests in mmu_context.h).
834 *
835 * Always invoked with interrupts disabled.
836 */
837void get_new_mmu_context(struct mm_struct *mm)
838{
839 unsigned long ctx, new_ctx;
840 unsigned long orig_pgsz_bits;
841
842 spin_lock(&ctx_alloc_lock);
843retry:
844 /* wrap might have happened, test again if our context became valid */
845 if (unlikely(CTX_VALID(mm->context)))
846 goto out;
847 orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
848 ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
849 new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
850 if (new_ctx >= (1 << CTX_NR_BITS)) {
851 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
852 if (new_ctx >= ctx) {
853 mmu_context_wrap();
854 goto retry;
855 }
856 }
857 if (mm->context.sparc64_ctx_val)
858 cpumask_clear(mm_cpumask(mm));
859 mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
860 new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
861 tlb_context_cache = new_ctx;
862 mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
863out:
864 spin_unlock(&ctx_alloc_lock);
865}
866
867static int numa_enabled = 1;
868static int numa_debug;
869
870static int __init early_numa(char *p)
871{
872 if (!p)
873 return 0;
874
875 if (strstr(p, "off"))
876 numa_enabled = 0;
877
878 if (strstr(p, "debug"))
879 numa_debug = 1;
880
881 return 0;
882}
883early_param("numa", early_numa);
884
885#define numadbg(f, a...) \
886do { if (numa_debug) \
887 printk(KERN_INFO f, ## a); \
888} while (0)
889
890static void __init find_ramdisk(unsigned long phys_base)
891{
892#ifdef CONFIG_BLK_DEV_INITRD
893 if (sparc_ramdisk_image || sparc_ramdisk_image64) {
894 unsigned long ramdisk_image;
895
896 /* Older versions of the bootloader only supported a
897 * 32-bit physical address for the ramdisk image
898 * location, stored at sparc_ramdisk_image. Newer
899 * SILO versions set sparc_ramdisk_image to zero and
900 * provide a full 64-bit physical address at
901 * sparc_ramdisk_image64.
902 */
903 ramdisk_image = sparc_ramdisk_image;
904 if (!ramdisk_image)
905 ramdisk_image = sparc_ramdisk_image64;
906
907 /* Another bootloader quirk. The bootloader normalizes
908 * the physical address to KERNBASE, so we have to
909 * factor that back out and add in the lowest valid
910 * physical page address to get the true physical address.
911 */
912 ramdisk_image -= KERNBASE;
913 ramdisk_image += phys_base;
914
915 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
916 ramdisk_image, sparc_ramdisk_size);
917
918 initrd_start = ramdisk_image;
919 initrd_end = ramdisk_image + sparc_ramdisk_size;
920
921 memblock_reserve(initrd_start, sparc_ramdisk_size);
922
923 initrd_start += PAGE_OFFSET;
924 initrd_end += PAGE_OFFSET;
925 }
926#endif
927}
928
929struct node_mem_mask {
930 unsigned long mask;
931 unsigned long match;
932};
933static struct node_mem_mask node_masks[MAX_NUMNODES];
934static int num_node_masks;
935
936#ifdef CONFIG_NEED_MULTIPLE_NODES
937
938struct mdesc_mlgroup {
939 u64 node;
940 u64 latency;
941 u64 match;
942 u64 mask;
943};
944
945static struct mdesc_mlgroup *mlgroups;
946static int num_mlgroups;
947
948int numa_cpu_lookup_table[NR_CPUS];
949cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
950
951struct mdesc_mblock {
952 u64 base;
953 u64 size;
954 u64 offset; /* RA-to-PA */
955};
956static struct mdesc_mblock *mblocks;
957static int num_mblocks;
958
959static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr)
960{
961 struct mdesc_mblock *m = NULL;
962 int i;
963
964 for (i = 0; i < num_mblocks; i++) {
965 m = &mblocks[i];
966
967 if (addr >= m->base &&
968 addr < (m->base + m->size)) {
969 break;
970 }
971 }
972
973 return m;
974}
975
976static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid)
977{
978 int prev_nid, new_nid;
979
980 prev_nid = -1;
981 for ( ; start < end; start += PAGE_SIZE) {
982 for (new_nid = 0; new_nid < num_node_masks; new_nid++) {
983 struct node_mem_mask *p = &node_masks[new_nid];
984
985 if ((start & p->mask) == p->match) {
986 if (prev_nid == -1)
987 prev_nid = new_nid;
988 break;
989 }
990 }
991
992 if (new_nid == num_node_masks) {
993 prev_nid = 0;
994 WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.",
995 start);
996 break;
997 }
998
999 if (prev_nid != new_nid)
1000 break;
1001 }
1002 *nid = prev_nid;
1003
1004 return start > end ? end : start;
1005}
1006
1007static u64 __init memblock_nid_range(u64 start, u64 end, int *nid)
1008{
1009 u64 ret_end, pa_start, m_mask, m_match, m_end;
1010 struct mdesc_mblock *mblock;
1011 int _nid, i;
1012
1013 if (tlb_type != hypervisor)
1014 return memblock_nid_range_sun4u(start, end, nid);
1015
1016 mblock = addr_to_mblock(start);
1017 if (!mblock) {
1018 WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]",
1019 start);
1020
1021 _nid = 0;
1022 ret_end = end;
1023 goto done;
1024 }
1025
1026 pa_start = start + mblock->offset;
1027 m_match = 0;
1028 m_mask = 0;
1029
1030 for (_nid = 0; _nid < num_node_masks; _nid++) {
1031 struct node_mem_mask *const m = &node_masks[_nid];
1032
1033 if ((pa_start & m->mask) == m->match) {
1034 m_match = m->match;
1035 m_mask = m->mask;
1036 break;
1037 }
1038 }
1039
1040 if (num_node_masks == _nid) {
1041 /* We could not find NUMA group, so default to 0, but lets
1042 * search for latency group, so we could calculate the correct
1043 * end address that we return
1044 */
1045 _nid = 0;
1046
1047 for (i = 0; i < num_mlgroups; i++) {
1048 struct mdesc_mlgroup *const m = &mlgroups[i];
1049
1050 if ((pa_start & m->mask) == m->match) {
1051 m_match = m->match;
1052 m_mask = m->mask;
1053 break;
1054 }
1055 }
1056
1057 if (i == num_mlgroups) {
1058 WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]",
1059 start);
1060
1061 ret_end = end;
1062 goto done;
1063 }
1064 }
1065
1066 /*
1067 * Each latency group has match and mask, and each memory block has an
1068 * offset. An address belongs to a latency group if its address matches
1069 * the following formula: ((addr + offset) & mask) == match
1070 * It is, however, slow to check every single page if it matches a
1071 * particular latency group. As optimization we calculate end value by
1072 * using bit arithmetics.
1073 */
1074 m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset;
1075 m_end += pa_start & ~((1ul << fls64(m_mask)) - 1);
1076 ret_end = m_end > end ? end : m_end;
1077
1078done:
1079 *nid = _nid;
1080 return ret_end;
1081}
1082#endif
1083
1084/* This must be invoked after performing all of the necessary
1085 * memblock_set_node() calls for 'nid'. We need to be able to get
1086 * correct data from get_pfn_range_for_nid().
1087 */
1088static void __init allocate_node_data(int nid)
1089{
1090 struct pglist_data *p;
1091 unsigned long start_pfn, end_pfn;
1092#ifdef CONFIG_NEED_MULTIPLE_NODES
1093 unsigned long paddr;
1094
1095 paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
1096 if (!paddr) {
1097 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
1098 prom_halt();
1099 }
1100 NODE_DATA(nid) = __va(paddr);
1101 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
1102
1103 NODE_DATA(nid)->node_id = nid;
1104#endif
1105
1106 p = NODE_DATA(nid);
1107
1108 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1109 p->node_start_pfn = start_pfn;
1110 p->node_spanned_pages = end_pfn - start_pfn;
1111}
1112
1113static void init_node_masks_nonnuma(void)
1114{
1115#ifdef CONFIG_NEED_MULTIPLE_NODES
1116 int i;
1117#endif
1118
1119 numadbg("Initializing tables for non-numa.\n");
1120
1121 node_masks[0].mask = 0;
1122 node_masks[0].match = 0;
1123 num_node_masks = 1;
1124
1125#ifdef CONFIG_NEED_MULTIPLE_NODES
1126 for (i = 0; i < NR_CPUS; i++)
1127 numa_cpu_lookup_table[i] = 0;
1128
1129 cpumask_setall(&numa_cpumask_lookup_table[0]);
1130#endif
1131}
1132
1133#ifdef CONFIG_NEED_MULTIPLE_NODES
1134struct pglist_data *node_data[MAX_NUMNODES];
1135
1136EXPORT_SYMBOL(numa_cpu_lookup_table);
1137EXPORT_SYMBOL(numa_cpumask_lookup_table);
1138EXPORT_SYMBOL(node_data);
1139
1140static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
1141 u32 cfg_handle)
1142{
1143 u64 arc;
1144
1145 mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
1146 u64 target = mdesc_arc_target(md, arc);
1147 const u64 *val;
1148
1149 val = mdesc_get_property(md, target,
1150 "cfg-handle", NULL);
1151 if (val && *val == cfg_handle)
1152 return 0;
1153 }
1154 return -ENODEV;
1155}
1156
1157static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
1158 u32 cfg_handle)
1159{
1160 u64 arc, candidate, best_latency = ~(u64)0;
1161
1162 candidate = MDESC_NODE_NULL;
1163 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1164 u64 target = mdesc_arc_target(md, arc);
1165 const char *name = mdesc_node_name(md, target);
1166 const u64 *val;
1167
1168 if (strcmp(name, "pio-latency-group"))
1169 continue;
1170
1171 val = mdesc_get_property(md, target, "latency", NULL);
1172 if (!val)
1173 continue;
1174
1175 if (*val < best_latency) {
1176 candidate = target;
1177 best_latency = *val;
1178 }
1179 }
1180
1181 if (candidate == MDESC_NODE_NULL)
1182 return -ENODEV;
1183
1184 return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
1185}
1186
1187int of_node_to_nid(struct device_node *dp)
1188{
1189 const struct linux_prom64_registers *regs;
1190 struct mdesc_handle *md;
1191 u32 cfg_handle;
1192 int count, nid;
1193 u64 grp;
1194
1195 /* This is the right thing to do on currently supported
1196 * SUN4U NUMA platforms as well, as the PCI controller does
1197 * not sit behind any particular memory controller.
1198 */
1199 if (!mlgroups)
1200 return -1;
1201
1202 regs = of_get_property(dp, "reg", NULL);
1203 if (!regs)
1204 return -1;
1205
1206 cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1207
1208 md = mdesc_grab();
1209
1210 count = 0;
1211 nid = -1;
1212 mdesc_for_each_node_by_name(md, grp, "group") {
1213 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1214 nid = count;
1215 break;
1216 }
1217 count++;
1218 }
1219
1220 mdesc_release(md);
1221
1222 return nid;
1223}
1224
1225static void __init add_node_ranges(void)
1226{
1227 struct memblock_region *reg;
1228 unsigned long prev_max;
1229
1230memblock_resized:
1231 prev_max = memblock.memory.max;
1232
1233 for_each_memblock(memory, reg) {
1234 unsigned long size = reg->size;
1235 unsigned long start, end;
1236
1237 start = reg->base;
1238 end = start + size;
1239 while (start < end) {
1240 unsigned long this_end;
1241 int nid;
1242
1243 this_end = memblock_nid_range(start, end, &nid);
1244
1245 numadbg("Setting memblock NUMA node nid[%d] "
1246 "start[%lx] end[%lx]\n",
1247 nid, start, this_end);
1248
1249 memblock_set_node(start, this_end - start,
1250 &memblock.memory, nid);
1251 if (memblock.memory.max != prev_max)
1252 goto memblock_resized;
1253 start = this_end;
1254 }
1255 }
1256}
1257
1258static int __init grab_mlgroups(struct mdesc_handle *md)
1259{
1260 unsigned long paddr;
1261 int count = 0;
1262 u64 node;
1263
1264 mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1265 count++;
1266 if (!count)
1267 return -ENOENT;
1268
1269 paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1270 SMP_CACHE_BYTES);
1271 if (!paddr)
1272 return -ENOMEM;
1273
1274 mlgroups = __va(paddr);
1275 num_mlgroups = count;
1276
1277 count = 0;
1278 mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1279 struct mdesc_mlgroup *m = &mlgroups[count++];
1280 const u64 *val;
1281
1282 m->node = node;
1283
1284 val = mdesc_get_property(md, node, "latency", NULL);
1285 m->latency = *val;
1286 val = mdesc_get_property(md, node, "address-match", NULL);
1287 m->match = *val;
1288 val = mdesc_get_property(md, node, "address-mask", NULL);
1289 m->mask = *val;
1290
1291 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1292 "match[%llx] mask[%llx]\n",
1293 count - 1, m->node, m->latency, m->match, m->mask);
1294 }
1295
1296 return 0;
1297}
1298
1299static int __init grab_mblocks(struct mdesc_handle *md)
1300{
1301 unsigned long paddr;
1302 int count = 0;
1303 u64 node;
1304
1305 mdesc_for_each_node_by_name(md, node, "mblock")
1306 count++;
1307 if (!count)
1308 return -ENOENT;
1309
1310 paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1311 SMP_CACHE_BYTES);
1312 if (!paddr)
1313 return -ENOMEM;
1314
1315 mblocks = __va(paddr);
1316 num_mblocks = count;
1317
1318 count = 0;
1319 mdesc_for_each_node_by_name(md, node, "mblock") {
1320 struct mdesc_mblock *m = &mblocks[count++];
1321 const u64 *val;
1322
1323 val = mdesc_get_property(md, node, "base", NULL);
1324 m->base = *val;
1325 val = mdesc_get_property(md, node, "size", NULL);
1326 m->size = *val;
1327 val = mdesc_get_property(md, node,
1328 "address-congruence-offset", NULL);
1329
1330 /* The address-congruence-offset property is optional.
1331 * Explicity zero it be identifty this.
1332 */
1333 if (val)
1334 m->offset = *val;
1335 else
1336 m->offset = 0UL;
1337
1338 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1339 count - 1, m->base, m->size, m->offset);
1340 }
1341
1342 return 0;
1343}
1344
1345static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1346 u64 grp, cpumask_t *mask)
1347{
1348 u64 arc;
1349
1350 cpumask_clear(mask);
1351
1352 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1353 u64 target = mdesc_arc_target(md, arc);
1354 const char *name = mdesc_node_name(md, target);
1355 const u64 *id;
1356
1357 if (strcmp(name, "cpu"))
1358 continue;
1359 id = mdesc_get_property(md, target, "id", NULL);
1360 if (*id < nr_cpu_ids)
1361 cpumask_set_cpu(*id, mask);
1362 }
1363}
1364
1365static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1366{
1367 int i;
1368
1369 for (i = 0; i < num_mlgroups; i++) {
1370 struct mdesc_mlgroup *m = &mlgroups[i];
1371 if (m->node == node)
1372 return m;
1373 }
1374 return NULL;
1375}
1376
1377int __node_distance(int from, int to)
1378{
1379 if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1380 pr_warn("Returning default NUMA distance value for %d->%d\n",
1381 from, to);
1382 return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1383 }
1384 return numa_latency[from][to];
1385}
1386
1387static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1388{
1389 int i;
1390
1391 for (i = 0; i < MAX_NUMNODES; i++) {
1392 struct node_mem_mask *n = &node_masks[i];
1393
1394 if ((grp->mask == n->mask) && (grp->match == n->match))
1395 break;
1396 }
1397 return i;
1398}
1399
1400static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1401 u64 grp, int index)
1402{
1403 u64 arc;
1404
1405 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1406 int tnode;
1407 u64 target = mdesc_arc_target(md, arc);
1408 struct mdesc_mlgroup *m = find_mlgroup(target);
1409
1410 if (!m)
1411 continue;
1412 tnode = find_best_numa_node_for_mlgroup(m);
1413 if (tnode == MAX_NUMNODES)
1414 continue;
1415 numa_latency[index][tnode] = m->latency;
1416 }
1417}
1418
1419static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1420 int index)
1421{
1422 struct mdesc_mlgroup *candidate = NULL;
1423 u64 arc, best_latency = ~(u64)0;
1424 struct node_mem_mask *n;
1425
1426 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1427 u64 target = mdesc_arc_target(md, arc);
1428 struct mdesc_mlgroup *m = find_mlgroup(target);
1429 if (!m)
1430 continue;
1431 if (m->latency < best_latency) {
1432 candidate = m;
1433 best_latency = m->latency;
1434 }
1435 }
1436 if (!candidate)
1437 return -ENOENT;
1438
1439 if (num_node_masks != index) {
1440 printk(KERN_ERR "Inconsistent NUMA state, "
1441 "index[%d] != num_node_masks[%d]\n",
1442 index, num_node_masks);
1443 return -EINVAL;
1444 }
1445
1446 n = &node_masks[num_node_masks++];
1447
1448 n->mask = candidate->mask;
1449 n->match = candidate->match;
1450
1451 numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n",
1452 index, n->mask, n->match, candidate->latency);
1453
1454 return 0;
1455}
1456
1457static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1458 int index)
1459{
1460 cpumask_t mask;
1461 int cpu;
1462
1463 numa_parse_mdesc_group_cpus(md, grp, &mask);
1464
1465 for_each_cpu(cpu, &mask)
1466 numa_cpu_lookup_table[cpu] = index;
1467 cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1468
1469 if (numa_debug) {
1470 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1471 for_each_cpu(cpu, &mask)
1472 printk("%d ", cpu);
1473 printk("]\n");
1474 }
1475
1476 return numa_attach_mlgroup(md, grp, index);
1477}
1478
1479static int __init numa_parse_mdesc(void)
1480{
1481 struct mdesc_handle *md = mdesc_grab();
1482 int i, j, err, count;
1483 u64 node;
1484
1485 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1486 if (node == MDESC_NODE_NULL) {
1487 mdesc_release(md);
1488 return -ENOENT;
1489 }
1490
1491 err = grab_mblocks(md);
1492 if (err < 0)
1493 goto out;
1494
1495 err = grab_mlgroups(md);
1496 if (err < 0)
1497 goto out;
1498
1499 count = 0;
1500 mdesc_for_each_node_by_name(md, node, "group") {
1501 err = numa_parse_mdesc_group(md, node, count);
1502 if (err < 0)
1503 break;
1504 count++;
1505 }
1506
1507 count = 0;
1508 mdesc_for_each_node_by_name(md, node, "group") {
1509 find_numa_latencies_for_group(md, node, count);
1510 count++;
1511 }
1512
1513 /* Normalize numa latency matrix according to ACPI SLIT spec. */
1514 for (i = 0; i < MAX_NUMNODES; i++) {
1515 u64 self_latency = numa_latency[i][i];
1516
1517 for (j = 0; j < MAX_NUMNODES; j++) {
1518 numa_latency[i][j] =
1519 (numa_latency[i][j] * LOCAL_DISTANCE) /
1520 self_latency;
1521 }
1522 }
1523
1524 add_node_ranges();
1525
1526 for (i = 0; i < num_node_masks; i++) {
1527 allocate_node_data(i);
1528 node_set_online(i);
1529 }
1530
1531 err = 0;
1532out:
1533 mdesc_release(md);
1534 return err;
1535}
1536
1537static int __init numa_parse_jbus(void)
1538{
1539 unsigned long cpu, index;
1540
1541 /* NUMA node id is encoded in bits 36 and higher, and there is
1542 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1543 */
1544 index = 0;
1545 for_each_present_cpu(cpu) {
1546 numa_cpu_lookup_table[cpu] = index;
1547 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1548 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1549 node_masks[index].match = cpu << 36UL;
1550
1551 index++;
1552 }
1553 num_node_masks = index;
1554
1555 add_node_ranges();
1556
1557 for (index = 0; index < num_node_masks; index++) {
1558 allocate_node_data(index);
1559 node_set_online(index);
1560 }
1561
1562 return 0;
1563}
1564
1565static int __init numa_parse_sun4u(void)
1566{
1567 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1568 unsigned long ver;
1569
1570 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1571 if ((ver >> 32UL) == __JALAPENO_ID ||
1572 (ver >> 32UL) == __SERRANO_ID)
1573 return numa_parse_jbus();
1574 }
1575 return -1;
1576}
1577
1578static int __init bootmem_init_numa(void)
1579{
1580 int i, j;
1581 int err = -1;
1582
1583 numadbg("bootmem_init_numa()\n");
1584
1585 /* Some sane defaults for numa latency values */
1586 for (i = 0; i < MAX_NUMNODES; i++) {
1587 for (j = 0; j < MAX_NUMNODES; j++)
1588 numa_latency[i][j] = (i == j) ?
1589 LOCAL_DISTANCE : REMOTE_DISTANCE;
1590 }
1591
1592 if (numa_enabled) {
1593 if (tlb_type == hypervisor)
1594 err = numa_parse_mdesc();
1595 else
1596 err = numa_parse_sun4u();
1597 }
1598 return err;
1599}
1600
1601#else
1602
1603static int bootmem_init_numa(void)
1604{
1605 return -1;
1606}
1607
1608#endif
1609
1610static void __init bootmem_init_nonnuma(void)
1611{
1612 unsigned long top_of_ram = memblock_end_of_DRAM();
1613 unsigned long total_ram = memblock_phys_mem_size();
1614
1615 numadbg("bootmem_init_nonnuma()\n");
1616
1617 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1618 top_of_ram, total_ram);
1619 printk(KERN_INFO "Memory hole size: %ldMB\n",
1620 (top_of_ram - total_ram) >> 20);
1621
1622 init_node_masks_nonnuma();
1623 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1624 allocate_node_data(0);
1625 node_set_online(0);
1626}
1627
1628static unsigned long __init bootmem_init(unsigned long phys_base)
1629{
1630 unsigned long end_pfn;
1631
1632 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1633 max_pfn = max_low_pfn = end_pfn;
1634 min_low_pfn = (phys_base >> PAGE_SHIFT);
1635
1636 if (bootmem_init_numa() < 0)
1637 bootmem_init_nonnuma();
1638
1639 /* Dump memblock with node info. */
1640 memblock_dump_all();
1641
1642 /* XXX cpu notifier XXX */
1643
1644 sparse_memory_present_with_active_regions(MAX_NUMNODES);
1645 sparse_init();
1646
1647 return end_pfn;
1648}
1649
1650static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1651static int pall_ents __initdata;
1652
1653static unsigned long max_phys_bits = 40;
1654
1655bool kern_addr_valid(unsigned long addr)
1656{
1657 pgd_t *pgd;
1658 pud_t *pud;
1659 pmd_t *pmd;
1660 pte_t *pte;
1661
1662 if ((long)addr < 0L) {
1663 unsigned long pa = __pa(addr);
1664
1665 if ((pa >> max_phys_bits) != 0UL)
1666 return false;
1667
1668 return pfn_valid(pa >> PAGE_SHIFT);
1669 }
1670
1671 if (addr >= (unsigned long) KERNBASE &&
1672 addr < (unsigned long)&_end)
1673 return true;
1674
1675 pgd = pgd_offset_k(addr);
1676 if (pgd_none(*pgd))
1677 return 0;
1678
1679 pud = pud_offset(pgd, addr);
1680 if (pud_none(*pud))
1681 return 0;
1682
1683 if (pud_large(*pud))
1684 return pfn_valid(pud_pfn(*pud));
1685
1686 pmd = pmd_offset(pud, addr);
1687 if (pmd_none(*pmd))
1688 return 0;
1689
1690 if (pmd_large(*pmd))
1691 return pfn_valid(pmd_pfn(*pmd));
1692
1693 pte = pte_offset_kernel(pmd, addr);
1694 if (pte_none(*pte))
1695 return 0;
1696
1697 return pfn_valid(pte_pfn(*pte));
1698}
1699EXPORT_SYMBOL(kern_addr_valid);
1700
1701static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1702 unsigned long vend,
1703 pud_t *pud)
1704{
1705 const unsigned long mask16gb = (1UL << 34) - 1UL;
1706 u64 pte_val = vstart;
1707
1708 /* Each PUD is 8GB */
1709 if ((vstart & mask16gb) ||
1710 (vend - vstart <= mask16gb)) {
1711 pte_val ^= kern_linear_pte_xor[2];
1712 pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1713
1714 return vstart + PUD_SIZE;
1715 }
1716
1717 pte_val ^= kern_linear_pte_xor[3];
1718 pte_val |= _PAGE_PUD_HUGE;
1719
1720 vend = vstart + mask16gb + 1UL;
1721 while (vstart < vend) {
1722 pud_val(*pud) = pte_val;
1723
1724 pte_val += PUD_SIZE;
1725 vstart += PUD_SIZE;
1726 pud++;
1727 }
1728 return vstart;
1729}
1730
1731static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1732 bool guard)
1733{
1734 if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1735 return true;
1736
1737 return false;
1738}
1739
1740static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1741 unsigned long vend,
1742 pmd_t *pmd)
1743{
1744 const unsigned long mask256mb = (1UL << 28) - 1UL;
1745 const unsigned long mask2gb = (1UL << 31) - 1UL;
1746 u64 pte_val = vstart;
1747
1748 /* Each PMD is 8MB */
1749 if ((vstart & mask256mb) ||
1750 (vend - vstart <= mask256mb)) {
1751 pte_val ^= kern_linear_pte_xor[0];
1752 pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1753
1754 return vstart + PMD_SIZE;
1755 }
1756
1757 if ((vstart & mask2gb) ||
1758 (vend - vstart <= mask2gb)) {
1759 pte_val ^= kern_linear_pte_xor[1];
1760 pte_val |= _PAGE_PMD_HUGE;
1761 vend = vstart + mask256mb + 1UL;
1762 } else {
1763 pte_val ^= kern_linear_pte_xor[2];
1764 pte_val |= _PAGE_PMD_HUGE;
1765 vend = vstart + mask2gb + 1UL;
1766 }
1767
1768 while (vstart < vend) {
1769 pmd_val(*pmd) = pte_val;
1770
1771 pte_val += PMD_SIZE;
1772 vstart += PMD_SIZE;
1773 pmd++;
1774 }
1775
1776 return vstart;
1777}
1778
1779static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1780 bool guard)
1781{
1782 if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1783 return true;
1784
1785 return false;
1786}
1787
1788static unsigned long __ref kernel_map_range(unsigned long pstart,
1789 unsigned long pend, pgprot_t prot,
1790 bool use_huge)
1791{
1792 unsigned long vstart = PAGE_OFFSET + pstart;
1793 unsigned long vend = PAGE_OFFSET + pend;
1794 unsigned long alloc_bytes = 0UL;
1795
1796 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1797 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1798 vstart, vend);
1799 prom_halt();
1800 }
1801
1802 while (vstart < vend) {
1803 unsigned long this_end, paddr = __pa(vstart);
1804 pgd_t *pgd = pgd_offset_k(vstart);
1805 pud_t *pud;
1806 pmd_t *pmd;
1807 pte_t *pte;
1808
1809 if (pgd_none(*pgd)) {
1810 pud_t *new;
1811
1812 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1813 alloc_bytes += PAGE_SIZE;
1814 pgd_populate(&init_mm, pgd, new);
1815 }
1816 pud = pud_offset(pgd, vstart);
1817 if (pud_none(*pud)) {
1818 pmd_t *new;
1819
1820 if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1821 vstart = kernel_map_hugepud(vstart, vend, pud);
1822 continue;
1823 }
1824 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1825 alloc_bytes += PAGE_SIZE;
1826 pud_populate(&init_mm, pud, new);
1827 }
1828
1829 pmd = pmd_offset(pud, vstart);
1830 if (pmd_none(*pmd)) {
1831 pte_t *new;
1832
1833 if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1834 vstart = kernel_map_hugepmd(vstart, vend, pmd);
1835 continue;
1836 }
1837 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1838 alloc_bytes += PAGE_SIZE;
1839 pmd_populate_kernel(&init_mm, pmd, new);
1840 }
1841
1842 pte = pte_offset_kernel(pmd, vstart);
1843 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1844 if (this_end > vend)
1845 this_end = vend;
1846
1847 while (vstart < this_end) {
1848 pte_val(*pte) = (paddr | pgprot_val(prot));
1849
1850 vstart += PAGE_SIZE;
1851 paddr += PAGE_SIZE;
1852 pte++;
1853 }
1854 }
1855
1856 return alloc_bytes;
1857}
1858
1859static void __init flush_all_kernel_tsbs(void)
1860{
1861 int i;
1862
1863 for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1864 struct tsb *ent = &swapper_tsb[i];
1865
1866 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1867 }
1868#ifndef CONFIG_DEBUG_PAGEALLOC
1869 for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1870 struct tsb *ent = &swapper_4m_tsb[i];
1871
1872 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1873 }
1874#endif
1875}
1876
1877extern unsigned int kvmap_linear_patch[1];
1878
1879static void __init kernel_physical_mapping_init(void)
1880{
1881 unsigned long i, mem_alloced = 0UL;
1882 bool use_huge = true;
1883
1884#ifdef CONFIG_DEBUG_PAGEALLOC
1885 use_huge = false;
1886#endif
1887 for (i = 0; i < pall_ents; i++) {
1888 unsigned long phys_start, phys_end;
1889
1890 phys_start = pall[i].phys_addr;
1891 phys_end = phys_start + pall[i].reg_size;
1892
1893 mem_alloced += kernel_map_range(phys_start, phys_end,
1894 PAGE_KERNEL, use_huge);
1895 }
1896
1897 printk("Allocated %ld bytes for kernel page tables.\n",
1898 mem_alloced);
1899
1900 kvmap_linear_patch[0] = 0x01000000; /* nop */
1901 flushi(&kvmap_linear_patch[0]);
1902
1903 flush_all_kernel_tsbs();
1904
1905 __flush_tlb_all();
1906}
1907
1908#ifdef CONFIG_DEBUG_PAGEALLOC
1909void __kernel_map_pages(struct page *page, int numpages, int enable)
1910{
1911 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1912 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1913
1914 kernel_map_range(phys_start, phys_end,
1915 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1916
1917 flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1918 PAGE_OFFSET + phys_end);
1919
1920 /* we should perform an IPI and flush all tlbs,
1921 * but that can deadlock->flush only current cpu.
1922 */
1923 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1924 PAGE_OFFSET + phys_end);
1925}
1926#endif
1927
1928unsigned long __init find_ecache_flush_span(unsigned long size)
1929{
1930 int i;
1931
1932 for (i = 0; i < pavail_ents; i++) {
1933 if (pavail[i].reg_size >= size)
1934 return pavail[i].phys_addr;
1935 }
1936
1937 return ~0UL;
1938}
1939
1940unsigned long PAGE_OFFSET;
1941EXPORT_SYMBOL(PAGE_OFFSET);
1942
1943unsigned long VMALLOC_END = 0x0000010000000000UL;
1944EXPORT_SYMBOL(VMALLOC_END);
1945
1946unsigned long sparc64_va_hole_top = 0xfffff80000000000UL;
1947unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1948
1949static void __init setup_page_offset(void)
1950{
1951 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1952 /* Cheetah/Panther support a full 64-bit virtual
1953 * address, so we can use all that our page tables
1954 * support.
1955 */
1956 sparc64_va_hole_top = 0xfff0000000000000UL;
1957 sparc64_va_hole_bottom = 0x0010000000000000UL;
1958
1959 max_phys_bits = 42;
1960 } else if (tlb_type == hypervisor) {
1961 switch (sun4v_chip_type) {
1962 case SUN4V_CHIP_NIAGARA1:
1963 case SUN4V_CHIP_NIAGARA2:
1964 /* T1 and T2 support 48-bit virtual addresses. */
1965 sparc64_va_hole_top = 0xffff800000000000UL;
1966 sparc64_va_hole_bottom = 0x0000800000000000UL;
1967
1968 max_phys_bits = 39;
1969 break;
1970 case SUN4V_CHIP_NIAGARA3:
1971 /* T3 supports 48-bit virtual addresses. */
1972 sparc64_va_hole_top = 0xffff800000000000UL;
1973 sparc64_va_hole_bottom = 0x0000800000000000UL;
1974
1975 max_phys_bits = 43;
1976 break;
1977 case SUN4V_CHIP_NIAGARA4:
1978 case SUN4V_CHIP_NIAGARA5:
1979 case SUN4V_CHIP_SPARC64X:
1980 case SUN4V_CHIP_SPARC_M6:
1981 /* T4 and later support 52-bit virtual addresses. */
1982 sparc64_va_hole_top = 0xfff8000000000000UL;
1983 sparc64_va_hole_bottom = 0x0008000000000000UL;
1984 max_phys_bits = 47;
1985 break;
1986 case SUN4V_CHIP_SPARC_M7:
1987 case SUN4V_CHIP_SPARC_SN:
1988 /* M7 and later support 52-bit virtual addresses. */
1989 sparc64_va_hole_top = 0xfff8000000000000UL;
1990 sparc64_va_hole_bottom = 0x0008000000000000UL;
1991 max_phys_bits = 49;
1992 break;
1993 case SUN4V_CHIP_SPARC_M8:
1994 default:
1995 /* M8 and later support 54-bit virtual addresses.
1996 * However, restricting M8 and above VA bits to 53
1997 * as 4-level page table cannot support more than
1998 * 53 VA bits.
1999 */
2000 sparc64_va_hole_top = 0xfff0000000000000UL;
2001 sparc64_va_hole_bottom = 0x0010000000000000UL;
2002 max_phys_bits = 51;
2003 break;
2004 }
2005 }
2006
2007 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
2008 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
2009 max_phys_bits);
2010 prom_halt();
2011 }
2012
2013 PAGE_OFFSET = sparc64_va_hole_top;
2014 VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
2015 (sparc64_va_hole_bottom >> 2));
2016
2017 pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
2018 PAGE_OFFSET, max_phys_bits);
2019 pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
2020 VMALLOC_START, VMALLOC_END);
2021 pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
2022 VMEMMAP_BASE, VMEMMAP_BASE << 1);
2023}
2024
2025static void __init tsb_phys_patch(void)
2026{
2027 struct tsb_ldquad_phys_patch_entry *pquad;
2028 struct tsb_phys_patch_entry *p;
2029
2030 pquad = &__tsb_ldquad_phys_patch;
2031 while (pquad < &__tsb_ldquad_phys_patch_end) {
2032 unsigned long addr = pquad->addr;
2033
2034 if (tlb_type == hypervisor)
2035 *(unsigned int *) addr = pquad->sun4v_insn;
2036 else
2037 *(unsigned int *) addr = pquad->sun4u_insn;
2038 wmb();
2039 __asm__ __volatile__("flush %0"
2040 : /* no outputs */
2041 : "r" (addr));
2042
2043 pquad++;
2044 }
2045
2046 p = &__tsb_phys_patch;
2047 while (p < &__tsb_phys_patch_end) {
2048 unsigned long addr = p->addr;
2049
2050 *(unsigned int *) addr = p->insn;
2051 wmb();
2052 __asm__ __volatile__("flush %0"
2053 : /* no outputs */
2054 : "r" (addr));
2055
2056 p++;
2057 }
2058}
2059
2060/* Don't mark as init, we give this to the Hypervisor. */
2061#ifndef CONFIG_DEBUG_PAGEALLOC
2062#define NUM_KTSB_DESCR 2
2063#else
2064#define NUM_KTSB_DESCR 1
2065#endif
2066static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
2067
2068/* The swapper TSBs are loaded with a base sequence of:
2069 *
2070 * sethi %uhi(SYMBOL), REG1
2071 * sethi %hi(SYMBOL), REG2
2072 * or REG1, %ulo(SYMBOL), REG1
2073 * or REG2, %lo(SYMBOL), REG2
2074 * sllx REG1, 32, REG1
2075 * or REG1, REG2, REG1
2076 *
2077 * When we use physical addressing for the TSB accesses, we patch the
2078 * first four instructions in the above sequence.
2079 */
2080
2081static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
2082{
2083 unsigned long high_bits, low_bits;
2084
2085 high_bits = (pa >> 32) & 0xffffffff;
2086 low_bits = (pa >> 0) & 0xffffffff;
2087
2088 while (start < end) {
2089 unsigned int *ia = (unsigned int *)(unsigned long)*start;
2090
2091 ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
2092 __asm__ __volatile__("flush %0" : : "r" (ia));
2093
2094 ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
2095 __asm__ __volatile__("flush %0" : : "r" (ia + 1));
2096
2097 ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
2098 __asm__ __volatile__("flush %0" : : "r" (ia + 2));
2099
2100 ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
2101 __asm__ __volatile__("flush %0" : : "r" (ia + 3));
2102
2103 start++;
2104 }
2105}
2106
2107static void ktsb_phys_patch(void)
2108{
2109 extern unsigned int __swapper_tsb_phys_patch;
2110 extern unsigned int __swapper_tsb_phys_patch_end;
2111 unsigned long ktsb_pa;
2112
2113 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2114 patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
2115 &__swapper_tsb_phys_patch_end, ktsb_pa);
2116#ifndef CONFIG_DEBUG_PAGEALLOC
2117 {
2118 extern unsigned int __swapper_4m_tsb_phys_patch;
2119 extern unsigned int __swapper_4m_tsb_phys_patch_end;
2120 ktsb_pa = (kern_base +
2121 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2122 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
2123 &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
2124 }
2125#endif
2126}
2127
2128static void __init sun4v_ktsb_init(void)
2129{
2130 unsigned long ktsb_pa;
2131
2132 /* First KTSB for PAGE_SIZE mappings. */
2133 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2134
2135 switch (PAGE_SIZE) {
2136 case 8 * 1024:
2137 default:
2138 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
2139 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
2140 break;
2141
2142 case 64 * 1024:
2143 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
2144 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
2145 break;
2146
2147 case 512 * 1024:
2148 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
2149 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
2150 break;
2151
2152 case 4 * 1024 * 1024:
2153 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
2154 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
2155 break;
2156 }
2157
2158 ktsb_descr[0].assoc = 1;
2159 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
2160 ktsb_descr[0].ctx_idx = 0;
2161 ktsb_descr[0].tsb_base = ktsb_pa;
2162 ktsb_descr[0].resv = 0;
2163
2164#ifndef CONFIG_DEBUG_PAGEALLOC
2165 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */
2166 ktsb_pa = (kern_base +
2167 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2168
2169 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
2170 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
2171 HV_PGSZ_MASK_256MB |
2172 HV_PGSZ_MASK_2GB |
2173 HV_PGSZ_MASK_16GB) &
2174 cpu_pgsz_mask);
2175 ktsb_descr[1].assoc = 1;
2176 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
2177 ktsb_descr[1].ctx_idx = 0;
2178 ktsb_descr[1].tsb_base = ktsb_pa;
2179 ktsb_descr[1].resv = 0;
2180#endif
2181}
2182
2183void sun4v_ktsb_register(void)
2184{
2185 unsigned long pa, ret;
2186
2187 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
2188
2189 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
2190 if (ret != 0) {
2191 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
2192 "errors with %lx\n", pa, ret);
2193 prom_halt();
2194 }
2195}
2196
2197static void __init sun4u_linear_pte_xor_finalize(void)
2198{
2199#ifndef CONFIG_DEBUG_PAGEALLOC
2200 /* This is where we would add Panther support for
2201 * 32MB and 256MB pages.
2202 */
2203#endif
2204}
2205
2206static void __init sun4v_linear_pte_xor_finalize(void)
2207{
2208 unsigned long pagecv_flag;
2209
2210 /* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
2211 * enables MCD error. Do not set bit 9 on M7 processor.
2212 */
2213 switch (sun4v_chip_type) {
2214 case SUN4V_CHIP_SPARC_M7:
2215 case SUN4V_CHIP_SPARC_M8:
2216 case SUN4V_CHIP_SPARC_SN:
2217 pagecv_flag = 0x00;
2218 break;
2219 default:
2220 pagecv_flag = _PAGE_CV_4V;
2221 break;
2222 }
2223#ifndef CONFIG_DEBUG_PAGEALLOC
2224 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
2225 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2226 PAGE_OFFSET;
2227 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2228 _PAGE_P_4V | _PAGE_W_4V);
2229 } else {
2230 kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2231 }
2232
2233 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2234 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2235 PAGE_OFFSET;
2236 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2237 _PAGE_P_4V | _PAGE_W_4V);
2238 } else {
2239 kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2240 }
2241
2242 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2243 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2244 PAGE_OFFSET;
2245 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2246 _PAGE_P_4V | _PAGE_W_4V);
2247 } else {
2248 kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2249 }
2250#endif
2251}
2252
2253/* paging_init() sets up the page tables */
2254
2255static unsigned long last_valid_pfn;
2256
2257static void sun4u_pgprot_init(void);
2258static void sun4v_pgprot_init(void);
2259
2260static phys_addr_t __init available_memory(void)
2261{
2262 phys_addr_t available = 0ULL;
2263 phys_addr_t pa_start, pa_end;
2264 u64 i;
2265
2266 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2267 &pa_end, NULL)
2268 available = available + (pa_end - pa_start);
2269
2270 return available;
2271}
2272
2273#define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U)
2274#define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V)
2275#define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2276#define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2277#define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2278#define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2279
2280/* We need to exclude reserved regions. This exclusion will include
2281 * vmlinux and initrd. To be more precise the initrd size could be used to
2282 * compute a new lower limit because it is freed later during initialization.
2283 */
2284static void __init reduce_memory(phys_addr_t limit_ram)
2285{
2286 phys_addr_t avail_ram = available_memory();
2287 phys_addr_t pa_start, pa_end;
2288 u64 i;
2289
2290 if (limit_ram >= avail_ram)
2291 return;
2292
2293 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2294 &pa_end, NULL) {
2295 phys_addr_t region_size = pa_end - pa_start;
2296 phys_addr_t clip_start = pa_start;
2297
2298 avail_ram = avail_ram - region_size;
2299 /* Are we consuming too much? */
2300 if (avail_ram < limit_ram) {
2301 phys_addr_t give_back = limit_ram - avail_ram;
2302
2303 region_size = region_size - give_back;
2304 clip_start = clip_start + give_back;
2305 }
2306
2307 memblock_remove(clip_start, region_size);
2308
2309 if (avail_ram <= limit_ram)
2310 break;
2311 i = 0UL;
2312 }
2313}
2314
2315void __init paging_init(void)
2316{
2317 unsigned long end_pfn, shift, phys_base;
2318 unsigned long real_end, i;
2319
2320 setup_page_offset();
2321
2322 /* These build time checkes make sure that the dcache_dirty_cpu()
2323 * page->flags usage will work.
2324 *
2325 * When a page gets marked as dcache-dirty, we store the
2326 * cpu number starting at bit 32 in the page->flags. Also,
2327 * functions like clear_dcache_dirty_cpu use the cpu mask
2328 * in 13-bit signed-immediate instruction fields.
2329 */
2330
2331 /*
2332 * Page flags must not reach into upper 32 bits that are used
2333 * for the cpu number
2334 */
2335 BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2336
2337 /*
2338 * The bit fields placed in the high range must not reach below
2339 * the 32 bit boundary. Otherwise we cannot place the cpu field
2340 * at the 32 bit boundary.
2341 */
2342 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2343 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2344
2345 BUILD_BUG_ON(NR_CPUS > 4096);
2346
2347 kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2348 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2349
2350 /* Invalidate both kernel TSBs. */
2351 memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2352#ifndef CONFIG_DEBUG_PAGEALLOC
2353 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2354#endif
2355
2356 /* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2357 * bit on M7 processor. This is a conflicting usage of the same
2358 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2359 * Detection error on all pages and this will lead to problems
2360 * later. Kernel does not run with MCD enabled and hence rest
2361 * of the required steps to fully configure memory corruption
2362 * detection are not taken. We need to ensure TTE.mcde is not
2363 * set on M7 processor. Compute the value of cacheability
2364 * flag for use later taking this into consideration.
2365 */
2366 switch (sun4v_chip_type) {
2367 case SUN4V_CHIP_SPARC_M7:
2368 case SUN4V_CHIP_SPARC_M8:
2369 case SUN4V_CHIP_SPARC_SN:
2370 page_cache4v_flag = _PAGE_CP_4V;
2371 break;
2372 default:
2373 page_cache4v_flag = _PAGE_CACHE_4V;
2374 break;
2375 }
2376
2377 if (tlb_type == hypervisor)
2378 sun4v_pgprot_init();
2379 else
2380 sun4u_pgprot_init();
2381
2382 if (tlb_type == cheetah_plus ||
2383 tlb_type == hypervisor) {
2384 tsb_phys_patch();
2385 ktsb_phys_patch();
2386 }
2387
2388 if (tlb_type == hypervisor)
2389 sun4v_patch_tlb_handlers();
2390
2391 /* Find available physical memory...
2392 *
2393 * Read it twice in order to work around a bug in openfirmware.
2394 * The call to grab this table itself can cause openfirmware to
2395 * allocate memory, which in turn can take away some space from
2396 * the list of available memory. Reading it twice makes sure
2397 * we really do get the final value.
2398 */
2399 read_obp_translations();
2400 read_obp_memory("reg", &pall[0], &pall_ents);
2401 read_obp_memory("available", &pavail[0], &pavail_ents);
2402 read_obp_memory("available", &pavail[0], &pavail_ents);
2403
2404 phys_base = 0xffffffffffffffffUL;
2405 for (i = 0; i < pavail_ents; i++) {
2406 phys_base = min(phys_base, pavail[i].phys_addr);
2407 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2408 }
2409
2410 memblock_reserve(kern_base, kern_size);
2411
2412 find_ramdisk(phys_base);
2413
2414 if (cmdline_memory_size)
2415 reduce_memory(cmdline_memory_size);
2416
2417 memblock_allow_resize();
2418 memblock_dump_all();
2419
2420 set_bit(0, mmu_context_bmap);
2421
2422 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2423
2424 real_end = (unsigned long)_end;
2425 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2426 printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2427 num_kernel_image_mappings);
2428
2429 /* Set kernel pgd to upper alias so physical page computations
2430 * work.
2431 */
2432 init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2433
2434 memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2435
2436 inherit_prom_mappings();
2437
2438 /* Ok, we can use our TLB miss and window trap handlers safely. */
2439 setup_tba();
2440
2441 __flush_tlb_all();
2442
2443 prom_build_devicetree();
2444 of_populate_present_mask();
2445#ifndef CONFIG_SMP
2446 of_fill_in_cpu_data();
2447#endif
2448
2449 if (tlb_type == hypervisor) {
2450 sun4v_mdesc_init();
2451 mdesc_populate_present_mask(cpu_all_mask);
2452#ifndef CONFIG_SMP
2453 mdesc_fill_in_cpu_data(cpu_all_mask);
2454#endif
2455 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2456
2457 sun4v_linear_pte_xor_finalize();
2458
2459 sun4v_ktsb_init();
2460 sun4v_ktsb_register();
2461 } else {
2462 unsigned long impl, ver;
2463
2464 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2465 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2466
2467 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2468 impl = ((ver >> 32) & 0xffff);
2469 if (impl == PANTHER_IMPL)
2470 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2471 HV_PGSZ_MASK_256MB);
2472
2473 sun4u_linear_pte_xor_finalize();
2474 }
2475
2476 /* Flush the TLBs and the 4M TSB so that the updated linear
2477 * pte XOR settings are realized for all mappings.
2478 */
2479 __flush_tlb_all();
2480#ifndef CONFIG_DEBUG_PAGEALLOC
2481 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2482#endif
2483 __flush_tlb_all();
2484
2485 /* Setup bootmem... */
2486 last_valid_pfn = end_pfn = bootmem_init(phys_base);
2487
2488 kernel_physical_mapping_init();
2489
2490 {
2491 unsigned long max_zone_pfns[MAX_NR_ZONES];
2492
2493 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2494
2495 max_zone_pfns[ZONE_NORMAL] = end_pfn;
2496
2497 free_area_init_nodes(max_zone_pfns);
2498 }
2499
2500 printk("Booting Linux...\n");
2501}
2502
2503int page_in_phys_avail(unsigned long paddr)
2504{
2505 int i;
2506
2507 paddr &= PAGE_MASK;
2508
2509 for (i = 0; i < pavail_ents; i++) {
2510 unsigned long start, end;
2511
2512 start = pavail[i].phys_addr;
2513 end = start + pavail[i].reg_size;
2514
2515 if (paddr >= start && paddr < end)
2516 return 1;
2517 }
2518 if (paddr >= kern_base && paddr < (kern_base + kern_size))
2519 return 1;
2520#ifdef CONFIG_BLK_DEV_INITRD
2521 if (paddr >= __pa(initrd_start) &&
2522 paddr < __pa(PAGE_ALIGN(initrd_end)))
2523 return 1;
2524#endif
2525
2526 return 0;
2527}
2528
2529static void __init register_page_bootmem_info(void)
2530{
2531#ifdef CONFIG_NEED_MULTIPLE_NODES
2532 int i;
2533
2534 for_each_online_node(i)
2535 if (NODE_DATA(i)->node_spanned_pages)
2536 register_page_bootmem_info_node(NODE_DATA(i));
2537#endif
2538}
2539void __init mem_init(void)
2540{
2541 high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2542
2543 free_all_bootmem();
2544
2545 /*
2546 * Must be done after boot memory is put on freelist, because here we
2547 * might set fields in deferred struct pages that have not yet been
2548 * initialized, and free_all_bootmem() initializes all the reserved
2549 * deferred pages for us.
2550 */
2551 register_page_bootmem_info();
2552
2553 /*
2554 * Set up the zero page, mark it reserved, so that page count
2555 * is not manipulated when freeing the page from user ptes.
2556 */
2557 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2558 if (mem_map_zero == NULL) {
2559 prom_printf("paging_init: Cannot alloc zero page.\n");
2560 prom_halt();
2561 }
2562 mark_page_reserved(mem_map_zero);
2563
2564 mem_init_print_info(NULL);
2565
2566 if (tlb_type == cheetah || tlb_type == cheetah_plus)
2567 cheetah_ecache_flush_init();
2568}
2569
2570void free_initmem(void)
2571{
2572 unsigned long addr, initend;
2573 int do_free = 1;
2574
2575 /* If the physical memory maps were trimmed by kernel command
2576 * line options, don't even try freeing this initmem stuff up.
2577 * The kernel image could have been in the trimmed out region
2578 * and if so the freeing below will free invalid page structs.
2579 */
2580 if (cmdline_memory_size)
2581 do_free = 0;
2582
2583 /*
2584 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2585 */
2586 addr = PAGE_ALIGN((unsigned long)(__init_begin));
2587 initend = (unsigned long)(__init_end) & PAGE_MASK;
2588 for (; addr < initend; addr += PAGE_SIZE) {
2589 unsigned long page;
2590
2591 page = (addr +
2592 ((unsigned long) __va(kern_base)) -
2593 ((unsigned long) KERNBASE));
2594 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2595
2596 if (do_free)
2597 free_reserved_page(virt_to_page(page));
2598 }
2599}
2600
2601#ifdef CONFIG_BLK_DEV_INITRD
2602void free_initrd_mem(unsigned long start, unsigned long end)
2603{
2604 free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2605 "initrd");
2606}
2607#endif
2608
2609pgprot_t PAGE_KERNEL __read_mostly;
2610EXPORT_SYMBOL(PAGE_KERNEL);
2611
2612pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2613pgprot_t PAGE_COPY __read_mostly;
2614
2615pgprot_t PAGE_SHARED __read_mostly;
2616EXPORT_SYMBOL(PAGE_SHARED);
2617
2618unsigned long pg_iobits __read_mostly;
2619
2620unsigned long _PAGE_IE __read_mostly;
2621EXPORT_SYMBOL(_PAGE_IE);
2622
2623unsigned long _PAGE_E __read_mostly;
2624EXPORT_SYMBOL(_PAGE_E);
2625
2626unsigned long _PAGE_CACHE __read_mostly;
2627EXPORT_SYMBOL(_PAGE_CACHE);
2628
2629#ifdef CONFIG_SPARSEMEM_VMEMMAP
2630int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2631 int node, struct vmem_altmap *altmap)
2632{
2633 unsigned long pte_base;
2634
2635 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2636 _PAGE_CP_4U | _PAGE_CV_4U |
2637 _PAGE_P_4U | _PAGE_W_4U);
2638 if (tlb_type == hypervisor)
2639 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2640 page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2641
2642 pte_base |= _PAGE_PMD_HUGE;
2643
2644 vstart = vstart & PMD_MASK;
2645 vend = ALIGN(vend, PMD_SIZE);
2646 for (; vstart < vend; vstart += PMD_SIZE) {
2647 pgd_t *pgd = vmemmap_pgd_populate(vstart, node);
2648 unsigned long pte;
2649 pud_t *pud;
2650 pmd_t *pmd;
2651
2652 if (!pgd)
2653 return -ENOMEM;
2654
2655 pud = vmemmap_pud_populate(pgd, vstart, node);
2656 if (!pud)
2657 return -ENOMEM;
2658
2659 pmd = pmd_offset(pud, vstart);
2660 pte = pmd_val(*pmd);
2661 if (!(pte & _PAGE_VALID)) {
2662 void *block = vmemmap_alloc_block(PMD_SIZE, node);
2663
2664 if (!block)
2665 return -ENOMEM;
2666
2667 pmd_val(*pmd) = pte_base | __pa(block);
2668 }
2669 }
2670
2671 return 0;
2672}
2673
2674void vmemmap_free(unsigned long start, unsigned long end,
2675 struct vmem_altmap *altmap)
2676{
2677}
2678#endif /* CONFIG_SPARSEMEM_VMEMMAP */
2679
2680static void prot_init_common(unsigned long page_none,
2681 unsigned long page_shared,
2682 unsigned long page_copy,
2683 unsigned long page_readonly,
2684 unsigned long page_exec_bit)
2685{
2686 PAGE_COPY = __pgprot(page_copy);
2687 PAGE_SHARED = __pgprot(page_shared);
2688
2689 protection_map[0x0] = __pgprot(page_none);
2690 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2691 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2692 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2693 protection_map[0x4] = __pgprot(page_readonly);
2694 protection_map[0x5] = __pgprot(page_readonly);
2695 protection_map[0x6] = __pgprot(page_copy);
2696 protection_map[0x7] = __pgprot(page_copy);
2697 protection_map[0x8] = __pgprot(page_none);
2698 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2699 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2700 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2701 protection_map[0xc] = __pgprot(page_readonly);
2702 protection_map[0xd] = __pgprot(page_readonly);
2703 protection_map[0xe] = __pgprot(page_shared);
2704 protection_map[0xf] = __pgprot(page_shared);
2705}
2706
2707static void __init sun4u_pgprot_init(void)
2708{
2709 unsigned long page_none, page_shared, page_copy, page_readonly;
2710 unsigned long page_exec_bit;
2711 int i;
2712
2713 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2714 _PAGE_CACHE_4U | _PAGE_P_4U |
2715 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2716 _PAGE_EXEC_4U);
2717 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2718 _PAGE_CACHE_4U | _PAGE_P_4U |
2719 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2720 _PAGE_EXEC_4U | _PAGE_L_4U);
2721
2722 _PAGE_IE = _PAGE_IE_4U;
2723 _PAGE_E = _PAGE_E_4U;
2724 _PAGE_CACHE = _PAGE_CACHE_4U;
2725
2726 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2727 __ACCESS_BITS_4U | _PAGE_E_4U);
2728
2729#ifdef CONFIG_DEBUG_PAGEALLOC
2730 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2731#else
2732 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2733 PAGE_OFFSET;
2734#endif
2735 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2736 _PAGE_P_4U | _PAGE_W_4U);
2737
2738 for (i = 1; i < 4; i++)
2739 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2740
2741 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2742 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2743 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2744
2745
2746 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2747 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2748 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2749 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2750 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2751 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2752 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2753
2754 page_exec_bit = _PAGE_EXEC_4U;
2755
2756 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2757 page_exec_bit);
2758}
2759
2760static void __init sun4v_pgprot_init(void)
2761{
2762 unsigned long page_none, page_shared, page_copy, page_readonly;
2763 unsigned long page_exec_bit;
2764 int i;
2765
2766 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2767 page_cache4v_flag | _PAGE_P_4V |
2768 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2769 _PAGE_EXEC_4V);
2770 PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2771
2772 _PAGE_IE = _PAGE_IE_4V;
2773 _PAGE_E = _PAGE_E_4V;
2774 _PAGE_CACHE = page_cache4v_flag;
2775
2776#ifdef CONFIG_DEBUG_PAGEALLOC
2777 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2778#else
2779 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2780 PAGE_OFFSET;
2781#endif
2782 kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2783 _PAGE_W_4V);
2784
2785 for (i = 1; i < 4; i++)
2786 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2787
2788 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2789 __ACCESS_BITS_4V | _PAGE_E_4V);
2790
2791 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2792 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2793 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2794 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2795
2796 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2797 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2798 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2799 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2800 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2801 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2802 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2803
2804 page_exec_bit = _PAGE_EXEC_4V;
2805
2806 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2807 page_exec_bit);
2808}
2809
2810unsigned long pte_sz_bits(unsigned long sz)
2811{
2812 if (tlb_type == hypervisor) {
2813 switch (sz) {
2814 case 8 * 1024:
2815 default:
2816 return _PAGE_SZ8K_4V;
2817 case 64 * 1024:
2818 return _PAGE_SZ64K_4V;
2819 case 512 * 1024:
2820 return _PAGE_SZ512K_4V;
2821 case 4 * 1024 * 1024:
2822 return _PAGE_SZ4MB_4V;
2823 }
2824 } else {
2825 switch (sz) {
2826 case 8 * 1024:
2827 default:
2828 return _PAGE_SZ8K_4U;
2829 case 64 * 1024:
2830 return _PAGE_SZ64K_4U;
2831 case 512 * 1024:
2832 return _PAGE_SZ512K_4U;
2833 case 4 * 1024 * 1024:
2834 return _PAGE_SZ4MB_4U;
2835 }
2836 }
2837}
2838
2839pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2840{
2841 pte_t pte;
2842
2843 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot));
2844 pte_val(pte) |= (((unsigned long)space) << 32);
2845 pte_val(pte) |= pte_sz_bits(page_size);
2846
2847 return pte;
2848}
2849
2850static unsigned long kern_large_tte(unsigned long paddr)
2851{
2852 unsigned long val;
2853
2854 val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2855 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2856 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2857 if (tlb_type == hypervisor)
2858 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2859 page_cache4v_flag | _PAGE_P_4V |
2860 _PAGE_EXEC_4V | _PAGE_W_4V);
2861
2862 return val | paddr;
2863}
2864
2865/* If not locked, zap it. */
2866void __flush_tlb_all(void)
2867{
2868 unsigned long pstate;
2869 int i;
2870
2871 __asm__ __volatile__("flushw\n\t"
2872 "rdpr %%pstate, %0\n\t"
2873 "wrpr %0, %1, %%pstate"
2874 : "=r" (pstate)
2875 : "i" (PSTATE_IE));
2876 if (tlb_type == hypervisor) {
2877 sun4v_mmu_demap_all();
2878 } else if (tlb_type == spitfire) {
2879 for (i = 0; i < 64; i++) {
2880 /* Spitfire Errata #32 workaround */
2881 /* NOTE: Always runs on spitfire, so no
2882 * cheetah+ page size encodings.
2883 */
2884 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2885 "flush %%g6"
2886 : /* No outputs */
2887 : "r" (0),
2888 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2889
2890 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2891 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2892 "membar #Sync"
2893 : /* no outputs */
2894 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2895 spitfire_put_dtlb_data(i, 0x0UL);
2896 }
2897
2898 /* Spitfire Errata #32 workaround */
2899 /* NOTE: Always runs on spitfire, so no
2900 * cheetah+ page size encodings.
2901 */
2902 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2903 "flush %%g6"
2904 : /* No outputs */
2905 : "r" (0),
2906 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2907
2908 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2909 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2910 "membar #Sync"
2911 : /* no outputs */
2912 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2913 spitfire_put_itlb_data(i, 0x0UL);
2914 }
2915 }
2916 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2917 cheetah_flush_dtlb_all();
2918 cheetah_flush_itlb_all();
2919 }
2920 __asm__ __volatile__("wrpr %0, 0, %%pstate"
2921 : : "r" (pstate));
2922}
2923
2924pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2925 unsigned long address)
2926{
2927 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2928 pte_t *pte = NULL;
2929
2930 if (page)
2931 pte = (pte_t *) page_address(page);
2932
2933 return pte;
2934}
2935
2936pgtable_t pte_alloc_one(struct mm_struct *mm,
2937 unsigned long address)
2938{
2939 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2940 if (!page)
2941 return NULL;
2942 if (!pgtable_page_ctor(page)) {
2943 free_unref_page(page);
2944 return NULL;
2945 }
2946 return (pte_t *) page_address(page);
2947}
2948
2949void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2950{
2951 free_page((unsigned long)pte);
2952}
2953
2954static void __pte_free(pgtable_t pte)
2955{
2956 struct page *page = virt_to_page(pte);
2957
2958 pgtable_page_dtor(page);
2959 __free_page(page);
2960}
2961
2962void pte_free(struct mm_struct *mm, pgtable_t pte)
2963{
2964 __pte_free(pte);
2965}
2966
2967void pgtable_free(void *table, bool is_page)
2968{
2969 if (is_page)
2970 __pte_free(table);
2971 else
2972 kmem_cache_free(pgtable_cache, table);
2973}
2974
2975#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2976void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2977 pmd_t *pmd)
2978{
2979 unsigned long pte, flags;
2980 struct mm_struct *mm;
2981 pmd_t entry = *pmd;
2982
2983 if (!pmd_large(entry) || !pmd_young(entry))
2984 return;
2985
2986 pte = pmd_val(entry);
2987
2988 /* Don't insert a non-valid PMD into the TSB, we'll deadlock. */
2989 if (!(pte & _PAGE_VALID))
2990 return;
2991
2992 /* We are fabricating 8MB pages using 4MB real hw pages. */
2993 pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2994
2995 mm = vma->vm_mm;
2996
2997 spin_lock_irqsave(&mm->context.lock, flags);
2998
2999 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
3000 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
3001 addr, pte);
3002
3003 spin_unlock_irqrestore(&mm->context.lock, flags);
3004}
3005#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3006
3007#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
3008static void context_reload(void *__data)
3009{
3010 struct mm_struct *mm = __data;
3011
3012 if (mm == current->mm)
3013 load_secondary_context(mm);
3014}
3015
3016void hugetlb_setup(struct pt_regs *regs)
3017{
3018 struct mm_struct *mm = current->mm;
3019 struct tsb_config *tp;
3020
3021 if (faulthandler_disabled() || !mm) {
3022 const struct exception_table_entry *entry;
3023
3024 entry = search_exception_tables(regs->tpc);
3025 if (entry) {
3026 regs->tpc = entry->fixup;
3027 regs->tnpc = regs->tpc + 4;
3028 return;
3029 }
3030 pr_alert("Unexpected HugeTLB setup in atomic context.\n");
3031 die_if_kernel("HugeTSB in atomic", regs);
3032 }
3033
3034 tp = &mm->context.tsb_block[MM_TSB_HUGE];
3035 if (likely(tp->tsb == NULL))
3036 tsb_grow(mm, MM_TSB_HUGE, 0);
3037
3038 tsb_context_switch(mm);
3039 smp_tsb_sync(mm);
3040
3041 /* On UltraSPARC-III+ and later, configure the second half of
3042 * the Data-TLB for huge pages.
3043 */
3044 if (tlb_type == cheetah_plus) {
3045 bool need_context_reload = false;
3046 unsigned long ctx;
3047
3048 spin_lock_irq(&ctx_alloc_lock);
3049 ctx = mm->context.sparc64_ctx_val;
3050 ctx &= ~CTX_PGSZ_MASK;
3051 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
3052 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
3053
3054 if (ctx != mm->context.sparc64_ctx_val) {
3055 /* When changing the page size fields, we
3056 * must perform a context flush so that no
3057 * stale entries match. This flush must
3058 * occur with the original context register
3059 * settings.
3060 */
3061 do_flush_tlb_mm(mm);
3062
3063 /* Reload the context register of all processors
3064 * also executing in this address space.
3065 */
3066 mm->context.sparc64_ctx_val = ctx;
3067 need_context_reload = true;
3068 }
3069 spin_unlock_irq(&ctx_alloc_lock);
3070
3071 if (need_context_reload)
3072 on_each_cpu(context_reload, mm, 0);
3073 }
3074}
3075#endif
3076
3077static struct resource code_resource = {
3078 .name = "Kernel code",
3079 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3080};
3081
3082static struct resource data_resource = {
3083 .name = "Kernel data",
3084 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3085};
3086
3087static struct resource bss_resource = {
3088 .name = "Kernel bss",
3089 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3090};
3091
3092static inline resource_size_t compute_kern_paddr(void *addr)
3093{
3094 return (resource_size_t) (addr - KERNBASE + kern_base);
3095}
3096
3097static void __init kernel_lds_init(void)
3098{
3099 code_resource.start = compute_kern_paddr(_text);
3100 code_resource.end = compute_kern_paddr(_etext - 1);
3101 data_resource.start = compute_kern_paddr(_etext);
3102 data_resource.end = compute_kern_paddr(_edata - 1);
3103 bss_resource.start = compute_kern_paddr(__bss_start);
3104 bss_resource.end = compute_kern_paddr(_end - 1);
3105}
3106
3107static int __init report_memory(void)
3108{
3109 int i;
3110 struct resource *res;
3111
3112 kernel_lds_init();
3113
3114 for (i = 0; i < pavail_ents; i++) {
3115 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
3116
3117 if (!res) {
3118 pr_warn("Failed to allocate source.\n");
3119 break;
3120 }
3121
3122 res->name = "System RAM";
3123 res->start = pavail[i].phys_addr;
3124 res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
3125 res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
3126
3127 if (insert_resource(&iomem_resource, res) < 0) {
3128 pr_warn("Resource insertion failed.\n");
3129 break;
3130 }
3131
3132 insert_resource(res, &code_resource);
3133 insert_resource(res, &data_resource);
3134 insert_resource(res, &bss_resource);
3135 }
3136
3137 return 0;
3138}
3139arch_initcall(report_memory);
3140
3141#ifdef CONFIG_SMP
3142#define do_flush_tlb_kernel_range smp_flush_tlb_kernel_range
3143#else
3144#define do_flush_tlb_kernel_range __flush_tlb_kernel_range
3145#endif
3146
3147void flush_tlb_kernel_range(unsigned long start, unsigned long end)
3148{
3149 if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
3150 if (start < LOW_OBP_ADDRESS) {
3151 flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
3152 do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
3153 }
3154 if (end > HI_OBP_ADDRESS) {
3155 flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
3156 do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
3157 }
3158 } else {
3159 flush_tsb_kernel_range(start, end);
3160 do_flush_tlb_kernel_range(start, end);
3161 }
3162}
3163
3164void copy_user_highpage(struct page *to, struct page *from,
3165 unsigned long vaddr, struct vm_area_struct *vma)
3166{
3167 char *vfrom, *vto;
3168
3169 vfrom = kmap_atomic(from);
3170 vto = kmap_atomic(to);
3171 copy_user_page(vto, vfrom, vaddr, to);
3172 kunmap_atomic(vto);
3173 kunmap_atomic(vfrom);
3174
3175 /* If this page has ADI enabled, copy over any ADI tags
3176 * as well
3177 */
3178 if (vma->vm_flags & VM_SPARC_ADI) {
3179 unsigned long pfrom, pto, i, adi_tag;
3180
3181 pfrom = page_to_phys(from);
3182 pto = page_to_phys(to);
3183
3184 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3185 asm volatile("ldxa [%1] %2, %0\n\t"
3186 : "=r" (adi_tag)
3187 : "r" (i), "i" (ASI_MCD_REAL));
3188 asm volatile("stxa %0, [%1] %2\n\t"
3189 :
3190 : "r" (adi_tag), "r" (pto),
3191 "i" (ASI_MCD_REAL));
3192 pto += adi_blksize();
3193 }
3194 asm volatile("membar #Sync\n\t");
3195 }
3196}
3197EXPORT_SYMBOL(copy_user_highpage);
3198
3199void copy_highpage(struct page *to, struct page *from)
3200{
3201 char *vfrom, *vto;
3202
3203 vfrom = kmap_atomic(from);
3204 vto = kmap_atomic(to);
3205 copy_page(vto, vfrom);
3206 kunmap_atomic(vto);
3207 kunmap_atomic(vfrom);
3208
3209 /* If this platform is ADI enabled, copy any ADI tags
3210 * as well
3211 */
3212 if (adi_capable()) {
3213 unsigned long pfrom, pto, i, adi_tag;
3214
3215 pfrom = page_to_phys(from);
3216 pto = page_to_phys(to);
3217
3218 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3219 asm volatile("ldxa [%1] %2, %0\n\t"
3220 : "=r" (adi_tag)
3221 : "r" (i), "i" (ASI_MCD_REAL));
3222 asm volatile("stxa %0, [%1] %2\n\t"
3223 :
3224 : "r" (adi_tag), "r" (pto),
3225 "i" (ASI_MCD_REAL));
3226 pto += adi_blksize();
3227 }
3228 asm volatile("membar #Sync\n\t");
3229 }
3230}
3231EXPORT_SYMBOL(copy_highpage);
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