arch/x86/include/asm/mmu_context.h

   1 #ifndef _ASM_X86_MMU_CONTEXT_H
   2 #define _ASM_X86_MMU_CONTEXT_H
   3
   4 #include <asm/desc.h>
   5 #include <linux/atomic.h>
   6 #include <linux/mm_types.h>
   7
   8 #include <trace/events/tlb.h>
   9
  10 #include <asm/pgalloc.h>
  11 #include <asm/tlbflush.h>
  12 #include <asm/paravirt.h>
  13 #include <asm/mpx.h>
  14 #ifndef CONFIG_PARAVIRT
  15 static inline void paravirt_activate_mm(struct mm_struct *prev,
  16                                         struct mm_struct *next)
  17 {
  18 }
  19 #endif  /* !CONFIG_PARAVIRT */
  20
  21 #ifdef CONFIG_PERF_EVENTS
  22 extern struct static_key rdpmc_always_available;
  23
  24 static inline void load_mm_cr4(struct mm_struct *mm)
  25 {
  26         if (static_key_false(&rdpmc_always_available) ||
  27             atomic_read(&mm->context.perf_rdpmc_allowed))
  28                 cr4_set_bits(X86_CR4_PCE);
  29         else
  30                 cr4_clear_bits(X86_CR4_PCE);
  31 }
  32 #else
  33 static inline void load_mm_cr4(struct mm_struct *mm) {}
  34 #endif
  35
  36 #ifdef CONFIG_MODIFY_LDT_SYSCALL
  37 /*
  38  * ldt_structs can be allocated, used, and freed, but they are never
  39  * modified while live.
  40  */
  41 struct ldt_struct {
  42         /*
  43          * Xen requires page-aligned LDTs with special permissions.  This is
  44          * needed to prevent us from installing evil descriptors such as
  45          * call gates.  On native, we could merge the ldt_struct and LDT
  46          * allocations, but it's not worth trying to optimize.
  47          */
  48         struct desc_struct *entries;
  49         int size;
  50 };
  51
  52 /*
  53  * Used for LDT copy/destruction.
  54  */
  55 int init_new_context(struct task_struct *tsk, struct mm_struct *mm);
  56 void destroy_context(struct mm_struct *mm);
  57 #else   /* CONFIG_MODIFY_LDT_SYSCALL */
  58 static inline int init_new_context(struct task_struct *tsk,
  59                                    struct mm_struct *mm)
  60 {
  61         return 0;
  62 }
  63 static inline void destroy_context(struct mm_struct *mm) {}
  64 #endif
  65
  66 static inline void load_mm_ldt(struct mm_struct *mm)
  67 {
  68 #ifdef CONFIG_MODIFY_LDT_SYSCALL
  69         struct ldt_struct *ldt;
  70
  71         /* lockless_dereference synchronizes with smp_store_release */
  72         ldt = lockless_dereference(mm->context.ldt);
  73
  74         /*
  75          * Any change to mm->context.ldt is followed by an IPI to all
  76          * CPUs with the mm active.  The LDT will not be freed until
  77          * after the IPI is handled by all such CPUs.  This means that,
  78          * if the ldt_struct changes before we return, the values we see
  79          * will be safe, and the new values will be loaded before we run
  80          * any user code.
  81          *
  82          * NB: don't try to convert this to use RCU without extreme care.
  83          * We would still need IRQs off, because we don't want to change
  84          * the local LDT after an IPI loaded a newer value than the one
  85          * that we can see.
  86          */
  87
  88         if (unlikely(ldt))
  89                 set_ldt(ldt->entries, ldt->size);
  90         else
  91                 clear_LDT();
  92 #else
  93         clear_LDT();
  94 #endif
  95
  96         DEBUG_LOCKS_WARN_ON(preemptible());
  97 }
  98
  99 static inline void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
 100 {
 101 #ifdef CONFIG_SMP
 102         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
 103                 this_cpu_write(cpu_tlbstate.state, TLBSTATE_LAZY);
 104 #endif
 105 }
 106
 107 static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
 108                              struct task_struct *tsk)
 109 {
 110         unsigned cpu = smp_processor_id();
 111
 112         if (likely(prev != next)) {
 113 #ifdef CONFIG_SMP
 114                 this_cpu_write(cpu_tlbstate.state, TLBSTATE_OK);
 115                 this_cpu_write(cpu_tlbstate.active_mm, next);
 116 #endif
 117                 cpumask_set_cpu(cpu, mm_cpumask(next));
 118
 119                 /*
 120                  * Re-load page tables.
 121                  *
 122                  * This logic has an ordering constraint:
 123                  *
 124                  *  CPU 0: Write to a PTE for 'next'
 125                  *  CPU 0: load bit 1 in mm_cpumask.  if nonzero, send IPI.
 126                  *  CPU 1: set bit 1 in next's mm_cpumask
 127                  *  CPU 1: load from the PTE that CPU 0 writes (implicit)
 128                  *
 129                  * We need to prevent an outcome in which CPU 1 observes
 130                  * the new PTE value and CPU 0 observes bit 1 clear in
 131                  * mm_cpumask.  (If that occurs, then the IPI will never
 132                  * be sent, and CPU 0's TLB will contain a stale entry.)
 133                  *
 134                  * The bad outcome can occur if either CPU's load is
 135                  * reordered before that CPU's store, so both CPUs must
 136                  * execute full barriers to prevent this from happening.
 137                  *
 138                  * Thus, switch_mm needs a full barrier between the
 139                  * store to mm_cpumask and any operation that could load
 140                  * from next->pgd.  TLB fills are special and can happen
 141                  * due to instruction fetches or for no reason at all,
 142                  * and neither LOCK nor MFENCE orders them.
 143                  * Fortunately, load_cr3() is serializing and gives the
 144                  * ordering guarantee we need.
 145                  *
 146                  */
 147                 load_cr3(next->pgd);
 148
 149                 trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
 150
 151                 /* Stop flush ipis for the previous mm */
 152                 cpumask_clear_cpu(cpu, mm_cpumask(prev));
 153
 154                 /* Load per-mm CR4 state */
 155                 load_mm_cr4(next);
 156
 157 #ifdef CONFIG_MODIFY_LDT_SYSCALL
 158                 /*
 159                  * Load the LDT, if the LDT is different.
 160                  *
 161                  * It's possible that prev->context.ldt doesn't match
 162                  * the LDT register.  This can happen if leave_mm(prev)
 163                  * was called and then modify_ldt changed
 164                  * prev->context.ldt but suppressed an IPI to this CPU.
 165                  * In this case, prev->context.ldt != NULL, because we
 166                  * never set context.ldt to NULL while the mm still
 167                  * exists.  That means that next->context.ldt !=
 168                  * prev->context.ldt, because mms never share an LDT.
 169                  */
 170                 if (unlikely(prev->context.ldt != next->context.ldt))
 171                         load_mm_ldt(next);
 172 #endif
 173         }
 174 #ifdef CONFIG_SMP
 175           else {
 176                 this_cpu_write(cpu_tlbstate.state, TLBSTATE_OK);
 177                 BUG_ON(this_cpu_read(cpu_tlbstate.active_mm) != next);
 178
 179                 if (!cpumask_test_cpu(cpu, mm_cpumask(next))) {
 180                         /*
 181                          * On established mms, the mm_cpumask is only changed
 182                          * from irq context, from ptep_clear_flush() while in
 183                          * lazy tlb mode, and here. Irqs are blocked during
 184                          * schedule, protecting us from simultaneous changes.
 185                          */
 186                         cpumask_set_cpu(cpu, mm_cpumask(next));
 187
 188                         /*
 189                          * We were in lazy tlb mode and leave_mm disabled
 190                          * tlb flush IPI delivery. We must reload CR3
 191                          * to make sure to use no freed page tables.
 192                          *
 193                          * As above, load_cr3() is serializing and orders TLB
 194                          * fills with respect to the mm_cpumask write.
 195                          */
 196                         load_cr3(next->pgd);
 197                         trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
 198                         load_mm_cr4(next);
 199                         load_mm_ldt(next);
 200                 }
 201         }
 202 #endif
 203 }
 204
 205 #define activate_mm(prev, next)                 \
 206 do {                                            \
 207         paravirt_activate_mm((prev), (next));   \
 208         switch_mm((prev), (next), NULL);        \
 209 } while (0);
 210
 211 #ifdef CONFIG_X86_32
 212 #define deactivate_mm(tsk, mm)                  \
 213 do {                                            \
 214         lazy_load_gs(0);                        \
 215 } while (0)
 216 #else
 217 #define deactivate_mm(tsk, mm)                  \
 218 do {                                            \
 219         load_gs_index(0);                       \
 220         loadsegment(fs, 0);                     \
 221 } while (0)
 222 #endif
 223
 224 static inline void arch_dup_mmap(struct mm_struct *oldmm,
 225                                  struct mm_struct *mm)
 226 {
 227         paravirt_arch_dup_mmap(oldmm, mm);
 228 }
 229
 230 static inline void arch_exit_mmap(struct mm_struct *mm)
 231 {
 232         paravirt_arch_exit_mmap(mm);
 233 }
 234
 235 #ifdef CONFIG_X86_64
 236 static inline bool is_64bit_mm(struct mm_struct *mm)
 237 {
 238         return  !config_enabled(CONFIG_IA32_EMULATION) ||
 239                 !(mm->context.ia32_compat == TIF_IA32);
 240 }
 241 #else
 242 static inline bool is_64bit_mm(struct mm_struct *mm)
 243 {
 244         return false;
 245 }
 246 #endif
 247
 248 static inline void arch_bprm_mm_init(struct mm_struct *mm,
 249                 struct vm_area_struct *vma)
 250 {
 251         mpx_mm_init(mm);
 252 }
 253
 254 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma,
 255                               unsigned long start, unsigned long end)
 256 {
 257         /*
 258          * mpx_notify_unmap() goes and reads a rarely-hot
 259          * cacheline in the mm_struct.  That can be expensive
 260          * enough to be seen in profiles.
 261          *
 262          * The mpx_notify_unmap() call and its contents have been
 263          * observed to affect munmap() performance on hardware
 264          * where MPX is not present.
 265          *
 266          * The unlikely() optimizes for the fast case: no MPX
 267          * in the CPU, or no MPX use in the process.  Even if
 268          * we get this wrong (in the unlikely event that MPX
 269          * is widely enabled on some system) the overhead of
 270          * MPX itself (reading bounds tables) is expected to
 271          * overwhelm the overhead of getting this unlikely()
 272          * consistently wrong.
 273          */
 274         if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX)))
 275                 mpx_notify_unmap(mm, vma, start, end);
 276 }
 277
 278 static inline int vma_pkey(struct vm_area_struct *vma)
 279 {
 280         u16 pkey = 0;
 281 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
 282         unsigned long vma_pkey_mask = VM_PKEY_BIT0 | VM_PKEY_BIT1 |
 283                                       VM_PKEY_BIT2 | VM_PKEY_BIT3;
 284         pkey = (vma->vm_flags & vma_pkey_mask) >> VM_PKEY_SHIFT;
 285 #endif
 286         return pkey;
 287 }
 288
 289 static inline bool __pkru_allows_pkey(u16 pkey, bool write)
 290 {
 291         u32 pkru = read_pkru();
 292
 293         if (!__pkru_allows_read(pkru, pkey))
 294                 return false;
 295         if (write && !__pkru_allows_write(pkru, pkey))
 296                 return false;
 297
 298         return true;
 299 }
 300
 301 /*
 302  * We only want to enforce protection keys on the current process
 303  * because we effectively have no access to PKRU for other
 304  * processes or any way to tell *which * PKRU in a threaded
 305  * process we could use.
 306  *
 307  * So do not enforce things if the VMA is not from the current
 308  * mm, or if we are in a kernel thread.
 309  */
 310 static inline bool vma_is_foreign(struct vm_area_struct *vma)
 311 {
 312         if (!current->mm)
 313                 return true;
 314         /*
 315          * Should PKRU be enforced on the access to this VMA?  If
 316          * the VMA is from another process, then PKRU has no
 317          * relevance and should not be enforced.
 318          */
 319         if (current->mm != vma->vm_mm)
 320                 return true;
 321
 322         return false;
 323 }
 324
 325 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma, bool write)
 326 {
 327         /* allow access if the VMA is not one from this process */
 328         if (vma_is_foreign(vma))
 329                 return true;
 330         return __pkru_allows_pkey(vma_pkey(vma), write);
 331 }
 332
 333 static inline bool arch_pte_access_permitted(pte_t pte, bool write)
 334 {
 335         return __pkru_allows_pkey(pte_flags_pkey(pte_flags(pte)), write);
 336 }
 337
 338 #endif /* _ASM_X86_MMU_CONTEXT_H */