Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame^] | 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* arch/sparc64/mm/tlb.c |
| 3 | * |
| 4 | * Copyright (C) 2004 David S. Miller <davem@redhat.com> |
| 5 | */ |
| 6 | |
| 7 | #include <linux/kernel.h> |
| 8 | #include <linux/percpu.h> |
| 9 | #include <linux/mm.h> |
| 10 | #include <linux/swap.h> |
| 11 | #include <linux/preempt.h> |
| 12 | |
| 13 | #include <asm/pgtable.h> |
| 14 | #include <asm/pgalloc.h> |
| 15 | #include <asm/tlbflush.h> |
| 16 | #include <asm/cacheflush.h> |
| 17 | #include <asm/mmu_context.h> |
| 18 | #include <asm/tlb.h> |
| 19 | |
| 20 | /* Heavily inspired by the ppc64 code. */ |
| 21 | |
| 22 | static DEFINE_PER_CPU(struct tlb_batch, tlb_batch); |
| 23 | |
| 24 | void flush_tlb_pending(void) |
| 25 | { |
| 26 | struct tlb_batch *tb = &get_cpu_var(tlb_batch); |
| 27 | struct mm_struct *mm = tb->mm; |
| 28 | |
| 29 | if (!tb->tlb_nr) |
| 30 | goto out; |
| 31 | |
| 32 | flush_tsb_user(tb); |
| 33 | |
| 34 | if (CTX_VALID(mm->context)) { |
| 35 | if (tb->tlb_nr == 1) { |
| 36 | global_flush_tlb_page(mm, tb->vaddrs[0]); |
| 37 | } else { |
| 38 | #ifdef CONFIG_SMP |
| 39 | smp_flush_tlb_pending(tb->mm, tb->tlb_nr, |
| 40 | &tb->vaddrs[0]); |
| 41 | #else |
| 42 | __flush_tlb_pending(CTX_HWBITS(tb->mm->context), |
| 43 | tb->tlb_nr, &tb->vaddrs[0]); |
| 44 | #endif |
| 45 | } |
| 46 | } |
| 47 | |
| 48 | tb->tlb_nr = 0; |
| 49 | |
| 50 | out: |
| 51 | put_cpu_var(tlb_batch); |
| 52 | } |
| 53 | |
| 54 | void arch_enter_lazy_mmu_mode(void) |
| 55 | { |
| 56 | struct tlb_batch *tb = this_cpu_ptr(&tlb_batch); |
| 57 | |
| 58 | tb->active = 1; |
| 59 | } |
| 60 | |
| 61 | void arch_leave_lazy_mmu_mode(void) |
| 62 | { |
| 63 | struct tlb_batch *tb = this_cpu_ptr(&tlb_batch); |
| 64 | |
| 65 | if (tb->tlb_nr) |
| 66 | flush_tlb_pending(); |
| 67 | tb->active = 0; |
| 68 | } |
| 69 | |
| 70 | static void tlb_batch_add_one(struct mm_struct *mm, unsigned long vaddr, |
| 71 | bool exec, unsigned int hugepage_shift) |
| 72 | { |
| 73 | struct tlb_batch *tb = &get_cpu_var(tlb_batch); |
| 74 | unsigned long nr; |
| 75 | |
| 76 | vaddr &= PAGE_MASK; |
| 77 | if (exec) |
| 78 | vaddr |= 0x1UL; |
| 79 | |
| 80 | nr = tb->tlb_nr; |
| 81 | |
| 82 | if (unlikely(nr != 0 && mm != tb->mm)) { |
| 83 | flush_tlb_pending(); |
| 84 | nr = 0; |
| 85 | } |
| 86 | |
| 87 | if (!tb->active) { |
| 88 | flush_tsb_user_page(mm, vaddr, hugepage_shift); |
| 89 | global_flush_tlb_page(mm, vaddr); |
| 90 | goto out; |
| 91 | } |
| 92 | |
| 93 | if (nr == 0) { |
| 94 | tb->mm = mm; |
| 95 | tb->hugepage_shift = hugepage_shift; |
| 96 | } |
| 97 | |
| 98 | if (tb->hugepage_shift != hugepage_shift) { |
| 99 | flush_tlb_pending(); |
| 100 | tb->hugepage_shift = hugepage_shift; |
| 101 | nr = 0; |
| 102 | } |
| 103 | |
| 104 | tb->vaddrs[nr] = vaddr; |
| 105 | tb->tlb_nr = ++nr; |
| 106 | if (nr >= TLB_BATCH_NR) |
| 107 | flush_tlb_pending(); |
| 108 | |
| 109 | out: |
| 110 | put_cpu_var(tlb_batch); |
| 111 | } |
| 112 | |
| 113 | void tlb_batch_add(struct mm_struct *mm, unsigned long vaddr, |
| 114 | pte_t *ptep, pte_t orig, int fullmm, |
| 115 | unsigned int hugepage_shift) |
| 116 | { |
| 117 | if (tlb_type != hypervisor && |
| 118 | pte_dirty(orig)) { |
| 119 | unsigned long paddr, pfn = pte_pfn(orig); |
| 120 | struct address_space *mapping; |
| 121 | struct page *page; |
| 122 | |
| 123 | if (!pfn_valid(pfn)) |
| 124 | goto no_cache_flush; |
| 125 | |
| 126 | page = pfn_to_page(pfn); |
| 127 | if (PageReserved(page)) |
| 128 | goto no_cache_flush; |
| 129 | |
| 130 | /* A real file page? */ |
| 131 | mapping = page_mapping_file(page); |
| 132 | if (!mapping) |
| 133 | goto no_cache_flush; |
| 134 | |
| 135 | paddr = (unsigned long) page_address(page); |
| 136 | if ((paddr ^ vaddr) & (1 << 13)) |
| 137 | flush_dcache_page_all(mm, page); |
| 138 | } |
| 139 | |
| 140 | no_cache_flush: |
| 141 | if (!fullmm) |
| 142 | tlb_batch_add_one(mm, vaddr, pte_exec(orig), hugepage_shift); |
| 143 | } |
| 144 | |
| 145 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 146 | static void tlb_batch_pmd_scan(struct mm_struct *mm, unsigned long vaddr, |
| 147 | pmd_t pmd) |
| 148 | { |
| 149 | unsigned long end; |
| 150 | pte_t *pte; |
| 151 | |
| 152 | pte = pte_offset_map(&pmd, vaddr); |
| 153 | end = vaddr + HPAGE_SIZE; |
| 154 | while (vaddr < end) { |
| 155 | if (pte_val(*pte) & _PAGE_VALID) { |
| 156 | bool exec = pte_exec(*pte); |
| 157 | |
| 158 | tlb_batch_add_one(mm, vaddr, exec, PAGE_SHIFT); |
| 159 | } |
| 160 | pte++; |
| 161 | vaddr += PAGE_SIZE; |
| 162 | } |
| 163 | pte_unmap(pte); |
| 164 | } |
| 165 | |
| 166 | |
| 167 | static void __set_pmd_acct(struct mm_struct *mm, unsigned long addr, |
| 168 | pmd_t orig, pmd_t pmd) |
| 169 | { |
| 170 | if (mm == &init_mm) |
| 171 | return; |
| 172 | |
| 173 | if ((pmd_val(pmd) ^ pmd_val(orig)) & _PAGE_PMD_HUGE) { |
| 174 | /* |
| 175 | * Note that this routine only sets pmds for THP pages. |
| 176 | * Hugetlb pages are handled elsewhere. We need to check |
| 177 | * for huge zero page. Huge zero pages are like hugetlb |
| 178 | * pages in that there is no RSS, but there is the need |
| 179 | * for TSB entries. So, huge zero page counts go into |
| 180 | * hugetlb_pte_count. |
| 181 | */ |
| 182 | if (pmd_val(pmd) & _PAGE_PMD_HUGE) { |
| 183 | if (is_huge_zero_page(pmd_page(pmd))) |
| 184 | mm->context.hugetlb_pte_count++; |
| 185 | else |
| 186 | mm->context.thp_pte_count++; |
| 187 | } else { |
| 188 | if (is_huge_zero_page(pmd_page(orig))) |
| 189 | mm->context.hugetlb_pte_count--; |
| 190 | else |
| 191 | mm->context.thp_pte_count--; |
| 192 | } |
| 193 | |
| 194 | /* Do not try to allocate the TSB hash table if we |
| 195 | * don't have one already. We have various locks held |
| 196 | * and thus we'll end up doing a GFP_KERNEL allocation |
| 197 | * in an atomic context. |
| 198 | * |
| 199 | * Instead, we let the first TLB miss on a hugepage |
| 200 | * take care of this. |
| 201 | */ |
| 202 | } |
| 203 | |
| 204 | if (!pmd_none(orig)) { |
| 205 | addr &= HPAGE_MASK; |
| 206 | if (pmd_trans_huge(orig)) { |
| 207 | pte_t orig_pte = __pte(pmd_val(orig)); |
| 208 | bool exec = pte_exec(orig_pte); |
| 209 | |
| 210 | tlb_batch_add_one(mm, addr, exec, REAL_HPAGE_SHIFT); |
| 211 | tlb_batch_add_one(mm, addr + REAL_HPAGE_SIZE, exec, |
| 212 | REAL_HPAGE_SHIFT); |
| 213 | } else { |
| 214 | tlb_batch_pmd_scan(mm, addr, orig); |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | void set_pmd_at(struct mm_struct *mm, unsigned long addr, |
| 220 | pmd_t *pmdp, pmd_t pmd) |
| 221 | { |
| 222 | pmd_t orig = *pmdp; |
| 223 | |
| 224 | *pmdp = pmd; |
| 225 | __set_pmd_acct(mm, addr, orig, pmd); |
| 226 | } |
| 227 | |
| 228 | static inline pmd_t pmdp_establish(struct vm_area_struct *vma, |
| 229 | unsigned long address, pmd_t *pmdp, pmd_t pmd) |
| 230 | { |
| 231 | pmd_t old; |
| 232 | |
| 233 | do { |
| 234 | old = *pmdp; |
| 235 | } while (cmpxchg64(&pmdp->pmd, old.pmd, pmd.pmd) != old.pmd); |
| 236 | __set_pmd_acct(vma->vm_mm, address, old, pmd); |
| 237 | |
| 238 | return old; |
| 239 | } |
| 240 | |
| 241 | /* |
| 242 | * This routine is only called when splitting a THP |
| 243 | */ |
| 244 | pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address, |
| 245 | pmd_t *pmdp) |
| 246 | { |
| 247 | pmd_t old, entry; |
| 248 | |
| 249 | entry = __pmd(pmd_val(*pmdp) & ~_PAGE_VALID); |
| 250 | old = pmdp_establish(vma, address, pmdp, entry); |
| 251 | flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE); |
| 252 | |
| 253 | /* |
| 254 | * set_pmd_at() will not be called in a way to decrement |
| 255 | * thp_pte_count when splitting a THP, so do it now. |
| 256 | * Sanity check pmd before doing the actual decrement. |
| 257 | */ |
| 258 | if ((pmd_val(entry) & _PAGE_PMD_HUGE) && |
| 259 | !is_huge_zero_page(pmd_page(entry))) |
| 260 | (vma->vm_mm)->context.thp_pte_count--; |
| 261 | |
| 262 | return old; |
| 263 | } |
| 264 | |
| 265 | void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp, |
| 266 | pgtable_t pgtable) |
| 267 | { |
| 268 | struct list_head *lh = (struct list_head *) pgtable; |
| 269 | |
| 270 | assert_spin_locked(&mm->page_table_lock); |
| 271 | |
| 272 | /* FIFO */ |
| 273 | if (!pmd_huge_pte(mm, pmdp)) |
| 274 | INIT_LIST_HEAD(lh); |
| 275 | else |
| 276 | list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp)); |
| 277 | pmd_huge_pte(mm, pmdp) = pgtable; |
| 278 | } |
| 279 | |
| 280 | pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp) |
| 281 | { |
| 282 | struct list_head *lh; |
| 283 | pgtable_t pgtable; |
| 284 | |
| 285 | assert_spin_locked(&mm->page_table_lock); |
| 286 | |
| 287 | /* FIFO */ |
| 288 | pgtable = pmd_huge_pte(mm, pmdp); |
| 289 | lh = (struct list_head *) pgtable; |
| 290 | if (list_empty(lh)) |
| 291 | pmd_huge_pte(mm, pmdp) = NULL; |
| 292 | else { |
| 293 | pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next; |
| 294 | list_del(lh); |
| 295 | } |
| 296 | pte_val(pgtable[0]) = 0; |
| 297 | pte_val(pgtable[1]) = 0; |
| 298 | |
| 299 | return pgtable; |
| 300 | } |
| 301 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ |