Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2#include <linux/init.h>
   3
   4#include <linux/mm.h>
   5#include <linux/spinlock.h>
   6#include <linux/smp.h>
   7#include <linux/interrupt.h>
   8#include <linux/export.h>
   9#include <linux/cpu.h>
  10#include <linux/debugfs.h>
  11#include <linux/sched/smt.h>
  12#include <linux/task_work.h>
 
 
  13
  14#include <asm/tlbflush.h>
  15#include <asm/mmu_context.h>
  16#include <asm/nospec-branch.h>
  17#include <asm/cache.h>
  18#include <asm/cacheflush.h>
  19#include <asm/apic.h>
  20#include <asm/perf_event.h>
 
  21
  22#include "mm_internal.h"
  23
  24#ifdef CONFIG_PARAVIRT
  25# define STATIC_NOPV
  26#else
  27# define STATIC_NOPV			static
  28# define __flush_tlb_local		native_flush_tlb_local
  29# define __flush_tlb_global		native_flush_tlb_global
  30# define __flush_tlb_one_user(addr)	native_flush_tlb_one_user(addr)
  31# define __flush_tlb_multi(msk, info)	native_flush_tlb_multi(msk, info)
  32#endif
  33
  34/*
  35 *	TLB flushing, formerly SMP-only
  36 *		c/o Linus Torvalds.
  37 *
  38 *	These mean you can really definitely utterly forget about
  39 *	writing to user space from interrupts. (Its not allowed anyway).
  40 *
  41 *	Optimizations Manfred Spraul <manfred@colorfullife.com>
  42 *
  43 *	More scalable flush, from Andi Kleen
  44 *
  45 *	Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi
  46 */
  47
  48/*
  49 * Bits to mangle the TIF_SPEC_* state into the mm pointer which is
  50 * stored in cpu_tlb_state.last_user_mm_spec.
  51 */
  52#define LAST_USER_MM_IBPB	0x1UL
  53#define LAST_USER_MM_L1D_FLUSH	0x2UL
  54#define LAST_USER_MM_SPEC_MASK	(LAST_USER_MM_IBPB | LAST_USER_MM_L1D_FLUSH)
  55
  56/* Bits to set when tlbstate and flush is (re)initialized */
  57#define LAST_USER_MM_INIT	LAST_USER_MM_IBPB
  58
  59/*
  60 * The x86 feature is called PCID (Process Context IDentifier). It is similar
  61 * to what is traditionally called ASID on the RISC processors.
  62 *
  63 * We don't use the traditional ASID implementation, where each process/mm gets
  64 * its own ASID and flush/restart when we run out of ASID space.
  65 *
  66 * Instead we have a small per-cpu array of ASIDs and cache the last few mm's
  67 * that came by on this CPU, allowing cheaper switch_mm between processes on
  68 * this CPU.
  69 *
  70 * We end up with different spaces for different things. To avoid confusion we
  71 * use different names for each of them:
  72 *
  73 * ASID  - [0, TLB_NR_DYN_ASIDS-1]
  74 *         the canonical identifier for an mm
  75 *
  76 * kPCID - [1, TLB_NR_DYN_ASIDS]
  77 *         the value we write into the PCID part of CR3; corresponds to the
  78 *         ASID+1, because PCID 0 is special.
  79 *
  80 * uPCID - [2048 + 1, 2048 + TLB_NR_DYN_ASIDS]
  81 *         for KPTI each mm has two address spaces and thus needs two
  82 *         PCID values, but we can still do with a single ASID denomination
  83 *         for each mm. Corresponds to kPCID + 2048.
  84 *
  85 */
  86
  87/* There are 12 bits of space for ASIDS in CR3 */
  88#define CR3_HW_ASID_BITS		12
  89
  90/*
  91 * When enabled, PAGE_TABLE_ISOLATION consumes a single bit for
  92 * user/kernel switches
  93 */
  94#ifdef CONFIG_PAGE_TABLE_ISOLATION
  95# define PTI_CONSUMED_PCID_BITS	1
  96#else
  97# define PTI_CONSUMED_PCID_BITS	0
  98#endif
  99
 100#define CR3_AVAIL_PCID_BITS (X86_CR3_PCID_BITS - PTI_CONSUMED_PCID_BITS)
 101
 102/*
 103 * ASIDs are zero-based: 0->MAX_AVAIL_ASID are valid.  -1 below to account
 104 * for them being zero-based.  Another -1 is because PCID 0 is reserved for
 105 * use by non-PCID-aware users.
 106 */
 107#define MAX_ASID_AVAILABLE ((1 << CR3_AVAIL_PCID_BITS) - 2)
 108
 109/*
 110 * Given @asid, compute kPCID
 111 */
 112static inline u16 kern_pcid(u16 asid)
 113{
 114	VM_WARN_ON_ONCE(asid > MAX_ASID_AVAILABLE);
 115
 116#ifdef CONFIG_PAGE_TABLE_ISOLATION
 117	/*
 118	 * Make sure that the dynamic ASID space does not conflict with the
 119	 * bit we are using to switch between user and kernel ASIDs.
 120	 */
 121	BUILD_BUG_ON(TLB_NR_DYN_ASIDS >= (1 << X86_CR3_PTI_PCID_USER_BIT));
 122
 123	/*
 124	 * The ASID being passed in here should have respected the
 125	 * MAX_ASID_AVAILABLE and thus never have the switch bit set.
 126	 */
 127	VM_WARN_ON_ONCE(asid & (1 << X86_CR3_PTI_PCID_USER_BIT));
 128#endif
 129	/*
 130	 * The dynamically-assigned ASIDs that get passed in are small
 131	 * (<TLB_NR_DYN_ASIDS).  They never have the high switch bit set,
 132	 * so do not bother to clear it.
 133	 *
 134	 * If PCID is on, ASID-aware code paths put the ASID+1 into the
 135	 * PCID bits.  This serves two purposes.  It prevents a nasty
 136	 * situation in which PCID-unaware code saves CR3, loads some other
 137	 * value (with PCID == 0), and then restores CR3, thus corrupting
 138	 * the TLB for ASID 0 if the saved ASID was nonzero.  It also means
 139	 * that any bugs involving loading a PCID-enabled CR3 with
 140	 * CR4.PCIDE off will trigger deterministically.
 141	 */
 142	return asid + 1;
 143}
 144
 145/*
 146 * Given @asid, compute uPCID
 147 */
 148static inline u16 user_pcid(u16 asid)
 149{
 150	u16 ret = kern_pcid(asid);
 151#ifdef CONFIG_PAGE_TABLE_ISOLATION
 152	ret |= 1 << X86_CR3_PTI_PCID_USER_BIT;
 153#endif
 154	return ret;
 155}
 156
 157static inline unsigned long build_cr3(pgd_t *pgd, u16 asid)
 158{
 
 
 159	if (static_cpu_has(X86_FEATURE_PCID)) {
 160		return __sme_pa(pgd) | kern_pcid(asid);
 161	} else {
 162		VM_WARN_ON_ONCE(asid != 0);
 163		return __sme_pa(pgd);
 164	}
 
 
 165}
 166
 167static inline unsigned long build_cr3_noflush(pgd_t *pgd, u16 asid)
 
 168{
 169	VM_WARN_ON_ONCE(asid > MAX_ASID_AVAILABLE);
 170	/*
 171	 * Use boot_cpu_has() instead of this_cpu_has() as this function
 172	 * might be called during early boot. This should work even after
 173	 * boot because all CPU's the have same capabilities:
 174	 */
 175	VM_WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_PCID));
 176	return __sme_pa(pgd) | kern_pcid(asid) | CR3_NOFLUSH;
 177}
 178
 179/*
 180 * We get here when we do something requiring a TLB invalidation
 181 * but could not go invalidate all of the contexts.  We do the
 182 * necessary invalidation by clearing out the 'ctx_id' which
 183 * forces a TLB flush when the context is loaded.
 184 */
 185static void clear_asid_other(void)
 186{
 187	u16 asid;
 188
 189	/*
 190	 * This is only expected to be set if we have disabled
 191	 * kernel _PAGE_GLOBAL pages.
 192	 */
 193	if (!static_cpu_has(X86_FEATURE_PTI)) {
 194		WARN_ON_ONCE(1);
 195		return;
 196	}
 197
 198	for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
 199		/* Do not need to flush the current asid */
 200		if (asid == this_cpu_read(cpu_tlbstate.loaded_mm_asid))
 201			continue;
 202		/*
 203		 * Make sure the next time we go to switch to
 204		 * this asid, we do a flush:
 205		 */
 206		this_cpu_write(cpu_tlbstate.ctxs[asid].ctx_id, 0);
 207	}
 208	this_cpu_write(cpu_tlbstate.invalidate_other, false);
 209}
 210
 211atomic64_t last_mm_ctx_id = ATOMIC64_INIT(1);
 212
 213
 214static void choose_new_asid(struct mm_struct *next, u64 next_tlb_gen,
 215			    u16 *new_asid, bool *need_flush)
 216{
 217	u16 asid;
 218
 219	if (!static_cpu_has(X86_FEATURE_PCID)) {
 220		*new_asid = 0;
 221		*need_flush = true;
 222		return;
 223	}
 224
 225	if (this_cpu_read(cpu_tlbstate.invalidate_other))
 226		clear_asid_other();
 227
 228	for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
 229		if (this_cpu_read(cpu_tlbstate.ctxs[asid].ctx_id) !=
 230		    next->context.ctx_id)
 231			continue;
 232
 233		*new_asid = asid;
 234		*need_flush = (this_cpu_read(cpu_tlbstate.ctxs[asid].tlb_gen) <
 235			       next_tlb_gen);
 236		return;
 237	}
 238
 239	/*
 240	 * We don't currently own an ASID slot on this CPU.
 241	 * Allocate a slot.
 242	 */
 243	*new_asid = this_cpu_add_return(cpu_tlbstate.next_asid, 1) - 1;
 244	if (*new_asid >= TLB_NR_DYN_ASIDS) {
 245		*new_asid = 0;
 246		this_cpu_write(cpu_tlbstate.next_asid, 1);
 247	}
 248	*need_flush = true;
 249}
 250
 251/*
 252 * Given an ASID, flush the corresponding user ASID.  We can delay this
 253 * until the next time we switch to it.
 254 *
 255 * See SWITCH_TO_USER_CR3.
 256 */
 257static inline void invalidate_user_asid(u16 asid)
 258{
 259	/* There is no user ASID if address space separation is off */
 260	if (!IS_ENABLED(CONFIG_PAGE_TABLE_ISOLATION))
 261		return;
 262
 263	/*
 264	 * We only have a single ASID if PCID is off and the CR3
 265	 * write will have flushed it.
 266	 */
 267	if (!cpu_feature_enabled(X86_FEATURE_PCID))
 268		return;
 269
 270	if (!static_cpu_has(X86_FEATURE_PTI))
 271		return;
 272
 273	__set_bit(kern_pcid(asid),
 274		  (unsigned long *)this_cpu_ptr(&cpu_tlbstate.user_pcid_flush_mask));
 275}
 276
 277static void load_new_mm_cr3(pgd_t *pgdir, u16 new_asid, bool need_flush)
 
 278{
 279	unsigned long new_mm_cr3;
 280
 281	if (need_flush) {
 282		invalidate_user_asid(new_asid);
 283		new_mm_cr3 = build_cr3(pgdir, new_asid);
 284	} else {
 285		new_mm_cr3 = build_cr3_noflush(pgdir, new_asid);
 286	}
 287
 288	/*
 289	 * Caution: many callers of this function expect
 290	 * that load_cr3() is serializing and orders TLB
 291	 * fills with respect to the mm_cpumask writes.
 292	 */
 293	write_cr3(new_mm_cr3);
 294}
 295
 296void leave_mm(int cpu)
 297{
 298	struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
 299
 300	/*
 301	 * It's plausible that we're in lazy TLB mode while our mm is init_mm.
 302	 * If so, our callers still expect us to flush the TLB, but there
 303	 * aren't any user TLB entries in init_mm to worry about.
 304	 *
 305	 * This needs to happen before any other sanity checks due to
 306	 * intel_idle's shenanigans.
 307	 */
 308	if (loaded_mm == &init_mm)
 309		return;
 310
 311	/* Warn if we're not lazy. */
 312	WARN_ON(!this_cpu_read(cpu_tlbstate_shared.is_lazy));
 313
 314	switch_mm(NULL, &init_mm, NULL);
 315}
 316EXPORT_SYMBOL_GPL(leave_mm);
 317
 318void switch_mm(struct mm_struct *prev, struct mm_struct *next,
 319	       struct task_struct *tsk)
 320{
 321	unsigned long flags;
 322
 323	local_irq_save(flags);
 324	switch_mm_irqs_off(prev, next, tsk);
 325	local_irq_restore(flags);
 326}
 327
 328/*
 329 * Invoked from return to user/guest by a task that opted-in to L1D
 330 * flushing but ended up running on an SMT enabled core due to wrong
 331 * affinity settings or CPU hotplug. This is part of the paranoid L1D flush
 332 * contract which this task requested.
 333 */
 334static void l1d_flush_force_sigbus(struct callback_head *ch)
 335{
 336	force_sig(SIGBUS);
 337}
 338
 339static void l1d_flush_evaluate(unsigned long prev_mm, unsigned long next_mm,
 340				struct task_struct *next)
 341{
 342	/* Flush L1D if the outgoing task requests it */
 343	if (prev_mm & LAST_USER_MM_L1D_FLUSH)
 344		wrmsrl(MSR_IA32_FLUSH_CMD, L1D_FLUSH);
 345
 346	/* Check whether the incoming task opted in for L1D flush */
 347	if (likely(!(next_mm & LAST_USER_MM_L1D_FLUSH)))
 348		return;
 349
 350	/*
 351	 * Validate that it is not running on an SMT sibling as this would
 352	 * make the excercise pointless because the siblings share L1D. If
 353	 * it runs on a SMT sibling, notify it with SIGBUS on return to
 354	 * user/guest
 355	 */
 356	if (this_cpu_read(cpu_info.smt_active)) {
 357		clear_ti_thread_flag(&next->thread_info, TIF_SPEC_L1D_FLUSH);
 358		next->l1d_flush_kill.func = l1d_flush_force_sigbus;
 359		task_work_add(next, &next->l1d_flush_kill, TWA_RESUME);
 360	}
 361}
 362
 363static unsigned long mm_mangle_tif_spec_bits(struct task_struct *next)
 364{
 365	unsigned long next_tif = read_task_thread_flags(next);
 366	unsigned long spec_bits = (next_tif >> TIF_SPEC_IB) & LAST_USER_MM_SPEC_MASK;
 367
 368	/*
 369	 * Ensure that the bit shift above works as expected and the two flags
 370	 * end up in bit 0 and 1.
 371	 */
 372	BUILD_BUG_ON(TIF_SPEC_L1D_FLUSH != TIF_SPEC_IB + 1);
 373
 374	return (unsigned long)next->mm | spec_bits;
 375}
 376
 377static void cond_mitigation(struct task_struct *next)
 378{
 379	unsigned long prev_mm, next_mm;
 380
 381	if (!next || !next->mm)
 382		return;
 383
 384	next_mm = mm_mangle_tif_spec_bits(next);
 385	prev_mm = this_cpu_read(cpu_tlbstate.last_user_mm_spec);
 386
 387	/*
 388	 * Avoid user/user BTB poisoning by flushing the branch predictor
 389	 * when switching between processes. This stops one process from
 390	 * doing Spectre-v2 attacks on another.
 391	 *
 392	 * Both, the conditional and the always IBPB mode use the mm
 393	 * pointer to avoid the IBPB when switching between tasks of the
 394	 * same process. Using the mm pointer instead of mm->context.ctx_id
 395	 * opens a hypothetical hole vs. mm_struct reuse, which is more or
 396	 * less impossible to control by an attacker. Aside of that it
 397	 * would only affect the first schedule so the theoretically
 398	 * exposed data is not really interesting.
 399	 */
 400	if (static_branch_likely(&switch_mm_cond_ibpb)) {
 401		/*
 402		 * This is a bit more complex than the always mode because
 403		 * it has to handle two cases:
 404		 *
 405		 * 1) Switch from a user space task (potential attacker)
 406		 *    which has TIF_SPEC_IB set to a user space task
 407		 *    (potential victim) which has TIF_SPEC_IB not set.
 408		 *
 409		 * 2) Switch from a user space task (potential attacker)
 410		 *    which has TIF_SPEC_IB not set to a user space task
 411		 *    (potential victim) which has TIF_SPEC_IB set.
 412		 *
 413		 * This could be done by unconditionally issuing IBPB when
 414		 * a task which has TIF_SPEC_IB set is either scheduled in
 415		 * or out. Though that results in two flushes when:
 416		 *
 417		 * - the same user space task is scheduled out and later
 418		 *   scheduled in again and only a kernel thread ran in
 419		 *   between.
 420		 *
 421		 * - a user space task belonging to the same process is
 422		 *   scheduled in after a kernel thread ran in between
 423		 *
 424		 * - a user space task belonging to the same process is
 425		 *   scheduled in immediately.
 426		 *
 427		 * Optimize this with reasonably small overhead for the
 428		 * above cases. Mangle the TIF_SPEC_IB bit into the mm
 429		 * pointer of the incoming task which is stored in
 430		 * cpu_tlbstate.last_user_mm_spec for comparison.
 431		 *
 432		 * Issue IBPB only if the mm's are different and one or
 433		 * both have the IBPB bit set.
 434		 */
 435		if (next_mm != prev_mm &&
 436		    (next_mm | prev_mm) & LAST_USER_MM_IBPB)
 437			indirect_branch_prediction_barrier();
 438	}
 439
 440	if (static_branch_unlikely(&switch_mm_always_ibpb)) {
 441		/*
 442		 * Only flush when switching to a user space task with a
 443		 * different context than the user space task which ran
 444		 * last on this CPU.
 445		 */
 446		if ((prev_mm & ~LAST_USER_MM_SPEC_MASK) !=
 447					(unsigned long)next->mm)
 448			indirect_branch_prediction_barrier();
 449	}
 450
 451	if (static_branch_unlikely(&switch_mm_cond_l1d_flush)) {
 452		/*
 453		 * Flush L1D when the outgoing task requested it and/or
 454		 * check whether the incoming task requested L1D flushing
 455		 * and ended up on an SMT sibling.
 456		 */
 457		if (unlikely((prev_mm | next_mm) & LAST_USER_MM_L1D_FLUSH))
 458			l1d_flush_evaluate(prev_mm, next_mm, next);
 459	}
 460
 461	this_cpu_write(cpu_tlbstate.last_user_mm_spec, next_mm);
 462}
 463
 464#ifdef CONFIG_PERF_EVENTS
 465static inline void cr4_update_pce_mm(struct mm_struct *mm)
 466{
 467	if (static_branch_unlikely(&rdpmc_always_available_key) ||
 468	    (!static_branch_unlikely(&rdpmc_never_available_key) &&
 469	     atomic_read(&mm->context.perf_rdpmc_allowed))) {
 470		/*
 471		 * Clear the existing dirty counters to
 472		 * prevent the leak for an RDPMC task.
 473		 */
 474		perf_clear_dirty_counters();
 475		cr4_set_bits_irqsoff(X86_CR4_PCE);
 476	} else
 477		cr4_clear_bits_irqsoff(X86_CR4_PCE);
 478}
 479
 480void cr4_update_pce(void *ignored)
 481{
 482	cr4_update_pce_mm(this_cpu_read(cpu_tlbstate.loaded_mm));
 483}
 484
 485#else
 486static inline void cr4_update_pce_mm(struct mm_struct *mm) { }
 487#endif
 488
 489void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
 
 
 
 
 
 
 490			struct task_struct *tsk)
 491{
 492	struct mm_struct *real_prev = this_cpu_read(cpu_tlbstate.loaded_mm);
 493	u16 prev_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
 494	bool was_lazy = this_cpu_read(cpu_tlbstate_shared.is_lazy);
 495	unsigned cpu = smp_processor_id();
 
 496	u64 next_tlb_gen;
 497	bool need_flush;
 498	u16 new_asid;
 499
 500	/*
 501	 * NB: The scheduler will call us with prev == next when switching
 502	 * from lazy TLB mode to normal mode if active_mm isn't changing.
 503	 * When this happens, we don't assume that CR3 (and hence
 504	 * cpu_tlbstate.loaded_mm) matches next.
 505	 *
 506	 * NB: leave_mm() calls us with prev == NULL and tsk == NULL.
 507	 */
 508
 509	/* We don't want flush_tlb_func() to run concurrently with us. */
 510	if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 511		WARN_ON_ONCE(!irqs_disabled());
 512
 513	/*
 514	 * Verify that CR3 is what we think it is.  This will catch
 515	 * hypothetical buggy code that directly switches to swapper_pg_dir
 516	 * without going through leave_mm() / switch_mm_irqs_off() or that
 517	 * does something like write_cr3(read_cr3_pa()).
 518	 *
 519	 * Only do this check if CONFIG_DEBUG_VM=y because __read_cr3()
 520	 * isn't free.
 521	 */
 522#ifdef CONFIG_DEBUG_VM
 523	if (WARN_ON_ONCE(__read_cr3() != build_cr3(real_prev->pgd, prev_asid))) {
 
 524		/*
 525		 * If we were to BUG here, we'd be very likely to kill
 526		 * the system so hard that we don't see the call trace.
 527		 * Try to recover instead by ignoring the error and doing
 528		 * a global flush to minimize the chance of corruption.
 529		 *
 530		 * (This is far from being a fully correct recovery.
 531		 *  Architecturally, the CPU could prefetch something
 532		 *  back into an incorrect ASID slot and leave it there
 533		 *  to cause trouble down the road.  It's better than
 534		 *  nothing, though.)
 535		 */
 536		__flush_tlb_all();
 537	}
 538#endif
 539	if (was_lazy)
 540		this_cpu_write(cpu_tlbstate_shared.is_lazy, false);
 541
 542	/*
 543	 * The membarrier system call requires a full memory barrier and
 544	 * core serialization before returning to user-space, after
 545	 * storing to rq->curr, when changing mm.  This is because
 546	 * membarrier() sends IPIs to all CPUs that are in the target mm
 547	 * to make them issue memory barriers.  However, if another CPU
 548	 * switches to/from the target mm concurrently with
 549	 * membarrier(), it can cause that CPU not to receive an IPI
 550	 * when it really should issue a memory barrier.  Writing to CR3
 551	 * provides that full memory barrier and core serializing
 552	 * instruction.
 553	 */
 554	if (real_prev == next) {
 
 555		VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[prev_asid].ctx_id) !=
 556			   next->context.ctx_id);
 557
 558		/*
 
 
 
 
 
 559		 * Even in lazy TLB mode, the CPU should stay set in the
 560		 * mm_cpumask. The TLB shootdown code can figure out from
 561		 * cpu_tlbstate_shared.is_lazy whether or not to send an IPI.
 562		 */
 563		if (WARN_ON_ONCE(real_prev != &init_mm &&
 564				 !cpumask_test_cpu(cpu, mm_cpumask(next))))
 565			cpumask_set_cpu(cpu, mm_cpumask(next));
 566
 567		/*
 568		 * If the CPU is not in lazy TLB mode, we are just switching
 569		 * from one thread in a process to another thread in the same
 570		 * process. No TLB flush required.
 571		 */
 572		if (!was_lazy)
 573			return;
 574
 575		/*
 576		 * Read the tlb_gen to check whether a flush is needed.
 577		 * If the TLB is up to date, just use it.
 578		 * The barrier synchronizes with the tlb_gen increment in
 579		 * the TLB shootdown code.
 580		 */
 581		smp_mb();
 582		next_tlb_gen = atomic64_read(&next->context.tlb_gen);
 583		if (this_cpu_read(cpu_tlbstate.ctxs[prev_asid].tlb_gen) ==
 584				next_tlb_gen)
 585			return;
 586
 587		/*
 588		 * TLB contents went out of date while we were in lazy
 589		 * mode. Fall through to the TLB switching code below.
 590		 */
 591		new_asid = prev_asid;
 592		need_flush = true;
 593	} else {
 594		/*
 595		 * Apply process to process speculation vulnerability
 596		 * mitigations if applicable.
 597		 */
 598		cond_mitigation(tsk);
 599
 600		/*
 601		 * Stop remote flushes for the previous mm.
 602		 * Skip kernel threads; we never send init_mm TLB flushing IPIs,
 603		 * but the bitmap manipulation can cause cache line contention.
 604		 */
 605		if (real_prev != &init_mm) {
 606			VM_WARN_ON_ONCE(!cpumask_test_cpu(cpu,
 607						mm_cpumask(real_prev)));
 608			cpumask_clear_cpu(cpu, mm_cpumask(real_prev));
 609		}
 610
 611		/*
 612		 * Start remote flushes and then read tlb_gen.
 613		 */
 614		if (next != &init_mm)
 615			cpumask_set_cpu(cpu, mm_cpumask(next));
 616		next_tlb_gen = atomic64_read(&next->context.tlb_gen);
 617
 618		choose_new_asid(next, next_tlb_gen, &new_asid, &need_flush);
 619
 620		/* Let nmi_uaccess_okay() know that we're changing CR3. */
 621		this_cpu_write(cpu_tlbstate.loaded_mm, LOADED_MM_SWITCHING);
 622		barrier();
 623	}
 624
 
 625	if (need_flush) {
 626		this_cpu_write(cpu_tlbstate.ctxs[new_asid].ctx_id, next->context.ctx_id);
 627		this_cpu_write(cpu_tlbstate.ctxs[new_asid].tlb_gen, next_tlb_gen);
 628		load_new_mm_cr3(next->pgd, new_asid, true);
 629
 630		trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
 631	} else {
 632		/* The new ASID is already up to date. */
 633		load_new_mm_cr3(next->pgd, new_asid, false);
 634
 635		trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, 0);
 636	}
 637
 638	/* Make sure we write CR3 before loaded_mm. */
 639	barrier();
 640
 641	this_cpu_write(cpu_tlbstate.loaded_mm, next);
 642	this_cpu_write(cpu_tlbstate.loaded_mm_asid, new_asid);
 
 643
 644	if (next != real_prev) {
 645		cr4_update_pce_mm(next);
 646		switch_ldt(real_prev, next);
 647	}
 648}
 649
 650/*
 651 * Please ignore the name of this function.  It should be called
 652 * switch_to_kernel_thread().
 653 *
 654 * enter_lazy_tlb() is a hint from the scheduler that we are entering a
 655 * kernel thread or other context without an mm.  Acceptable implementations
 656 * include doing nothing whatsoever, switching to init_mm, or various clever
 657 * lazy tricks to try to minimize TLB flushes.
 658 *
 659 * The scheduler reserves the right to call enter_lazy_tlb() several times
 660 * in a row.  It will notify us that we're going back to a real mm by
 661 * calling switch_mm_irqs_off().
 662 */
 663void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
 664{
 665	if (this_cpu_read(cpu_tlbstate.loaded_mm) == &init_mm)
 666		return;
 667
 668	this_cpu_write(cpu_tlbstate_shared.is_lazy, true);
 669}
 670
 671/*
 672 * Call this when reinitializing a CPU.  It fixes the following potential
 673 * problems:
 674 *
 675 * - The ASID changed from what cpu_tlbstate thinks it is (most likely
 676 *   because the CPU was taken down and came back up with CR3's PCID
 677 *   bits clear.  CPU hotplug can do this.
 678 *
 679 * - The TLB contains junk in slots corresponding to inactive ASIDs.
 680 *
 681 * - The CPU went so far out to lunch that it may have missed a TLB
 682 *   flush.
 683 */
 684void initialize_tlbstate_and_flush(void)
 685{
 686	int i;
 687	struct mm_struct *mm = this_cpu_read(cpu_tlbstate.loaded_mm);
 688	u64 tlb_gen = atomic64_read(&init_mm.context.tlb_gen);
 
 689	unsigned long cr3 = __read_cr3();
 690
 691	/* Assert that CR3 already references the right mm. */
 692	WARN_ON((cr3 & CR3_ADDR_MASK) != __pa(mm->pgd));
 693
 
 
 
 
 694	/*
 695	 * Assert that CR4.PCIDE is set if needed.  (CR4.PCIDE initialization
 696	 * doesn't work like other CR4 bits because it can only be set from
 697	 * long mode.)
 698	 */
 699	WARN_ON(boot_cpu_has(X86_FEATURE_PCID) &&
 700		!(cr4_read_shadow() & X86_CR4_PCIDE));
 701
 702	/* Force ASID 0 and force a TLB flush. */
 703	write_cr3(build_cr3(mm->pgd, 0));
 704
 705	/* Reinitialize tlbstate. */
 706	this_cpu_write(cpu_tlbstate.last_user_mm_spec, LAST_USER_MM_INIT);
 707	this_cpu_write(cpu_tlbstate.loaded_mm_asid, 0);
 708	this_cpu_write(cpu_tlbstate.next_asid, 1);
 709	this_cpu_write(cpu_tlbstate.ctxs[0].ctx_id, mm->context.ctx_id);
 710	this_cpu_write(cpu_tlbstate.ctxs[0].tlb_gen, tlb_gen);
 
 711
 712	for (i = 1; i < TLB_NR_DYN_ASIDS; i++)
 713		this_cpu_write(cpu_tlbstate.ctxs[i].ctx_id, 0);
 714}
 715
 716/*
 717 * flush_tlb_func()'s memory ordering requirement is that any
 718 * TLB fills that happen after we flush the TLB are ordered after we
 719 * read active_mm's tlb_gen.  We don't need any explicit barriers
 720 * because all x86 flush operations are serializing and the
 721 * atomic64_read operation won't be reordered by the compiler.
 722 */
 723static void flush_tlb_func(void *info)
 724{
 725	/*
 726	 * We have three different tlb_gen values in here.  They are:
 727	 *
 728	 * - mm_tlb_gen:     the latest generation.
 729	 * - local_tlb_gen:  the generation that this CPU has already caught
 730	 *                   up to.
 731	 * - f->new_tlb_gen: the generation that the requester of the flush
 732	 *                   wants us to catch up to.
 733	 */
 734	const struct flush_tlb_info *f = info;
 735	struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
 736	u32 loaded_mm_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
 737	u64 local_tlb_gen = this_cpu_read(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen);
 738	bool local = smp_processor_id() == f->initiating_cpu;
 739	unsigned long nr_invalidate = 0;
 740	u64 mm_tlb_gen;
 741
 742	/* This code cannot presently handle being reentered. */
 743	VM_WARN_ON(!irqs_disabled());
 744
 745	if (!local) {
 746		inc_irq_stat(irq_tlb_count);
 747		count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 748
 749		/* Can only happen on remote CPUs */
 750		if (f->mm && f->mm != loaded_mm)
 751			return;
 752	}
 753
 754	if (unlikely(loaded_mm == &init_mm))
 755		return;
 756
 757	VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[loaded_mm_asid].ctx_id) !=
 758		   loaded_mm->context.ctx_id);
 759
 760	if (this_cpu_read(cpu_tlbstate_shared.is_lazy)) {
 761		/*
 762		 * We're in lazy mode.  We need to at least flush our
 763		 * paging-structure cache to avoid speculatively reading
 764		 * garbage into our TLB.  Since switching to init_mm is barely
 765		 * slower than a minimal flush, just switch to init_mm.
 766		 *
 767		 * This should be rare, with native_flush_tlb_multi() skipping
 768		 * IPIs to lazy TLB mode CPUs.
 769		 */
 770		switch_mm_irqs_off(NULL, &init_mm, NULL);
 771		return;
 772	}
 773
 774	if (unlikely(f->new_tlb_gen != TLB_GENERATION_INVALID &&
 775		     f->new_tlb_gen <= local_tlb_gen)) {
 776		/*
 777		 * The TLB is already up to date in respect to f->new_tlb_gen.
 778		 * While the core might be still behind mm_tlb_gen, checking
 779		 * mm_tlb_gen unnecessarily would have negative caching effects
 780		 * so avoid it.
 781		 */
 782		return;
 783	}
 784
 785	/*
 786	 * Defer mm_tlb_gen reading as long as possible to avoid cache
 787	 * contention.
 788	 */
 789	mm_tlb_gen = atomic64_read(&loaded_mm->context.tlb_gen);
 790
 791	if (unlikely(local_tlb_gen == mm_tlb_gen)) {
 792		/*
 793		 * There's nothing to do: we're already up to date.  This can
 794		 * happen if two concurrent flushes happen -- the first flush to
 795		 * be handled can catch us all the way up, leaving no work for
 796		 * the second flush.
 797		 */
 798		goto done;
 799	}
 800
 801	WARN_ON_ONCE(local_tlb_gen > mm_tlb_gen);
 802	WARN_ON_ONCE(f->new_tlb_gen > mm_tlb_gen);
 803
 804	/*
 805	 * If we get to this point, we know that our TLB is out of date.
 806	 * This does not strictly imply that we need to flush (it's
 807	 * possible that f->new_tlb_gen <= local_tlb_gen), but we're
 808	 * going to need to flush in the very near future, so we might
 809	 * as well get it over with.
 810	 *
 811	 * The only question is whether to do a full or partial flush.
 812	 *
 813	 * We do a partial flush if requested and two extra conditions
 814	 * are met:
 815	 *
 816	 * 1. f->new_tlb_gen == local_tlb_gen + 1.  We have an invariant that
 817	 *    we've always done all needed flushes to catch up to
 818	 *    local_tlb_gen.  If, for example, local_tlb_gen == 2 and
 819	 *    f->new_tlb_gen == 3, then we know that the flush needed to bring
 820	 *    us up to date for tlb_gen 3 is the partial flush we're
 821	 *    processing.
 822	 *
 823	 *    As an example of why this check is needed, suppose that there
 824	 *    are two concurrent flushes.  The first is a full flush that
 825	 *    changes context.tlb_gen from 1 to 2.  The second is a partial
 826	 *    flush that changes context.tlb_gen from 2 to 3.  If they get
 827	 *    processed on this CPU in reverse order, we'll see
 828	 *     local_tlb_gen == 1, mm_tlb_gen == 3, and end != TLB_FLUSH_ALL.
 829	 *    If we were to use __flush_tlb_one_user() and set local_tlb_gen to
 830	 *    3, we'd be break the invariant: we'd update local_tlb_gen above
 831	 *    1 without the full flush that's needed for tlb_gen 2.
 832	 *
 833	 * 2. f->new_tlb_gen == mm_tlb_gen.  This is purely an optimization.
 834	 *    Partial TLB flushes are not all that much cheaper than full TLB
 835	 *    flushes, so it seems unlikely that it would be a performance win
 836	 *    to do a partial flush if that won't bring our TLB fully up to
 837	 *    date.  By doing a full flush instead, we can increase
 838	 *    local_tlb_gen all the way to mm_tlb_gen and we can probably
 839	 *    avoid another flush in the very near future.
 840	 */
 841	if (f->end != TLB_FLUSH_ALL &&
 842	    f->new_tlb_gen == local_tlb_gen + 1 &&
 843	    f->new_tlb_gen == mm_tlb_gen) {
 844		/* Partial flush */
 845		unsigned long addr = f->start;
 846
 847		/* Partial flush cannot have invalid generations */
 848		VM_WARN_ON(f->new_tlb_gen == TLB_GENERATION_INVALID);
 849
 850		/* Partial flush must have valid mm */
 851		VM_WARN_ON(f->mm == NULL);
 852
 853		nr_invalidate = (f->end - f->start) >> f->stride_shift;
 854
 855		while (addr < f->end) {
 856			flush_tlb_one_user(addr);
 857			addr += 1UL << f->stride_shift;
 858		}
 859		if (local)
 860			count_vm_tlb_events(NR_TLB_LOCAL_FLUSH_ONE, nr_invalidate);
 861	} else {
 862		/* Full flush. */
 863		nr_invalidate = TLB_FLUSH_ALL;
 864
 865		flush_tlb_local();
 866		if (local)
 867			count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 868	}
 869
 870	/* Both paths above update our state to mm_tlb_gen. */
 871	this_cpu_write(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen, mm_tlb_gen);
 872
 873	/* Tracing is done in a unified manner to reduce the code size */
 874done:
 875	trace_tlb_flush(!local ? TLB_REMOTE_SHOOTDOWN :
 876				(f->mm == NULL) ? TLB_LOCAL_SHOOTDOWN :
 877						  TLB_LOCAL_MM_SHOOTDOWN,
 878			nr_invalidate);
 879}
 880
 881static bool tlb_is_not_lazy(int cpu, void *data)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 882{
 883	return !per_cpu(cpu_tlbstate_shared.is_lazy, cpu);
 
 
 
 
 884}
 885
 886DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state_shared, cpu_tlbstate_shared);
 887EXPORT_PER_CPU_SYMBOL(cpu_tlbstate_shared);
 888
 889STATIC_NOPV void native_flush_tlb_multi(const struct cpumask *cpumask,
 890					 const struct flush_tlb_info *info)
 891{
 892	/*
 893	 * Do accounting and tracing. Note that there are (and have always been)
 894	 * cases in which a remote TLB flush will be traced, but eventually
 895	 * would not happen.
 896	 */
 897	count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 898	if (info->end == TLB_FLUSH_ALL)
 899		trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL);
 900	else
 901		trace_tlb_flush(TLB_REMOTE_SEND_IPI,
 902				(info->end - info->start) >> PAGE_SHIFT);
 903
 904	/*
 905	 * If no page tables were freed, we can skip sending IPIs to
 906	 * CPUs in lazy TLB mode. They will flush the CPU themselves
 907	 * at the next context switch.
 908	 *
 909	 * However, if page tables are getting freed, we need to send the
 910	 * IPI everywhere, to prevent CPUs in lazy TLB mode from tripping
 911	 * up on the new contents of what used to be page tables, while
 912	 * doing a speculative memory access.
 913	 */
 914	if (info->freed_tables)
 915		on_each_cpu_mask(cpumask, flush_tlb_func, (void *)info, true);
 916	else
 917		on_each_cpu_cond_mask(tlb_is_not_lazy, flush_tlb_func,
 918				(void *)info, 1, cpumask);
 919}
 920
 921void flush_tlb_multi(const struct cpumask *cpumask,
 922		      const struct flush_tlb_info *info)
 923{
 924	__flush_tlb_multi(cpumask, info);
 925}
 926
 927/*
 928 * See Documentation/x86/tlb.rst for details.  We choose 33
 929 * because it is large enough to cover the vast majority (at
 930 * least 95%) of allocations, and is small enough that we are
 931 * confident it will not cause too much overhead.  Each single
 932 * flush is about 100 ns, so this caps the maximum overhead at
 933 * _about_ 3,000 ns.
 934 *
 935 * This is in units of pages.
 936 */
 937unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
 938
 939static DEFINE_PER_CPU_SHARED_ALIGNED(struct flush_tlb_info, flush_tlb_info);
 940
 941#ifdef CONFIG_DEBUG_VM
 942static DEFINE_PER_CPU(unsigned int, flush_tlb_info_idx);
 943#endif
 944
 945static struct flush_tlb_info *get_flush_tlb_info(struct mm_struct *mm,
 946			unsigned long start, unsigned long end,
 947			unsigned int stride_shift, bool freed_tables,
 948			u64 new_tlb_gen)
 949{
 950	struct flush_tlb_info *info = this_cpu_ptr(&flush_tlb_info);
 951
 952#ifdef CONFIG_DEBUG_VM
 953	/*
 954	 * Ensure that the following code is non-reentrant and flush_tlb_info
 955	 * is not overwritten. This means no TLB flushing is initiated by
 956	 * interrupt handlers and machine-check exception handlers.
 957	 */
 958	BUG_ON(this_cpu_inc_return(flush_tlb_info_idx) != 1);
 959#endif
 960
 961	info->start		= start;
 962	info->end		= end;
 963	info->mm		= mm;
 964	info->stride_shift	= stride_shift;
 965	info->freed_tables	= freed_tables;
 966	info->new_tlb_gen	= new_tlb_gen;
 967	info->initiating_cpu	= smp_processor_id();
 
 968
 969	return info;
 970}
 971
 972static void put_flush_tlb_info(void)
 973{
 974#ifdef CONFIG_DEBUG_VM
 975	/* Complete reentrancy prevention checks */
 976	barrier();
 977	this_cpu_dec(flush_tlb_info_idx);
 978#endif
 979}
 980
 981void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
 982				unsigned long end, unsigned int stride_shift,
 983				bool freed_tables)
 984{
 985	struct flush_tlb_info *info;
 986	u64 new_tlb_gen;
 987	int cpu;
 988
 989	cpu = get_cpu();
 990
 991	/* Should we flush just the requested range? */
 992	if ((end == TLB_FLUSH_ALL) ||
 993	    ((end - start) >> stride_shift) > tlb_single_page_flush_ceiling) {
 994		start = 0;
 995		end = TLB_FLUSH_ALL;
 996	}
 997
 998	/* This is also a barrier that synchronizes with switch_mm(). */
 999	new_tlb_gen = inc_mm_tlb_gen(mm);
1000
1001	info = get_flush_tlb_info(mm, start, end, stride_shift, freed_tables,
1002				  new_tlb_gen);
1003
1004	/*
1005	 * flush_tlb_multi() is not optimized for the common case in which only
1006	 * a local TLB flush is needed. Optimize this use-case by calling
1007	 * flush_tlb_func_local() directly in this case.
1008	 */
1009	if (cpumask_any_but(mm_cpumask(mm), cpu) < nr_cpu_ids) {
 
1010		flush_tlb_multi(mm_cpumask(mm), info);
1011	} else if (mm == this_cpu_read(cpu_tlbstate.loaded_mm)) {
1012		lockdep_assert_irqs_enabled();
1013		local_irq_disable();
1014		flush_tlb_func(info);
1015		local_irq_enable();
1016	}
1017
1018	put_flush_tlb_info();
1019	put_cpu();
 
1020}
1021
1022
1023static void do_flush_tlb_all(void *info)
1024{
1025	count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
1026	__flush_tlb_all();
1027}
1028
1029void flush_tlb_all(void)
1030{
1031	count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
1032	on_each_cpu(do_flush_tlb_all, NULL, 1);
1033}
1034
1035static void do_kernel_range_flush(void *info)
1036{
1037	struct flush_tlb_info *f = info;
1038	unsigned long addr;
1039
1040	/* flush range by one by one 'invlpg' */
1041	for (addr = f->start; addr < f->end; addr += PAGE_SIZE)
1042		flush_tlb_one_kernel(addr);
1043}
1044
1045void flush_tlb_kernel_range(unsigned long start, unsigned long end)
1046{
1047	/* Balance as user space task's flush, a bit conservative */
1048	if (end == TLB_FLUSH_ALL ||
1049	    (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) {
1050		on_each_cpu(do_flush_tlb_all, NULL, 1);
1051	} else {
1052		struct flush_tlb_info *info;
1053
1054		preempt_disable();
1055		info = get_flush_tlb_info(NULL, start, end, 0, false,
1056					  TLB_GENERATION_INVALID);
1057
1058		on_each_cpu(do_kernel_range_flush, info, 1);
1059
1060		put_flush_tlb_info();
1061		preempt_enable();
1062	}
1063}
1064
1065/*
1066 * This can be used from process context to figure out what the value of
1067 * CR3 is without needing to do a (slow) __read_cr3().
1068 *
1069 * It's intended to be used for code like KVM that sneakily changes CR3
1070 * and needs to restore it.  It needs to be used very carefully.
1071 */
1072unsigned long __get_current_cr3_fast(void)
1073{
1074	unsigned long cr3 = build_cr3(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd,
1075		this_cpu_read(cpu_tlbstate.loaded_mm_asid));
 
 
1076
1077	/* For now, be very restrictive about when this can be called. */
1078	VM_WARN_ON(in_nmi() || preemptible());
1079
1080	VM_BUG_ON(cr3 != __read_cr3());
1081	return cr3;
1082}
1083EXPORT_SYMBOL_GPL(__get_current_cr3_fast);
1084
1085/*
1086 * Flush one page in the kernel mapping
1087 */
1088void flush_tlb_one_kernel(unsigned long addr)
1089{
1090	count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
1091
1092	/*
1093	 * If PTI is off, then __flush_tlb_one_user() is just INVLPG or its
1094	 * paravirt equivalent.  Even with PCID, this is sufficient: we only
1095	 * use PCID if we also use global PTEs for the kernel mapping, and
1096	 * INVLPG flushes global translations across all address spaces.
1097	 *
1098	 * If PTI is on, then the kernel is mapped with non-global PTEs, and
1099	 * __flush_tlb_one_user() will flush the given address for the current
1100	 * kernel address space and for its usermode counterpart, but it does
1101	 * not flush it for other address spaces.
1102	 */
1103	flush_tlb_one_user(addr);
1104
1105	if (!static_cpu_has(X86_FEATURE_PTI))
1106		return;
1107
1108	/*
1109	 * See above.  We need to propagate the flush to all other address
1110	 * spaces.  In principle, we only need to propagate it to kernelmode
1111	 * address spaces, but the extra bookkeeping we would need is not
1112	 * worth it.
1113	 */
1114	this_cpu_write(cpu_tlbstate.invalidate_other, true);
1115}
1116
1117/*
1118 * Flush one page in the user mapping
1119 */
1120STATIC_NOPV void native_flush_tlb_one_user(unsigned long addr)
1121{
1122	u32 loaded_mm_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
 
1123
1124	asm volatile("invlpg (%0)" ::"r" (addr) : "memory");
 
1125
 
1126	if (!static_cpu_has(X86_FEATURE_PTI))
1127		return;
1128
 
 
 
1129	/*
1130	 * Some platforms #GP if we call invpcid(type=1/2) before CR4.PCIDE=1.
1131	 * Just use invalidate_user_asid() in case we are called early.
 
1132	 */
1133	if (!this_cpu_has(X86_FEATURE_INVPCID_SINGLE))
1134		invalidate_user_asid(loaded_mm_asid);
1135	else
1136		invpcid_flush_one(user_pcid(loaded_mm_asid), addr);
 
 
1137}
1138
1139void flush_tlb_one_user(unsigned long addr)
1140{
1141	__flush_tlb_one_user(addr);
1142}
1143
1144/*
1145 * Flush everything
1146 */
1147STATIC_NOPV void native_flush_tlb_global(void)
1148{
1149	unsigned long flags;
1150
1151	if (static_cpu_has(X86_FEATURE_INVPCID)) {
1152		/*
1153		 * Using INVPCID is considerably faster than a pair of writes
1154		 * to CR4 sandwiched inside an IRQ flag save/restore.
1155		 *
1156		 * Note, this works with CR4.PCIDE=0 or 1.
1157		 */
1158		invpcid_flush_all();
1159		return;
1160	}
1161
1162	/*
1163	 * Read-modify-write to CR4 - protect it from preemption and
1164	 * from interrupts. (Use the raw variant because this code can
1165	 * be called from deep inside debugging code.)
1166	 */
1167	raw_local_irq_save(flags);
1168
1169	__native_tlb_flush_global(this_cpu_read(cpu_tlbstate.cr4));
1170
1171	raw_local_irq_restore(flags);
1172}
1173
1174/*
1175 * Flush the entire current user mapping
1176 */
1177STATIC_NOPV void native_flush_tlb_local(void)
1178{
1179	/*
1180	 * Preemption or interrupts must be disabled to protect the access
1181	 * to the per CPU variable and to prevent being preempted between
1182	 * read_cr3() and write_cr3().
1183	 */
1184	WARN_ON_ONCE(preemptible());
1185
1186	invalidate_user_asid(this_cpu_read(cpu_tlbstate.loaded_mm_asid));
1187
1188	/* If current->mm == NULL then the read_cr3() "borrows" an mm */
1189	native_write_cr3(__native_read_cr3());
1190}
1191
1192void flush_tlb_local(void)
1193{
1194	__flush_tlb_local();
1195}
1196
1197/*
1198 * Flush everything
1199 */
1200void __flush_tlb_all(void)
1201{
1202	/*
1203	 * This is to catch users with enabled preemption and the PGE feature
1204	 * and don't trigger the warning in __native_flush_tlb().
1205	 */
1206	VM_WARN_ON_ONCE(preemptible());
1207
1208	if (boot_cpu_has(X86_FEATURE_PGE)) {
1209		__flush_tlb_global();
1210	} else {
1211		/*
1212		 * !PGE -> !PCID (setup_pcid()), thus every flush is total.
1213		 */
1214		flush_tlb_local();
1215	}
1216}
1217EXPORT_SYMBOL_GPL(__flush_tlb_all);
1218
1219void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
1220{
1221	struct flush_tlb_info *info;
1222
1223	int cpu = get_cpu();
1224
1225	info = get_flush_tlb_info(NULL, 0, TLB_FLUSH_ALL, 0, false,
1226				  TLB_GENERATION_INVALID);
1227	/*
1228	 * flush_tlb_multi() is not optimized for the common case in which only
1229	 * a local TLB flush is needed. Optimize this use-case by calling
1230	 * flush_tlb_func_local() directly in this case.
1231	 */
1232	if (cpumask_any_but(&batch->cpumask, cpu) < nr_cpu_ids) {
1233		flush_tlb_multi(&batch->cpumask, info);
1234	} else if (cpumask_test_cpu(cpu, &batch->cpumask)) {
1235		lockdep_assert_irqs_enabled();
1236		local_irq_disable();
1237		flush_tlb_func(info);
1238		local_irq_enable();
1239	}
1240
1241	cpumask_clear(&batch->cpumask);
1242
1243	put_flush_tlb_info();
1244	put_cpu();
1245}
1246
1247/*
1248 * Blindly accessing user memory from NMI context can be dangerous
1249 * if we're in the middle of switching the current user task or
1250 * switching the loaded mm.  It can also be dangerous if we
1251 * interrupted some kernel code that was temporarily using a
1252 * different mm.
1253 */
1254bool nmi_uaccess_okay(void)
1255{
1256	struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
1257	struct mm_struct *current_mm = current->mm;
1258
1259	VM_WARN_ON_ONCE(!loaded_mm);
1260
1261	/*
1262	 * The condition we want to check is
1263	 * current_mm->pgd == __va(read_cr3_pa()).  This may be slow, though,
1264	 * if we're running in a VM with shadow paging, and nmi_uaccess_okay()
1265	 * is supposed to be reasonably fast.
1266	 *
1267	 * Instead, we check the almost equivalent but somewhat conservative
1268	 * condition below, and we rely on the fact that switch_mm_irqs_off()
1269	 * sets loaded_mm to LOADED_MM_SWITCHING before writing to CR3.
1270	 */
1271	if (loaded_mm != current_mm)
1272		return false;
1273
1274	VM_WARN_ON_ONCE(current_mm->pgd != __va(read_cr3_pa()));
1275
1276	return true;
1277}
1278
1279static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
1280			     size_t count, loff_t *ppos)
1281{
1282	char buf[32];
1283	unsigned int len;
1284
1285	len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
1286	return simple_read_from_buffer(user_buf, count, ppos, buf, len);
1287}
1288
1289static ssize_t tlbflush_write_file(struct file *file,
1290		 const char __user *user_buf, size_t count, loff_t *ppos)
1291{
1292	char buf[32];
1293	ssize_t len;
1294	int ceiling;
1295
1296	len = min(count, sizeof(buf) - 1);
1297	if (copy_from_user(buf, user_buf, len))
1298		return -EFAULT;
1299
1300	buf[len] = '\0';
1301	if (kstrtoint(buf, 0, &ceiling))
1302		return -EINVAL;
1303
1304	if (ceiling < 0)
1305		return -EINVAL;
1306
1307	tlb_single_page_flush_ceiling = ceiling;
1308	return count;
1309}
1310
1311static const struct file_operations fops_tlbflush = {
1312	.read = tlbflush_read_file,
1313	.write = tlbflush_write_file,
1314	.llseek = default_llseek,
1315};
1316
1317static int __init create_tlb_single_page_flush_ceiling(void)
1318{
1319	debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
1320			    arch_debugfs_dir, NULL, &fops_tlbflush);
1321	return 0;
1322}
1323late_initcall(create_tlb_single_page_flush_ceiling);
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-only
   2#include <linux/init.h>
   3
   4#include <linux/mm.h>
   5#include <linux/spinlock.h>
   6#include <linux/smp.h>
   7#include <linux/interrupt.h>
   8#include <linux/export.h>
   9#include <linux/cpu.h>
  10#include <linux/debugfs.h>
  11#include <linux/sched/smt.h>
  12#include <linux/task_work.h>
  13#include <linux/mmu_notifier.h>
  14#include <linux/mmu_context.h>
  15
  16#include <asm/tlbflush.h>
  17#include <asm/mmu_context.h>
  18#include <asm/nospec-branch.h>
  19#include <asm/cache.h>
  20#include <asm/cacheflush.h>
  21#include <asm/apic.h>
  22#include <asm/perf_event.h>
  23#include <asm/tlb.h>
  24
  25#include "mm_internal.h"
  26
  27#ifdef CONFIG_PARAVIRT
  28# define STATIC_NOPV
  29#else
  30# define STATIC_NOPV			static
  31# define __flush_tlb_local		native_flush_tlb_local
  32# define __flush_tlb_global		native_flush_tlb_global
  33# define __flush_tlb_one_user(addr)	native_flush_tlb_one_user(addr)
  34# define __flush_tlb_multi(msk, info)	native_flush_tlb_multi(msk, info)
  35#endif
  36
  37/*
  38 *	TLB flushing, formerly SMP-only
  39 *		c/o Linus Torvalds.
  40 *
  41 *	These mean you can really definitely utterly forget about
  42 *	writing to user space from interrupts. (Its not allowed anyway).
  43 *
  44 *	Optimizations Manfred Spraul <manfred@colorfullife.com>
  45 *
  46 *	More scalable flush, from Andi Kleen
  47 *
  48 *	Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi
  49 */
  50
  51/*
  52 * Bits to mangle the TIF_SPEC_* state into the mm pointer which is
  53 * stored in cpu_tlb_state.last_user_mm_spec.
  54 */
  55#define LAST_USER_MM_IBPB	0x1UL
  56#define LAST_USER_MM_L1D_FLUSH	0x2UL
  57#define LAST_USER_MM_SPEC_MASK	(LAST_USER_MM_IBPB | LAST_USER_MM_L1D_FLUSH)
  58
  59/* Bits to set when tlbstate and flush is (re)initialized */
  60#define LAST_USER_MM_INIT	LAST_USER_MM_IBPB
  61
  62/*
  63 * The x86 feature is called PCID (Process Context IDentifier). It is similar
  64 * to what is traditionally called ASID on the RISC processors.
  65 *
  66 * We don't use the traditional ASID implementation, where each process/mm gets
  67 * its own ASID and flush/restart when we run out of ASID space.
  68 *
  69 * Instead we have a small per-cpu array of ASIDs and cache the last few mm's
  70 * that came by on this CPU, allowing cheaper switch_mm between processes on
  71 * this CPU.
  72 *
  73 * We end up with different spaces for different things. To avoid confusion we
  74 * use different names for each of them:
  75 *
  76 * ASID  - [0, TLB_NR_DYN_ASIDS-1]
  77 *         the canonical identifier for an mm
  78 *
  79 * kPCID - [1, TLB_NR_DYN_ASIDS]
  80 *         the value we write into the PCID part of CR3; corresponds to the
  81 *         ASID+1, because PCID 0 is special.
  82 *
  83 * uPCID - [2048 + 1, 2048 + TLB_NR_DYN_ASIDS]
  84 *         for KPTI each mm has two address spaces and thus needs two
  85 *         PCID values, but we can still do with a single ASID denomination
  86 *         for each mm. Corresponds to kPCID + 2048.
  87 *
  88 */
  89
 
 
 
  90/*
  91 * When enabled, MITIGATION_PAGE_TABLE_ISOLATION consumes a single bit for
  92 * user/kernel switches
  93 */
  94#ifdef CONFIG_MITIGATION_PAGE_TABLE_ISOLATION
  95# define PTI_CONSUMED_PCID_BITS	1
  96#else
  97# define PTI_CONSUMED_PCID_BITS	0
  98#endif
  99
 100#define CR3_AVAIL_PCID_BITS (X86_CR3_PCID_BITS - PTI_CONSUMED_PCID_BITS)
 101
 102/*
 103 * ASIDs are zero-based: 0->MAX_AVAIL_ASID are valid.  -1 below to account
 104 * for them being zero-based.  Another -1 is because PCID 0 is reserved for
 105 * use by non-PCID-aware users.
 106 */
 107#define MAX_ASID_AVAILABLE ((1 << CR3_AVAIL_PCID_BITS) - 2)
 108
 109/*
 110 * Given @asid, compute kPCID
 111 */
 112static inline u16 kern_pcid(u16 asid)
 113{
 114	VM_WARN_ON_ONCE(asid > MAX_ASID_AVAILABLE);
 115
 116#ifdef CONFIG_MITIGATION_PAGE_TABLE_ISOLATION
 117	/*
 118	 * Make sure that the dynamic ASID space does not conflict with the
 119	 * bit we are using to switch between user and kernel ASIDs.
 120	 */
 121	BUILD_BUG_ON(TLB_NR_DYN_ASIDS >= (1 << X86_CR3_PTI_PCID_USER_BIT));
 122
 123	/*
 124	 * The ASID being passed in here should have respected the
 125	 * MAX_ASID_AVAILABLE and thus never have the switch bit set.
 126	 */
 127	VM_WARN_ON_ONCE(asid & (1 << X86_CR3_PTI_PCID_USER_BIT));
 128#endif
 129	/*
 130	 * The dynamically-assigned ASIDs that get passed in are small
 131	 * (<TLB_NR_DYN_ASIDS).  They never have the high switch bit set,
 132	 * so do not bother to clear it.
 133	 *
 134	 * If PCID is on, ASID-aware code paths put the ASID+1 into the
 135	 * PCID bits.  This serves two purposes.  It prevents a nasty
 136	 * situation in which PCID-unaware code saves CR3, loads some other
 137	 * value (with PCID == 0), and then restores CR3, thus corrupting
 138	 * the TLB for ASID 0 if the saved ASID was nonzero.  It also means
 139	 * that any bugs involving loading a PCID-enabled CR3 with
 140	 * CR4.PCIDE off will trigger deterministically.
 141	 */
 142	return asid + 1;
 143}
 144
 145/*
 146 * Given @asid, compute uPCID
 147 */
 148static inline u16 user_pcid(u16 asid)
 149{
 150	u16 ret = kern_pcid(asid);
 151#ifdef CONFIG_MITIGATION_PAGE_TABLE_ISOLATION
 152	ret |= 1 << X86_CR3_PTI_PCID_USER_BIT;
 153#endif
 154	return ret;
 155}
 156
 157static inline unsigned long build_cr3(pgd_t *pgd, u16 asid, unsigned long lam)
 158{
 159	unsigned long cr3 = __sme_pa(pgd) | lam;
 160
 161	if (static_cpu_has(X86_FEATURE_PCID)) {
 162		cr3 |= kern_pcid(asid);
 163	} else {
 164		VM_WARN_ON_ONCE(asid != 0);
 
 165	}
 166
 167	return cr3;
 168}
 169
 170static inline unsigned long build_cr3_noflush(pgd_t *pgd, u16 asid,
 171					      unsigned long lam)
 172{
 
 173	/*
 174	 * Use boot_cpu_has() instead of this_cpu_has() as this function
 175	 * might be called during early boot. This should work even after
 176	 * boot because all CPU's the have same capabilities:
 177	 */
 178	VM_WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_PCID));
 179	return build_cr3(pgd, asid, lam) | CR3_NOFLUSH;
 180}
 181
 182/*
 183 * We get here when we do something requiring a TLB invalidation
 184 * but could not go invalidate all of the contexts.  We do the
 185 * necessary invalidation by clearing out the 'ctx_id' which
 186 * forces a TLB flush when the context is loaded.
 187 */
 188static void clear_asid_other(void)
 189{
 190	u16 asid;
 191
 192	/*
 193	 * This is only expected to be set if we have disabled
 194	 * kernel _PAGE_GLOBAL pages.
 195	 */
 196	if (!static_cpu_has(X86_FEATURE_PTI)) {
 197		WARN_ON_ONCE(1);
 198		return;
 199	}
 200
 201	for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
 202		/* Do not need to flush the current asid */
 203		if (asid == this_cpu_read(cpu_tlbstate.loaded_mm_asid))
 204			continue;
 205		/*
 206		 * Make sure the next time we go to switch to
 207		 * this asid, we do a flush:
 208		 */
 209		this_cpu_write(cpu_tlbstate.ctxs[asid].ctx_id, 0);
 210	}
 211	this_cpu_write(cpu_tlbstate.invalidate_other, false);
 212}
 213
 214atomic64_t last_mm_ctx_id = ATOMIC64_INIT(1);
 215
 216
 217static void choose_new_asid(struct mm_struct *next, u64 next_tlb_gen,
 218			    u16 *new_asid, bool *need_flush)
 219{
 220	u16 asid;
 221
 222	if (!static_cpu_has(X86_FEATURE_PCID)) {
 223		*new_asid = 0;
 224		*need_flush = true;
 225		return;
 226	}
 227
 228	if (this_cpu_read(cpu_tlbstate.invalidate_other))
 229		clear_asid_other();
 230
 231	for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
 232		if (this_cpu_read(cpu_tlbstate.ctxs[asid].ctx_id) !=
 233		    next->context.ctx_id)
 234			continue;
 235
 236		*new_asid = asid;
 237		*need_flush = (this_cpu_read(cpu_tlbstate.ctxs[asid].tlb_gen) <
 238			       next_tlb_gen);
 239		return;
 240	}
 241
 242	/*
 243	 * We don't currently own an ASID slot on this CPU.
 244	 * Allocate a slot.
 245	 */
 246	*new_asid = this_cpu_add_return(cpu_tlbstate.next_asid, 1) - 1;
 247	if (*new_asid >= TLB_NR_DYN_ASIDS) {
 248		*new_asid = 0;
 249		this_cpu_write(cpu_tlbstate.next_asid, 1);
 250	}
 251	*need_flush = true;
 252}
 253
 254/*
 255 * Given an ASID, flush the corresponding user ASID.  We can delay this
 256 * until the next time we switch to it.
 257 *
 258 * See SWITCH_TO_USER_CR3.
 259 */
 260static inline void invalidate_user_asid(u16 asid)
 261{
 262	/* There is no user ASID if address space separation is off */
 263	if (!IS_ENABLED(CONFIG_MITIGATION_PAGE_TABLE_ISOLATION))
 264		return;
 265
 266	/*
 267	 * We only have a single ASID if PCID is off and the CR3
 268	 * write will have flushed it.
 269	 */
 270	if (!cpu_feature_enabled(X86_FEATURE_PCID))
 271		return;
 272
 273	if (!static_cpu_has(X86_FEATURE_PTI))
 274		return;
 275
 276	__set_bit(kern_pcid(asid),
 277		  (unsigned long *)this_cpu_ptr(&cpu_tlbstate.user_pcid_flush_mask));
 278}
 279
 280static void load_new_mm_cr3(pgd_t *pgdir, u16 new_asid, unsigned long lam,
 281			    bool need_flush)
 282{
 283	unsigned long new_mm_cr3;
 284
 285	if (need_flush) {
 286		invalidate_user_asid(new_asid);
 287		new_mm_cr3 = build_cr3(pgdir, new_asid, lam);
 288	} else {
 289		new_mm_cr3 = build_cr3_noflush(pgdir, new_asid, lam);
 290	}
 291
 292	/*
 293	 * Caution: many callers of this function expect
 294	 * that load_cr3() is serializing and orders TLB
 295	 * fills with respect to the mm_cpumask writes.
 296	 */
 297	write_cr3(new_mm_cr3);
 298}
 299
 300void leave_mm(void)
 301{
 302	struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
 303
 304	/*
 305	 * It's plausible that we're in lazy TLB mode while our mm is init_mm.
 306	 * If so, our callers still expect us to flush the TLB, but there
 307	 * aren't any user TLB entries in init_mm to worry about.
 308	 *
 309	 * This needs to happen before any other sanity checks due to
 310	 * intel_idle's shenanigans.
 311	 */
 312	if (loaded_mm == &init_mm)
 313		return;
 314
 315	/* Warn if we're not lazy. */
 316	WARN_ON(!this_cpu_read(cpu_tlbstate_shared.is_lazy));
 317
 318	switch_mm(NULL, &init_mm, NULL);
 319}
 320EXPORT_SYMBOL_GPL(leave_mm);
 321
 322void switch_mm(struct mm_struct *prev, struct mm_struct *next,
 323	       struct task_struct *tsk)
 324{
 325	unsigned long flags;
 326
 327	local_irq_save(flags);
 328	switch_mm_irqs_off(NULL, next, tsk);
 329	local_irq_restore(flags);
 330}
 331
 332/*
 333 * Invoked from return to user/guest by a task that opted-in to L1D
 334 * flushing but ended up running on an SMT enabled core due to wrong
 335 * affinity settings or CPU hotplug. This is part of the paranoid L1D flush
 336 * contract which this task requested.
 337 */
 338static void l1d_flush_force_sigbus(struct callback_head *ch)
 339{
 340	force_sig(SIGBUS);
 341}
 342
 343static void l1d_flush_evaluate(unsigned long prev_mm, unsigned long next_mm,
 344				struct task_struct *next)
 345{
 346	/* Flush L1D if the outgoing task requests it */
 347	if (prev_mm & LAST_USER_MM_L1D_FLUSH)
 348		wrmsrl(MSR_IA32_FLUSH_CMD, L1D_FLUSH);
 349
 350	/* Check whether the incoming task opted in for L1D flush */
 351	if (likely(!(next_mm & LAST_USER_MM_L1D_FLUSH)))
 352		return;
 353
 354	/*
 355	 * Validate that it is not running on an SMT sibling as this would
 356	 * make the exercise pointless because the siblings share L1D. If
 357	 * it runs on a SMT sibling, notify it with SIGBUS on return to
 358	 * user/guest
 359	 */
 360	if (this_cpu_read(cpu_info.smt_active)) {
 361		clear_ti_thread_flag(&next->thread_info, TIF_SPEC_L1D_FLUSH);
 362		next->l1d_flush_kill.func = l1d_flush_force_sigbus;
 363		task_work_add(next, &next->l1d_flush_kill, TWA_RESUME);
 364	}
 365}
 366
 367static unsigned long mm_mangle_tif_spec_bits(struct task_struct *next)
 368{
 369	unsigned long next_tif = read_task_thread_flags(next);
 370	unsigned long spec_bits = (next_tif >> TIF_SPEC_IB) & LAST_USER_MM_SPEC_MASK;
 371
 372	/*
 373	 * Ensure that the bit shift above works as expected and the two flags
 374	 * end up in bit 0 and 1.
 375	 */
 376	BUILD_BUG_ON(TIF_SPEC_L1D_FLUSH != TIF_SPEC_IB + 1);
 377
 378	return (unsigned long)next->mm | spec_bits;
 379}
 380
 381static void cond_mitigation(struct task_struct *next)
 382{
 383	unsigned long prev_mm, next_mm;
 384
 385	if (!next || !next->mm)
 386		return;
 387
 388	next_mm = mm_mangle_tif_spec_bits(next);
 389	prev_mm = this_cpu_read(cpu_tlbstate.last_user_mm_spec);
 390
 391	/*
 392	 * Avoid user/user BTB poisoning by flushing the branch predictor
 393	 * when switching between processes. This stops one process from
 394	 * doing Spectre-v2 attacks on another.
 395	 *
 396	 * Both, the conditional and the always IBPB mode use the mm
 397	 * pointer to avoid the IBPB when switching between tasks of the
 398	 * same process. Using the mm pointer instead of mm->context.ctx_id
 399	 * opens a hypothetical hole vs. mm_struct reuse, which is more or
 400	 * less impossible to control by an attacker. Aside of that it
 401	 * would only affect the first schedule so the theoretically
 402	 * exposed data is not really interesting.
 403	 */
 404	if (static_branch_likely(&switch_mm_cond_ibpb)) {
 405		/*
 406		 * This is a bit more complex than the always mode because
 407		 * it has to handle two cases:
 408		 *
 409		 * 1) Switch from a user space task (potential attacker)
 410		 *    which has TIF_SPEC_IB set to a user space task
 411		 *    (potential victim) which has TIF_SPEC_IB not set.
 412		 *
 413		 * 2) Switch from a user space task (potential attacker)
 414		 *    which has TIF_SPEC_IB not set to a user space task
 415		 *    (potential victim) which has TIF_SPEC_IB set.
 416		 *
 417		 * This could be done by unconditionally issuing IBPB when
 418		 * a task which has TIF_SPEC_IB set is either scheduled in
 419		 * or out. Though that results in two flushes when:
 420		 *
 421		 * - the same user space task is scheduled out and later
 422		 *   scheduled in again and only a kernel thread ran in
 423		 *   between.
 424		 *
 425		 * - a user space task belonging to the same process is
 426		 *   scheduled in after a kernel thread ran in between
 427		 *
 428		 * - a user space task belonging to the same process is
 429		 *   scheduled in immediately.
 430		 *
 431		 * Optimize this with reasonably small overhead for the
 432		 * above cases. Mangle the TIF_SPEC_IB bit into the mm
 433		 * pointer of the incoming task which is stored in
 434		 * cpu_tlbstate.last_user_mm_spec for comparison.
 435		 *
 436		 * Issue IBPB only if the mm's are different and one or
 437		 * both have the IBPB bit set.
 438		 */
 439		if (next_mm != prev_mm &&
 440		    (next_mm | prev_mm) & LAST_USER_MM_IBPB)
 441			indirect_branch_prediction_barrier();
 442	}
 443
 444	if (static_branch_unlikely(&switch_mm_always_ibpb)) {
 445		/*
 446		 * Only flush when switching to a user space task with a
 447		 * different context than the user space task which ran
 448		 * last on this CPU.
 449		 */
 450		if ((prev_mm & ~LAST_USER_MM_SPEC_MASK) !=
 451					(unsigned long)next->mm)
 452			indirect_branch_prediction_barrier();
 453	}
 454
 455	if (static_branch_unlikely(&switch_mm_cond_l1d_flush)) {
 456		/*
 457		 * Flush L1D when the outgoing task requested it and/or
 458		 * check whether the incoming task requested L1D flushing
 459		 * and ended up on an SMT sibling.
 460		 */
 461		if (unlikely((prev_mm | next_mm) & LAST_USER_MM_L1D_FLUSH))
 462			l1d_flush_evaluate(prev_mm, next_mm, next);
 463	}
 464
 465	this_cpu_write(cpu_tlbstate.last_user_mm_spec, next_mm);
 466}
 467
 468#ifdef CONFIG_PERF_EVENTS
 469static inline void cr4_update_pce_mm(struct mm_struct *mm)
 470{
 471	if (static_branch_unlikely(&rdpmc_always_available_key) ||
 472	    (!static_branch_unlikely(&rdpmc_never_available_key) &&
 473	     atomic_read(&mm->context.perf_rdpmc_allowed))) {
 474		/*
 475		 * Clear the existing dirty counters to
 476		 * prevent the leak for an RDPMC task.
 477		 */
 478		perf_clear_dirty_counters();
 479		cr4_set_bits_irqsoff(X86_CR4_PCE);
 480	} else
 481		cr4_clear_bits_irqsoff(X86_CR4_PCE);
 482}
 483
 484void cr4_update_pce(void *ignored)
 485{
 486	cr4_update_pce_mm(this_cpu_read(cpu_tlbstate.loaded_mm));
 487}
 488
 489#else
 490static inline void cr4_update_pce_mm(struct mm_struct *mm) { }
 491#endif
 492
 493/*
 494 * This optimizes when not actually switching mm's.  Some architectures use the
 495 * 'unused' argument for this optimization, but x86 must use
 496 * 'cpu_tlbstate.loaded_mm' instead because it does not always keep
 497 * 'current->active_mm' up to date.
 498 */
 499void switch_mm_irqs_off(struct mm_struct *unused, struct mm_struct *next,
 500			struct task_struct *tsk)
 501{
 502	struct mm_struct *prev = this_cpu_read(cpu_tlbstate.loaded_mm);
 503	u16 prev_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
 504	bool was_lazy = this_cpu_read(cpu_tlbstate_shared.is_lazy);
 505	unsigned cpu = smp_processor_id();
 506	unsigned long new_lam;
 507	u64 next_tlb_gen;
 508	bool need_flush;
 509	u16 new_asid;
 510
 
 
 
 
 
 
 
 
 
 511	/* We don't want flush_tlb_func() to run concurrently with us. */
 512	if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 513		WARN_ON_ONCE(!irqs_disabled());
 514
 515	/*
 516	 * Verify that CR3 is what we think it is.  This will catch
 517	 * hypothetical buggy code that directly switches to swapper_pg_dir
 518	 * without going through leave_mm() / switch_mm_irqs_off() or that
 519	 * does something like write_cr3(read_cr3_pa()).
 520	 *
 521	 * Only do this check if CONFIG_DEBUG_VM=y because __read_cr3()
 522	 * isn't free.
 523	 */
 524#ifdef CONFIG_DEBUG_VM
 525	if (WARN_ON_ONCE(__read_cr3() != build_cr3(prev->pgd, prev_asid,
 526						   tlbstate_lam_cr3_mask()))) {
 527		/*
 528		 * If we were to BUG here, we'd be very likely to kill
 529		 * the system so hard that we don't see the call trace.
 530		 * Try to recover instead by ignoring the error and doing
 531		 * a global flush to minimize the chance of corruption.
 532		 *
 533		 * (This is far from being a fully correct recovery.
 534		 *  Architecturally, the CPU could prefetch something
 535		 *  back into an incorrect ASID slot and leave it there
 536		 *  to cause trouble down the road.  It's better than
 537		 *  nothing, though.)
 538		 */
 539		__flush_tlb_all();
 540	}
 541#endif
 542	if (was_lazy)
 543		this_cpu_write(cpu_tlbstate_shared.is_lazy, false);
 544
 545	/*
 546	 * The membarrier system call requires a full memory barrier and
 547	 * core serialization before returning to user-space, after
 548	 * storing to rq->curr, when changing mm.  This is because
 549	 * membarrier() sends IPIs to all CPUs that are in the target mm
 550	 * to make them issue memory barriers.  However, if another CPU
 551	 * switches to/from the target mm concurrently with
 552	 * membarrier(), it can cause that CPU not to receive an IPI
 553	 * when it really should issue a memory barrier.  Writing to CR3
 554	 * provides that full memory barrier and core serializing
 555	 * instruction.
 556	 */
 557	if (prev == next) {
 558		/* Not actually switching mm's */
 559		VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[prev_asid].ctx_id) !=
 560			   next->context.ctx_id);
 561
 562		/*
 563		 * If this races with another thread that enables lam, 'new_lam'
 564		 * might not match tlbstate_lam_cr3_mask().
 565		 */
 566
 567		/*
 568		 * Even in lazy TLB mode, the CPU should stay set in the
 569		 * mm_cpumask. The TLB shootdown code can figure out from
 570		 * cpu_tlbstate_shared.is_lazy whether or not to send an IPI.
 571		 */
 572		if (IS_ENABLED(CONFIG_DEBUG_VM) && WARN_ON_ONCE(prev != &init_mm &&
 573				 !cpumask_test_cpu(cpu, mm_cpumask(next))))
 574			cpumask_set_cpu(cpu, mm_cpumask(next));
 575
 576		/*
 577		 * If the CPU is not in lazy TLB mode, we are just switching
 578		 * from one thread in a process to another thread in the same
 579		 * process. No TLB flush required.
 580		 */
 581		if (!was_lazy)
 582			return;
 583
 584		/*
 585		 * Read the tlb_gen to check whether a flush is needed.
 586		 * If the TLB is up to date, just use it.
 587		 * The barrier synchronizes with the tlb_gen increment in
 588		 * the TLB shootdown code.
 589		 */
 590		smp_mb();
 591		next_tlb_gen = atomic64_read(&next->context.tlb_gen);
 592		if (this_cpu_read(cpu_tlbstate.ctxs[prev_asid].tlb_gen) ==
 593				next_tlb_gen)
 594			return;
 595
 596		/*
 597		 * TLB contents went out of date while we were in lazy
 598		 * mode. Fall through to the TLB switching code below.
 599		 */
 600		new_asid = prev_asid;
 601		need_flush = true;
 602	} else {
 603		/*
 604		 * Apply process to process speculation vulnerability
 605		 * mitigations if applicable.
 606		 */
 607		cond_mitigation(tsk);
 608
 609		/*
 610		 * Stop remote flushes for the previous mm.
 611		 * Skip kernel threads; we never send init_mm TLB flushing IPIs,
 612		 * but the bitmap manipulation can cause cache line contention.
 613		 */
 614		if (prev != &init_mm) {
 615			VM_WARN_ON_ONCE(!cpumask_test_cpu(cpu,
 616						mm_cpumask(prev)));
 617			cpumask_clear_cpu(cpu, mm_cpumask(prev));
 618		}
 619
 620		/* Start receiving IPIs and then read tlb_gen (and LAM below) */
 
 
 621		if (next != &init_mm)
 622			cpumask_set_cpu(cpu, mm_cpumask(next));
 623		next_tlb_gen = atomic64_read(&next->context.tlb_gen);
 624
 625		choose_new_asid(next, next_tlb_gen, &new_asid, &need_flush);
 626
 627		/* Let nmi_uaccess_okay() know that we're changing CR3. */
 628		this_cpu_write(cpu_tlbstate.loaded_mm, LOADED_MM_SWITCHING);
 629		barrier();
 630	}
 631
 632	new_lam = mm_lam_cr3_mask(next);
 633	if (need_flush) {
 634		this_cpu_write(cpu_tlbstate.ctxs[new_asid].ctx_id, next->context.ctx_id);
 635		this_cpu_write(cpu_tlbstate.ctxs[new_asid].tlb_gen, next_tlb_gen);
 636		load_new_mm_cr3(next->pgd, new_asid, new_lam, true);
 637
 638		trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
 639	} else {
 640		/* The new ASID is already up to date. */
 641		load_new_mm_cr3(next->pgd, new_asid, new_lam, false);
 642
 643		trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, 0);
 644	}
 645
 646	/* Make sure we write CR3 before loaded_mm. */
 647	barrier();
 648
 649	this_cpu_write(cpu_tlbstate.loaded_mm, next);
 650	this_cpu_write(cpu_tlbstate.loaded_mm_asid, new_asid);
 651	cpu_tlbstate_update_lam(new_lam, mm_untag_mask(next));
 652
 653	if (next != prev) {
 654		cr4_update_pce_mm(next);
 655		switch_ldt(prev, next);
 656	}
 657}
 658
 659/*
 660 * Please ignore the name of this function.  It should be called
 661 * switch_to_kernel_thread().
 662 *
 663 * enter_lazy_tlb() is a hint from the scheduler that we are entering a
 664 * kernel thread or other context without an mm.  Acceptable implementations
 665 * include doing nothing whatsoever, switching to init_mm, or various clever
 666 * lazy tricks to try to minimize TLB flushes.
 667 *
 668 * The scheduler reserves the right to call enter_lazy_tlb() several times
 669 * in a row.  It will notify us that we're going back to a real mm by
 670 * calling switch_mm_irqs_off().
 671 */
 672void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
 673{
 674	if (this_cpu_read(cpu_tlbstate.loaded_mm) == &init_mm)
 675		return;
 676
 677	this_cpu_write(cpu_tlbstate_shared.is_lazy, true);
 678}
 679
 680/*
 681 * Call this when reinitializing a CPU.  It fixes the following potential
 682 * problems:
 683 *
 684 * - The ASID changed from what cpu_tlbstate thinks it is (most likely
 685 *   because the CPU was taken down and came back up with CR3's PCID
 686 *   bits clear.  CPU hotplug can do this.
 687 *
 688 * - The TLB contains junk in slots corresponding to inactive ASIDs.
 689 *
 690 * - The CPU went so far out to lunch that it may have missed a TLB
 691 *   flush.
 692 */
 693void initialize_tlbstate_and_flush(void)
 694{
 695	int i;
 696	struct mm_struct *mm = this_cpu_read(cpu_tlbstate.loaded_mm);
 697	u64 tlb_gen = atomic64_read(&init_mm.context.tlb_gen);
 698	unsigned long lam = mm_lam_cr3_mask(mm);
 699	unsigned long cr3 = __read_cr3();
 700
 701	/* Assert that CR3 already references the right mm. */
 702	WARN_ON((cr3 & CR3_ADDR_MASK) != __pa(mm->pgd));
 703
 704	/* LAM expected to be disabled */
 705	WARN_ON(cr3 & (X86_CR3_LAM_U48 | X86_CR3_LAM_U57));
 706	WARN_ON(lam);
 707
 708	/*
 709	 * Assert that CR4.PCIDE is set if needed.  (CR4.PCIDE initialization
 710	 * doesn't work like other CR4 bits because it can only be set from
 711	 * long mode.)
 712	 */
 713	WARN_ON(boot_cpu_has(X86_FEATURE_PCID) &&
 714		!(cr4_read_shadow() & X86_CR4_PCIDE));
 715
 716	/* Disable LAM, force ASID 0 and force a TLB flush. */
 717	write_cr3(build_cr3(mm->pgd, 0, 0));
 718
 719	/* Reinitialize tlbstate. */
 720	this_cpu_write(cpu_tlbstate.last_user_mm_spec, LAST_USER_MM_INIT);
 721	this_cpu_write(cpu_tlbstate.loaded_mm_asid, 0);
 722	this_cpu_write(cpu_tlbstate.next_asid, 1);
 723	this_cpu_write(cpu_tlbstate.ctxs[0].ctx_id, mm->context.ctx_id);
 724	this_cpu_write(cpu_tlbstate.ctxs[0].tlb_gen, tlb_gen);
 725	cpu_tlbstate_update_lam(lam, mm_untag_mask(mm));
 726
 727	for (i = 1; i < TLB_NR_DYN_ASIDS; i++)
 728		this_cpu_write(cpu_tlbstate.ctxs[i].ctx_id, 0);
 729}
 730
 731/*
 732 * flush_tlb_func()'s memory ordering requirement is that any
 733 * TLB fills that happen after we flush the TLB are ordered after we
 734 * read active_mm's tlb_gen.  We don't need any explicit barriers
 735 * because all x86 flush operations are serializing and the
 736 * atomic64_read operation won't be reordered by the compiler.
 737 */
 738static void flush_tlb_func(void *info)
 739{
 740	/*
 741	 * We have three different tlb_gen values in here.  They are:
 742	 *
 743	 * - mm_tlb_gen:     the latest generation.
 744	 * - local_tlb_gen:  the generation that this CPU has already caught
 745	 *                   up to.
 746	 * - f->new_tlb_gen: the generation that the requester of the flush
 747	 *                   wants us to catch up to.
 748	 */
 749	const struct flush_tlb_info *f = info;
 750	struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
 751	u32 loaded_mm_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
 752	u64 local_tlb_gen = this_cpu_read(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen);
 753	bool local = smp_processor_id() == f->initiating_cpu;
 754	unsigned long nr_invalidate = 0;
 755	u64 mm_tlb_gen;
 756
 757	/* This code cannot presently handle being reentered. */
 758	VM_WARN_ON(!irqs_disabled());
 759
 760	if (!local) {
 761		inc_irq_stat(irq_tlb_count);
 762		count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 763
 764		/* Can only happen on remote CPUs */
 765		if (f->mm && f->mm != loaded_mm)
 766			return;
 767	}
 768
 769	if (unlikely(loaded_mm == &init_mm))
 770		return;
 771
 772	VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[loaded_mm_asid].ctx_id) !=
 773		   loaded_mm->context.ctx_id);
 774
 775	if (this_cpu_read(cpu_tlbstate_shared.is_lazy)) {
 776		/*
 777		 * We're in lazy mode.  We need to at least flush our
 778		 * paging-structure cache to avoid speculatively reading
 779		 * garbage into our TLB.  Since switching to init_mm is barely
 780		 * slower than a minimal flush, just switch to init_mm.
 781		 *
 782		 * This should be rare, with native_flush_tlb_multi() skipping
 783		 * IPIs to lazy TLB mode CPUs.
 784		 */
 785		switch_mm_irqs_off(NULL, &init_mm, NULL);
 786		return;
 787	}
 788
 789	if (unlikely(f->new_tlb_gen != TLB_GENERATION_INVALID &&
 790		     f->new_tlb_gen <= local_tlb_gen)) {
 791		/*
 792		 * The TLB is already up to date in respect to f->new_tlb_gen.
 793		 * While the core might be still behind mm_tlb_gen, checking
 794		 * mm_tlb_gen unnecessarily would have negative caching effects
 795		 * so avoid it.
 796		 */
 797		return;
 798	}
 799
 800	/*
 801	 * Defer mm_tlb_gen reading as long as possible to avoid cache
 802	 * contention.
 803	 */
 804	mm_tlb_gen = atomic64_read(&loaded_mm->context.tlb_gen);
 805
 806	if (unlikely(local_tlb_gen == mm_tlb_gen)) {
 807		/*
 808		 * There's nothing to do: we're already up to date.  This can
 809		 * happen if two concurrent flushes happen -- the first flush to
 810		 * be handled can catch us all the way up, leaving no work for
 811		 * the second flush.
 812		 */
 813		goto done;
 814	}
 815
 816	WARN_ON_ONCE(local_tlb_gen > mm_tlb_gen);
 817	WARN_ON_ONCE(f->new_tlb_gen > mm_tlb_gen);
 818
 819	/*
 820	 * If we get to this point, we know that our TLB is out of date.
 821	 * This does not strictly imply that we need to flush (it's
 822	 * possible that f->new_tlb_gen <= local_tlb_gen), but we're
 823	 * going to need to flush in the very near future, so we might
 824	 * as well get it over with.
 825	 *
 826	 * The only question is whether to do a full or partial flush.
 827	 *
 828	 * We do a partial flush if requested and two extra conditions
 829	 * are met:
 830	 *
 831	 * 1. f->new_tlb_gen == local_tlb_gen + 1.  We have an invariant that
 832	 *    we've always done all needed flushes to catch up to
 833	 *    local_tlb_gen.  If, for example, local_tlb_gen == 2 and
 834	 *    f->new_tlb_gen == 3, then we know that the flush needed to bring
 835	 *    us up to date for tlb_gen 3 is the partial flush we're
 836	 *    processing.
 837	 *
 838	 *    As an example of why this check is needed, suppose that there
 839	 *    are two concurrent flushes.  The first is a full flush that
 840	 *    changes context.tlb_gen from 1 to 2.  The second is a partial
 841	 *    flush that changes context.tlb_gen from 2 to 3.  If they get
 842	 *    processed on this CPU in reverse order, we'll see
 843	 *     local_tlb_gen == 1, mm_tlb_gen == 3, and end != TLB_FLUSH_ALL.
 844	 *    If we were to use __flush_tlb_one_user() and set local_tlb_gen to
 845	 *    3, we'd be break the invariant: we'd update local_tlb_gen above
 846	 *    1 without the full flush that's needed for tlb_gen 2.
 847	 *
 848	 * 2. f->new_tlb_gen == mm_tlb_gen.  This is purely an optimization.
 849	 *    Partial TLB flushes are not all that much cheaper than full TLB
 850	 *    flushes, so it seems unlikely that it would be a performance win
 851	 *    to do a partial flush if that won't bring our TLB fully up to
 852	 *    date.  By doing a full flush instead, we can increase
 853	 *    local_tlb_gen all the way to mm_tlb_gen and we can probably
 854	 *    avoid another flush in the very near future.
 855	 */
 856	if (f->end != TLB_FLUSH_ALL &&
 857	    f->new_tlb_gen == local_tlb_gen + 1 &&
 858	    f->new_tlb_gen == mm_tlb_gen) {
 859		/* Partial flush */
 860		unsigned long addr = f->start;
 861
 862		/* Partial flush cannot have invalid generations */
 863		VM_WARN_ON(f->new_tlb_gen == TLB_GENERATION_INVALID);
 864
 865		/* Partial flush must have valid mm */
 866		VM_WARN_ON(f->mm == NULL);
 867
 868		nr_invalidate = (f->end - f->start) >> f->stride_shift;
 869
 870		while (addr < f->end) {
 871			flush_tlb_one_user(addr);
 872			addr += 1UL << f->stride_shift;
 873		}
 874		if (local)
 875			count_vm_tlb_events(NR_TLB_LOCAL_FLUSH_ONE, nr_invalidate);
 876	} else {
 877		/* Full flush. */
 878		nr_invalidate = TLB_FLUSH_ALL;
 879
 880		flush_tlb_local();
 881		if (local)
 882			count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 883	}
 884
 885	/* Both paths above update our state to mm_tlb_gen. */
 886	this_cpu_write(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen, mm_tlb_gen);
 887
 888	/* Tracing is done in a unified manner to reduce the code size */
 889done:
 890	trace_tlb_flush(!local ? TLB_REMOTE_SHOOTDOWN :
 891				(f->mm == NULL) ? TLB_LOCAL_SHOOTDOWN :
 892						  TLB_LOCAL_MM_SHOOTDOWN,
 893			nr_invalidate);
 894}
 895
 896static bool should_flush_tlb(int cpu, void *data)
 897{
 898	struct flush_tlb_info *info = data;
 899
 900	/* Lazy TLB will get flushed at the next context switch. */
 901	if (per_cpu(cpu_tlbstate_shared.is_lazy, cpu))
 902		return false;
 903
 904	/* No mm means kernel memory flush. */
 905	if (!info->mm)
 906		return true;
 907
 908	/* The target mm is loaded, and the CPU is not lazy. */
 909	if (per_cpu(cpu_tlbstate.loaded_mm, cpu) == info->mm)
 910		return true;
 911
 912	/* In cpumask, but not the loaded mm? Periodically remove by flushing. */
 913	if (info->trim_cpumask)
 914		return true;
 915
 916	return false;
 917}
 918
 919static bool should_trim_cpumask(struct mm_struct *mm)
 920{
 921	if (time_after(jiffies, READ_ONCE(mm->context.next_trim_cpumask))) {
 922		WRITE_ONCE(mm->context.next_trim_cpumask, jiffies + HZ);
 923		return true;
 924	}
 925	return false;
 926}
 927
 928DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state_shared, cpu_tlbstate_shared);
 929EXPORT_PER_CPU_SYMBOL(cpu_tlbstate_shared);
 930
 931STATIC_NOPV void native_flush_tlb_multi(const struct cpumask *cpumask,
 932					 const struct flush_tlb_info *info)
 933{
 934	/*
 935	 * Do accounting and tracing. Note that there are (and have always been)
 936	 * cases in which a remote TLB flush will be traced, but eventually
 937	 * would not happen.
 938	 */
 939	count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 940	if (info->end == TLB_FLUSH_ALL)
 941		trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL);
 942	else
 943		trace_tlb_flush(TLB_REMOTE_SEND_IPI,
 944				(info->end - info->start) >> PAGE_SHIFT);
 945
 946	/*
 947	 * If no page tables were freed, we can skip sending IPIs to
 948	 * CPUs in lazy TLB mode. They will flush the CPU themselves
 949	 * at the next context switch.
 950	 *
 951	 * However, if page tables are getting freed, we need to send the
 952	 * IPI everywhere, to prevent CPUs in lazy TLB mode from tripping
 953	 * up on the new contents of what used to be page tables, while
 954	 * doing a speculative memory access.
 955	 */
 956	if (info->freed_tables)
 957		on_each_cpu_mask(cpumask, flush_tlb_func, (void *)info, true);
 958	else
 959		on_each_cpu_cond_mask(should_flush_tlb, flush_tlb_func,
 960				(void *)info, 1, cpumask);
 961}
 962
 963void flush_tlb_multi(const struct cpumask *cpumask,
 964		      const struct flush_tlb_info *info)
 965{
 966	__flush_tlb_multi(cpumask, info);
 967}
 968
 969/*
 970 * See Documentation/arch/x86/tlb.rst for details.  We choose 33
 971 * because it is large enough to cover the vast majority (at
 972 * least 95%) of allocations, and is small enough that we are
 973 * confident it will not cause too much overhead.  Each single
 974 * flush is about 100 ns, so this caps the maximum overhead at
 975 * _about_ 3,000 ns.
 976 *
 977 * This is in units of pages.
 978 */
 979unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
 980
 981static DEFINE_PER_CPU_SHARED_ALIGNED(struct flush_tlb_info, flush_tlb_info);
 982
 983#ifdef CONFIG_DEBUG_VM
 984static DEFINE_PER_CPU(unsigned int, flush_tlb_info_idx);
 985#endif
 986
 987static struct flush_tlb_info *get_flush_tlb_info(struct mm_struct *mm,
 988			unsigned long start, unsigned long end,
 989			unsigned int stride_shift, bool freed_tables,
 990			u64 new_tlb_gen)
 991{
 992	struct flush_tlb_info *info = this_cpu_ptr(&flush_tlb_info);
 993
 994#ifdef CONFIG_DEBUG_VM
 995	/*
 996	 * Ensure that the following code is non-reentrant and flush_tlb_info
 997	 * is not overwritten. This means no TLB flushing is initiated by
 998	 * interrupt handlers and machine-check exception handlers.
 999	 */
1000	BUG_ON(this_cpu_inc_return(flush_tlb_info_idx) != 1);
1001#endif
1002
1003	info->start		= start;
1004	info->end		= end;
1005	info->mm		= mm;
1006	info->stride_shift	= stride_shift;
1007	info->freed_tables	= freed_tables;
1008	info->new_tlb_gen	= new_tlb_gen;
1009	info->initiating_cpu	= smp_processor_id();
1010	info->trim_cpumask	= 0;
1011
1012	return info;
1013}
1014
1015static void put_flush_tlb_info(void)
1016{
1017#ifdef CONFIG_DEBUG_VM
1018	/* Complete reentrancy prevention checks */
1019	barrier();
1020	this_cpu_dec(flush_tlb_info_idx);
1021#endif
1022}
1023
1024void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
1025				unsigned long end, unsigned int stride_shift,
1026				bool freed_tables)
1027{
1028	struct flush_tlb_info *info;
1029	u64 new_tlb_gen;
1030	int cpu;
1031
1032	cpu = get_cpu();
1033
1034	/* Should we flush just the requested range? */
1035	if ((end == TLB_FLUSH_ALL) ||
1036	    ((end - start) >> stride_shift) > tlb_single_page_flush_ceiling) {
1037		start = 0;
1038		end = TLB_FLUSH_ALL;
1039	}
1040
1041	/* This is also a barrier that synchronizes with switch_mm(). */
1042	new_tlb_gen = inc_mm_tlb_gen(mm);
1043
1044	info = get_flush_tlb_info(mm, start, end, stride_shift, freed_tables,
1045				  new_tlb_gen);
1046
1047	/*
1048	 * flush_tlb_multi() is not optimized for the common case in which only
1049	 * a local TLB flush is needed. Optimize this use-case by calling
1050	 * flush_tlb_func_local() directly in this case.
1051	 */
1052	if (cpumask_any_but(mm_cpumask(mm), cpu) < nr_cpu_ids) {
1053		info->trim_cpumask = should_trim_cpumask(mm);
1054		flush_tlb_multi(mm_cpumask(mm), info);
1055	} else if (mm == this_cpu_read(cpu_tlbstate.loaded_mm)) {
1056		lockdep_assert_irqs_enabled();
1057		local_irq_disable();
1058		flush_tlb_func(info);
1059		local_irq_enable();
1060	}
1061
1062	put_flush_tlb_info();
1063	put_cpu();
1064	mmu_notifier_arch_invalidate_secondary_tlbs(mm, start, end);
1065}
1066
1067
1068static void do_flush_tlb_all(void *info)
1069{
1070	count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
1071	__flush_tlb_all();
1072}
1073
1074void flush_tlb_all(void)
1075{
1076	count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
1077	on_each_cpu(do_flush_tlb_all, NULL, 1);
1078}
1079
1080static void do_kernel_range_flush(void *info)
1081{
1082	struct flush_tlb_info *f = info;
1083	unsigned long addr;
1084
1085	/* flush range by one by one 'invlpg' */
1086	for (addr = f->start; addr < f->end; addr += PAGE_SIZE)
1087		flush_tlb_one_kernel(addr);
1088}
1089
1090void flush_tlb_kernel_range(unsigned long start, unsigned long end)
1091{
1092	/* Balance as user space task's flush, a bit conservative */
1093	if (end == TLB_FLUSH_ALL ||
1094	    (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) {
1095		on_each_cpu(do_flush_tlb_all, NULL, 1);
1096	} else {
1097		struct flush_tlb_info *info;
1098
1099		preempt_disable();
1100		info = get_flush_tlb_info(NULL, start, end, 0, false,
1101					  TLB_GENERATION_INVALID);
1102
1103		on_each_cpu(do_kernel_range_flush, info, 1);
1104
1105		put_flush_tlb_info();
1106		preempt_enable();
1107	}
1108}
1109
1110/*
1111 * This can be used from process context to figure out what the value of
1112 * CR3 is without needing to do a (slow) __read_cr3().
1113 *
1114 * It's intended to be used for code like KVM that sneakily changes CR3
1115 * and needs to restore it.  It needs to be used very carefully.
1116 */
1117unsigned long __get_current_cr3_fast(void)
1118{
1119	unsigned long cr3 =
1120		build_cr3(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd,
1121			  this_cpu_read(cpu_tlbstate.loaded_mm_asid),
1122			  tlbstate_lam_cr3_mask());
1123
1124	/* For now, be very restrictive about when this can be called. */
1125	VM_WARN_ON(in_nmi() || preemptible());
1126
1127	VM_BUG_ON(cr3 != __read_cr3());
1128	return cr3;
1129}
1130EXPORT_SYMBOL_GPL(__get_current_cr3_fast);
1131
1132/*
1133 * Flush one page in the kernel mapping
1134 */
1135void flush_tlb_one_kernel(unsigned long addr)
1136{
1137	count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
1138
1139	/*
1140	 * If PTI is off, then __flush_tlb_one_user() is just INVLPG or its
1141	 * paravirt equivalent.  Even with PCID, this is sufficient: we only
1142	 * use PCID if we also use global PTEs for the kernel mapping, and
1143	 * INVLPG flushes global translations across all address spaces.
1144	 *
1145	 * If PTI is on, then the kernel is mapped with non-global PTEs, and
1146	 * __flush_tlb_one_user() will flush the given address for the current
1147	 * kernel address space and for its usermode counterpart, but it does
1148	 * not flush it for other address spaces.
1149	 */
1150	flush_tlb_one_user(addr);
1151
1152	if (!static_cpu_has(X86_FEATURE_PTI))
1153		return;
1154
1155	/*
1156	 * See above.  We need to propagate the flush to all other address
1157	 * spaces.  In principle, we only need to propagate it to kernelmode
1158	 * address spaces, but the extra bookkeeping we would need is not
1159	 * worth it.
1160	 */
1161	this_cpu_write(cpu_tlbstate.invalidate_other, true);
1162}
1163
1164/*
1165 * Flush one page in the user mapping
1166 */
1167STATIC_NOPV void native_flush_tlb_one_user(unsigned long addr)
1168{
1169	u32 loaded_mm_asid;
1170	bool cpu_pcide;
1171
1172	/* Flush 'addr' from the kernel PCID: */
1173	invlpg(addr);
1174
1175	/* If PTI is off there is no user PCID and nothing to flush. */
1176	if (!static_cpu_has(X86_FEATURE_PTI))
1177		return;
1178
1179	loaded_mm_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
1180	cpu_pcide      = this_cpu_read(cpu_tlbstate.cr4) & X86_CR4_PCIDE;
1181
1182	/*
1183	 * invpcid_flush_one(pcid>0) will #GP if CR4.PCIDE==0.  Check
1184	 * 'cpu_pcide' to ensure that *this* CPU will not trigger those
1185	 * #GP's even if called before CR4.PCIDE has been initialized.
1186	 */
1187	if (boot_cpu_has(X86_FEATURE_INVPCID) && cpu_pcide)
 
 
1188		invpcid_flush_one(user_pcid(loaded_mm_asid), addr);
1189	else
1190		invalidate_user_asid(loaded_mm_asid);
1191}
1192
1193void flush_tlb_one_user(unsigned long addr)
1194{
1195	__flush_tlb_one_user(addr);
1196}
1197
1198/*
1199 * Flush everything
1200 */
1201STATIC_NOPV void native_flush_tlb_global(void)
1202{
1203	unsigned long flags;
1204
1205	if (static_cpu_has(X86_FEATURE_INVPCID)) {
1206		/*
1207		 * Using INVPCID is considerably faster than a pair of writes
1208		 * to CR4 sandwiched inside an IRQ flag save/restore.
1209		 *
1210		 * Note, this works with CR4.PCIDE=0 or 1.
1211		 */
1212		invpcid_flush_all();
1213		return;
1214	}
1215
1216	/*
1217	 * Read-modify-write to CR4 - protect it from preemption and
1218	 * from interrupts. (Use the raw variant because this code can
1219	 * be called from deep inside debugging code.)
1220	 */
1221	raw_local_irq_save(flags);
1222
1223	__native_tlb_flush_global(this_cpu_read(cpu_tlbstate.cr4));
1224
1225	raw_local_irq_restore(flags);
1226}
1227
1228/*
1229 * Flush the entire current user mapping
1230 */
1231STATIC_NOPV void native_flush_tlb_local(void)
1232{
1233	/*
1234	 * Preemption or interrupts must be disabled to protect the access
1235	 * to the per CPU variable and to prevent being preempted between
1236	 * read_cr3() and write_cr3().
1237	 */
1238	WARN_ON_ONCE(preemptible());
1239
1240	invalidate_user_asid(this_cpu_read(cpu_tlbstate.loaded_mm_asid));
1241
1242	/* If current->mm == NULL then the read_cr3() "borrows" an mm */
1243	native_write_cr3(__native_read_cr3());
1244}
1245
1246void flush_tlb_local(void)
1247{
1248	__flush_tlb_local();
1249}
1250
1251/*
1252 * Flush everything
1253 */
1254void __flush_tlb_all(void)
1255{
1256	/*
1257	 * This is to catch users with enabled preemption and the PGE feature
1258	 * and don't trigger the warning in __native_flush_tlb().
1259	 */
1260	VM_WARN_ON_ONCE(preemptible());
1261
1262	if (cpu_feature_enabled(X86_FEATURE_PGE)) {
1263		__flush_tlb_global();
1264	} else {
1265		/*
1266		 * !PGE -> !PCID (setup_pcid()), thus every flush is total.
1267		 */
1268		flush_tlb_local();
1269	}
1270}
1271EXPORT_SYMBOL_GPL(__flush_tlb_all);
1272
1273void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
1274{
1275	struct flush_tlb_info *info;
1276
1277	int cpu = get_cpu();
1278
1279	info = get_flush_tlb_info(NULL, 0, TLB_FLUSH_ALL, 0, false,
1280				  TLB_GENERATION_INVALID);
1281	/*
1282	 * flush_tlb_multi() is not optimized for the common case in which only
1283	 * a local TLB flush is needed. Optimize this use-case by calling
1284	 * flush_tlb_func_local() directly in this case.
1285	 */
1286	if (cpumask_any_but(&batch->cpumask, cpu) < nr_cpu_ids) {
1287		flush_tlb_multi(&batch->cpumask, info);
1288	} else if (cpumask_test_cpu(cpu, &batch->cpumask)) {
1289		lockdep_assert_irqs_enabled();
1290		local_irq_disable();
1291		flush_tlb_func(info);
1292		local_irq_enable();
1293	}
1294
1295	cpumask_clear(&batch->cpumask);
1296
1297	put_flush_tlb_info();
1298	put_cpu();
1299}
1300
1301/*
1302 * Blindly accessing user memory from NMI context can be dangerous
1303 * if we're in the middle of switching the current user task or
1304 * switching the loaded mm.  It can also be dangerous if we
1305 * interrupted some kernel code that was temporarily using a
1306 * different mm.
1307 */
1308bool nmi_uaccess_okay(void)
1309{
1310	struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
1311	struct mm_struct *current_mm = current->mm;
1312
1313	VM_WARN_ON_ONCE(!loaded_mm);
1314
1315	/*
1316	 * The condition we want to check is
1317	 * current_mm->pgd == __va(read_cr3_pa()).  This may be slow, though,
1318	 * if we're running in a VM with shadow paging, and nmi_uaccess_okay()
1319	 * is supposed to be reasonably fast.
1320	 *
1321	 * Instead, we check the almost equivalent but somewhat conservative
1322	 * condition below, and we rely on the fact that switch_mm_irqs_off()
1323	 * sets loaded_mm to LOADED_MM_SWITCHING before writing to CR3.
1324	 */
1325	if (loaded_mm != current_mm)
1326		return false;
1327
1328	VM_WARN_ON_ONCE(current_mm->pgd != __va(read_cr3_pa()));
1329
1330	return true;
1331}
1332
1333static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
1334			     size_t count, loff_t *ppos)
1335{
1336	char buf[32];
1337	unsigned int len;
1338
1339	len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
1340	return simple_read_from_buffer(user_buf, count, ppos, buf, len);
1341}
1342
1343static ssize_t tlbflush_write_file(struct file *file,
1344		 const char __user *user_buf, size_t count, loff_t *ppos)
1345{
1346	char buf[32];
1347	ssize_t len;
1348	int ceiling;
1349
1350	len = min(count, sizeof(buf) - 1);
1351	if (copy_from_user(buf, user_buf, len))
1352		return -EFAULT;
1353
1354	buf[len] = '\0';
1355	if (kstrtoint(buf, 0, &ceiling))
1356		return -EINVAL;
1357
1358	if (ceiling < 0)
1359		return -EINVAL;
1360
1361	tlb_single_page_flush_ceiling = ceiling;
1362	return count;
1363}
1364
1365static const struct file_operations fops_tlbflush = {
1366	.read = tlbflush_read_file,
1367	.write = tlbflush_write_file,
1368	.llseek = default_llseek,
1369};
1370
1371static int __init create_tlb_single_page_flush_ceiling(void)
1372{
1373	debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
1374			    arch_debugfs_dir, NULL, &fops_tlbflush);
1375	return 0;
1376}
1377late_initcall(create_tlb_single_page_flush_ceiling);