1 /* 2 * This file contains common routines for dealing with free of page tables 3 * Along with common page table handling code 4 * 5 * Derived from arch/powerpc/mm/tlb_64.c: 6 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org) 7 * 8 * Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au) 9 * and Cort Dougan (PReP) (cort@cs.nmt.edu) 10 * Copyright (C) 1996 Paul Mackerras 11 * 12 * Derived from "arch/i386/mm/init.c" 13 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 14 * 15 * Dave Engebretsen <engebret@us.ibm.com> 16 * Rework for PPC64 port. 17 * 18 * This program is free software; you can redistribute it and/or 19 * modify it under the terms of the GNU General Public License 20 * as published by the Free Software Foundation; either version 21 * 2 of the License, or (at your option) any later version. 22 */ 23 24 #include <linux/kernel.h> 25 #include <linux/gfp.h> 26 #include <linux/mm.h> 27 #include <linux/init.h> 28 #include <linux/percpu.h> 29 #include <linux/hardirq.h> 30 #include <asm/pgalloc.h> 31 #include <asm/tlbflush.h> 32 #include <asm/tlb.h> 33 34 #include "mmu_decl.h" 35 36 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers); 37 38 #ifdef CONFIG_SMP 39 40 /* 41 * Handle batching of page table freeing on SMP. Page tables are 42 * queued up and send to be freed later by RCU in order to avoid 43 * freeing a page table page that is being walked without locks 44 */ 45 46 static DEFINE_PER_CPU(struct pte_freelist_batch *, pte_freelist_cur); 47 static unsigned long pte_freelist_forced_free; 48 49 struct pte_freelist_batch 50 { 51 struct rcu_head rcu; 52 unsigned int index; 53 unsigned long tables[0]; 54 }; 55 56 #define PTE_FREELIST_SIZE \ 57 ((PAGE_SIZE - sizeof(struct pte_freelist_batch)) \ 58 / sizeof(unsigned long)) 59 60 static void pte_free_smp_sync(void *arg) 61 { 62 /* Do nothing, just ensure we sync with all CPUs */ 63 } 64 65 /* This is only called when we are critically out of memory 66 * (and fail to get a page in pte_free_tlb). 67 */ 68 static void pgtable_free_now(void *table, unsigned shift) 69 { 70 pte_freelist_forced_free++; 71 72 smp_call_function(pte_free_smp_sync, NULL, 1); 73 74 pgtable_free(table, shift); 75 } 76 77 static void pte_free_rcu_callback(struct rcu_head *head) 78 { 79 struct pte_freelist_batch *batch = 80 container_of(head, struct pte_freelist_batch, rcu); 81 unsigned int i; 82 83 for (i = 0; i < batch->index; i++) { 84 void *table = (void *)(batch->tables[i] & ~MAX_PGTABLE_INDEX_SIZE); 85 unsigned shift = batch->tables[i] & MAX_PGTABLE_INDEX_SIZE; 86 87 pgtable_free(table, shift); 88 } 89 90 free_page((unsigned long)batch); 91 } 92 93 static void pte_free_submit(struct pte_freelist_batch *batch) 94 { 95 INIT_RCU_HEAD(&batch->rcu); 96 call_rcu(&batch->rcu, pte_free_rcu_callback); 97 } 98 99 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, unsigned shift) 100 { 101 /* This is safe since tlb_gather_mmu has disabled preemption */ 102 struct pte_freelist_batch **batchp = &__get_cpu_var(pte_freelist_cur); 103 unsigned long pgf; 104 105 if (atomic_read(&tlb->mm->mm_users) < 2 || 106 cpumask_equal(mm_cpumask(tlb->mm), cpumask_of(smp_processor_id()))){ 107 pgtable_free(table, shift); 108 return; 109 } 110 111 if (*batchp == NULL) { 112 *batchp = (struct pte_freelist_batch *)__get_free_page(GFP_ATOMIC); 113 if (*batchp == NULL) { 114 pgtable_free_now(table, shift); 115 return; 116 } 117 (*batchp)->index = 0; 118 } 119 BUG_ON(shift > MAX_PGTABLE_INDEX_SIZE); 120 pgf = (unsigned long)table | shift; 121 (*batchp)->tables[(*batchp)->index++] = pgf; 122 if ((*batchp)->index == PTE_FREELIST_SIZE) { 123 pte_free_submit(*batchp); 124 *batchp = NULL; 125 } 126 } 127 128 void pte_free_finish(void) 129 { 130 /* This is safe since tlb_gather_mmu has disabled preemption */ 131 struct pte_freelist_batch **batchp = &__get_cpu_var(pte_freelist_cur); 132 133 if (*batchp == NULL) 134 return; 135 pte_free_submit(*batchp); 136 *batchp = NULL; 137 } 138 139 #endif /* CONFIG_SMP */ 140 141 static inline int is_exec_fault(void) 142 { 143 return current->thread.regs && TRAP(current->thread.regs) == 0x400; 144 } 145 146 /* We only try to do i/d cache coherency on stuff that looks like 147 * reasonably "normal" PTEs. We currently require a PTE to be present 148 * and we avoid _PAGE_SPECIAL and _PAGE_NO_CACHE. We also only do that 149 * on userspace PTEs 150 */ 151 static inline int pte_looks_normal(pte_t pte) 152 { 153 return (pte_val(pte) & 154 (_PAGE_PRESENT | _PAGE_SPECIAL | _PAGE_NO_CACHE | _PAGE_USER)) == 155 (_PAGE_PRESENT | _PAGE_USER); 156 } 157 158 struct page * maybe_pte_to_page(pte_t pte) 159 { 160 unsigned long pfn = pte_pfn(pte); 161 struct page *page; 162 163 if (unlikely(!pfn_valid(pfn))) 164 return NULL; 165 page = pfn_to_page(pfn); 166 if (PageReserved(page)) 167 return NULL; 168 return page; 169 } 170 171 #if defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0 172 173 /* Server-style MMU handles coherency when hashing if HW exec permission 174 * is supposed per page (currently 64-bit only). If not, then, we always 175 * flush the cache for valid PTEs in set_pte. Embedded CPU without HW exec 176 * support falls into the same category. 177 */ 178 179 static pte_t set_pte_filter(pte_t pte, unsigned long addr) 180 { 181 pte = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS); 182 if (pte_looks_normal(pte) && !(cpu_has_feature(CPU_FTR_COHERENT_ICACHE) || 183 cpu_has_feature(CPU_FTR_NOEXECUTE))) { 184 struct page *pg = maybe_pte_to_page(pte); 185 if (!pg) 186 return pte; 187 if (!test_bit(PG_arch_1, &pg->flags)) { 188 #ifdef CONFIG_8xx 189 /* On 8xx, cache control instructions (particularly 190 * "dcbst" from flush_dcache_icache) fault as write 191 * operation if there is an unpopulated TLB entry 192 * for the address in question. To workaround that, 193 * we invalidate the TLB here, thus avoiding dcbst 194 * misbehaviour. 195 */ 196 /* 8xx doesn't care about PID, size or ind args */ 197 _tlbil_va(addr, 0, 0, 0); 198 #endif /* CONFIG_8xx */ 199 flush_dcache_icache_page(pg); 200 set_bit(PG_arch_1, &pg->flags); 201 } 202 } 203 return pte; 204 } 205 206 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma, 207 int dirty) 208 { 209 return pte; 210 } 211 212 #else /* defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0 */ 213 214 /* Embedded type MMU with HW exec support. This is a bit more complicated 215 * as we don't have two bits to spare for _PAGE_EXEC and _PAGE_HWEXEC so 216 * instead we "filter out" the exec permission for non clean pages. 217 */ 218 static pte_t set_pte_filter(pte_t pte, unsigned long addr) 219 { 220 struct page *pg; 221 222 /* No exec permission in the first place, move on */ 223 if (!(pte_val(pte) & _PAGE_EXEC) || !pte_looks_normal(pte)) 224 return pte; 225 226 /* If you set _PAGE_EXEC on weird pages you're on your own */ 227 pg = maybe_pte_to_page(pte); 228 if (unlikely(!pg)) 229 return pte; 230 231 /* If the page clean, we move on */ 232 if (test_bit(PG_arch_1, &pg->flags)) 233 return pte; 234 235 /* If it's an exec fault, we flush the cache and make it clean */ 236 if (is_exec_fault()) { 237 flush_dcache_icache_page(pg); 238 set_bit(PG_arch_1, &pg->flags); 239 return pte; 240 } 241 242 /* Else, we filter out _PAGE_EXEC */ 243 return __pte(pte_val(pte) & ~_PAGE_EXEC); 244 } 245 246 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma, 247 int dirty) 248 { 249 struct page *pg; 250 251 /* So here, we only care about exec faults, as we use them 252 * to recover lost _PAGE_EXEC and perform I$/D$ coherency 253 * if necessary. Also if _PAGE_EXEC is already set, same deal, 254 * we just bail out 255 */ 256 if (dirty || (pte_val(pte) & _PAGE_EXEC) || !is_exec_fault()) 257 return pte; 258 259 #ifdef CONFIG_DEBUG_VM 260 /* So this is an exec fault, _PAGE_EXEC is not set. If it was 261 * an error we would have bailed out earlier in do_page_fault() 262 * but let's make sure of it 263 */ 264 if (WARN_ON(!(vma->vm_flags & VM_EXEC))) 265 return pte; 266 #endif /* CONFIG_DEBUG_VM */ 267 268 /* If you set _PAGE_EXEC on weird pages you're on your own */ 269 pg = maybe_pte_to_page(pte); 270 if (unlikely(!pg)) 271 goto bail; 272 273 /* If the page is already clean, we move on */ 274 if (test_bit(PG_arch_1, &pg->flags)) 275 goto bail; 276 277 /* Clean the page and set PG_arch_1 */ 278 flush_dcache_icache_page(pg); 279 set_bit(PG_arch_1, &pg->flags); 280 281 bail: 282 return __pte(pte_val(pte) | _PAGE_EXEC); 283 } 284 285 #endif /* !(defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0) */ 286 287 /* 288 * set_pte stores a linux PTE into the linux page table. 289 */ 290 void set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep, 291 pte_t pte) 292 { 293 #ifdef CONFIG_DEBUG_VM 294 WARN_ON(pte_present(*ptep)); 295 #endif 296 /* Note: mm->context.id might not yet have been assigned as 297 * this context might not have been activated yet when this 298 * is called. 299 */ 300 pte = set_pte_filter(pte, addr); 301 302 /* Perform the setting of the PTE */ 303 __set_pte_at(mm, addr, ptep, pte, 0); 304 } 305 306 /* 307 * This is called when relaxing access to a PTE. It's also called in the page 308 * fault path when we don't hit any of the major fault cases, ie, a minor 309 * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have 310 * handled those two for us, we additionally deal with missing execute 311 * permission here on some processors 312 */ 313 int ptep_set_access_flags(struct vm_area_struct *vma, unsigned long address, 314 pte_t *ptep, pte_t entry, int dirty) 315 { 316 int changed; 317 entry = set_access_flags_filter(entry, vma, dirty); 318 changed = !pte_same(*(ptep), entry); 319 if (changed) { 320 if (!(vma->vm_flags & VM_HUGETLB)) 321 assert_pte_locked(vma->vm_mm, address); 322 __ptep_set_access_flags(ptep, entry); 323 flush_tlb_page_nohash(vma, address); 324 } 325 return changed; 326 } 327 328 #ifdef CONFIG_DEBUG_VM 329 void assert_pte_locked(struct mm_struct *mm, unsigned long addr) 330 { 331 pgd_t *pgd; 332 pud_t *pud; 333 pmd_t *pmd; 334 335 if (mm == &init_mm) 336 return; 337 pgd = mm->pgd + pgd_index(addr); 338 BUG_ON(pgd_none(*pgd)); 339 pud = pud_offset(pgd, addr); 340 BUG_ON(pud_none(*pud)); 341 pmd = pmd_offset(pud, addr); 342 BUG_ON(!pmd_present(*pmd)); 343 assert_spin_locked(pte_lockptr(mm, pmd)); 344 } 345 #endif /* CONFIG_DEBUG_VM */ 346 347