?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
drm/panthor: Don't add write fences to the shared BOs
[drm/drm-misc.git] / arch / sparc / mm / init_64.c
blob21f8cbbd0581c77b9ceb198b712c7eaa84b2483b
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
56 unsigned long kern_linear_pte_xor[4] __read_mostly;
57 static 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 */
85 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
86 #endif
87 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
88
89 static unsigned long cpu_pgsz_mask;
90
91 #define MAX_BANKS 1024
92
93 static struct linux_prom64_registers pavail[MAX_BANKS];
94 static int pavail_ents;
95
96 u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
97
98 static 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
109 static 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(&regs[i], &regs[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. */
172 unsigned long kern_base __read_mostly;
173 unsigned long kern_size __read_mostly;
174
175 /* Initial ramdisk setup */
176 extern unsigned long sparc_ramdisk_image64;
177 extern unsigned int sparc_ramdisk_image;
178 extern unsigned int sparc_ramdisk_size;
179
180 struct page *mem_map_zero __read_mostly;
181 EXPORT_SYMBOL(mem_map_zero);
182
183 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
184
185 unsigned long sparc64_kern_pri_context __read_mostly;
186 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
187 unsigned long sparc64_kern_sec_context __read_mostly;
188
189 int num_kernel_image_mappings;
190
191 #ifdef CONFIG_DEBUG_DCFLUSH
192 atomic_t dcpage_flushes = ATOMIC_INIT(0);
193 #ifdef CONFIG_SMP
194 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
195 #endif
196 #endif
197
198 inline 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
229 static 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
250 static 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
274 static 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
284 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
285
286 static 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 */
317 static 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
334 static 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
344 arch_initcall(hugetlbpage_init);
345
346 static 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
358 bool __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
397 void 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
463 void 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
501 out:
502 put_cpu();
503 }
504 EXPORT_SYMBOL(flush_dcache_folio);
505
506 void __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 }
529 EXPORT_SYMBOL(flush_icache_range);
530
531 void 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
571 struct linux_prom_translation prom_trans[512] __read_mostly;
572 unsigned int prom_trans_ents __read_mostly;
573
574 unsigned 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 */
580 static inline int in_obp_range(unsigned long vaddr)
581 {
582 return (vaddr >= LOW_OBP_ADDRESS &&
583 vaddr < HI_OBP_ADDRESS);
584 }
585
586 static 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[]'. */
598 static 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
664 static 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
677 static unsigned long kern_large_tte(unsigned long paddr);
678
679 static 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
716 static 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
724 void 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
733 void __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 }
756 EXPORT_SYMBOL(__flush_dcache_range);
757
758 /* get_new_mmu_context() uses "cache + 1". */
759 DEFINE_SPINLOCK(ctx_alloc_lock);
760 unsigned 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)
763 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
764 DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0};
765
766 static 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 */
823 void 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);
829 retry:
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;
849 out:
850 spin_unlock(&ctx_alloc_lock);
851 }
852
853 static int numa_enabled = 1;
854 static int numa_debug;
855
856 static 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 }
869 early_param("numa", early_numa);
870
871 #define numadbg(f, a...) \
872 do { if (numa_debug) \
873 printk(KERN_INFO f, ## a); \
874 } while (0)
875
876 static 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
915 struct node_mem_mask {
916 unsigned long mask;
917 unsigned long match;
918 };
919 static struct node_mem_mask node_masks[MAX_NUMNODES];
920 static int num_node_masks;
921
922 #ifdef CONFIG_NUMA
923
924 struct mdesc_mlgroup {
925 u64 node;
926 u64 latency;
927 u64 match;
928 u64 mask;
929 };
930
931 static struct mdesc_mlgroup *mlgroups;
932 static int num_mlgroups;
933
934 int numa_cpu_lookup_table[NR_CPUS];
935 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
936
937 struct mdesc_mblock {
938 u64 base;
939 u64 size;
940 u64 offset; /* RA-to-PA */
941 };
942 static struct mdesc_mblock *mblocks;
943 static int num_mblocks;
944
945 static 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
962 static 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
993 static 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
1064 done:
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 */
1074 static 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
1092 static 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
1114 EXPORT_SYMBOL(numa_cpu_lookup_table);
1115 EXPORT_SYMBOL(numa_cpumask_lookup_table);
1116
1117 static 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
1134 static 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
1164 int 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
1202 static void __init add_node_ranges(void)
1203 {
1204 phys_addr_t start, end;
1205 unsigned long prev_max;
1206 u64 i;
1207
1208 memblock_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
1231 static 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
1272 static 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
1318 static 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
1338 static 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
1350 int __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 }
1359 EXPORT_SYMBOL(__node_distance);
1360
1361 static 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
1374 static 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
1393 static 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
1431 static 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
1453 static 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;
1506 out:
1507 mdesc_release(md);
1508 return err;
1509 }
1510
1511 static 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
1539 static 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
1552 static 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
1577 static int bootmem_init_numa(void)
1578 {
1579 return -1;
1580 }
1581
1582 #endif
1583
1584 static 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
1602 static 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
1623 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1624 static int pall_ents __initdata;
1625
1626 static unsigned long max_phys_bits = 40;
1627
1628 bool 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
1678 static 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
1708 static 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
1717 static 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
1756 static 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
1765 static 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
1858 err_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
1864 static 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
1882 extern unsigned int kvmap_linear_patch[1];
1883
1884 static 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
1914 void __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
1933 unsigned 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
1945 unsigned long PAGE_OFFSET;
1946 EXPORT_SYMBOL(PAGE_OFFSET);
1947
1948 unsigned long VMALLOC_END = 0x0000010000000000UL;
1949 EXPORT_SYMBOL(VMALLOC_END);
1950
1951 unsigned long sparc64_va_hole_top = 0xfffff80000000000UL;
1952 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1953
1954 static 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
2030 static 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
2071 static 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
2086 static 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
2112 static 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
2133 static 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
2188 void 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
2202 static 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
2211 static 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
2260 static unsigned long last_valid_pfn;
2261
2262 static void sun4u_pgprot_init(void);
2263 static 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 */
2276 static 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
2282 void __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
2470 int 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
2496 static 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 }
2506 void __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
2536 void 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
2567 pgprot_t PAGE_KERNEL __read_mostly;
2568 EXPORT_SYMBOL(PAGE_KERNEL);
2569
2570 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2571 pgprot_t PAGE_COPY __read_mostly;
2572
2573 pgprot_t PAGE_SHARED __read_mostly;
2574 EXPORT_SYMBOL(PAGE_SHARED);
2575
2576 unsigned long pg_iobits __read_mostly;
2577
2578 unsigned long _PAGE_IE __read_mostly;
2579 EXPORT_SYMBOL(_PAGE_IE);
2580
2581 unsigned long _PAGE_E __read_mostly;
2582 EXPORT_SYMBOL(_PAGE_E);
2583
2584 unsigned long _PAGE_CACHE __read_mostly;
2585 EXPORT_SYMBOL(_PAGE_CACHE);
2586
2587 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2588 int __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() */
2639 static pgprot_t protection_map[16] __ro_after_init;
2640
2641 static 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
2668 static 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
2721 static 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
2771 unsigned 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
2800 pte_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
2811 static 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. */
2827 void __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
2885 pte_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
2896 pgtable_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
2909 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2910 {
2911 free_page((unsigned long)pte);
2912 }
2913
2914 static 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
2922 void pte_free(struct mm_struct *mm, pgtable_t pte)
2923 {
2924 __pte_free(pte);
2925 }
2926
2927 void 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
2936 static 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
2944 void 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
2952 void 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)
2984 static 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
2992 void 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
3053 static struct resource code_resource = {
3054 .name = "Kernel code",
3055 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3056 };
3057
3058 static struct resource data_resource = {
3059 .name = "Kernel data",
3060 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3061 };
3062
3063 static struct resource bss_resource = {
3064 .name = "Kernel bss",
3065 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3066 };
3067
3068 static inline resource_size_t compute_kern_paddr(void *addr)
3069 {
3070 return (resource_size_t) (addr - KERNBASE + kern_base);
3071 }
3072
3073 static 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
3083 static 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 }
3115 arch_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
3123 void 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
3140 void 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 }
3173 EXPORT_SYMBOL(copy_user_highpage);
3174
3175 void 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 }
3207 EXPORT_SYMBOL(copy_highpage);
3208
3209 pgprot_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 }
3219 EXPORT_SYMBOL(vm_get_page_prot);
Kernel DRM miscellaneous fixes and cross-tree changes