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 #include <linux/pkeys.h>
   8
   9 #include <trace/events/tlb.h>
  10
  11 #include <asm/pgalloc.h>
  12 #include <asm/tlbflush.h>
  13 #include <asm/paravirt.h>
  14 #include <asm/mpx.h>
  15 #ifndef CONFIG_PARAVIRT
  16 static inline void paravirt_activate_mm(struct mm_struct *prev,
  17                                         struct mm_struct *next)
  18 {
  19 }
  20 #endif  /* !CONFIG_PARAVIRT */
  21
  22 #ifdef CONFIG_PERF_EVENTS
  23 extern struct static_key rdpmc_always_available;
  24
  25 static inline void load_mm_cr4(struct mm_struct *mm)
  26 {
  27         if (static_key_false(&rdpmc_always_available) ||
  28             atomic_read(&mm->context.perf_rdpmc_allowed))
  29                 cr4_set_bits(X86_CR4_PCE);
  30         else
  31                 cr4_clear_bits(X86_CR4_PCE);
  32 }
  33 #else
  34 static inline void load_mm_cr4(struct mm_struct *mm) {}
  35 #endif
  36
  37 #ifdef CONFIG_MODIFY_LDT_SYSCALL
  38 /*
  39  * ldt_structs can be allocated, used, and freed, but they are never
  40  * modified while live.
  41  */
  42 struct ldt_struct {
  43         /*
  44          * Xen requires page-aligned LDTs with special permissions.  This is
  45          * needed to prevent us from installing evil descriptors such as
  46          * call gates.  On native, we could merge the ldt_struct and LDT
  47          * allocations, but it's not worth trying to optimize.
  48          */
  49         struct desc_struct *entries;
  50         unsigned int nr_entries;
  51 };
  52
  53 /*
  54  * Used for LDT copy/destruction.
  55  */
  56 int init_new_context_ldt(struct task_struct *tsk, struct mm_struct *mm);
  57 void destroy_context_ldt(struct mm_struct *mm);
  58 #else   /* CONFIG_MODIFY_LDT_SYSCALL */
  59 static inline int init_new_context_ldt(struct task_struct *tsk,
  60                                        struct mm_struct *mm)
  61 {
  62         return 0;
  63 }
  64 static inline void destroy_context_ldt(struct mm_struct *mm) {}
  65 #endif
  66
  67 static inline void load_mm_ldt(struct mm_struct *mm)
  68 {
  69 #ifdef CONFIG_MODIFY_LDT_SYSCALL
  70         struct ldt_struct *ldt;
  71
  72         /* lockless_dereference synchronizes with smp_store_release */
  73         ldt = lockless_dereference(mm->context.ldt);
  74
  75         /*
  76          * Any change to mm->context.ldt is followed by an IPI to all
  77          * CPUs with the mm active.  The LDT will not be freed until
  78          * after the IPI is handled by all such CPUs.  This means that,
  79          * if the ldt_struct changes before we return, the values we see
  80          * will be safe, and the new values will be loaded before we run
  81          * any user code.
  82          *
  83          * NB: don't try to convert this to use RCU without extreme care.
  84          * We would still need IRQs off, because we don't want to change
  85          * the local LDT after an IPI loaded a newer value than the one
  86          * that we can see.
  87          */
  88
  89         if (unlikely(ldt))
  90                 set_ldt(ldt->entries, ldt->nr_entries);
  91         else
  92                 clear_LDT();
  93 #else
  94         clear_LDT();
  95 #endif
  96 }
  97
  98 static inline void switch_ldt(struct mm_struct *prev, struct mm_struct *next)
  99 {
 100 #ifdef CONFIG_MODIFY_LDT_SYSCALL
 101         /*
 102          * Load the LDT if either the old or new mm had an LDT.
 103          *
 104          * An mm will never go from having an LDT to not having an LDT.  Two
 105          * mms never share an LDT, so we don't gain anything by checking to
 106          * see whether the LDT changed.  There's also no guarantee that
 107          * prev->context.ldt actually matches LDTR, but, if LDTR is non-NULL,
 108          * then prev->context.ldt will also be non-NULL.
 109          *
 110          * If we really cared, we could optimize the case where prev == next
 111          * and we're exiting lazy mode.  Most of the time, if this happens,
 112          * we don't actually need to reload LDTR, but modify_ldt() is mostly
 113          * used by legacy code and emulators where we don't need this level of
 114          * performance.
 115          *
 116          * This uses | instead of || because it generates better code.
 117          */
 118         if (unlikely((unsigned long)prev->context.ldt |
 119                      (unsigned long)next->context.ldt))
 120                 load_mm_ldt(next);
 121 #endif
 122
 123         DEBUG_LOCKS_WARN_ON(preemptible());
 124 }
 125
 126 static inline void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
 127 {
 128         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
 129                 this_cpu_write(cpu_tlbstate.state, TLBSTATE_LAZY);
 130 }
 131
 132 static inline int init_new_context(struct task_struct *tsk,
 133                                    struct mm_struct *mm)
 134 {
 135         #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
 136         if (cpu_feature_enabled(X86_FEATURE_OSPKE)) {
 137                 /* pkey 0 is the default and always allocated */
 138                 mm->context.pkey_allocation_map = 0x1;
 139                 /* -1 means unallocated or invalid */
 140                 mm->context.execute_only_pkey = -1;
 141         }
 142         #endif
 143         init_new_context_ldt(tsk, mm);
 144
 145         return 0;
 146 }
 147 static inline void destroy_context(struct mm_struct *mm)
 148 {
 149         destroy_context_ldt(mm);
 150 }
 151
 152 extern void switch_mm(struct mm_struct *prev, struct mm_struct *next,
 153                       struct task_struct *tsk);
 154
 155 extern void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
 156                                struct task_struct *tsk);
 157 #define switch_mm_irqs_off switch_mm_irqs_off
 158
 159 #define activate_mm(prev, next)                 \
 160 do {                                            \
 161         paravirt_activate_mm((prev), (next));   \
 162         switch_mm((prev), (next), NULL);        \
 163 } while (0);
 164
 165 #ifdef CONFIG_X86_32
 166 #define deactivate_mm(tsk, mm)                  \
 167 do {                                            \
 168         lazy_load_gs(0);                        \
 169 } while (0)
 170 #else
 171 #define deactivate_mm(tsk, mm)                  \
 172 do {                                            \
 173         load_gs_index(0);                       \
 174         loadsegment(fs, 0);                     \
 175 } while (0)
 176 #endif
 177
 178 static inline void arch_dup_mmap(struct mm_struct *oldmm,
 179                                  struct mm_struct *mm)
 180 {
 181         paravirt_arch_dup_mmap(oldmm, mm);
 182 }
 183
 184 static inline void arch_exit_mmap(struct mm_struct *mm)
 185 {
 186         paravirt_arch_exit_mmap(mm);
 187 }
 188
 189 #ifdef CONFIG_X86_64
 190 static inline bool is_64bit_mm(struct mm_struct *mm)
 191 {
 192         return  !IS_ENABLED(CONFIG_IA32_EMULATION) ||
 193                 !(mm->context.ia32_compat == TIF_IA32);
 194 }
 195 #else
 196 static inline bool is_64bit_mm(struct mm_struct *mm)
 197 {
 198         return false;
 199 }
 200 #endif
 201
 202 static inline void arch_bprm_mm_init(struct mm_struct *mm,
 203                 struct vm_area_struct *vma)
 204 {
 205         mpx_mm_init(mm);
 206 }
 207
 208 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma,
 209                               unsigned long start, unsigned long end)
 210 {
 211         /*
 212          * mpx_notify_unmap() goes and reads a rarely-hot
 213          * cacheline in the mm_struct.  That can be expensive
 214          * enough to be seen in profiles.
 215          *
 216          * The mpx_notify_unmap() call and its contents have been
 217          * observed to affect munmap() performance on hardware
 218          * where MPX is not present.
 219          *
 220          * The unlikely() optimizes for the fast case: no MPX
 221          * in the CPU, or no MPX use in the process.  Even if
 222          * we get this wrong (in the unlikely event that MPX
 223          * is widely enabled on some system) the overhead of
 224          * MPX itself (reading bounds tables) is expected to
 225          * overwhelm the overhead of getting this unlikely()
 226          * consistently wrong.
 227          */
 228         if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX)))
 229                 mpx_notify_unmap(mm, vma, start, end);
 230 }
 231
 232 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
 233 static inline int vma_pkey(struct vm_area_struct *vma)
 234 {
 235         unsigned long vma_pkey_mask = VM_PKEY_BIT0 | VM_PKEY_BIT1 |
 236                                       VM_PKEY_BIT2 | VM_PKEY_BIT3;
 237
 238         return (vma->vm_flags & vma_pkey_mask) >> VM_PKEY_SHIFT;
 239 }
 240 #else
 241 static inline int vma_pkey(struct vm_area_struct *vma)
 242 {
 243         return 0;
 244 }
 245 #endif
 246
 247 /*
 248  * We only want to enforce protection keys on the current process
 249  * because we effectively have no access to PKRU for other
 250  * processes or any way to tell *which * PKRU in a threaded
 251  * process we could use.
 252  *
 253  * So do not enforce things if the VMA is not from the current
 254  * mm, or if we are in a kernel thread.
 255  */
 256 static inline bool vma_is_foreign(struct vm_area_struct *vma)
 257 {
 258         if (!current->mm)
 259                 return true;
 260         /*
 261          * Should PKRU be enforced on the access to this VMA?  If
 262          * the VMA is from another process, then PKRU has no
 263          * relevance and should not be enforced.
 264          */
 265         if (current->mm != vma->vm_mm)
 266                 return true;
 267
 268         return false;
 269 }
 270
 271 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma,
 272                 bool write, bool execute, bool foreign)
 273 {
 274         /* pkeys never affect instruction fetches */
 275         if (execute)
 276                 return true;
 277         /* allow access if the VMA is not one from this process */
 278         if (foreign || vma_is_foreign(vma))
 279                 return true;
 280         return __pkru_allows_pkey(vma_pkey(vma), write);
 281 }
 282
 283
 284 /*
 285  * This can be used from process context to figure out what the value of
 286  * CR3 is without needing to do a (slow) __read_cr3().
 287  *
 288  * It's intended to be used for code like KVM that sneakily changes CR3
 289  * and needs to restore it.  It needs to be used very carefully.
 290  */
 291 static inline unsigned long __get_current_cr3_fast(void)
 292 {
 293         unsigned long cr3 = __pa(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd);
 294
 295         /* For now, be very restrictive about when this can be called. */
 296         VM_WARN_ON(in_nmi() || preemptible());
 297
 298         VM_BUG_ON(cr3 != __read_cr3());
 299         return cr3;
 300 }
 301
 302 #endif /* _ASM_X86_MMU_CONTEXT_H */