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/kexec.h>
57 #include <linux/numa.h>
58 #include <linux/pgtable.h>
59 #include <linux/overflow.h>
60 #include <linux/stackprotector.h>
61 #include <linux/cpuhotplug.h>
62 #include <linux/mc146818rtc.h>
65 #include <asm/cacheinfo.h>
69 #include <asm/realmode.h>
72 #include <asm/tlbflush.h>
74 #include <asm/mwait.h>
76 #include <asm/io_apic.h>
77 #include <asm/fpu/api.h>
78 #include <asm/setup.h>
79 #include <asm/uv/uv.h>
80 #include <asm/microcode.h>
81 #include <asm/i8259.h>
83 #include <asm/qspinlock.h>
84 #include <asm/intel-family.h>
85 #include <asm/cpu_device_id.h>
86 #include <asm/spec-ctrl.h>
87 #include <asm/hw_irq.h>
88 #include <asm/stackprotector.h>
90 #include <asm/spec-ctrl.h>
92 /* representing HT siblings of each logical CPU */
93 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
94 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
96 /* representing HT and core siblings of each logical CPU */
97 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
98 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
100 /* representing HT, core, and die siblings of each logical CPU */
101 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
102 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
104 /* CPUs which are the primary SMT threads */
105 struct cpumask __cpu_primary_thread_mask __read_mostly;
107 /* Representing CPUs for which sibling maps can be computed */
108 static cpumask_var_t cpu_sibling_setup_mask;
110 struct mwait_cpu_dead {
111 unsigned int control;
115 #define CPUDEAD_MWAIT_WAIT 0xDEADBEEF
116 #define CPUDEAD_MWAIT_KEXEC_HLT 0x4A17DEAD
119 * Cache line aligned data for mwait_play_dead(). Separate on purpose so
120 * that it's unlikely to be touched by other CPUs.
122 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
124 /* Maximum number of SMT threads on any online core */
125 int __read_mostly __max_smt_threads = 1;
127 /* Flag to indicate if a complete sched domain rebuild is required */
128 bool x86_topology_update;
130 int arch_update_cpu_topology(void)
132 int retval = x86_topology_update;
134 x86_topology_update = false;
138 static unsigned int smpboot_warm_reset_vector_count;
140 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
144 spin_lock_irqsave(&rtc_lock, flags);
145 if (!smpboot_warm_reset_vector_count++) {
146 CMOS_WRITE(0xa, 0xf);
147 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
148 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
150 spin_unlock_irqrestore(&rtc_lock, flags);
153 static inline void smpboot_restore_warm_reset_vector(void)
158 * Paranoid: Set warm reset code and vector here back
161 spin_lock_irqsave(&rtc_lock, flags);
162 if (!--smpboot_warm_reset_vector_count) {
164 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
166 spin_unlock_irqrestore(&rtc_lock, flags);
170 /* Run the next set of setup steps for the upcoming CPU */
171 static void ap_starting(void)
173 int cpuid = smp_processor_id();
175 /* Mop up eventual mwait_play_dead() wreckage */
176 this_cpu_write(mwait_cpu_dead.status, 0);
177 this_cpu_write(mwait_cpu_dead.control, 0);
180 * If woken up by an INIT in an 82489DX configuration the alive
181 * synchronization guarantees that the CPU does not reach this
182 * point before an INIT_deassert IPI reaches the local APIC, so it
183 * is now safe to touch the local APIC.
185 * Set up this CPU, first the APIC, which is probably redundant on
190 /* Save the processor parameters. */
191 smp_store_cpu_info(cpuid);
194 * The topology information must be up to date before
195 * notify_cpu_starting().
197 set_cpu_sibling_map(cpuid);
199 ap_init_aperfmperf();
201 pr_debug("Stack at about %p\n", &cpuid);
206 * This runs the AP through all the cpuhp states to its target
207 * state CPUHP_ONLINE.
209 notify_cpu_starting(cpuid);
212 static void ap_calibrate_delay(void)
215 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
216 * smp_store_cpu_info() stored a value that is close but not as
217 * accurate as the value just calculated.
219 * As this is invoked after the TSC synchronization check,
220 * calibrate_delay_is_known() will skip the calibration routine
221 * when TSC is synchronized across sockets.
224 cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
228 * Activate a secondary processor.
230 static void notrace start_secondary(void *unused)
233 * Don't put *anything* except direct CPU state initialization
234 * before cpu_init(), SMP booting is too fragile that we want to
235 * limit the things done here to the most necessary things.
240 * 32-bit specific. 64-bit reaches this code with the correct page
241 * table established. Yet another historical divergence.
243 if (IS_ENABLED(CONFIG_X86_32)) {
244 /* switch away from the initial page table */
245 load_cr3(swapper_pg_dir);
249 cpu_init_exception_handling();
252 * Load the microcode before reaching the AP alive synchronization
253 * point below so it is not part of the full per CPU serialized
254 * bringup part when "parallel" bringup is enabled.
256 * That's even safe when hyperthreading is enabled in the CPU as
257 * the core code starts the primary threads first and leaves the
258 * secondary threads waiting for SIPI. Loading microcode on
259 * physical cores concurrently is a safe operation.
261 * This covers both the Intel specific issue that concurrent
262 * microcode loading on SMT siblings must be prohibited and the
263 * vendor independent issue`that microcode loading which changes
264 * CPUID, MSRs etc. must be strictly serialized to maintain
265 * software state correctness.
270 * Synchronization point with the hotplug core. Sets this CPUs
271 * synchronization state to ALIVE and spin-waits for the control CPU to
272 * release this CPU for further bringup.
274 cpuhp_ap_sync_alive();
278 rcutree_report_cpu_starting(raw_smp_processor_id());
279 x86_cpuinit.early_percpu_clock_init();
283 /* Check TSC synchronization with the control CPU. */
284 check_tsc_sync_target();
287 * Calibrate the delay loop after the TSC synchronization check.
288 * This allows to skip the calibration when TSC is synchronized
291 ap_calibrate_delay();
293 speculative_store_bypass_ht_init();
296 * Lock vector_lock, set CPU online and bring the vector
297 * allocator online. Online must be set with vector_lock held
298 * to prevent a concurrent irq setup/teardown from seeing a
299 * half valid vector space.
302 set_cpu_online(smp_processor_id(), true);
304 unlock_vector_lock();
305 x86_platform.nmi_init();
307 /* enable local interrupts */
310 x86_cpuinit.setup_percpu_clockev();
313 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
317 * The bootstrap kernel entry code has set these up. Save them for
320 void smp_store_cpu_info(int id)
322 struct cpuinfo_x86 *c = &cpu_data(id);
324 /* Copy boot_cpu_data only on the first bringup */
329 * During boot time, CPU0 has this setup already. Save the info when
332 identify_secondary_cpu(c);
333 c->initialized = true;
337 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
339 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
341 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
345 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
347 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
349 return !WARN_ONCE(!topology_same_node(c, o),
350 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
351 "[node: %d != %d]. Ignoring dependency.\n",
352 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
355 #define link_mask(mfunc, c1, c2) \
357 cpumask_set_cpu((c1), mfunc(c2)); \
358 cpumask_set_cpu((c2), mfunc(c1)); \
361 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
363 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
364 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
366 if (c->topo.pkg_id == o->topo.pkg_id &&
367 c->topo.die_id == o->topo.die_id &&
368 c->topo.amd_node_id == o->topo.amd_node_id &&
369 per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) {
370 if (c->topo.core_id == o->topo.core_id)
371 return topology_sane(c, o, "smt");
373 if ((c->topo.cu_id != 0xff) &&
374 (o->topo.cu_id != 0xff) &&
375 (c->topo.cu_id == o->topo.cu_id))
376 return topology_sane(c, o, "smt");
379 } else if (c->topo.pkg_id == o->topo.pkg_id &&
380 c->topo.die_id == o->topo.die_id &&
381 c->topo.core_id == o->topo.core_id) {
382 return topology_sane(c, o, "smt");
388 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
390 if (c->topo.pkg_id != o->topo.pkg_id || c->topo.die_id != o->topo.die_id)
393 if (cpu_feature_enabled(X86_FEATURE_TOPOEXT) && topology_amd_nodes_per_pkg() > 1)
394 return c->topo.amd_node_id == o->topo.amd_node_id;
399 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
401 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
403 /* If the arch didn't set up l2c_id, fall back to SMT */
404 if (per_cpu_l2c_id(cpu1) == BAD_APICID)
405 return match_smt(c, o);
407 /* Do not match if L2 cache id does not match: */
408 if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2))
411 return topology_sane(c, o, "l2c");
415 * Unlike the other levels, we do not enforce keeping a
416 * multicore group inside a NUMA node. If this happens, we will
417 * discard the MC level of the topology later.
419 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
421 if (c->topo.pkg_id == o->topo.pkg_id)
427 * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
429 * Any Intel CPU that has multiple nodes per package and does not
430 * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
432 * When in SNC mode, these CPUs enumerate an LLC that is shared
433 * by multiple NUMA nodes. The LLC is shared for off-package data
434 * access but private to the NUMA node (half of the package) for
435 * on-package access. CPUID (the source of the information about
436 * the LLC) can only enumerate the cache as shared or unshared,
437 * but not this particular configuration.
440 static const struct x86_cpu_id intel_cod_cpu[] = {
441 X86_MATCH_VFM(INTEL_HASWELL_X, 0), /* COD */
442 X86_MATCH_VFM(INTEL_BROADWELL_X, 0), /* COD */
443 X86_MATCH_VFM(INTEL_ANY, 1), /* SNC */
447 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
449 const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
450 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
451 bool intel_snc = id && id->driver_data;
453 /* Do not match if we do not have a valid APICID for cpu: */
454 if (per_cpu_llc_id(cpu1) == BAD_APICID)
457 /* Do not match if LLC id does not match: */
458 if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2))
462 * Allow the SNC topology without warning. Return of false
463 * means 'c' does not share the LLC of 'o'. This will be
464 * reflected to userspace.
466 if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
469 return topology_sane(c, o, "llc");
473 static inline int x86_sched_itmt_flags(void)
475 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
478 #ifdef CONFIG_SCHED_MC
479 static int x86_core_flags(void)
481 return cpu_core_flags() | x86_sched_itmt_flags();
484 #ifdef CONFIG_SCHED_SMT
485 static int x86_smt_flags(void)
487 return cpu_smt_flags();
490 #ifdef CONFIG_SCHED_CLUSTER
491 static int x86_cluster_flags(void)
493 return cpu_cluster_flags() | x86_sched_itmt_flags();
497 static int x86_die_flags(void)
499 if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
500 return x86_sched_itmt_flags();
506 * Set if a package/die has multiple NUMA nodes inside.
507 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
508 * Sub-NUMA Clustering have this.
510 static bool x86_has_numa_in_package;
512 static struct sched_domain_topology_level x86_topology[6];
514 static void __init build_sched_topology(void)
518 #ifdef CONFIG_SCHED_SMT
519 x86_topology[i++] = (struct sched_domain_topology_level){
520 cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT)
523 #ifdef CONFIG_SCHED_CLUSTER
524 x86_topology[i++] = (struct sched_domain_topology_level){
525 cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS)
528 #ifdef CONFIG_SCHED_MC
529 x86_topology[i++] = (struct sched_domain_topology_level){
530 cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC)
534 * When there is NUMA topology inside the package skip the PKG domain
535 * since the NUMA domains will auto-magically create the right spanning
536 * domains based on the SLIT.
538 if (!x86_has_numa_in_package) {
539 x86_topology[i++] = (struct sched_domain_topology_level){
540 cpu_cpu_mask, x86_die_flags, SD_INIT_NAME(PKG)
545 * There must be one trailing NULL entry left.
547 BUG_ON(i >= ARRAY_SIZE(x86_topology)-1);
549 set_sched_topology(x86_topology);
552 void set_cpu_sibling_map(int cpu)
554 bool has_smt = __max_threads_per_core > 1;
555 bool has_mp = has_smt || topology_num_cores_per_package() > 1;
556 struct cpuinfo_x86 *c = &cpu_data(cpu);
557 struct cpuinfo_x86 *o;
560 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
563 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
564 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
565 cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
566 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
567 cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
572 for_each_cpu(i, cpu_sibling_setup_mask) {
575 if (match_pkg(c, o) && !topology_same_node(c, o))
576 x86_has_numa_in_package = true;
578 if ((i == cpu) || (has_smt && match_smt(c, o)))
579 link_mask(topology_sibling_cpumask, cpu, i);
581 if ((i == cpu) || (has_mp && match_llc(c, o)))
582 link_mask(cpu_llc_shared_mask, cpu, i);
584 if ((i == cpu) || (has_mp && match_l2c(c, o)))
585 link_mask(cpu_l2c_shared_mask, cpu, i);
587 if ((i == cpu) || (has_mp && match_die(c, o)))
588 link_mask(topology_die_cpumask, cpu, i);
591 threads = cpumask_weight(topology_sibling_cpumask(cpu));
592 if (threads > __max_smt_threads)
593 __max_smt_threads = threads;
595 for_each_cpu(i, topology_sibling_cpumask(cpu))
596 cpu_data(i).smt_active = threads > 1;
599 * This needs a separate iteration over the cpus because we rely on all
600 * topology_sibling_cpumask links to be set-up.
602 for_each_cpu(i, cpu_sibling_setup_mask) {
605 if ((i == cpu) || (has_mp && match_pkg(c, o))) {
606 link_mask(topology_core_cpumask, cpu, i);
609 * Does this new cpu bringup a new core?
613 * for each core in package, increment
614 * the booted_cores for this new cpu
617 topology_sibling_cpumask(i)) == i)
620 * increment the core count for all
621 * the other cpus in this package
624 cpu_data(i).booted_cores++;
625 } else if (i != cpu && !c->booted_cores)
626 c->booted_cores = cpu_data(i).booted_cores;
631 /* maps the cpu to the sched domain representing multi-core */
632 const struct cpumask *cpu_coregroup_mask(int cpu)
634 return cpu_llc_shared_mask(cpu);
637 const struct cpumask *cpu_clustergroup_mask(int cpu)
639 return cpu_l2c_shared_mask(cpu);
641 EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
643 static void impress_friends(void)
646 unsigned long bogosum = 0;
648 * Allow the user to impress friends.
650 pr_debug("Before bogomips\n");
651 for_each_online_cpu(cpu)
652 bogosum += cpu_data(cpu).loops_per_jiffy;
654 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
657 (bogosum/(5000/HZ))%100);
659 pr_debug("Before bogocount - setting activated=1\n");
663 * The Multiprocessor Specification 1.4 (1997) example code suggests
664 * that there should be a 10ms delay between the BSP asserting INIT
665 * and de-asserting INIT, when starting a remote processor.
666 * But that slows boot and resume on modern processors, which include
667 * many cores and don't require that delay.
669 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
670 * Modern processor families are quirked to remove the delay entirely.
672 #define UDELAY_10MS_DEFAULT 10000
674 static unsigned int init_udelay = UINT_MAX;
676 static int __init cpu_init_udelay(char *str)
678 get_option(&str, &init_udelay);
682 early_param("cpu_init_udelay", cpu_init_udelay);
684 static void __init smp_quirk_init_udelay(void)
686 /* if cmdline changed it from default, leave it alone */
687 if (init_udelay != UINT_MAX)
690 /* if modern processor, use no delay */
691 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
692 ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
693 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
697 /* else, use legacy delay */
698 init_udelay = UDELAY_10MS_DEFAULT;
702 * Wake up AP by INIT, INIT, STARTUP sequence.
704 static void send_init_sequence(u32 phys_apicid)
706 int maxlvt = lapic_get_maxlvt();
708 /* Be paranoid about clearing APIC errors. */
709 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
710 /* Due to the Pentium erratum 3AP. */
712 apic_write(APIC_ESR, 0);
716 /* Assert INIT on the target CPU */
717 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
718 safe_apic_wait_icr_idle();
722 /* Deassert INIT on the target CPU */
723 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
724 safe_apic_wait_icr_idle();
728 * Wake up AP by INIT, INIT, STARTUP sequence.
730 static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip)
732 unsigned long send_status = 0, accept_status = 0;
733 int num_starts, j, maxlvt;
736 maxlvt = lapic_get_maxlvt();
737 send_init_sequence(phys_apicid);
742 * Should we send STARTUP IPIs ?
744 * Determine this based on the APIC version.
745 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
747 if (APIC_INTEGRATED(boot_cpu_apic_version))
753 * Run STARTUP IPI loop.
755 pr_debug("#startup loops: %d\n", num_starts);
757 for (j = 1; j <= num_starts; j++) {
758 pr_debug("Sending STARTUP #%d\n", j);
759 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
760 apic_write(APIC_ESR, 0);
762 pr_debug("After apic_write\n");
769 /* Boot on the stack */
770 /* Kick the second */
771 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
775 * Give the other CPU some time to accept the IPI.
777 if (init_udelay == 0)
782 pr_debug("Startup point 1\n");
784 pr_debug("Waiting for send to finish...\n");
785 send_status = safe_apic_wait_icr_idle();
788 * Give the other CPU some time to accept the IPI.
790 if (init_udelay == 0)
795 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
796 apic_write(APIC_ESR, 0);
797 accept_status = (apic_read(APIC_ESR) & 0xEF);
798 if (send_status || accept_status)
801 pr_debug("After Startup\n");
804 pr_err("APIC never delivered???\n");
806 pr_err("APIC delivery error (%lx)\n", accept_status);
809 return (send_status | accept_status);
812 /* reduce the number of lines printed when booting a large cpu count system */
813 static void announce_cpu(int cpu, int apicid)
815 static int width, node_width, first = 1;
816 static int current_node = NUMA_NO_NODE;
817 int node = early_cpu_to_node(cpu);
820 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
823 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
825 if (system_state < SYSTEM_RUNNING) {
827 pr_info("x86: Booting SMP configuration:\n");
829 if (node != current_node) {
830 if (current_node > (-1))
834 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
835 node_width - num_digits(node), " ", node);
838 /* Add padding for the BSP */
840 pr_cont("%*s", width + 1, " ");
843 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
845 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
849 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
853 /* Just in case we booted with a single CPU. */
854 alternatives_enable_smp();
856 per_cpu(pcpu_hot.current_task, cpu) = idle;
857 cpu_init_stack_canary(cpu, idle);
859 /* Initialize the interrupt stack(s) */
860 ret = irq_init_percpu_irqstack(cpu);
865 /* Stack for startup_32 can be just as for start_secondary onwards */
866 per_cpu(pcpu_hot.top_of_stack, cpu) = task_top_of_stack(idle);
872 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
873 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
874 * Returns zero if startup was successfully sent, else error code from
875 * ->wakeup_secondary_cpu.
877 static int do_boot_cpu(u32 apicid, int cpu, struct task_struct *idle)
879 unsigned long start_ip = real_mode_header->trampoline_start;
883 /* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
884 if (apic->wakeup_secondary_cpu_64)
885 start_ip = real_mode_header->trampoline_start64;
887 idle->thread.sp = (unsigned long)task_pt_regs(idle);
888 initial_code = (unsigned long)start_secondary;
890 if (IS_ENABLED(CONFIG_X86_32)) {
891 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
892 initial_stack = idle->thread.sp;
893 } else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
894 smpboot_control = cpu;
897 /* Enable the espfix hack for this CPU */
900 /* So we see what's up */
901 announce_cpu(cpu, apicid);
904 * This grunge runs the startup process for
905 * the targeted processor.
907 if (x86_platform.legacy.warm_reset) {
909 pr_debug("Setting warm reset code and vector.\n");
911 smpboot_setup_warm_reset_vector(start_ip);
913 * Be paranoid about clearing APIC errors.
915 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
916 apic_write(APIC_ESR, 0);
924 * Wake up a CPU in difference cases:
925 * - Use a method from the APIC driver if one defined, with wakeup
926 * straight to 64-bit mode preferred over wakeup to RM.
928 * - Use an INIT boot APIC message
930 if (apic->wakeup_secondary_cpu_64)
931 ret = apic->wakeup_secondary_cpu_64(apicid, start_ip);
932 else if (apic->wakeup_secondary_cpu)
933 ret = apic->wakeup_secondary_cpu(apicid, start_ip);
935 ret = wakeup_secondary_cpu_via_init(apicid, start_ip);
937 /* If the wakeup mechanism failed, cleanup the warm reset vector */
939 arch_cpuhp_cleanup_kick_cpu(cpu);
943 int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
945 u32 apicid = apic->cpu_present_to_apicid(cpu);
948 lockdep_assert_irqs_enabled();
950 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
952 if (apicid == BAD_APICID || !apic_id_valid(apicid)) {
953 pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid);
957 if (!test_bit(apicid, phys_cpu_present_map)) {
958 pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid);
963 * Save current MTRR state in case it was changed since early boot
964 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
968 /* the FPU context is blank, nobody can own it */
969 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
971 err = common_cpu_up(cpu, tidle);
975 err = do_boot_cpu(apicid, cpu, tidle);
977 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
982 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
984 return smp_ops.kick_ap_alive(cpu, tidle);
987 void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
989 /* Cleanup possible dangling ends... */
990 if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
991 smpboot_restore_warm_reset_vector();
994 void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
996 if (smp_ops.cleanup_dead_cpu)
997 smp_ops.cleanup_dead_cpu(cpu);
999 if (system_state == SYSTEM_RUNNING)
1000 pr_info("CPU %u is now offline\n", cpu);
1003 void arch_cpuhp_sync_state_poll(void)
1005 if (smp_ops.poll_sync_state)
1006 smp_ops.poll_sync_state();
1010 * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1012 void __init arch_disable_smp_support(void)
1014 disable_ioapic_support();
1018 * Fall back to non SMP mode after errors.
1020 * RED-PEN audit/test this more. I bet there is more state messed up here.
1022 static __init void disable_smp(void)
1024 pr_info("SMP disabled\n");
1026 disable_ioapic_support();
1027 topology_reset_possible_cpus_up();
1029 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1030 cpumask_set_cpu(0, topology_core_cpumask(0));
1031 cpumask_set_cpu(0, topology_die_cpumask(0));
1034 void __init smp_prepare_cpus_common(void)
1036 unsigned int cpu, node;
1038 /* Mark all except the boot CPU as hotpluggable */
1039 for_each_possible_cpu(cpu) {
1041 per_cpu(cpu_info.cpu_index, cpu) = nr_cpu_ids;
1044 for_each_possible_cpu(cpu) {
1045 node = cpu_to_node(cpu);
1047 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu), GFP_KERNEL, node);
1048 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu), GFP_KERNEL, node);
1049 zalloc_cpumask_var_node(&per_cpu(cpu_die_map, cpu), GFP_KERNEL, node);
1050 zalloc_cpumask_var_node(&per_cpu(cpu_llc_shared_map, cpu), GFP_KERNEL, node);
1051 zalloc_cpumask_var_node(&per_cpu(cpu_l2c_shared_map, cpu), GFP_KERNEL, node);
1054 set_cpu_sibling_map(0);
1057 void __init smp_prepare_boot_cpu(void)
1059 smp_ops.smp_prepare_boot_cpu();
1062 #ifdef CONFIG_X86_64
1063 /* Establish whether parallel bringup can be supported. */
1064 bool __init arch_cpuhp_init_parallel_bringup(void)
1066 if (!x86_cpuinit.parallel_bringup) {
1067 pr_info("Parallel CPU startup disabled by the platform\n");
1071 smpboot_control = STARTUP_READ_APICID;
1072 pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1078 * Prepare for SMP bootup.
1079 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1080 * for common interface support.
1082 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1084 smp_prepare_cpus_common();
1086 switch (apic_intr_mode) {
1088 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1091 case APIC_SYMMETRIC_IO_NO_ROUTING:
1093 /* Setup local timer */
1094 x86_init.timers.setup_percpu_clockev();
1096 case APIC_VIRTUAL_WIRE:
1097 case APIC_SYMMETRIC_IO:
1101 /* Setup local timer */
1102 x86_init.timers.setup_percpu_clockev();
1105 print_cpu_info(&cpu_data(0));
1109 smp_quirk_init_udelay();
1111 speculative_store_bypass_ht_init();
1113 snp_set_wakeup_secondary_cpu();
1116 void arch_thaw_secondary_cpus_begin(void)
1118 set_cache_aps_delayed_init(true);
1121 void arch_thaw_secondary_cpus_end(void)
1127 * Early setup to make printk work.
1129 void __init native_smp_prepare_boot_cpu(void)
1131 int me = smp_processor_id();
1133 /* SMP handles this from setup_per_cpu_areas() */
1134 if (!IS_ENABLED(CONFIG_SMP))
1135 switch_gdt_and_percpu_base(me);
1137 native_pv_lock_init();
1140 void __init native_smp_cpus_done(unsigned int max_cpus)
1142 pr_debug("Boot done\n");
1144 build_sched_topology();
1150 /* correctly size the local cpu masks */
1151 void __init setup_cpu_local_masks(void)
1153 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1156 #ifdef CONFIG_HOTPLUG_CPU
1158 /* Recompute SMT state for all CPUs on offline */
1159 static void recompute_smt_state(void)
1161 int max_threads, cpu;
1164 for_each_online_cpu (cpu) {
1165 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1167 if (threads > max_threads)
1168 max_threads = threads;
1170 __max_smt_threads = max_threads;
1173 static void remove_siblinginfo(int cpu)
1176 struct cpuinfo_x86 *c = &cpu_data(cpu);
1178 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1179 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1181 * last thread sibling in this cpu core going down
1183 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1184 cpu_data(sibling).booted_cores--;
1187 for_each_cpu(sibling, topology_die_cpumask(cpu))
1188 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1190 for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1191 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1192 if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1193 cpu_data(sibling).smt_active = false;
1196 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1197 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1198 for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1199 cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1200 cpumask_clear(cpu_llc_shared_mask(cpu));
1201 cpumask_clear(cpu_l2c_shared_mask(cpu));
1202 cpumask_clear(topology_sibling_cpumask(cpu));
1203 cpumask_clear(topology_core_cpumask(cpu));
1204 cpumask_clear(topology_die_cpumask(cpu));
1205 c->topo.core_id = 0;
1206 c->booted_cores = 0;
1207 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1208 recompute_smt_state();
1211 static void remove_cpu_from_maps(int cpu)
1213 set_cpu_online(cpu, false);
1214 numa_remove_cpu(cpu);
1217 void cpu_disable_common(void)
1219 int cpu = smp_processor_id();
1221 remove_siblinginfo(cpu);
1223 /* It's now safe to remove this processor from the online map */
1225 remove_cpu_from_maps(cpu);
1226 unlock_vector_lock();
1231 int native_cpu_disable(void)
1235 ret = lapic_can_unplug_cpu();
1239 cpu_disable_common();
1242 * Disable the local APIC. Otherwise IPI broadcasts will reach
1243 * it. It still responds normally to INIT, NMI, SMI, and SIPI
1246 * Disabling the APIC must happen after cpu_disable_common()
1247 * which invokes fixup_irqs().
1249 * Disabling the APIC preserves already set bits in IRR, but
1250 * an interrupt arriving after disabling the local APIC does not
1251 * set the corresponding IRR bit.
1253 * fixup_irqs() scans IRR for set bits so it can raise a not
1254 * yet handled interrupt on the new destination CPU via an IPI
1255 * but obviously it can't do so for IRR bits which are not set.
1256 * IOW, interrupts arriving after disabling the local APIC will
1259 apic_soft_disable();
1264 void play_dead_common(void)
1268 cpuhp_ap_report_dead();
1270 local_irq_disable();
1274 * We need to flush the caches before going to sleep, lest we have
1275 * dirty data in our caches when we come back up.
1277 static inline void mwait_play_dead(void)
1279 struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1280 unsigned int eax, ebx, ecx, edx;
1281 unsigned int highest_cstate = 0;
1282 unsigned int highest_subcstate = 0;
1285 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1286 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1288 if (!this_cpu_has(X86_FEATURE_MWAIT))
1290 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1292 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1295 eax = CPUID_MWAIT_LEAF;
1297 native_cpuid(&eax, &ebx, &ecx, &edx);
1300 * eax will be 0 if EDX enumeration is not valid.
1301 * Initialized below to cstate, sub_cstate value when EDX is valid.
1303 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1306 edx >>= MWAIT_SUBSTATE_SIZE;
1307 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1308 if (edx & MWAIT_SUBSTATE_MASK) {
1310 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1313 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1314 (highest_subcstate - 1);
1317 /* Set up state for the kexec() hack below */
1318 md->status = CPUDEAD_MWAIT_WAIT;
1319 md->control = CPUDEAD_MWAIT_WAIT;
1325 * The CLFLUSH is a workaround for erratum AAI65 for
1326 * the Xeon 7400 series. It's not clear it is actually
1327 * needed, but it should be harmless in either case.
1328 * The WBINVD is insufficient due to the spurious-wakeup
1329 * case where we return around the loop.
1334 __monitor(md, 0, 0);
1338 if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1340 * Kexec is about to happen. Don't go back into mwait() as
1341 * the kexec kernel might overwrite text and data including
1342 * page tables and stack. So mwait() would resume when the
1343 * monitor cache line is written to and then the CPU goes
1344 * south due to overwritten text, page tables and stack.
1346 * Note: This does _NOT_ protect against a stray MCE, NMI,
1347 * SMI. They will resume execution at the instruction
1348 * following the HLT instruction and run into the problem
1349 * which this is trying to prevent.
1351 WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1359 * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1360 * mwait_play_dead().
1362 void smp_kick_mwait_play_dead(void)
1364 u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1365 struct mwait_cpu_dead *md;
1366 unsigned int cpu, i;
1368 for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1369 md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1371 /* Does it sit in mwait_play_dead() ? */
1372 if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1375 /* Wait up to 5ms */
1376 for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1377 /* Bring it out of mwait */
1378 WRITE_ONCE(md->control, newstate);
1382 if (READ_ONCE(md->status) != newstate)
1383 pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1387 void __noreturn hlt_play_dead(void)
1389 if (__this_cpu_read(cpu_info.x86) >= 4)
1397 * native_play_dead() is essentially a __noreturn function, but it can't
1398 * be marked as such as the compiler may complain about it.
1400 void native_play_dead(void)
1402 if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1403 __update_spec_ctrl(0);
1406 tboot_shutdown(TB_SHUTDOWN_WFS);
1409 if (cpuidle_play_dead())
1413 #else /* ... !CONFIG_HOTPLUG_CPU */
1414 int native_cpu_disable(void)
1419 void native_play_dead(void)