Linux Audio

Check our new training course

Loading...
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  Copyright (C) 1995  Linus Torvalds
   4 *
   5 *  Pentium III FXSR, SSE support
   6 *	Gareth Hughes <gareth@valinux.com>, May 2000
   7 *
   8 *  X86-64 port
   9 *	Andi Kleen.
  10 *
  11 *	CPU hotplug support - ashok.raj@intel.com
  12 */
  13
  14/*
  15 * This file handles the architecture-dependent parts of process handling..
  16 */
  17
  18#include <linux/cpu.h>
  19#include <linux/errno.h>
  20#include <linux/sched.h>
  21#include <linux/sched/task.h>
  22#include <linux/sched/task_stack.h>
  23#include <linux/fs.h>
  24#include <linux/kernel.h>
  25#include <linux/mm.h>
  26#include <linux/elfcore.h>
  27#include <linux/smp.h>
  28#include <linux/slab.h>
  29#include <linux/user.h>
  30#include <linux/interrupt.h>
  31#include <linux/delay.h>
  32#include <linux/export.h>
  33#include <linux/ptrace.h>
  34#include <linux/notifier.h>
  35#include <linux/kprobes.h>
  36#include <linux/kdebug.h>
  37#include <linux/prctl.h>
  38#include <linux/uaccess.h>
  39#include <linux/io.h>
  40#include <linux/ftrace.h>
  41#include <linux/syscalls.h>
  42#include <linux/iommu.h>
  43
  44#include <asm/processor.h>
  45#include <asm/pkru.h>
  46#include <asm/fpu/sched.h>
  47#include <asm/mmu_context.h>
  48#include <asm/prctl.h>
  49#include <asm/desc.h>
  50#include <asm/proto.h>
  51#include <asm/ia32.h>
  52#include <asm/debugreg.h>
  53#include <asm/switch_to.h>
  54#include <asm/xen/hypervisor.h>
  55#include <asm/vdso.h>
  56#include <asm/resctrl.h>
  57#include <asm/unistd.h>
  58#include <asm/fsgsbase.h>
  59#include <asm/fred.h>
  60#ifdef CONFIG_IA32_EMULATION
  61/* Not included via unistd.h */
  62#include <asm/unistd_32_ia32.h>
  63#endif
  64
  65#include "process.h"
  66
  67/* Prints also some state that isn't saved in the pt_regs */
  68void __show_regs(struct pt_regs *regs, enum show_regs_mode mode,
  69		 const char *log_lvl)
  70{
  71	unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
  72	unsigned long d0, d1, d2, d3, d6, d7;
  73	unsigned int fsindex, gsindex;
  74	unsigned int ds, es;
  75
  76	show_iret_regs(regs, log_lvl);
  77
  78	if (regs->orig_ax != -1)
  79		pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
  80	else
  81		pr_cont("\n");
  82
  83	printk("%sRAX: %016lx RBX: %016lx RCX: %016lx\n",
  84	       log_lvl, regs->ax, regs->bx, regs->cx);
  85	printk("%sRDX: %016lx RSI: %016lx RDI: %016lx\n",
  86	       log_lvl, regs->dx, regs->si, regs->di);
  87	printk("%sRBP: %016lx R08: %016lx R09: %016lx\n",
  88	       log_lvl, regs->bp, regs->r8, regs->r9);
  89	printk("%sR10: %016lx R11: %016lx R12: %016lx\n",
  90	       log_lvl, regs->r10, regs->r11, regs->r12);
  91	printk("%sR13: %016lx R14: %016lx R15: %016lx\n",
  92	       log_lvl, regs->r13, regs->r14, regs->r15);
  93
  94	if (mode == SHOW_REGS_SHORT)
  95		return;
  96
  97	if (mode == SHOW_REGS_USER) {
  98		rdmsrl(MSR_FS_BASE, fs);
  99		rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
 100		printk("%sFS:  %016lx GS:  %016lx\n",
 101		       log_lvl, fs, shadowgs);
 102		return;
 103	}
 104
 105	asm("movl %%ds,%0" : "=r" (ds));
 106	asm("movl %%es,%0" : "=r" (es));
 107	asm("movl %%fs,%0" : "=r" (fsindex));
 108	asm("movl %%gs,%0" : "=r" (gsindex));
 109
 110	rdmsrl(MSR_FS_BASE, fs);
 111	rdmsrl(MSR_GS_BASE, gs);
 112	rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
 113
 114	cr0 = read_cr0();
 115	cr2 = read_cr2();
 116	cr3 = __read_cr3();
 117	cr4 = __read_cr4();
 118
 119	printk("%sFS:  %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
 120	       log_lvl, fs, fsindex, gs, gsindex, shadowgs);
 121	printk("%sCS:  %04x DS: %04x ES: %04x CR0: %016lx\n",
 122		log_lvl, regs->cs, ds, es, cr0);
 123	printk("%sCR2: %016lx CR3: %016lx CR4: %016lx\n",
 124		log_lvl, cr2, cr3, cr4);
 125
 126	get_debugreg(d0, 0);
 127	get_debugreg(d1, 1);
 128	get_debugreg(d2, 2);
 129	get_debugreg(d3, 3);
 130	get_debugreg(d6, 6);
 131	get_debugreg(d7, 7);
 132
 133	/* Only print out debug registers if they are in their non-default state. */
 134	if (!((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
 135	    (d6 == DR6_RESERVED) && (d7 == 0x400))) {
 136		printk("%sDR0: %016lx DR1: %016lx DR2: %016lx\n",
 137		       log_lvl, d0, d1, d2);
 138		printk("%sDR3: %016lx DR6: %016lx DR7: %016lx\n",
 139		       log_lvl, d3, d6, d7);
 140	}
 141
 142	if (cr4 & X86_CR4_PKE)
 143		printk("%sPKRU: %08x\n", log_lvl, read_pkru());
 144}
 145
 146void release_thread(struct task_struct *dead_task)
 147{
 148	WARN_ON(dead_task->mm);
 149}
 150
 151enum which_selector {
 152	FS,
 153	GS
 154};
 155
 156/*
 157 * Out of line to be protected from kprobes and tracing. If this would be
 158 * traced or probed than any access to a per CPU variable happens with
 159 * the wrong GS.
 160 *
 161 * It is not used on Xen paravirt. When paravirt support is needed, it
 162 * needs to be renamed with native_ prefix.
 163 */
 164static noinstr unsigned long __rdgsbase_inactive(void)
 165{
 166	unsigned long gsbase;
 167
 168	lockdep_assert_irqs_disabled();
 169
 170	/*
 171	 * SWAPGS is no longer needed thus NOT allowed with FRED because
 172	 * FRED transitions ensure that an operating system can _always_
 173	 * operate with its own GS base address:
 174	 * - For events that occur in ring 3, FRED event delivery swaps
 175	 *   the GS base address with the IA32_KERNEL_GS_BASE MSR.
 176	 * - ERETU (the FRED transition that returns to ring 3) also swaps
 177	 *   the GS base address with the IA32_KERNEL_GS_BASE MSR.
 178	 *
 179	 * And the operating system can still setup the GS segment for a
 180	 * user thread without the need of loading a user thread GS with:
 181	 * - Using LKGS, available with FRED, to modify other attributes
 182	 *   of the GS segment without compromising its ability always to
 183	 *   operate with its own GS base address.
 184	 * - Accessing the GS segment base address for a user thread as
 185	 *   before using RDMSR or WRMSR on the IA32_KERNEL_GS_BASE MSR.
 186	 *
 187	 * Note, LKGS loads the GS base address into the IA32_KERNEL_GS_BASE
 188	 * MSR instead of the GS segment’s descriptor cache. As such, the
 189	 * operating system never changes its runtime GS base address.
 190	 */
 191	if (!cpu_feature_enabled(X86_FEATURE_FRED) &&
 192	    !cpu_feature_enabled(X86_FEATURE_XENPV)) {
 193		native_swapgs();
 194		gsbase = rdgsbase();
 195		native_swapgs();
 196	} else {
 197		instrumentation_begin();
 198		rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
 199		instrumentation_end();
 200	}
 201
 202	return gsbase;
 203}
 204
 205/*
 206 * Out of line to be protected from kprobes and tracing. If this would be
 207 * traced or probed than any access to a per CPU variable happens with
 208 * the wrong GS.
 209 *
 210 * It is not used on Xen paravirt. When paravirt support is needed, it
 211 * needs to be renamed with native_ prefix.
 212 */
 213static noinstr void __wrgsbase_inactive(unsigned long gsbase)
 214{
 215	lockdep_assert_irqs_disabled();
 216
 217	if (!cpu_feature_enabled(X86_FEATURE_FRED) &&
 218	    !cpu_feature_enabled(X86_FEATURE_XENPV)) {
 219		native_swapgs();
 220		wrgsbase(gsbase);
 221		native_swapgs();
 222	} else {
 223		instrumentation_begin();
 224		wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
 225		instrumentation_end();
 226	}
 227}
 228
 229/*
 230 * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
 231 * not available.  The goal is to be reasonably fast on non-FSGSBASE systems.
 232 * It's forcibly inlined because it'll generate better code and this function
 233 * is hot.
 234 */
 235static __always_inline void save_base_legacy(struct task_struct *prev_p,
 236					     unsigned short selector,
 237					     enum which_selector which)
 238{
 239	if (likely(selector == 0)) {
 240		/*
 241		 * On Intel (without X86_BUG_NULL_SEG), the segment base could
 242		 * be the pre-existing saved base or it could be zero.  On AMD
 243		 * (with X86_BUG_NULL_SEG), the segment base could be almost
 244		 * anything.
 245		 *
 246		 * This branch is very hot (it's hit twice on almost every
 247		 * context switch between 64-bit programs), and avoiding
 248		 * the RDMSR helps a lot, so we just assume that whatever
 249		 * value is already saved is correct.  This matches historical
 250		 * Linux behavior, so it won't break existing applications.
 251		 *
 252		 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
 253		 * report that the base is zero, it needs to actually be zero:
 254		 * see the corresponding logic in load_seg_legacy.
 255		 */
 256	} else {
 257		/*
 258		 * If the selector is 1, 2, or 3, then the base is zero on
 259		 * !X86_BUG_NULL_SEG CPUs and could be anything on
 260		 * X86_BUG_NULL_SEG CPUs.  In the latter case, Linux
 261		 * has never attempted to preserve the base across context
 262		 * switches.
 263		 *
 264		 * If selector > 3, then it refers to a real segment, and
 265		 * saving the base isn't necessary.
 266		 */
 267		if (which == FS)
 268			prev_p->thread.fsbase = 0;
 269		else
 270			prev_p->thread.gsbase = 0;
 271	}
 272}
 273
 274static __always_inline void save_fsgs(struct task_struct *task)
 275{
 276	savesegment(fs, task->thread.fsindex);
 277	savesegment(gs, task->thread.gsindex);
 278	if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
 279		/*
 280		 * If FSGSBASE is enabled, we can't make any useful guesses
 281		 * about the base, and user code expects us to save the current
 282		 * value.  Fortunately, reading the base directly is efficient.
 283		 */
 284		task->thread.fsbase = rdfsbase();
 285		task->thread.gsbase = __rdgsbase_inactive();
 286	} else {
 287		save_base_legacy(task, task->thread.fsindex, FS);
 288		save_base_legacy(task, task->thread.gsindex, GS);
 289	}
 290}
 291
 292/*
 293 * While a process is running,current->thread.fsbase and current->thread.gsbase
 294 * may not match the corresponding CPU registers (see save_base_legacy()).
 295 */
 296void current_save_fsgs(void)
 297{
 298	unsigned long flags;
 299
 300	/* Interrupts need to be off for FSGSBASE */
 301	local_irq_save(flags);
 302	save_fsgs(current);
 303	local_irq_restore(flags);
 304}
 305#if IS_ENABLED(CONFIG_KVM)
 306EXPORT_SYMBOL_GPL(current_save_fsgs);
 307#endif
 308
 309static __always_inline void loadseg(enum which_selector which,
 310				    unsigned short sel)
 311{
 312	if (which == FS)
 313		loadsegment(fs, sel);
 314	else
 315		load_gs_index(sel);
 316}
 317
 318static __always_inline void load_seg_legacy(unsigned short prev_index,
 319					    unsigned long prev_base,
 320					    unsigned short next_index,
 321					    unsigned long next_base,
 322					    enum which_selector which)
 323{
 324	if (likely(next_index <= 3)) {
 325		/*
 326		 * The next task is using 64-bit TLS, is not using this
 327		 * segment at all, or is having fun with arcane CPU features.
 328		 */
 329		if (next_base == 0) {
 330			/*
 331			 * Nasty case: on AMD CPUs, we need to forcibly zero
 332			 * the base.
 333			 */
 334			if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
 335				loadseg(which, __USER_DS);
 336				loadseg(which, next_index);
 337			} else {
 338				/*
 339				 * We could try to exhaustively detect cases
 340				 * under which we can skip the segment load,
 341				 * but there's really only one case that matters
 342				 * for performance: if both the previous and
 343				 * next states are fully zeroed, we can skip
 344				 * the load.
 345				 *
 346				 * (This assumes that prev_base == 0 has no
 347				 * false positives.  This is the case on
 348				 * Intel-style CPUs.)
 349				 */
 350				if (likely(prev_index | next_index | prev_base))
 351					loadseg(which, next_index);
 352			}
 353		} else {
 354			if (prev_index != next_index)
 355				loadseg(which, next_index);
 356			wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
 357			       next_base);
 358		}
 359	} else {
 360		/*
 361		 * The next task is using a real segment.  Loading the selector
 362		 * is sufficient.
 363		 */
 364		loadseg(which, next_index);
 365	}
 366}
 367
 368/*
 369 * Store prev's PKRU value and load next's PKRU value if they differ. PKRU
 370 * is not XSTATE managed on context switch because that would require a
 371 * lookup in the task's FPU xsave buffer and require to keep that updated
 372 * in various places.
 373 */
 374static __always_inline void x86_pkru_load(struct thread_struct *prev,
 375					  struct thread_struct *next)
 376{
 377	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
 378		return;
 379
 380	/* Stash the prev task's value: */
 381	prev->pkru = rdpkru();
 382
 383	/*
 384	 * PKRU writes are slightly expensive.  Avoid them when not
 385	 * strictly necessary:
 386	 */
 387	if (prev->pkru != next->pkru)
 388		wrpkru(next->pkru);
 389}
 390
 391static __always_inline void x86_fsgsbase_load(struct thread_struct *prev,
 392					      struct thread_struct *next)
 393{
 394	if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
 395		/* Update the FS and GS selectors if they could have changed. */
 396		if (unlikely(prev->fsindex || next->fsindex))
 397			loadseg(FS, next->fsindex);
 398		if (unlikely(prev->gsindex || next->gsindex))
 399			loadseg(GS, next->gsindex);
 400
 401		/* Update the bases. */
 402		wrfsbase(next->fsbase);
 403		__wrgsbase_inactive(next->gsbase);
 404	} else {
 405		load_seg_legacy(prev->fsindex, prev->fsbase,
 406				next->fsindex, next->fsbase, FS);
 407		load_seg_legacy(prev->gsindex, prev->gsbase,
 408				next->gsindex, next->gsbase, GS);
 409	}
 410}
 411
 412unsigned long x86_fsgsbase_read_task(struct task_struct *task,
 413				     unsigned short selector)
 414{
 415	unsigned short idx = selector >> 3;
 416	unsigned long base;
 417
 418	if (likely((selector & SEGMENT_TI_MASK) == 0)) {
 419		if (unlikely(idx >= GDT_ENTRIES))
 420			return 0;
 421
 422		/*
 423		 * There are no user segments in the GDT with nonzero bases
 424		 * other than the TLS segments.
 425		 */
 426		if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
 427			return 0;
 428
 429		idx -= GDT_ENTRY_TLS_MIN;
 430		base = get_desc_base(&task->thread.tls_array[idx]);
 431	} else {
 432#ifdef CONFIG_MODIFY_LDT_SYSCALL
 433		struct ldt_struct *ldt;
 434
 435		/*
 436		 * If performance here mattered, we could protect the LDT
 437		 * with RCU.  This is a slow path, though, so we can just
 438		 * take the mutex.
 439		 */
 440		mutex_lock(&task->mm->context.lock);
 441		ldt = task->mm->context.ldt;
 442		if (unlikely(!ldt || idx >= ldt->nr_entries))
 443			base = 0;
 444		else
 445			base = get_desc_base(ldt->entries + idx);
 446		mutex_unlock(&task->mm->context.lock);
 447#else
 448		base = 0;
 449#endif
 450	}
 451
 452	return base;
 453}
 454
 455unsigned long x86_gsbase_read_cpu_inactive(void)
 456{
 457	unsigned long gsbase;
 458
 459	if (boot_cpu_has(X86_FEATURE_FSGSBASE)) {
 460		unsigned long flags;
 461
 462		local_irq_save(flags);
 463		gsbase = __rdgsbase_inactive();
 464		local_irq_restore(flags);
 465	} else {
 466		rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
 467	}
 468
 469	return gsbase;
 470}
 471
 472void x86_gsbase_write_cpu_inactive(unsigned long gsbase)
 473{
 474	if (boot_cpu_has(X86_FEATURE_FSGSBASE)) {
 475		unsigned long flags;
 476
 477		local_irq_save(flags);
 478		__wrgsbase_inactive(gsbase);
 479		local_irq_restore(flags);
 480	} else {
 481		wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
 482	}
 483}
 484
 485unsigned long x86_fsbase_read_task(struct task_struct *task)
 486{
 487	unsigned long fsbase;
 488
 489	if (task == current)
 490		fsbase = x86_fsbase_read_cpu();
 491	else if (boot_cpu_has(X86_FEATURE_FSGSBASE) ||
 492		 (task->thread.fsindex == 0))
 493		fsbase = task->thread.fsbase;
 494	else
 495		fsbase = x86_fsgsbase_read_task(task, task->thread.fsindex);
 496
 497	return fsbase;
 498}
 499
 500unsigned long x86_gsbase_read_task(struct task_struct *task)
 501{
 502	unsigned long gsbase;
 503
 504	if (task == current)
 505		gsbase = x86_gsbase_read_cpu_inactive();
 506	else if (boot_cpu_has(X86_FEATURE_FSGSBASE) ||
 507		 (task->thread.gsindex == 0))
 508		gsbase = task->thread.gsbase;
 509	else
 510		gsbase = x86_fsgsbase_read_task(task, task->thread.gsindex);
 511
 512	return gsbase;
 513}
 514
 515void x86_fsbase_write_task(struct task_struct *task, unsigned long fsbase)
 516{
 517	WARN_ON_ONCE(task == current);
 518
 519	task->thread.fsbase = fsbase;
 520}
 521
 522void x86_gsbase_write_task(struct task_struct *task, unsigned long gsbase)
 523{
 524	WARN_ON_ONCE(task == current);
 525
 526	task->thread.gsbase = gsbase;
 527}
 528
 529static void
 530start_thread_common(struct pt_regs *regs, unsigned long new_ip,
 531		    unsigned long new_sp,
 532		    u16 _cs, u16 _ss, u16 _ds)
 533{
 534	WARN_ON_ONCE(regs != current_pt_regs());
 535
 536	if (static_cpu_has(X86_BUG_NULL_SEG)) {
 537		/* Loading zero below won't clear the base. */
 538		loadsegment(fs, __USER_DS);
 539		load_gs_index(__USER_DS);
 540	}
 541
 542	reset_thread_features();
 543
 544	loadsegment(fs, 0);
 545	loadsegment(es, _ds);
 546	loadsegment(ds, _ds);
 547	load_gs_index(0);
 548
 549	regs->ip	= new_ip;
 550	regs->sp	= new_sp;
 551	regs->csx	= _cs;
 552	regs->ssx	= _ss;
 553	/*
 554	 * Allow single-step trap and NMI when starting a new task, thus
 555	 * once the new task enters user space, single-step trap and NMI
 556	 * are both enabled immediately.
 557	 *
 558	 * Entering a new task is logically speaking a return from a
 559	 * system call (exec, fork, clone, etc.). As such, if ptrace
 560	 * enables single stepping a single step exception should be
 561	 * allowed to trigger immediately upon entering user space.
 562	 * This is not optional.
 563	 *
 564	 * NMI should *never* be disabled in user space. As such, this
 565	 * is an optional, opportunistic way to catch errors.
 566	 *
 567	 * Paranoia: High-order 48 bits above the lowest 16 bit SS are
 568	 * discarded by the legacy IRET instruction on all Intel, AMD,
 569	 * and Cyrix/Centaur/VIA CPUs, thus can be set unconditionally,
 570	 * even when FRED is not enabled. But we choose the safer side
 571	 * to use these bits only when FRED is enabled.
 572	 */
 573	if (cpu_feature_enabled(X86_FEATURE_FRED)) {
 574		regs->fred_ss.swevent	= true;
 575		regs->fred_ss.nmi	= true;
 576	}
 577
 578	regs->flags	= X86_EFLAGS_IF | X86_EFLAGS_FIXED;
 579}
 580
 581void
 582start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
 583{
 584	start_thread_common(regs, new_ip, new_sp,
 585			    __USER_CS, __USER_DS, 0);
 586}
 587EXPORT_SYMBOL_GPL(start_thread);
 588
 589#ifdef CONFIG_COMPAT
 590void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp, bool x32)
 591{
 592	start_thread_common(regs, new_ip, new_sp,
 593			    x32 ? __USER_CS : __USER32_CS,
 
 594			    __USER_DS, __USER_DS);
 595}
 596#endif
 597
 598/*
 599 *	switch_to(x,y) should switch tasks from x to y.
 600 *
 601 * This could still be optimized:
 602 * - fold all the options into a flag word and test it with a single test.
 603 * - could test fs/gs bitsliced
 604 *
 605 * Kprobes not supported here. Set the probe on schedule instead.
 606 * Function graph tracer not supported too.
 607 */
 608__no_kmsan_checks
 609__visible __notrace_funcgraph struct task_struct *
 610__switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 611{
 612	struct thread_struct *prev = &prev_p->thread;
 613	struct thread_struct *next = &next_p->thread;
 
 
 614	int cpu = smp_processor_id();
 615
 616	WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
 617		     this_cpu_read(pcpu_hot.hardirq_stack_inuse));
 618
 619	if (!test_tsk_thread_flag(prev_p, TIF_NEED_FPU_LOAD))
 620		switch_fpu_prepare(prev_p, cpu);
 621
 622	/* We must save %fs and %gs before load_TLS() because
 623	 * %fs and %gs may be cleared by load_TLS().
 624	 *
 625	 * (e.g. xen_load_tls())
 626	 */
 627	save_fsgs(prev_p);
 628
 629	/*
 630	 * Load TLS before restoring any segments so that segment loads
 631	 * reference the correct GDT entries.
 632	 */
 633	load_TLS(next, cpu);
 634
 635	/*
 636	 * Leave lazy mode, flushing any hypercalls made here.  This
 637	 * must be done after loading TLS entries in the GDT but before
 638	 * loading segments that might reference them.
 639	 */
 640	arch_end_context_switch(next_p);
 641
 642	/* Switch DS and ES.
 643	 *
 644	 * Reading them only returns the selectors, but writing them (if
 645	 * nonzero) loads the full descriptor from the GDT or LDT.  The
 646	 * LDT for next is loaded in switch_mm, and the GDT is loaded
 647	 * above.
 648	 *
 649	 * We therefore need to write new values to the segment
 650	 * registers on every context switch unless both the new and old
 651	 * values are zero.
 652	 *
 653	 * Note that we don't need to do anything for CS and SS, as
 654	 * those are saved and restored as part of pt_regs.
 655	 */
 656	savesegment(es, prev->es);
 657	if (unlikely(next->es | prev->es))
 658		loadsegment(es, next->es);
 659
 660	savesegment(ds, prev->ds);
 661	if (unlikely(next->ds | prev->ds))
 662		loadsegment(ds, next->ds);
 663
 664	x86_fsgsbase_load(prev, next);
 665
 666	x86_pkru_load(prev, next);
 667
 668	/*
 669	 * Switch the PDA and FPU contexts.
 670	 */
 671	raw_cpu_write(pcpu_hot.current_task, next_p);
 672	raw_cpu_write(pcpu_hot.top_of_stack, task_top_of_stack(next_p));
 673
 674	switch_fpu_finish(next_p);
 675
 676	/* Reload sp0. */
 677	update_task_stack(next_p);
 678
 679	switch_to_extra(prev_p, next_p);
 680
 681	if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
 682		/*
 683		 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
 684		 * does not update the cached descriptor.  As a result, if we
 685		 * do SYSRET while SS is NULL, we'll end up in user mode with
 686		 * SS apparently equal to __USER_DS but actually unusable.
 687		 *
 688		 * The straightforward workaround would be to fix it up just
 689		 * before SYSRET, but that would slow down the system call
 690		 * fast paths.  Instead, we ensure that SS is never NULL in
 691		 * system call context.  We do this by replacing NULL SS
 692		 * selectors at every context switch.  SYSCALL sets up a valid
 693		 * SS, so the only way to get NULL is to re-enter the kernel
 694		 * from CPL 3 through an interrupt.  Since that can't happen
 695		 * in the same task as a running syscall, we are guaranteed to
 696		 * context switch between every interrupt vector entry and a
 697		 * subsequent SYSRET.
 698		 *
 699		 * We read SS first because SS reads are much faster than
 700		 * writes.  Out of caution, we force SS to __KERNEL_DS even if
 701		 * it previously had a different non-NULL value.
 702		 */
 703		unsigned short ss_sel;
 704		savesegment(ss, ss_sel);
 705		if (ss_sel != __KERNEL_DS)
 706			loadsegment(ss, __KERNEL_DS);
 707	}
 708
 709	/* Load the Intel cache allocation PQR MSR. */
 710	resctrl_sched_in(next_p);
 711
 712	return prev_p;
 713}
 714
 715void set_personality_64bit(void)
 716{
 717	/* inherit personality from parent */
 718
 719	/* Make sure to be in 64bit mode */
 
 720	clear_thread_flag(TIF_ADDR32);
 
 721	/* Pretend that this comes from a 64bit execve */
 722	task_pt_regs(current)->orig_ax = __NR_execve;
 723	current_thread_info()->status &= ~TS_COMPAT;
 
 
 724	if (current->mm)
 725		__set_bit(MM_CONTEXT_HAS_VSYSCALL, &current->mm->context.flags);
 726
 727	/* TBD: overwrites user setup. Should have two bits.
 728	   But 64bit processes have always behaved this way,
 729	   so it's not too bad. The main problem is just that
 730	   32bit children are affected again. */
 731	current->personality &= ~READ_IMPLIES_EXEC;
 732}
 733
 734static void __set_personality_x32(void)
 735{
 736#ifdef CONFIG_X86_X32_ABI
 
 
 737	if (current->mm)
 738		current->mm->context.flags = 0;
 739
 740	current->personality &= ~READ_IMPLIES_EXEC;
 741	/*
 742	 * in_32bit_syscall() uses the presence of the x32 syscall bit
 743	 * flag to determine compat status.  The x86 mmap() code relies on
 744	 * the syscall bitness so set x32 syscall bit right here to make
 745	 * in_32bit_syscall() work during exec().
 746	 *
 747	 * Pretend to come from a x32 execve.
 748	 */
 749	task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
 750	current_thread_info()->status &= ~TS_COMPAT;
 751#endif
 752}
 753
 754static void __set_personality_ia32(void)
 755{
 756#ifdef CONFIG_IA32_EMULATION
 757	if (current->mm) {
 758		/*
 759		 * uprobes applied to this MM need to know this and
 760		 * cannot use user_64bit_mode() at that time.
 761		 */
 762		__set_bit(MM_CONTEXT_UPROBE_IA32, &current->mm->context.flags);
 763	}
 764
 765	current->personality |= force_personality32;
 766	/* Prepare the first "return" to user space */
 767	task_pt_regs(current)->orig_ax = __NR_ia32_execve;
 768	current_thread_info()->status |= TS_COMPAT;
 769#endif
 770}
 771
 772void set_personality_ia32(bool x32)
 773{
 774	/* Make sure to be in 32bit mode */
 775	set_thread_flag(TIF_ADDR32);
 776
 777	if (x32)
 778		__set_personality_x32();
 779	else
 780		__set_personality_ia32();
 781}
 782EXPORT_SYMBOL_GPL(set_personality_ia32);
 783
 784#ifdef CONFIG_CHECKPOINT_RESTORE
 785static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
 786{
 787	int ret;
 788
 789	ret = map_vdso_once(image, addr);
 790	if (ret)
 791		return ret;
 792
 793	return (long)image->size;
 794}
 795#endif
 796
 797#ifdef CONFIG_ADDRESS_MASKING
 798
 799#define LAM_U57_BITS 6
 800
 801static void enable_lam_func(void *__mm)
 802{
 803	struct mm_struct *mm = __mm;
 804	unsigned long lam;
 805
 806	if (this_cpu_read(cpu_tlbstate.loaded_mm) == mm) {
 807		lam = mm_lam_cr3_mask(mm);
 808		write_cr3(__read_cr3() | lam);
 809		cpu_tlbstate_update_lam(lam, mm_untag_mask(mm));
 810	}
 811}
 812
 813static void mm_enable_lam(struct mm_struct *mm)
 814{
 815	mm->context.lam_cr3_mask = X86_CR3_LAM_U57;
 816	mm->context.untag_mask =  ~GENMASK(62, 57);
 817
 818	/*
 819	 * Even though the process must still be single-threaded at this
 820	 * point, kernel threads may be using the mm.  IPI those kernel
 821	 * threads if they exist.
 822	 */
 823	on_each_cpu_mask(mm_cpumask(mm), enable_lam_func, mm, true);
 824	set_bit(MM_CONTEXT_LOCK_LAM, &mm->context.flags);
 825}
 826
 827static int prctl_enable_tagged_addr(struct mm_struct *mm, unsigned long nr_bits)
 828{
 829	if (!cpu_feature_enabled(X86_FEATURE_LAM))
 830		return -ENODEV;
 831
 832	/* PTRACE_ARCH_PRCTL */
 833	if (current->mm != mm)
 834		return -EINVAL;
 835
 836	if (mm_valid_pasid(mm) &&
 837	    !test_bit(MM_CONTEXT_FORCE_TAGGED_SVA, &mm->context.flags))
 838		return -EINVAL;
 839
 840	if (mmap_write_lock_killable(mm))
 841		return -EINTR;
 842
 843	/*
 844	 * MM_CONTEXT_LOCK_LAM is set on clone.  Prevent LAM from
 845	 * being enabled unless the process is single threaded:
 846	 */
 847	if (test_bit(MM_CONTEXT_LOCK_LAM, &mm->context.flags)) {
 848		mmap_write_unlock(mm);
 849		return -EBUSY;
 850	}
 851
 852	if (!nr_bits || nr_bits > LAM_U57_BITS) {
 853		mmap_write_unlock(mm);
 854		return -EINVAL;
 855	}
 856
 857	mm_enable_lam(mm);
 858
 859	mmap_write_unlock(mm);
 860
 861	return 0;
 862}
 863#endif
 864
 865long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
 866{
 867	int ret = 0;
 868
 869	switch (option) {
 870	case ARCH_SET_GS: {
 871		if (unlikely(arg2 >= TASK_SIZE_MAX))
 872			return -EPERM;
 873
 874		preempt_disable();
 875		/*
 876		 * ARCH_SET_GS has always overwritten the index
 877		 * and the base. Zero is the most sensible value
 878		 * to put in the index, and is the only value that
 879		 * makes any sense if FSGSBASE is unavailable.
 880		 */
 881		if (task == current) {
 882			loadseg(GS, 0);
 883			x86_gsbase_write_cpu_inactive(arg2);
 884
 885			/*
 886			 * On non-FSGSBASE systems, save_base_legacy() expects
 887			 * that we also fill in thread.gsbase.
 888			 */
 889			task->thread.gsbase = arg2;
 890
 891		} else {
 892			task->thread.gsindex = 0;
 893			x86_gsbase_write_task(task, arg2);
 894		}
 895		preempt_enable();
 896		break;
 897	}
 898	case ARCH_SET_FS: {
 899		/*
 900		 * Not strictly needed for %fs, but do it for symmetry
 901		 * with %gs
 902		 */
 903		if (unlikely(arg2 >= TASK_SIZE_MAX))
 904			return -EPERM;
 905
 906		preempt_disable();
 907		/*
 908		 * Set the selector to 0 for the same reason
 909		 * as %gs above.
 910		 */
 911		if (task == current) {
 912			loadseg(FS, 0);
 913			x86_fsbase_write_cpu(arg2);
 914
 915			/*
 916			 * On non-FSGSBASE systems, save_base_legacy() expects
 917			 * that we also fill in thread.fsbase.
 918			 */
 919			task->thread.fsbase = arg2;
 920		} else {
 921			task->thread.fsindex = 0;
 922			x86_fsbase_write_task(task, arg2);
 923		}
 924		preempt_enable();
 925		break;
 926	}
 927	case ARCH_GET_FS: {
 928		unsigned long base = x86_fsbase_read_task(task);
 929
 930		ret = put_user(base, (unsigned long __user *)arg2);
 931		break;
 932	}
 933	case ARCH_GET_GS: {
 934		unsigned long base = x86_gsbase_read_task(task);
 935
 936		ret = put_user(base, (unsigned long __user *)arg2);
 937		break;
 938	}
 939
 940#ifdef CONFIG_CHECKPOINT_RESTORE
 941# ifdef CONFIG_X86_X32_ABI
 942	case ARCH_MAP_VDSO_X32:
 943		return prctl_map_vdso(&vdso_image_x32, arg2);
 944# endif
 945# if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
 946	case ARCH_MAP_VDSO_32:
 947		return prctl_map_vdso(&vdso_image_32, arg2);
 948# endif
 949	case ARCH_MAP_VDSO_64:
 950		return prctl_map_vdso(&vdso_image_64, arg2);
 951#endif
 952#ifdef CONFIG_ADDRESS_MASKING
 953	case ARCH_GET_UNTAG_MASK:
 954		return put_user(task->mm->context.untag_mask,
 955				(unsigned long __user *)arg2);
 956	case ARCH_ENABLE_TAGGED_ADDR:
 957		return prctl_enable_tagged_addr(task->mm, arg2);
 958	case ARCH_FORCE_TAGGED_SVA:
 959		if (current != task)
 960			return -EINVAL;
 961		set_bit(MM_CONTEXT_FORCE_TAGGED_SVA, &task->mm->context.flags);
 962		return 0;
 963	case ARCH_GET_MAX_TAG_BITS:
 964		if (!cpu_feature_enabled(X86_FEATURE_LAM))
 965			return put_user(0, (unsigned long __user *)arg2);
 966		else
 967			return put_user(LAM_U57_BITS, (unsigned long __user *)arg2);
 968#endif
 969	case ARCH_SHSTK_ENABLE:
 970	case ARCH_SHSTK_DISABLE:
 971	case ARCH_SHSTK_LOCK:
 972	case ARCH_SHSTK_UNLOCK:
 973	case ARCH_SHSTK_STATUS:
 974		return shstk_prctl(task, option, arg2);
 975	default:
 976		ret = -EINVAL;
 977		break;
 978	}
 979
 980	return ret;
 981}
 982
 983SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
 984{
 985	long ret;
 986
 987	ret = do_arch_prctl_64(current, option, arg2);
 988	if (ret == -EINVAL)
 989		ret = do_arch_prctl_common(option, arg2);
 990
 991	return ret;
 992}
 993
 994#ifdef CONFIG_IA32_EMULATION
 995COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
 996{
 997	return do_arch_prctl_common(option, arg2);
 998}
 999#endif
1000
1001unsigned long KSTK_ESP(struct task_struct *task)
1002{
1003	return task_pt_regs(task)->sp;
1004}
v5.9
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 *  Copyright (C) 1995  Linus Torvalds
  4 *
  5 *  Pentium III FXSR, SSE support
  6 *	Gareth Hughes <gareth@valinux.com>, May 2000
  7 *
  8 *  X86-64 port
  9 *	Andi Kleen.
 10 *
 11 *	CPU hotplug support - ashok.raj@intel.com
 12 */
 13
 14/*
 15 * This file handles the architecture-dependent parts of process handling..
 16 */
 17
 18#include <linux/cpu.h>
 19#include <linux/errno.h>
 20#include <linux/sched.h>
 21#include <linux/sched/task.h>
 22#include <linux/sched/task_stack.h>
 23#include <linux/fs.h>
 24#include <linux/kernel.h>
 25#include <linux/mm.h>
 26#include <linux/elfcore.h>
 27#include <linux/smp.h>
 28#include <linux/slab.h>
 29#include <linux/user.h>
 30#include <linux/interrupt.h>
 31#include <linux/delay.h>
 32#include <linux/export.h>
 33#include <linux/ptrace.h>
 34#include <linux/notifier.h>
 35#include <linux/kprobes.h>
 36#include <linux/kdebug.h>
 37#include <linux/prctl.h>
 38#include <linux/uaccess.h>
 39#include <linux/io.h>
 40#include <linux/ftrace.h>
 41#include <linux/syscalls.h>
 
 42
 43#include <asm/processor.h>
 44#include <asm/fpu/internal.h>
 
 45#include <asm/mmu_context.h>
 46#include <asm/prctl.h>
 47#include <asm/desc.h>
 48#include <asm/proto.h>
 49#include <asm/ia32.h>
 50#include <asm/debugreg.h>
 51#include <asm/switch_to.h>
 52#include <asm/xen/hypervisor.h>
 53#include <asm/vdso.h>
 54#include <asm/resctrl.h>
 55#include <asm/unistd.h>
 56#include <asm/fsgsbase.h>
 
 57#ifdef CONFIG_IA32_EMULATION
 58/* Not included via unistd.h */
 59#include <asm/unistd_32_ia32.h>
 60#endif
 61
 62#include "process.h"
 63
 64/* Prints also some state that isn't saved in the pt_regs */
 65void __show_regs(struct pt_regs *regs, enum show_regs_mode mode,
 66		 const char *log_lvl)
 67{
 68	unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
 69	unsigned long d0, d1, d2, d3, d6, d7;
 70	unsigned int fsindex, gsindex;
 71	unsigned int ds, es;
 72
 73	show_iret_regs(regs, log_lvl);
 74
 75	if (regs->orig_ax != -1)
 76		pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
 77	else
 78		pr_cont("\n");
 79
 80	printk("%sRAX: %016lx RBX: %016lx RCX: %016lx\n",
 81	       log_lvl, regs->ax, regs->bx, regs->cx);
 82	printk("%sRDX: %016lx RSI: %016lx RDI: %016lx\n",
 83	       log_lvl, regs->dx, regs->si, regs->di);
 84	printk("%sRBP: %016lx R08: %016lx R09: %016lx\n",
 85	       log_lvl, regs->bp, regs->r8, regs->r9);
 86	printk("%sR10: %016lx R11: %016lx R12: %016lx\n",
 87	       log_lvl, regs->r10, regs->r11, regs->r12);
 88	printk("%sR13: %016lx R14: %016lx R15: %016lx\n",
 89	       log_lvl, regs->r13, regs->r14, regs->r15);
 90
 91	if (mode == SHOW_REGS_SHORT)
 92		return;
 93
 94	if (mode == SHOW_REGS_USER) {
 95		rdmsrl(MSR_FS_BASE, fs);
 96		rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
 97		printk("%sFS:  %016lx GS:  %016lx\n",
 98		       log_lvl, fs, shadowgs);
 99		return;
100	}
101
102	asm("movl %%ds,%0" : "=r" (ds));
103	asm("movl %%es,%0" : "=r" (es));
104	asm("movl %%fs,%0" : "=r" (fsindex));
105	asm("movl %%gs,%0" : "=r" (gsindex));
106
107	rdmsrl(MSR_FS_BASE, fs);
108	rdmsrl(MSR_GS_BASE, gs);
109	rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
110
111	cr0 = read_cr0();
112	cr2 = read_cr2();
113	cr3 = __read_cr3();
114	cr4 = __read_cr4();
115
116	printk("%sFS:  %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
117	       log_lvl, fs, fsindex, gs, gsindex, shadowgs);
118	printk("%sCS:  %04lx DS: %04x ES: %04x CR0: %016lx\n",
119		log_lvl, regs->cs, ds, es, cr0);
120	printk("%sCR2: %016lx CR3: %016lx CR4: %016lx\n",
121		log_lvl, cr2, cr3, cr4);
122
123	get_debugreg(d0, 0);
124	get_debugreg(d1, 1);
125	get_debugreg(d2, 2);
126	get_debugreg(d3, 3);
127	get_debugreg(d6, 6);
128	get_debugreg(d7, 7);
129
130	/* Only print out debug registers if they are in their non-default state. */
131	if (!((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
132	    (d6 == DR6_RESERVED) && (d7 == 0x400))) {
133		printk("%sDR0: %016lx DR1: %016lx DR2: %016lx\n",
134		       log_lvl, d0, d1, d2);
135		printk("%sDR3: %016lx DR6: %016lx DR7: %016lx\n",
136		       log_lvl, d3, d6, d7);
137	}
138
139	if (boot_cpu_has(X86_FEATURE_OSPKE))
140		printk("%sPKRU: %08x\n", log_lvl, read_pkru());
141}
142
143void release_thread(struct task_struct *dead_task)
144{
145	WARN_ON(dead_task->mm);
146}
147
148enum which_selector {
149	FS,
150	GS
151};
152
153/*
154 * Out of line to be protected from kprobes and tracing. If this would be
155 * traced or probed than any access to a per CPU variable happens with
156 * the wrong GS.
157 *
158 * It is not used on Xen paravirt. When paravirt support is needed, it
159 * needs to be renamed with native_ prefix.
160 */
161static noinstr unsigned long __rdgsbase_inactive(void)
162{
163	unsigned long gsbase;
164
165	lockdep_assert_irqs_disabled();
166
167	if (!static_cpu_has(X86_FEATURE_XENPV)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
168		native_swapgs();
169		gsbase = rdgsbase();
170		native_swapgs();
171	} else {
172		instrumentation_begin();
173		rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
174		instrumentation_end();
175	}
176
177	return gsbase;
178}
179
180/*
181 * Out of line to be protected from kprobes and tracing. If this would be
182 * traced or probed than any access to a per CPU variable happens with
183 * the wrong GS.
184 *
185 * It is not used on Xen paravirt. When paravirt support is needed, it
186 * needs to be renamed with native_ prefix.
187 */
188static noinstr void __wrgsbase_inactive(unsigned long gsbase)
189{
190	lockdep_assert_irqs_disabled();
191
192	if (!static_cpu_has(X86_FEATURE_XENPV)) {
 
193		native_swapgs();
194		wrgsbase(gsbase);
195		native_swapgs();
196	} else {
197		instrumentation_begin();
198		wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
199		instrumentation_end();
200	}
201}
202
203/*
204 * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
205 * not available.  The goal is to be reasonably fast on non-FSGSBASE systems.
206 * It's forcibly inlined because it'll generate better code and this function
207 * is hot.
208 */
209static __always_inline void save_base_legacy(struct task_struct *prev_p,
210					     unsigned short selector,
211					     enum which_selector which)
212{
213	if (likely(selector == 0)) {
214		/*
215		 * On Intel (without X86_BUG_NULL_SEG), the segment base could
216		 * be the pre-existing saved base or it could be zero.  On AMD
217		 * (with X86_BUG_NULL_SEG), the segment base could be almost
218		 * anything.
219		 *
220		 * This branch is very hot (it's hit twice on almost every
221		 * context switch between 64-bit programs), and avoiding
222		 * the RDMSR helps a lot, so we just assume that whatever
223		 * value is already saved is correct.  This matches historical
224		 * Linux behavior, so it won't break existing applications.
225		 *
226		 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
227		 * report that the base is zero, it needs to actually be zero:
228		 * see the corresponding logic in load_seg_legacy.
229		 */
230	} else {
231		/*
232		 * If the selector is 1, 2, or 3, then the base is zero on
233		 * !X86_BUG_NULL_SEG CPUs and could be anything on
234		 * X86_BUG_NULL_SEG CPUs.  In the latter case, Linux
235		 * has never attempted to preserve the base across context
236		 * switches.
237		 *
238		 * If selector > 3, then it refers to a real segment, and
239		 * saving the base isn't necessary.
240		 */
241		if (which == FS)
242			prev_p->thread.fsbase = 0;
243		else
244			prev_p->thread.gsbase = 0;
245	}
246}
247
248static __always_inline void save_fsgs(struct task_struct *task)
249{
250	savesegment(fs, task->thread.fsindex);
251	savesegment(gs, task->thread.gsindex);
252	if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
253		/*
254		 * If FSGSBASE is enabled, we can't make any useful guesses
255		 * about the base, and user code expects us to save the current
256		 * value.  Fortunately, reading the base directly is efficient.
257		 */
258		task->thread.fsbase = rdfsbase();
259		task->thread.gsbase = __rdgsbase_inactive();
260	} else {
261		save_base_legacy(task, task->thread.fsindex, FS);
262		save_base_legacy(task, task->thread.gsindex, GS);
263	}
264}
265
266/*
267 * While a process is running,current->thread.fsbase and current->thread.gsbase
268 * may not match the corresponding CPU registers (see save_base_legacy()).
269 */
270void current_save_fsgs(void)
271{
272	unsigned long flags;
273
274	/* Interrupts need to be off for FSGSBASE */
275	local_irq_save(flags);
276	save_fsgs(current);
277	local_irq_restore(flags);
278}
279#if IS_ENABLED(CONFIG_KVM)
280EXPORT_SYMBOL_GPL(current_save_fsgs);
281#endif
282
283static __always_inline void loadseg(enum which_selector which,
284				    unsigned short sel)
285{
286	if (which == FS)
287		loadsegment(fs, sel);
288	else
289		load_gs_index(sel);
290}
291
292static __always_inline void load_seg_legacy(unsigned short prev_index,
293					    unsigned long prev_base,
294					    unsigned short next_index,
295					    unsigned long next_base,
296					    enum which_selector which)
297{
298	if (likely(next_index <= 3)) {
299		/*
300		 * The next task is using 64-bit TLS, is not using this
301		 * segment at all, or is having fun with arcane CPU features.
302		 */
303		if (next_base == 0) {
304			/*
305			 * Nasty case: on AMD CPUs, we need to forcibly zero
306			 * the base.
307			 */
308			if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
309				loadseg(which, __USER_DS);
310				loadseg(which, next_index);
311			} else {
312				/*
313				 * We could try to exhaustively detect cases
314				 * under which we can skip the segment load,
315				 * but there's really only one case that matters
316				 * for performance: if both the previous and
317				 * next states are fully zeroed, we can skip
318				 * the load.
319				 *
320				 * (This assumes that prev_base == 0 has no
321				 * false positives.  This is the case on
322				 * Intel-style CPUs.)
323				 */
324				if (likely(prev_index | next_index | prev_base))
325					loadseg(which, next_index);
326			}
327		} else {
328			if (prev_index != next_index)
329				loadseg(which, next_index);
330			wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
331			       next_base);
332		}
333	} else {
334		/*
335		 * The next task is using a real segment.  Loading the selector
336		 * is sufficient.
337		 */
338		loadseg(which, next_index);
339	}
340}
341
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
342static __always_inline void x86_fsgsbase_load(struct thread_struct *prev,
343					      struct thread_struct *next)
344{
345	if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
346		/* Update the FS and GS selectors if they could have changed. */
347		if (unlikely(prev->fsindex || next->fsindex))
348			loadseg(FS, next->fsindex);
349		if (unlikely(prev->gsindex || next->gsindex))
350			loadseg(GS, next->gsindex);
351
352		/* Update the bases. */
353		wrfsbase(next->fsbase);
354		__wrgsbase_inactive(next->gsbase);
355	} else {
356		load_seg_legacy(prev->fsindex, prev->fsbase,
357				next->fsindex, next->fsbase, FS);
358		load_seg_legacy(prev->gsindex, prev->gsbase,
359				next->gsindex, next->gsbase, GS);
360	}
361}
362
363unsigned long x86_fsgsbase_read_task(struct task_struct *task,
364				     unsigned short selector)
365{
366	unsigned short idx = selector >> 3;
367	unsigned long base;
368
369	if (likely((selector & SEGMENT_TI_MASK) == 0)) {
370		if (unlikely(idx >= GDT_ENTRIES))
371			return 0;
372
373		/*
374		 * There are no user segments in the GDT with nonzero bases
375		 * other than the TLS segments.
376		 */
377		if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
378			return 0;
379
380		idx -= GDT_ENTRY_TLS_MIN;
381		base = get_desc_base(&task->thread.tls_array[idx]);
382	} else {
383#ifdef CONFIG_MODIFY_LDT_SYSCALL
384		struct ldt_struct *ldt;
385
386		/*
387		 * If performance here mattered, we could protect the LDT
388		 * with RCU.  This is a slow path, though, so we can just
389		 * take the mutex.
390		 */
391		mutex_lock(&task->mm->context.lock);
392		ldt = task->mm->context.ldt;
393		if (unlikely(!ldt || idx >= ldt->nr_entries))
394			base = 0;
395		else
396			base = get_desc_base(ldt->entries + idx);
397		mutex_unlock(&task->mm->context.lock);
398#else
399		base = 0;
400#endif
401	}
402
403	return base;
404}
405
406unsigned long x86_gsbase_read_cpu_inactive(void)
407{
408	unsigned long gsbase;
409
410	if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
411		unsigned long flags;
412
413		local_irq_save(flags);
414		gsbase = __rdgsbase_inactive();
415		local_irq_restore(flags);
416	} else {
417		rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
418	}
419
420	return gsbase;
421}
422
423void x86_gsbase_write_cpu_inactive(unsigned long gsbase)
424{
425	if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
426		unsigned long flags;
427
428		local_irq_save(flags);
429		__wrgsbase_inactive(gsbase);
430		local_irq_restore(flags);
431	} else {
432		wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
433	}
434}
435
436unsigned long x86_fsbase_read_task(struct task_struct *task)
437{
438	unsigned long fsbase;
439
440	if (task == current)
441		fsbase = x86_fsbase_read_cpu();
442	else if (static_cpu_has(X86_FEATURE_FSGSBASE) ||
443		 (task->thread.fsindex == 0))
444		fsbase = task->thread.fsbase;
445	else
446		fsbase = x86_fsgsbase_read_task(task, task->thread.fsindex);
447
448	return fsbase;
449}
450
451unsigned long x86_gsbase_read_task(struct task_struct *task)
452{
453	unsigned long gsbase;
454
455	if (task == current)
456		gsbase = x86_gsbase_read_cpu_inactive();
457	else if (static_cpu_has(X86_FEATURE_FSGSBASE) ||
458		 (task->thread.gsindex == 0))
459		gsbase = task->thread.gsbase;
460	else
461		gsbase = x86_fsgsbase_read_task(task, task->thread.gsindex);
462
463	return gsbase;
464}
465
466void x86_fsbase_write_task(struct task_struct *task, unsigned long fsbase)
467{
468	WARN_ON_ONCE(task == current);
469
470	task->thread.fsbase = fsbase;
471}
472
473void x86_gsbase_write_task(struct task_struct *task, unsigned long gsbase)
474{
475	WARN_ON_ONCE(task == current);
476
477	task->thread.gsbase = gsbase;
478}
479
480static void
481start_thread_common(struct pt_regs *regs, unsigned long new_ip,
482		    unsigned long new_sp,
483		    unsigned int _cs, unsigned int _ss, unsigned int _ds)
484{
485	WARN_ON_ONCE(regs != current_pt_regs());
486
487	if (static_cpu_has(X86_BUG_NULL_SEG)) {
488		/* Loading zero below won't clear the base. */
489		loadsegment(fs, __USER_DS);
490		load_gs_index(__USER_DS);
491	}
492
 
 
493	loadsegment(fs, 0);
494	loadsegment(es, _ds);
495	loadsegment(ds, _ds);
496	load_gs_index(0);
497
498	regs->ip		= new_ip;
499	regs->sp		= new_sp;
500	regs->cs		= _cs;
501	regs->ss		= _ss;
502	regs->flags		= X86_EFLAGS_IF;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
503}
504
505void
506start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
507{
508	start_thread_common(regs, new_ip, new_sp,
509			    __USER_CS, __USER_DS, 0);
510}
511EXPORT_SYMBOL_GPL(start_thread);
512
513#ifdef CONFIG_COMPAT
514void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp)
515{
516	start_thread_common(regs, new_ip, new_sp,
517			    test_thread_flag(TIF_X32)
518			    ? __USER_CS : __USER32_CS,
519			    __USER_DS, __USER_DS);
520}
521#endif
522
523/*
524 *	switch_to(x,y) should switch tasks from x to y.
525 *
526 * This could still be optimized:
527 * - fold all the options into a flag word and test it with a single test.
528 * - could test fs/gs bitsliced
529 *
530 * Kprobes not supported here. Set the probe on schedule instead.
531 * Function graph tracer not supported too.
532 */
 
533__visible __notrace_funcgraph struct task_struct *
534__switch_to(struct task_struct *prev_p, struct task_struct *next_p)
535{
536	struct thread_struct *prev = &prev_p->thread;
537	struct thread_struct *next = &next_p->thread;
538	struct fpu *prev_fpu = &prev->fpu;
539	struct fpu *next_fpu = &next->fpu;
540	int cpu = smp_processor_id();
541
542	WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
543		     this_cpu_read(irq_count) != -1);
544
545	if (!test_thread_flag(TIF_NEED_FPU_LOAD))
546		switch_fpu_prepare(prev_fpu, cpu);
547
548	/* We must save %fs and %gs before load_TLS() because
549	 * %fs and %gs may be cleared by load_TLS().
550	 *
551	 * (e.g. xen_load_tls())
552	 */
553	save_fsgs(prev_p);
554
555	/*
556	 * Load TLS before restoring any segments so that segment loads
557	 * reference the correct GDT entries.
558	 */
559	load_TLS(next, cpu);
560
561	/*
562	 * Leave lazy mode, flushing any hypercalls made here.  This
563	 * must be done after loading TLS entries in the GDT but before
564	 * loading segments that might reference them.
565	 */
566	arch_end_context_switch(next_p);
567
568	/* Switch DS and ES.
569	 *
570	 * Reading them only returns the selectors, but writing them (if
571	 * nonzero) loads the full descriptor from the GDT or LDT.  The
572	 * LDT for next is loaded in switch_mm, and the GDT is loaded
573	 * above.
574	 *
575	 * We therefore need to write new values to the segment
576	 * registers on every context switch unless both the new and old
577	 * values are zero.
578	 *
579	 * Note that we don't need to do anything for CS and SS, as
580	 * those are saved and restored as part of pt_regs.
581	 */
582	savesegment(es, prev->es);
583	if (unlikely(next->es | prev->es))
584		loadsegment(es, next->es);
585
586	savesegment(ds, prev->ds);
587	if (unlikely(next->ds | prev->ds))
588		loadsegment(ds, next->ds);
589
590	x86_fsgsbase_load(prev, next);
591
 
 
592	/*
593	 * Switch the PDA and FPU contexts.
594	 */
595	this_cpu_write(current_task, next_p);
596	this_cpu_write(cpu_current_top_of_stack, task_top_of_stack(next_p));
597
598	switch_fpu_finish(next_fpu);
599
600	/* Reload sp0. */
601	update_task_stack(next_p);
602
603	switch_to_extra(prev_p, next_p);
604
605	if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
606		/*
607		 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
608		 * does not update the cached descriptor.  As a result, if we
609		 * do SYSRET while SS is NULL, we'll end up in user mode with
610		 * SS apparently equal to __USER_DS but actually unusable.
611		 *
612		 * The straightforward workaround would be to fix it up just
613		 * before SYSRET, but that would slow down the system call
614		 * fast paths.  Instead, we ensure that SS is never NULL in
615		 * system call context.  We do this by replacing NULL SS
616		 * selectors at every context switch.  SYSCALL sets up a valid
617		 * SS, so the only way to get NULL is to re-enter the kernel
618		 * from CPL 3 through an interrupt.  Since that can't happen
619		 * in the same task as a running syscall, we are guaranteed to
620		 * context switch between every interrupt vector entry and a
621		 * subsequent SYSRET.
622		 *
623		 * We read SS first because SS reads are much faster than
624		 * writes.  Out of caution, we force SS to __KERNEL_DS even if
625		 * it previously had a different non-NULL value.
626		 */
627		unsigned short ss_sel;
628		savesegment(ss, ss_sel);
629		if (ss_sel != __KERNEL_DS)
630			loadsegment(ss, __KERNEL_DS);
631	}
632
633	/* Load the Intel cache allocation PQR MSR. */
634	resctrl_sched_in();
635
636	return prev_p;
637}
638
639void set_personality_64bit(void)
640{
641	/* inherit personality from parent */
642
643	/* Make sure to be in 64bit mode */
644	clear_thread_flag(TIF_IA32);
645	clear_thread_flag(TIF_ADDR32);
646	clear_thread_flag(TIF_X32);
647	/* Pretend that this comes from a 64bit execve */
648	task_pt_regs(current)->orig_ax = __NR_execve;
649	current_thread_info()->status &= ~TS_COMPAT;
650
651	/* Ensure the corresponding mm is not marked. */
652	if (current->mm)
653		current->mm->context.ia32_compat = 0;
654
655	/* TBD: overwrites user setup. Should have two bits.
656	   But 64bit processes have always behaved this way,
657	   so it's not too bad. The main problem is just that
658	   32bit children are affected again. */
659	current->personality &= ~READ_IMPLIES_EXEC;
660}
661
662static void __set_personality_x32(void)
663{
664#ifdef CONFIG_X86_X32
665	clear_thread_flag(TIF_IA32);
666	set_thread_flag(TIF_X32);
667	if (current->mm)
668		current->mm->context.ia32_compat = TIF_X32;
 
669	current->personality &= ~READ_IMPLIES_EXEC;
670	/*
671	 * in_32bit_syscall() uses the presence of the x32 syscall bit
672	 * flag to determine compat status.  The x86 mmap() code relies on
673	 * the syscall bitness so set x32 syscall bit right here to make
674	 * in_32bit_syscall() work during exec().
675	 *
676	 * Pretend to come from a x32 execve.
677	 */
678	task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
679	current_thread_info()->status &= ~TS_COMPAT;
680#endif
681}
682
683static void __set_personality_ia32(void)
684{
685#ifdef CONFIG_IA32_EMULATION
686	set_thread_flag(TIF_IA32);
687	clear_thread_flag(TIF_X32);
688	if (current->mm)
689		current->mm->context.ia32_compat = TIF_IA32;
 
 
 
 
690	current->personality |= force_personality32;
691	/* Prepare the first "return" to user space */
692	task_pt_regs(current)->orig_ax = __NR_ia32_execve;
693	current_thread_info()->status |= TS_COMPAT;
694#endif
695}
696
697void set_personality_ia32(bool x32)
698{
699	/* Make sure to be in 32bit mode */
700	set_thread_flag(TIF_ADDR32);
701
702	if (x32)
703		__set_personality_x32();
704	else
705		__set_personality_ia32();
706}
707EXPORT_SYMBOL_GPL(set_personality_ia32);
708
709#ifdef CONFIG_CHECKPOINT_RESTORE
710static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
711{
712	int ret;
713
714	ret = map_vdso_once(image, addr);
715	if (ret)
716		return ret;
717
718	return (long)image->size;
719}
720#endif
721
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
722long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
723{
724	int ret = 0;
725
726	switch (option) {
727	case ARCH_SET_GS: {
728		if (unlikely(arg2 >= TASK_SIZE_MAX))
729			return -EPERM;
730
731		preempt_disable();
732		/*
733		 * ARCH_SET_GS has always overwritten the index
734		 * and the base. Zero is the most sensible value
735		 * to put in the index, and is the only value that
736		 * makes any sense if FSGSBASE is unavailable.
737		 */
738		if (task == current) {
739			loadseg(GS, 0);
740			x86_gsbase_write_cpu_inactive(arg2);
741
742			/*
743			 * On non-FSGSBASE systems, save_base_legacy() expects
744			 * that we also fill in thread.gsbase.
745			 */
746			task->thread.gsbase = arg2;
747
748		} else {
749			task->thread.gsindex = 0;
750			x86_gsbase_write_task(task, arg2);
751		}
752		preempt_enable();
753		break;
754	}
755	case ARCH_SET_FS: {
756		/*
757		 * Not strictly needed for %fs, but do it for symmetry
758		 * with %gs
759		 */
760		if (unlikely(arg2 >= TASK_SIZE_MAX))
761			return -EPERM;
762
763		preempt_disable();
764		/*
765		 * Set the selector to 0 for the same reason
766		 * as %gs above.
767		 */
768		if (task == current) {
769			loadseg(FS, 0);
770			x86_fsbase_write_cpu(arg2);
771
772			/*
773			 * On non-FSGSBASE systems, save_base_legacy() expects
774			 * that we also fill in thread.fsbase.
775			 */
776			task->thread.fsbase = arg2;
777		} else {
778			task->thread.fsindex = 0;
779			x86_fsbase_write_task(task, arg2);
780		}
781		preempt_enable();
782		break;
783	}
784	case ARCH_GET_FS: {
785		unsigned long base = x86_fsbase_read_task(task);
786
787		ret = put_user(base, (unsigned long __user *)arg2);
788		break;
789	}
790	case ARCH_GET_GS: {
791		unsigned long base = x86_gsbase_read_task(task);
792
793		ret = put_user(base, (unsigned long __user *)arg2);
794		break;
795	}
796
797#ifdef CONFIG_CHECKPOINT_RESTORE
798# ifdef CONFIG_X86_X32_ABI
799	case ARCH_MAP_VDSO_X32:
800		return prctl_map_vdso(&vdso_image_x32, arg2);
801# endif
802# if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
803	case ARCH_MAP_VDSO_32:
804		return prctl_map_vdso(&vdso_image_32, arg2);
805# endif
806	case ARCH_MAP_VDSO_64:
807		return prctl_map_vdso(&vdso_image_64, arg2);
808#endif
809
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
810	default:
811		ret = -EINVAL;
812		break;
813	}
814
815	return ret;
816}
817
818SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
819{
820	long ret;
821
822	ret = do_arch_prctl_64(current, option, arg2);
823	if (ret == -EINVAL)
824		ret = do_arch_prctl_common(current, option, arg2);
825
826	return ret;
827}
828
829#ifdef CONFIG_IA32_EMULATION
830COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
831{
832	return do_arch_prctl_common(current, option, arg2);
833}
834#endif
835
836unsigned long KSTK_ESP(struct task_struct *task)
837{
838	return task_pt_regs(task)->sp;
839}