?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
mm: fix exec activate_mm vs TLB shootdown and lazy tlb switching race
[linux/fpc-iii.git] / arch / sparc / mm / init_64.c
blob76977296dc9c6410820ce50270725bf777b4adfc
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
57 unsigned long kern_linear_pte_xor[4] __read_mostly;
58 static 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 */
86 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
87 #endif
88 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
89
90 static unsigned long cpu_pgsz_mask;
91
92 #define MAX_BANKS 1024
93
94 static struct linux_prom64_registers pavail[MAX_BANKS];
95 static int pavail_ents;
96
97 u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
98
99 static 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
110 static 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(&regs[i], &regs[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. */
173 unsigned long kern_base __read_mostly;
174 unsigned long kern_size __read_mostly;
175
176 /* Initial ramdisk setup */
177 extern unsigned long sparc_ramdisk_image64;
178 extern unsigned int sparc_ramdisk_image;
179 extern unsigned int sparc_ramdisk_size;
180
181 struct page *mem_map_zero __read_mostly;
182 EXPORT_SYMBOL(mem_map_zero);
183
184 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
185
186 unsigned long sparc64_kern_pri_context __read_mostly;
187 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
188 unsigned long sparc64_kern_sec_context __read_mostly;
189
190 int num_kernel_image_mappings;
191
192 #ifdef CONFIG_DEBUG_DCFLUSH
193 atomic_t dcpage_flushes = ATOMIC_INIT(0);
194 #ifdef CONFIG_SMP
195 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
196 #endif
197 #endif
198
199 inline 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(page) != NULL));
210 #else
211 if (page_mapping(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
225 static 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
246 static 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
270 static 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
280 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
281
282 static 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 */
312 static 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
329 static 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
340 static 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
350 arch_initcall(hugetlbpage_init);
351
352 static 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
364 static 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
411 out:
412 return rc;
413 }
414 __setup("hugepagesz=", setup_hugepagesz);
415 #endif /* CONFIG_HUGETLB_PAGE */
416
417 void 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
476 void 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(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
513 out:
514 put_cpu();
515 }
516 EXPORT_SYMBOL(flush_dcache_page);
517
518 void __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 }
544 EXPORT_SYMBOL(flush_icache_range);
545
546 void 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
586 struct linux_prom_translation prom_trans[512] __read_mostly;
587 unsigned int prom_trans_ents __read_mostly;
588
589 unsigned 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 */
595 static inline int in_obp_range(unsigned long vaddr)
596 {
597 return (vaddr >= LOW_OBP_ADDRESS &&
598 vaddr < HI_OBP_ADDRESS);
599 }
600
601 static 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[]'. */
613 static 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
679 static 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
692 static unsigned long kern_large_tte(unsigned long paddr);
693
694 static 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
731 static 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
739 void prom_world(int enter)
740 {
741 if (!enter)
742 set_fs(get_fs());
743
744 __asm__ __volatile__("flushw");
745 }
746
747 void __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 }
770 EXPORT_SYMBOL(__flush_dcache_range);
771
772 /* get_new_mmu_context() uses "cache + 1". */
773 DEFINE_SPINLOCK(ctx_alloc_lock);
774 unsigned 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)
777 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
778 DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0};
779
780 static 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 */
837 void 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);
843 retry:
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;
863 out:
864 spin_unlock(&ctx_alloc_lock);
865 }
866
867 static int numa_enabled = 1;
868 static int numa_debug;
869
870 static 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 }
883 early_param("numa", early_numa);
884
885 #define numadbg(f, a...) \
886 do { if (numa_debug) \
887 printk(KERN_INFO f, ## a); \
888 } while (0)
889
890 static 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
929 struct node_mem_mask {
930 unsigned long mask;
931 unsigned long match;
932 };
933 static struct node_mem_mask node_masks[MAX_NUMNODES];
934 static int num_node_masks;
935
936 #ifdef CONFIG_NEED_MULTIPLE_NODES
937
938 struct mdesc_mlgroup {
939 u64 node;
940 u64 latency;
941 u64 match;
942 u64 mask;
943 };
944
945 static struct mdesc_mlgroup *mlgroups;
946 static int num_mlgroups;
947
948 int numa_cpu_lookup_table[NR_CPUS];
949 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
950
951 struct mdesc_mblock {
952 u64 base;
953 u64 size;
954 u64 offset; /* RA-to-PA */
955 };
956 static struct mdesc_mblock *mblocks;
957 static int num_mblocks;
958
959 static 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
976 static 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
1007 static 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
1078 done:
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 */
1088 static 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
1113 static 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
1134 struct pglist_data *node_data[MAX_NUMNODES];
1135
1136 EXPORT_SYMBOL(numa_cpu_lookup_table);
1137 EXPORT_SYMBOL(numa_cpumask_lookup_table);
1138 EXPORT_SYMBOL(node_data);
1139
1140 static 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
1157 static 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
1187 int 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
1225 static void __init add_node_ranges(void)
1226 {
1227 struct memblock_region *reg;
1228 unsigned long prev_max;
1229
1230 memblock_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
1258 static 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
1299 static 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
1345 static 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
1365 static 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
1377 int __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 EXPORT_SYMBOL(__node_distance);
1387
1388 static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1389 {
1390 int i;
1391
1392 for (i = 0; i < MAX_NUMNODES; i++) {
1393 struct node_mem_mask *n = &node_masks[i];
1394
1395 if ((grp->mask == n->mask) && (grp->match == n->match))
1396 break;
1397 }
1398 return i;
1399 }
1400
1401 static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1402 u64 grp, int index)
1403 {
1404 u64 arc;
1405
1406 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1407 int tnode;
1408 u64 target = mdesc_arc_target(md, arc);
1409 struct mdesc_mlgroup *m = find_mlgroup(target);
1410
1411 if (!m)
1412 continue;
1413 tnode = find_best_numa_node_for_mlgroup(m);
1414 if (tnode == MAX_NUMNODES)
1415 continue;
1416 numa_latency[index][tnode] = m->latency;
1417 }
1418 }
1419
1420 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1421 int index)
1422 {
1423 struct mdesc_mlgroup *candidate = NULL;
1424 u64 arc, best_latency = ~(u64)0;
1425 struct node_mem_mask *n;
1426
1427 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1428 u64 target = mdesc_arc_target(md, arc);
1429 struct mdesc_mlgroup *m = find_mlgroup(target);
1430 if (!m)
1431 continue;
1432 if (m->latency < best_latency) {
1433 candidate = m;
1434 best_latency = m->latency;
1435 }
1436 }
1437 if (!candidate)
1438 return -ENOENT;
1439
1440 if (num_node_masks != index) {
1441 printk(KERN_ERR "Inconsistent NUMA state, "
1442 "index[%d] != num_node_masks[%d]\n",
1443 index, num_node_masks);
1444 return -EINVAL;
1445 }
1446
1447 n = &node_masks[num_node_masks++];
1448
1449 n->mask = candidate->mask;
1450 n->match = candidate->match;
1451
1452 numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n",
1453 index, n->mask, n->match, candidate->latency);
1454
1455 return 0;
1456 }
1457
1458 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1459 int index)
1460 {
1461 cpumask_t mask;
1462 int cpu;
1463
1464 numa_parse_mdesc_group_cpus(md, grp, &mask);
1465
1466 for_each_cpu(cpu, &mask)
1467 numa_cpu_lookup_table[cpu] = index;
1468 cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1469
1470 if (numa_debug) {
1471 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1472 for_each_cpu(cpu, &mask)
1473 printk("%d ", cpu);
1474 printk("]\n");
1475 }
1476
1477 return numa_attach_mlgroup(md, grp, index);
1478 }
1479
1480 static int __init numa_parse_mdesc(void)
1481 {
1482 struct mdesc_handle *md = mdesc_grab();
1483 int i, j, err, count;
1484 u64 node;
1485
1486 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1487 if (node == MDESC_NODE_NULL) {
1488 mdesc_release(md);
1489 return -ENOENT;
1490 }
1491
1492 err = grab_mblocks(md);
1493 if (err < 0)
1494 goto out;
1495
1496 err = grab_mlgroups(md);
1497 if (err < 0)
1498 goto out;
1499
1500 count = 0;
1501 mdesc_for_each_node_by_name(md, node, "group") {
1502 err = numa_parse_mdesc_group(md, node, count);
1503 if (err < 0)
1504 break;
1505 count++;
1506 }
1507
1508 count = 0;
1509 mdesc_for_each_node_by_name(md, node, "group") {
1510 find_numa_latencies_for_group(md, node, count);
1511 count++;
1512 }
1513
1514 /* Normalize numa latency matrix according to ACPI SLIT spec. */
1515 for (i = 0; i < MAX_NUMNODES; i++) {
1516 u64 self_latency = numa_latency[i][i];
1517
1518 for (j = 0; j < MAX_NUMNODES; j++) {
1519 numa_latency[i][j] =
1520 (numa_latency[i][j] * LOCAL_DISTANCE) /
1521 self_latency;
1522 }
1523 }
1524
1525 add_node_ranges();
1526
1527 for (i = 0; i < num_node_masks; i++) {
1528 allocate_node_data(i);
1529 node_set_online(i);
1530 }
1531
1532 err = 0;
1533 out:
1534 mdesc_release(md);
1535 return err;
1536 }
1537
1538 static int __init numa_parse_jbus(void)
1539 {
1540 unsigned long cpu, index;
1541
1542 /* NUMA node id is encoded in bits 36 and higher, and there is
1543 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1544 */
1545 index = 0;
1546 for_each_present_cpu(cpu) {
1547 numa_cpu_lookup_table[cpu] = index;
1548 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1549 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1550 node_masks[index].match = cpu << 36UL;
1551
1552 index++;
1553 }
1554 num_node_masks = index;
1555
1556 add_node_ranges();
1557
1558 for (index = 0; index < num_node_masks; index++) {
1559 allocate_node_data(index);
1560 node_set_online(index);
1561 }
1562
1563 return 0;
1564 }
1565
1566 static int __init numa_parse_sun4u(void)
1567 {
1568 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1569 unsigned long ver;
1570
1571 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1572 if ((ver >> 32UL) == __JALAPENO_ID ||
1573 (ver >> 32UL) == __SERRANO_ID)
1574 return numa_parse_jbus();
1575 }
1576 return -1;
1577 }
1578
1579 static int __init bootmem_init_numa(void)
1580 {
1581 int i, j;
1582 int err = -1;
1583
1584 numadbg("bootmem_init_numa()\n");
1585
1586 /* Some sane defaults for numa latency values */
1587 for (i = 0; i < MAX_NUMNODES; i++) {
1588 for (j = 0; j < MAX_NUMNODES; j++)
1589 numa_latency[i][j] = (i == j) ?
1590 LOCAL_DISTANCE : REMOTE_DISTANCE;
1591 }
1592
1593 if (numa_enabled) {
1594 if (tlb_type == hypervisor)
1595 err = numa_parse_mdesc();
1596 else
1597 err = numa_parse_sun4u();
1598 }
1599 return err;
1600 }
1601
1602 #else
1603
1604 static int bootmem_init_numa(void)
1605 {
1606 return -1;
1607 }
1608
1609 #endif
1610
1611 static void __init bootmem_init_nonnuma(void)
1612 {
1613 unsigned long top_of_ram = memblock_end_of_DRAM();
1614 unsigned long total_ram = memblock_phys_mem_size();
1615
1616 numadbg("bootmem_init_nonnuma()\n");
1617
1618 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1619 top_of_ram, total_ram);
1620 printk(KERN_INFO "Memory hole size: %ldMB\n",
1621 (top_of_ram - total_ram) >> 20);
1622
1623 init_node_masks_nonnuma();
1624 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1625 allocate_node_data(0);
1626 node_set_online(0);
1627 }
1628
1629 static unsigned long __init bootmem_init(unsigned long phys_base)
1630 {
1631 unsigned long end_pfn;
1632
1633 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1634 max_pfn = max_low_pfn = end_pfn;
1635 min_low_pfn = (phys_base >> PAGE_SHIFT);
1636
1637 if (bootmem_init_numa() < 0)
1638 bootmem_init_nonnuma();
1639
1640 /* Dump memblock with node info. */
1641 memblock_dump_all();
1642
1643 /* XXX cpu notifier XXX */
1644
1645 sparse_memory_present_with_active_regions(MAX_NUMNODES);
1646 sparse_init();
1647
1648 return end_pfn;
1649 }
1650
1651 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1652 static int pall_ents __initdata;
1653
1654 static unsigned long max_phys_bits = 40;
1655
1656 bool kern_addr_valid(unsigned long addr)
1657 {
1658 pgd_t *pgd;
1659 pud_t *pud;
1660 pmd_t *pmd;
1661 pte_t *pte;
1662
1663 if ((long)addr < 0L) {
1664 unsigned long pa = __pa(addr);
1665
1666 if ((pa >> max_phys_bits) != 0UL)
1667 return false;
1668
1669 return pfn_valid(pa >> PAGE_SHIFT);
1670 }
1671
1672 if (addr >= (unsigned long) KERNBASE &&
1673 addr < (unsigned long)&_end)
1674 return true;
1675
1676 pgd = pgd_offset_k(addr);
1677 if (pgd_none(*pgd))
1678 return 0;
1679
1680 pud = pud_offset(pgd, addr);
1681 if (pud_none(*pud))
1682 return 0;
1683
1684 if (pud_large(*pud))
1685 return pfn_valid(pud_pfn(*pud));
1686
1687 pmd = pmd_offset(pud, addr);
1688 if (pmd_none(*pmd))
1689 return 0;
1690
1691 if (pmd_large(*pmd))
1692 return pfn_valid(pmd_pfn(*pmd));
1693
1694 pte = pte_offset_kernel(pmd, addr);
1695 if (pte_none(*pte))
1696 return 0;
1697
1698 return pfn_valid(pte_pfn(*pte));
1699 }
1700 EXPORT_SYMBOL(kern_addr_valid);
1701
1702 static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1703 unsigned long vend,
1704 pud_t *pud)
1705 {
1706 const unsigned long mask16gb = (1UL << 34) - 1UL;
1707 u64 pte_val = vstart;
1708
1709 /* Each PUD is 8GB */
1710 if ((vstart & mask16gb) ||
1711 (vend - vstart <= mask16gb)) {
1712 pte_val ^= kern_linear_pte_xor[2];
1713 pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1714
1715 return vstart + PUD_SIZE;
1716 }
1717
1718 pte_val ^= kern_linear_pte_xor[3];
1719 pte_val |= _PAGE_PUD_HUGE;
1720
1721 vend = vstart + mask16gb + 1UL;
1722 while (vstart < vend) {
1723 pud_val(*pud) = pte_val;
1724
1725 pte_val += PUD_SIZE;
1726 vstart += PUD_SIZE;
1727 pud++;
1728 }
1729 return vstart;
1730 }
1731
1732 static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1733 bool guard)
1734 {
1735 if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1736 return true;
1737
1738 return false;
1739 }
1740
1741 static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1742 unsigned long vend,
1743 pmd_t *pmd)
1744 {
1745 const unsigned long mask256mb = (1UL << 28) - 1UL;
1746 const unsigned long mask2gb = (1UL << 31) - 1UL;
1747 u64 pte_val = vstart;
1748
1749 /* Each PMD is 8MB */
1750 if ((vstart & mask256mb) ||
1751 (vend - vstart <= mask256mb)) {
1752 pte_val ^= kern_linear_pte_xor[0];
1753 pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1754
1755 return vstart + PMD_SIZE;
1756 }
1757
1758 if ((vstart & mask2gb) ||
1759 (vend - vstart <= mask2gb)) {
1760 pte_val ^= kern_linear_pte_xor[1];
1761 pte_val |= _PAGE_PMD_HUGE;
1762 vend = vstart + mask256mb + 1UL;
1763 } else {
1764 pte_val ^= kern_linear_pte_xor[2];
1765 pte_val |= _PAGE_PMD_HUGE;
1766 vend = vstart + mask2gb + 1UL;
1767 }
1768
1769 while (vstart < vend) {
1770 pmd_val(*pmd) = pte_val;
1771
1772 pte_val += PMD_SIZE;
1773 vstart += PMD_SIZE;
1774 pmd++;
1775 }
1776
1777 return vstart;
1778 }
1779
1780 static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1781 bool guard)
1782 {
1783 if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1784 return true;
1785
1786 return false;
1787 }
1788
1789 static unsigned long __ref kernel_map_range(unsigned long pstart,
1790 unsigned long pend, pgprot_t prot,
1791 bool use_huge)
1792 {
1793 unsigned long vstart = PAGE_OFFSET + pstart;
1794 unsigned long vend = PAGE_OFFSET + pend;
1795 unsigned long alloc_bytes = 0UL;
1796
1797 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1798 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1799 vstart, vend);
1800 prom_halt();
1801 }
1802
1803 while (vstart < vend) {
1804 unsigned long this_end, paddr = __pa(vstart);
1805 pgd_t *pgd = pgd_offset_k(vstart);
1806 pud_t *pud;
1807 pmd_t *pmd;
1808 pte_t *pte;
1809
1810 if (pgd_none(*pgd)) {
1811 pud_t *new;
1812
1813 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1814 alloc_bytes += PAGE_SIZE;
1815 pgd_populate(&init_mm, pgd, new);
1816 }
1817 pud = pud_offset(pgd, vstart);
1818 if (pud_none(*pud)) {
1819 pmd_t *new;
1820
1821 if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1822 vstart = kernel_map_hugepud(vstart, vend, pud);
1823 continue;
1824 }
1825 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1826 alloc_bytes += PAGE_SIZE;
1827 pud_populate(&init_mm, pud, new);
1828 }
1829
1830 pmd = pmd_offset(pud, vstart);
1831 if (pmd_none(*pmd)) {
1832 pte_t *new;
1833
1834 if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1835 vstart = kernel_map_hugepmd(vstart, vend, pmd);
1836 continue;
1837 }
1838 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1839 alloc_bytes += PAGE_SIZE;
1840 pmd_populate_kernel(&init_mm, pmd, new);
1841 }
1842
1843 pte = pte_offset_kernel(pmd, vstart);
1844 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1845 if (this_end > vend)
1846 this_end = vend;
1847
1848 while (vstart < this_end) {
1849 pte_val(*pte) = (paddr | pgprot_val(prot));
1850
1851 vstart += PAGE_SIZE;
1852 paddr += PAGE_SIZE;
1853 pte++;
1854 }
1855 }
1856
1857 return alloc_bytes;
1858 }
1859
1860 static void __init flush_all_kernel_tsbs(void)
1861 {
1862 int i;
1863
1864 for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1865 struct tsb *ent = &swapper_tsb[i];
1866
1867 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1868 }
1869 #ifndef CONFIG_DEBUG_PAGEALLOC
1870 for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1871 struct tsb *ent = &swapper_4m_tsb[i];
1872
1873 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1874 }
1875 #endif
1876 }
1877
1878 extern unsigned int kvmap_linear_patch[1];
1879
1880 static void __init kernel_physical_mapping_init(void)
1881 {
1882 unsigned long i, mem_alloced = 0UL;
1883 bool use_huge = true;
1884
1885 #ifdef CONFIG_DEBUG_PAGEALLOC
1886 use_huge = false;
1887 #endif
1888 for (i = 0; i < pall_ents; i++) {
1889 unsigned long phys_start, phys_end;
1890
1891 phys_start = pall[i].phys_addr;
1892 phys_end = phys_start + pall[i].reg_size;
1893
1894 mem_alloced += kernel_map_range(phys_start, phys_end,
1895 PAGE_KERNEL, use_huge);
1896 }
1897
1898 printk("Allocated %ld bytes for kernel page tables.\n",
1899 mem_alloced);
1900
1901 kvmap_linear_patch[0] = 0x01000000; /* nop */
1902 flushi(&kvmap_linear_patch[0]);
1903
1904 flush_all_kernel_tsbs();
1905
1906 __flush_tlb_all();
1907 }
1908
1909 #ifdef CONFIG_DEBUG_PAGEALLOC
1910 void __kernel_map_pages(struct page *page, int numpages, int enable)
1911 {
1912 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1913 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1914
1915 kernel_map_range(phys_start, phys_end,
1916 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1917
1918 flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1919 PAGE_OFFSET + phys_end);
1920
1921 /* we should perform an IPI and flush all tlbs,
1922 * but that can deadlock->flush only current cpu.
1923 */
1924 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1925 PAGE_OFFSET + phys_end);
1926 }
1927 #endif
1928
1929 unsigned long __init find_ecache_flush_span(unsigned long size)
1930 {
1931 int i;
1932
1933 for (i = 0; i < pavail_ents; i++) {
1934 if (pavail[i].reg_size >= size)
1935 return pavail[i].phys_addr;
1936 }
1937
1938 return ~0UL;
1939 }
1940
1941 unsigned long PAGE_OFFSET;
1942 EXPORT_SYMBOL(PAGE_OFFSET);
1943
1944 unsigned long VMALLOC_END = 0x0000010000000000UL;
1945 EXPORT_SYMBOL(VMALLOC_END);
1946
1947 unsigned long sparc64_va_hole_top = 0xfffff80000000000UL;
1948 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1949
1950 static void __init setup_page_offset(void)
1951 {
1952 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1953 /* Cheetah/Panther support a full 64-bit virtual
1954 * address, so we can use all that our page tables
1955 * support.
1956 */
1957 sparc64_va_hole_top = 0xfff0000000000000UL;
1958 sparc64_va_hole_bottom = 0x0010000000000000UL;
1959
1960 max_phys_bits = 42;
1961 } else if (tlb_type == hypervisor) {
1962 switch (sun4v_chip_type) {
1963 case SUN4V_CHIP_NIAGARA1:
1964 case SUN4V_CHIP_NIAGARA2:
1965 /* T1 and T2 support 48-bit virtual addresses. */
1966 sparc64_va_hole_top = 0xffff800000000000UL;
1967 sparc64_va_hole_bottom = 0x0000800000000000UL;
1968
1969 max_phys_bits = 39;
1970 break;
1971 case SUN4V_CHIP_NIAGARA3:
1972 /* T3 supports 48-bit virtual addresses. */
1973 sparc64_va_hole_top = 0xffff800000000000UL;
1974 sparc64_va_hole_bottom = 0x0000800000000000UL;
1975
1976 max_phys_bits = 43;
1977 break;
1978 case SUN4V_CHIP_NIAGARA4:
1979 case SUN4V_CHIP_NIAGARA5:
1980 case SUN4V_CHIP_SPARC64X:
1981 case SUN4V_CHIP_SPARC_M6:
1982 /* T4 and later support 52-bit virtual addresses. */
1983 sparc64_va_hole_top = 0xfff8000000000000UL;
1984 sparc64_va_hole_bottom = 0x0008000000000000UL;
1985 max_phys_bits = 47;
1986 break;
1987 case SUN4V_CHIP_SPARC_M7:
1988 case SUN4V_CHIP_SPARC_SN:
1989 /* M7 and later support 52-bit virtual addresses. */
1990 sparc64_va_hole_top = 0xfff8000000000000UL;
1991 sparc64_va_hole_bottom = 0x0008000000000000UL;
1992 max_phys_bits = 49;
1993 break;
1994 case SUN4V_CHIP_SPARC_M8:
1995 default:
1996 /* M8 and later support 54-bit virtual addresses.
1997 * However, restricting M8 and above VA bits to 53
1998 * as 4-level page table cannot support more than
1999 * 53 VA bits.
2000 */
2001 sparc64_va_hole_top = 0xfff0000000000000UL;
2002 sparc64_va_hole_bottom = 0x0010000000000000UL;
2003 max_phys_bits = 51;
2004 break;
2005 }
2006 }
2007
2008 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
2009 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
2010 max_phys_bits);
2011 prom_halt();
2012 }
2013
2014 PAGE_OFFSET = sparc64_va_hole_top;
2015 VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
2016 (sparc64_va_hole_bottom >> 2));
2017
2018 pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
2019 PAGE_OFFSET, max_phys_bits);
2020 pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
2021 VMALLOC_START, VMALLOC_END);
2022 pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
2023 VMEMMAP_BASE, VMEMMAP_BASE << 1);
2024 }
2025
2026 static void __init tsb_phys_patch(void)
2027 {
2028 struct tsb_ldquad_phys_patch_entry *pquad;
2029 struct tsb_phys_patch_entry *p;
2030
2031 pquad = &__tsb_ldquad_phys_patch;
2032 while (pquad < &__tsb_ldquad_phys_patch_end) {
2033 unsigned long addr = pquad->addr;
2034
2035 if (tlb_type == hypervisor)
2036 *(unsigned int *) addr = pquad->sun4v_insn;
2037 else
2038 *(unsigned int *) addr = pquad->sun4u_insn;
2039 wmb();
2040 __asm__ __volatile__("flush %0"
2041 : /* no outputs */
2042 : "r" (addr));
2043
2044 pquad++;
2045 }
2046
2047 p = &__tsb_phys_patch;
2048 while (p < &__tsb_phys_patch_end) {
2049 unsigned long addr = p->addr;
2050
2051 *(unsigned int *) addr = p->insn;
2052 wmb();
2053 __asm__ __volatile__("flush %0"
2054 : /* no outputs */
2055 : "r" (addr));
2056
2057 p++;
2058 }
2059 }
2060
2061 /* Don't mark as init, we give this to the Hypervisor. */
2062 #ifndef CONFIG_DEBUG_PAGEALLOC
2063 #define NUM_KTSB_DESCR 2
2064 #else
2065 #define NUM_KTSB_DESCR 1
2066 #endif
2067 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
2068
2069 /* The swapper TSBs are loaded with a base sequence of:
2070 *
2071 * sethi %uhi(SYMBOL), REG1
2072 * sethi %hi(SYMBOL), REG2
2073 * or REG1, %ulo(SYMBOL), REG1
2074 * or REG2, %lo(SYMBOL), REG2
2075 * sllx REG1, 32, REG1
2076 * or REG1, REG2, REG1
2077 *
2078 * When we use physical addressing for the TSB accesses, we patch the
2079 * first four instructions in the above sequence.
2080 */
2081
2082 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
2083 {
2084 unsigned long high_bits, low_bits;
2085
2086 high_bits = (pa >> 32) & 0xffffffff;
2087 low_bits = (pa >> 0) & 0xffffffff;
2088
2089 while (start < end) {
2090 unsigned int *ia = (unsigned int *)(unsigned long)*start;
2091
2092 ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
2093 __asm__ __volatile__("flush %0" : : "r" (ia));
2094
2095 ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
2096 __asm__ __volatile__("flush %0" : : "r" (ia + 1));
2097
2098 ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
2099 __asm__ __volatile__("flush %0" : : "r" (ia + 2));
2100
2101 ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
2102 __asm__ __volatile__("flush %0" : : "r" (ia + 3));
2103
2104 start++;
2105 }
2106 }
2107
2108 static void ktsb_phys_patch(void)
2109 {
2110 extern unsigned int __swapper_tsb_phys_patch;
2111 extern unsigned int __swapper_tsb_phys_patch_end;
2112 unsigned long ktsb_pa;
2113
2114 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2115 patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
2116 &__swapper_tsb_phys_patch_end, ktsb_pa);
2117 #ifndef CONFIG_DEBUG_PAGEALLOC
2118 {
2119 extern unsigned int __swapper_4m_tsb_phys_patch;
2120 extern unsigned int __swapper_4m_tsb_phys_patch_end;
2121 ktsb_pa = (kern_base +
2122 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2123 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
2124 &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
2125 }
2126 #endif
2127 }
2128
2129 static void __init sun4v_ktsb_init(void)
2130 {
2131 unsigned long ktsb_pa;
2132
2133 /* First KTSB for PAGE_SIZE mappings. */
2134 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2135
2136 switch (PAGE_SIZE) {
2137 case 8 * 1024:
2138 default:
2139 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
2140 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
2141 break;
2142
2143 case 64 * 1024:
2144 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
2145 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
2146 break;
2147
2148 case 512 * 1024:
2149 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
2150 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
2151 break;
2152
2153 case 4 * 1024 * 1024:
2154 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
2155 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
2156 break;
2157 }
2158
2159 ktsb_descr[0].assoc = 1;
2160 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
2161 ktsb_descr[0].ctx_idx = 0;
2162 ktsb_descr[0].tsb_base = ktsb_pa;
2163 ktsb_descr[0].resv = 0;
2164
2165 #ifndef CONFIG_DEBUG_PAGEALLOC
2166 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */
2167 ktsb_pa = (kern_base +
2168 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2169
2170 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
2171 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
2172 HV_PGSZ_MASK_256MB |
2173 HV_PGSZ_MASK_2GB |
2174 HV_PGSZ_MASK_16GB) &
2175 cpu_pgsz_mask);
2176 ktsb_descr[1].assoc = 1;
2177 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
2178 ktsb_descr[1].ctx_idx = 0;
2179 ktsb_descr[1].tsb_base = ktsb_pa;
2180 ktsb_descr[1].resv = 0;
2181 #endif
2182 }
2183
2184 void sun4v_ktsb_register(void)
2185 {
2186 unsigned long pa, ret;
2187
2188 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
2189
2190 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
2191 if (ret != 0) {
2192 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
2193 "errors with %lx\n", pa, ret);
2194 prom_halt();
2195 }
2196 }
2197
2198 static void __init sun4u_linear_pte_xor_finalize(void)
2199 {
2200 #ifndef CONFIG_DEBUG_PAGEALLOC
2201 /* This is where we would add Panther support for
2202 * 32MB and 256MB pages.
2203 */
2204 #endif
2205 }
2206
2207 static void __init sun4v_linear_pte_xor_finalize(void)
2208 {
2209 unsigned long pagecv_flag;
2210
2211 /* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
2212 * enables MCD error. Do not set bit 9 on M7 processor.
2213 */
2214 switch (sun4v_chip_type) {
2215 case SUN4V_CHIP_SPARC_M7:
2216 case SUN4V_CHIP_SPARC_M8:
2217 case SUN4V_CHIP_SPARC_SN:
2218 pagecv_flag = 0x00;
2219 break;
2220 default:
2221 pagecv_flag = _PAGE_CV_4V;
2222 break;
2223 }
2224 #ifndef CONFIG_DEBUG_PAGEALLOC
2225 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
2226 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2227 PAGE_OFFSET;
2228 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2229 _PAGE_P_4V | _PAGE_W_4V);
2230 } else {
2231 kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2232 }
2233
2234 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2235 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2236 PAGE_OFFSET;
2237 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2238 _PAGE_P_4V | _PAGE_W_4V);
2239 } else {
2240 kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2241 }
2242
2243 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2244 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2245 PAGE_OFFSET;
2246 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2247 _PAGE_P_4V | _PAGE_W_4V);
2248 } else {
2249 kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2250 }
2251 #endif
2252 }
2253
2254 /* paging_init() sets up the page tables */
2255
2256 static unsigned long last_valid_pfn;
2257
2258 static void sun4u_pgprot_init(void);
2259 static void sun4v_pgprot_init(void);
2260
2261 static phys_addr_t __init available_memory(void)
2262 {
2263 phys_addr_t available = 0ULL;
2264 phys_addr_t pa_start, pa_end;
2265 u64 i;
2266
2267 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2268 &pa_end, NULL)
2269 available = available + (pa_end - pa_start);
2270
2271 return available;
2272 }
2273
2274 #define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U)
2275 #define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V)
2276 #define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2277 #define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2278 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2279 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2280
2281 /* We need to exclude reserved regions. This exclusion will include
2282 * vmlinux and initrd. To be more precise the initrd size could be used to
2283 * compute a new lower limit because it is freed later during initialization.
2284 */
2285 static void __init reduce_memory(phys_addr_t limit_ram)
2286 {
2287 phys_addr_t avail_ram = available_memory();
2288 phys_addr_t pa_start, pa_end;
2289 u64 i;
2290
2291 if (limit_ram >= avail_ram)
2292 return;
2293
2294 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2295 &pa_end, NULL) {
2296 phys_addr_t region_size = pa_end - pa_start;
2297 phys_addr_t clip_start = pa_start;
2298
2299 avail_ram = avail_ram - region_size;
2300 /* Are we consuming too much? */
2301 if (avail_ram < limit_ram) {
2302 phys_addr_t give_back = limit_ram - avail_ram;
2303
2304 region_size = region_size - give_back;
2305 clip_start = clip_start + give_back;
2306 }
2307
2308 memblock_remove(clip_start, region_size);
2309
2310 if (avail_ram <= limit_ram)
2311 break;
2312 i = 0UL;
2313 }
2314 }
2315
2316 void __init paging_init(void)
2317 {
2318 unsigned long end_pfn, shift, phys_base;
2319 unsigned long real_end, i;
2320
2321 setup_page_offset();
2322
2323 /* These build time checkes make sure that the dcache_dirty_cpu()
2324 * page->flags usage will work.
2325 *
2326 * When a page gets marked as dcache-dirty, we store the
2327 * cpu number starting at bit 32 in the page->flags. Also,
2328 * functions like clear_dcache_dirty_cpu use the cpu mask
2329 * in 13-bit signed-immediate instruction fields.
2330 */
2331
2332 /*
2333 * Page flags must not reach into upper 32 bits that are used
2334 * for the cpu number
2335 */
2336 BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2337
2338 /*
2339 * The bit fields placed in the high range must not reach below
2340 * the 32 bit boundary. Otherwise we cannot place the cpu field
2341 * at the 32 bit boundary.
2342 */
2343 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2344 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2345
2346 BUILD_BUG_ON(NR_CPUS > 4096);
2347
2348 kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2349 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2350
2351 /* Invalidate both kernel TSBs. */
2352 memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2353 #ifndef CONFIG_DEBUG_PAGEALLOC
2354 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2355 #endif
2356
2357 /* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2358 * bit on M7 processor. This is a conflicting usage of the same
2359 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2360 * Detection error on all pages and this will lead to problems
2361 * later. Kernel does not run with MCD enabled and hence rest
2362 * of the required steps to fully configure memory corruption
2363 * detection are not taken. We need to ensure TTE.mcde is not
2364 * set on M7 processor. Compute the value of cacheability
2365 * flag for use later taking this into consideration.
2366 */
2367 switch (sun4v_chip_type) {
2368 case SUN4V_CHIP_SPARC_M7:
2369 case SUN4V_CHIP_SPARC_M8:
2370 case SUN4V_CHIP_SPARC_SN:
2371 page_cache4v_flag = _PAGE_CP_4V;
2372 break;
2373 default:
2374 page_cache4v_flag = _PAGE_CACHE_4V;
2375 break;
2376 }
2377
2378 if (tlb_type == hypervisor)
2379 sun4v_pgprot_init();
2380 else
2381 sun4u_pgprot_init();
2382
2383 if (tlb_type == cheetah_plus ||
2384 tlb_type == hypervisor) {
2385 tsb_phys_patch();
2386 ktsb_phys_patch();
2387 }
2388
2389 if (tlb_type == hypervisor)
2390 sun4v_patch_tlb_handlers();
2391
2392 /* Find available physical memory...
2393 *
2394 * Read it twice in order to work around a bug in openfirmware.
2395 * The call to grab this table itself can cause openfirmware to
2396 * allocate memory, which in turn can take away some space from
2397 * the list of available memory. Reading it twice makes sure
2398 * we really do get the final value.
2399 */
2400 read_obp_translations();
2401 read_obp_memory("reg", &pall[0], &pall_ents);
2402 read_obp_memory("available", &pavail[0], &pavail_ents);
2403 read_obp_memory("available", &pavail[0], &pavail_ents);
2404
2405 phys_base = 0xffffffffffffffffUL;
2406 for (i = 0; i < pavail_ents; i++) {
2407 phys_base = min(phys_base, pavail[i].phys_addr);
2408 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2409 }
2410
2411 memblock_reserve(kern_base, kern_size);
2412
2413 find_ramdisk(phys_base);
2414
2415 if (cmdline_memory_size)
2416 reduce_memory(cmdline_memory_size);
2417
2418 memblock_allow_resize();
2419 memblock_dump_all();
2420
2421 set_bit(0, mmu_context_bmap);
2422
2423 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2424
2425 real_end = (unsigned long)_end;
2426 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2427 printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2428 num_kernel_image_mappings);
2429
2430 /* Set kernel pgd to upper alias so physical page computations
2431 * work.
2432 */
2433 init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2434
2435 memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2436
2437 inherit_prom_mappings();
2438
2439 /* Ok, we can use our TLB miss and window trap handlers safely. */
2440 setup_tba();
2441
2442 __flush_tlb_all();
2443
2444 prom_build_devicetree();
2445 of_populate_present_mask();
2446 #ifndef CONFIG_SMP
2447 of_fill_in_cpu_data();
2448 #endif
2449
2450 if (tlb_type == hypervisor) {
2451 sun4v_mdesc_init();
2452 mdesc_populate_present_mask(cpu_all_mask);
2453 #ifndef CONFIG_SMP
2454 mdesc_fill_in_cpu_data(cpu_all_mask);
2455 #endif
2456 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2457
2458 sun4v_linear_pte_xor_finalize();
2459
2460 sun4v_ktsb_init();
2461 sun4v_ktsb_register();
2462 } else {
2463 unsigned long impl, ver;
2464
2465 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2466 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2467
2468 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2469 impl = ((ver >> 32) & 0xffff);
2470 if (impl == PANTHER_IMPL)
2471 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2472 HV_PGSZ_MASK_256MB);
2473
2474 sun4u_linear_pte_xor_finalize();
2475 }
2476
2477 /* Flush the TLBs and the 4M TSB so that the updated linear
2478 * pte XOR settings are realized for all mappings.
2479 */
2480 __flush_tlb_all();
2481 #ifndef CONFIG_DEBUG_PAGEALLOC
2482 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2483 #endif
2484 __flush_tlb_all();
2485
2486 /* Setup bootmem... */
2487 last_valid_pfn = end_pfn = bootmem_init(phys_base);
2488
2489 kernel_physical_mapping_init();
2490
2491 {
2492 unsigned long max_zone_pfns[MAX_NR_ZONES];
2493
2494 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2495
2496 max_zone_pfns[ZONE_NORMAL] = end_pfn;
2497
2498 free_area_init_nodes(max_zone_pfns);
2499 }
2500
2501 printk("Booting Linux...\n");
2502 }
2503
2504 int page_in_phys_avail(unsigned long paddr)
2505 {
2506 int i;
2507
2508 paddr &= PAGE_MASK;
2509
2510 for (i = 0; i < pavail_ents; i++) {
2511 unsigned long start, end;
2512
2513 start = pavail[i].phys_addr;
2514 end = start + pavail[i].reg_size;
2515
2516 if (paddr >= start && paddr < end)
2517 return 1;
2518 }
2519 if (paddr >= kern_base && paddr < (kern_base + kern_size))
2520 return 1;
2521 #ifdef CONFIG_BLK_DEV_INITRD
2522 if (paddr >= __pa(initrd_start) &&
2523 paddr < __pa(PAGE_ALIGN(initrd_end)))
2524 return 1;
2525 #endif
2526
2527 return 0;
2528 }
2529
2530 static void __init register_page_bootmem_info(void)
2531 {
2532 #ifdef CONFIG_NEED_MULTIPLE_NODES
2533 int i;
2534
2535 for_each_online_node(i)
2536 if (NODE_DATA(i)->node_spanned_pages)
2537 register_page_bootmem_info_node(NODE_DATA(i));
2538 #endif
2539 }
2540 void __init mem_init(void)
2541 {
2542 high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2543
2544 free_all_bootmem();
2545
2546 /*
2547 * Must be done after boot memory is put on freelist, because here we
2548 * might set fields in deferred struct pages that have not yet been
2549 * initialized, and free_all_bootmem() initializes all the reserved
2550 * deferred pages for us.
2551 */
2552 register_page_bootmem_info();
2553
2554 /*
2555 * Set up the zero page, mark it reserved, so that page count
2556 * is not manipulated when freeing the page from user ptes.
2557 */
2558 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2559 if (mem_map_zero == NULL) {
2560 prom_printf("paging_init: Cannot alloc zero page.\n");
2561 prom_halt();
2562 }
2563 mark_page_reserved(mem_map_zero);
2564
2565 mem_init_print_info(NULL);
2566
2567 if (tlb_type == cheetah || tlb_type == cheetah_plus)
2568 cheetah_ecache_flush_init();
2569 }
2570
2571 void free_initmem(void)
2572 {
2573 unsigned long addr, initend;
2574 int do_free = 1;
2575
2576 /* If the physical memory maps were trimmed by kernel command
2577 * line options, don't even try freeing this initmem stuff up.
2578 * The kernel image could have been in the trimmed out region
2579 * and if so the freeing below will free invalid page structs.
2580 */
2581 if (cmdline_memory_size)
2582 do_free = 0;
2583
2584 /*
2585 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2586 */
2587 addr = PAGE_ALIGN((unsigned long)(__init_begin));
2588 initend = (unsigned long)(__init_end) & PAGE_MASK;
2589 for (; addr < initend; addr += PAGE_SIZE) {
2590 unsigned long page;
2591
2592 page = (addr +
2593 ((unsigned long) __va(kern_base)) -
2594 ((unsigned long) KERNBASE));
2595 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2596
2597 if (do_free)
2598 free_reserved_page(virt_to_page(page));
2599 }
2600 }
2601
2602 #ifdef CONFIG_BLK_DEV_INITRD
2603 void free_initrd_mem(unsigned long start, unsigned long end)
2604 {
2605 free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2606 "initrd");
2607 }
2608 #endif
2609
2610 pgprot_t PAGE_KERNEL __read_mostly;
2611 EXPORT_SYMBOL(PAGE_KERNEL);
2612
2613 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2614 pgprot_t PAGE_COPY __read_mostly;
2615
2616 pgprot_t PAGE_SHARED __read_mostly;
2617 EXPORT_SYMBOL(PAGE_SHARED);
2618
2619 unsigned long pg_iobits __read_mostly;
2620
2621 unsigned long _PAGE_IE __read_mostly;
2622 EXPORT_SYMBOL(_PAGE_IE);
2623
2624 unsigned long _PAGE_E __read_mostly;
2625 EXPORT_SYMBOL(_PAGE_E);
2626
2627 unsigned long _PAGE_CACHE __read_mostly;
2628 EXPORT_SYMBOL(_PAGE_CACHE);
2629
2630 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2631 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2632 int node)
2633 {
2634 unsigned long pte_base;
2635
2636 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2637 _PAGE_CP_4U | _PAGE_CV_4U |
2638 _PAGE_P_4U | _PAGE_W_4U);
2639 if (tlb_type == hypervisor)
2640 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2641 page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2642
2643 pte_base |= _PAGE_PMD_HUGE;
2644
2645 vstart = vstart & PMD_MASK;
2646 vend = ALIGN(vend, PMD_SIZE);
2647 for (; vstart < vend; vstart += PMD_SIZE) {
2648 pgd_t *pgd = pgd_offset_k(vstart);
2649 unsigned long pte;
2650 pud_t *pud;
2651 pmd_t *pmd;
2652
2653 if (pgd_none(*pgd)) {
2654 pud_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2655
2656 if (!new)
2657 return -ENOMEM;
2658 pgd_populate(&init_mm, pgd, new);
2659 }
2660
2661 pud = pud_offset(pgd, vstart);
2662 if (pud_none(*pud)) {
2663 pmd_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2664
2665 if (!new)
2666 return -ENOMEM;
2667 pud_populate(&init_mm, pud, new);
2668 }
2669
2670 pmd = pmd_offset(pud, vstart);
2671
2672 pte = pmd_val(*pmd);
2673 if (!(pte & _PAGE_VALID)) {
2674 void *block = vmemmap_alloc_block(PMD_SIZE, node);
2675
2676 if (!block)
2677 return -ENOMEM;
2678
2679 pmd_val(*pmd) = pte_base | __pa(block);
2680 }
2681 }
2682
2683 return 0;
2684 }
2685
2686 void vmemmap_free(unsigned long start, unsigned long end)
2687 {
2688 }
2689 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2690
2691 static void prot_init_common(unsigned long page_none,
2692 unsigned long page_shared,
2693 unsigned long page_copy,
2694 unsigned long page_readonly,
2695 unsigned long page_exec_bit)
2696 {
2697 PAGE_COPY = __pgprot(page_copy);
2698 PAGE_SHARED = __pgprot(page_shared);
2699
2700 protection_map[0x0] = __pgprot(page_none);
2701 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2702 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2703 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2704 protection_map[0x4] = __pgprot(page_readonly);
2705 protection_map[0x5] = __pgprot(page_readonly);
2706 protection_map[0x6] = __pgprot(page_copy);
2707 protection_map[0x7] = __pgprot(page_copy);
2708 protection_map[0x8] = __pgprot(page_none);
2709 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2710 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2711 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2712 protection_map[0xc] = __pgprot(page_readonly);
2713 protection_map[0xd] = __pgprot(page_readonly);
2714 protection_map[0xe] = __pgprot(page_shared);
2715 protection_map[0xf] = __pgprot(page_shared);
2716 }
2717
2718 static void __init sun4u_pgprot_init(void)
2719 {
2720 unsigned long page_none, page_shared, page_copy, page_readonly;
2721 unsigned long page_exec_bit;
2722 int i;
2723
2724 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2725 _PAGE_CACHE_4U | _PAGE_P_4U |
2726 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2727 _PAGE_EXEC_4U);
2728 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2729 _PAGE_CACHE_4U | _PAGE_P_4U |
2730 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2731 _PAGE_EXEC_4U | _PAGE_L_4U);
2732
2733 _PAGE_IE = _PAGE_IE_4U;
2734 _PAGE_E = _PAGE_E_4U;
2735 _PAGE_CACHE = _PAGE_CACHE_4U;
2736
2737 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2738 __ACCESS_BITS_4U | _PAGE_E_4U);
2739
2740 #ifdef CONFIG_DEBUG_PAGEALLOC
2741 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2742 #else
2743 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2744 PAGE_OFFSET;
2745 #endif
2746 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2747 _PAGE_P_4U | _PAGE_W_4U);
2748
2749 for (i = 1; i < 4; i++)
2750 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2751
2752 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2753 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2754 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2755
2756
2757 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2758 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2759 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2760 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2761 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2762 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2763 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2764
2765 page_exec_bit = _PAGE_EXEC_4U;
2766
2767 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2768 page_exec_bit);
2769 }
2770
2771 static void __init sun4v_pgprot_init(void)
2772 {
2773 unsigned long page_none, page_shared, page_copy, page_readonly;
2774 unsigned long page_exec_bit;
2775 int i;
2776
2777 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2778 page_cache4v_flag | _PAGE_P_4V |
2779 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2780 _PAGE_EXEC_4V);
2781 PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2782
2783 _PAGE_IE = _PAGE_IE_4V;
2784 _PAGE_E = _PAGE_E_4V;
2785 _PAGE_CACHE = page_cache4v_flag;
2786
2787 #ifdef CONFIG_DEBUG_PAGEALLOC
2788 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2789 #else
2790 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2791 PAGE_OFFSET;
2792 #endif
2793 kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2794 _PAGE_W_4V);
2795
2796 for (i = 1; i < 4; i++)
2797 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2798
2799 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2800 __ACCESS_BITS_4V | _PAGE_E_4V);
2801
2802 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2803 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2804 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2805 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2806
2807 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2808 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2809 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2810 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2811 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2812 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2813 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2814
2815 page_exec_bit = _PAGE_EXEC_4V;
2816
2817 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2818 page_exec_bit);
2819 }
2820
2821 unsigned long pte_sz_bits(unsigned long sz)
2822 {
2823 if (tlb_type == hypervisor) {
2824 switch (sz) {
2825 case 8 * 1024:
2826 default:
2827 return _PAGE_SZ8K_4V;
2828 case 64 * 1024:
2829 return _PAGE_SZ64K_4V;
2830 case 512 * 1024:
2831 return _PAGE_SZ512K_4V;
2832 case 4 * 1024 * 1024:
2833 return _PAGE_SZ4MB_4V;
2834 }
2835 } else {
2836 switch (sz) {
2837 case 8 * 1024:
2838 default:
2839 return _PAGE_SZ8K_4U;
2840 case 64 * 1024:
2841 return _PAGE_SZ64K_4U;
2842 case 512 * 1024:
2843 return _PAGE_SZ512K_4U;
2844 case 4 * 1024 * 1024:
2845 return _PAGE_SZ4MB_4U;
2846 }
2847 }
2848 }
2849
2850 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2851 {
2852 pte_t pte;
2853
2854 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot));
2855 pte_val(pte) |= (((unsigned long)space) << 32);
2856 pte_val(pte) |= pte_sz_bits(page_size);
2857
2858 return pte;
2859 }
2860
2861 static unsigned long kern_large_tte(unsigned long paddr)
2862 {
2863 unsigned long val;
2864
2865 val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2866 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2867 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2868 if (tlb_type == hypervisor)
2869 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2870 page_cache4v_flag | _PAGE_P_4V |
2871 _PAGE_EXEC_4V | _PAGE_W_4V);
2872
2873 return val | paddr;
2874 }
2875
2876 /* If not locked, zap it. */
2877 void __flush_tlb_all(void)
2878 {
2879 unsigned long pstate;
2880 int i;
2881
2882 __asm__ __volatile__("flushw\n\t"
2883 "rdpr %%pstate, %0\n\t"
2884 "wrpr %0, %1, %%pstate"
2885 : "=r" (pstate)
2886 : "i" (PSTATE_IE));
2887 if (tlb_type == hypervisor) {
2888 sun4v_mmu_demap_all();
2889 } else if (tlb_type == spitfire) {
2890 for (i = 0; i < 64; i++) {
2891 /* Spitfire Errata #32 workaround */
2892 /* NOTE: Always runs on spitfire, so no
2893 * cheetah+ page size encodings.
2894 */
2895 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2896 "flush %%g6"
2897 : /* No outputs */
2898 : "r" (0),
2899 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2900
2901 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2902 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2903 "membar #Sync"
2904 : /* no outputs */
2905 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2906 spitfire_put_dtlb_data(i, 0x0UL);
2907 }
2908
2909 /* Spitfire Errata #32 workaround */
2910 /* NOTE: Always runs on spitfire, so no
2911 * cheetah+ page size encodings.
2912 */
2913 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2914 "flush %%g6"
2915 : /* No outputs */
2916 : "r" (0),
2917 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2918
2919 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2920 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2921 "membar #Sync"
2922 : /* no outputs */
2923 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2924 spitfire_put_itlb_data(i, 0x0UL);
2925 }
2926 }
2927 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2928 cheetah_flush_dtlb_all();
2929 cheetah_flush_itlb_all();
2930 }
2931 __asm__ __volatile__("wrpr %0, 0, %%pstate"
2932 : : "r" (pstate));
2933 }
2934
2935 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2936 unsigned long address)
2937 {
2938 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2939 pte_t *pte = NULL;
2940
2941 if (page)
2942 pte = (pte_t *) page_address(page);
2943
2944 return pte;
2945 }
2946
2947 pgtable_t pte_alloc_one(struct mm_struct *mm,
2948 unsigned long address)
2949 {
2950 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2951 if (!page)
2952 return NULL;
2953 if (!pgtable_page_ctor(page)) {
2954 free_hot_cold_page(page, 0);
2955 return NULL;
2956 }
2957 return (pte_t *) page_address(page);
2958 }
2959
2960 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2961 {
2962 free_page((unsigned long)pte);
2963 }
2964
2965 static void __pte_free(pgtable_t pte)
2966 {
2967 struct page *page = virt_to_page(pte);
2968
2969 pgtable_page_dtor(page);
2970 __free_page(page);
2971 }
2972
2973 void pte_free(struct mm_struct *mm, pgtable_t pte)
2974 {
2975 __pte_free(pte);
2976 }
2977
2978 void pgtable_free(void *table, bool is_page)
2979 {
2980 if (is_page)
2981 __pte_free(table);
2982 else
2983 kmem_cache_free(pgtable_cache, table);
2984 }
2985
2986 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2987 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2988 pmd_t *pmd)
2989 {
2990 unsigned long pte, flags;
2991 struct mm_struct *mm;
2992 pmd_t entry = *pmd;
2993
2994 if (!pmd_large(entry) || !pmd_young(entry))
2995 return;
2996
2997 pte = pmd_val(entry);
2998
2999 /* Don't insert a non-valid PMD into the TSB, we'll deadlock. */
3000 if (!(pte & _PAGE_VALID))
3001 return;
3002
3003 /* We are fabricating 8MB pages using 4MB real hw pages. */
3004 pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
3005
3006 mm = vma->vm_mm;
3007
3008 spin_lock_irqsave(&mm->context.lock, flags);
3009
3010 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
3011 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
3012 addr, pte);
3013
3014 spin_unlock_irqrestore(&mm->context.lock, flags);
3015 }
3016 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3017
3018 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
3019 static void context_reload(void *__data)
3020 {
3021 struct mm_struct *mm = __data;
3022
3023 if (mm == current->mm)
3024 load_secondary_context(mm);
3025 }
3026
3027 void hugetlb_setup(struct pt_regs *regs)
3028 {
3029 struct mm_struct *mm = current->mm;
3030 struct tsb_config *tp;
3031
3032 if (faulthandler_disabled() || !mm) {
3033 const struct exception_table_entry *entry;
3034
3035 entry = search_exception_tables(regs->tpc);
3036 if (entry) {
3037 regs->tpc = entry->fixup;
3038 regs->tnpc = regs->tpc + 4;
3039 return;
3040 }
3041 pr_alert("Unexpected HugeTLB setup in atomic context.\n");
3042 die_if_kernel("HugeTSB in atomic", regs);
3043 }
3044
3045 tp = &mm->context.tsb_block[MM_TSB_HUGE];
3046 if (likely(tp->tsb == NULL))
3047 tsb_grow(mm, MM_TSB_HUGE, 0);
3048
3049 tsb_context_switch(mm);
3050 smp_tsb_sync(mm);
3051
3052 /* On UltraSPARC-III+ and later, configure the second half of
3053 * the Data-TLB for huge pages.
3054 */
3055 if (tlb_type == cheetah_plus) {
3056 bool need_context_reload = false;
3057 unsigned long ctx;
3058
3059 spin_lock_irq(&ctx_alloc_lock);
3060 ctx = mm->context.sparc64_ctx_val;
3061 ctx &= ~CTX_PGSZ_MASK;
3062 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
3063 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
3064
3065 if (ctx != mm->context.sparc64_ctx_val) {
3066 /* When changing the page size fields, we
3067 * must perform a context flush so that no
3068 * stale entries match. This flush must
3069 * occur with the original context register
3070 * settings.
3071 */
3072 do_flush_tlb_mm(mm);
3073
3074 /* Reload the context register of all processors
3075 * also executing in this address space.
3076 */
3077 mm->context.sparc64_ctx_val = ctx;
3078 need_context_reload = true;
3079 }
3080 spin_unlock_irq(&ctx_alloc_lock);
3081
3082 if (need_context_reload)
3083 on_each_cpu(context_reload, mm, 0);
3084 }
3085 }
3086 #endif
3087
3088 static struct resource code_resource = {
3089 .name = "Kernel code",
3090 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3091 };
3092
3093 static struct resource data_resource = {
3094 .name = "Kernel data",
3095 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3096 };
3097
3098 static struct resource bss_resource = {
3099 .name = "Kernel bss",
3100 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3101 };
3102
3103 static inline resource_size_t compute_kern_paddr(void *addr)
3104 {
3105 return (resource_size_t) (addr - KERNBASE + kern_base);
3106 }
3107
3108 static void __init kernel_lds_init(void)
3109 {
3110 code_resource.start = compute_kern_paddr(_text);
3111 code_resource.end = compute_kern_paddr(_etext - 1);
3112 data_resource.start = compute_kern_paddr(_etext);
3113 data_resource.end = compute_kern_paddr(_edata - 1);
3114 bss_resource.start = compute_kern_paddr(__bss_start);
3115 bss_resource.end = compute_kern_paddr(_end - 1);
3116 }
3117
3118 static int __init report_memory(void)
3119 {
3120 int i;
3121 struct resource *res;
3122
3123 kernel_lds_init();
3124
3125 for (i = 0; i < pavail_ents; i++) {
3126 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
3127
3128 if (!res) {
3129 pr_warn("Failed to allocate source.\n");
3130 break;
3131 }
3132
3133 res->name = "System RAM";
3134 res->start = pavail[i].phys_addr;
3135 res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
3136 res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
3137
3138 if (insert_resource(&iomem_resource, res) < 0) {
3139 pr_warn("Resource insertion failed.\n");
3140 break;
3141 }
3142
3143 insert_resource(res, &code_resource);
3144 insert_resource(res, &data_resource);
3145 insert_resource(res, &bss_resource);
3146 }
3147
3148 return 0;
3149 }
3150 arch_initcall(report_memory);
3151
3152 #ifdef CONFIG_SMP
3153 #define do_flush_tlb_kernel_range smp_flush_tlb_kernel_range
3154 #else
3155 #define do_flush_tlb_kernel_range __flush_tlb_kernel_range
3156 #endif
3157
3158 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
3159 {
3160 if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
3161 if (start < LOW_OBP_ADDRESS) {
3162 flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
3163 do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
3164 }
3165 if (end > HI_OBP_ADDRESS) {
3166 flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
3167 do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
3168 }
3169 } else {
3170 flush_tsb_kernel_range(start, end);
3171 do_flush_tlb_kernel_range(start, end);
3172 }
3173 }
Linux with added FPC-III support