Commit | Line | Data |
---|---|---|
867e359b CM |
1 | /* |
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> | |
867e359b CM |
41 | #include <asm/pgtable.h> |
42 | #include <asm/pgalloc.h> | |
43 | #include <asm/dma.h> | |
44 | #include <asm/fixmap.h> | |
45 | #include <asm/tlb.h> | |
46 | #include <asm/tlbflush.h> | |
47 | #include <asm/sections.h> | |
48 | #include <asm/setup.h> | |
49 | #include <asm/homecache.h> | |
50 | #include <hv/hypervisor.h> | |
51 | #include <arch/chip.h> | |
52 | ||
53 | #include "migrate.h" | |
54 | ||
867e359b CM |
55 | #define clear_pgd(pmdptr) (*(pmdptr) = hv_pte(0)) |
56 | ||
0707ad30 | 57 | #ifndef __tilegx__ |
867e359b | 58 | unsigned long VMALLOC_RESERVE = CONFIG_VMALLOC_RESERVE; |
00dce031 | 59 | EXPORT_SYMBOL(VMALLOC_RESERVE); |
0707ad30 | 60 | #endif |
867e359b | 61 | |
867e359b CM |
62 | /* Create an L2 page table */ |
63 | static pte_t * __init alloc_pte(void) | |
64 | { | |
65 | return __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE, HV_PAGE_TABLE_ALIGN, 0); | |
66 | } | |
67 | ||
68 | /* | |
69 | * L2 page tables per controller. We allocate these all at once from | |
70 | * the bootmem allocator and store them here. This saves on kernel L2 | |
71 | * page table memory, compared to allocating a full 64K page per L2 | |
72 | * page table, and also means that in cases where we use huge pages, | |
73 | * we are guaranteed to later be able to shatter those huge pages and | |
74 | * switch to using these page tables instead, without requiring | |
75 | * further allocation. Each l2_ptes[] entry points to the first page | |
76 | * table for the first hugepage-size piece of memory on the | |
77 | * controller; other page tables are just indexed directly, i.e. the | |
78 | * L2 page tables are contiguous in memory for each controller. | |
79 | */ | |
80 | static pte_t *l2_ptes[MAX_NUMNODES]; | |
81 | static int num_l2_ptes[MAX_NUMNODES]; | |
82 | ||
83 | static void init_prealloc_ptes(int node, int pages) | |
84 | { | |
d5d14ed6 | 85 | BUG_ON(pages & (PTRS_PER_PTE - 1)); |
867e359b CM |
86 | if (pages) { |
87 | num_l2_ptes[node] = pages; | |
88 | l2_ptes[node] = __alloc_bootmem(pages * sizeof(pte_t), | |
89 | HV_PAGE_TABLE_ALIGN, 0); | |
90 | } | |
91 | } | |
92 | ||
93 | pte_t *get_prealloc_pte(unsigned long pfn) | |
94 | { | |
95 | int node = pfn_to_nid(pfn); | |
96 | pfn &= ~(-1UL << (NR_PA_HIGHBIT_SHIFT - PAGE_SHIFT)); | |
97 | BUG_ON(node >= MAX_NUMNODES); | |
98 | BUG_ON(pfn >= num_l2_ptes[node]); | |
99 | return &l2_ptes[node][pfn]; | |
100 | } | |
101 | ||
102 | /* | |
103 | * What caching do we expect pages from the heap to have when | |
104 | * they are allocated during bootup? (Once we've installed the | |
105 | * "real" swapper_pg_dir.) | |
106 | */ | |
107 | static int initial_heap_home(void) | |
108 | { | |
867e359b CM |
109 | if (hash_default) |
110 | return PAGE_HOME_HASH; | |
867e359b CM |
111 | return smp_processor_id(); |
112 | } | |
113 | ||
114 | /* | |
115 | * Place a pointer to an L2 page table in a middle page | |
116 | * directory entry. | |
117 | */ | |
118 | static void __init assign_pte(pmd_t *pmd, pte_t *page_table) | |
119 | { | |
120 | phys_addr_t pa = __pa(page_table); | |
121 | unsigned long l2_ptfn = pa >> HV_LOG2_PAGE_TABLE_ALIGN; | |
122 | pte_t pteval = hv_pte_set_ptfn(__pgprot(_PAGE_TABLE), l2_ptfn); | |
123 | BUG_ON((pa & (HV_PAGE_TABLE_ALIGN-1)) != 0); | |
124 | pteval = pte_set_home(pteval, initial_heap_home()); | |
125 | *(pte_t *)pmd = pteval; | |
126 | if (page_table != (pte_t *)pmd_page_vaddr(*pmd)) | |
127 | BUG(); | |
128 | } | |
129 | ||
130 | #ifdef __tilegx__ | |
131 | ||
867e359b CM |
132 | static inline pmd_t *alloc_pmd(void) |
133 | { | |
d5d14ed6 | 134 | return __alloc_bootmem(L1_KERNEL_PGTABLE_SIZE, HV_PAGE_TABLE_ALIGN, 0); |
867e359b CM |
135 | } |
136 | ||
137 | static inline void assign_pmd(pud_t *pud, pmd_t *pmd) | |
138 | { | |
139 | assign_pte((pmd_t *)pud, (pte_t *)pmd); | |
140 | } | |
141 | ||
142 | #endif /* __tilegx__ */ | |
143 | ||
144 | /* Replace the given pmd with a full PTE table. */ | |
145 | void __init shatter_pmd(pmd_t *pmd) | |
146 | { | |
147 | pte_t *pte = get_prealloc_pte(pte_pfn(*(pte_t *)pmd)); | |
148 | assign_pte(pmd, pte); | |
149 | } | |
150 | ||
bbaa22c3 CM |
151 | #ifdef __tilegx__ |
152 | static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va) | |
153 | { | |
154 | pud_t *pud = pud_offset(&pgtables[pgd_index(va)], va); | |
155 | if (pud_none(*pud)) | |
156 | assign_pmd(pud, alloc_pmd()); | |
157 | return pmd_offset(pud, va); | |
158 | } | |
159 | #else | |
160 | static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va) | |
161 | { | |
162 | return pmd_offset(pud_offset(&pgtables[pgd_index(va)], va), va); | |
163 | } | |
164 | #endif | |
165 | ||
867e359b CM |
166 | /* |
167 | * This function initializes a certain range of kernel virtual memory | |
168 | * with new bootmem page tables, everywhere page tables are missing in | |
169 | * the given range. | |
170 | */ | |
171 | ||
172 | /* | |
173 | * NOTE: The pagetables are allocated contiguous on the physical space | |
174 | * so we can cache the place of the first one and move around without | |
175 | * checking the pgd every time. | |
176 | */ | |
177 | static void __init page_table_range_init(unsigned long start, | |
bbaa22c3 | 178 | unsigned long end, pgd_t *pgd) |
867e359b | 179 | { |
867e359b | 180 | unsigned long vaddr; |
bbaa22c3 CM |
181 | start = round_down(start, PMD_SIZE); |
182 | end = round_up(end, PMD_SIZE); | |
183 | for (vaddr = start; vaddr < end; vaddr += PMD_SIZE) { | |
184 | pmd_t *pmd = get_pmd(pgd, vaddr); | |
867e359b CM |
185 | if (pmd_none(*pmd)) |
186 | assign_pte(pmd, alloc_pte()); | |
867e359b CM |
187 | } |
188 | } | |
867e359b CM |
189 | |
190 | ||
867e359b CM |
191 | static int __initdata ktext_hash = 1; /* .text pages */ |
192 | static int __initdata kdata_hash = 1; /* .data and .bss pages */ | |
193 | int __write_once hash_default = 1; /* kernel allocator pages */ | |
194 | EXPORT_SYMBOL(hash_default); | |
195 | int __write_once kstack_hash = 1; /* if no homecaching, use h4h */ | |
867e359b CM |
196 | |
197 | /* | |
198 | * CPUs to use to for striping the pages of kernel data. If hash-for-home | |
199 | * is available, this is only relevant if kcache_hash sets up the | |
200 | * .data and .bss to be page-homed, and we don't want the default mode | |
201 | * of using the full set of kernel cpus for the striping. | |
202 | */ | |
203 | static __initdata struct cpumask kdata_mask; | |
204 | static __initdata int kdata_arg_seen; | |
205 | ||
206 | int __write_once kdata_huge; /* if no homecaching, small pages */ | |
207 | ||
208 | ||
209 | /* Combine a generic pgprot_t with cache home to get a cache-aware pgprot. */ | |
210 | static pgprot_t __init construct_pgprot(pgprot_t prot, int home) | |
211 | { | |
212 | prot = pte_set_home(prot, home); | |
867e359b CM |
213 | if (home == PAGE_HOME_IMMUTABLE) { |
214 | if (ktext_hash) | |
215 | prot = hv_pte_set_mode(prot, HV_PTE_MODE_CACHE_HASH_L3); | |
216 | else | |
217 | prot = hv_pte_set_mode(prot, HV_PTE_MODE_CACHE_NO_L3); | |
218 | } | |
867e359b CM |
219 | return prot; |
220 | } | |
221 | ||
222 | /* | |
223 | * For a given kernel data VA, how should it be cached? | |
224 | * We return the complete pgprot_t with caching bits set. | |
225 | */ | |
226 | static pgprot_t __init init_pgprot(ulong address) | |
227 | { | |
228 | int cpu; | |
229 | unsigned long page; | |
acbde1db | 230 | enum { CODE_DELTA = MEM_SV_START - PAGE_OFFSET }; |
867e359b | 231 | |
867e359b CM |
232 | /* For kdata=huge, everything is just hash-for-home. */ |
233 | if (kdata_huge) | |
234 | return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH); | |
867e359b CM |
235 | |
236 | /* We map the aliased pages of permanent text inaccessible. */ | |
237 | if (address < (ulong) _sinittext - CODE_DELTA) | |
238 | return PAGE_NONE; | |
239 | ||
d7c96611 | 240 | /* We map read-only data non-coherent for performance. */ |
867e359b CM |
241 | if ((address >= (ulong) __start_rodata && |
242 | address < (ulong) __end_rodata) || | |
243 | address == (ulong) empty_zero_page) { | |
244 | return construct_pgprot(PAGE_KERNEL_RO, PAGE_HOME_IMMUTABLE); | |
245 | } | |
246 | ||
867e359b | 247 | #ifndef __tilegx__ |
867e359b CM |
248 | /* Force the atomic_locks[] array page to be hash-for-home. */ |
249 | if (address == (ulong) atomic_locks) | |
250 | return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH); | |
867e359b CM |
251 | #endif |
252 | ||
253 | /* | |
254 | * Everything else that isn't data or bss is heap, so mark it | |
255 | * with the initial heap home (hash-for-home, or this cpu). This | |
0707ad30 CM |
256 | * includes any addresses after the loaded image and any address before |
257 | * _einitdata, since we already captured the case of text before | |
258 | * _sinittext, and __pa(einittext) is approximately __pa(sinitdata). | |
867e359b CM |
259 | * |
260 | * All the LOWMEM pages that we mark this way will get their | |
261 | * struct page homecache properly marked later, in set_page_homes(). | |
262 | * The HIGHMEM pages we leave with a default zero for their | |
263 | * homes, but with a zero free_time we don't have to actually | |
264 | * do a flush action the first time we use them, either. | |
265 | */ | |
0707ad30 | 266 | if (address >= (ulong) _end || address < (ulong) _einitdata) |
867e359b CM |
267 | return construct_pgprot(PAGE_KERNEL, initial_heap_home()); |
268 | ||
867e359b CM |
269 | /* Use hash-for-home if requested for data/bss. */ |
270 | if (kdata_hash) | |
271 | return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH); | |
867e359b CM |
272 | |
273 | /* | |
274 | * Otherwise we just hand out consecutive cpus. To avoid | |
275 | * requiring this function to hold state, we just walk forward from | |
e540e835 WSH |
276 | * __end_rodata by PAGE_SIZE, skipping the readonly and init data, to |
277 | * reach the requested address, while walking cpu home around | |
278 | * kdata_mask. This is typically no more than a dozen or so iterations. | |
867e359b | 279 | */ |
ce61cdc2 | 280 | page = (((ulong)__end_rodata) + PAGE_SIZE - 1) & PAGE_MASK; |
0707ad30 CM |
281 | BUG_ON(address < page || address >= (ulong)_end); |
282 | cpu = cpumask_first(&kdata_mask); | |
283 | for (; page < address; page += PAGE_SIZE) { | |
284 | if (page >= (ulong)&init_thread_union && | |
285 | page < (ulong)&init_thread_union + THREAD_SIZE) | |
286 | continue; | |
867e359b | 287 | if (page == (ulong)empty_zero_page) |
0707ad30 | 288 | continue; |
867e359b | 289 | #ifndef __tilegx__ |
867e359b | 290 | if (page == (ulong)atomic_locks) |
0707ad30 | 291 | continue; |
867e359b | 292 | #endif |
0707ad30 CM |
293 | cpu = cpumask_next(cpu, &kdata_mask); |
294 | if (cpu == NR_CPUS) | |
295 | cpu = cpumask_first(&kdata_mask); | |
867e359b CM |
296 | } |
297 | return construct_pgprot(PAGE_KERNEL, cpu); | |
298 | } | |
299 | ||
300 | /* | |
301 | * This function sets up how we cache the kernel text. If we have | |
302 | * hash-for-home support, normally that is used instead (see the | |
303 | * kcache_hash boot flag for more information). But if we end up | |
304 | * using a page-based caching technique, this option sets up the | |
305 | * details of that. In addition, the "ktext=nocache" option may | |
306 | * always be used to disable local caching of text pages, if desired. | |
307 | */ | |
308 | ||
309 | static int __initdata ktext_arg_seen; | |
310 | static int __initdata ktext_small; | |
311 | static int __initdata ktext_local; | |
312 | static int __initdata ktext_all; | |
313 | static int __initdata ktext_nondataplane; | |
314 | static int __initdata ktext_nocache; | |
315 | static struct cpumask __initdata ktext_mask; | |
316 | ||
317 | static int __init setup_ktext(char *str) | |
318 | { | |
319 | if (str == NULL) | |
320 | return -EINVAL; | |
321 | ||
322 | /* If you have a leading "nocache", turn off ktext caching */ | |
323 | if (strncmp(str, "nocache", 7) == 0) { | |
324 | ktext_nocache = 1; | |
0707ad30 | 325 | pr_info("ktext: disabling local caching of kernel text\n"); |
867e359b CM |
326 | str += 7; |
327 | if (*str == ',') | |
328 | ++str; | |
329 | if (*str == '\0') | |
330 | return 0; | |
331 | } | |
332 | ||
333 | ktext_arg_seen = 1; | |
334 | ||
d7c96611 | 335 | /* Default setting: use a huge page */ |
867e359b | 336 | if (strcmp(str, "huge") == 0) |
0707ad30 | 337 | pr_info("ktext: using one huge locally cached page\n"); |
867e359b CM |
338 | |
339 | /* Pay TLB cost but get no cache benefit: cache small pages locally */ | |
340 | else if (strcmp(str, "local") == 0) { | |
341 | ktext_small = 1; | |
342 | ktext_local = 1; | |
0707ad30 | 343 | pr_info("ktext: using small pages with local caching\n"); |
867e359b CM |
344 | } |
345 | ||
346 | /* Neighborhood cache ktext pages on all cpus. */ | |
347 | else if (strcmp(str, "all") == 0) { | |
348 | ktext_small = 1; | |
349 | ktext_all = 1; | |
0707ad30 | 350 | pr_info("ktext: using maximal caching neighborhood\n"); |
867e359b CM |
351 | } |
352 | ||
353 | ||
354 | /* Neighborhood ktext pages on specified mask */ | |
355 | else if (cpulist_parse(str, &ktext_mask) == 0) { | |
356 | char buf[NR_CPUS * 5]; | |
357 | cpulist_scnprintf(buf, sizeof(buf), &ktext_mask); | |
358 | if (cpumask_weight(&ktext_mask) > 1) { | |
359 | ktext_small = 1; | |
0707ad30 | 360 | pr_info("ktext: using caching neighborhood %s " |
867e359b CM |
361 | "with small pages\n", buf); |
362 | } else { | |
0707ad30 | 363 | pr_info("ktext: caching on cpu %s with one huge page\n", |
867e359b CM |
364 | buf); |
365 | } | |
366 | } | |
367 | ||
368 | else if (*str) | |
369 | return -EINVAL; | |
370 | ||
371 | return 0; | |
372 | } | |
373 | ||
374 | early_param("ktext", setup_ktext); | |
375 | ||
376 | ||
377 | static inline pgprot_t ktext_set_nocache(pgprot_t prot) | |
378 | { | |
379 | if (!ktext_nocache) | |
380 | prot = hv_pte_set_nc(prot); | |
867e359b CM |
381 | else |
382 | prot = hv_pte_set_no_alloc_l2(prot); | |
867e359b CM |
383 | return prot; |
384 | } | |
385 | ||
867e359b CM |
386 | /* Temporary page table we use for staging. */ |
387 | static pgd_t pgtables[PTRS_PER_PGD] | |
2cb82400 | 388 | __attribute__((aligned(HV_PAGE_TABLE_ALIGN))); |
867e359b CM |
389 | |
390 | /* | |
391 | * This maps the physical memory to kernel virtual address space, a total | |
392 | * of max_low_pfn pages, by creating page tables starting from address | |
393 | * PAGE_OFFSET. | |
394 | * | |
395 | * This routine transitions us from using a set of compiled-in large | |
396 | * pages to using some more precise caching, including removing access | |
397 | * to code pages mapped at PAGE_OFFSET (executed only at MEM_SV_START) | |
398 | * marking read-only data as locally cacheable, striping the remaining | |
399 | * .data and .bss across all the available tiles, and removing access | |
400 | * to pages above the top of RAM (thus ensuring a page fault from a bad | |
401 | * virtual address rather than a hypervisor shoot down for accessing | |
402 | * memory outside the assigned limits). | |
403 | */ | |
404 | static void __init kernel_physical_mapping_init(pgd_t *pgd_base) | |
405 | { | |
51007004 | 406 | unsigned long long irqmask; |
867e359b CM |
407 | unsigned long address, pfn; |
408 | pmd_t *pmd; | |
409 | pte_t *pte; | |
410 | int pte_ofs; | |
411 | const struct cpumask *my_cpu_mask = cpumask_of(smp_processor_id()); | |
412 | struct cpumask kstripe_mask; | |
413 | int rc, i; | |
414 | ||
867e359b | 415 | if (ktext_arg_seen && ktext_hash) { |
0707ad30 CM |
416 | pr_warning("warning: \"ktext\" boot argument ignored" |
417 | " if \"kcache_hash\" sets up text hash-for-home\n"); | |
867e359b CM |
418 | ktext_small = 0; |
419 | } | |
420 | ||
421 | if (kdata_arg_seen && kdata_hash) { | |
0707ad30 CM |
422 | pr_warning("warning: \"kdata\" boot argument ignored" |
423 | " if \"kcache_hash\" sets up data hash-for-home\n"); | |
867e359b CM |
424 | } |
425 | ||
426 | if (kdata_huge && !hash_default) { | |
0707ad30 CM |
427 | pr_warning("warning: disabling \"kdata=huge\"; requires" |
428 | " kcache_hash=all or =allbutstack\n"); | |
867e359b CM |
429 | kdata_huge = 0; |
430 | } | |
867e359b CM |
431 | |
432 | /* | |
433 | * Set up a mask for cpus to use for kernel striping. | |
434 | * This is normally all cpus, but minus dataplane cpus if any. | |
435 | * If the dataplane covers the whole chip, we stripe over | |
436 | * the whole chip too. | |
437 | */ | |
438 | cpumask_copy(&kstripe_mask, cpu_possible_mask); | |
439 | if (!kdata_arg_seen) | |
440 | kdata_mask = kstripe_mask; | |
441 | ||
442 | /* Allocate and fill in L2 page tables */ | |
443 | for (i = 0; i < MAX_NUMNODES; ++i) { | |
444 | #ifdef CONFIG_HIGHMEM | |
445 | unsigned long end_pfn = node_lowmem_end_pfn[i]; | |
446 | #else | |
447 | unsigned long end_pfn = node_end_pfn[i]; | |
448 | #endif | |
449 | unsigned long end_huge_pfn = 0; | |
450 | ||
451 | /* Pre-shatter the last huge page to allow per-cpu pages. */ | |
452 | if (kdata_huge) | |
453 | end_huge_pfn = end_pfn - (HPAGE_SIZE >> PAGE_SHIFT); | |
454 | ||
455 | pfn = node_start_pfn[i]; | |
456 | ||
457 | /* Allocate enough memory to hold L2 page tables for node. */ | |
458 | init_prealloc_ptes(i, end_pfn - pfn); | |
459 | ||
460 | address = (unsigned long) pfn_to_kaddr(pfn); | |
461 | while (pfn < end_pfn) { | |
462 | BUG_ON(address & (HPAGE_SIZE-1)); | |
463 | pmd = get_pmd(pgtables, address); | |
464 | pte = get_prealloc_pte(pfn); | |
465 | if (pfn < end_huge_pfn) { | |
466 | pgprot_t prot = init_pgprot(address); | |
467 | *(pte_t *)pmd = pte_mkhuge(pfn_pte(pfn, prot)); | |
468 | for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE; | |
469 | pfn++, pte_ofs++, address += PAGE_SIZE) | |
470 | pte[pte_ofs] = pfn_pte(pfn, prot); | |
471 | } else { | |
472 | if (kdata_huge) | |
473 | printk(KERN_DEBUG "pre-shattered huge" | |
474 | " page at %#lx\n", address); | |
475 | for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE; | |
476 | pfn++, pte_ofs++, address += PAGE_SIZE) { | |
477 | pgprot_t prot = init_pgprot(address); | |
478 | pte[pte_ofs] = pfn_pte(pfn, prot); | |
479 | } | |
480 | assign_pte(pmd, pte); | |
481 | } | |
482 | } | |
483 | } | |
484 | ||
485 | /* | |
486 | * Set or check ktext_map now that we have cpu_possible_mask | |
487 | * and kstripe_mask to work with. | |
488 | */ | |
489 | if (ktext_all) | |
490 | cpumask_copy(&ktext_mask, cpu_possible_mask); | |
491 | else if (ktext_nondataplane) | |
492 | ktext_mask = kstripe_mask; | |
493 | else if (!cpumask_empty(&ktext_mask)) { | |
494 | /* Sanity-check any mask that was requested */ | |
495 | struct cpumask bad; | |
496 | cpumask_andnot(&bad, &ktext_mask, cpu_possible_mask); | |
497 | cpumask_and(&ktext_mask, &ktext_mask, cpu_possible_mask); | |
498 | if (!cpumask_empty(&bad)) { | |
499 | char buf[NR_CPUS * 5]; | |
500 | cpulist_scnprintf(buf, sizeof(buf), &bad); | |
0707ad30 | 501 | pr_info("ktext: not using unavailable cpus %s\n", buf); |
867e359b CM |
502 | } |
503 | if (cpumask_empty(&ktext_mask)) { | |
0707ad30 CM |
504 | pr_warning("ktext: no valid cpus; caching on %d.\n", |
505 | smp_processor_id()); | |
867e359b CM |
506 | cpumask_copy(&ktext_mask, |
507 | cpumask_of(smp_processor_id())); | |
508 | } | |
509 | } | |
510 | ||
acbde1db | 511 | address = MEM_SV_START; |
867e359b | 512 | pmd = get_pmd(pgtables, address); |
7a7039ee | 513 | pfn = 0; /* code starts at PA 0 */ |
867e359b CM |
514 | if (ktext_small) { |
515 | /* Allocate an L2 PTE for the kernel text */ | |
516 | int cpu = 0; | |
517 | pgprot_t prot = construct_pgprot(PAGE_KERNEL_EXEC, | |
518 | PAGE_HOME_IMMUTABLE); | |
519 | ||
520 | if (ktext_local) { | |
521 | if (ktext_nocache) | |
522 | prot = hv_pte_set_mode(prot, | |
523 | HV_PTE_MODE_UNCACHED); | |
524 | else | |
525 | prot = hv_pte_set_mode(prot, | |
526 | HV_PTE_MODE_CACHE_NO_L3); | |
527 | } else { | |
528 | prot = hv_pte_set_mode(prot, | |
529 | HV_PTE_MODE_CACHE_TILE_L3); | |
530 | cpu = cpumask_first(&ktext_mask); | |
531 | ||
532 | prot = ktext_set_nocache(prot); | |
533 | } | |
534 | ||
40a3b8df | 535 | BUG_ON(address != (unsigned long)_text); |
7a7039ee CM |
536 | pte = NULL; |
537 | for (; address < (unsigned long)_einittext; | |
538 | pfn++, address += PAGE_SIZE) { | |
539 | pte_ofs = pte_index(address); | |
540 | if (pte_ofs == 0) { | |
541 | if (pte) | |
542 | assign_pte(pmd++, pte); | |
543 | pte = alloc_pte(); | |
544 | } | |
867e359b CM |
545 | if (!ktext_local) { |
546 | prot = set_remote_cache_cpu(prot, cpu); | |
547 | cpu = cpumask_next(cpu, &ktext_mask); | |
548 | if (cpu == NR_CPUS) | |
549 | cpu = cpumask_first(&ktext_mask); | |
550 | } | |
551 | pte[pte_ofs] = pfn_pte(pfn, prot); | |
552 | } | |
7a7039ee CM |
553 | if (pte) |
554 | assign_pte(pmd, pte); | |
867e359b CM |
555 | } else { |
556 | pte_t pteval = pfn_pte(0, PAGE_KERNEL_EXEC); | |
557 | pteval = pte_mkhuge(pteval); | |
867e359b CM |
558 | if (ktext_hash) { |
559 | pteval = hv_pte_set_mode(pteval, | |
560 | HV_PTE_MODE_CACHE_HASH_L3); | |
561 | pteval = ktext_set_nocache(pteval); | |
562 | } else | |
867e359b CM |
563 | if (cpumask_weight(&ktext_mask) == 1) { |
564 | pteval = set_remote_cache_cpu(pteval, | |
565 | cpumask_first(&ktext_mask)); | |
566 | pteval = hv_pte_set_mode(pteval, | |
567 | HV_PTE_MODE_CACHE_TILE_L3); | |
568 | pteval = ktext_set_nocache(pteval); | |
569 | } else if (ktext_nocache) | |
570 | pteval = hv_pte_set_mode(pteval, | |
571 | HV_PTE_MODE_UNCACHED); | |
572 | else | |
573 | pteval = hv_pte_set_mode(pteval, | |
574 | HV_PTE_MODE_CACHE_NO_L3); | |
7a7039ee CM |
575 | for (; address < (unsigned long)_einittext; |
576 | pfn += PFN_DOWN(HPAGE_SIZE), address += HPAGE_SIZE) | |
577 | *(pte_t *)(pmd++) = pfn_pte(pfn, pteval); | |
867e359b CM |
578 | } |
579 | ||
580 | /* Set swapper_pgprot here so it is flushed to memory right away. */ | |
581 | swapper_pgprot = init_pgprot((unsigned long)swapper_pg_dir); | |
582 | ||
583 | /* | |
584 | * Since we may be changing the caching of the stack and page | |
585 | * table itself, we invoke an assembly helper to do the | |
586 | * following steps: | |
587 | * | |
588 | * - flush the cache so we start with an empty slate | |
589 | * - install pgtables[] as the real page table | |
590 | * - flush the TLB so the new page table takes effect | |
591 | */ | |
51007004 CM |
592 | irqmask = interrupt_mask_save_mask(); |
593 | interrupt_mask_set_mask(-1ULL); | |
867e359b CM |
594 | rc = flush_and_install_context(__pa(pgtables), |
595 | init_pgprot((unsigned long)pgtables), | |
596 | __get_cpu_var(current_asid), | |
597 | cpumask_bits(my_cpu_mask)); | |
51007004 | 598 | interrupt_mask_restore_mask(irqmask); |
867e359b CM |
599 | BUG_ON(rc != 0); |
600 | ||
601 | /* Copy the page table back to the normal swapper_pg_dir. */ | |
602 | memcpy(pgd_base, pgtables, sizeof(pgtables)); | |
603 | __install_page_table(pgd_base, __get_cpu_var(current_asid), | |
604 | swapper_pgprot); | |
401586e9 CM |
605 | |
606 | /* | |
607 | * We just read swapper_pgprot and thus brought it into the cache, | |
608 | * with its new home & caching mode. When we start the other CPUs, | |
609 | * they're going to reference swapper_pgprot via their initial fake | |
610 | * VA-is-PA mappings, which cache everything locally. At that | |
611 | * time, if it's in our cache with a conflicting home, the | |
612 | * simulator's coherence checker will complain. So, flush it out | |
613 | * of our cache; we're not going to ever use it again anyway. | |
614 | */ | |
615 | __insn_finv(&swapper_pgprot); | |
867e359b CM |
616 | } |
617 | ||
618 | /* | |
619 | * devmem_is_allowed() checks to see if /dev/mem access to a certain address | |
620 | * is valid. The argument is a physical page number. | |
621 | * | |
622 | * On Tile, the only valid things for which we can just hand out unchecked | |
623 | * PTEs are the kernel code and data. Anything else might change its | |
624 | * homing with time, and we wouldn't know to adjust the /dev/mem PTEs. | |
625 | * Note that init_thread_union is released to heap soon after boot, | |
626 | * so we include it in the init data. | |
627 | * | |
628 | * For TILE-Gx, we might want to consider allowing access to PA | |
629 | * regions corresponding to PCI space, etc. | |
630 | */ | |
631 | int devmem_is_allowed(unsigned long pagenr) | |
632 | { | |
633 | return pagenr < kaddr_to_pfn(_end) && | |
634 | !(pagenr >= kaddr_to_pfn(&init_thread_union) || | |
635 | pagenr < kaddr_to_pfn(_einitdata)) && | |
636 | !(pagenr >= kaddr_to_pfn(_sinittext) || | |
637 | pagenr <= kaddr_to_pfn(_einittext-1)); | |
638 | } | |
639 | ||
640 | #ifdef CONFIG_HIGHMEM | |
641 | static void __init permanent_kmaps_init(pgd_t *pgd_base) | |
642 | { | |
643 | pgd_t *pgd; | |
644 | pud_t *pud; | |
645 | pmd_t *pmd; | |
646 | pte_t *pte; | |
647 | unsigned long vaddr; | |
648 | ||
649 | vaddr = PKMAP_BASE; | |
650 | page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base); | |
651 | ||
652 | pgd = swapper_pg_dir + pgd_index(vaddr); | |
653 | pud = pud_offset(pgd, vaddr); | |
654 | pmd = pmd_offset(pud, vaddr); | |
655 | pte = pte_offset_kernel(pmd, vaddr); | |
656 | pkmap_page_table = pte; | |
657 | } | |
658 | #endif /* CONFIG_HIGHMEM */ | |
659 | ||
660 | ||
621b1955 | 661 | #ifndef CONFIG_64BIT |
867e359b CM |
662 | static void __init init_free_pfn_range(unsigned long start, unsigned long end) |
663 | { | |
664 | unsigned long pfn; | |
665 | struct page *page = pfn_to_page(start); | |
666 | ||
667 | for (pfn = start; pfn < end; ) { | |
668 | /* Optimize by freeing pages in large batches */ | |
669 | int order = __ffs(pfn); | |
670 | int count, i; | |
671 | struct page *p; | |
672 | ||
673 | if (order >= MAX_ORDER) | |
674 | order = MAX_ORDER-1; | |
675 | count = 1 << order; | |
676 | while (pfn + count > end) { | |
677 | count >>= 1; | |
678 | --order; | |
679 | } | |
680 | for (p = page, i = 0; i < count; ++i, ++p) { | |
681 | __ClearPageReserved(p); | |
682 | /* | |
683 | * Hacky direct set to avoid unnecessary | |
684 | * lock take/release for EVERY page here. | |
685 | */ | |
686 | p->_count.counter = 0; | |
687 | p->_mapcount.counter = -1; | |
688 | } | |
689 | init_page_count(page); | |
690 | __free_pages(page, order); | |
abd1b6d6 | 691 | adjust_managed_page_count(page, count); |
867e359b CM |
692 | |
693 | page += count; | |
694 | pfn += count; | |
695 | } | |
696 | } | |
697 | ||
698 | static void __init set_non_bootmem_pages_init(void) | |
699 | { | |
700 | struct zone *z; | |
701 | for_each_zone(z) { | |
702 | unsigned long start, end; | |
703 | int nid = z->zone_pgdat->node_id; | |
eef015c8 | 704 | #ifdef CONFIG_HIGHMEM |
0707ad30 | 705 | int idx = zone_idx(z); |
eef015c8 | 706 | #endif |
867e359b CM |
707 | |
708 | start = z->zone_start_pfn; | |
867e359b | 709 | end = start + z->spanned_pages; |
eef015c8 CM |
710 | start = max(start, node_free_pfn[nid]); |
711 | start = max(start, max_low_pfn); | |
712 | ||
867e359b | 713 | #ifdef CONFIG_HIGHMEM |
0707ad30 | 714 | if (idx == ZONE_HIGHMEM) |
867e359b CM |
715 | totalhigh_pages += z->spanned_pages; |
716 | #endif | |
717 | if (kdata_huge) { | |
718 | unsigned long percpu_pfn = node_percpu_pfn[nid]; | |
719 | if (start < percpu_pfn && end > percpu_pfn) | |
720 | end = percpu_pfn; | |
721 | } | |
722 | #ifdef CONFIG_PCI | |
723 | if (start <= pci_reserve_start_pfn && | |
724 | end > pci_reserve_start_pfn) { | |
725 | if (end > pci_reserve_end_pfn) | |
726 | init_free_pfn_range(pci_reserve_end_pfn, end); | |
727 | end = pci_reserve_start_pfn; | |
728 | } | |
729 | #endif | |
730 | init_free_pfn_range(start, end); | |
731 | } | |
732 | } | |
621b1955 | 733 | #endif |
867e359b CM |
734 | |
735 | /* | |
736 | * paging_init() sets up the page tables - note that all of lowmem is | |
737 | * already mapped by head.S. | |
738 | */ | |
739 | void __init paging_init(void) | |
740 | { | |
867e359b CM |
741 | #ifdef __tilegx__ |
742 | pud_t *pud; | |
743 | #endif | |
744 | pgd_t *pgd_base = swapper_pg_dir; | |
745 | ||
746 | kernel_physical_mapping_init(pgd_base); | |
747 | ||
084fe6a0 | 748 | /* Fixed mappings, only the page table structure has to be created. */ |
bbaa22c3 CM |
749 | page_table_range_init(fix_to_virt(__end_of_fixed_addresses - 1), |
750 | FIXADDR_TOP, pgd_base); | |
751 | ||
752 | #ifdef CONFIG_HIGHMEM | |
867e359b CM |
753 | permanent_kmaps_init(pgd_base); |
754 | #endif | |
755 | ||
756 | #ifdef __tilegx__ | |
757 | /* | |
758 | * Since GX allocates just one pmd_t array worth of vmalloc space, | |
759 | * we go ahead and allocate it statically here, then share it | |
760 | * globally. As a result we don't have to worry about any task | |
761 | * changing init_mm once we get up and running, and there's no | |
762 | * need for e.g. vmalloc_sync_all(). | |
763 | */ | |
d5d14ed6 | 764 | BUILD_BUG_ON(pgd_index(VMALLOC_START) != pgd_index(VMALLOC_END - 1)); |
867e359b CM |
765 | pud = pud_offset(pgd_base + pgd_index(VMALLOC_START), VMALLOC_START); |
766 | assign_pmd(pud, alloc_pmd()); | |
767 | #endif | |
768 | } | |
769 | ||
770 | ||
771 | /* | |
772 | * Walk the kernel page tables and derive the page_home() from | |
773 | * the PTEs, so that set_pte() can properly validate the caching | |
774 | * of all PTEs it sees. | |
775 | */ | |
776 | void __init set_page_homes(void) | |
777 | { | |
778 | } | |
779 | ||
780 | static void __init set_max_mapnr_init(void) | |
781 | { | |
782 | #ifdef CONFIG_FLATMEM | |
783 | max_mapnr = max_low_pfn; | |
784 | #endif | |
785 | } | |
786 | ||
787 | void __init mem_init(void) | |
788 | { | |
867e359b CM |
789 | int i; |
790 | #ifndef __tilegx__ | |
791 | void *last; | |
792 | #endif | |
793 | ||
794 | #ifdef CONFIG_FLATMEM | |
d1afa65c | 795 | BUG_ON(!mem_map); |
867e359b CM |
796 | #endif |
797 | ||
798 | #ifdef CONFIG_HIGHMEM | |
799 | /* check that fixmap and pkmap do not overlap */ | |
800 | if (PKMAP_ADDR(LAST_PKMAP-1) >= FIXADDR_START) { | |
0707ad30 | 801 | pr_err("fixmap and kmap areas overlap" |
867e359b | 802 | " - this will crash\n"); |
0707ad30 | 803 | pr_err("pkstart: %lxh pkend: %lxh fixstart %lxh\n", |
867e359b CM |
804 | PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP-1), |
805 | FIXADDR_START); | |
806 | BUG(); | |
807 | } | |
808 | #endif | |
809 | ||
810 | set_max_mapnr_init(); | |
811 | ||
812 | /* this will put all bootmem onto the freelists */ | |
0c988534 | 813 | free_all_bootmem(); |
867e359b | 814 | |
621b1955 | 815 | #ifndef CONFIG_64BIT |
867e359b CM |
816 | /* count all remaining LOWMEM and give all HIGHMEM to page allocator */ |
817 | set_non_bootmem_pages_init(); | |
621b1955 | 818 | #endif |
867e359b | 819 | |
3f29c331 | 820 | mem_init_print_info(NULL); |
867e359b CM |
821 | |
822 | /* | |
823 | * In debug mode, dump some interesting memory mappings. | |
824 | */ | |
825 | #ifdef CONFIG_HIGHMEM | |
826 | printk(KERN_DEBUG " KMAP %#lx - %#lx\n", | |
827 | FIXADDR_START, FIXADDR_TOP + PAGE_SIZE - 1); | |
828 | printk(KERN_DEBUG " PKMAP %#lx - %#lx\n", | |
829 | PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP) - 1); | |
867e359b CM |
830 | #endif |
831 | printk(KERN_DEBUG " VMALLOC %#lx - %#lx\n", | |
832 | _VMALLOC_START, _VMALLOC_END - 1); | |
833 | #ifdef __tilegx__ | |
834 | for (i = MAX_NUMNODES-1; i >= 0; --i) { | |
835 | struct pglist_data *node = &node_data[i]; | |
836 | if (node->node_present_pages) { | |
837 | unsigned long start = (unsigned long) | |
838 | pfn_to_kaddr(node->node_start_pfn); | |
839 | unsigned long end = start + | |
840 | (node->node_present_pages << PAGE_SHIFT); | |
841 | printk(KERN_DEBUG " MEM%d %#lx - %#lx\n", | |
842 | i, start, end - 1); | |
843 | } | |
844 | } | |
845 | #else | |
846 | last = high_memory; | |
847 | for (i = MAX_NUMNODES-1; i >= 0; --i) { | |
848 | if ((unsigned long)vbase_map[i] != -1UL) { | |
849 | printk(KERN_DEBUG " LOWMEM%d %#lx - %#lx\n", | |
850 | i, (unsigned long) (vbase_map[i]), | |
851 | (unsigned long) (last-1)); | |
852 | last = vbase_map[i]; | |
853 | } | |
854 | } | |
855 | #endif | |
856 | ||
857 | #ifndef __tilegx__ | |
858 | /* | |
859 | * Convert from using one lock for all atomic operations to | |
860 | * one per cpu. | |
861 | */ | |
862 | __init_atomic_per_cpu(); | |
863 | #endif | |
864 | } | |
865 | ||
866 | /* | |
867 | * this is for the non-NUMA, single node SMP system case. | |
868 | * Specifically, in the case of x86, we will always add | |
869 | * memory to the highmem for now. | |
870 | */ | |
871 | #ifndef CONFIG_NEED_MULTIPLE_NODES | |
872 | int arch_add_memory(u64 start, u64 size) | |
873 | { | |
874 | struct pglist_data *pgdata = &contig_page_data; | |
875 | struct zone *zone = pgdata->node_zones + MAX_NR_ZONES-1; | |
876 | unsigned long start_pfn = start >> PAGE_SHIFT; | |
877 | unsigned long nr_pages = size >> PAGE_SHIFT; | |
878 | ||
879 | return __add_pages(zone, start_pfn, nr_pages); | |
880 | } | |
881 | ||
882 | int remove_memory(u64 start, u64 size) | |
883 | { | |
884 | return -EINVAL; | |
885 | } | |
24d335ca WC |
886 | |
887 | #ifdef CONFIG_MEMORY_HOTREMOVE | |
888 | int arch_remove_memory(u64 start, u64 size) | |
889 | { | |
890 | /* TODO */ | |
891 | return -EBUSY; | |
892 | } | |
893 | #endif | |
867e359b CM |
894 | #endif |
895 | ||
896 | struct kmem_cache *pgd_cache; | |
897 | ||
898 | void __init pgtable_cache_init(void) | |
899 | { | |
76c567fb | 900 | pgd_cache = kmem_cache_create("pgd", SIZEOF_PGD, SIZEOF_PGD, 0, NULL); |
867e359b CM |
901 | if (!pgd_cache) |
902 | panic("pgtable_cache_init(): Cannot create pgd cache"); | |
903 | } | |
904 | ||
867e359b CM |
905 | #ifdef CONFIG_DEBUG_PAGEALLOC |
906 | static long __write_once initfree; | |
907 | #else | |
908 | static long __write_once initfree = 1; | |
909 | #endif | |
910 | ||
911 | /* Select whether to free (1) or mark unusable (0) the __init pages. */ | |
912 | static int __init set_initfree(char *str) | |
913 | { | |
d59e609d | 914 | long val; |
b2dfa048 | 915 | if (kstrtol(str, 0, &val) == 0) { |
d59e609d CM |
916 | initfree = val; |
917 | pr_info("initfree: %s free init pages\n", | |
918 | initfree ? "will" : "won't"); | |
919 | } | |
867e359b CM |
920 | return 1; |
921 | } | |
922 | __setup("initfree=", set_initfree); | |
923 | ||
924 | static void free_init_pages(char *what, unsigned long begin, unsigned long end) | |
925 | { | |
926 | unsigned long addr = (unsigned long) begin; | |
927 | ||
928 | if (kdata_huge && !initfree) { | |
0707ad30 CM |
929 | pr_warning("Warning: ignoring initfree=0:" |
930 | " incompatible with kdata=huge\n"); | |
867e359b CM |
931 | initfree = 1; |
932 | } | |
933 | end = (end + PAGE_SIZE - 1) & PAGE_MASK; | |
934 | local_flush_tlb_pages(NULL, begin, PAGE_SIZE, end - begin); | |
935 | for (addr = begin; addr < end; addr += PAGE_SIZE) { | |
936 | /* | |
937 | * Note we just reset the home here directly in the | |
938 | * page table. We know this is safe because our caller | |
939 | * just flushed the caches on all the other cpus, | |
940 | * and they won't be touching any of these pages. | |
941 | */ | |
942 | int pfn = kaddr_to_pfn((void *)addr); | |
943 | struct page *page = pfn_to_page(pfn); | |
640710a3 | 944 | pte_t *ptep = virt_to_kpte(addr); |
867e359b CM |
945 | if (!initfree) { |
946 | /* | |
947 | * If debugging page accesses then do not free | |
948 | * this memory but mark them not present - any | |
949 | * buggy init-section access will create a | |
950 | * kernel page fault: | |
951 | */ | |
952 | pte_clear(&init_mm, addr, ptep); | |
953 | continue; | |
954 | } | |
867e359b CM |
955 | if (pte_huge(*ptep)) |
956 | BUG_ON(!kdata_huge); | |
957 | else | |
958 | set_pte_at(&init_mm, addr, ptep, | |
959 | pfn_pte(pfn, PAGE_KERNEL)); | |
960 | memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE); | |
abd1b6d6 | 961 | free_reserved_page(page); |
867e359b | 962 | } |
0707ad30 | 963 | pr_info("Freeing %s: %ldk freed\n", what, (end - begin) >> 10); |
867e359b CM |
964 | } |
965 | ||
966 | void free_initmem(void) | |
967 | { | |
acbde1db | 968 | const unsigned long text_delta = MEM_SV_START - PAGE_OFFSET; |
867e359b CM |
969 | |
970 | /* | |
ce61cdc2 | 971 | * Evict the cache on all cores to avoid incoherence. |
d7c96611 | 972 | * We are guaranteed that no one will touch the init pages any more. |
867e359b CM |
973 | */ |
974 | homecache_evict(&cpu_cacheable_map); | |
975 | ||
976 | /* Free the data pages that we won't use again after init. */ | |
977 | free_init_pages("unused kernel data", | |
978 | (unsigned long)_sinitdata, | |
979 | (unsigned long)_einitdata); | |
980 | ||
981 | /* | |
982 | * Free the pages mapped from 0xc0000000 that correspond to code | |
acbde1db | 983 | * pages from MEM_SV_START that we won't use again after init. |
867e359b CM |
984 | */ |
985 | free_init_pages("unused kernel text", | |
986 | (unsigned long)_sinittext - text_delta, | |
987 | (unsigned long)_einittext - text_delta); | |
867e359b CM |
988 | /* Do a global TLB flush so everyone sees the changes. */ |
989 | flush_tlb_all(); | |
990 | } |