1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/init.h>
5 #include <linux/spinlock.h>
7 #include <linux/interrupt.h>
8 #include <linux/export.h>
10 #include <linux/debugfs.h>
11 #include <linux/sched/smt.h>
12 #include <linux/task_work.h>
13 #include <linux/mmu_notifier.h>
15 #include <asm/tlbflush.h>
16 #include <asm/mmu_context.h>
17 #include <asm/nospec-branch.h>
18 #include <asm/cache.h>
19 #include <asm/cacheflush.h>
21 #include <asm/perf_event.h>
23 #include "mm_internal.h"
25 #ifdef CONFIG_PARAVIRT
28 # define STATIC_NOPV static
29 # define __flush_tlb_local native_flush_tlb_local
30 # define __flush_tlb_global native_flush_tlb_global
31 # define __flush_tlb_one_user(addr) native_flush_tlb_one_user(addr)
32 # define __flush_tlb_multi(msk, info) native_flush_tlb_multi(msk, info)
36 * TLB flushing, formerly SMP-only
39 * These mean you can really definitely utterly forget about
40 * writing to user space from interrupts. (Its not allowed anyway).
42 * Optimizations Manfred Spraul <manfred@colorfullife.com>
44 * More scalable flush, from Andi Kleen
46 * Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi
50 * Bits to mangle the TIF_SPEC_* state into the mm pointer which is
51 * stored in cpu_tlb_state.last_user_mm_spec.
53 #define LAST_USER_MM_IBPB 0x1UL
54 #define LAST_USER_MM_L1D_FLUSH 0x2UL
55 #define LAST_USER_MM_SPEC_MASK (LAST_USER_MM_IBPB | LAST_USER_MM_L1D_FLUSH)
57 /* Bits to set when tlbstate and flush is (re)initialized */
58 #define LAST_USER_MM_INIT LAST_USER_MM_IBPB
61 * The x86 feature is called PCID (Process Context IDentifier). It is similar
62 * to what is traditionally called ASID on the RISC processors.
64 * We don't use the traditional ASID implementation, where each process/mm gets
65 * its own ASID and flush/restart when we run out of ASID space.
67 * Instead we have a small per-cpu array of ASIDs and cache the last few mm's
68 * that came by on this CPU, allowing cheaper switch_mm between processes on
71 * We end up with different spaces for different things. To avoid confusion we
72 * use different names for each of them:
74 * ASID - [0, TLB_NR_DYN_ASIDS-1]
75 * the canonical identifier for an mm
77 * kPCID - [1, TLB_NR_DYN_ASIDS]
78 * the value we write into the PCID part of CR3; corresponds to the
79 * ASID+1, because PCID 0 is special.
81 * uPCID - [2048 + 1, 2048 + TLB_NR_DYN_ASIDS]
82 * for KPTI each mm has two address spaces and thus needs two
83 * PCID values, but we can still do with a single ASID denomination
84 * for each mm. Corresponds to kPCID + 2048.
88 /* There are 12 bits of space for ASIDS in CR3 */
89 #define CR3_HW_ASID_BITS 12
92 * When enabled, PAGE_TABLE_ISOLATION consumes a single bit for
93 * user/kernel switches
95 #ifdef CONFIG_PAGE_TABLE_ISOLATION
96 # define PTI_CONSUMED_PCID_BITS 1
98 # define PTI_CONSUMED_PCID_BITS 0
101 #define CR3_AVAIL_PCID_BITS (X86_CR3_PCID_BITS - PTI_CONSUMED_PCID_BITS)
104 * ASIDs are zero-based: 0->MAX_AVAIL_ASID are valid. -1 below to account
105 * for them being zero-based. Another -1 is because PCID 0 is reserved for
106 * use by non-PCID-aware users.
108 #define MAX_ASID_AVAILABLE ((1 << CR3_AVAIL_PCID_BITS) - 2)
111 * Given @asid, compute kPCID
113 static inline u16 kern_pcid(u16 asid)
115 VM_WARN_ON_ONCE(asid > MAX_ASID_AVAILABLE);
117 #ifdef CONFIG_PAGE_TABLE_ISOLATION
119 * Make sure that the dynamic ASID space does not conflict with the
120 * bit we are using to switch between user and kernel ASIDs.
122 BUILD_BUG_ON(TLB_NR_DYN_ASIDS >= (1 << X86_CR3_PTI_PCID_USER_BIT));
125 * The ASID being passed in here should have respected the
126 * MAX_ASID_AVAILABLE and thus never have the switch bit set.
128 VM_WARN_ON_ONCE(asid & (1 << X86_CR3_PTI_PCID_USER_BIT));
131 * The dynamically-assigned ASIDs that get passed in are small
132 * (<TLB_NR_DYN_ASIDS). They never have the high switch bit set,
133 * so do not bother to clear it.
135 * If PCID is on, ASID-aware code paths put the ASID+1 into the
136 * PCID bits. This serves two purposes. It prevents a nasty
137 * situation in which PCID-unaware code saves CR3, loads some other
138 * value (with PCID == 0), and then restores CR3, thus corrupting
139 * the TLB for ASID 0 if the saved ASID was nonzero. It also means
140 * that any bugs involving loading a PCID-enabled CR3 with
141 * CR4.PCIDE off will trigger deterministically.
147 * Given @asid, compute uPCID
149 static inline u16 user_pcid(u16 asid)
151 u16 ret = kern_pcid(asid);
152 #ifdef CONFIG_PAGE_TABLE_ISOLATION
153 ret |= 1 << X86_CR3_PTI_PCID_USER_BIT;
158 static inline unsigned long build_cr3(pgd_t *pgd, u16 asid, unsigned long lam)
160 unsigned long cr3 = __sme_pa(pgd) | lam;
162 if (static_cpu_has(X86_FEATURE_PCID)) {
163 VM_WARN_ON_ONCE(asid > MAX_ASID_AVAILABLE);
164 cr3 |= kern_pcid(asid);
166 VM_WARN_ON_ONCE(asid != 0);
172 static inline unsigned long build_cr3_noflush(pgd_t *pgd, u16 asid,
176 * Use boot_cpu_has() instead of this_cpu_has() as this function
177 * might be called during early boot. This should work even after
178 * boot because all CPU's the have same capabilities:
180 VM_WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_PCID));
181 return build_cr3(pgd, asid, lam) | CR3_NOFLUSH;
185 * We get here when we do something requiring a TLB invalidation
186 * but could not go invalidate all of the contexts. We do the
187 * necessary invalidation by clearing out the 'ctx_id' which
188 * forces a TLB flush when the context is loaded.
190 static void clear_asid_other(void)
195 * This is only expected to be set if we have disabled
196 * kernel _PAGE_GLOBAL pages.
198 if (!static_cpu_has(X86_FEATURE_PTI)) {
203 for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
204 /* Do not need to flush the current asid */
205 if (asid == this_cpu_read(cpu_tlbstate.loaded_mm_asid))
208 * Make sure the next time we go to switch to
209 * this asid, we do a flush:
211 this_cpu_write(cpu_tlbstate.ctxs[asid].ctx_id, 0);
213 this_cpu_write(cpu_tlbstate.invalidate_other, false);
216 atomic64_t last_mm_ctx_id = ATOMIC64_INIT(1);
219 static void choose_new_asid(struct mm_struct *next, u64 next_tlb_gen,
220 u16 *new_asid, bool *need_flush)
224 if (!static_cpu_has(X86_FEATURE_PCID)) {
230 if (this_cpu_read(cpu_tlbstate.invalidate_other))
233 for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
234 if (this_cpu_read(cpu_tlbstate.ctxs[asid].ctx_id) !=
235 next->context.ctx_id)
239 *need_flush = (this_cpu_read(cpu_tlbstate.ctxs[asid].tlb_gen) <
245 * We don't currently own an ASID slot on this CPU.
248 *new_asid = this_cpu_add_return(cpu_tlbstate.next_asid, 1) - 1;
249 if (*new_asid >= TLB_NR_DYN_ASIDS) {
251 this_cpu_write(cpu_tlbstate.next_asid, 1);
257 * Given an ASID, flush the corresponding user ASID. We can delay this
258 * until the next time we switch to it.
260 * See SWITCH_TO_USER_CR3.
262 static inline void invalidate_user_asid(u16 asid)
264 /* There is no user ASID if address space separation is off */
265 if (!IS_ENABLED(CONFIG_PAGE_TABLE_ISOLATION))
269 * We only have a single ASID if PCID is off and the CR3
270 * write will have flushed it.
272 if (!cpu_feature_enabled(X86_FEATURE_PCID))
275 if (!static_cpu_has(X86_FEATURE_PTI))
278 __set_bit(kern_pcid(asid),
279 (unsigned long *)this_cpu_ptr(&cpu_tlbstate.user_pcid_flush_mask));
282 static void load_new_mm_cr3(pgd_t *pgdir, u16 new_asid, unsigned long lam,
285 unsigned long new_mm_cr3;
288 invalidate_user_asid(new_asid);
289 new_mm_cr3 = build_cr3(pgdir, new_asid, lam);
291 new_mm_cr3 = build_cr3_noflush(pgdir, new_asid, lam);
295 * Caution: many callers of this function expect
296 * that load_cr3() is serializing and orders TLB
297 * fills with respect to the mm_cpumask writes.
299 write_cr3(new_mm_cr3);
302 void leave_mm(int cpu)
304 struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
307 * It's plausible that we're in lazy TLB mode while our mm is init_mm.
308 * If so, our callers still expect us to flush the TLB, but there
309 * aren't any user TLB entries in init_mm to worry about.
311 * This needs to happen before any other sanity checks due to
312 * intel_idle's shenanigans.
314 if (loaded_mm == &init_mm)
317 /* Warn if we're not lazy. */
318 WARN_ON(!this_cpu_read(cpu_tlbstate_shared.is_lazy));
320 switch_mm(NULL, &init_mm, NULL);
322 EXPORT_SYMBOL_GPL(leave_mm);
324 void switch_mm(struct mm_struct *prev, struct mm_struct *next,
325 struct task_struct *tsk)
329 local_irq_save(flags);
330 switch_mm_irqs_off(prev, next, tsk);
331 local_irq_restore(flags);
335 * Invoked from return to user/guest by a task that opted-in to L1D
336 * flushing but ended up running on an SMT enabled core due to wrong
337 * affinity settings or CPU hotplug. This is part of the paranoid L1D flush
338 * contract which this task requested.
340 static void l1d_flush_force_sigbus(struct callback_head *ch)
345 static void l1d_flush_evaluate(unsigned long prev_mm, unsigned long next_mm,
346 struct task_struct *next)
348 /* Flush L1D if the outgoing task requests it */
349 if (prev_mm & LAST_USER_MM_L1D_FLUSH)
350 wrmsrl(MSR_IA32_FLUSH_CMD, L1D_FLUSH);
352 /* Check whether the incoming task opted in for L1D flush */
353 if (likely(!(next_mm & LAST_USER_MM_L1D_FLUSH)))
357 * Validate that it is not running on an SMT sibling as this would
358 * make the excercise pointless because the siblings share L1D. If
359 * it runs on a SMT sibling, notify it with SIGBUS on return to
362 if (this_cpu_read(cpu_info.smt_active)) {
363 clear_ti_thread_flag(&next->thread_info, TIF_SPEC_L1D_FLUSH);
364 next->l1d_flush_kill.func = l1d_flush_force_sigbus;
365 task_work_add(next, &next->l1d_flush_kill, TWA_RESUME);
369 static unsigned long mm_mangle_tif_spec_bits(struct task_struct *next)
371 unsigned long next_tif = read_task_thread_flags(next);
372 unsigned long spec_bits = (next_tif >> TIF_SPEC_IB) & LAST_USER_MM_SPEC_MASK;
375 * Ensure that the bit shift above works as expected and the two flags
376 * end up in bit 0 and 1.
378 BUILD_BUG_ON(TIF_SPEC_L1D_FLUSH != TIF_SPEC_IB + 1);
380 return (unsigned long)next->mm | spec_bits;
383 static void cond_mitigation(struct task_struct *next)
385 unsigned long prev_mm, next_mm;
387 if (!next || !next->mm)
390 next_mm = mm_mangle_tif_spec_bits(next);
391 prev_mm = this_cpu_read(cpu_tlbstate.last_user_mm_spec);
394 * Avoid user/user BTB poisoning by flushing the branch predictor
395 * when switching between processes. This stops one process from
396 * doing Spectre-v2 attacks on another.
398 * Both, the conditional and the always IBPB mode use the mm
399 * pointer to avoid the IBPB when switching between tasks of the
400 * same process. Using the mm pointer instead of mm->context.ctx_id
401 * opens a hypothetical hole vs. mm_struct reuse, which is more or
402 * less impossible to control by an attacker. Aside of that it
403 * would only affect the first schedule so the theoretically
404 * exposed data is not really interesting.
406 if (static_branch_likely(&switch_mm_cond_ibpb)) {
408 * This is a bit more complex than the always mode because
409 * it has to handle two cases:
411 * 1) Switch from a user space task (potential attacker)
412 * which has TIF_SPEC_IB set to a user space task
413 * (potential victim) which has TIF_SPEC_IB not set.
415 * 2) Switch from a user space task (potential attacker)
416 * which has TIF_SPEC_IB not set to a user space task
417 * (potential victim) which has TIF_SPEC_IB set.
419 * This could be done by unconditionally issuing IBPB when
420 * a task which has TIF_SPEC_IB set is either scheduled in
421 * or out. Though that results in two flushes when:
423 * - the same user space task is scheduled out and later
424 * scheduled in again and only a kernel thread ran in
427 * - a user space task belonging to the same process is
428 * scheduled in after a kernel thread ran in between
430 * - a user space task belonging to the same process is
431 * scheduled in immediately.
433 * Optimize this with reasonably small overhead for the
434 * above cases. Mangle the TIF_SPEC_IB bit into the mm
435 * pointer of the incoming task which is stored in
436 * cpu_tlbstate.last_user_mm_spec for comparison.
438 * Issue IBPB only if the mm's are different and one or
439 * both have the IBPB bit set.
441 if (next_mm != prev_mm &&
442 (next_mm | prev_mm) & LAST_USER_MM_IBPB)
443 indirect_branch_prediction_barrier();
446 if (static_branch_unlikely(&switch_mm_always_ibpb)) {
448 * Only flush when switching to a user space task with a
449 * different context than the user space task which ran
452 if ((prev_mm & ~LAST_USER_MM_SPEC_MASK) !=
453 (unsigned long)next->mm)
454 indirect_branch_prediction_barrier();
457 if (static_branch_unlikely(&switch_mm_cond_l1d_flush)) {
459 * Flush L1D when the outgoing task requested it and/or
460 * check whether the incoming task requested L1D flushing
461 * and ended up on an SMT sibling.
463 if (unlikely((prev_mm | next_mm) & LAST_USER_MM_L1D_FLUSH))
464 l1d_flush_evaluate(prev_mm, next_mm, next);
467 this_cpu_write(cpu_tlbstate.last_user_mm_spec, next_mm);
470 #ifdef CONFIG_PERF_EVENTS
471 static inline void cr4_update_pce_mm(struct mm_struct *mm)
473 if (static_branch_unlikely(&rdpmc_always_available_key) ||
474 (!static_branch_unlikely(&rdpmc_never_available_key) &&
475 atomic_read(&mm->context.perf_rdpmc_allowed))) {
477 * Clear the existing dirty counters to
478 * prevent the leak for an RDPMC task.
480 perf_clear_dirty_counters();
481 cr4_set_bits_irqsoff(X86_CR4_PCE);
483 cr4_clear_bits_irqsoff(X86_CR4_PCE);
486 void cr4_update_pce(void *ignored)
488 cr4_update_pce_mm(this_cpu_read(cpu_tlbstate.loaded_mm));
492 static inline void cr4_update_pce_mm(struct mm_struct *mm) { }
495 void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
496 struct task_struct *tsk)
498 struct mm_struct *real_prev = this_cpu_read(cpu_tlbstate.loaded_mm);
499 u16 prev_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
500 unsigned long new_lam = mm_lam_cr3_mask(next);
501 bool was_lazy = this_cpu_read(cpu_tlbstate_shared.is_lazy);
502 unsigned cpu = smp_processor_id();
508 * NB: The scheduler will call us with prev == next when switching
509 * from lazy TLB mode to normal mode if active_mm isn't changing.
510 * When this happens, we don't assume that CR3 (and hence
511 * cpu_tlbstate.loaded_mm) matches next.
513 * NB: leave_mm() calls us with prev == NULL and tsk == NULL.
516 /* We don't want flush_tlb_func() to run concurrently with us. */
517 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
518 WARN_ON_ONCE(!irqs_disabled());
521 * Verify that CR3 is what we think it is. This will catch
522 * hypothetical buggy code that directly switches to swapper_pg_dir
523 * without going through leave_mm() / switch_mm_irqs_off() or that
524 * does something like write_cr3(read_cr3_pa()).
526 * Only do this check if CONFIG_DEBUG_VM=y because __read_cr3()
529 #ifdef CONFIG_DEBUG_VM
530 if (WARN_ON_ONCE(__read_cr3() != build_cr3(real_prev->pgd, prev_asid,
531 tlbstate_lam_cr3_mask()))) {
533 * If we were to BUG here, we'd be very likely to kill
534 * the system so hard that we don't see the call trace.
535 * Try to recover instead by ignoring the error and doing
536 * a global flush to minimize the chance of corruption.
538 * (This is far from being a fully correct recovery.
539 * Architecturally, the CPU could prefetch something
540 * back into an incorrect ASID slot and leave it there
541 * to cause trouble down the road. It's better than
548 this_cpu_write(cpu_tlbstate_shared.is_lazy, false);
551 * The membarrier system call requires a full memory barrier and
552 * core serialization before returning to user-space, after
553 * storing to rq->curr, when changing mm. This is because
554 * membarrier() sends IPIs to all CPUs that are in the target mm
555 * to make them issue memory barriers. However, if another CPU
556 * switches to/from the target mm concurrently with
557 * membarrier(), it can cause that CPU not to receive an IPI
558 * when it really should issue a memory barrier. Writing to CR3
559 * provides that full memory barrier and core serializing
562 if (real_prev == next) {
563 /* Not actually switching mm's */
564 VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[prev_asid].ctx_id) !=
565 next->context.ctx_id);
568 * If this races with another thread that enables lam, 'new_lam'
569 * might not match tlbstate_lam_cr3_mask().
573 * Even in lazy TLB mode, the CPU should stay set in the
574 * mm_cpumask. The TLB shootdown code can figure out from
575 * cpu_tlbstate_shared.is_lazy whether or not to send an IPI.
577 if (WARN_ON_ONCE(real_prev != &init_mm &&
578 !cpumask_test_cpu(cpu, mm_cpumask(next))))
579 cpumask_set_cpu(cpu, mm_cpumask(next));
582 * If the CPU is not in lazy TLB mode, we are just switching
583 * from one thread in a process to another thread in the same
584 * process. No TLB flush required.
590 * Read the tlb_gen to check whether a flush is needed.
591 * If the TLB is up to date, just use it.
592 * The barrier synchronizes with the tlb_gen increment in
593 * the TLB shootdown code.
596 next_tlb_gen = atomic64_read(&next->context.tlb_gen);
597 if (this_cpu_read(cpu_tlbstate.ctxs[prev_asid].tlb_gen) ==
602 * TLB contents went out of date while we were in lazy
603 * mode. Fall through to the TLB switching code below.
605 new_asid = prev_asid;
609 * Apply process to process speculation vulnerability
610 * mitigations if applicable.
612 cond_mitigation(tsk);
615 * Stop remote flushes for the previous mm.
616 * Skip kernel threads; we never send init_mm TLB flushing IPIs,
617 * but the bitmap manipulation can cause cache line contention.
619 if (real_prev != &init_mm) {
620 VM_WARN_ON_ONCE(!cpumask_test_cpu(cpu,
621 mm_cpumask(real_prev)));
622 cpumask_clear_cpu(cpu, mm_cpumask(real_prev));
626 * Start remote flushes and then read tlb_gen.
628 if (next != &init_mm)
629 cpumask_set_cpu(cpu, mm_cpumask(next));
630 next_tlb_gen = atomic64_read(&next->context.tlb_gen);
632 choose_new_asid(next, next_tlb_gen, &new_asid, &need_flush);
634 /* Let nmi_uaccess_okay() know that we're changing CR3. */
635 this_cpu_write(cpu_tlbstate.loaded_mm, LOADED_MM_SWITCHING);
639 set_tlbstate_lam_mode(next);
641 this_cpu_write(cpu_tlbstate.ctxs[new_asid].ctx_id, next->context.ctx_id);
642 this_cpu_write(cpu_tlbstate.ctxs[new_asid].tlb_gen, next_tlb_gen);
643 load_new_mm_cr3(next->pgd, new_asid, new_lam, true);
645 trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
647 /* The new ASID is already up to date. */
648 load_new_mm_cr3(next->pgd, new_asid, new_lam, false);
650 trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, 0);
653 /* Make sure we write CR3 before loaded_mm. */
656 this_cpu_write(cpu_tlbstate.loaded_mm, next);
657 this_cpu_write(cpu_tlbstate.loaded_mm_asid, new_asid);
659 if (next != real_prev) {
660 cr4_update_pce_mm(next);
661 switch_ldt(real_prev, next);
666 * Please ignore the name of this function. It should be called
667 * switch_to_kernel_thread().
669 * enter_lazy_tlb() is a hint from the scheduler that we are entering a
670 * kernel thread or other context without an mm. Acceptable implementations
671 * include doing nothing whatsoever, switching to init_mm, or various clever
672 * lazy tricks to try to minimize TLB flushes.
674 * The scheduler reserves the right to call enter_lazy_tlb() several times
675 * in a row. It will notify us that we're going back to a real mm by
676 * calling switch_mm_irqs_off().
678 void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
680 if (this_cpu_read(cpu_tlbstate.loaded_mm) == &init_mm)
683 this_cpu_write(cpu_tlbstate_shared.is_lazy, true);
687 * Call this when reinitializing a CPU. It fixes the following potential
690 * - The ASID changed from what cpu_tlbstate thinks it is (most likely
691 * because the CPU was taken down and came back up with CR3's PCID
692 * bits clear. CPU hotplug can do this.
694 * - The TLB contains junk in slots corresponding to inactive ASIDs.
696 * - The CPU went so far out to lunch that it may have missed a TLB
699 void initialize_tlbstate_and_flush(void)
702 struct mm_struct *mm = this_cpu_read(cpu_tlbstate.loaded_mm);
703 u64 tlb_gen = atomic64_read(&init_mm.context.tlb_gen);
704 unsigned long cr3 = __read_cr3();
706 /* Assert that CR3 already references the right mm. */
707 WARN_ON((cr3 & CR3_ADDR_MASK) != __pa(mm->pgd));
709 /* LAM expected to be disabled */
710 WARN_ON(cr3 & (X86_CR3_LAM_U48 | X86_CR3_LAM_U57));
711 WARN_ON(mm_lam_cr3_mask(mm));
714 * Assert that CR4.PCIDE is set if needed. (CR4.PCIDE initialization
715 * doesn't work like other CR4 bits because it can only be set from
718 WARN_ON(boot_cpu_has(X86_FEATURE_PCID) &&
719 !(cr4_read_shadow() & X86_CR4_PCIDE));
721 /* Disable LAM, force ASID 0 and force a TLB flush. */
722 write_cr3(build_cr3(mm->pgd, 0, 0));
724 /* Reinitialize tlbstate. */
725 this_cpu_write(cpu_tlbstate.last_user_mm_spec, LAST_USER_MM_INIT);
726 this_cpu_write(cpu_tlbstate.loaded_mm_asid, 0);
727 this_cpu_write(cpu_tlbstate.next_asid, 1);
728 this_cpu_write(cpu_tlbstate.ctxs[0].ctx_id, mm->context.ctx_id);
729 this_cpu_write(cpu_tlbstate.ctxs[0].tlb_gen, tlb_gen);
730 set_tlbstate_lam_mode(mm);
732 for (i = 1; i < TLB_NR_DYN_ASIDS; i++)
733 this_cpu_write(cpu_tlbstate.ctxs[i].ctx_id, 0);
737 * flush_tlb_func()'s memory ordering requirement is that any
738 * TLB fills that happen after we flush the TLB are ordered after we
739 * read active_mm's tlb_gen. We don't need any explicit barriers
740 * because all x86 flush operations are serializing and the
741 * atomic64_read operation won't be reordered by the compiler.
743 static void flush_tlb_func(void *info)
746 * We have three different tlb_gen values in here. They are:
748 * - mm_tlb_gen: the latest generation.
749 * - local_tlb_gen: the generation that this CPU has already caught
751 * - f->new_tlb_gen: the generation that the requester of the flush
752 * wants us to catch up to.
754 const struct flush_tlb_info *f = info;
755 struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
756 u32 loaded_mm_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
757 u64 local_tlb_gen = this_cpu_read(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen);
758 bool local = smp_processor_id() == f->initiating_cpu;
759 unsigned long nr_invalidate = 0;
762 /* This code cannot presently handle being reentered. */
763 VM_WARN_ON(!irqs_disabled());
766 inc_irq_stat(irq_tlb_count);
767 count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
769 /* Can only happen on remote CPUs */
770 if (f->mm && f->mm != loaded_mm)
774 if (unlikely(loaded_mm == &init_mm))
777 VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[loaded_mm_asid].ctx_id) !=
778 loaded_mm->context.ctx_id);
780 if (this_cpu_read(cpu_tlbstate_shared.is_lazy)) {
782 * We're in lazy mode. We need to at least flush our
783 * paging-structure cache to avoid speculatively reading
784 * garbage into our TLB. Since switching to init_mm is barely
785 * slower than a minimal flush, just switch to init_mm.
787 * This should be rare, with native_flush_tlb_multi() skipping
788 * IPIs to lazy TLB mode CPUs.
790 switch_mm_irqs_off(NULL, &init_mm, NULL);
794 if (unlikely(f->new_tlb_gen != TLB_GENERATION_INVALID &&
795 f->new_tlb_gen <= local_tlb_gen)) {
797 * The TLB is already up to date in respect to f->new_tlb_gen.
798 * While the core might be still behind mm_tlb_gen, checking
799 * mm_tlb_gen unnecessarily would have negative caching effects
806 * Defer mm_tlb_gen reading as long as possible to avoid cache
809 mm_tlb_gen = atomic64_read(&loaded_mm->context.tlb_gen);
811 if (unlikely(local_tlb_gen == mm_tlb_gen)) {
813 * There's nothing to do: we're already up to date. This can
814 * happen if two concurrent flushes happen -- the first flush to
815 * be handled can catch us all the way up, leaving no work for
821 WARN_ON_ONCE(local_tlb_gen > mm_tlb_gen);
822 WARN_ON_ONCE(f->new_tlb_gen > mm_tlb_gen);
825 * If we get to this point, we know that our TLB is out of date.
826 * This does not strictly imply that we need to flush (it's
827 * possible that f->new_tlb_gen <= local_tlb_gen), but we're
828 * going to need to flush in the very near future, so we might
829 * as well get it over with.
831 * The only question is whether to do a full or partial flush.
833 * We do a partial flush if requested and two extra conditions
836 * 1. f->new_tlb_gen == local_tlb_gen + 1. We have an invariant that
837 * we've always done all needed flushes to catch up to
838 * local_tlb_gen. If, for example, local_tlb_gen == 2 and
839 * f->new_tlb_gen == 3, then we know that the flush needed to bring
840 * us up to date for tlb_gen 3 is the partial flush we're
843 * As an example of why this check is needed, suppose that there
844 * are two concurrent flushes. The first is a full flush that
845 * changes context.tlb_gen from 1 to 2. The second is a partial
846 * flush that changes context.tlb_gen from 2 to 3. If they get
847 * processed on this CPU in reverse order, we'll see
848 * local_tlb_gen == 1, mm_tlb_gen == 3, and end != TLB_FLUSH_ALL.
849 * If we were to use __flush_tlb_one_user() and set local_tlb_gen to
850 * 3, we'd be break the invariant: we'd update local_tlb_gen above
851 * 1 without the full flush that's needed for tlb_gen 2.
853 * 2. f->new_tlb_gen == mm_tlb_gen. This is purely an optimization.
854 * Partial TLB flushes are not all that much cheaper than full TLB
855 * flushes, so it seems unlikely that it would be a performance win
856 * to do a partial flush if that won't bring our TLB fully up to
857 * date. By doing a full flush instead, we can increase
858 * local_tlb_gen all the way to mm_tlb_gen and we can probably
859 * avoid another flush in the very near future.
861 if (f->end != TLB_FLUSH_ALL &&
862 f->new_tlb_gen == local_tlb_gen + 1 &&
863 f->new_tlb_gen == mm_tlb_gen) {
865 unsigned long addr = f->start;
867 /* Partial flush cannot have invalid generations */
868 VM_WARN_ON(f->new_tlb_gen == TLB_GENERATION_INVALID);
870 /* Partial flush must have valid mm */
871 VM_WARN_ON(f->mm == NULL);
873 nr_invalidate = (f->end - f->start) >> f->stride_shift;
875 while (addr < f->end) {
876 flush_tlb_one_user(addr);
877 addr += 1UL << f->stride_shift;
880 count_vm_tlb_events(NR_TLB_LOCAL_FLUSH_ONE, nr_invalidate);
883 nr_invalidate = TLB_FLUSH_ALL;
887 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
890 /* Both paths above update our state to mm_tlb_gen. */
891 this_cpu_write(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen, mm_tlb_gen);
893 /* Tracing is done in a unified manner to reduce the code size */
895 trace_tlb_flush(!local ? TLB_REMOTE_SHOOTDOWN :
896 (f->mm == NULL) ? TLB_LOCAL_SHOOTDOWN :
897 TLB_LOCAL_MM_SHOOTDOWN,
901 static bool tlb_is_not_lazy(int cpu, void *data)
903 return !per_cpu(cpu_tlbstate_shared.is_lazy, cpu);
906 DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state_shared, cpu_tlbstate_shared);
907 EXPORT_PER_CPU_SYMBOL(cpu_tlbstate_shared);
909 STATIC_NOPV void native_flush_tlb_multi(const struct cpumask *cpumask,
910 const struct flush_tlb_info *info)
913 * Do accounting and tracing. Note that there are (and have always been)
914 * cases in which a remote TLB flush will be traced, but eventually
917 count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
918 if (info->end == TLB_FLUSH_ALL)
919 trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL);
921 trace_tlb_flush(TLB_REMOTE_SEND_IPI,
922 (info->end - info->start) >> PAGE_SHIFT);
925 * If no page tables were freed, we can skip sending IPIs to
926 * CPUs in lazy TLB mode. They will flush the CPU themselves
927 * at the next context switch.
929 * However, if page tables are getting freed, we need to send the
930 * IPI everywhere, to prevent CPUs in lazy TLB mode from tripping
931 * up on the new contents of what used to be page tables, while
932 * doing a speculative memory access.
934 if (info->freed_tables)
935 on_each_cpu_mask(cpumask, flush_tlb_func, (void *)info, true);
937 on_each_cpu_cond_mask(tlb_is_not_lazy, flush_tlb_func,
938 (void *)info, 1, cpumask);
941 void flush_tlb_multi(const struct cpumask *cpumask,
942 const struct flush_tlb_info *info)
944 __flush_tlb_multi(cpumask, info);
948 * See Documentation/arch/x86/tlb.rst for details. We choose 33
949 * because it is large enough to cover the vast majority (at
950 * least 95%) of allocations, and is small enough that we are
951 * confident it will not cause too much overhead. Each single
952 * flush is about 100 ns, so this caps the maximum overhead at
955 * This is in units of pages.
957 unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
959 static DEFINE_PER_CPU_SHARED_ALIGNED(struct flush_tlb_info, flush_tlb_info);
961 #ifdef CONFIG_DEBUG_VM
962 static DEFINE_PER_CPU(unsigned int, flush_tlb_info_idx);
965 static struct flush_tlb_info *get_flush_tlb_info(struct mm_struct *mm,
966 unsigned long start, unsigned long end,
967 unsigned int stride_shift, bool freed_tables,
970 struct flush_tlb_info *info = this_cpu_ptr(&flush_tlb_info);
972 #ifdef CONFIG_DEBUG_VM
974 * Ensure that the following code is non-reentrant and flush_tlb_info
975 * is not overwritten. This means no TLB flushing is initiated by
976 * interrupt handlers and machine-check exception handlers.
978 BUG_ON(this_cpu_inc_return(flush_tlb_info_idx) != 1);
984 info->stride_shift = stride_shift;
985 info->freed_tables = freed_tables;
986 info->new_tlb_gen = new_tlb_gen;
987 info->initiating_cpu = smp_processor_id();
992 static void put_flush_tlb_info(void)
994 #ifdef CONFIG_DEBUG_VM
995 /* Complete reentrancy prevention checks */
997 this_cpu_dec(flush_tlb_info_idx);
1001 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
1002 unsigned long end, unsigned int stride_shift,
1005 struct flush_tlb_info *info;
1011 /* Should we flush just the requested range? */
1012 if ((end == TLB_FLUSH_ALL) ||
1013 ((end - start) >> stride_shift) > tlb_single_page_flush_ceiling) {
1015 end = TLB_FLUSH_ALL;
1018 /* This is also a barrier that synchronizes with switch_mm(). */
1019 new_tlb_gen = inc_mm_tlb_gen(mm);
1021 info = get_flush_tlb_info(mm, start, end, stride_shift, freed_tables,
1025 * flush_tlb_multi() is not optimized for the common case in which only
1026 * a local TLB flush is needed. Optimize this use-case by calling
1027 * flush_tlb_func_local() directly in this case.
1029 if (cpumask_any_but(mm_cpumask(mm), cpu) < nr_cpu_ids) {
1030 flush_tlb_multi(mm_cpumask(mm), info);
1031 } else if (mm == this_cpu_read(cpu_tlbstate.loaded_mm)) {
1032 lockdep_assert_irqs_enabled();
1033 local_irq_disable();
1034 flush_tlb_func(info);
1038 put_flush_tlb_info();
1040 mmu_notifier_arch_invalidate_secondary_tlbs(mm, start, end);
1044 static void do_flush_tlb_all(void *info)
1046 count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
1050 void flush_tlb_all(void)
1052 count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
1053 on_each_cpu(do_flush_tlb_all, NULL, 1);
1056 static void do_kernel_range_flush(void *info)
1058 struct flush_tlb_info *f = info;
1061 /* flush range by one by one 'invlpg' */
1062 for (addr = f->start; addr < f->end; addr += PAGE_SIZE)
1063 flush_tlb_one_kernel(addr);
1066 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
1068 /* Balance as user space task's flush, a bit conservative */
1069 if (end == TLB_FLUSH_ALL ||
1070 (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) {
1071 on_each_cpu(do_flush_tlb_all, NULL, 1);
1073 struct flush_tlb_info *info;
1076 info = get_flush_tlb_info(NULL, start, end, 0, false,
1077 TLB_GENERATION_INVALID);
1079 on_each_cpu(do_kernel_range_flush, info, 1);
1081 put_flush_tlb_info();
1087 * This can be used from process context to figure out what the value of
1088 * CR3 is without needing to do a (slow) __read_cr3().
1090 * It's intended to be used for code like KVM that sneakily changes CR3
1091 * and needs to restore it. It needs to be used very carefully.
1093 unsigned long __get_current_cr3_fast(void)
1096 build_cr3(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd,
1097 this_cpu_read(cpu_tlbstate.loaded_mm_asid),
1098 tlbstate_lam_cr3_mask());
1100 /* For now, be very restrictive about when this can be called. */
1101 VM_WARN_ON(in_nmi() || preemptible());
1103 VM_BUG_ON(cr3 != __read_cr3());
1106 EXPORT_SYMBOL_GPL(__get_current_cr3_fast);
1109 * Flush one page in the kernel mapping
1111 void flush_tlb_one_kernel(unsigned long addr)
1113 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
1116 * If PTI is off, then __flush_tlb_one_user() is just INVLPG or its
1117 * paravirt equivalent. Even with PCID, this is sufficient: we only
1118 * use PCID if we also use global PTEs for the kernel mapping, and
1119 * INVLPG flushes global translations across all address spaces.
1121 * If PTI is on, then the kernel is mapped with non-global PTEs, and
1122 * __flush_tlb_one_user() will flush the given address for the current
1123 * kernel address space and for its usermode counterpart, but it does
1124 * not flush it for other address spaces.
1126 flush_tlb_one_user(addr);
1128 if (!static_cpu_has(X86_FEATURE_PTI))
1132 * See above. We need to propagate the flush to all other address
1133 * spaces. In principle, we only need to propagate it to kernelmode
1134 * address spaces, but the extra bookkeeping we would need is not
1137 this_cpu_write(cpu_tlbstate.invalidate_other, true);
1141 * Flush one page in the user mapping
1143 STATIC_NOPV void native_flush_tlb_one_user(unsigned long addr)
1148 /* Flush 'addr' from the kernel PCID: */
1149 asm volatile("invlpg (%0)" ::"r" (addr) : "memory");
1151 /* If PTI is off there is no user PCID and nothing to flush. */
1152 if (!static_cpu_has(X86_FEATURE_PTI))
1155 loaded_mm_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
1156 cpu_pcide = this_cpu_read(cpu_tlbstate.cr4) & X86_CR4_PCIDE;
1159 * invpcid_flush_one(pcid>0) will #GP if CR4.PCIDE==0. Check
1160 * 'cpu_pcide' to ensure that *this* CPU will not trigger those
1161 * #GP's even if called before CR4.PCIDE has been initialized.
1163 if (boot_cpu_has(X86_FEATURE_INVPCID) && cpu_pcide)
1164 invpcid_flush_one(user_pcid(loaded_mm_asid), addr);
1166 invalidate_user_asid(loaded_mm_asid);
1169 void flush_tlb_one_user(unsigned long addr)
1171 __flush_tlb_one_user(addr);
1177 STATIC_NOPV void native_flush_tlb_global(void)
1179 unsigned long flags;
1181 if (static_cpu_has(X86_FEATURE_INVPCID)) {
1183 * Using INVPCID is considerably faster than a pair of writes
1184 * to CR4 sandwiched inside an IRQ flag save/restore.
1186 * Note, this works with CR4.PCIDE=0 or 1.
1188 invpcid_flush_all();
1193 * Read-modify-write to CR4 - protect it from preemption and
1194 * from interrupts. (Use the raw variant because this code can
1195 * be called from deep inside debugging code.)
1197 raw_local_irq_save(flags);
1199 __native_tlb_flush_global(this_cpu_read(cpu_tlbstate.cr4));
1201 raw_local_irq_restore(flags);
1205 * Flush the entire current user mapping
1207 STATIC_NOPV void native_flush_tlb_local(void)
1210 * Preemption or interrupts must be disabled to protect the access
1211 * to the per CPU variable and to prevent being preempted between
1212 * read_cr3() and write_cr3().
1214 WARN_ON_ONCE(preemptible());
1216 invalidate_user_asid(this_cpu_read(cpu_tlbstate.loaded_mm_asid));
1218 /* If current->mm == NULL then the read_cr3() "borrows" an mm */
1219 native_write_cr3(__native_read_cr3());
1222 void flush_tlb_local(void)
1224 __flush_tlb_local();
1230 void __flush_tlb_all(void)
1233 * This is to catch users with enabled preemption and the PGE feature
1234 * and don't trigger the warning in __native_flush_tlb().
1236 VM_WARN_ON_ONCE(preemptible());
1238 if (cpu_feature_enabled(X86_FEATURE_PGE)) {
1239 __flush_tlb_global();
1242 * !PGE -> !PCID (setup_pcid()), thus every flush is total.
1247 EXPORT_SYMBOL_GPL(__flush_tlb_all);
1249 void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
1251 struct flush_tlb_info *info;
1253 int cpu = get_cpu();
1255 info = get_flush_tlb_info(NULL, 0, TLB_FLUSH_ALL, 0, false,
1256 TLB_GENERATION_INVALID);
1258 * flush_tlb_multi() is not optimized for the common case in which only
1259 * a local TLB flush is needed. Optimize this use-case by calling
1260 * flush_tlb_func_local() directly in this case.
1262 if (cpumask_any_but(&batch->cpumask, cpu) < nr_cpu_ids) {
1263 flush_tlb_multi(&batch->cpumask, info);
1264 } else if (cpumask_test_cpu(cpu, &batch->cpumask)) {
1265 lockdep_assert_irqs_enabled();
1266 local_irq_disable();
1267 flush_tlb_func(info);
1271 cpumask_clear(&batch->cpumask);
1273 put_flush_tlb_info();
1278 * Blindly accessing user memory from NMI context can be dangerous
1279 * if we're in the middle of switching the current user task or
1280 * switching the loaded mm. It can also be dangerous if we
1281 * interrupted some kernel code that was temporarily using a
1284 bool nmi_uaccess_okay(void)
1286 struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
1287 struct mm_struct *current_mm = current->mm;
1289 VM_WARN_ON_ONCE(!loaded_mm);
1292 * The condition we want to check is
1293 * current_mm->pgd == __va(read_cr3_pa()). This may be slow, though,
1294 * if we're running in a VM with shadow paging, and nmi_uaccess_okay()
1295 * is supposed to be reasonably fast.
1297 * Instead, we check the almost equivalent but somewhat conservative
1298 * condition below, and we rely on the fact that switch_mm_irqs_off()
1299 * sets loaded_mm to LOADED_MM_SWITCHING before writing to CR3.
1301 if (loaded_mm != current_mm)
1304 VM_WARN_ON_ONCE(current_mm->pgd != __va(read_cr3_pa()));
1309 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
1310 size_t count, loff_t *ppos)
1315 len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
1316 return simple_read_from_buffer(user_buf, count, ppos, buf, len);
1319 static ssize_t tlbflush_write_file(struct file *file,
1320 const char __user *user_buf, size_t count, loff_t *ppos)
1326 len = min(count, sizeof(buf) - 1);
1327 if (copy_from_user(buf, user_buf, len))
1331 if (kstrtoint(buf, 0, &ceiling))
1337 tlb_single_page_flush_ceiling = ceiling;
1341 static const struct file_operations fops_tlbflush = {
1342 .read = tlbflush_read_file,
1343 .write = tlbflush_write_file,
1344 .llseek = default_llseek,
1347 static int __init create_tlb_single_page_flush_ceiling(void)
1349 debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
1350 arch_debugfs_dir, NULL, &fops_tlbflush);
1353 late_initcall(create_tlb_single_page_flush_ceiling);