1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef _ASM_X86_PGTABLE_3LEVEL_H 3 #define _ASM_X86_PGTABLE_3LEVEL_H 4 5 #include <asm/atomic64_32.h> 6 7 /* 8 * Intel Physical Address Extension (PAE) Mode - three-level page 9 * tables on PPro+ CPUs. 10 * 11 * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com> 12 */ 13 14 #define pte_ERROR(e) \ 15 pr_err("%s:%d: bad pte %p(%08lx%08lx)\n", \ 16 __FILE__, __LINE__, &(e), (e).pte_high, (e).pte_low) 17 #define pmd_ERROR(e) \ 18 pr_err("%s:%d: bad pmd %p(%016Lx)\n", \ 19 __FILE__, __LINE__, &(e), pmd_val(e)) 20 #define pgd_ERROR(e) \ 21 pr_err("%s:%d: bad pgd %p(%016Lx)\n", \ 22 __FILE__, __LINE__, &(e), pgd_val(e)) 23 24 /* Rules for using set_pte: the pte being assigned *must* be 25 * either not present or in a state where the hardware will 26 * not attempt to update the pte. In places where this is 27 * not possible, use pte_get_and_clear to obtain the old pte 28 * value and then use set_pte to update it. -ben 29 */ 30 static inline void native_set_pte(pte_t *ptep, pte_t pte) 31 { 32 ptep->pte_high = pte.pte_high; 33 smp_wmb(); 34 ptep->pte_low = pte.pte_low; 35 } 36 37 #define pmd_read_atomic pmd_read_atomic 38 /* 39 * pte_offset_map_lock on 32bit PAE kernels was reading the pmd_t with 40 * a "*pmdp" dereference done by gcc. Problem is, in certain places 41 * where pte_offset_map_lock is called, concurrent page faults are 42 * allowed, if the mmap_sem is hold for reading. An example is mincore 43 * vs page faults vs MADV_DONTNEED. On the page fault side 44 * pmd_populate rightfully does a set_64bit, but if we're reading the 45 * pmd_t with a "*pmdp" on the mincore side, a SMP race can happen 46 * because gcc will not read the 64bit of the pmd atomically. To fix 47 * this all places running pmd_offset_map_lock() while holding the 48 * mmap_sem in read mode, shall read the pmdp pointer using this 49 * function to know if the pmd is null nor not, and in turn to know if 50 * they can run pmd_offset_map_lock or pmd_trans_huge or other pmd 51 * operations. 52 * 53 * Without THP if the mmap_sem is hold for reading, the pmd can only 54 * transition from null to not null while pmd_read_atomic runs. So 55 * we can always return atomic pmd values with this function. 56 * 57 * With THP if the mmap_sem is hold for reading, the pmd can become 58 * trans_huge or none or point to a pte (and in turn become "stable") 59 * at any time under pmd_read_atomic. We could read it really 60 * atomically here with a atomic64_read for the THP enabled case (and 61 * it would be a whole lot simpler), but to avoid using cmpxchg8b we 62 * only return an atomic pmdval if the low part of the pmdval is later 63 * found stable (i.e. pointing to a pte). And we're returning a none 64 * pmdval if the low part of the pmd is none. In some cases the high 65 * and low part of the pmdval returned may not be consistent if THP is 66 * enabled (the low part may point to previously mapped hugepage, 67 * while the high part may point to a more recently mapped hugepage), 68 * but pmd_none_or_trans_huge_or_clear_bad() only needs the low part 69 * of the pmd to be read atomically to decide if the pmd is unstable 70 * or not, with the only exception of when the low part of the pmd is 71 * zero in which case we return a none pmd. 72 */ 73 static inline pmd_t pmd_read_atomic(pmd_t *pmdp) 74 { 75 pmdval_t ret; 76 u32 *tmp = (u32 *)pmdp; 77 78 ret = (pmdval_t) (*tmp); 79 if (ret) { 80 /* 81 * If the low part is null, we must not read the high part 82 * or we can end up with a partial pmd. 83 */ 84 smp_rmb(); 85 ret |= ((pmdval_t)*(tmp + 1)) << 32; 86 } 87 88 return (pmd_t) { ret }; 89 } 90 91 static inline void native_set_pte_atomic(pte_t *ptep, pte_t pte) 92 { 93 set_64bit((unsigned long long *)(ptep), native_pte_val(pte)); 94 } 95 96 static inline void native_set_pmd(pmd_t *pmdp, pmd_t pmd) 97 { 98 set_64bit((unsigned long long *)(pmdp), native_pmd_val(pmd)); 99 } 100 101 static inline void native_set_pud(pud_t *pudp, pud_t pud) 102 { 103 #ifdef CONFIG_PAGE_TABLE_ISOLATION 104 pud.p4d.pgd = pti_set_user_pgtbl(&pudp->p4d.pgd, pud.p4d.pgd); 105 #endif 106 set_64bit((unsigned long long *)(pudp), native_pud_val(pud)); 107 } 108 109 /* 110 * For PTEs and PDEs, we must clear the P-bit first when clearing a page table 111 * entry, so clear the bottom half first and enforce ordering with a compiler 112 * barrier. 113 */ 114 static inline void native_pte_clear(struct mm_struct *mm, unsigned long addr, 115 pte_t *ptep) 116 { 117 ptep->pte_low = 0; 118 smp_wmb(); 119 ptep->pte_high = 0; 120 } 121 122 static inline void native_pmd_clear(pmd_t *pmd) 123 { 124 u32 *tmp = (u32 *)pmd; 125 *tmp = 0; 126 smp_wmb(); 127 *(tmp + 1) = 0; 128 } 129 130 static inline void native_pud_clear(pud_t *pudp) 131 { 132 } 133 134 static inline void pud_clear(pud_t *pudp) 135 { 136 set_pud(pudp, __pud(0)); 137 138 /* 139 * According to Intel App note "TLBs, Paging-Structure Caches, 140 * and Their Invalidation", April 2007, document 317080-001, 141 * section 8.1: in PAE mode we explicitly have to flush the 142 * TLB via cr3 if the top-level pgd is changed... 143 * 144 * Currently all places where pud_clear() is called either have 145 * flush_tlb_mm() followed or don't need TLB flush (x86_64 code or 146 * pud_clear_bad()), so we don't need TLB flush here. 147 */ 148 } 149 150 #ifdef CONFIG_SMP 151 static inline pte_t native_ptep_get_and_clear(pte_t *ptep) 152 { 153 pte_t res; 154 155 res.pte = (pteval_t)arch_atomic64_xchg((atomic64_t *)ptep, 0); 156 157 return res; 158 } 159 #else 160 #define native_ptep_get_and_clear(xp) native_local_ptep_get_and_clear(xp) 161 #endif 162 163 union split_pmd { 164 struct { 165 u32 pmd_low; 166 u32 pmd_high; 167 }; 168 pmd_t pmd; 169 }; 170 171 #ifdef CONFIG_SMP 172 static inline pmd_t native_pmdp_get_and_clear(pmd_t *pmdp) 173 { 174 union split_pmd res, *orig = (union split_pmd *)pmdp; 175 176 /* xchg acts as a barrier before setting of the high bits */ 177 res.pmd_low = xchg(&orig->pmd_low, 0); 178 res.pmd_high = orig->pmd_high; 179 orig->pmd_high = 0; 180 181 return res.pmd; 182 } 183 #else 184 #define native_pmdp_get_and_clear(xp) native_local_pmdp_get_and_clear(xp) 185 #endif 186 187 #ifndef pmdp_establish 188 #define pmdp_establish pmdp_establish 189 static inline pmd_t pmdp_establish(struct vm_area_struct *vma, 190 unsigned long address, pmd_t *pmdp, pmd_t pmd) 191 { 192 pmd_t old; 193 194 /* 195 * If pmd has present bit cleared we can get away without expensive 196 * cmpxchg64: we can update pmdp half-by-half without racing with 197 * anybody. 198 */ 199 if (!(pmd_val(pmd) & _PAGE_PRESENT)) { 200 union split_pmd old, new, *ptr; 201 202 ptr = (union split_pmd *)pmdp; 203 204 new.pmd = pmd; 205 206 /* xchg acts as a barrier before setting of the high bits */ 207 old.pmd_low = xchg(&ptr->pmd_low, new.pmd_low); 208 old.pmd_high = ptr->pmd_high; 209 ptr->pmd_high = new.pmd_high; 210 return old.pmd; 211 } 212 213 do { 214 old = *pmdp; 215 } while (cmpxchg64(&pmdp->pmd, old.pmd, pmd.pmd) != old.pmd); 216 217 return old; 218 } 219 #endif 220 221 #ifdef CONFIG_SMP 222 union split_pud { 223 struct { 224 u32 pud_low; 225 u32 pud_high; 226 }; 227 pud_t pud; 228 }; 229 230 static inline pud_t native_pudp_get_and_clear(pud_t *pudp) 231 { 232 union split_pud res, *orig = (union split_pud *)pudp; 233 234 #ifdef CONFIG_PAGE_TABLE_ISOLATION 235 pti_set_user_pgtbl(&pudp->p4d.pgd, __pgd(0)); 236 #endif 237 238 /* xchg acts as a barrier before setting of the high bits */ 239 res.pud_low = xchg(&orig->pud_low, 0); 240 res.pud_high = orig->pud_high; 241 orig->pud_high = 0; 242 243 return res.pud; 244 } 245 #else 246 #define native_pudp_get_and_clear(xp) native_local_pudp_get_and_clear(xp) 247 #endif 248 249 /* Encode and de-code a swap entry */ 250 #define SWP_TYPE_BITS 5 251 252 #define SWP_OFFSET_FIRST_BIT (_PAGE_BIT_PROTNONE + 1) 253 254 /* We always extract/encode the offset by shifting it all the way up, and then down again */ 255 #define SWP_OFFSET_SHIFT (SWP_OFFSET_FIRST_BIT + SWP_TYPE_BITS) 256 257 #define MAX_SWAPFILES_CHECK() BUILD_BUG_ON(MAX_SWAPFILES_SHIFT > 5) 258 #define __swp_type(x) (((x).val) & 0x1f) 259 #define __swp_offset(x) ((x).val >> 5) 260 #define __swp_entry(type, offset) ((swp_entry_t){(type) | (offset) << 5}) 261 262 /* 263 * Normally, __swp_entry() converts from arch-independent swp_entry_t to 264 * arch-dependent swp_entry_t, and __swp_entry_to_pte() just stores the result 265 * to pte. But here we have 32bit swp_entry_t and 64bit pte, and need to use the 266 * whole 64 bits. Thus, we shift the "real" arch-dependent conversion to 267 * __swp_entry_to_pte() through the following helper macro based on 64bit 268 * __swp_entry(). 269 */ 270 #define __swp_pteval_entry(type, offset) ((pteval_t) { \ 271 (~(pteval_t)(offset) << SWP_OFFSET_SHIFT >> SWP_TYPE_BITS) \ 272 | ((pteval_t)(type) << (64 - SWP_TYPE_BITS)) }) 273 274 #define __swp_entry_to_pte(x) ((pte_t){ .pte = \ 275 __swp_pteval_entry(__swp_type(x), __swp_offset(x)) }) 276 /* 277 * Analogically, __pte_to_swp_entry() doesn't just extract the arch-dependent 278 * swp_entry_t, but also has to convert it from 64bit to the 32bit 279 * intermediate representation, using the following macros based on 64bit 280 * __swp_type() and __swp_offset(). 281 */ 282 #define __pteval_swp_type(x) ((unsigned long)((x).pte >> (64 - SWP_TYPE_BITS))) 283 #define __pteval_swp_offset(x) ((unsigned long)(~((x).pte) << SWP_TYPE_BITS >> SWP_OFFSET_SHIFT)) 284 285 #define __pte_to_swp_entry(pte) (__swp_entry(__pteval_swp_type(pte), \ 286 __pteval_swp_offset(pte))) 287 288 #define gup_get_pte gup_get_pte 289 /* 290 * WARNING: only to be used in the get_user_pages_fast() implementation. 291 * 292 * With get_user_pages_fast(), we walk down the pagetables without taking 293 * any locks. For this we would like to load the pointers atomically, 294 * but that is not possible (without expensive cmpxchg8b) on PAE. What 295 * we do have is the guarantee that a PTE will only either go from not 296 * present to present, or present to not present or both -- it will not 297 * switch to a completely different present page without a TLB flush in 298 * between; something that we are blocking by holding interrupts off. 299 * 300 * Setting ptes from not present to present goes: 301 * 302 * ptep->pte_high = h; 303 * smp_wmb(); 304 * ptep->pte_low = l; 305 * 306 * And present to not present goes: 307 * 308 * ptep->pte_low = 0; 309 * smp_wmb(); 310 * ptep->pte_high = 0; 311 * 312 * We must ensure here that the load of pte_low sees 'l' iff pte_high 313 * sees 'h'. We load pte_high *after* loading pte_low, which ensures we 314 * don't see an older value of pte_high. *Then* we recheck pte_low, 315 * which ensures that we haven't picked up a changed pte high. We might 316 * have gotten rubbish values from pte_low and pte_high, but we are 317 * guaranteed that pte_low will not have the present bit set *unless* 318 * it is 'l'. Because get_user_pages_fast() only operates on present ptes 319 * we're safe. 320 */ 321 static inline pte_t gup_get_pte(pte_t *ptep) 322 { 323 pte_t pte; 324 325 do { 326 pte.pte_low = ptep->pte_low; 327 smp_rmb(); 328 pte.pte_high = ptep->pte_high; 329 smp_rmb(); 330 } while (unlikely(pte.pte_low != ptep->pte_low)); 331 332 return pte; 333 } 334 335 #include <asm/pgtable-invert.h> 336 337 #endif /* _ASM_X86_PGTABLE_3LEVEL_H */ 338