1 #ifndef _ASM_X86_TLBFLUSH_H
2 #define _ASM_X86_TLBFLUSH_H
3
4 #include <linux/mm.h>
5 #include <linux/sched.h>
6
7 #include <asm/processor.h>
8 #include <asm/cpufeature.h>
9 #include <asm/special_insns.h>
10 #include <asm/smp.h>
11
__invpcid(unsigned long pcid,unsigned long addr,unsigned long type)12 static inline void __invpcid(unsigned long pcid, unsigned long addr,
13 unsigned long type)
14 {
15 struct { u64 d[2]; } desc = { { pcid, addr } };
16
17 /*
18 * The memory clobber is because the whole point is to invalidate
19 * stale TLB entries and, especially if we're flushing global
20 * mappings, we don't want the compiler to reorder any subsequent
21 * memory accesses before the TLB flush.
22 *
23 * The hex opcode is invpcid (%ecx), %eax in 32-bit mode and
24 * invpcid (%rcx), %rax in long mode.
25 */
26 asm volatile (".byte 0x66, 0x0f, 0x38, 0x82, 0x01"
27 : : "m" (desc), "a" (type), "c" (&desc) : "memory");
28 }
29
30 #define INVPCID_TYPE_INDIV_ADDR 0
31 #define INVPCID_TYPE_SINGLE_CTXT 1
32 #define INVPCID_TYPE_ALL_INCL_GLOBAL 2
33 #define INVPCID_TYPE_ALL_NON_GLOBAL 3
34
35 /* Flush all mappings for a given pcid and addr, not including globals. */
invpcid_flush_one(unsigned long pcid,unsigned long addr)36 static inline void invpcid_flush_one(unsigned long pcid,
37 unsigned long addr)
38 {
39 __invpcid(pcid, addr, INVPCID_TYPE_INDIV_ADDR);
40 }
41
42 /* Flush all mappings for a given PCID, not including globals. */
invpcid_flush_single_context(unsigned long pcid)43 static inline void invpcid_flush_single_context(unsigned long pcid)
44 {
45 __invpcid(pcid, 0, INVPCID_TYPE_SINGLE_CTXT);
46 }
47
48 /* Flush all mappings, including globals, for all PCIDs. */
invpcid_flush_all(void)49 static inline void invpcid_flush_all(void)
50 {
51 __invpcid(0, 0, INVPCID_TYPE_ALL_INCL_GLOBAL);
52 }
53
54 /* Flush all mappings for all PCIDs except globals. */
invpcid_flush_all_nonglobals(void)55 static inline void invpcid_flush_all_nonglobals(void)
56 {
57 __invpcid(0, 0, INVPCID_TYPE_ALL_NON_GLOBAL);
58 }
59
60 #ifdef CONFIG_PARAVIRT
61 #include <asm/paravirt.h>
62 #else
63 #define __flush_tlb() __native_flush_tlb()
64 #define __flush_tlb_global() __native_flush_tlb_global()
65 #define __flush_tlb_single(addr) __native_flush_tlb_single(addr)
66 #endif
67
68 struct tlb_state {
69 struct mm_struct *active_mm;
70 int state;
71
72 /* Last user mm for optimizing IBPB */
73 union {
74 struct mm_struct *last_user_mm;
75 unsigned long last_user_mm_ibpb;
76 };
77
78 /*
79 * Access to this CR4 shadow and to H/W CR4 is protected by
80 * disabling interrupts when modifying either one.
81 */
82 unsigned long cr4;
83 };
84 DECLARE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate);
85
86 /* Initialize cr4 shadow for this CPU. */
cr4_init_shadow(void)87 static inline void cr4_init_shadow(void)
88 {
89 this_cpu_write(cpu_tlbstate.cr4, __read_cr4_safe());
90 }
91
92 /* Set in this cpu's CR4. */
cr4_set_bits(unsigned long mask)93 static inline void cr4_set_bits(unsigned long mask)
94 {
95 unsigned long cr4;
96
97 cr4 = this_cpu_read(cpu_tlbstate.cr4);
98 if ((cr4 | mask) != cr4) {
99 cr4 |= mask;
100 this_cpu_write(cpu_tlbstate.cr4, cr4);
101 __write_cr4(cr4);
102 }
103 }
104
105 /* Clear in this cpu's CR4. */
cr4_clear_bits(unsigned long mask)106 static inline void cr4_clear_bits(unsigned long mask)
107 {
108 unsigned long cr4;
109
110 cr4 = this_cpu_read(cpu_tlbstate.cr4);
111 if ((cr4 & ~mask) != cr4) {
112 cr4 &= ~mask;
113 this_cpu_write(cpu_tlbstate.cr4, cr4);
114 __write_cr4(cr4);
115 }
116 }
117
cr4_toggle_bits(unsigned long mask)118 static inline void cr4_toggle_bits(unsigned long mask)
119 {
120 unsigned long cr4;
121
122 cr4 = this_cpu_read(cpu_tlbstate.cr4);
123 cr4 ^= mask;
124 this_cpu_write(cpu_tlbstate.cr4, cr4);
125 __write_cr4(cr4);
126 }
127
128 /* Read the CR4 shadow. */
cr4_read_shadow(void)129 static inline unsigned long cr4_read_shadow(void)
130 {
131 return this_cpu_read(cpu_tlbstate.cr4);
132 }
133
134 /*
135 * Save some of cr4 feature set we're using (e.g. Pentium 4MB
136 * enable and PPro Global page enable), so that any CPU's that boot
137 * up after us can get the correct flags. This should only be used
138 * during boot on the boot cpu.
139 */
140 extern unsigned long mmu_cr4_features;
141 extern u32 *trampoline_cr4_features;
142
cr4_set_bits_and_update_boot(unsigned long mask)143 static inline void cr4_set_bits_and_update_boot(unsigned long mask)
144 {
145 mmu_cr4_features |= mask;
146 if (trampoline_cr4_features)
147 *trampoline_cr4_features = mmu_cr4_features;
148 cr4_set_bits(mask);
149 }
150
151 /*
152 * Declare a couple of kaiser interfaces here for convenience,
153 * to avoid the need for asm/kaiser.h in unexpected places.
154 */
155 #ifdef CONFIG_PAGE_TABLE_ISOLATION
156 extern int kaiser_enabled;
157 extern void kaiser_setup_pcid(void);
158 extern void kaiser_flush_tlb_on_return_to_user(void);
159 #else
160 #define kaiser_enabled 0
kaiser_setup_pcid(void)161 static inline void kaiser_setup_pcid(void)
162 {
163 }
kaiser_flush_tlb_on_return_to_user(void)164 static inline void kaiser_flush_tlb_on_return_to_user(void)
165 {
166 }
167 #endif
168
__native_flush_tlb(void)169 static inline void __native_flush_tlb(void)
170 {
171 /*
172 * If current->mm == NULL then we borrow a mm which may change during a
173 * task switch and therefore we must not be preempted while we write CR3
174 * back:
175 */
176 preempt_disable();
177 if (kaiser_enabled)
178 kaiser_flush_tlb_on_return_to_user();
179 native_write_cr3(native_read_cr3());
180 preempt_enable();
181 }
182
__native_flush_tlb_global_irq_disabled(void)183 static inline void __native_flush_tlb_global_irq_disabled(void)
184 {
185 unsigned long cr4;
186
187 cr4 = this_cpu_read(cpu_tlbstate.cr4);
188 if (cr4 & X86_CR4_PGE) {
189 /* clear PGE and flush TLB of all entries */
190 native_write_cr4(cr4 & ~X86_CR4_PGE);
191 /* restore PGE as it was before */
192 native_write_cr4(cr4);
193 } else {
194 /* do it with cr3, letting kaiser flush user PCID */
195 __native_flush_tlb();
196 }
197 }
198
__native_flush_tlb_global(void)199 static inline void __native_flush_tlb_global(void)
200 {
201 unsigned long flags;
202
203 if (this_cpu_has(X86_FEATURE_INVPCID)) {
204 /*
205 * Using INVPCID is considerably faster than a pair of writes
206 * to CR4 sandwiched inside an IRQ flag save/restore.
207 *
208 * Note, this works with CR4.PCIDE=0 or 1.
209 */
210 invpcid_flush_all();
211 return;
212 }
213
214 /*
215 * Read-modify-write to CR4 - protect it from preemption and
216 * from interrupts. (Use the raw variant because this code can
217 * be called from deep inside debugging code.)
218 */
219 raw_local_irq_save(flags);
220 __native_flush_tlb_global_irq_disabled();
221 raw_local_irq_restore(flags);
222 }
223
__native_flush_tlb_single(unsigned long addr)224 static inline void __native_flush_tlb_single(unsigned long addr)
225 {
226 /*
227 * SIMICS #GP's if you run INVPCID with type 2/3
228 * and X86_CR4_PCIDE clear. Shame!
229 *
230 * The ASIDs used below are hard-coded. But, we must not
231 * call invpcid(type=1/2) before CR4.PCIDE=1. Just call
232 * invlpg in the case we are called early.
233 */
234
235 if (!this_cpu_has(X86_FEATURE_INVPCID_SINGLE)) {
236 if (kaiser_enabled)
237 kaiser_flush_tlb_on_return_to_user();
238 asm volatile("invlpg (%0)" ::"r" (addr) : "memory");
239 return;
240 }
241 /* Flush the address out of both PCIDs. */
242 /*
243 * An optimization here might be to determine addresses
244 * that are only kernel-mapped and only flush the kernel
245 * ASID. But, userspace flushes are probably much more
246 * important performance-wise.
247 *
248 * In the KAISER disabled case, do an INVLPG to make sure
249 * the mapping is flushed in case it is a global one.
250 */
251 if (kaiser_enabled) {
252 invpcid_flush_one(X86_CR3_PCID_ASID_USER, addr);
253 invpcid_flush_one(X86_CR3_PCID_ASID_KERN, addr);
254 } else {
255 asm volatile("invlpg (%0)" ::"r" (addr) : "memory");
256 }
257 }
258
__flush_tlb_all(void)259 static inline void __flush_tlb_all(void)
260 {
261 __flush_tlb_global();
262 /*
263 * Note: if we somehow had PCID but not PGE, then this wouldn't work --
264 * we'd end up flushing kernel translations for the current ASID but
265 * we might fail to flush kernel translations for other cached ASIDs.
266 *
267 * To avoid this issue, we force PCID off if PGE is off.
268 */
269 }
270
__flush_tlb_one(unsigned long addr)271 static inline void __flush_tlb_one(unsigned long addr)
272 {
273 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
274 __flush_tlb_single(addr);
275 }
276
277 #define TLB_FLUSH_ALL -1UL
278
279 /*
280 * TLB flushing:
281 *
282 * - flush_tlb_all() flushes all processes TLBs
283 * - flush_tlb_mm(mm) flushes the specified mm context TLB's
284 * - flush_tlb_page(vma, vmaddr) flushes one page
285 * - flush_tlb_range(vma, start, end) flushes a range of pages
286 * - flush_tlb_kernel_range(start, end) flushes a range of kernel pages
287 * - flush_tlb_others(cpumask, mm, start, end) flushes TLBs on other cpus
288 *
289 * ..but the i386 has somewhat limited tlb flushing capabilities,
290 * and page-granular flushes are available only on i486 and up.
291 */
292
293 #define local_flush_tlb() __flush_tlb()
294
295 #define flush_tlb_mm(mm) flush_tlb_mm_range(mm, 0UL, TLB_FLUSH_ALL, 0UL)
296
297 #define flush_tlb_range(vma, start, end) \
298 flush_tlb_mm_range(vma->vm_mm, start, end, vma->vm_flags)
299
300 extern void flush_tlb_all(void);
301 extern void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
302 unsigned long end, unsigned long vmflag);
303 extern void flush_tlb_kernel_range(unsigned long start, unsigned long end);
304
flush_tlb_page(struct vm_area_struct * vma,unsigned long a)305 static inline void flush_tlb_page(struct vm_area_struct *vma, unsigned long a)
306 {
307 flush_tlb_mm_range(vma->vm_mm, a, a + PAGE_SIZE, VM_NONE);
308 }
309
310 void native_flush_tlb_others(const struct cpumask *cpumask,
311 struct mm_struct *mm,
312 unsigned long start, unsigned long end);
313
314 #define TLBSTATE_OK 1
315 #define TLBSTATE_LAZY 2
316
reset_lazy_tlbstate(void)317 static inline void reset_lazy_tlbstate(void)
318 {
319 this_cpu_write(cpu_tlbstate.state, 0);
320 this_cpu_write(cpu_tlbstate.active_mm, &init_mm);
321 }
322
323 #ifndef CONFIG_PARAVIRT
324 #define flush_tlb_others(mask, mm, start, end) \
325 native_flush_tlb_others(mask, mm, start, end)
326 #endif
327
328 #endif /* _ASM_X86_TLBFLUSH_H */
329