1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * x86 SMP booting functions
5 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
6 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
7 * Copyright 2001 Andi Kleen, SuSE Labs.
9 * Much of the core SMP work is based on previous work by Thomas Radke, to
10 * whom a great many thanks are extended.
12 * Thanks to Intel for making available several different Pentium,
13 * Pentium Pro and Pentium-II/Xeon MP machines.
14 * Original development of Linux SMP code supported by Caldera.
17 * Felix Koop : NR_CPUS used properly
18 * Jose Renau : Handle single CPU case.
19 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
20 * Greg Wright : Fix for kernel stacks panic.
21 * Erich Boleyn : MP v1.4 and additional changes.
22 * Matthias Sattler : Changes for 2.1 kernel map.
23 * Michel Lespinasse : Changes for 2.1 kernel map.
24 * Michael Chastain : Change trampoline.S to gnu as.
25 * Alan Cox : Dumb bug: 'B' step PPro's are fine
26 * Ingo Molnar : Added APIC timers, based on code
28 * Ingo Molnar : various cleanups and rewrites
29 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
30 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
31 * Andi Kleen : Changed for SMP boot into long mode.
32 * Martin J. Bligh : Added support for multi-quad systems
33 * Dave Jones : Report invalid combinations of Athlon CPUs.
34 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
35 * Andi Kleen : Converted to new state machine.
36 * Ashok Raj : CPU hotplug support
37 * Glauber Costa : i386 and x86_64 integration
40 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
42 #include <linux/init.h>
43 #include <linux/smp.h>
44 #include <linux/export.h>
45 #include <linux/sched.h>
46 #include <linux/sched/topology.h>
47 #include <linux/sched/hotplug.h>
48 #include <linux/sched/task_stack.h>
49 #include <linux/percpu.h>
50 #include <linux/memblock.h>
51 #include <linux/err.h>
52 #include <linux/nmi.h>
53 #include <linux/tboot.h>
54 #include <linux/gfp.h>
55 #include <linux/cpuidle.h>
56 #include <linux/numa.h>
57 #include <linux/pgtable.h>
58 #include <linux/overflow.h>
59 #include <linux/syscore_ops.h>
65 #include <asm/realmode.h>
68 #include <asm/tlbflush.h>
70 #include <asm/mwait.h>
72 #include <asm/io_apic.h>
73 #include <asm/fpu/internal.h>
74 #include <asm/setup.h>
75 #include <asm/uv/uv.h>
76 #include <linux/mc146818rtc.h>
77 #include <asm/i8259.h>
79 #include <asm/qspinlock.h>
80 #include <asm/intel-family.h>
81 #include <asm/cpu_device_id.h>
82 #include <asm/spec-ctrl.h>
83 #include <asm/hw_irq.h>
84 #include <asm/stackprotector.h>
86 #ifdef CONFIG_ACPI_CPPC_LIB
87 #include <acpi/cppc_acpi.h>
90 /* representing HT siblings of each logical CPU */
91 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
92 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
94 /* representing HT and core siblings of each logical CPU */
95 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
96 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
98 /* representing HT, core, and die siblings of each logical CPU */
99 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
100 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
102 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
104 /* Per CPU bogomips and other parameters */
105 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
106 EXPORT_PER_CPU_SYMBOL(cpu_info);
108 /* Logical package management. We might want to allocate that dynamically */
109 unsigned int __max_logical_packages __read_mostly;
110 EXPORT_SYMBOL(__max_logical_packages);
111 static unsigned int logical_packages __read_mostly;
112 static unsigned int logical_die __read_mostly;
114 /* Maximum number of SMT threads on any online core */
115 int __read_mostly __max_smt_threads = 1;
117 /* Flag to indicate if a complete sched domain rebuild is required */
118 bool x86_topology_update;
120 int arch_update_cpu_topology(void)
122 int retval = x86_topology_update;
124 x86_topology_update = false;
128 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
132 spin_lock_irqsave(&rtc_lock, flags);
133 CMOS_WRITE(0xa, 0xf);
134 spin_unlock_irqrestore(&rtc_lock, flags);
135 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
137 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
141 static inline void smpboot_restore_warm_reset_vector(void)
146 * Paranoid: Set warm reset code and vector here back
149 spin_lock_irqsave(&rtc_lock, flags);
151 spin_unlock_irqrestore(&rtc_lock, flags);
153 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
156 static void init_freq_invariance(bool secondary, bool cppc_ready);
159 * Report back to the Boot Processor during boot time or to the caller processor
162 static void smp_callin(void)
167 * If waken up by an INIT in an 82489DX configuration
168 * cpu_callout_mask guarantees we don't get here before
169 * an INIT_deassert IPI reaches our local APIC, so it is
170 * now safe to touch our local APIC.
172 cpuid = smp_processor_id();
175 * the boot CPU has finished the init stage and is spinning
176 * on callin_map until we finish. We are free to set up this
177 * CPU, first the APIC. (this is probably redundant on most
183 * Save our processor parameters. Note: this information
184 * is needed for clock calibration.
186 smp_store_cpu_info(cpuid);
189 * The topology information must be up to date before
190 * calibrate_delay() and notify_cpu_starting().
192 set_cpu_sibling_map(raw_smp_processor_id());
194 init_freq_invariance(true, false);
198 * Update loops_per_jiffy in cpu_data. Previous call to
199 * smp_store_cpu_info() stored a value that is close but not as
200 * accurate as the value just calculated.
203 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
204 pr_debug("Stack at about %p\n", &cpuid);
208 notify_cpu_starting(cpuid);
211 * Allow the master to continue.
213 cpumask_set_cpu(cpuid, cpu_callin_mask);
216 static int cpu0_logical_apicid;
217 static int enable_start_cpu0;
219 * Activate a secondary processor.
221 static void notrace start_secondary(void *unused)
224 * Don't put *anything* except direct CPU state initialization
225 * before cpu_init(), SMP booting is too fragile that we want to
226 * limit the things done here to the most necessary things.
231 /* switch away from the initial page table */
232 load_cr3(swapper_pg_dir);
235 cpu_init_secondary();
236 rcu_cpu_starting(raw_smp_processor_id());
237 x86_cpuinit.early_percpu_clock_init();
240 enable_start_cpu0 = 0;
242 /* otherwise gcc will move up smp_processor_id before the cpu_init */
245 * Check TSC synchronization with the boot CPU:
247 check_tsc_sync_target();
249 speculative_store_bypass_ht_init();
252 * Lock vector_lock, set CPU online and bring the vector
253 * allocator online. Online must be set with vector_lock held
254 * to prevent a concurrent irq setup/teardown from seeing a
255 * half valid vector space.
258 set_cpu_online(smp_processor_id(), true);
260 unlock_vector_lock();
261 cpu_set_state_online(smp_processor_id());
262 x86_platform.nmi_init();
264 /* enable local interrupts */
267 x86_cpuinit.setup_percpu_clockev();
270 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
274 * topology_is_primary_thread - Check whether CPU is the primary SMT thread
277 bool topology_is_primary_thread(unsigned int cpu)
279 return apic_id_is_primary_thread(per_cpu(x86_cpu_to_apicid, cpu));
283 * topology_smt_supported - Check whether SMT is supported by the CPUs
285 bool topology_smt_supported(void)
287 return smp_num_siblings > 1;
291 * topology_phys_to_logical_pkg - Map a physical package id to a logical
293 * Returns logical package id or -1 if not found
295 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
299 for_each_possible_cpu(cpu) {
300 struct cpuinfo_x86 *c = &cpu_data(cpu);
302 if (c->initialized && c->phys_proc_id == phys_pkg)
303 return c->logical_proc_id;
307 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
309 * topology_phys_to_logical_die - Map a physical die id to logical
311 * Returns logical die id or -1 if not found
313 int topology_phys_to_logical_die(unsigned int die_id, unsigned int cur_cpu)
316 int proc_id = cpu_data(cur_cpu).phys_proc_id;
318 for_each_possible_cpu(cpu) {
319 struct cpuinfo_x86 *c = &cpu_data(cpu);
321 if (c->initialized && c->cpu_die_id == die_id &&
322 c->phys_proc_id == proc_id)
323 return c->logical_die_id;
327 EXPORT_SYMBOL(topology_phys_to_logical_die);
330 * topology_update_package_map - Update the physical to logical package map
331 * @pkg: The physical package id as retrieved via CPUID
332 * @cpu: The cpu for which this is updated
334 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
338 /* Already available somewhere? */
339 new = topology_phys_to_logical_pkg(pkg);
343 new = logical_packages++;
345 pr_info("CPU %u Converting physical %u to logical package %u\n",
349 cpu_data(cpu).logical_proc_id = new;
353 * topology_update_die_map - Update the physical to logical die map
354 * @die: The die id as retrieved via CPUID
355 * @cpu: The cpu for which this is updated
357 int topology_update_die_map(unsigned int die, unsigned int cpu)
361 /* Already available somewhere? */
362 new = topology_phys_to_logical_die(die, cpu);
368 pr_info("CPU %u Converting physical %u to logical die %u\n",
372 cpu_data(cpu).logical_die_id = new;
376 void __init smp_store_boot_cpu_info(void)
378 int id = 0; /* CPU 0 */
379 struct cpuinfo_x86 *c = &cpu_data(id);
383 topology_update_package_map(c->phys_proc_id, id);
384 topology_update_die_map(c->cpu_die_id, id);
385 c->initialized = true;
389 * The bootstrap kernel entry code has set these up. Save them for
392 void smp_store_cpu_info(int id)
394 struct cpuinfo_x86 *c = &cpu_data(id);
396 /* Copy boot_cpu_data only on the first bringup */
401 * During boot time, CPU0 has this setup already. Save the info when
402 * bringing up AP or offlined CPU0.
404 identify_secondary_cpu(c);
405 c->initialized = true;
409 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
411 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
413 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
417 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
419 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
421 return !WARN_ONCE(!topology_same_node(c, o),
422 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
423 "[node: %d != %d]. Ignoring dependency.\n",
424 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
427 #define link_mask(mfunc, c1, c2) \
429 cpumask_set_cpu((c1), mfunc(c2)); \
430 cpumask_set_cpu((c2), mfunc(c1)); \
433 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
435 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
436 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
438 if (c->phys_proc_id == o->phys_proc_id &&
439 c->cpu_die_id == o->cpu_die_id &&
440 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
441 if (c->cpu_core_id == o->cpu_core_id)
442 return topology_sane(c, o, "smt");
444 if ((c->cu_id != 0xff) &&
445 (o->cu_id != 0xff) &&
446 (c->cu_id == o->cu_id))
447 return topology_sane(c, o, "smt");
450 } else if (c->phys_proc_id == o->phys_proc_id &&
451 c->cpu_die_id == o->cpu_die_id &&
452 c->cpu_core_id == o->cpu_core_id) {
453 return topology_sane(c, o, "smt");
459 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
461 if (c->phys_proc_id == o->phys_proc_id &&
462 c->cpu_die_id == o->cpu_die_id)
468 * Unlike the other levels, we do not enforce keeping a
469 * multicore group inside a NUMA node. If this happens, we will
470 * discard the MC level of the topology later.
472 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
474 if (c->phys_proc_id == o->phys_proc_id)
480 * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
482 * Any Intel CPU that has multiple nodes per package and does not
483 * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
485 * When in SNC mode, these CPUs enumerate an LLC that is shared
486 * by multiple NUMA nodes. The LLC is shared for off-package data
487 * access but private to the NUMA node (half of the package) for
488 * on-package access. CPUID (the source of the information about
489 * the LLC) can only enumerate the cache as shared or unshared,
490 * but not this particular configuration.
493 static const struct x86_cpu_id intel_cod_cpu[] = {
494 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, 0), /* COD */
495 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, 0), /* COD */
496 X86_MATCH_INTEL_FAM6_MODEL(ANY, 1), /* SNC */
500 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
502 const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
503 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
504 bool intel_snc = id && id->driver_data;
506 /* Do not match if we do not have a valid APICID for cpu: */
507 if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
510 /* Do not match if LLC id does not match: */
511 if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
515 * Allow the SNC topology without warning. Return of false
516 * means 'c' does not share the LLC of 'o'. This will be
517 * reflected to userspace.
519 if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
522 return topology_sane(c, o, "llc");
526 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
527 static inline int x86_sched_itmt_flags(void)
529 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
532 #ifdef CONFIG_SCHED_MC
533 static int x86_core_flags(void)
535 return cpu_core_flags() | x86_sched_itmt_flags();
538 #ifdef CONFIG_SCHED_SMT
539 static int x86_smt_flags(void)
541 return cpu_smt_flags() | x86_sched_itmt_flags();
546 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
547 #ifdef CONFIG_SCHED_SMT
548 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
550 #ifdef CONFIG_SCHED_MC
551 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
556 static struct sched_domain_topology_level x86_topology[] = {
557 #ifdef CONFIG_SCHED_SMT
558 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
560 #ifdef CONFIG_SCHED_MC
561 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
563 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
568 * Set if a package/die has multiple NUMA nodes inside.
569 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
570 * Sub-NUMA Clustering have this.
572 static bool x86_has_numa_in_package;
574 void set_cpu_sibling_map(int cpu)
576 bool has_smt = smp_num_siblings > 1;
577 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
578 struct cpuinfo_x86 *c = &cpu_data(cpu);
579 struct cpuinfo_x86 *o;
582 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
585 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
586 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
587 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
588 cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
593 for_each_cpu(i, cpu_sibling_setup_mask) {
596 if (match_pkg(c, o) && !topology_same_node(c, o))
597 x86_has_numa_in_package = true;
599 if ((i == cpu) || (has_smt && match_smt(c, o)))
600 link_mask(topology_sibling_cpumask, cpu, i);
602 if ((i == cpu) || (has_mp && match_llc(c, o)))
603 link_mask(cpu_llc_shared_mask, cpu, i);
605 if ((i == cpu) || (has_mp && match_die(c, o)))
606 link_mask(topology_die_cpumask, cpu, i);
609 threads = cpumask_weight(topology_sibling_cpumask(cpu));
610 if (threads > __max_smt_threads)
611 __max_smt_threads = threads;
613 for_each_cpu(i, topology_sibling_cpumask(cpu))
614 cpu_data(i).smt_active = threads > 1;
617 * This needs a separate iteration over the cpus because we rely on all
618 * topology_sibling_cpumask links to be set-up.
620 for_each_cpu(i, cpu_sibling_setup_mask) {
623 if ((i == cpu) || (has_mp && match_pkg(c, o))) {
624 link_mask(topology_core_cpumask, cpu, i);
627 * Does this new cpu bringup a new core?
631 * for each core in package, increment
632 * the booted_cores for this new cpu
635 topology_sibling_cpumask(i)) == i)
638 * increment the core count for all
639 * the other cpus in this package
642 cpu_data(i).booted_cores++;
643 } else if (i != cpu && !c->booted_cores)
644 c->booted_cores = cpu_data(i).booted_cores;
649 /* maps the cpu to the sched domain representing multi-core */
650 const struct cpumask *cpu_coregroup_mask(int cpu)
652 return cpu_llc_shared_mask(cpu);
655 static void impress_friends(void)
658 unsigned long bogosum = 0;
660 * Allow the user to impress friends.
662 pr_debug("Before bogomips\n");
663 for_each_possible_cpu(cpu)
664 if (cpumask_test_cpu(cpu, cpu_callout_mask))
665 bogosum += cpu_data(cpu).loops_per_jiffy;
666 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
669 (bogosum/(5000/HZ))%100);
671 pr_debug("Before bogocount - setting activated=1\n");
674 void __inquire_remote_apic(int apicid)
676 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
677 const char * const names[] = { "ID", "VERSION", "SPIV" };
681 pr_info("Inquiring remote APIC 0x%x...\n", apicid);
683 for (i = 0; i < ARRAY_SIZE(regs); i++) {
684 pr_info("... APIC 0x%x %s: ", apicid, names[i]);
689 status = safe_apic_wait_icr_idle();
691 pr_cont("a previous APIC delivery may have failed\n");
693 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
698 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
699 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
702 case APIC_ICR_RR_VALID:
703 status = apic_read(APIC_RRR);
704 pr_cont("%08x\n", status);
713 * The Multiprocessor Specification 1.4 (1997) example code suggests
714 * that there should be a 10ms delay between the BSP asserting INIT
715 * and de-asserting INIT, when starting a remote processor.
716 * But that slows boot and resume on modern processors, which include
717 * many cores and don't require that delay.
719 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
720 * Modern processor families are quirked to remove the delay entirely.
722 #define UDELAY_10MS_DEFAULT 10000
724 static unsigned int init_udelay = UINT_MAX;
726 static int __init cpu_init_udelay(char *str)
728 get_option(&str, &init_udelay);
732 early_param("cpu_init_udelay", cpu_init_udelay);
734 static void __init smp_quirk_init_udelay(void)
736 /* if cmdline changed it from default, leave it alone */
737 if (init_udelay != UINT_MAX)
740 /* if modern processor, use no delay */
741 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
742 ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
743 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
747 /* else, use legacy delay */
748 init_udelay = UDELAY_10MS_DEFAULT;
752 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
753 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
754 * won't ... remember to clear down the APIC, etc later.
757 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
759 u32 dm = apic->dest_mode_logical ? APIC_DEST_LOGICAL : APIC_DEST_PHYSICAL;
760 unsigned long send_status, accept_status = 0;
764 /* Boot on the stack */
765 /* Kick the second */
766 apic_icr_write(APIC_DM_NMI | dm, apicid);
768 pr_debug("Waiting for send to finish...\n");
769 send_status = safe_apic_wait_icr_idle();
772 * Give the other CPU some time to accept the IPI.
775 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
776 maxlvt = lapic_get_maxlvt();
777 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
778 apic_write(APIC_ESR, 0);
779 accept_status = (apic_read(APIC_ESR) & 0xEF);
781 pr_debug("NMI sent\n");
784 pr_err("APIC never delivered???\n");
786 pr_err("APIC delivery error (%lx)\n", accept_status);
788 return (send_status | accept_status);
792 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
794 unsigned long send_status = 0, accept_status = 0;
795 int maxlvt, num_starts, j;
797 maxlvt = lapic_get_maxlvt();
800 * Be paranoid about clearing APIC errors.
802 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
803 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
804 apic_write(APIC_ESR, 0);
808 pr_debug("Asserting INIT\n");
811 * Turn INIT on target chip
816 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
819 pr_debug("Waiting for send to finish...\n");
820 send_status = safe_apic_wait_icr_idle();
824 pr_debug("Deasserting INIT\n");
828 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
830 pr_debug("Waiting for send to finish...\n");
831 send_status = safe_apic_wait_icr_idle();
836 * Should we send STARTUP IPIs ?
838 * Determine this based on the APIC version.
839 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
841 if (APIC_INTEGRATED(boot_cpu_apic_version))
847 * Run STARTUP IPI loop.
849 pr_debug("#startup loops: %d\n", num_starts);
851 for (j = 1; j <= num_starts; j++) {
852 pr_debug("Sending STARTUP #%d\n", j);
853 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
854 apic_write(APIC_ESR, 0);
856 pr_debug("After apic_write\n");
863 /* Boot on the stack */
864 /* Kick the second */
865 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
869 * Give the other CPU some time to accept the IPI.
871 if (init_udelay == 0)
876 pr_debug("Startup point 1\n");
878 pr_debug("Waiting for send to finish...\n");
879 send_status = safe_apic_wait_icr_idle();
882 * Give the other CPU some time to accept the IPI.
884 if (init_udelay == 0)
889 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
890 apic_write(APIC_ESR, 0);
891 accept_status = (apic_read(APIC_ESR) & 0xEF);
892 if (send_status || accept_status)
895 pr_debug("After Startup\n");
898 pr_err("APIC never delivered???\n");
900 pr_err("APIC delivery error (%lx)\n", accept_status);
902 return (send_status | accept_status);
905 /* reduce the number of lines printed when booting a large cpu count system */
906 static void announce_cpu(int cpu, int apicid)
908 static int current_node = NUMA_NO_NODE;
909 int node = early_cpu_to_node(cpu);
910 static int width, node_width;
913 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
916 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
919 printk(KERN_INFO "x86: Booting SMP configuration:\n");
921 if (system_state < SYSTEM_RUNNING) {
922 if (node != current_node) {
923 if (current_node > (-1))
927 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
928 node_width - num_digits(node), " ", node);
931 /* Add padding for the BSP */
933 pr_cont("%*s", width + 1, " ");
935 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
938 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
942 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
946 cpu = smp_processor_id();
947 if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
954 * Wake up AP by INIT, INIT, STARTUP sequence.
956 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
957 * boot-strap code which is not a desired behavior for waking up BSP. To
958 * void the boot-strap code, wake up CPU0 by NMI instead.
960 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
961 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
962 * We'll change this code in the future to wake up hard offlined CPU0 if
963 * real platform and request are available.
966 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
967 int *cpu0_nmi_registered)
975 * Wake up AP by INIT, INIT, STARTUP sequence.
978 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
983 * Wake up BSP by nmi.
985 * Register a NMI handler to help wake up CPU0.
987 boot_error = register_nmi_handler(NMI_LOCAL,
988 wakeup_cpu0_nmi, 0, "wake_cpu0");
991 enable_start_cpu0 = 1;
992 *cpu0_nmi_registered = 1;
993 id = apic->dest_mode_logical ? cpu0_logical_apicid : apicid;
994 boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
1003 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
1007 /* Just in case we booted with a single CPU. */
1008 alternatives_enable_smp();
1010 per_cpu(current_task, cpu) = idle;
1011 cpu_init_stack_canary(cpu, idle);
1013 /* Initialize the interrupt stack(s) */
1014 ret = irq_init_percpu_irqstack(cpu);
1018 #ifdef CONFIG_X86_32
1019 /* Stack for startup_32 can be just as for start_secondary onwards */
1020 per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
1022 initial_gs = per_cpu_offset(cpu);
1028 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
1029 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
1030 * Returns zero if CPU booted OK, else error code from
1031 * ->wakeup_secondary_cpu.
1033 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
1034 int *cpu0_nmi_registered)
1036 /* start_ip had better be page-aligned! */
1037 unsigned long start_ip = real_mode_header->trampoline_start;
1039 unsigned long boot_error = 0;
1040 unsigned long timeout;
1042 idle->thread.sp = (unsigned long)task_pt_regs(idle);
1043 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
1044 initial_code = (unsigned long)start_secondary;
1045 initial_stack = idle->thread.sp;
1047 /* Enable the espfix hack for this CPU */
1048 init_espfix_ap(cpu);
1050 /* So we see what's up */
1051 announce_cpu(cpu, apicid);
1054 * This grunge runs the startup process for
1055 * the targeted processor.
1058 if (x86_platform.legacy.warm_reset) {
1060 pr_debug("Setting warm reset code and vector.\n");
1062 smpboot_setup_warm_reset_vector(start_ip);
1064 * Be paranoid about clearing APIC errors.
1066 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1067 apic_write(APIC_ESR, 0);
1068 apic_read(APIC_ESR);
1073 * AP might wait on cpu_callout_mask in cpu_init() with
1074 * cpu_initialized_mask set if previous attempt to online
1075 * it timed-out. Clear cpu_initialized_mask so that after
1076 * INIT/SIPI it could start with a clean state.
1078 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1082 * Wake up a CPU in difference cases:
1083 * - Use the method in the APIC driver if it's defined
1085 * - Use an INIT boot APIC message for APs or NMI for BSP.
1087 if (apic->wakeup_secondary_cpu)
1088 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1090 boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1091 cpu0_nmi_registered);
1095 * Wait 10s total for first sign of life from AP
1098 timeout = jiffies + 10*HZ;
1099 while (time_before(jiffies, timeout)) {
1100 if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1102 * Tell AP to proceed with initialization
1104 cpumask_set_cpu(cpu, cpu_callout_mask);
1114 * Wait till AP completes initial initialization
1116 while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1118 * Allow other tasks to run while we wait for the
1119 * AP to come online. This also gives a chance
1120 * for the MTRR work(triggered by the AP coming online)
1121 * to be completed in the stop machine context.
1127 if (x86_platform.legacy.warm_reset) {
1129 * Cleanup possible dangling ends...
1131 smpboot_restore_warm_reset_vector();
1137 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1139 int apicid = apic->cpu_present_to_apicid(cpu);
1140 int cpu0_nmi_registered = 0;
1141 unsigned long flags;
1144 lockdep_assert_irqs_enabled();
1146 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1148 if (apicid == BAD_APICID ||
1149 !physid_isset(apicid, phys_cpu_present_map) ||
1150 !apic->apic_id_valid(apicid)) {
1151 pr_err("%s: bad cpu %d\n", __func__, cpu);
1156 * Already booted CPU?
1158 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1159 pr_debug("do_boot_cpu %d Already started\n", cpu);
1164 * Save current MTRR state in case it was changed since early boot
1165 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1169 /* x86 CPUs take themselves offline, so delayed offline is OK. */
1170 err = cpu_check_up_prepare(cpu);
1171 if (err && err != -EBUSY)
1174 /* the FPU context is blank, nobody can own it */
1175 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1177 err = common_cpu_up(cpu, tidle);
1181 err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1183 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1189 * Check TSC synchronization with the AP (keep irqs disabled
1192 local_irq_save(flags);
1193 check_tsc_sync_source(cpu);
1194 local_irq_restore(flags);
1196 while (!cpu_online(cpu)) {
1198 touch_nmi_watchdog();
1203 * Clean up the nmi handler. Do this after the callin and callout sync
1204 * to avoid impact of possible long unregister time.
1206 if (cpu0_nmi_registered)
1207 unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1213 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1215 void arch_disable_smp_support(void)
1217 disable_ioapic_support();
1221 * Fall back to non SMP mode after errors.
1223 * RED-PEN audit/test this more. I bet there is more state messed up here.
1225 static __init void disable_smp(void)
1227 pr_info("SMP disabled\n");
1229 disable_ioapic_support();
1231 init_cpu_present(cpumask_of(0));
1232 init_cpu_possible(cpumask_of(0));
1234 if (smp_found_config)
1235 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1237 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1238 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1239 cpumask_set_cpu(0, topology_core_cpumask(0));
1240 cpumask_set_cpu(0, topology_die_cpumask(0));
1244 * Various sanity checks.
1246 static void __init smp_sanity_check(void)
1250 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1251 if (def_to_bigsmp && nr_cpu_ids > 8) {
1255 pr_warn("More than 8 CPUs detected - skipping them\n"
1256 "Use CONFIG_X86_BIGSMP\n");
1259 for_each_present_cpu(cpu) {
1261 set_cpu_present(cpu, false);
1266 for_each_possible_cpu(cpu) {
1268 set_cpu_possible(cpu, false);
1276 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1277 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1278 hard_smp_processor_id());
1280 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1284 * Should not be necessary because the MP table should list the boot
1285 * CPU too, but we do it for the sake of robustness anyway.
1287 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1288 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1289 boot_cpu_physical_apicid);
1290 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1295 static void __init smp_cpu_index_default(void)
1298 struct cpuinfo_x86 *c;
1300 for_each_possible_cpu(i) {
1302 /* mark all to hotplug */
1303 c->cpu_index = nr_cpu_ids;
1307 static void __init smp_get_logical_apicid(void)
1310 cpu0_logical_apicid = apic_read(APIC_LDR);
1312 cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1316 * Prepare for SMP bootup.
1317 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1318 * for common interface support.
1320 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1324 smp_cpu_index_default();
1327 * Setup boot CPU information
1329 smp_store_boot_cpu_info(); /* Final full version of the data */
1330 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1333 for_each_possible_cpu(i) {
1334 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1335 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1336 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1337 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1341 * Set 'default' x86 topology, this matches default_topology() in that
1342 * it has NUMA nodes as a topology level. See also
1343 * native_smp_cpus_done().
1345 * Must be done before set_cpus_sibling_map() is ran.
1347 set_sched_topology(x86_topology);
1349 set_cpu_sibling_map(0);
1350 init_freq_invariance(false, false);
1353 switch (apic_intr_mode) {
1355 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1358 case APIC_SYMMETRIC_IO_NO_ROUTING:
1360 /* Setup local timer */
1361 x86_init.timers.setup_percpu_clockev();
1363 case APIC_VIRTUAL_WIRE:
1364 case APIC_SYMMETRIC_IO:
1368 /* Setup local timer */
1369 x86_init.timers.setup_percpu_clockev();
1371 smp_get_logical_apicid();
1374 print_cpu_info(&cpu_data(0));
1378 set_mtrr_aps_delayed_init();
1380 smp_quirk_init_udelay();
1382 speculative_store_bypass_ht_init();
1385 void arch_thaw_secondary_cpus_begin(void)
1387 set_mtrr_aps_delayed_init();
1390 void arch_thaw_secondary_cpus_end(void)
1396 * Early setup to make printk work.
1398 void __init native_smp_prepare_boot_cpu(void)
1400 int me = smp_processor_id();
1401 switch_to_new_gdt(me);
1402 /* already set me in cpu_online_mask in boot_cpu_init() */
1403 cpumask_set_cpu(me, cpu_callout_mask);
1404 cpu_set_state_online(me);
1405 native_pv_lock_init();
1408 void __init calculate_max_logical_packages(void)
1413 * Today neither Intel nor AMD support heterogeneous systems so
1414 * extrapolate the boot cpu's data to all packages.
1416 ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1417 __max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1418 pr_info("Max logical packages: %u\n", __max_logical_packages);
1421 void __init native_smp_cpus_done(unsigned int max_cpus)
1423 pr_debug("Boot done\n");
1425 calculate_max_logical_packages();
1427 if (x86_has_numa_in_package)
1428 set_sched_topology(x86_numa_in_package_topology);
1435 static int __initdata setup_possible_cpus = -1;
1436 static int __init _setup_possible_cpus(char *str)
1438 get_option(&str, &setup_possible_cpus);
1441 early_param("possible_cpus", _setup_possible_cpus);
1445 * cpu_possible_mask should be static, it cannot change as cpu's
1446 * are onlined, or offlined. The reason is per-cpu data-structures
1447 * are allocated by some modules at init time, and don't expect to
1448 * do this dynamically on cpu arrival/departure.
1449 * cpu_present_mask on the other hand can change dynamically.
1450 * In case when cpu_hotplug is not compiled, then we resort to current
1451 * behaviour, which is cpu_possible == cpu_present.
1454 * Three ways to find out the number of additional hotplug CPUs:
1455 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1456 * - The user can overwrite it with possible_cpus=NUM
1457 * - Otherwise don't reserve additional CPUs.
1458 * We do this because additional CPUs waste a lot of memory.
1461 __init void prefill_possible_map(void)
1465 /* No boot processor was found in mptable or ACPI MADT */
1466 if (!num_processors) {
1467 if (boot_cpu_has(X86_FEATURE_APIC)) {
1468 int apicid = boot_cpu_physical_apicid;
1469 int cpu = hard_smp_processor_id();
1471 pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1473 /* Make sure boot cpu is enumerated */
1474 if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1475 apic->apic_id_valid(apicid))
1476 generic_processor_info(apicid, boot_cpu_apic_version);
1479 if (!num_processors)
1483 i = setup_max_cpus ?: 1;
1484 if (setup_possible_cpus == -1) {
1485 possible = num_processors;
1486 #ifdef CONFIG_HOTPLUG_CPU
1488 possible += disabled_cpus;
1494 possible = setup_possible_cpus;
1496 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1498 /* nr_cpu_ids could be reduced via nr_cpus= */
1499 if (possible > nr_cpu_ids) {
1500 pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1501 possible, nr_cpu_ids);
1502 possible = nr_cpu_ids;
1505 #ifdef CONFIG_HOTPLUG_CPU
1506 if (!setup_max_cpus)
1509 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1510 possible, setup_max_cpus);
1514 nr_cpu_ids = possible;
1516 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1517 possible, max_t(int, possible - num_processors, 0));
1519 reset_cpu_possible_mask();
1521 for (i = 0; i < possible; i++)
1522 set_cpu_possible(i, true);
1525 #ifdef CONFIG_HOTPLUG_CPU
1527 /* Recompute SMT state for all CPUs on offline */
1528 static void recompute_smt_state(void)
1530 int max_threads, cpu;
1533 for_each_online_cpu (cpu) {
1534 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1536 if (threads > max_threads)
1537 max_threads = threads;
1539 __max_smt_threads = max_threads;
1542 static void remove_siblinginfo(int cpu)
1545 struct cpuinfo_x86 *c = &cpu_data(cpu);
1547 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1548 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1550 * last thread sibling in this cpu core going down
1552 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1553 cpu_data(sibling).booted_cores--;
1556 for_each_cpu(sibling, topology_die_cpumask(cpu))
1557 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1559 for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1560 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1561 if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1562 cpu_data(sibling).smt_active = false;
1565 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1566 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1567 cpumask_clear(cpu_llc_shared_mask(cpu));
1568 cpumask_clear(topology_sibling_cpumask(cpu));
1569 cpumask_clear(topology_core_cpumask(cpu));
1570 cpumask_clear(topology_die_cpumask(cpu));
1572 c->booted_cores = 0;
1573 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1574 recompute_smt_state();
1577 static void remove_cpu_from_maps(int cpu)
1579 set_cpu_online(cpu, false);
1580 cpumask_clear_cpu(cpu, cpu_callout_mask);
1581 cpumask_clear_cpu(cpu, cpu_callin_mask);
1582 /* was set by cpu_init() */
1583 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1584 numa_remove_cpu(cpu);
1587 void cpu_disable_common(void)
1589 int cpu = smp_processor_id();
1591 remove_siblinginfo(cpu);
1593 /* It's now safe to remove this processor from the online map */
1595 remove_cpu_from_maps(cpu);
1596 unlock_vector_lock();
1601 int native_cpu_disable(void)
1605 ret = lapic_can_unplug_cpu();
1609 cpu_disable_common();
1612 * Disable the local APIC. Otherwise IPI broadcasts will reach
1613 * it. It still responds normally to INIT, NMI, SMI, and SIPI
1616 * Disabling the APIC must happen after cpu_disable_common()
1617 * which invokes fixup_irqs().
1619 * Disabling the APIC preserves already set bits in IRR, but
1620 * an interrupt arriving after disabling the local APIC does not
1621 * set the corresponding IRR bit.
1623 * fixup_irqs() scans IRR for set bits so it can raise a not
1624 * yet handled interrupt on the new destination CPU via an IPI
1625 * but obviously it can't do so for IRR bits which are not set.
1626 * IOW, interrupts arriving after disabling the local APIC will
1629 apic_soft_disable();
1634 int common_cpu_die(unsigned int cpu)
1638 /* We don't do anything here: idle task is faking death itself. */
1640 /* They ack this in play_dead() by setting CPU_DEAD */
1641 if (cpu_wait_death(cpu, 5)) {
1642 if (system_state == SYSTEM_RUNNING)
1643 pr_info("CPU %u is now offline\n", cpu);
1645 pr_err("CPU %u didn't die...\n", cpu);
1652 void native_cpu_die(unsigned int cpu)
1654 common_cpu_die(cpu);
1657 void play_dead_common(void)
1662 (void)cpu_report_death();
1665 * With physical CPU hotplug, we should halt the cpu
1667 local_irq_disable();
1671 * cond_wakeup_cpu0 - Wake up CPU0 if needed.
1673 * If NMI wants to wake up CPU0, start CPU0.
1675 void cond_wakeup_cpu0(void)
1677 if (smp_processor_id() == 0 && enable_start_cpu0)
1680 EXPORT_SYMBOL_GPL(cond_wakeup_cpu0);
1683 * We need to flush the caches before going to sleep, lest we have
1684 * dirty data in our caches when we come back up.
1686 static inline void mwait_play_dead(void)
1688 unsigned int eax, ebx, ecx, edx;
1689 unsigned int highest_cstate = 0;
1690 unsigned int highest_subcstate = 0;
1694 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1695 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1697 if (!this_cpu_has(X86_FEATURE_MWAIT))
1699 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1701 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1704 eax = CPUID_MWAIT_LEAF;
1706 native_cpuid(&eax, &ebx, &ecx, &edx);
1709 * eax will be 0 if EDX enumeration is not valid.
1710 * Initialized below to cstate, sub_cstate value when EDX is valid.
1712 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1715 edx >>= MWAIT_SUBSTATE_SIZE;
1716 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1717 if (edx & MWAIT_SUBSTATE_MASK) {
1719 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1722 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1723 (highest_subcstate - 1);
1727 * This should be a memory location in a cache line which is
1728 * unlikely to be touched by other processors. The actual
1729 * content is immaterial as it is not actually modified in any way.
1731 mwait_ptr = ¤t_thread_info()->flags;
1737 * The CLFLUSH is a workaround for erratum AAI65 for
1738 * the Xeon 7400 series. It's not clear it is actually
1739 * needed, but it should be harmless in either case.
1740 * The WBINVD is insufficient due to the spurious-wakeup
1741 * case where we return around the loop.
1746 __monitor(mwait_ptr, 0, 0);
1754 void hlt_play_dead(void)
1756 if (__this_cpu_read(cpu_info.x86) >= 4)
1766 void native_play_dead(void)
1769 tboot_shutdown(TB_SHUTDOWN_WFS);
1771 mwait_play_dead(); /* Only returns on failure */
1772 if (cpuidle_play_dead())
1776 #else /* ... !CONFIG_HOTPLUG_CPU */
1777 int native_cpu_disable(void)
1782 void native_cpu_die(unsigned int cpu)
1784 /* We said "no" in __cpu_disable */
1788 void native_play_dead(void)
1795 #ifdef CONFIG_X86_64
1797 * APERF/MPERF frequency ratio computation.
1799 * The scheduler wants to do frequency invariant accounting and needs a <1
1800 * ratio to account for the 'current' frequency, corresponding to
1801 * freq_curr / freq_max.
1803 * Since the frequency freq_curr on x86 is controlled by micro-controller and
1804 * our P-state setting is little more than a request/hint, we need to observe
1805 * the effective frequency 'BusyMHz', i.e. the average frequency over a time
1806 * interval after discarding idle time. This is given by:
1808 * BusyMHz = delta_APERF / delta_MPERF * freq_base
1810 * where freq_base is the max non-turbo P-state.
1812 * The freq_max term has to be set to a somewhat arbitrary value, because we
1813 * can't know which turbo states will be available at a given point in time:
1814 * it all depends on the thermal headroom of the entire package. We set it to
1815 * the turbo level with 4 cores active.
1817 * Benchmarks show that's a good compromise between the 1C turbo ratio
1818 * (freq_curr/freq_max would rarely reach 1) and something close to freq_base,
1819 * which would ignore the entire turbo range (a conspicuous part, making
1820 * freq_curr/freq_max always maxed out).
1822 * An exception to the heuristic above is the Atom uarch, where we choose the
1823 * highest turbo level for freq_max since Atom's are generally oriented towards
1826 * Setting freq_max to anything less than the 1C turbo ratio makes the ratio
1827 * freq_curr / freq_max to eventually grow >1, in which case we clip it to 1.
1830 DEFINE_STATIC_KEY_FALSE(arch_scale_freq_key);
1832 static DEFINE_PER_CPU(u64, arch_prev_aperf);
1833 static DEFINE_PER_CPU(u64, arch_prev_mperf);
1834 static u64 arch_turbo_freq_ratio = SCHED_CAPACITY_SCALE;
1835 static u64 arch_max_freq_ratio = SCHED_CAPACITY_SCALE;
1837 void arch_set_max_freq_ratio(bool turbo_disabled)
1839 arch_max_freq_ratio = turbo_disabled ? SCHED_CAPACITY_SCALE :
1840 arch_turbo_freq_ratio;
1842 EXPORT_SYMBOL_GPL(arch_set_max_freq_ratio);
1844 static bool turbo_disabled(void)
1849 err = rdmsrl_safe(MSR_IA32_MISC_ENABLE, &misc_en);
1853 return (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
1856 static bool slv_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq)
1860 err = rdmsrl_safe(MSR_ATOM_CORE_RATIOS, base_freq);
1864 err = rdmsrl_safe(MSR_ATOM_CORE_TURBO_RATIOS, turbo_freq);
1868 *base_freq = (*base_freq >> 16) & 0x3F; /* max P state */
1869 *turbo_freq = *turbo_freq & 0x3F; /* 1C turbo */
1874 #define X86_MATCH(model) \
1875 X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, \
1876 INTEL_FAM6_##model, X86_FEATURE_APERFMPERF, NULL)
1878 static const struct x86_cpu_id has_knl_turbo_ratio_limits[] = {
1879 X86_MATCH(XEON_PHI_KNL),
1880 X86_MATCH(XEON_PHI_KNM),
1884 static const struct x86_cpu_id has_skx_turbo_ratio_limits[] = {
1885 X86_MATCH(SKYLAKE_X),
1889 static const struct x86_cpu_id has_glm_turbo_ratio_limits[] = {
1890 X86_MATCH(ATOM_GOLDMONT),
1891 X86_MATCH(ATOM_GOLDMONT_D),
1892 X86_MATCH(ATOM_GOLDMONT_PLUS),
1896 static bool knl_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq,
1897 int num_delta_fratio)
1899 int fratio, delta_fratio, found;
1903 err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
1907 *base_freq = (*base_freq >> 8) & 0xFF; /* max P state */
1909 err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
1913 fratio = (msr >> 8) & 0xFF;
1917 if (found >= num_delta_fratio) {
1918 *turbo_freq = fratio;
1922 delta_fratio = (msr >> (i + 5)) & 0x7;
1926 fratio -= delta_fratio;
1935 static bool skx_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq, int size)
1941 err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
1945 *base_freq = (*base_freq >> 8) & 0xFF; /* max P state */
1947 err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &ratios);
1951 err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT1, &counts);
1955 for (i = 0; i < 64; i += 8) {
1956 group_size = (counts >> i) & 0xFF;
1957 if (group_size >= size) {
1958 *turbo_freq = (ratios >> i) & 0xFF;
1966 static bool core_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq)
1971 err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
1975 err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
1979 *base_freq = (*base_freq >> 8) & 0xFF; /* max P state */
1980 *turbo_freq = (msr >> 24) & 0xFF; /* 4C turbo */
1982 /* The CPU may have less than 4 cores */
1984 *turbo_freq = msr & 0xFF; /* 1C turbo */
1989 static bool intel_set_max_freq_ratio(void)
1991 u64 base_freq, turbo_freq;
1994 if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
1997 if (x86_match_cpu(has_glm_turbo_ratio_limits) &&
1998 skx_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
2001 if (x86_match_cpu(has_knl_turbo_ratio_limits) &&
2002 knl_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
2005 if (x86_match_cpu(has_skx_turbo_ratio_limits) &&
2006 skx_set_max_freq_ratio(&base_freq, &turbo_freq, 4))
2009 if (core_set_max_freq_ratio(&base_freq, &turbo_freq))
2016 * Some hypervisors advertise X86_FEATURE_APERFMPERF
2017 * but then fill all MSR's with zeroes.
2018 * Some CPUs have turbo boost but don't declare any turbo ratio
2019 * in MSR_TURBO_RATIO_LIMIT.
2021 if (!base_freq || !turbo_freq) {
2022 pr_debug("Couldn't determine cpu base or turbo frequency, necessary for scale-invariant accounting.\n");
2026 turbo_ratio = div_u64(turbo_freq * SCHED_CAPACITY_SCALE, base_freq);
2028 pr_debug("Non-zero turbo and base frequencies led to a 0 ratio.\n");
2032 arch_turbo_freq_ratio = turbo_ratio;
2033 arch_set_max_freq_ratio(turbo_disabled());
2038 #ifdef CONFIG_ACPI_CPPC_LIB
2039 static bool amd_set_max_freq_ratio(void)
2041 struct cppc_perf_caps perf_caps;
2042 u64 highest_perf, nominal_perf;
2046 rc = cppc_get_perf_caps(0, &perf_caps);
2048 pr_debug("Could not retrieve perf counters (%d)\n", rc);
2052 highest_perf = amd_get_highest_perf();
2053 nominal_perf = perf_caps.nominal_perf;
2055 if (!highest_perf || !nominal_perf) {
2056 pr_debug("Could not retrieve highest or nominal performance\n");
2060 perf_ratio = div_u64(highest_perf * SCHED_CAPACITY_SCALE, nominal_perf);
2061 /* midpoint between max_boost and max_P */
2062 perf_ratio = (perf_ratio + SCHED_CAPACITY_SCALE) >> 1;
2064 pr_debug("Non-zero highest/nominal perf values led to a 0 ratio\n");
2068 arch_turbo_freq_ratio = perf_ratio;
2069 arch_set_max_freq_ratio(false);
2074 static bool amd_set_max_freq_ratio(void)
2080 static void init_counter_refs(void)
2084 rdmsrl(MSR_IA32_APERF, aperf);
2085 rdmsrl(MSR_IA32_MPERF, mperf);
2087 this_cpu_write(arch_prev_aperf, aperf);
2088 this_cpu_write(arch_prev_mperf, mperf);
2091 #ifdef CONFIG_PM_SLEEP
2092 static struct syscore_ops freq_invariance_syscore_ops = {
2093 .resume = init_counter_refs,
2096 static void register_freq_invariance_syscore_ops(void)
2098 /* Bail out if registered already. */
2099 if (freq_invariance_syscore_ops.node.prev)
2102 register_syscore_ops(&freq_invariance_syscore_ops);
2105 static inline void register_freq_invariance_syscore_ops(void) {}
2108 static void init_freq_invariance(bool secondary, bool cppc_ready)
2112 if (!boot_cpu_has(X86_FEATURE_APERFMPERF))
2116 if (static_branch_likely(&arch_scale_freq_key)) {
2117 init_counter_refs();
2122 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
2123 ret = intel_set_max_freq_ratio();
2124 else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
2128 ret = amd_set_max_freq_ratio();
2132 init_counter_refs();
2133 static_branch_enable(&arch_scale_freq_key);
2134 register_freq_invariance_syscore_ops();
2135 pr_info("Estimated ratio of average max frequency by base frequency (times 1024): %llu\n", arch_max_freq_ratio);
2137 pr_debug("Couldn't determine max cpu frequency, necessary for scale-invariant accounting.\n");
2141 #ifdef CONFIG_ACPI_CPPC_LIB
2142 static DEFINE_MUTEX(freq_invariance_lock);
2144 void init_freq_invariance_cppc(void)
2146 static bool secondary;
2148 mutex_lock(&freq_invariance_lock);
2150 init_freq_invariance(secondary, true);
2153 mutex_unlock(&freq_invariance_lock);
2157 static void disable_freq_invariance_workfn(struct work_struct *work)
2159 static_branch_disable(&arch_scale_freq_key);
2162 static DECLARE_WORK(disable_freq_invariance_work,
2163 disable_freq_invariance_workfn);
2165 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
2167 void arch_scale_freq_tick(void)
2169 u64 freq_scale = SCHED_CAPACITY_SCALE;
2173 if (!arch_scale_freq_invariant())
2176 rdmsrl(MSR_IA32_APERF, aperf);
2177 rdmsrl(MSR_IA32_MPERF, mperf);
2179 acnt = aperf - this_cpu_read(arch_prev_aperf);
2180 mcnt = mperf - this_cpu_read(arch_prev_mperf);
2182 this_cpu_write(arch_prev_aperf, aperf);
2183 this_cpu_write(arch_prev_mperf, mperf);
2185 if (check_shl_overflow(acnt, 2*SCHED_CAPACITY_SHIFT, &acnt))
2188 if (check_mul_overflow(mcnt, arch_max_freq_ratio, &mcnt) || !mcnt)
2191 freq_scale = div64_u64(acnt, mcnt);
2195 if (freq_scale > SCHED_CAPACITY_SCALE)
2196 freq_scale = SCHED_CAPACITY_SCALE;
2198 this_cpu_write(arch_freq_scale, freq_scale);
2202 pr_warn("Scheduler frequency invariance went wobbly, disabling!\n");
2203 schedule_work(&disable_freq_invariance_work);
2206 static inline void init_freq_invariance(bool secondary, bool cppc_ready)
2209 #endif /* CONFIG_X86_64 */