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// SPDX-License-Identifier: GPL-2.0
2/*
3 * arch/sparc64/mm/init.c
4 *
5 * Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
6 * Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
7 */
8
9#include <linux/extable.h>
10#include <linux/kernel.h>
11#include <linux/sched.h>
12#include <linux/string.h>
13#include <linux/init.h>
14#include <linux/memblock.h>
15#include <linux/mm.h>
16#include <linux/hugetlb.h>
17#include <linux/initrd.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/poison.h>
21#include <linux/fs.h>
22#include <linux/seq_file.h>
23#include <linux/kprobes.h>
24#include <linux/cache.h>
25#include <linux/sort.h>
26#include <linux/ioport.h>
27#include <linux/percpu.h>
28#include <linux/mmzone.h>
29#include <linux/gfp.h>
30#include <linux/bootmem_info.h>
31
32#include <asm/head.h>
33#include <asm/page.h>
34#include <asm/pgalloc.h>
35#include <asm/oplib.h>
36#include <asm/iommu.h>
37#include <asm/io.h>
38#include <linux/uaccess.h>
39#include <asm/mmu_context.h>
40#include <asm/tlbflush.h>
41#include <asm/dma.h>
42#include <asm/starfire.h>
43#include <asm/tlb.h>
44#include <asm/spitfire.h>
45#include <asm/sections.h>
46#include <asm/tsb.h>
47#include <asm/hypervisor.h>
48#include <asm/prom.h>
49#include <asm/mdesc.h>
50#include <asm/cpudata.h>
51#include <asm/setup.h>
52#include <asm/irq.h>
53
54#include "init_64.h"
55
56unsigned long kern_linear_pte_xor[4] __read_mostly;
57static unsigned long page_cache4v_flag;
58
59/* A bitmap, two bits for every 256MB of physical memory. These two
60 * bits determine what page size we use for kernel linear
61 * translations. They form an index into kern_linear_pte_xor[]. The
62 * value in the indexed slot is XOR'd with the TLB miss virtual
63 * address to form the resulting TTE. The mapping is:
64 *
65 * 0 ==> 4MB
66 * 1 ==> 256MB
67 * 2 ==> 2GB
68 * 3 ==> 16GB
69 *
70 * All sun4v chips support 256MB pages. Only SPARC-T4 and later
71 * support 2GB pages, and hopefully future cpus will support the 16GB
72 * pages as well. For slots 2 and 3, we encode a 256MB TTE xor there
73 * if these larger page sizes are not supported by the cpu.
74 *
75 * It would be nice to determine this from the machine description
76 * 'cpu' properties, but we need to have this table setup before the
77 * MDESC is initialized.
78 */
79
80#ifndef CONFIG_DEBUG_PAGEALLOC
81/* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
82 * Space is allocated for this right after the trap table in
83 * arch/sparc64/kernel/head.S
84 */
85extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
86#endif
87extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
88
89static unsigned long cpu_pgsz_mask;
90
91#define MAX_BANKS 1024
92
93static struct linux_prom64_registers pavail[MAX_BANKS];
94static int pavail_ents;
95
96u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
97
98static int cmp_p64(const void *a, const void *b)
99{
100 const struct linux_prom64_registers *x = a, *y = b;
101
102 if (x->phys_addr > y->phys_addr)
103 return 1;
104 if (x->phys_addr < y->phys_addr)
105 return -1;
106 return 0;
107}
108
109static void __init read_obp_memory(const char *property,
110 struct linux_prom64_registers *regs,
111 int *num_ents)
112{
113 phandle node = prom_finddevice("/memory");
114 int prop_size = prom_getproplen(node, property);
115 int ents, ret, i;
116
117 ents = prop_size / sizeof(struct linux_prom64_registers);
118 if (ents > MAX_BANKS) {
119 prom_printf("The machine has more %s property entries than "
120 "this kernel can support (%d).\n",
121 property, MAX_BANKS);
122 prom_halt();
123 }
124
125 ret = prom_getproperty(node, property, (char *) regs, prop_size);
126 if (ret == -1) {
127 prom_printf("Couldn't get %s property from /memory.\n",
128 property);
129 prom_halt();
130 }
131
132 /* Sanitize what we got from the firmware, by page aligning
133 * everything.
134 */
135 for (i = 0; i < ents; i++) {
136 unsigned long base, size;
137
138 base = regs[i].phys_addr;
139 size = regs[i].reg_size;
140
141 size &= PAGE_MASK;
142 if (base & ~PAGE_MASK) {
143 unsigned long new_base = PAGE_ALIGN(base);
144
145 size -= new_base - base;
146 if ((long) size < 0L)
147 size = 0UL;
148 base = new_base;
149 }
150 if (size == 0UL) {
151 /* If it is empty, simply get rid of it.
152 * This simplifies the logic of the other
153 * functions that process these arrays.
154 */
155 memmove(®s[i], ®s[i + 1],
156 (ents - i - 1) * sizeof(regs[0]));
157 i--;
158 ents--;
159 continue;
160 }
161 regs[i].phys_addr = base;
162 regs[i].reg_size = size;
163 }
164
165 *num_ents = ents;
166
167 sort(regs, ents, sizeof(struct linux_prom64_registers),
168 cmp_p64, NULL);
169}
170
171/* Kernel physical address base and size in bytes. */
172unsigned long kern_base __read_mostly;
173unsigned long kern_size __read_mostly;
174
175/* Initial ramdisk setup */
176extern unsigned long sparc_ramdisk_image64;
177extern unsigned int sparc_ramdisk_image;
178extern unsigned int sparc_ramdisk_size;
179
180struct page *mem_map_zero __read_mostly;
181EXPORT_SYMBOL(mem_map_zero);
182
183unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
184
185unsigned long sparc64_kern_pri_context __read_mostly;
186unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
187unsigned long sparc64_kern_sec_context __read_mostly;
188
189int num_kernel_image_mappings;
190
191#ifdef CONFIG_DEBUG_DCFLUSH
192atomic_t dcpage_flushes = ATOMIC_INIT(0);
193#ifdef CONFIG_SMP
194atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
195#endif
196#endif
197
198inline void flush_dcache_folio_impl(struct folio *folio)
199{
200 unsigned int i, nr = folio_nr_pages(folio);
201
202 BUG_ON(tlb_type == hypervisor);
203#ifdef CONFIG_DEBUG_DCFLUSH
204 atomic_inc(&dcpage_flushes);
205#endif
206
207#ifdef DCACHE_ALIASING_POSSIBLE
208 for (i = 0; i < nr; i++)
209 __flush_dcache_page(folio_address(folio) + i * PAGE_SIZE,
210 ((tlb_type == spitfire) &&
211 folio_flush_mapping(folio) != NULL));
212#else
213 if (folio_flush_mapping(folio) != NULL &&
214 tlb_type == spitfire) {
215 for (i = 0; i < nr; i++)
216 __flush_icache_page((pfn + i) * PAGE_SIZE);
217 }
218#endif
219}
220
221#define PG_dcache_dirty PG_arch_1
222#define PG_dcache_cpu_shift 32UL
223#define PG_dcache_cpu_mask \
224 ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
225
226#define dcache_dirty_cpu(folio) \
227 (((folio)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
228
229static inline void set_dcache_dirty(struct folio *folio, int this_cpu)
230{
231 unsigned long mask = this_cpu;
232 unsigned long non_cpu_bits;
233
234 non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
235 mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
236
237 __asm__ __volatile__("1:\n\t"
238 "ldx [%2], %%g7\n\t"
239 "and %%g7, %1, %%g1\n\t"
240 "or %%g1, %0, %%g1\n\t"
241 "casx [%2], %%g7, %%g1\n\t"
242 "cmp %%g7, %%g1\n\t"
243 "bne,pn %%xcc, 1b\n\t"
244 " nop"
245 : /* no outputs */
246 : "r" (mask), "r" (non_cpu_bits), "r" (&folio->flags)
247 : "g1", "g7");
248}
249
250static inline void clear_dcache_dirty_cpu(struct folio *folio, unsigned long cpu)
251{
252 unsigned long mask = (1UL << PG_dcache_dirty);
253
254 __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
255 "1:\n\t"
256 "ldx [%2], %%g7\n\t"
257 "srlx %%g7, %4, %%g1\n\t"
258 "and %%g1, %3, %%g1\n\t"
259 "cmp %%g1, %0\n\t"
260 "bne,pn %%icc, 2f\n\t"
261 " andn %%g7, %1, %%g1\n\t"
262 "casx [%2], %%g7, %%g1\n\t"
263 "cmp %%g7, %%g1\n\t"
264 "bne,pn %%xcc, 1b\n\t"
265 " nop\n"
266 "2:"
267 : /* no outputs */
268 : "r" (cpu), "r" (mask), "r" (&folio->flags),
269 "i" (PG_dcache_cpu_mask),
270 "i" (PG_dcache_cpu_shift)
271 : "g1", "g7");
272}
273
274static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
275{
276 unsigned long tsb_addr = (unsigned long) ent;
277
278 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
279 tsb_addr = __pa(tsb_addr);
280
281 __tsb_insert(tsb_addr, tag, pte);
282}
283
284unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
285
286static void flush_dcache(unsigned long pfn)
287{
288 struct page *page;
289
290 page = pfn_to_page(pfn);
291 if (page) {
292 struct folio *folio = page_folio(page);
293 unsigned long pg_flags;
294
295 pg_flags = folio->flags;
296 if (pg_flags & (1UL << PG_dcache_dirty)) {
297 int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
298 PG_dcache_cpu_mask);
299 int this_cpu = get_cpu();
300
301 /* This is just to optimize away some function calls
302 * in the SMP case.
303 */
304 if (cpu == this_cpu)
305 flush_dcache_folio_impl(folio);
306 else
307 smp_flush_dcache_folio_impl(folio, cpu);
308
309 clear_dcache_dirty_cpu(folio, cpu);
310
311 put_cpu();
312 }
313 }
314}
315
316/* mm->context.lock must be held */
317static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
318 unsigned long tsb_hash_shift, unsigned long address,
319 unsigned long tte)
320{
321 struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
322 unsigned long tag;
323
324 if (unlikely(!tsb))
325 return;
326
327 tsb += ((address >> tsb_hash_shift) &
328 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
329 tag = (address >> 22UL);
330 tsb_insert(tsb, tag, tte);
331}
332
333#ifdef CONFIG_HUGETLB_PAGE
334static int __init hugetlbpage_init(void)
335{
336 hugetlb_add_hstate(HPAGE_64K_SHIFT - PAGE_SHIFT);
337 hugetlb_add_hstate(HPAGE_SHIFT - PAGE_SHIFT);
338 hugetlb_add_hstate(HPAGE_256MB_SHIFT - PAGE_SHIFT);
339 hugetlb_add_hstate(HPAGE_2GB_SHIFT - PAGE_SHIFT);
340
341 return 0;
342}
343
344arch_initcall(hugetlbpage_init);
345
346static void __init pud_huge_patch(void)
347{
348 struct pud_huge_patch_entry *p;
349 unsigned long addr;
350
351 p = &__pud_huge_patch;
352 addr = p->addr;
353 *(unsigned int *)addr = p->insn;
354
355 __asm__ __volatile__("flush %0" : : "r" (addr));
356}
357
358bool __init arch_hugetlb_valid_size(unsigned long size)
359{
360 unsigned int hugepage_shift = ilog2(size);
361 unsigned short hv_pgsz_idx;
362 unsigned int hv_pgsz_mask;
363
364 switch (hugepage_shift) {
365 case HPAGE_16GB_SHIFT:
366 hv_pgsz_mask = HV_PGSZ_MASK_16GB;
367 hv_pgsz_idx = HV_PGSZ_IDX_16GB;
368 pud_huge_patch();
369 break;
370 case HPAGE_2GB_SHIFT:
371 hv_pgsz_mask = HV_PGSZ_MASK_2GB;
372 hv_pgsz_idx = HV_PGSZ_IDX_2GB;
373 break;
374 case HPAGE_256MB_SHIFT:
375 hv_pgsz_mask = HV_PGSZ_MASK_256MB;
376 hv_pgsz_idx = HV_PGSZ_IDX_256MB;
377 break;
378 case HPAGE_SHIFT:
379 hv_pgsz_mask = HV_PGSZ_MASK_4MB;
380 hv_pgsz_idx = HV_PGSZ_IDX_4MB;
381 break;
382 case HPAGE_64K_SHIFT:
383 hv_pgsz_mask = HV_PGSZ_MASK_64K;
384 hv_pgsz_idx = HV_PGSZ_IDX_64K;
385 break;
386 default:
387 hv_pgsz_mask = 0;
388 }
389
390 if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U)
391 return false;
392
393 return true;
394}
395#endif /* CONFIG_HUGETLB_PAGE */
396
397void update_mmu_cache_range(struct vm_fault *vmf, struct vm_area_struct *vma,
398 unsigned long address, pte_t *ptep, unsigned int nr)
399{
400 struct mm_struct *mm;
401 unsigned long flags;
402 bool is_huge_tsb;
403 pte_t pte = *ptep;
404 unsigned int i;
405
406 if (tlb_type != hypervisor) {
407 unsigned long pfn = pte_pfn(pte);
408
409 if (pfn_valid(pfn))
410 flush_dcache(pfn);
411 }
412
413 mm = vma->vm_mm;
414
415 /* Don't insert a non-valid PTE into the TSB, we'll deadlock. */
416 if (!pte_accessible(mm, pte))
417 return;
418
419 spin_lock_irqsave(&mm->context.lock, flags);
420
421 is_huge_tsb = false;
422#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
423 if (mm->context.hugetlb_pte_count || mm->context.thp_pte_count) {
424 unsigned long hugepage_size = PAGE_SIZE;
425
426 if (is_vm_hugetlb_page(vma))
427 hugepage_size = huge_page_size(hstate_vma(vma));
428
429 if (hugepage_size >= PUD_SIZE) {
430 unsigned long mask = 0x1ffc00000UL;
431
432 /* Transfer bits [32:22] from address to resolve
433 * at 4M granularity.
434 */
435 pte_val(pte) &= ~mask;
436 pte_val(pte) |= (address & mask);
437 } else if (hugepage_size >= PMD_SIZE) {
438 /* We are fabricating 8MB pages using 4MB
439 * real hw pages.
440 */
441 pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT));
442 }
443
444 if (hugepage_size >= PMD_SIZE) {
445 __update_mmu_tsb_insert(mm, MM_TSB_HUGE,
446 REAL_HPAGE_SHIFT, address, pte_val(pte));
447 is_huge_tsb = true;
448 }
449 }
450#endif
451 if (!is_huge_tsb) {
452 for (i = 0; i < nr; i++) {
453 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
454 address, pte_val(pte));
455 address += PAGE_SIZE;
456 pte_val(pte) += PAGE_SIZE;
457 }
458 }
459
460 spin_unlock_irqrestore(&mm->context.lock, flags);
461}
462
463void flush_dcache_folio(struct folio *folio)
464{
465 unsigned long pfn = folio_pfn(folio);
466 struct address_space *mapping;
467 int this_cpu;
468
469 if (tlb_type == hypervisor)
470 return;
471
472 /* Do not bother with the expensive D-cache flush if it
473 * is merely the zero page. The 'bigcore' testcase in GDB
474 * causes this case to run millions of times.
475 */
476 if (is_zero_pfn(pfn))
477 return;
478
479 this_cpu = get_cpu();
480
481 mapping = folio_flush_mapping(folio);
482 if (mapping && !mapping_mapped(mapping)) {
483 bool dirty = test_bit(PG_dcache_dirty, &folio->flags);
484 if (dirty) {
485 int dirty_cpu = dcache_dirty_cpu(folio);
486
487 if (dirty_cpu == this_cpu)
488 goto out;
489 smp_flush_dcache_folio_impl(folio, dirty_cpu);
490 }
491 set_dcache_dirty(folio, this_cpu);
492 } else {
493 /* We could delay the flush for the !folio_mapping
494 * case too. But that case is for exec env/arg
495 * pages and those are %99 certainly going to get
496 * faulted into the tlb (and thus flushed) anyways.
497 */
498 flush_dcache_folio_impl(folio);
499 }
500
501out:
502 put_cpu();
503}
504EXPORT_SYMBOL(flush_dcache_folio);
505
506void __kprobes flush_icache_range(unsigned long start, unsigned long end)
507{
508 /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
509 if (tlb_type == spitfire) {
510 unsigned long kaddr;
511
512 /* This code only runs on Spitfire cpus so this is
513 * why we can assume _PAGE_PADDR_4U.
514 */
515 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
516 unsigned long paddr, mask = _PAGE_PADDR_4U;
517
518 if (kaddr >= PAGE_OFFSET)
519 paddr = kaddr & mask;
520 else {
521 pte_t *ptep = virt_to_kpte(kaddr);
522
523 paddr = pte_val(*ptep) & mask;
524 }
525 __flush_icache_page(paddr);
526 }
527 }
528}
529EXPORT_SYMBOL(flush_icache_range);
530
531void mmu_info(struct seq_file *m)
532{
533 static const char *pgsz_strings[] = {
534 "8K", "64K", "512K", "4MB", "32MB",
535 "256MB", "2GB", "16GB",
536 };
537 int i, printed;
538
539 if (tlb_type == cheetah)
540 seq_printf(m, "MMU Type\t: Cheetah\n");
541 else if (tlb_type == cheetah_plus)
542 seq_printf(m, "MMU Type\t: Cheetah+\n");
543 else if (tlb_type == spitfire)
544 seq_printf(m, "MMU Type\t: Spitfire\n");
545 else if (tlb_type == hypervisor)
546 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
547 else
548 seq_printf(m, "MMU Type\t: ???\n");
549
550 seq_printf(m, "MMU PGSZs\t: ");
551 printed = 0;
552 for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
553 if (cpu_pgsz_mask & (1UL << i)) {
554 seq_printf(m, "%s%s",
555 printed ? "," : "", pgsz_strings[i]);
556 printed++;
557 }
558 }
559 seq_putc(m, '\n');
560
561#ifdef CONFIG_DEBUG_DCFLUSH
562 seq_printf(m, "DCPageFlushes\t: %d\n",
563 atomic_read(&dcpage_flushes));
564#ifdef CONFIG_SMP
565 seq_printf(m, "DCPageFlushesXC\t: %d\n",
566 atomic_read(&dcpage_flushes_xcall));
567#endif /* CONFIG_SMP */
568#endif /* CONFIG_DEBUG_DCFLUSH */
569}
570
571struct linux_prom_translation prom_trans[512] __read_mostly;
572unsigned int prom_trans_ents __read_mostly;
573
574unsigned long kern_locked_tte_data;
575
576/* The obp translations are saved based on 8k pagesize, since obp can
577 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
578 * HI_OBP_ADDRESS range are handled in ktlb.S.
579 */
580static inline int in_obp_range(unsigned long vaddr)
581{
582 return (vaddr >= LOW_OBP_ADDRESS &&
583 vaddr < HI_OBP_ADDRESS);
584}
585
586static int cmp_ptrans(const void *a, const void *b)
587{
588 const struct linux_prom_translation *x = a, *y = b;
589
590 if (x->virt > y->virt)
591 return 1;
592 if (x->virt < y->virt)
593 return -1;
594 return 0;
595}
596
597/* Read OBP translations property into 'prom_trans[]'. */
598static void __init read_obp_translations(void)
599{
600 int n, node, ents, first, last, i;
601
602 node = prom_finddevice("/virtual-memory");
603 n = prom_getproplen(node, "translations");
604 if (unlikely(n == 0 || n == -1)) {
605 prom_printf("prom_mappings: Couldn't get size.\n");
606 prom_halt();
607 }
608 if (unlikely(n > sizeof(prom_trans))) {
609 prom_printf("prom_mappings: Size %d is too big.\n", n);
610 prom_halt();
611 }
612
613 if ((n = prom_getproperty(node, "translations",
614 (char *)&prom_trans[0],
615 sizeof(prom_trans))) == -1) {
616 prom_printf("prom_mappings: Couldn't get property.\n");
617 prom_halt();
618 }
619
620 n = n / sizeof(struct linux_prom_translation);
621
622 ents = n;
623
624 sort(prom_trans, ents, sizeof(struct linux_prom_translation),
625 cmp_ptrans, NULL);
626
627 /* Now kick out all the non-OBP entries. */
628 for (i = 0; i < ents; i++) {
629 if (in_obp_range(prom_trans[i].virt))
630 break;
631 }
632 first = i;
633 for (; i < ents; i++) {
634 if (!in_obp_range(prom_trans[i].virt))
635 break;
636 }
637 last = i;
638
639 for (i = 0; i < (last - first); i++) {
640 struct linux_prom_translation *src = &prom_trans[i + first];
641 struct linux_prom_translation *dest = &prom_trans[i];
642
643 *dest = *src;
644 }
645 for (; i < ents; i++) {
646 struct linux_prom_translation *dest = &prom_trans[i];
647 dest->virt = dest->size = dest->data = 0x0UL;
648 }
649
650 prom_trans_ents = last - first;
651
652 if (tlb_type == spitfire) {
653 /* Clear diag TTE bits. */
654 for (i = 0; i < prom_trans_ents; i++)
655 prom_trans[i].data &= ~0x0003fe0000000000UL;
656 }
657
658 /* Force execute bit on. */
659 for (i = 0; i < prom_trans_ents; i++)
660 prom_trans[i].data |= (tlb_type == hypervisor ?
661 _PAGE_EXEC_4V : _PAGE_EXEC_4U);
662}
663
664static void __init hypervisor_tlb_lock(unsigned long vaddr,
665 unsigned long pte,
666 unsigned long mmu)
667{
668 unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
669
670 if (ret != 0) {
671 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
672 "errors with %lx\n", vaddr, 0, pte, mmu, ret);
673 prom_halt();
674 }
675}
676
677static unsigned long kern_large_tte(unsigned long paddr);
678
679static void __init remap_kernel(void)
680{
681 unsigned long phys_page, tte_vaddr, tte_data;
682 int i, tlb_ent = sparc64_highest_locked_tlbent();
683
684 tte_vaddr = (unsigned long) KERNBASE;
685 phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
686 tte_data = kern_large_tte(phys_page);
687
688 kern_locked_tte_data = tte_data;
689
690 /* Now lock us into the TLBs via Hypervisor or OBP. */
691 if (tlb_type == hypervisor) {
692 for (i = 0; i < num_kernel_image_mappings; i++) {
693 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
694 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
695 tte_vaddr += 0x400000;
696 tte_data += 0x400000;
697 }
698 } else {
699 for (i = 0; i < num_kernel_image_mappings; i++) {
700 prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
701 prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
702 tte_vaddr += 0x400000;
703 tte_data += 0x400000;
704 }
705 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
706 }
707 if (tlb_type == cheetah_plus) {
708 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
709 CTX_CHEETAH_PLUS_NUC);
710 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
711 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
712 }
713}
714
715
716static void __init inherit_prom_mappings(void)
717{
718 /* Now fixup OBP's idea about where we really are mapped. */
719 printk("Remapping the kernel... ");
720 remap_kernel();
721 printk("done.\n");
722}
723
724void prom_world(int enter)
725{
726 /*
727 * No need to change the address space any more, just flush
728 * the register windows
729 */
730 __asm__ __volatile__("flushw");
731}
732
733void __flush_dcache_range(unsigned long start, unsigned long end)
734{
735 unsigned long va;
736
737 if (tlb_type == spitfire) {
738 int n = 0;
739
740 for (va = start; va < end; va += 32) {
741 spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
742 if (++n >= 512)
743 break;
744 }
745 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
746 start = __pa(start);
747 end = __pa(end);
748 for (va = start; va < end; va += 32)
749 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
750 "membar #Sync"
751 : /* no outputs */
752 : "r" (va),
753 "i" (ASI_DCACHE_INVALIDATE));
754 }
755}
756EXPORT_SYMBOL(__flush_dcache_range);
757
758/* get_new_mmu_context() uses "cache + 1". */
759DEFINE_SPINLOCK(ctx_alloc_lock);
760unsigned long tlb_context_cache = CTX_FIRST_VERSION;
761#define MAX_CTX_NR (1UL << CTX_NR_BITS)
762#define CTX_BMAP_SLOTS BITS_TO_LONGS(MAX_CTX_NR)
763DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
764DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0};
765
766static void mmu_context_wrap(void)
767{
768 unsigned long old_ver = tlb_context_cache & CTX_VERSION_MASK;
769 unsigned long new_ver, new_ctx, old_ctx;
770 struct mm_struct *mm;
771 int cpu;
772
773 bitmap_zero(mmu_context_bmap, 1 << CTX_NR_BITS);
774
775 /* Reserve kernel context */
776 set_bit(0, mmu_context_bmap);
777
778 new_ver = (tlb_context_cache & CTX_VERSION_MASK) + CTX_FIRST_VERSION;
779 if (unlikely(new_ver == 0))
780 new_ver = CTX_FIRST_VERSION;
781 tlb_context_cache = new_ver;
782
783 /*
784 * Make sure that any new mm that are added into per_cpu_secondary_mm,
785 * are going to go through get_new_mmu_context() path.
786 */
787 mb();
788
789 /*
790 * Updated versions to current on those CPUs that had valid secondary
791 * contexts
792 */
793 for_each_online_cpu(cpu) {
794 /*
795 * If a new mm is stored after we took this mm from the array,
796 * it will go into get_new_mmu_context() path, because we
797 * already bumped the version in tlb_context_cache.
798 */
799 mm = per_cpu(per_cpu_secondary_mm, cpu);
800
801 if (unlikely(!mm || mm == &init_mm))
802 continue;
803
804 old_ctx = mm->context.sparc64_ctx_val;
805 if (likely((old_ctx & CTX_VERSION_MASK) == old_ver)) {
806 new_ctx = (old_ctx & ~CTX_VERSION_MASK) | new_ver;
807 set_bit(new_ctx & CTX_NR_MASK, mmu_context_bmap);
808 mm->context.sparc64_ctx_val = new_ctx;
809 }
810 }
811}
812
813/* Caller does TLB context flushing on local CPU if necessary.
814 * The caller also ensures that CTX_VALID(mm->context) is false.
815 *
816 * We must be careful about boundary cases so that we never
817 * let the user have CTX 0 (nucleus) or we ever use a CTX
818 * version of zero (and thus NO_CONTEXT would not be caught
819 * by version mis-match tests in mmu_context.h).
820 *
821 * Always invoked with interrupts disabled.
822 */
823void get_new_mmu_context(struct mm_struct *mm)
824{
825 unsigned long ctx, new_ctx;
826 unsigned long orig_pgsz_bits;
827
828 spin_lock(&ctx_alloc_lock);
829retry:
830 /* wrap might have happened, test again if our context became valid */
831 if (unlikely(CTX_VALID(mm->context)))
832 goto out;
833 orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
834 ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
835 new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
836 if (new_ctx >= (1 << CTX_NR_BITS)) {
837 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
838 if (new_ctx >= ctx) {
839 mmu_context_wrap();
840 goto retry;
841 }
842 }
843 if (mm->context.sparc64_ctx_val)
844 cpumask_clear(mm_cpumask(mm));
845 mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
846 new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
847 tlb_context_cache = new_ctx;
848 mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
849out:
850 spin_unlock(&ctx_alloc_lock);
851}
852
853static int numa_enabled = 1;
854static int numa_debug;
855
856static int __init early_numa(char *p)
857{
858 if (!p)
859 return 0;
860
861 if (strstr(p, "off"))
862 numa_enabled = 0;
863
864 if (strstr(p, "debug"))
865 numa_debug = 1;
866
867 return 0;
868}
869early_param("numa", early_numa);
870
871#define numadbg(f, a...) \
872do { if (numa_debug) \
873 printk(KERN_INFO f, ## a); \
874} while (0)
875
876static void __init find_ramdisk(unsigned long phys_base)
877{
878#ifdef CONFIG_BLK_DEV_INITRD
879 if (sparc_ramdisk_image || sparc_ramdisk_image64) {
880 unsigned long ramdisk_image;
881
882 /* Older versions of the bootloader only supported a
883 * 32-bit physical address for the ramdisk image
884 * location, stored at sparc_ramdisk_image. Newer
885 * SILO versions set sparc_ramdisk_image to zero and
886 * provide a full 64-bit physical address at
887 * sparc_ramdisk_image64.
888 */
889 ramdisk_image = sparc_ramdisk_image;
890 if (!ramdisk_image)
891 ramdisk_image = sparc_ramdisk_image64;
892
893 /* Another bootloader quirk. The bootloader normalizes
894 * the physical address to KERNBASE, so we have to
895 * factor that back out and add in the lowest valid
896 * physical page address to get the true physical address.
897 */
898 ramdisk_image -= KERNBASE;
899 ramdisk_image += phys_base;
900
901 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
902 ramdisk_image, sparc_ramdisk_size);
903
904 initrd_start = ramdisk_image;
905 initrd_end = ramdisk_image + sparc_ramdisk_size;
906
907 memblock_reserve(initrd_start, sparc_ramdisk_size);
908
909 initrd_start += PAGE_OFFSET;
910 initrd_end += PAGE_OFFSET;
911 }
912#endif
913}
914
915struct node_mem_mask {
916 unsigned long mask;
917 unsigned long match;
918};
919static struct node_mem_mask node_masks[MAX_NUMNODES];
920static int num_node_masks;
921
922#ifdef CONFIG_NUMA
923
924struct mdesc_mlgroup {
925 u64 node;
926 u64 latency;
927 u64 match;
928 u64 mask;
929};
930
931static struct mdesc_mlgroup *mlgroups;
932static int num_mlgroups;
933
934int numa_cpu_lookup_table[NR_CPUS];
935cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
936
937struct mdesc_mblock {
938 u64 base;
939 u64 size;
940 u64 offset; /* RA-to-PA */
941};
942static struct mdesc_mblock *mblocks;
943static int num_mblocks;
944
945static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr)
946{
947 struct mdesc_mblock *m = NULL;
948 int i;
949
950 for (i = 0; i < num_mblocks; i++) {
951 m = &mblocks[i];
952
953 if (addr >= m->base &&
954 addr < (m->base + m->size)) {
955 break;
956 }
957 }
958
959 return m;
960}
961
962static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid)
963{
964 int prev_nid, new_nid;
965
966 prev_nid = NUMA_NO_NODE;
967 for ( ; start < end; start += PAGE_SIZE) {
968 for (new_nid = 0; new_nid < num_node_masks; new_nid++) {
969 struct node_mem_mask *p = &node_masks[new_nid];
970
971 if ((start & p->mask) == p->match) {
972 if (prev_nid == NUMA_NO_NODE)
973 prev_nid = new_nid;
974 break;
975 }
976 }
977
978 if (new_nid == num_node_masks) {
979 prev_nid = 0;
980 WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.",
981 start);
982 break;
983 }
984
985 if (prev_nid != new_nid)
986 break;
987 }
988 *nid = prev_nid;
989
990 return start > end ? end : start;
991}
992
993static u64 __init memblock_nid_range(u64 start, u64 end, int *nid)
994{
995 u64 ret_end, pa_start, m_mask, m_match, m_end;
996 struct mdesc_mblock *mblock;
997 int _nid, i;
998
999 if (tlb_type != hypervisor)
1000 return memblock_nid_range_sun4u(start, end, nid);
1001
1002 mblock = addr_to_mblock(start);
1003 if (!mblock) {
1004 WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]",
1005 start);
1006
1007 _nid = 0;
1008 ret_end = end;
1009 goto done;
1010 }
1011
1012 pa_start = start + mblock->offset;
1013 m_match = 0;
1014 m_mask = 0;
1015
1016 for (_nid = 0; _nid < num_node_masks; _nid++) {
1017 struct node_mem_mask *const m = &node_masks[_nid];
1018
1019 if ((pa_start & m->mask) == m->match) {
1020 m_match = m->match;
1021 m_mask = m->mask;
1022 break;
1023 }
1024 }
1025
1026 if (num_node_masks == _nid) {
1027 /* We could not find NUMA group, so default to 0, but lets
1028 * search for latency group, so we could calculate the correct
1029 * end address that we return
1030 */
1031 _nid = 0;
1032
1033 for (i = 0; i < num_mlgroups; i++) {
1034 struct mdesc_mlgroup *const m = &mlgroups[i];
1035
1036 if ((pa_start & m->mask) == m->match) {
1037 m_match = m->match;
1038 m_mask = m->mask;
1039 break;
1040 }
1041 }
1042
1043 if (i == num_mlgroups) {
1044 WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]",
1045 start);
1046
1047 ret_end = end;
1048 goto done;
1049 }
1050 }
1051
1052 /*
1053 * Each latency group has match and mask, and each memory block has an
1054 * offset. An address belongs to a latency group if its address matches
1055 * the following formula: ((addr + offset) & mask) == match
1056 * It is, however, slow to check every single page if it matches a
1057 * particular latency group. As optimization we calculate end value by
1058 * using bit arithmetics.
1059 */
1060 m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset;
1061 m_end += pa_start & ~((1ul << fls64(m_mask)) - 1);
1062 ret_end = m_end > end ? end : m_end;
1063
1064done:
1065 *nid = _nid;
1066 return ret_end;
1067}
1068#endif
1069
1070/* This must be invoked after performing all of the necessary
1071 * memblock_set_node() calls for 'nid'. We need to be able to get
1072 * correct data from get_pfn_range_for_nid().
1073 */
1074static void __init allocate_node_data(int nid)
1075{
1076 struct pglist_data *p;
1077 unsigned long start_pfn, end_pfn;
1078
1079#ifdef CONFIG_NUMA
1080 alloc_node_data(nid);
1081
1082 NODE_DATA(nid)->node_id = nid;
1083#endif
1084
1085 p = NODE_DATA(nid);
1086
1087 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1088 p->node_start_pfn = start_pfn;
1089 p->node_spanned_pages = end_pfn - start_pfn;
1090}
1091
1092static void init_node_masks_nonnuma(void)
1093{
1094#ifdef CONFIG_NUMA
1095 int i;
1096#endif
1097
1098 numadbg("Initializing tables for non-numa.\n");
1099
1100 node_masks[0].mask = 0;
1101 node_masks[0].match = 0;
1102 num_node_masks = 1;
1103
1104#ifdef CONFIG_NUMA
1105 for (i = 0; i < NR_CPUS; i++)
1106 numa_cpu_lookup_table[i] = 0;
1107
1108 cpumask_setall(&numa_cpumask_lookup_table[0]);
1109#endif
1110}
1111
1112#ifdef CONFIG_NUMA
1113
1114EXPORT_SYMBOL(numa_cpu_lookup_table);
1115EXPORT_SYMBOL(numa_cpumask_lookup_table);
1116
1117static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
1118 u32 cfg_handle)
1119{
1120 u64 arc;
1121
1122 mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
1123 u64 target = mdesc_arc_target(md, arc);
1124 const u64 *val;
1125
1126 val = mdesc_get_property(md, target,
1127 "cfg-handle", NULL);
1128 if (val && *val == cfg_handle)
1129 return 0;
1130 }
1131 return -ENODEV;
1132}
1133
1134static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
1135 u32 cfg_handle)
1136{
1137 u64 arc, candidate, best_latency = ~(u64)0;
1138
1139 candidate = MDESC_NODE_NULL;
1140 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1141 u64 target = mdesc_arc_target(md, arc);
1142 const char *name = mdesc_node_name(md, target);
1143 const u64 *val;
1144
1145 if (strcmp(name, "pio-latency-group"))
1146 continue;
1147
1148 val = mdesc_get_property(md, target, "latency", NULL);
1149 if (!val)
1150 continue;
1151
1152 if (*val < best_latency) {
1153 candidate = target;
1154 best_latency = *val;
1155 }
1156 }
1157
1158 if (candidate == MDESC_NODE_NULL)
1159 return -ENODEV;
1160
1161 return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
1162}
1163
1164int of_node_to_nid(struct device_node *dp)
1165{
1166 const struct linux_prom64_registers *regs;
1167 struct mdesc_handle *md;
1168 u32 cfg_handle;
1169 int count, nid;
1170 u64 grp;
1171
1172 /* This is the right thing to do on currently supported
1173 * SUN4U NUMA platforms as well, as the PCI controller does
1174 * not sit behind any particular memory controller.
1175 */
1176 if (!mlgroups)
1177 return -1;
1178
1179 regs = of_get_property(dp, "reg", NULL);
1180 if (!regs)
1181 return -1;
1182
1183 cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1184
1185 md = mdesc_grab();
1186
1187 count = 0;
1188 nid = NUMA_NO_NODE;
1189 mdesc_for_each_node_by_name(md, grp, "group") {
1190 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1191 nid = count;
1192 break;
1193 }
1194 count++;
1195 }
1196
1197 mdesc_release(md);
1198
1199 return nid;
1200}
1201
1202static void __init add_node_ranges(void)
1203{
1204 phys_addr_t start, end;
1205 unsigned long prev_max;
1206 u64 i;
1207
1208memblock_resized:
1209 prev_max = memblock.memory.max;
1210
1211 for_each_mem_range(i, &start, &end) {
1212 while (start < end) {
1213 unsigned long this_end;
1214 int nid;
1215
1216 this_end = memblock_nid_range(start, end, &nid);
1217
1218 numadbg("Setting memblock NUMA node nid[%d] "
1219 "start[%llx] end[%lx]\n",
1220 nid, start, this_end);
1221
1222 memblock_set_node(start, this_end - start,
1223 &memblock.memory, nid);
1224 if (memblock.memory.max != prev_max)
1225 goto memblock_resized;
1226 start = this_end;
1227 }
1228 }
1229}
1230
1231static int __init grab_mlgroups(struct mdesc_handle *md)
1232{
1233 unsigned long paddr;
1234 int count = 0;
1235 u64 node;
1236
1237 mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1238 count++;
1239 if (!count)
1240 return -ENOENT;
1241
1242 paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mlgroup),
1243 SMP_CACHE_BYTES);
1244 if (!paddr)
1245 return -ENOMEM;
1246
1247 mlgroups = __va(paddr);
1248 num_mlgroups = count;
1249
1250 count = 0;
1251 mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1252 struct mdesc_mlgroup *m = &mlgroups[count++];
1253 const u64 *val;
1254
1255 m->node = node;
1256
1257 val = mdesc_get_property(md, node, "latency", NULL);
1258 m->latency = *val;
1259 val = mdesc_get_property(md, node, "address-match", NULL);
1260 m->match = *val;
1261 val = mdesc_get_property(md, node, "address-mask", NULL);
1262 m->mask = *val;
1263
1264 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1265 "match[%llx] mask[%llx]\n",
1266 count - 1, m->node, m->latency, m->match, m->mask);
1267 }
1268
1269 return 0;
1270}
1271
1272static int __init grab_mblocks(struct mdesc_handle *md)
1273{
1274 unsigned long paddr;
1275 int count = 0;
1276 u64 node;
1277
1278 mdesc_for_each_node_by_name(md, node, "mblock")
1279 count++;
1280 if (!count)
1281 return -ENOENT;
1282
1283 paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mblock),
1284 SMP_CACHE_BYTES);
1285 if (!paddr)
1286 return -ENOMEM;
1287
1288 mblocks = __va(paddr);
1289 num_mblocks = count;
1290
1291 count = 0;
1292 mdesc_for_each_node_by_name(md, node, "mblock") {
1293 struct mdesc_mblock *m = &mblocks[count++];
1294 const u64 *val;
1295
1296 val = mdesc_get_property(md, node, "base", NULL);
1297 m->base = *val;
1298 val = mdesc_get_property(md, node, "size", NULL);
1299 m->size = *val;
1300 val = mdesc_get_property(md, node,
1301 "address-congruence-offset", NULL);
1302
1303 /* The address-congruence-offset property is optional.
1304 * Explicity zero it be identifty this.
1305 */
1306 if (val)
1307 m->offset = *val;
1308 else
1309 m->offset = 0UL;
1310
1311 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1312 count - 1, m->base, m->size, m->offset);
1313 }
1314
1315 return 0;
1316}
1317
1318static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1319 u64 grp, cpumask_t *mask)
1320{
1321 u64 arc;
1322
1323 cpumask_clear(mask);
1324
1325 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1326 u64 target = mdesc_arc_target(md, arc);
1327 const char *name = mdesc_node_name(md, target);
1328 const u64 *id;
1329
1330 if (strcmp(name, "cpu"))
1331 continue;
1332 id = mdesc_get_property(md, target, "id", NULL);
1333 if (*id < nr_cpu_ids)
1334 cpumask_set_cpu(*id, mask);
1335 }
1336}
1337
1338static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1339{
1340 int i;
1341
1342 for (i = 0; i < num_mlgroups; i++) {
1343 struct mdesc_mlgroup *m = &mlgroups[i];
1344 if (m->node == node)
1345 return m;
1346 }
1347 return NULL;
1348}
1349
1350int __node_distance(int from, int to)
1351{
1352 if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1353 pr_warn("Returning default NUMA distance value for %d->%d\n",
1354 from, to);
1355 return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1356 }
1357 return numa_latency[from][to];
1358}
1359EXPORT_SYMBOL(__node_distance);
1360
1361static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1362{
1363 int i;
1364
1365 for (i = 0; i < MAX_NUMNODES; i++) {
1366 struct node_mem_mask *n = &node_masks[i];
1367
1368 if ((grp->mask == n->mask) && (grp->match == n->match))
1369 break;
1370 }
1371 return i;
1372}
1373
1374static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1375 u64 grp, int index)
1376{
1377 u64 arc;
1378
1379 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1380 int tnode;
1381 u64 target = mdesc_arc_target(md, arc);
1382 struct mdesc_mlgroup *m = find_mlgroup(target);
1383
1384 if (!m)
1385 continue;
1386 tnode = find_best_numa_node_for_mlgroup(m);
1387 if (tnode == MAX_NUMNODES)
1388 continue;
1389 numa_latency[index][tnode] = m->latency;
1390 }
1391}
1392
1393static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1394 int index)
1395{
1396 struct mdesc_mlgroup *candidate = NULL;
1397 u64 arc, best_latency = ~(u64)0;
1398 struct node_mem_mask *n;
1399
1400 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1401 u64 target = mdesc_arc_target(md, arc);
1402 struct mdesc_mlgroup *m = find_mlgroup(target);
1403 if (!m)
1404 continue;
1405 if (m->latency < best_latency) {
1406 candidate = m;
1407 best_latency = m->latency;
1408 }
1409 }
1410 if (!candidate)
1411 return -ENOENT;
1412
1413 if (num_node_masks != index) {
1414 printk(KERN_ERR "Inconsistent NUMA state, "
1415 "index[%d] != num_node_masks[%d]\n",
1416 index, num_node_masks);
1417 return -EINVAL;
1418 }
1419
1420 n = &node_masks[num_node_masks++];
1421
1422 n->mask = candidate->mask;
1423 n->match = candidate->match;
1424
1425 numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n",
1426 index, n->mask, n->match, candidate->latency);
1427
1428 return 0;
1429}
1430
1431static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1432 int index)
1433{
1434 cpumask_t mask;
1435 int cpu;
1436
1437 numa_parse_mdesc_group_cpus(md, grp, &mask);
1438
1439 for_each_cpu(cpu, &mask)
1440 numa_cpu_lookup_table[cpu] = index;
1441 cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1442
1443 if (numa_debug) {
1444 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1445 for_each_cpu(cpu, &mask)
1446 printk("%d ", cpu);
1447 printk("]\n");
1448 }
1449
1450 return numa_attach_mlgroup(md, grp, index);
1451}
1452
1453static int __init numa_parse_mdesc(void)
1454{
1455 struct mdesc_handle *md = mdesc_grab();
1456 int i, j, err, count;
1457 u64 node;
1458
1459 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1460 if (node == MDESC_NODE_NULL) {
1461 mdesc_release(md);
1462 return -ENOENT;
1463 }
1464
1465 err = grab_mblocks(md);
1466 if (err < 0)
1467 goto out;
1468
1469 err = grab_mlgroups(md);
1470 if (err < 0)
1471 goto out;
1472
1473 count = 0;
1474 mdesc_for_each_node_by_name(md, node, "group") {
1475 err = numa_parse_mdesc_group(md, node, count);
1476 if (err < 0)
1477 break;
1478 count++;
1479 }
1480
1481 count = 0;
1482 mdesc_for_each_node_by_name(md, node, "group") {
1483 find_numa_latencies_for_group(md, node, count);
1484 count++;
1485 }
1486
1487 /* Normalize numa latency matrix according to ACPI SLIT spec. */
1488 for (i = 0; i < MAX_NUMNODES; i++) {
1489 u64 self_latency = numa_latency[i][i];
1490
1491 for (j = 0; j < MAX_NUMNODES; j++) {
1492 numa_latency[i][j] =
1493 (numa_latency[i][j] * LOCAL_DISTANCE) /
1494 self_latency;
1495 }
1496 }
1497
1498 add_node_ranges();
1499
1500 for (i = 0; i < num_node_masks; i++) {
1501 allocate_node_data(i);
1502 node_set_online(i);
1503 }
1504
1505 err = 0;
1506out:
1507 mdesc_release(md);
1508 return err;
1509}
1510
1511static int __init numa_parse_jbus(void)
1512{
1513 unsigned long cpu, index;
1514
1515 /* NUMA node id is encoded in bits 36 and higher, and there is
1516 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1517 */
1518 index = 0;
1519 for_each_present_cpu(cpu) {
1520 numa_cpu_lookup_table[cpu] = index;
1521 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1522 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1523 node_masks[index].match = cpu << 36UL;
1524
1525 index++;
1526 }
1527 num_node_masks = index;
1528
1529 add_node_ranges();
1530
1531 for (index = 0; index < num_node_masks; index++) {
1532 allocate_node_data(index);
1533 node_set_online(index);
1534 }
1535
1536 return 0;
1537}
1538
1539static int __init numa_parse_sun4u(void)
1540{
1541 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1542 unsigned long ver;
1543
1544 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1545 if ((ver >> 32UL) == __JALAPENO_ID ||
1546 (ver >> 32UL) == __SERRANO_ID)
1547 return numa_parse_jbus();
1548 }
1549 return -1;
1550}
1551
1552static int __init bootmem_init_numa(void)
1553{
1554 int i, j;
1555 int err = -1;
1556
1557 numadbg("bootmem_init_numa()\n");
1558
1559 /* Some sane defaults for numa latency values */
1560 for (i = 0; i < MAX_NUMNODES; i++) {
1561 for (j = 0; j < MAX_NUMNODES; j++)
1562 numa_latency[i][j] = (i == j) ?
1563 LOCAL_DISTANCE : REMOTE_DISTANCE;
1564 }
1565
1566 if (numa_enabled) {
1567 if (tlb_type == hypervisor)
1568 err = numa_parse_mdesc();
1569 else
1570 err = numa_parse_sun4u();
1571 }
1572 return err;
1573}
1574
1575#else
1576
1577static int bootmem_init_numa(void)
1578{
1579 return -1;
1580}
1581
1582#endif
1583
1584static void __init bootmem_init_nonnuma(void)
1585{
1586 unsigned long top_of_ram = memblock_end_of_DRAM();
1587 unsigned long total_ram = memblock_phys_mem_size();
1588
1589 numadbg("bootmem_init_nonnuma()\n");
1590
1591 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1592 top_of_ram, total_ram);
1593 printk(KERN_INFO "Memory hole size: %ldMB\n",
1594 (top_of_ram - total_ram) >> 20);
1595
1596 init_node_masks_nonnuma();
1597 memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
1598 allocate_node_data(0);
1599 node_set_online(0);
1600}
1601
1602static unsigned long __init bootmem_init(unsigned long phys_base)
1603{
1604 unsigned long end_pfn;
1605
1606 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1607 max_pfn = max_low_pfn = end_pfn;
1608 min_low_pfn = (phys_base >> PAGE_SHIFT);
1609
1610 if (bootmem_init_numa() < 0)
1611 bootmem_init_nonnuma();
1612
1613 /* Dump memblock with node info. */
1614 memblock_dump_all();
1615
1616 /* XXX cpu notifier XXX */
1617
1618 sparse_init();
1619
1620 return end_pfn;
1621}
1622
1623static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1624static int pall_ents __initdata;
1625
1626static unsigned long max_phys_bits = 40;
1627
1628bool kern_addr_valid(unsigned long addr)
1629{
1630 pgd_t *pgd;
1631 p4d_t *p4d;
1632 pud_t *pud;
1633 pmd_t *pmd;
1634 pte_t *pte;
1635
1636 if ((long)addr < 0L) {
1637 unsigned long pa = __pa(addr);
1638
1639 if ((pa >> max_phys_bits) != 0UL)
1640 return false;
1641
1642 return pfn_valid(pa >> PAGE_SHIFT);
1643 }
1644
1645 if (addr >= (unsigned long) KERNBASE &&
1646 addr < (unsigned long)&_end)
1647 return true;
1648
1649 pgd = pgd_offset_k(addr);
1650 if (pgd_none(*pgd))
1651 return false;
1652
1653 p4d = p4d_offset(pgd, addr);
1654 if (p4d_none(*p4d))
1655 return false;
1656
1657 pud = pud_offset(p4d, addr);
1658 if (pud_none(*pud))
1659 return false;
1660
1661 if (pud_leaf(*pud))
1662 return pfn_valid(pud_pfn(*pud));
1663
1664 pmd = pmd_offset(pud, addr);
1665 if (pmd_none(*pmd))
1666 return false;
1667
1668 if (pmd_leaf(*pmd))
1669 return pfn_valid(pmd_pfn(*pmd));
1670
1671 pte = pte_offset_kernel(pmd, addr);
1672 if (pte_none(*pte))
1673 return false;
1674
1675 return pfn_valid(pte_pfn(*pte));
1676}
1677
1678static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1679 unsigned long vend,
1680 pud_t *pud)
1681{
1682 const unsigned long mask16gb = (1UL << 34) - 1UL;
1683 u64 pte_val = vstart;
1684
1685 /* Each PUD is 8GB */
1686 if ((vstart & mask16gb) ||
1687 (vend - vstart <= mask16gb)) {
1688 pte_val ^= kern_linear_pte_xor[2];
1689 pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1690
1691 return vstart + PUD_SIZE;
1692 }
1693
1694 pte_val ^= kern_linear_pte_xor[3];
1695 pte_val |= _PAGE_PUD_HUGE;
1696
1697 vend = vstart + mask16gb + 1UL;
1698 while (vstart < vend) {
1699 pud_val(*pud) = pte_val;
1700
1701 pte_val += PUD_SIZE;
1702 vstart += PUD_SIZE;
1703 pud++;
1704 }
1705 return vstart;
1706}
1707
1708static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1709 bool guard)
1710{
1711 if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1712 return true;
1713
1714 return false;
1715}
1716
1717static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1718 unsigned long vend,
1719 pmd_t *pmd)
1720{
1721 const unsigned long mask256mb = (1UL << 28) - 1UL;
1722 const unsigned long mask2gb = (1UL << 31) - 1UL;
1723 u64 pte_val = vstart;
1724
1725 /* Each PMD is 8MB */
1726 if ((vstart & mask256mb) ||
1727 (vend - vstart <= mask256mb)) {
1728 pte_val ^= kern_linear_pte_xor[0];
1729 pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1730
1731 return vstart + PMD_SIZE;
1732 }
1733
1734 if ((vstart & mask2gb) ||
1735 (vend - vstart <= mask2gb)) {
1736 pte_val ^= kern_linear_pte_xor[1];
1737 pte_val |= _PAGE_PMD_HUGE;
1738 vend = vstart + mask256mb + 1UL;
1739 } else {
1740 pte_val ^= kern_linear_pte_xor[2];
1741 pte_val |= _PAGE_PMD_HUGE;
1742 vend = vstart + mask2gb + 1UL;
1743 }
1744
1745 while (vstart < vend) {
1746 pmd_val(*pmd) = pte_val;
1747
1748 pte_val += PMD_SIZE;
1749 vstart += PMD_SIZE;
1750 pmd++;
1751 }
1752
1753 return vstart;
1754}
1755
1756static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1757 bool guard)
1758{
1759 if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1760 return true;
1761
1762 return false;
1763}
1764
1765static unsigned long __ref kernel_map_range(unsigned long pstart,
1766 unsigned long pend, pgprot_t prot,
1767 bool use_huge)
1768{
1769 unsigned long vstart = PAGE_OFFSET + pstart;
1770 unsigned long vend = PAGE_OFFSET + pend;
1771 unsigned long alloc_bytes = 0UL;
1772
1773 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1774 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1775 vstart, vend);
1776 prom_halt();
1777 }
1778
1779 while (vstart < vend) {
1780 unsigned long this_end, paddr = __pa(vstart);
1781 pgd_t *pgd = pgd_offset_k(vstart);
1782 p4d_t *p4d;
1783 pud_t *pud;
1784 pmd_t *pmd;
1785 pte_t *pte;
1786
1787 if (pgd_none(*pgd)) {
1788 pud_t *new;
1789
1790 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1791 PAGE_SIZE);
1792 if (!new)
1793 goto err_alloc;
1794 alloc_bytes += PAGE_SIZE;
1795 pgd_populate(&init_mm, pgd, new);
1796 }
1797
1798 p4d = p4d_offset(pgd, vstart);
1799 if (p4d_none(*p4d)) {
1800 pud_t *new;
1801
1802 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1803 PAGE_SIZE);
1804 if (!new)
1805 goto err_alloc;
1806 alloc_bytes += PAGE_SIZE;
1807 p4d_populate(&init_mm, p4d, new);
1808 }
1809
1810 pud = pud_offset(p4d, vstart);
1811 if (pud_none(*pud)) {
1812 pmd_t *new;
1813
1814 if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1815 vstart = kernel_map_hugepud(vstart, vend, pud);
1816 continue;
1817 }
1818 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1819 PAGE_SIZE);
1820 if (!new)
1821 goto err_alloc;
1822 alloc_bytes += PAGE_SIZE;
1823 pud_populate(&init_mm, pud, new);
1824 }
1825
1826 pmd = pmd_offset(pud, vstart);
1827 if (pmd_none(*pmd)) {
1828 pte_t *new;
1829
1830 if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1831 vstart = kernel_map_hugepmd(vstart, vend, pmd);
1832 continue;
1833 }
1834 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1835 PAGE_SIZE);
1836 if (!new)
1837 goto err_alloc;
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
1858err_alloc:
1859 panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n",
1860 __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1861 return -ENOMEM;
1862}
1863
1864static void __init flush_all_kernel_tsbs(void)
1865{
1866 int i;
1867
1868 for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1869 struct tsb *ent = &swapper_tsb[i];
1870
1871 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1872 }
1873#ifndef CONFIG_DEBUG_PAGEALLOC
1874 for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1875 struct tsb *ent = &swapper_4m_tsb[i];
1876
1877 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1878 }
1879#endif
1880}
1881
1882extern unsigned int kvmap_linear_patch[1];
1883
1884static void __init kernel_physical_mapping_init(void)
1885{
1886 unsigned long i, mem_alloced = 0UL;
1887 bool use_huge = true;
1888
1889#ifdef CONFIG_DEBUG_PAGEALLOC
1890 use_huge = false;
1891#endif
1892 for (i = 0; i < pall_ents; i++) {
1893 unsigned long phys_start, phys_end;
1894
1895 phys_start = pall[i].phys_addr;
1896 phys_end = phys_start + pall[i].reg_size;
1897
1898 mem_alloced += kernel_map_range(phys_start, phys_end,
1899 PAGE_KERNEL, use_huge);
1900 }
1901
1902 printk("Allocated %ld bytes for kernel page tables.\n",
1903 mem_alloced);
1904
1905 kvmap_linear_patch[0] = 0x01000000; /* nop */
1906 flushi(&kvmap_linear_patch[0]);
1907
1908 flush_all_kernel_tsbs();
1909
1910 __flush_tlb_all();
1911}
1912
1913#ifdef CONFIG_DEBUG_PAGEALLOC
1914void __kernel_map_pages(struct page *page, int numpages, int enable)
1915{
1916 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1917 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1918
1919 kernel_map_range(phys_start, phys_end,
1920 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1921
1922 flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1923 PAGE_OFFSET + phys_end);
1924
1925 /* we should perform an IPI and flush all tlbs,
1926 * but that can deadlock->flush only current cpu.
1927 */
1928 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1929 PAGE_OFFSET + phys_end);
1930}
1931#endif
1932
1933unsigned long __init find_ecache_flush_span(unsigned long size)
1934{
1935 int i;
1936
1937 for (i = 0; i < pavail_ents; i++) {
1938 if (pavail[i].reg_size >= size)
1939 return pavail[i].phys_addr;
1940 }
1941
1942 return ~0UL;
1943}
1944
1945unsigned long PAGE_OFFSET;
1946EXPORT_SYMBOL(PAGE_OFFSET);
1947
1948unsigned long VMALLOC_END = 0x0000010000000000UL;
1949EXPORT_SYMBOL(VMALLOC_END);
1950
1951unsigned long sparc64_va_hole_top = 0xfffff80000000000UL;
1952unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1953
1954static void __init setup_page_offset(void)
1955{
1956 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1957 /* Cheetah/Panther support a full 64-bit virtual
1958 * address, so we can use all that our page tables
1959 * support.
1960 */
1961 sparc64_va_hole_top = 0xfff0000000000000UL;
1962 sparc64_va_hole_bottom = 0x0010000000000000UL;
1963
1964 max_phys_bits = 42;
1965 } else if (tlb_type == hypervisor) {
1966 switch (sun4v_chip_type) {
1967 case SUN4V_CHIP_NIAGARA1:
1968 case SUN4V_CHIP_NIAGARA2:
1969 /* T1 and T2 support 48-bit virtual addresses. */
1970 sparc64_va_hole_top = 0xffff800000000000UL;
1971 sparc64_va_hole_bottom = 0x0000800000000000UL;
1972
1973 max_phys_bits = 39;
1974 break;
1975 case SUN4V_CHIP_NIAGARA3:
1976 /* T3 supports 48-bit virtual addresses. */
1977 sparc64_va_hole_top = 0xffff800000000000UL;
1978 sparc64_va_hole_bottom = 0x0000800000000000UL;
1979
1980 max_phys_bits = 43;
1981 break;
1982 case SUN4V_CHIP_NIAGARA4:
1983 case SUN4V_CHIP_NIAGARA5:
1984 case SUN4V_CHIP_SPARC64X:
1985 case SUN4V_CHIP_SPARC_M6:
1986 /* T4 and later support 52-bit virtual addresses. */
1987 sparc64_va_hole_top = 0xfff8000000000000UL;
1988 sparc64_va_hole_bottom = 0x0008000000000000UL;
1989 max_phys_bits = 47;
1990 break;
1991 case SUN4V_CHIP_SPARC_M7:
1992 case SUN4V_CHIP_SPARC_SN:
1993 /* M7 and later support 52-bit virtual addresses. */
1994 sparc64_va_hole_top = 0xfff8000000000000UL;
1995 sparc64_va_hole_bottom = 0x0008000000000000UL;
1996 max_phys_bits = 49;
1997 break;
1998 case SUN4V_CHIP_SPARC_M8:
1999 default:
2000 /* M8 and later support 54-bit virtual addresses.
2001 * However, restricting M8 and above VA bits to 53
2002 * as 4-level page table cannot support more than
2003 * 53 VA bits.
2004 */
2005 sparc64_va_hole_top = 0xfff0000000000000UL;
2006 sparc64_va_hole_bottom = 0x0010000000000000UL;
2007 max_phys_bits = 51;
2008 break;
2009 }
2010 }
2011
2012 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
2013 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
2014 max_phys_bits);
2015 prom_halt();
2016 }
2017
2018 PAGE_OFFSET = sparc64_va_hole_top;
2019 VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
2020 (sparc64_va_hole_bottom >> 2));
2021
2022 pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
2023 PAGE_OFFSET, max_phys_bits);
2024 pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
2025 VMALLOC_START, VMALLOC_END);
2026 pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
2027 VMEMMAP_BASE, VMEMMAP_BASE << 1);
2028}
2029
2030static void __init tsb_phys_patch(void)
2031{
2032 struct tsb_ldquad_phys_patch_entry *pquad;
2033 struct tsb_phys_patch_entry *p;
2034
2035 pquad = &__tsb_ldquad_phys_patch;
2036 while (pquad < &__tsb_ldquad_phys_patch_end) {
2037 unsigned long addr = pquad->addr;
2038
2039 if (tlb_type == hypervisor)
2040 *(unsigned int *) addr = pquad->sun4v_insn;
2041 else
2042 *(unsigned int *) addr = pquad->sun4u_insn;
2043 wmb();
2044 __asm__ __volatile__("flush %0"
2045 : /* no outputs */
2046 : "r" (addr));
2047
2048 pquad++;
2049 }
2050
2051 p = &__tsb_phys_patch;
2052 while (p < &__tsb_phys_patch_end) {
2053 unsigned long addr = p->addr;
2054
2055 *(unsigned int *) addr = p->insn;
2056 wmb();
2057 __asm__ __volatile__("flush %0"
2058 : /* no outputs */
2059 : "r" (addr));
2060
2061 p++;
2062 }
2063}
2064
2065/* Don't mark as init, we give this to the Hypervisor. */
2066#ifndef CONFIG_DEBUG_PAGEALLOC
2067#define NUM_KTSB_DESCR 2
2068#else
2069#define NUM_KTSB_DESCR 1
2070#endif
2071static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
2072
2073/* The swapper TSBs are loaded with a base sequence of:
2074 *
2075 * sethi %uhi(SYMBOL), REG1
2076 * sethi %hi(SYMBOL), REG2
2077 * or REG1, %ulo(SYMBOL), REG1
2078 * or REG2, %lo(SYMBOL), REG2
2079 * sllx REG1, 32, REG1
2080 * or REG1, REG2, REG1
2081 *
2082 * When we use physical addressing for the TSB accesses, we patch the
2083 * first four instructions in the above sequence.
2084 */
2085
2086static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
2087{
2088 unsigned long high_bits, low_bits;
2089
2090 high_bits = (pa >> 32) & 0xffffffff;
2091 low_bits = (pa >> 0) & 0xffffffff;
2092
2093 while (start < end) {
2094 unsigned int *ia = (unsigned int *)(unsigned long)*start;
2095
2096 ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
2097 __asm__ __volatile__("flush %0" : : "r" (ia));
2098
2099 ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
2100 __asm__ __volatile__("flush %0" : : "r" (ia + 1));
2101
2102 ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
2103 __asm__ __volatile__("flush %0" : : "r" (ia + 2));
2104
2105 ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
2106 __asm__ __volatile__("flush %0" : : "r" (ia + 3));
2107
2108 start++;
2109 }
2110}
2111
2112static void ktsb_phys_patch(void)
2113{
2114 extern unsigned int __swapper_tsb_phys_patch;
2115 extern unsigned int __swapper_tsb_phys_patch_end;
2116 unsigned long ktsb_pa;
2117
2118 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2119 patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
2120 &__swapper_tsb_phys_patch_end, ktsb_pa);
2121#ifndef CONFIG_DEBUG_PAGEALLOC
2122 {
2123 extern unsigned int __swapper_4m_tsb_phys_patch;
2124 extern unsigned int __swapper_4m_tsb_phys_patch_end;
2125 ktsb_pa = (kern_base +
2126 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2127 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
2128 &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
2129 }
2130#endif
2131}
2132
2133static void __init sun4v_ktsb_init(void)
2134{
2135 unsigned long ktsb_pa;
2136
2137 /* First KTSB for PAGE_SIZE mappings. */
2138 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2139
2140 switch (PAGE_SIZE) {
2141 case 8 * 1024:
2142 default:
2143 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
2144 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
2145 break;
2146
2147 case 64 * 1024:
2148 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
2149 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
2150 break;
2151
2152 case 512 * 1024:
2153 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
2154 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
2155 break;
2156
2157 case 4 * 1024 * 1024:
2158 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
2159 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
2160 break;
2161 }
2162
2163 ktsb_descr[0].assoc = 1;
2164 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
2165 ktsb_descr[0].ctx_idx = 0;
2166 ktsb_descr[0].tsb_base = ktsb_pa;
2167 ktsb_descr[0].resv = 0;
2168
2169#ifndef CONFIG_DEBUG_PAGEALLOC
2170 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */
2171 ktsb_pa = (kern_base +
2172 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2173
2174 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
2175 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
2176 HV_PGSZ_MASK_256MB |
2177 HV_PGSZ_MASK_2GB |
2178 HV_PGSZ_MASK_16GB) &
2179 cpu_pgsz_mask);
2180 ktsb_descr[1].assoc = 1;
2181 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
2182 ktsb_descr[1].ctx_idx = 0;
2183 ktsb_descr[1].tsb_base = ktsb_pa;
2184 ktsb_descr[1].resv = 0;
2185#endif
2186}
2187
2188void sun4v_ktsb_register(void)
2189{
2190 unsigned long pa, ret;
2191
2192 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
2193
2194 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
2195 if (ret != 0) {
2196 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
2197 "errors with %lx\n", pa, ret);
2198 prom_halt();
2199 }
2200}
2201
2202static void __init sun4u_linear_pte_xor_finalize(void)
2203{
2204#ifndef CONFIG_DEBUG_PAGEALLOC
2205 /* This is where we would add Panther support for
2206 * 32MB and 256MB pages.
2207 */
2208#endif
2209}
2210
2211static void __init sun4v_linear_pte_xor_finalize(void)
2212{
2213 unsigned long pagecv_flag;
2214
2215 /* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
2216 * enables MCD error. Do not set bit 9 on M7 processor.
2217 */
2218 switch (sun4v_chip_type) {
2219 case SUN4V_CHIP_SPARC_M7:
2220 case SUN4V_CHIP_SPARC_M8:
2221 case SUN4V_CHIP_SPARC_SN:
2222 pagecv_flag = 0x00;
2223 break;
2224 default:
2225 pagecv_flag = _PAGE_CV_4V;
2226 break;
2227 }
2228#ifndef CONFIG_DEBUG_PAGEALLOC
2229 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
2230 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2231 PAGE_OFFSET;
2232 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2233 _PAGE_P_4V | _PAGE_W_4V);
2234 } else {
2235 kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2236 }
2237
2238 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2239 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2240 PAGE_OFFSET;
2241 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2242 _PAGE_P_4V | _PAGE_W_4V);
2243 } else {
2244 kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2245 }
2246
2247 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2248 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2249 PAGE_OFFSET;
2250 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2251 _PAGE_P_4V | _PAGE_W_4V);
2252 } else {
2253 kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2254 }
2255#endif
2256}
2257
2258/* paging_init() sets up the page tables */
2259
2260static unsigned long last_valid_pfn;
2261
2262static void sun4u_pgprot_init(void);
2263static void sun4v_pgprot_init(void);
2264
2265#define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U)
2266#define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V)
2267#define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2268#define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2269#define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2270#define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2271
2272/* We need to exclude reserved regions. This exclusion will include
2273 * vmlinux and initrd. To be more precise the initrd size could be used to
2274 * compute a new lower limit because it is freed later during initialization.
2275 */
2276static void __init reduce_memory(phys_addr_t limit_ram)
2277{
2278 limit_ram += memblock_reserved_size();
2279 memblock_enforce_memory_limit(limit_ram);
2280}
2281
2282void __init paging_init(void)
2283{
2284 unsigned long end_pfn, shift, phys_base;
2285 unsigned long real_end, i;
2286
2287 setup_page_offset();
2288
2289 /* These build time checkes make sure that the dcache_dirty_cpu()
2290 * folio->flags usage will work.
2291 *
2292 * When a page gets marked as dcache-dirty, we store the
2293 * cpu number starting at bit 32 in the folio->flags. Also,
2294 * functions like clear_dcache_dirty_cpu use the cpu mask
2295 * in 13-bit signed-immediate instruction fields.
2296 */
2297
2298 /*
2299 * Page flags must not reach into upper 32 bits that are used
2300 * for the cpu number
2301 */
2302 BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2303
2304 /*
2305 * The bit fields placed in the high range must not reach below
2306 * the 32 bit boundary. Otherwise we cannot place the cpu field
2307 * at the 32 bit boundary.
2308 */
2309 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2310 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2311
2312 BUILD_BUG_ON(NR_CPUS > 4096);
2313
2314 kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2315 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2316
2317 /* Invalidate both kernel TSBs. */
2318 memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2319#ifndef CONFIG_DEBUG_PAGEALLOC
2320 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2321#endif
2322
2323 /* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2324 * bit on M7 processor. This is a conflicting usage of the same
2325 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2326 * Detection error on all pages and this will lead to problems
2327 * later. Kernel does not run with MCD enabled and hence rest
2328 * of the required steps to fully configure memory corruption
2329 * detection are not taken. We need to ensure TTE.mcde is not
2330 * set on M7 processor. Compute the value of cacheability
2331 * flag for use later taking this into consideration.
2332 */
2333 switch (sun4v_chip_type) {
2334 case SUN4V_CHIP_SPARC_M7:
2335 case SUN4V_CHIP_SPARC_M8:
2336 case SUN4V_CHIP_SPARC_SN:
2337 page_cache4v_flag = _PAGE_CP_4V;
2338 break;
2339 default:
2340 page_cache4v_flag = _PAGE_CACHE_4V;
2341 break;
2342 }
2343
2344 if (tlb_type == hypervisor)
2345 sun4v_pgprot_init();
2346 else
2347 sun4u_pgprot_init();
2348
2349 if (tlb_type == cheetah_plus ||
2350 tlb_type == hypervisor) {
2351 tsb_phys_patch();
2352 ktsb_phys_patch();
2353 }
2354
2355 if (tlb_type == hypervisor)
2356 sun4v_patch_tlb_handlers();
2357
2358 /* Find available physical memory...
2359 *
2360 * Read it twice in order to work around a bug in openfirmware.
2361 * The call to grab this table itself can cause openfirmware to
2362 * allocate memory, which in turn can take away some space from
2363 * the list of available memory. Reading it twice makes sure
2364 * we really do get the final value.
2365 */
2366 read_obp_translations();
2367 read_obp_memory("reg", &pall[0], &pall_ents);
2368 read_obp_memory("available", &pavail[0], &pavail_ents);
2369 read_obp_memory("available", &pavail[0], &pavail_ents);
2370
2371 phys_base = 0xffffffffffffffffUL;
2372 for (i = 0; i < pavail_ents; i++) {
2373 phys_base = min(phys_base, pavail[i].phys_addr);
2374 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2375 }
2376
2377 memblock_reserve(kern_base, kern_size);
2378
2379 find_ramdisk(phys_base);
2380
2381 if (cmdline_memory_size)
2382 reduce_memory(cmdline_memory_size);
2383
2384 memblock_allow_resize();
2385 memblock_dump_all();
2386
2387 set_bit(0, mmu_context_bmap);
2388
2389 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2390
2391 real_end = (unsigned long)_end;
2392 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2393 printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2394 num_kernel_image_mappings);
2395
2396 /* Set kernel pgd to upper alias so physical page computations
2397 * work.
2398 */
2399 init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2400
2401 memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2402
2403 inherit_prom_mappings();
2404
2405 /* Ok, we can use our TLB miss and window trap handlers safely. */
2406 setup_tba();
2407
2408 __flush_tlb_all();
2409
2410 prom_build_devicetree();
2411 of_populate_present_mask();
2412#ifndef CONFIG_SMP
2413 of_fill_in_cpu_data();
2414#endif
2415
2416 if (tlb_type == hypervisor) {
2417 sun4v_mdesc_init();
2418 mdesc_populate_present_mask(cpu_all_mask);
2419#ifndef CONFIG_SMP
2420 mdesc_fill_in_cpu_data(cpu_all_mask);
2421#endif
2422 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2423
2424 sun4v_linear_pte_xor_finalize();
2425
2426 sun4v_ktsb_init();
2427 sun4v_ktsb_register();
2428 } else {
2429 unsigned long impl, ver;
2430
2431 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2432 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2433
2434 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2435 impl = ((ver >> 32) & 0xffff);
2436 if (impl == PANTHER_IMPL)
2437 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2438 HV_PGSZ_MASK_256MB);
2439
2440 sun4u_linear_pte_xor_finalize();
2441 }
2442
2443 /* Flush the TLBs and the 4M TSB so that the updated linear
2444 * pte XOR settings are realized for all mappings.
2445 */
2446 __flush_tlb_all();
2447#ifndef CONFIG_DEBUG_PAGEALLOC
2448 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2449#endif
2450 __flush_tlb_all();
2451
2452 /* Setup bootmem... */
2453 last_valid_pfn = end_pfn = bootmem_init(phys_base);
2454
2455 kernel_physical_mapping_init();
2456
2457 {
2458 unsigned long max_zone_pfns[MAX_NR_ZONES];
2459
2460 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2461
2462 max_zone_pfns[ZONE_NORMAL] = end_pfn;
2463
2464 free_area_init(max_zone_pfns);
2465 }
2466
2467 printk("Booting Linux...\n");
2468}
2469
2470int page_in_phys_avail(unsigned long paddr)
2471{
2472 int i;
2473
2474 paddr &= PAGE_MASK;
2475
2476 for (i = 0; i < pavail_ents; i++) {
2477 unsigned long start, end;
2478
2479 start = pavail[i].phys_addr;
2480 end = start + pavail[i].reg_size;
2481
2482 if (paddr >= start && paddr < end)
2483 return 1;
2484 }
2485 if (paddr >= kern_base && paddr < (kern_base + kern_size))
2486 return 1;
2487#ifdef CONFIG_BLK_DEV_INITRD
2488 if (paddr >= __pa(initrd_start) &&
2489 paddr < __pa(PAGE_ALIGN(initrd_end)))
2490 return 1;
2491#endif
2492
2493 return 0;
2494}
2495
2496static void __init register_page_bootmem_info(void)
2497{
2498#ifdef CONFIG_NUMA
2499 int i;
2500
2501 for_each_online_node(i)
2502 if (NODE_DATA(i)->node_spanned_pages)
2503 register_page_bootmem_info_node(NODE_DATA(i));
2504#endif
2505}
2506void __init mem_init(void)
2507{
2508 high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2509
2510 memblock_free_all();
2511
2512 /*
2513 * Must be done after boot memory is put on freelist, because here we
2514 * might set fields in deferred struct pages that have not yet been
2515 * initialized, and memblock_free_all() initializes all the reserved
2516 * deferred pages for us.
2517 */
2518 register_page_bootmem_info();
2519
2520 /*
2521 * Set up the zero page, mark it reserved, so that page count
2522 * is not manipulated when freeing the page from user ptes.
2523 */
2524 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2525 if (mem_map_zero == NULL) {
2526 prom_printf("paging_init: Cannot alloc zero page.\n");
2527 prom_halt();
2528 }
2529 mark_page_reserved(mem_map_zero);
2530
2531
2532 if (tlb_type == cheetah || tlb_type == cheetah_plus)
2533 cheetah_ecache_flush_init();
2534}
2535
2536void free_initmem(void)
2537{
2538 unsigned long addr, initend;
2539 int do_free = 1;
2540
2541 /* If the physical memory maps were trimmed by kernel command
2542 * line options, don't even try freeing this initmem stuff up.
2543 * The kernel image could have been in the trimmed out region
2544 * and if so the freeing below will free invalid page structs.
2545 */
2546 if (cmdline_memory_size)
2547 do_free = 0;
2548
2549 /*
2550 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2551 */
2552 addr = PAGE_ALIGN((unsigned long)(__init_begin));
2553 initend = (unsigned long)(__init_end) & PAGE_MASK;
2554 for (; addr < initend; addr += PAGE_SIZE) {
2555 unsigned long page;
2556
2557 page = (addr +
2558 ((unsigned long) __va(kern_base)) -
2559 ((unsigned long) KERNBASE));
2560 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2561
2562 if (do_free)
2563 free_reserved_page(virt_to_page(page));
2564 }
2565}
2566
2567pgprot_t PAGE_KERNEL __read_mostly;
2568EXPORT_SYMBOL(PAGE_KERNEL);
2569
2570pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2571pgprot_t PAGE_COPY __read_mostly;
2572
2573pgprot_t PAGE_SHARED __read_mostly;
2574EXPORT_SYMBOL(PAGE_SHARED);
2575
2576unsigned long pg_iobits __read_mostly;
2577
2578unsigned long _PAGE_IE __read_mostly;
2579EXPORT_SYMBOL(_PAGE_IE);
2580
2581unsigned long _PAGE_E __read_mostly;
2582EXPORT_SYMBOL(_PAGE_E);
2583
2584unsigned long _PAGE_CACHE __read_mostly;
2585EXPORT_SYMBOL(_PAGE_CACHE);
2586
2587#ifdef CONFIG_SPARSEMEM_VMEMMAP
2588int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2589 int node, struct vmem_altmap *altmap)
2590{
2591 unsigned long pte_base;
2592
2593 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2594 _PAGE_CP_4U | _PAGE_CV_4U |
2595 _PAGE_P_4U | _PAGE_W_4U);
2596 if (tlb_type == hypervisor)
2597 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2598 page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2599
2600 pte_base |= _PAGE_PMD_HUGE;
2601
2602 vstart = vstart & PMD_MASK;
2603 vend = ALIGN(vend, PMD_SIZE);
2604 for (; vstart < vend; vstart += PMD_SIZE) {
2605 pgd_t *pgd = vmemmap_pgd_populate(vstart, node);
2606 unsigned long pte;
2607 p4d_t *p4d;
2608 pud_t *pud;
2609 pmd_t *pmd;
2610
2611 if (!pgd)
2612 return -ENOMEM;
2613
2614 p4d = vmemmap_p4d_populate(pgd, vstart, node);
2615 if (!p4d)
2616 return -ENOMEM;
2617
2618 pud = vmemmap_pud_populate(p4d, vstart, node);
2619 if (!pud)
2620 return -ENOMEM;
2621
2622 pmd = pmd_offset(pud, vstart);
2623 pte = pmd_val(*pmd);
2624 if (!(pte & _PAGE_VALID)) {
2625 void *block = vmemmap_alloc_block(PMD_SIZE, node);
2626
2627 if (!block)
2628 return -ENOMEM;
2629
2630 pmd_val(*pmd) = pte_base | __pa(block);
2631 }
2632 }
2633
2634 return 0;
2635}
2636#endif /* CONFIG_SPARSEMEM_VMEMMAP */
2637
2638/* These are actually filled in at boot time by sun4{u,v}_pgprot_init() */
2639static pgprot_t protection_map[16] __ro_after_init;
2640
2641static void prot_init_common(unsigned long page_none,
2642 unsigned long page_shared,
2643 unsigned long page_copy,
2644 unsigned long page_readonly,
2645 unsigned long page_exec_bit)
2646{
2647 PAGE_COPY = __pgprot(page_copy);
2648 PAGE_SHARED = __pgprot(page_shared);
2649
2650 protection_map[0x0] = __pgprot(page_none);
2651 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2652 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2653 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2654 protection_map[0x4] = __pgprot(page_readonly);
2655 protection_map[0x5] = __pgprot(page_readonly);
2656 protection_map[0x6] = __pgprot(page_copy);
2657 protection_map[0x7] = __pgprot(page_copy);
2658 protection_map[0x8] = __pgprot(page_none);
2659 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2660 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2661 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2662 protection_map[0xc] = __pgprot(page_readonly);
2663 protection_map[0xd] = __pgprot(page_readonly);
2664 protection_map[0xe] = __pgprot(page_shared);
2665 protection_map[0xf] = __pgprot(page_shared);
2666}
2667
2668static void __init sun4u_pgprot_init(void)
2669{
2670 unsigned long page_none, page_shared, page_copy, page_readonly;
2671 unsigned long page_exec_bit;
2672 int i;
2673
2674 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2675 _PAGE_CACHE_4U | _PAGE_P_4U |
2676 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2677 _PAGE_EXEC_4U);
2678 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2679 _PAGE_CACHE_4U | _PAGE_P_4U |
2680 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2681 _PAGE_EXEC_4U | _PAGE_L_4U);
2682
2683 _PAGE_IE = _PAGE_IE_4U;
2684 _PAGE_E = _PAGE_E_4U;
2685 _PAGE_CACHE = _PAGE_CACHE_4U;
2686
2687 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2688 __ACCESS_BITS_4U | _PAGE_E_4U);
2689
2690#ifdef CONFIG_DEBUG_PAGEALLOC
2691 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2692#else
2693 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2694 PAGE_OFFSET;
2695#endif
2696 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2697 _PAGE_P_4U | _PAGE_W_4U);
2698
2699 for (i = 1; i < 4; i++)
2700 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2701
2702 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2703 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2704 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2705
2706
2707 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2708 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2709 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2710 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2711 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2712 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2713 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2714
2715 page_exec_bit = _PAGE_EXEC_4U;
2716
2717 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2718 page_exec_bit);
2719}
2720
2721static void __init sun4v_pgprot_init(void)
2722{
2723 unsigned long page_none, page_shared, page_copy, page_readonly;
2724 unsigned long page_exec_bit;
2725 int i;
2726
2727 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2728 page_cache4v_flag | _PAGE_P_4V |
2729 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2730 _PAGE_EXEC_4V);
2731 PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2732
2733 _PAGE_IE = _PAGE_IE_4V;
2734 _PAGE_E = _PAGE_E_4V;
2735 _PAGE_CACHE = page_cache4v_flag;
2736
2737#ifdef CONFIG_DEBUG_PAGEALLOC
2738 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2739#else
2740 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2741 PAGE_OFFSET;
2742#endif
2743 kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2744 _PAGE_W_4V);
2745
2746 for (i = 1; i < 4; i++)
2747 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2748
2749 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2750 __ACCESS_BITS_4V | _PAGE_E_4V);
2751
2752 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2753 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2754 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2755 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2756
2757 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2758 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2759 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2760 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2761 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2762 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2763 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2764
2765 page_exec_bit = _PAGE_EXEC_4V;
2766
2767 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2768 page_exec_bit);
2769}
2770
2771unsigned long pte_sz_bits(unsigned long sz)
2772{
2773 if (tlb_type == hypervisor) {
2774 switch (sz) {
2775 case 8 * 1024:
2776 default:
2777 return _PAGE_SZ8K_4V;
2778 case 64 * 1024:
2779 return _PAGE_SZ64K_4V;
2780 case 512 * 1024:
2781 return _PAGE_SZ512K_4V;
2782 case 4 * 1024 * 1024:
2783 return _PAGE_SZ4MB_4V;
2784 }
2785 } else {
2786 switch (sz) {
2787 case 8 * 1024:
2788 default:
2789 return _PAGE_SZ8K_4U;
2790 case 64 * 1024:
2791 return _PAGE_SZ64K_4U;
2792 case 512 * 1024:
2793 return _PAGE_SZ512K_4U;
2794 case 4 * 1024 * 1024:
2795 return _PAGE_SZ4MB_4U;
2796 }
2797 }
2798}
2799
2800pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2801{
2802 pte_t pte;
2803
2804 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot));
2805 pte_val(pte) |= (((unsigned long)space) << 32);
2806 pte_val(pte) |= pte_sz_bits(page_size);
2807
2808 return pte;
2809}
2810
2811static unsigned long kern_large_tte(unsigned long paddr)
2812{
2813 unsigned long val;
2814
2815 val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2816 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2817 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2818 if (tlb_type == hypervisor)
2819 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2820 page_cache4v_flag | _PAGE_P_4V |
2821 _PAGE_EXEC_4V | _PAGE_W_4V);
2822
2823 return val | paddr;
2824}
2825
2826/* If not locked, zap it. */
2827void __flush_tlb_all(void)
2828{
2829 unsigned long pstate;
2830 int i;
2831
2832 __asm__ __volatile__("flushw\n\t"
2833 "rdpr %%pstate, %0\n\t"
2834 "wrpr %0, %1, %%pstate"
2835 : "=r" (pstate)
2836 : "i" (PSTATE_IE));
2837 if (tlb_type == hypervisor) {
2838 sun4v_mmu_demap_all();
2839 } else if (tlb_type == spitfire) {
2840 for (i = 0; i < 64; i++) {
2841 /* Spitfire Errata #32 workaround */
2842 /* NOTE: Always runs on spitfire, so no
2843 * cheetah+ page size encodings.
2844 */
2845 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2846 "flush %%g6"
2847 : /* No outputs */
2848 : "r" (0),
2849 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2850
2851 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2852 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2853 "membar #Sync"
2854 : /* no outputs */
2855 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2856 spitfire_put_dtlb_data(i, 0x0UL);
2857 }
2858
2859 /* Spitfire Errata #32 workaround */
2860 /* NOTE: Always runs on spitfire, so no
2861 * cheetah+ page size encodings.
2862 */
2863 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2864 "flush %%g6"
2865 : /* No outputs */
2866 : "r" (0),
2867 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2868
2869 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2870 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2871 "membar #Sync"
2872 : /* no outputs */
2873 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2874 spitfire_put_itlb_data(i, 0x0UL);
2875 }
2876 }
2877 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2878 cheetah_flush_dtlb_all();
2879 cheetah_flush_itlb_all();
2880 }
2881 __asm__ __volatile__("wrpr %0, 0, %%pstate"
2882 : : "r" (pstate));
2883}
2884
2885pte_t *pte_alloc_one_kernel(struct mm_struct *mm)
2886{
2887 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2888 pte_t *pte = NULL;
2889
2890 if (page)
2891 pte = (pte_t *) page_address(page);
2892
2893 return pte;
2894}
2895
2896pgtable_t pte_alloc_one(struct mm_struct *mm)
2897{
2898 struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL | __GFP_ZERO, 0);
2899
2900 if (!ptdesc)
2901 return NULL;
2902 if (!pagetable_pte_ctor(ptdesc)) {
2903 pagetable_free(ptdesc);
2904 return NULL;
2905 }
2906 return ptdesc_address(ptdesc);
2907}
2908
2909void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2910{
2911 free_page((unsigned long)pte);
2912}
2913
2914static void __pte_free(pgtable_t pte)
2915{
2916 struct ptdesc *ptdesc = virt_to_ptdesc(pte);
2917
2918 pagetable_pte_dtor(ptdesc);
2919 pagetable_free(ptdesc);
2920}
2921
2922void pte_free(struct mm_struct *mm, pgtable_t pte)
2923{
2924 __pte_free(pte);
2925}
2926
2927void pgtable_free(void *table, bool is_page)
2928{
2929 if (is_page)
2930 __pte_free(table);
2931 else
2932 kmem_cache_free(pgtable_cache, table);
2933}
2934
2935#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2936static void pte_free_now(struct rcu_head *head)
2937{
2938 struct page *page;
2939
2940 page = container_of(head, struct page, rcu_head);
2941 __pte_free((pgtable_t)page_address(page));
2942}
2943
2944void pte_free_defer(struct mm_struct *mm, pgtable_t pgtable)
2945{
2946 struct page *page;
2947
2948 page = virt_to_page(pgtable);
2949 call_rcu(&page->rcu_head, pte_free_now);
2950}
2951
2952void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2953 pmd_t *pmd)
2954{
2955 unsigned long pte, flags;
2956 struct mm_struct *mm;
2957 pmd_t entry = *pmd;
2958
2959 if (!pmd_leaf(entry) || !pmd_young(entry))
2960 return;
2961
2962 pte = pmd_val(entry);
2963
2964 /* Don't insert a non-valid PMD into the TSB, we'll deadlock. */
2965 if (!(pte & _PAGE_VALID))
2966 return;
2967
2968 /* We are fabricating 8MB pages using 4MB real hw pages. */
2969 pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2970
2971 mm = vma->vm_mm;
2972
2973 spin_lock_irqsave(&mm->context.lock, flags);
2974
2975 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2976 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2977 addr, pte);
2978
2979 spin_unlock_irqrestore(&mm->context.lock, flags);
2980}
2981#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2982
2983#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2984static void context_reload(void *__data)
2985{
2986 struct mm_struct *mm = __data;
2987
2988 if (mm == current->mm)
2989 load_secondary_context(mm);
2990}
2991
2992void hugetlb_setup(struct pt_regs *regs)
2993{
2994 struct mm_struct *mm = current->mm;
2995 struct tsb_config *tp;
2996
2997 if (faulthandler_disabled() || !mm) {
2998 const struct exception_table_entry *entry;
2999
3000 entry = search_exception_tables(regs->tpc);
3001 if (entry) {
3002 regs->tpc = entry->fixup;
3003 regs->tnpc = regs->tpc + 4;
3004 return;
3005 }
3006 pr_alert("Unexpected HugeTLB setup in atomic context.\n");
3007 die_if_kernel("HugeTSB in atomic", regs);
3008 }
3009
3010 tp = &mm->context.tsb_block[MM_TSB_HUGE];
3011 if (likely(tp->tsb == NULL))
3012 tsb_grow(mm, MM_TSB_HUGE, 0);
3013
3014 tsb_context_switch(mm);
3015 smp_tsb_sync(mm);
3016
3017 /* On UltraSPARC-III+ and later, configure the second half of
3018 * the Data-TLB for huge pages.
3019 */
3020 if (tlb_type == cheetah_plus) {
3021 bool need_context_reload = false;
3022 unsigned long ctx;
3023
3024 spin_lock_irq(&ctx_alloc_lock);
3025 ctx = mm->context.sparc64_ctx_val;
3026 ctx &= ~CTX_PGSZ_MASK;
3027 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
3028 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
3029
3030 if (ctx != mm->context.sparc64_ctx_val) {
3031 /* When changing the page size fields, we
3032 * must perform a context flush so that no
3033 * stale entries match. This flush must
3034 * occur with the original context register
3035 * settings.
3036 */
3037 do_flush_tlb_mm(mm);
3038
3039 /* Reload the context register of all processors
3040 * also executing in this address space.
3041 */
3042 mm->context.sparc64_ctx_val = ctx;
3043 need_context_reload = true;
3044 }
3045 spin_unlock_irq(&ctx_alloc_lock);
3046
3047 if (need_context_reload)
3048 on_each_cpu(context_reload, mm, 0);
3049 }
3050}
3051#endif
3052
3053static struct resource code_resource = {
3054 .name = "Kernel code",
3055 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3056};
3057
3058static struct resource data_resource = {
3059 .name = "Kernel data",
3060 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3061};
3062
3063static struct resource bss_resource = {
3064 .name = "Kernel bss",
3065 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3066};
3067
3068static inline resource_size_t compute_kern_paddr(void *addr)
3069{
3070 return (resource_size_t) (addr - KERNBASE + kern_base);
3071}
3072
3073static void __init kernel_lds_init(void)
3074{
3075 code_resource.start = compute_kern_paddr(_text);
3076 code_resource.end = compute_kern_paddr(_etext - 1);
3077 data_resource.start = compute_kern_paddr(_etext);
3078 data_resource.end = compute_kern_paddr(_edata - 1);
3079 bss_resource.start = compute_kern_paddr(__bss_start);
3080 bss_resource.end = compute_kern_paddr(_end - 1);
3081}
3082
3083static int __init report_memory(void)
3084{
3085 int i;
3086 struct resource *res;
3087
3088 kernel_lds_init();
3089
3090 for (i = 0; i < pavail_ents; i++) {
3091 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
3092
3093 if (!res) {
3094 pr_warn("Failed to allocate source.\n");
3095 break;
3096 }
3097
3098 res->name = "System RAM";
3099 res->start = pavail[i].phys_addr;
3100 res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
3101 res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
3102
3103 if (insert_resource(&iomem_resource, res) < 0) {
3104 pr_warn("Resource insertion failed.\n");
3105 break;
3106 }
3107
3108 insert_resource(res, &code_resource);
3109 insert_resource(res, &data_resource);
3110 insert_resource(res, &bss_resource);
3111 }
3112
3113 return 0;
3114}
3115arch_initcall(report_memory);
3116
3117#ifdef CONFIG_SMP
3118#define do_flush_tlb_kernel_range smp_flush_tlb_kernel_range
3119#else
3120#define do_flush_tlb_kernel_range __flush_tlb_kernel_range
3121#endif
3122
3123void flush_tlb_kernel_range(unsigned long start, unsigned long end)
3124{
3125 if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
3126 if (start < LOW_OBP_ADDRESS) {
3127 flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
3128 do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
3129 }
3130 if (end > HI_OBP_ADDRESS) {
3131 flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
3132 do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
3133 }
3134 } else {
3135 flush_tsb_kernel_range(start, end);
3136 do_flush_tlb_kernel_range(start, end);
3137 }
3138}
3139
3140void copy_user_highpage(struct page *to, struct page *from,
3141 unsigned long vaddr, struct vm_area_struct *vma)
3142{
3143 char *vfrom, *vto;
3144
3145 vfrom = kmap_atomic(from);
3146 vto = kmap_atomic(to);
3147 copy_user_page(vto, vfrom, vaddr, to);
3148 kunmap_atomic(vto);
3149 kunmap_atomic(vfrom);
3150
3151 /* If this page has ADI enabled, copy over any ADI tags
3152 * as well
3153 */
3154 if (vma->vm_flags & VM_SPARC_ADI) {
3155 unsigned long pfrom, pto, i, adi_tag;
3156
3157 pfrom = page_to_phys(from);
3158 pto = page_to_phys(to);
3159
3160 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3161 asm volatile("ldxa [%1] %2, %0\n\t"
3162 : "=r" (adi_tag)
3163 : "r" (i), "i" (ASI_MCD_REAL));
3164 asm volatile("stxa %0, [%1] %2\n\t"
3165 :
3166 : "r" (adi_tag), "r" (pto),
3167 "i" (ASI_MCD_REAL));
3168 pto += adi_blksize();
3169 }
3170 asm volatile("membar #Sync\n\t");
3171 }
3172}
3173EXPORT_SYMBOL(copy_user_highpage);
3174
3175void copy_highpage(struct page *to, struct page *from)
3176{
3177 char *vfrom, *vto;
3178
3179 vfrom = kmap_atomic(from);
3180 vto = kmap_atomic(to);
3181 copy_page(vto, vfrom);
3182 kunmap_atomic(vto);
3183 kunmap_atomic(vfrom);
3184
3185 /* If this platform is ADI enabled, copy any ADI tags
3186 * as well
3187 */
3188 if (adi_capable()) {
3189 unsigned long pfrom, pto, i, adi_tag;
3190
3191 pfrom = page_to_phys(from);
3192 pto = page_to_phys(to);
3193
3194 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3195 asm volatile("ldxa [%1] %2, %0\n\t"
3196 : "=r" (adi_tag)
3197 : "r" (i), "i" (ASI_MCD_REAL));
3198 asm volatile("stxa %0, [%1] %2\n\t"
3199 :
3200 : "r" (adi_tag), "r" (pto),
3201 "i" (ASI_MCD_REAL));
3202 pto += adi_blksize();
3203 }
3204 asm volatile("membar #Sync\n\t");
3205 }
3206}
3207EXPORT_SYMBOL(copy_highpage);
3208
3209pgprot_t vm_get_page_prot(unsigned long vm_flags)
3210{
3211 unsigned long prot = pgprot_val(protection_map[vm_flags &
3212 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]);
3213
3214 if (vm_flags & VM_SPARC_ADI)
3215 prot |= _PAGE_MCD_4V;
3216
3217 return __pgprot(prot);
3218}
3219EXPORT_SYMBOL(vm_get_page_prot);