blob: 78e1982cb6c9dc0385e258b81bfc233f23baaaac [file] [log] [blame]
Chris Metcalf867e3592010-05-28 23:09:12 -04001/*
2 * Copyright (C) 1995 Linus Torvalds
3 * Copyright 2010 Tilera Corporation. All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License
7 * as published by the Free Software Foundation, version 2.
8 *
9 * This program is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
12 * NON INFRINGEMENT. See the GNU General Public License for
13 * more details.
14 */
15
16#include <linux/module.h>
17#include <linux/signal.h>
18#include <linux/sched.h>
19#include <linux/kernel.h>
20#include <linux/errno.h>
21#include <linux/string.h>
22#include <linux/types.h>
23#include <linux/ptrace.h>
24#include <linux/mman.h>
25#include <linux/mm.h>
26#include <linux/hugetlb.h>
27#include <linux/swap.h>
28#include <linux/smp.h>
29#include <linux/init.h>
30#include <linux/highmem.h>
31#include <linux/pagemap.h>
32#include <linux/poison.h>
33#include <linux/bootmem.h>
34#include <linux/slab.h>
35#include <linux/proc_fs.h>
36#include <linux/efi.h>
37#include <linux/memory_hotplug.h>
38#include <linux/uaccess.h>
39#include <asm/mmu_context.h>
40#include <asm/processor.h>
41#include <asm/system.h>
42#include <asm/pgtable.h>
43#include <asm/pgalloc.h>
44#include <asm/dma.h>
45#include <asm/fixmap.h>
46#include <asm/tlb.h>
47#include <asm/tlbflush.h>
48#include <asm/sections.h>
49#include <asm/setup.h>
50#include <asm/homecache.h>
51#include <hv/hypervisor.h>
52#include <arch/chip.h>
53
54#include "migrate.h"
55
56/*
57 * We could set FORCE_MAX_ZONEORDER to "(HPAGE_SHIFT - PAGE_SHIFT + 1)"
58 * in the Tile Kconfig, but this generates configure warnings.
59 * Do it here and force people to get it right to compile this file.
60 * The problem is that with 4KB small pages and 16MB huge pages,
61 * the default value doesn't allow us to group enough small pages
62 * together to make up a huge page.
63 */
64#if CONFIG_FORCE_MAX_ZONEORDER < HPAGE_SHIFT - PAGE_SHIFT + 1
65# error "Change FORCE_MAX_ZONEORDER in arch/tile/Kconfig to match page size"
66#endif
67
68#define clear_pgd(pmdptr) (*(pmdptr) = hv_pte(0))
69
Chris Metcalf0707ad32010-06-25 17:04:17 -040070#ifndef __tilegx__
Chris Metcalf867e3592010-05-28 23:09:12 -040071unsigned long VMALLOC_RESERVE = CONFIG_VMALLOC_RESERVE;
Chris Metcalf0707ad32010-06-25 17:04:17 -040072#endif
Chris Metcalf867e3592010-05-28 23:09:12 -040073
74DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
75
76/* Create an L2 page table */
77static pte_t * __init alloc_pte(void)
78{
79 return __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE, HV_PAGE_TABLE_ALIGN, 0);
80}
81
82/*
83 * L2 page tables per controller. We allocate these all at once from
84 * the bootmem allocator and store them here. This saves on kernel L2
85 * page table memory, compared to allocating a full 64K page per L2
86 * page table, and also means that in cases where we use huge pages,
87 * we are guaranteed to later be able to shatter those huge pages and
88 * switch to using these page tables instead, without requiring
89 * further allocation. Each l2_ptes[] entry points to the first page
90 * table for the first hugepage-size piece of memory on the
91 * controller; other page tables are just indexed directly, i.e. the
92 * L2 page tables are contiguous in memory for each controller.
93 */
94static pte_t *l2_ptes[MAX_NUMNODES];
95static int num_l2_ptes[MAX_NUMNODES];
96
97static void init_prealloc_ptes(int node, int pages)
98{
99 BUG_ON(pages & (HV_L2_ENTRIES-1));
100 if (pages) {
101 num_l2_ptes[node] = pages;
102 l2_ptes[node] = __alloc_bootmem(pages * sizeof(pte_t),
103 HV_PAGE_TABLE_ALIGN, 0);
104 }
105}
106
107pte_t *get_prealloc_pte(unsigned long pfn)
108{
109 int node = pfn_to_nid(pfn);
110 pfn &= ~(-1UL << (NR_PA_HIGHBIT_SHIFT - PAGE_SHIFT));
111 BUG_ON(node >= MAX_NUMNODES);
112 BUG_ON(pfn >= num_l2_ptes[node]);
113 return &l2_ptes[node][pfn];
114}
115
116/*
117 * What caching do we expect pages from the heap to have when
118 * they are allocated during bootup? (Once we've installed the
119 * "real" swapper_pg_dir.)
120 */
121static int initial_heap_home(void)
122{
123#if CHIP_HAS_CBOX_HOME_MAP()
124 if (hash_default)
125 return PAGE_HOME_HASH;
126#endif
127 return smp_processor_id();
128}
129
130/*
131 * Place a pointer to an L2 page table in a middle page
132 * directory entry.
133 */
134static void __init assign_pte(pmd_t *pmd, pte_t *page_table)
135{
136 phys_addr_t pa = __pa(page_table);
137 unsigned long l2_ptfn = pa >> HV_LOG2_PAGE_TABLE_ALIGN;
138 pte_t pteval = hv_pte_set_ptfn(__pgprot(_PAGE_TABLE), l2_ptfn);
139 BUG_ON((pa & (HV_PAGE_TABLE_ALIGN-1)) != 0);
140 pteval = pte_set_home(pteval, initial_heap_home());
141 *(pte_t *)pmd = pteval;
142 if (page_table != (pte_t *)pmd_page_vaddr(*pmd))
143 BUG();
144}
145
146#ifdef __tilegx__
147
148#if HV_L1_SIZE != HV_L2_SIZE
149# error Rework assumption that L1 and L2 page tables are same size.
150#endif
151
152/* Since pmd_t arrays and pte_t arrays are the same size, just use casts. */
153static inline pmd_t *alloc_pmd(void)
154{
155 return (pmd_t *)alloc_pte();
156}
157
158static inline void assign_pmd(pud_t *pud, pmd_t *pmd)
159{
160 assign_pte((pmd_t *)pud, (pte_t *)pmd);
161}
162
163#endif /* __tilegx__ */
164
165/* Replace the given pmd with a full PTE table. */
166void __init shatter_pmd(pmd_t *pmd)
167{
168 pte_t *pte = get_prealloc_pte(pte_pfn(*(pte_t *)pmd));
169 assign_pte(pmd, pte);
170}
171
172#ifdef CONFIG_HIGHMEM
173/*
174 * This function initializes a certain range of kernel virtual memory
175 * with new bootmem page tables, everywhere page tables are missing in
176 * the given range.
177 */
178
179/*
180 * NOTE: The pagetables are allocated contiguous on the physical space
181 * so we can cache the place of the first one and move around without
182 * checking the pgd every time.
183 */
184static void __init page_table_range_init(unsigned long start,
185 unsigned long end, pgd_t *pgd_base)
186{
187 pgd_t *pgd;
188 int pgd_idx;
189 unsigned long vaddr;
190
191 vaddr = start;
192 pgd_idx = pgd_index(vaddr);
193 pgd = pgd_base + pgd_idx;
194
195 for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
196 pmd_t *pmd = pmd_offset(pud_offset(pgd, vaddr), vaddr);
197 if (pmd_none(*pmd))
198 assign_pte(pmd, alloc_pte());
199 vaddr += PMD_SIZE;
200 }
201}
202#endif /* CONFIG_HIGHMEM */
203
204
205#if CHIP_HAS_CBOX_HOME_MAP()
206
207static int __initdata ktext_hash = 1; /* .text pages */
208static int __initdata kdata_hash = 1; /* .data and .bss pages */
209int __write_once hash_default = 1; /* kernel allocator pages */
210EXPORT_SYMBOL(hash_default);
211int __write_once kstack_hash = 1; /* if no homecaching, use h4h */
212#endif /* CHIP_HAS_CBOX_HOME_MAP */
213
214/*
215 * CPUs to use to for striping the pages of kernel data. If hash-for-home
216 * is available, this is only relevant if kcache_hash sets up the
217 * .data and .bss to be page-homed, and we don't want the default mode
218 * of using the full set of kernel cpus for the striping.
219 */
220static __initdata struct cpumask kdata_mask;
221static __initdata int kdata_arg_seen;
222
223int __write_once kdata_huge; /* if no homecaching, small pages */
224
225
226/* Combine a generic pgprot_t with cache home to get a cache-aware pgprot. */
227static pgprot_t __init construct_pgprot(pgprot_t prot, int home)
228{
229 prot = pte_set_home(prot, home);
230#if CHIP_HAS_CBOX_HOME_MAP()
231 if (home == PAGE_HOME_IMMUTABLE) {
232 if (ktext_hash)
233 prot = hv_pte_set_mode(prot, HV_PTE_MODE_CACHE_HASH_L3);
234 else
235 prot = hv_pte_set_mode(prot, HV_PTE_MODE_CACHE_NO_L3);
236 }
237#endif
238 return prot;
239}
240
241/*
242 * For a given kernel data VA, how should it be cached?
243 * We return the complete pgprot_t with caching bits set.
244 */
245static pgprot_t __init init_pgprot(ulong address)
246{
247 int cpu;
248 unsigned long page;
249 enum { CODE_DELTA = MEM_SV_INTRPT - PAGE_OFFSET };
250
251#if CHIP_HAS_CBOX_HOME_MAP()
252 /* For kdata=huge, everything is just hash-for-home. */
253 if (kdata_huge)
254 return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
255#endif
256
257 /* We map the aliased pages of permanent text inaccessible. */
258 if (address < (ulong) _sinittext - CODE_DELTA)
259 return PAGE_NONE;
260
261 /*
262 * We map read-only data non-coherent for performance. We could
263 * use neighborhood caching on TILE64, but it's not clear it's a win.
264 */
265 if ((address >= (ulong) __start_rodata &&
266 address < (ulong) __end_rodata) ||
267 address == (ulong) empty_zero_page) {
268 return construct_pgprot(PAGE_KERNEL_RO, PAGE_HOME_IMMUTABLE);
269 }
270
271 /* As a performance optimization, keep the boot init stack here. */
272 if (address >= (ulong)&init_thread_union &&
273 address < (ulong)&init_thread_union + THREAD_SIZE)
274 return construct_pgprot(PAGE_KERNEL, smp_processor_id());
275
276#ifndef __tilegx__
277#if !ATOMIC_LOCKS_FOUND_VIA_TABLE()
278 /* Force the atomic_locks[] array page to be hash-for-home. */
279 if (address == (ulong) atomic_locks)
280 return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
281#endif
282#endif
283
284 /*
285 * Everything else that isn't data or bss is heap, so mark it
286 * with the initial heap home (hash-for-home, or this cpu). This
Chris Metcalf0707ad32010-06-25 17:04:17 -0400287 * includes any addresses after the loaded image and any address before
288 * _einitdata, since we already captured the case of text before
289 * _sinittext, and __pa(einittext) is approximately __pa(sinitdata).
Chris Metcalf867e3592010-05-28 23:09:12 -0400290 *
291 * All the LOWMEM pages that we mark this way will get their
292 * struct page homecache properly marked later, in set_page_homes().
293 * The HIGHMEM pages we leave with a default zero for their
294 * homes, but with a zero free_time we don't have to actually
295 * do a flush action the first time we use them, either.
296 */
Chris Metcalf0707ad32010-06-25 17:04:17 -0400297 if (address >= (ulong) _end || address < (ulong) _einitdata)
Chris Metcalf867e3592010-05-28 23:09:12 -0400298 return construct_pgprot(PAGE_KERNEL, initial_heap_home());
299
300#if CHIP_HAS_CBOX_HOME_MAP()
301 /* Use hash-for-home if requested for data/bss. */
302 if (kdata_hash)
303 return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
304#endif
305
306 /*
Chris Metcalf0707ad32010-06-25 17:04:17 -0400307 * Make the w1data homed like heap to start with, to avoid
308 * making it part of the page-striped data area when we're just
309 * going to convert it to read-only soon anyway.
310 */
311 if (address >= (ulong)__w1data_begin && address < (ulong)__w1data_end)
312 return construct_pgprot(PAGE_KERNEL, initial_heap_home());
313
314 /*
Chris Metcalf867e3592010-05-28 23:09:12 -0400315 * Otherwise we just hand out consecutive cpus. To avoid
316 * requiring this function to hold state, we just walk forward from
317 * _sdata by PAGE_SIZE, skipping the readonly and init data, to reach
318 * the requested address, while walking cpu home around kdata_mask.
319 * This is typically no more than a dozen or so iterations.
320 */
Chris Metcalf0707ad32010-06-25 17:04:17 -0400321 page = (((ulong)__w1data_end) + PAGE_SIZE - 1) & PAGE_MASK;
322 BUG_ON(address < page || address >= (ulong)_end);
323 cpu = cpumask_first(&kdata_mask);
324 for (; page < address; page += PAGE_SIZE) {
325 if (page >= (ulong)&init_thread_union &&
326 page < (ulong)&init_thread_union + THREAD_SIZE)
327 continue;
Chris Metcalf867e3592010-05-28 23:09:12 -0400328 if (page == (ulong)empty_zero_page)
Chris Metcalf0707ad32010-06-25 17:04:17 -0400329 continue;
Chris Metcalf867e3592010-05-28 23:09:12 -0400330#ifndef __tilegx__
331#if !ATOMIC_LOCKS_FOUND_VIA_TABLE()
332 if (page == (ulong)atomic_locks)
Chris Metcalf0707ad32010-06-25 17:04:17 -0400333 continue;
Chris Metcalf867e3592010-05-28 23:09:12 -0400334#endif
335#endif
Chris Metcalf0707ad32010-06-25 17:04:17 -0400336 cpu = cpumask_next(cpu, &kdata_mask);
337 if (cpu == NR_CPUS)
338 cpu = cpumask_first(&kdata_mask);
Chris Metcalf867e3592010-05-28 23:09:12 -0400339 }
340 return construct_pgprot(PAGE_KERNEL, cpu);
341}
342
343/*
344 * This function sets up how we cache the kernel text. If we have
345 * hash-for-home support, normally that is used instead (see the
346 * kcache_hash boot flag for more information). But if we end up
347 * using a page-based caching technique, this option sets up the
348 * details of that. In addition, the "ktext=nocache" option may
349 * always be used to disable local caching of text pages, if desired.
350 */
351
352static int __initdata ktext_arg_seen;
353static int __initdata ktext_small;
354static int __initdata ktext_local;
355static int __initdata ktext_all;
356static int __initdata ktext_nondataplane;
357static int __initdata ktext_nocache;
358static struct cpumask __initdata ktext_mask;
359
360static int __init setup_ktext(char *str)
361{
362 if (str == NULL)
363 return -EINVAL;
364
365 /* If you have a leading "nocache", turn off ktext caching */
366 if (strncmp(str, "nocache", 7) == 0) {
367 ktext_nocache = 1;
Chris Metcalf0707ad32010-06-25 17:04:17 -0400368 pr_info("ktext: disabling local caching of kernel text\n");
Chris Metcalf867e3592010-05-28 23:09:12 -0400369 str += 7;
370 if (*str == ',')
371 ++str;
372 if (*str == '\0')
373 return 0;
374 }
375
376 ktext_arg_seen = 1;
377
378 /* Default setting on Tile64: use a huge page */
379 if (strcmp(str, "huge") == 0)
Chris Metcalf0707ad32010-06-25 17:04:17 -0400380 pr_info("ktext: using one huge locally cached page\n");
Chris Metcalf867e3592010-05-28 23:09:12 -0400381
382 /* Pay TLB cost but get no cache benefit: cache small pages locally */
383 else if (strcmp(str, "local") == 0) {
384 ktext_small = 1;
385 ktext_local = 1;
Chris Metcalf0707ad32010-06-25 17:04:17 -0400386 pr_info("ktext: using small pages with local caching\n");
Chris Metcalf867e3592010-05-28 23:09:12 -0400387 }
388
389 /* Neighborhood cache ktext pages on all cpus. */
390 else if (strcmp(str, "all") == 0) {
391 ktext_small = 1;
392 ktext_all = 1;
Chris Metcalf0707ad32010-06-25 17:04:17 -0400393 pr_info("ktext: using maximal caching neighborhood\n");
Chris Metcalf867e3592010-05-28 23:09:12 -0400394 }
395
396
397 /* Neighborhood ktext pages on specified mask */
398 else if (cpulist_parse(str, &ktext_mask) == 0) {
399 char buf[NR_CPUS * 5];
400 cpulist_scnprintf(buf, sizeof(buf), &ktext_mask);
401 if (cpumask_weight(&ktext_mask) > 1) {
402 ktext_small = 1;
Chris Metcalf0707ad32010-06-25 17:04:17 -0400403 pr_info("ktext: using caching neighborhood %s "
Chris Metcalf867e3592010-05-28 23:09:12 -0400404 "with small pages\n", buf);
405 } else {
Chris Metcalf0707ad32010-06-25 17:04:17 -0400406 pr_info("ktext: caching on cpu %s with one huge page\n",
Chris Metcalf867e3592010-05-28 23:09:12 -0400407 buf);
408 }
409 }
410
411 else if (*str)
412 return -EINVAL;
413
414 return 0;
415}
416
417early_param("ktext", setup_ktext);
418
419
420static inline pgprot_t ktext_set_nocache(pgprot_t prot)
421{
422 if (!ktext_nocache)
423 prot = hv_pte_set_nc(prot);
424#if CHIP_HAS_NC_AND_NOALLOC_BITS()
425 else
426 prot = hv_pte_set_no_alloc_l2(prot);
427#endif
428 return prot;
429}
430
431#ifndef __tilegx__
432static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
433{
434 return pmd_offset(pud_offset(&pgtables[pgd_index(va)], va), va);
435}
436#else
437static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
438{
439 pud_t *pud = pud_offset(&pgtables[pgd_index(va)], va);
440 if (pud_none(*pud))
441 assign_pmd(pud, alloc_pmd());
442 return pmd_offset(pud, va);
443}
444#endif
445
446/* Temporary page table we use for staging. */
447static pgd_t pgtables[PTRS_PER_PGD]
448 __attribute__((section(".init.page")));
449
450/*
451 * This maps the physical memory to kernel virtual address space, a total
452 * of max_low_pfn pages, by creating page tables starting from address
453 * PAGE_OFFSET.
454 *
455 * This routine transitions us from using a set of compiled-in large
456 * pages to using some more precise caching, including removing access
457 * to code pages mapped at PAGE_OFFSET (executed only at MEM_SV_START)
458 * marking read-only data as locally cacheable, striping the remaining
459 * .data and .bss across all the available tiles, and removing access
460 * to pages above the top of RAM (thus ensuring a page fault from a bad
461 * virtual address rather than a hypervisor shoot down for accessing
462 * memory outside the assigned limits).
463 */
464static void __init kernel_physical_mapping_init(pgd_t *pgd_base)
465{
466 unsigned long address, pfn;
467 pmd_t *pmd;
468 pte_t *pte;
469 int pte_ofs;
470 const struct cpumask *my_cpu_mask = cpumask_of(smp_processor_id());
471 struct cpumask kstripe_mask;
472 int rc, i;
473
474#if CHIP_HAS_CBOX_HOME_MAP()
475 if (ktext_arg_seen && ktext_hash) {
Chris Metcalf0707ad32010-06-25 17:04:17 -0400476 pr_warning("warning: \"ktext\" boot argument ignored"
477 " if \"kcache_hash\" sets up text hash-for-home\n");
Chris Metcalf867e3592010-05-28 23:09:12 -0400478 ktext_small = 0;
479 }
480
481 if (kdata_arg_seen && kdata_hash) {
Chris Metcalf0707ad32010-06-25 17:04:17 -0400482 pr_warning("warning: \"kdata\" boot argument ignored"
483 " if \"kcache_hash\" sets up data hash-for-home\n");
Chris Metcalf867e3592010-05-28 23:09:12 -0400484 }
485
486 if (kdata_huge && !hash_default) {
Chris Metcalf0707ad32010-06-25 17:04:17 -0400487 pr_warning("warning: disabling \"kdata=huge\"; requires"
488 " kcache_hash=all or =allbutstack\n");
Chris Metcalf867e3592010-05-28 23:09:12 -0400489 kdata_huge = 0;
490 }
491#endif
492
493 /*
494 * Set up a mask for cpus to use for kernel striping.
495 * This is normally all cpus, but minus dataplane cpus if any.
496 * If the dataplane covers the whole chip, we stripe over
497 * the whole chip too.
498 */
499 cpumask_copy(&kstripe_mask, cpu_possible_mask);
500 if (!kdata_arg_seen)
501 kdata_mask = kstripe_mask;
502
503 /* Allocate and fill in L2 page tables */
504 for (i = 0; i < MAX_NUMNODES; ++i) {
505#ifdef CONFIG_HIGHMEM
506 unsigned long end_pfn = node_lowmem_end_pfn[i];
507#else
508 unsigned long end_pfn = node_end_pfn[i];
509#endif
510 unsigned long end_huge_pfn = 0;
511
512 /* Pre-shatter the last huge page to allow per-cpu pages. */
513 if (kdata_huge)
514 end_huge_pfn = end_pfn - (HPAGE_SIZE >> PAGE_SHIFT);
515
516 pfn = node_start_pfn[i];
517
518 /* Allocate enough memory to hold L2 page tables for node. */
519 init_prealloc_ptes(i, end_pfn - pfn);
520
521 address = (unsigned long) pfn_to_kaddr(pfn);
522 while (pfn < end_pfn) {
523 BUG_ON(address & (HPAGE_SIZE-1));
524 pmd = get_pmd(pgtables, address);
525 pte = get_prealloc_pte(pfn);
526 if (pfn < end_huge_pfn) {
527 pgprot_t prot = init_pgprot(address);
528 *(pte_t *)pmd = pte_mkhuge(pfn_pte(pfn, prot));
529 for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE;
530 pfn++, pte_ofs++, address += PAGE_SIZE)
531 pte[pte_ofs] = pfn_pte(pfn, prot);
532 } else {
533 if (kdata_huge)
534 printk(KERN_DEBUG "pre-shattered huge"
535 " page at %#lx\n", address);
536 for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE;
537 pfn++, pte_ofs++, address += PAGE_SIZE) {
538 pgprot_t prot = init_pgprot(address);
539 pte[pte_ofs] = pfn_pte(pfn, prot);
540 }
541 assign_pte(pmd, pte);
542 }
543 }
544 }
545
546 /*
547 * Set or check ktext_map now that we have cpu_possible_mask
548 * and kstripe_mask to work with.
549 */
550 if (ktext_all)
551 cpumask_copy(&ktext_mask, cpu_possible_mask);
552 else if (ktext_nondataplane)
553 ktext_mask = kstripe_mask;
554 else if (!cpumask_empty(&ktext_mask)) {
555 /* Sanity-check any mask that was requested */
556 struct cpumask bad;
557 cpumask_andnot(&bad, &ktext_mask, cpu_possible_mask);
558 cpumask_and(&ktext_mask, &ktext_mask, cpu_possible_mask);
559 if (!cpumask_empty(&bad)) {
560 char buf[NR_CPUS * 5];
561 cpulist_scnprintf(buf, sizeof(buf), &bad);
Chris Metcalf0707ad32010-06-25 17:04:17 -0400562 pr_info("ktext: not using unavailable cpus %s\n", buf);
Chris Metcalf867e3592010-05-28 23:09:12 -0400563 }
564 if (cpumask_empty(&ktext_mask)) {
Chris Metcalf0707ad32010-06-25 17:04:17 -0400565 pr_warning("ktext: no valid cpus; caching on %d.\n",
566 smp_processor_id());
Chris Metcalf867e3592010-05-28 23:09:12 -0400567 cpumask_copy(&ktext_mask,
568 cpumask_of(smp_processor_id()));
569 }
570 }
571
572 address = MEM_SV_INTRPT;
573 pmd = get_pmd(pgtables, address);
574 if (ktext_small) {
575 /* Allocate an L2 PTE for the kernel text */
576 int cpu = 0;
577 pgprot_t prot = construct_pgprot(PAGE_KERNEL_EXEC,
578 PAGE_HOME_IMMUTABLE);
579
580 if (ktext_local) {
581 if (ktext_nocache)
582 prot = hv_pte_set_mode(prot,
583 HV_PTE_MODE_UNCACHED);
584 else
585 prot = hv_pte_set_mode(prot,
586 HV_PTE_MODE_CACHE_NO_L3);
587 } else {
588 prot = hv_pte_set_mode(prot,
589 HV_PTE_MODE_CACHE_TILE_L3);
590 cpu = cpumask_first(&ktext_mask);
591
592 prot = ktext_set_nocache(prot);
593 }
594
595 BUG_ON(address != (unsigned long)_stext);
596 pfn = 0; /* code starts at PA 0 */
597 pte = alloc_pte();
598 for (pte_ofs = 0; address < (unsigned long)_einittext;
599 pfn++, pte_ofs++, address += PAGE_SIZE) {
600 if (!ktext_local) {
601 prot = set_remote_cache_cpu(prot, cpu);
602 cpu = cpumask_next(cpu, &ktext_mask);
603 if (cpu == NR_CPUS)
604 cpu = cpumask_first(&ktext_mask);
605 }
606 pte[pte_ofs] = pfn_pte(pfn, prot);
607 }
608 assign_pte(pmd, pte);
609 } else {
610 pte_t pteval = pfn_pte(0, PAGE_KERNEL_EXEC);
611 pteval = pte_mkhuge(pteval);
612#if CHIP_HAS_CBOX_HOME_MAP()
613 if (ktext_hash) {
614 pteval = hv_pte_set_mode(pteval,
615 HV_PTE_MODE_CACHE_HASH_L3);
616 pteval = ktext_set_nocache(pteval);
617 } else
618#endif /* CHIP_HAS_CBOX_HOME_MAP() */
619 if (cpumask_weight(&ktext_mask) == 1) {
620 pteval = set_remote_cache_cpu(pteval,
621 cpumask_first(&ktext_mask));
622 pteval = hv_pte_set_mode(pteval,
623 HV_PTE_MODE_CACHE_TILE_L3);
624 pteval = ktext_set_nocache(pteval);
625 } else if (ktext_nocache)
626 pteval = hv_pte_set_mode(pteval,
627 HV_PTE_MODE_UNCACHED);
628 else
629 pteval = hv_pte_set_mode(pteval,
630 HV_PTE_MODE_CACHE_NO_L3);
631 *(pte_t *)pmd = pteval;
632 }
633
634 /* Set swapper_pgprot here so it is flushed to memory right away. */
635 swapper_pgprot = init_pgprot((unsigned long)swapper_pg_dir);
636
637 /*
638 * Since we may be changing the caching of the stack and page
639 * table itself, we invoke an assembly helper to do the
640 * following steps:
641 *
642 * - flush the cache so we start with an empty slate
643 * - install pgtables[] as the real page table
644 * - flush the TLB so the new page table takes effect
645 */
646 rc = flush_and_install_context(__pa(pgtables),
647 init_pgprot((unsigned long)pgtables),
648 __get_cpu_var(current_asid),
649 cpumask_bits(my_cpu_mask));
650 BUG_ON(rc != 0);
651
652 /* Copy the page table back to the normal swapper_pg_dir. */
653 memcpy(pgd_base, pgtables, sizeof(pgtables));
654 __install_page_table(pgd_base, __get_cpu_var(current_asid),
655 swapper_pgprot);
656}
657
658/*
659 * devmem_is_allowed() checks to see if /dev/mem access to a certain address
660 * is valid. The argument is a physical page number.
661 *
662 * On Tile, the only valid things for which we can just hand out unchecked
663 * PTEs are the kernel code and data. Anything else might change its
664 * homing with time, and we wouldn't know to adjust the /dev/mem PTEs.
665 * Note that init_thread_union is released to heap soon after boot,
666 * so we include it in the init data.
667 *
668 * For TILE-Gx, we might want to consider allowing access to PA
669 * regions corresponding to PCI space, etc.
670 */
671int devmem_is_allowed(unsigned long pagenr)
672{
673 return pagenr < kaddr_to_pfn(_end) &&
674 !(pagenr >= kaddr_to_pfn(&init_thread_union) ||
675 pagenr < kaddr_to_pfn(_einitdata)) &&
676 !(pagenr >= kaddr_to_pfn(_sinittext) ||
677 pagenr <= kaddr_to_pfn(_einittext-1));
678}
679
680#ifdef CONFIG_HIGHMEM
681static void __init permanent_kmaps_init(pgd_t *pgd_base)
682{
683 pgd_t *pgd;
684 pud_t *pud;
685 pmd_t *pmd;
686 pte_t *pte;
687 unsigned long vaddr;
688
689 vaddr = PKMAP_BASE;
690 page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
691
692 pgd = swapper_pg_dir + pgd_index(vaddr);
693 pud = pud_offset(pgd, vaddr);
694 pmd = pmd_offset(pud, vaddr);
695 pte = pte_offset_kernel(pmd, vaddr);
696 pkmap_page_table = pte;
697}
698#endif /* CONFIG_HIGHMEM */
699
700
701static void __init init_free_pfn_range(unsigned long start, unsigned long end)
702{
703 unsigned long pfn;
704 struct page *page = pfn_to_page(start);
705
706 for (pfn = start; pfn < end; ) {
707 /* Optimize by freeing pages in large batches */
708 int order = __ffs(pfn);
709 int count, i;
710 struct page *p;
711
712 if (order >= MAX_ORDER)
713 order = MAX_ORDER-1;
714 count = 1 << order;
715 while (pfn + count > end) {
716 count >>= 1;
717 --order;
718 }
719 for (p = page, i = 0; i < count; ++i, ++p) {
720 __ClearPageReserved(p);
721 /*
722 * Hacky direct set to avoid unnecessary
723 * lock take/release for EVERY page here.
724 */
725 p->_count.counter = 0;
726 p->_mapcount.counter = -1;
727 }
728 init_page_count(page);
729 __free_pages(page, order);
730 totalram_pages += count;
731
732 page += count;
733 pfn += count;
734 }
735}
736
737static void __init set_non_bootmem_pages_init(void)
738{
739 struct zone *z;
740 for_each_zone(z) {
741 unsigned long start, end;
742 int nid = z->zone_pgdat->node_id;
Chris Metcalf0707ad32010-06-25 17:04:17 -0400743 int idx = zone_idx(z);
Chris Metcalf867e3592010-05-28 23:09:12 -0400744
745 start = z->zone_start_pfn;
746 if (start == 0)
747 continue; /* bootmem */
748 end = start + z->spanned_pages;
Chris Metcalf0707ad32010-06-25 17:04:17 -0400749 if (idx == ZONE_NORMAL) {
Chris Metcalf867e3592010-05-28 23:09:12 -0400750 BUG_ON(start != node_start_pfn[nid]);
751 start = node_free_pfn[nid];
752 }
753#ifdef CONFIG_HIGHMEM
Chris Metcalf0707ad32010-06-25 17:04:17 -0400754 if (idx == ZONE_HIGHMEM)
Chris Metcalf867e3592010-05-28 23:09:12 -0400755 totalhigh_pages += z->spanned_pages;
756#endif
757 if (kdata_huge) {
758 unsigned long percpu_pfn = node_percpu_pfn[nid];
759 if (start < percpu_pfn && end > percpu_pfn)
760 end = percpu_pfn;
761 }
762#ifdef CONFIG_PCI
763 if (start <= pci_reserve_start_pfn &&
764 end > pci_reserve_start_pfn) {
765 if (end > pci_reserve_end_pfn)
766 init_free_pfn_range(pci_reserve_end_pfn, end);
767 end = pci_reserve_start_pfn;
768 }
769#endif
770 init_free_pfn_range(start, end);
771 }
772}
773
774/*
775 * paging_init() sets up the page tables - note that all of lowmem is
776 * already mapped by head.S.
777 */
778void __init paging_init(void)
779{
780#ifdef CONFIG_HIGHMEM
781 unsigned long vaddr, end;
782#endif
783#ifdef __tilegx__
784 pud_t *pud;
785#endif
786 pgd_t *pgd_base = swapper_pg_dir;
787
788 kernel_physical_mapping_init(pgd_base);
789
790#ifdef CONFIG_HIGHMEM
791 /*
792 * Fixed mappings, only the page table structure has to be
793 * created - mappings will be set by set_fixmap():
794 */
795 vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
796 end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
797 page_table_range_init(vaddr, end, pgd_base);
798 permanent_kmaps_init(pgd_base);
799#endif
800
801#ifdef __tilegx__
802 /*
803 * Since GX allocates just one pmd_t array worth of vmalloc space,
804 * we go ahead and allocate it statically here, then share it
805 * globally. As a result we don't have to worry about any task
806 * changing init_mm once we get up and running, and there's no
807 * need for e.g. vmalloc_sync_all().
808 */
809 BUILD_BUG_ON(pgd_index(VMALLOC_START) != pgd_index(VMALLOC_END));
810 pud = pud_offset(pgd_base + pgd_index(VMALLOC_START), VMALLOC_START);
811 assign_pmd(pud, alloc_pmd());
812#endif
813}
814
815
816/*
817 * Walk the kernel page tables and derive the page_home() from
818 * the PTEs, so that set_pte() can properly validate the caching
819 * of all PTEs it sees.
820 */
821void __init set_page_homes(void)
822{
823}
824
825static void __init set_max_mapnr_init(void)
826{
827#ifdef CONFIG_FLATMEM
828 max_mapnr = max_low_pfn;
829#endif
830}
831
832void __init mem_init(void)
833{
834 int codesize, datasize, initsize;
835 int i;
836#ifndef __tilegx__
837 void *last;
838#endif
839
840#ifdef CONFIG_FLATMEM
841 if (!mem_map)
842 BUG();
843#endif
844
845#ifdef CONFIG_HIGHMEM
846 /* check that fixmap and pkmap do not overlap */
847 if (PKMAP_ADDR(LAST_PKMAP-1) >= FIXADDR_START) {
Chris Metcalf0707ad32010-06-25 17:04:17 -0400848 pr_err("fixmap and kmap areas overlap"
Chris Metcalf867e3592010-05-28 23:09:12 -0400849 " - this will crash\n");
Chris Metcalf0707ad32010-06-25 17:04:17 -0400850 pr_err("pkstart: %lxh pkend: %lxh fixstart %lxh\n",
Chris Metcalf867e3592010-05-28 23:09:12 -0400851 PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP-1),
852 FIXADDR_START);
853 BUG();
854 }
855#endif
856
857 set_max_mapnr_init();
858
859 /* this will put all bootmem onto the freelists */
860 totalram_pages += free_all_bootmem();
861
862 /* count all remaining LOWMEM and give all HIGHMEM to page allocator */
863 set_non_bootmem_pages_init();
864
865 codesize = (unsigned long)&_etext - (unsigned long)&_text;
866 datasize = (unsigned long)&_end - (unsigned long)&_sdata;
867 initsize = (unsigned long)&_einittext - (unsigned long)&_sinittext;
868 initsize += (unsigned long)&_einitdata - (unsigned long)&_sinitdata;
869
Chris Metcalf0707ad32010-06-25 17:04:17 -0400870 pr_info("Memory: %luk/%luk available (%dk kernel code, %dk data, %dk init, %ldk highmem)\n",
Chris Metcalf867e3592010-05-28 23:09:12 -0400871 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
872 num_physpages << (PAGE_SHIFT-10),
873 codesize >> 10,
874 datasize >> 10,
875 initsize >> 10,
876 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
877 );
878
879 /*
880 * In debug mode, dump some interesting memory mappings.
881 */
882#ifdef CONFIG_HIGHMEM
883 printk(KERN_DEBUG " KMAP %#lx - %#lx\n",
884 FIXADDR_START, FIXADDR_TOP + PAGE_SIZE - 1);
885 printk(KERN_DEBUG " PKMAP %#lx - %#lx\n",
886 PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP) - 1);
887#endif
888#ifdef CONFIG_HUGEVMAP
889 printk(KERN_DEBUG " HUGEMAP %#lx - %#lx\n",
890 HUGE_VMAP_BASE, HUGE_VMAP_END - 1);
891#endif
892 printk(KERN_DEBUG " VMALLOC %#lx - %#lx\n",
893 _VMALLOC_START, _VMALLOC_END - 1);
894#ifdef __tilegx__
895 for (i = MAX_NUMNODES-1; i >= 0; --i) {
896 struct pglist_data *node = &node_data[i];
897 if (node->node_present_pages) {
898 unsigned long start = (unsigned long)
899 pfn_to_kaddr(node->node_start_pfn);
900 unsigned long end = start +
901 (node->node_present_pages << PAGE_SHIFT);
902 printk(KERN_DEBUG " MEM%d %#lx - %#lx\n",
903 i, start, end - 1);
904 }
905 }
906#else
907 last = high_memory;
908 for (i = MAX_NUMNODES-1; i >= 0; --i) {
909 if ((unsigned long)vbase_map[i] != -1UL) {
910 printk(KERN_DEBUG " LOWMEM%d %#lx - %#lx\n",
911 i, (unsigned long) (vbase_map[i]),
912 (unsigned long) (last-1));
913 last = vbase_map[i];
914 }
915 }
916#endif
917
918#ifndef __tilegx__
919 /*
920 * Convert from using one lock for all atomic operations to
921 * one per cpu.
922 */
923 __init_atomic_per_cpu();
924#endif
925}
926
927/*
928 * this is for the non-NUMA, single node SMP system case.
929 * Specifically, in the case of x86, we will always add
930 * memory to the highmem for now.
931 */
932#ifndef CONFIG_NEED_MULTIPLE_NODES
933int arch_add_memory(u64 start, u64 size)
934{
935 struct pglist_data *pgdata = &contig_page_data;
936 struct zone *zone = pgdata->node_zones + MAX_NR_ZONES-1;
937 unsigned long start_pfn = start >> PAGE_SHIFT;
938 unsigned long nr_pages = size >> PAGE_SHIFT;
939
940 return __add_pages(zone, start_pfn, nr_pages);
941}
942
943int remove_memory(u64 start, u64 size)
944{
945 return -EINVAL;
946}
947#endif
948
949struct kmem_cache *pgd_cache;
950
951void __init pgtable_cache_init(void)
952{
953 pgd_cache = kmem_cache_create("pgd",
954 PTRS_PER_PGD*sizeof(pgd_t),
955 PTRS_PER_PGD*sizeof(pgd_t),
956 0,
957 NULL);
958 if (!pgd_cache)
959 panic("pgtable_cache_init(): Cannot create pgd cache");
960}
961
962#if !CHIP_HAS_COHERENT_LOCAL_CACHE()
963/*
964 * The __w1data area holds data that is only written during initialization,
965 * and is read-only and thus freely cacheable thereafter. Fix the page
966 * table entries that cover that region accordingly.
967 */
968static void mark_w1data_ro(void)
969{
970 /* Loop over page table entries */
971 unsigned long addr = (unsigned long)__w1data_begin;
972 BUG_ON((addr & (PAGE_SIZE-1)) != 0);
973 for (; addr <= (unsigned long)__w1data_end - 1; addr += PAGE_SIZE) {
974 unsigned long pfn = kaddr_to_pfn((void *)addr);
Chris Metcalf867e3592010-05-28 23:09:12 -0400975 pte_t *ptep = virt_to_pte(NULL, addr);
976 BUG_ON(pte_huge(*ptep)); /* not relevant for kdata_huge */
977 set_pte_at(&init_mm, addr, ptep, pfn_pte(pfn, PAGE_KERNEL_RO));
978 }
979}
980#endif
981
982#ifdef CONFIG_DEBUG_PAGEALLOC
983static long __write_once initfree;
984#else
985static long __write_once initfree = 1;
986#endif
987
988/* Select whether to free (1) or mark unusable (0) the __init pages. */
989static int __init set_initfree(char *str)
990{
991 strict_strtol(str, 0, &initfree);
Chris Metcalf0707ad32010-06-25 17:04:17 -0400992 pr_info("initfree: %s free init pages\n", initfree ? "will" : "won't");
Chris Metcalf867e3592010-05-28 23:09:12 -0400993 return 1;
994}
995__setup("initfree=", set_initfree);
996
997static void free_init_pages(char *what, unsigned long begin, unsigned long end)
998{
999 unsigned long addr = (unsigned long) begin;
1000
1001 if (kdata_huge && !initfree) {
Chris Metcalf0707ad32010-06-25 17:04:17 -04001002 pr_warning("Warning: ignoring initfree=0:"
1003 " incompatible with kdata=huge\n");
Chris Metcalf867e3592010-05-28 23:09:12 -04001004 initfree = 1;
1005 }
1006 end = (end + PAGE_SIZE - 1) & PAGE_MASK;
1007 local_flush_tlb_pages(NULL, begin, PAGE_SIZE, end - begin);
1008 for (addr = begin; addr < end; addr += PAGE_SIZE) {
1009 /*
1010 * Note we just reset the home here directly in the
1011 * page table. We know this is safe because our caller
1012 * just flushed the caches on all the other cpus,
1013 * and they won't be touching any of these pages.
1014 */
1015 int pfn = kaddr_to_pfn((void *)addr);
1016 struct page *page = pfn_to_page(pfn);
1017 pte_t *ptep = virt_to_pte(NULL, addr);
1018 if (!initfree) {
1019 /*
1020 * If debugging page accesses then do not free
1021 * this memory but mark them not present - any
1022 * buggy init-section access will create a
1023 * kernel page fault:
1024 */
1025 pte_clear(&init_mm, addr, ptep);
1026 continue;
1027 }
1028 __ClearPageReserved(page);
1029 init_page_count(page);
1030 if (pte_huge(*ptep))
1031 BUG_ON(!kdata_huge);
1032 else
1033 set_pte_at(&init_mm, addr, ptep,
1034 pfn_pte(pfn, PAGE_KERNEL));
1035 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
1036 free_page(addr);
1037 totalram_pages++;
1038 }
Chris Metcalf0707ad32010-06-25 17:04:17 -04001039 pr_info("Freeing %s: %ldk freed\n", what, (end - begin) >> 10);
Chris Metcalf867e3592010-05-28 23:09:12 -04001040}
1041
1042void free_initmem(void)
1043{
1044 const unsigned long text_delta = MEM_SV_INTRPT - PAGE_OFFSET;
1045
1046 /*
1047 * Evict the dirty initdata on the boot cpu, evict the w1data
1048 * wherever it's homed, and evict all the init code everywhere.
1049 * We are guaranteed that no one will touch the init pages any
1050 * more, and although other cpus may be touching the w1data,
1051 * we only actually change the caching on tile64, which won't
1052 * be keeping local copies in the other tiles' caches anyway.
1053 */
1054 homecache_evict(&cpu_cacheable_map);
1055
1056 /* Free the data pages that we won't use again after init. */
1057 free_init_pages("unused kernel data",
1058 (unsigned long)_sinitdata,
1059 (unsigned long)_einitdata);
1060
1061 /*
1062 * Free the pages mapped from 0xc0000000 that correspond to code
Chris Metcalfa78c9422010-10-14 16:23:03 -04001063 * pages from MEM_SV_INTRPT that we won't use again after init.
Chris Metcalf867e3592010-05-28 23:09:12 -04001064 */
1065 free_init_pages("unused kernel text",
1066 (unsigned long)_sinittext - text_delta,
1067 (unsigned long)_einittext - text_delta);
1068
1069#if !CHIP_HAS_COHERENT_LOCAL_CACHE()
1070 /*
1071 * Upgrade the .w1data section to globally cached.
1072 * We don't do this on tilepro, since the cache architecture
1073 * pretty much makes it irrelevant, and in any case we end
1074 * up having racing issues with other tiles that may touch
1075 * the data after we flush the cache but before we update
1076 * the PTEs and flush the TLBs, causing sharer shootdowns
1077 * later. Even though this is to clean data, it seems like
1078 * an unnecessary complication.
1079 */
1080 mark_w1data_ro();
1081#endif
1082
1083 /* Do a global TLB flush so everyone sees the changes. */
1084 flush_tlb_all();
1085}