2 * x86 SMP booting functions
4 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6 * Copyright 2001 Andi Kleen, SuSE Labs.
8 * Much of the core SMP work is based on previous work by Thomas Radke, to
9 * whom a great many thanks are extended.
11 * Thanks to Intel for making available several different Pentium,
12 * Pentium Pro and Pentium-II/Xeon MP machines.
13 * Original development of Linux SMP code supported by Caldera.
15 * This code is released under the GNU General Public License version 2 or
19 * Felix Koop : NR_CPUS used properly
20 * Jose Renau : Handle single CPU case.
21 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
22 * Greg Wright : Fix for kernel stacks panic.
23 * Erich Boleyn : MP v1.4 and additional changes.
24 * Matthias Sattler : Changes for 2.1 kernel map.
25 * Michel Lespinasse : Changes for 2.1 kernel map.
26 * Michael Chastain : Change trampoline.S to gnu as.
27 * Alan Cox : Dumb bug: 'B' step PPro's are fine
28 * Ingo Molnar : Added APIC timers, based on code
30 * Ingo Molnar : various cleanups and rewrites
31 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
32 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
33 * Andi Kleen : Changed for SMP boot into long mode.
34 * Martin J. Bligh : Added support for multi-quad systems
35 * Dave Jones : Report invalid combinations of Athlon CPUs.
36 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
37 * Andi Kleen : Converted to new state machine.
38 * Ashok Raj : CPU hotplug support
39 * Glauber Costa : i386 and x86_64 integration
42 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
44 #include <linux/init.h>
45 #include <linux/smp.h>
46 #include <linux/module.h>
47 #include <linux/sched.h>
48 #include <linux/percpu.h>
49 #include <linux/bootmem.h>
50 #include <linux/err.h>
51 #include <linux/nmi.h>
52 #include <linux/tboot.h>
53 #include <linux/stackprotector.h>
54 #include <linux/gfp.h>
55 #include <linux/cpuidle.h>
62 #include <asm/realmode.h>
65 #include <asm/pgtable.h>
66 #include <asm/tlbflush.h>
68 #include <asm/mwait.h>
70 #include <asm/io_apic.h>
71 #include <asm/fpu/internal.h>
72 #include <asm/setup.h>
73 #include <asm/uv/uv.h>
74 #include <linux/mc146818rtc.h>
75 #include <asm/i8259.h>
76 #include <asm/realmode.h>
79 /* Number of siblings per CPU package */
80 int smp_num_siblings
= 1;
81 EXPORT_SYMBOL(smp_num_siblings
);
83 /* Last level cache ID of each logical CPU */
84 DEFINE_PER_CPU_READ_MOSTLY(u16
, cpu_llc_id
) = BAD_APICID
;
86 /* representing HT siblings of each logical CPU */
87 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t
, cpu_sibling_map
);
88 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map
);
90 /* representing HT and core siblings of each logical CPU */
91 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t
, cpu_core_map
);
92 EXPORT_PER_CPU_SYMBOL(cpu_core_map
);
94 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t
, cpu_llc_shared_map
);
96 /* Per CPU bogomips and other parameters */
97 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86
, cpu_info
);
98 EXPORT_PER_CPU_SYMBOL(cpu_info
);
100 /* Logical package management. We might want to allocate that dynamically */
101 static int *physical_to_logical_pkg __read_mostly
;
102 static unsigned long *physical_package_map __read_mostly
;;
103 static unsigned long *logical_package_map __read_mostly
;
104 static unsigned int max_physical_pkg_id __read_mostly
;
105 unsigned int __max_logical_packages __read_mostly
;
106 EXPORT_SYMBOL(__max_logical_packages
);
108 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip
)
112 spin_lock_irqsave(&rtc_lock
, flags
);
113 CMOS_WRITE(0xa, 0xf);
114 spin_unlock_irqrestore(&rtc_lock
, flags
);
117 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH
)) =
120 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW
)) =
125 static inline void smpboot_restore_warm_reset_vector(void)
130 * Install writable page 0 entry to set BIOS data area.
135 * Paranoid: Set warm reset code and vector here back
138 spin_lock_irqsave(&rtc_lock
, flags
);
140 spin_unlock_irqrestore(&rtc_lock
, flags
);
142 *((volatile u32
*)phys_to_virt(TRAMPOLINE_PHYS_LOW
)) = 0;
146 * Report back to the Boot Processor during boot time or to the caller processor
149 static void smp_callin(void)
154 * If waken up by an INIT in an 82489DX configuration
155 * cpu_callout_mask guarantees we don't get here before
156 * an INIT_deassert IPI reaches our local APIC, so it is
157 * now safe to touch our local APIC.
159 cpuid
= smp_processor_id();
162 * (This works even if the APIC is not enabled.)
164 phys_id
= read_apic_id();
167 * the boot CPU has finished the init stage and is spinning
168 * on callin_map until we finish. We are free to set up this
169 * CPU, first the APIC. (this is probably redundant on most
175 * Save our processor parameters. Note: this information
176 * is needed for clock calibration.
178 smp_store_cpu_info(cpuid
);
182 * Update loops_per_jiffy in cpu_data. Previous call to
183 * smp_store_cpu_info() stored a value that is close but not as
184 * accurate as the value just calculated.
187 cpu_data(cpuid
).loops_per_jiffy
= loops_per_jiffy
;
188 pr_debug("Stack at about %p\n", &cpuid
);
191 * This must be done before setting cpu_online_mask
192 * or calling notify_cpu_starting.
194 set_cpu_sibling_map(raw_smp_processor_id());
197 notify_cpu_starting(cpuid
);
200 * Allow the master to continue.
202 cpumask_set_cpu(cpuid
, cpu_callin_mask
);
205 static int cpu0_logical_apicid
;
206 static int enable_start_cpu0
;
208 * Activate a secondary processor.
210 static void notrace
start_secondary(void *unused
)
213 * Don't put *anything* before cpu_init(), SMP booting is too
214 * fragile that we want to limit the things done here to the
215 * most necessary things.
218 x86_cpuinit
.early_percpu_clock_init();
222 enable_start_cpu0
= 0;
225 /* switch away from the initial page table */
226 load_cr3(swapper_pg_dir
);
230 /* otherwise gcc will move up smp_processor_id before the cpu_init */
233 * Check TSC synchronization with the BP:
235 check_tsc_sync_target();
238 * Lock vector_lock and initialize the vectors on this cpu
239 * before setting the cpu online. We must set it online with
240 * vector_lock held to prevent a concurrent setup/teardown
241 * from seeing a half valid vector space.
244 setup_vector_irq(smp_processor_id());
245 set_cpu_online(smp_processor_id(), true);
246 unlock_vector_lock();
247 cpu_set_state_online(smp_processor_id());
248 x86_platform
.nmi_init();
250 /* enable local interrupts */
253 /* to prevent fake stack check failure in clock setup */
254 boot_init_stack_canary();
256 x86_cpuinit
.setup_percpu_clockev();
259 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE
);
262 int topology_update_package_map(unsigned int apicid
, unsigned int cpu
)
264 unsigned int new, pkg
= apicid
>> boot_cpu_data
.x86_coreid_bits
;
266 /* Called from early boot ? */
267 if (!physical_package_map
)
270 if (pkg
>= max_physical_pkg_id
)
273 /* Set the logical package id */
274 if (test_and_set_bit(pkg
, physical_package_map
))
277 new = find_first_zero_bit(logical_package_map
, __max_logical_packages
);
278 if (new >= __max_logical_packages
) {
279 physical_to_logical_pkg
[pkg
] = -1;
280 pr_warn("APIC(%x) Package %u exceeds logical package map\n",
284 set_bit(new, logical_package_map
);
285 pr_info("APIC(%x) Converting physical %u to logical package %u\n",
287 physical_to_logical_pkg
[pkg
] = new;
290 cpu_data(cpu
).logical_proc_id
= physical_to_logical_pkg
[pkg
];
295 * topology_phys_to_logical_pkg - Map a physical package id to a logical
297 * Returns logical package id or -1 if not found
299 int topology_phys_to_logical_pkg(unsigned int phys_pkg
)
301 if (phys_pkg
>= max_physical_pkg_id
)
303 return physical_to_logical_pkg
[phys_pkg
];
305 EXPORT_SYMBOL(topology_phys_to_logical_pkg
);
307 static void __init
smp_init_package_map(void)
309 unsigned int ncpus
, cpu
;
313 * Today neither Intel nor AMD support heterogenous systems. That
314 * might change in the future....
316 * While ideally we'd want '* smp_num_siblings' in the below @ncpus
317 * computation, this won't actually work since some Intel BIOSes
318 * report inconsistent HT data when they disable HT.
320 * In particular, they reduce the APIC-IDs to only include the cores,
321 * but leave the CPUID topology to say there are (2) siblings.
322 * This means we don't know how many threads there will be until
323 * after the APIC enumeration.
325 * By not including this we'll sometimes over-estimate the number of
326 * logical packages by the amount of !present siblings, but this is
327 * still better than MAX_LOCAL_APIC.
329 * We use total_cpus not nr_cpu_ids because nr_cpu_ids can be limited
330 * on the command line leading to a similar issue as the HT disable
331 * problem because the hyperthreads are usually enumerated after the
334 ncpus
= boot_cpu_data
.x86_max_cores
;
336 pr_warn("x86_max_cores == zero !?!?");
340 __max_logical_packages
= DIV_ROUND_UP(total_cpus
, ncpus
);
343 * Possibly larger than what we need as the number of apic ids per
344 * package can be smaller than the actual used apic ids.
346 max_physical_pkg_id
= DIV_ROUND_UP(MAX_LOCAL_APIC
, ncpus
);
347 size
= max_physical_pkg_id
* sizeof(unsigned int);
348 physical_to_logical_pkg
= kmalloc(size
, GFP_KERNEL
);
349 memset(physical_to_logical_pkg
, 0xff, size
);
350 size
= BITS_TO_LONGS(max_physical_pkg_id
) * sizeof(unsigned long);
351 physical_package_map
= kzalloc(size
, GFP_KERNEL
);
352 size
= BITS_TO_LONGS(__max_logical_packages
) * sizeof(unsigned long);
353 logical_package_map
= kzalloc(size
, GFP_KERNEL
);
355 pr_info("Max logical packages: %u\n", __max_logical_packages
);
357 for_each_present_cpu(cpu
) {
358 unsigned int apicid
= apic
->cpu_present_to_apicid(cpu
);
360 if (apicid
== BAD_APICID
|| !apic
->apic_id_valid(apicid
))
362 if (!topology_update_package_map(apicid
, cpu
))
364 pr_warn("CPU %u APICId %x disabled\n", cpu
, apicid
);
365 per_cpu(x86_bios_cpu_apicid
, cpu
) = BAD_APICID
;
366 set_cpu_possible(cpu
, false);
367 set_cpu_present(cpu
, false);
371 void __init
smp_store_boot_cpu_info(void)
373 int id
= 0; /* CPU 0 */
374 struct cpuinfo_x86
*c
= &cpu_data(id
);
378 smp_init_package_map();
382 * The bootstrap kernel entry code has set these up. Save them for
385 void smp_store_cpu_info(int id
)
387 struct cpuinfo_x86
*c
= &cpu_data(id
);
392 * During boot time, CPU0 has this setup already. Save the info when
393 * bringing up AP or offlined CPU0.
395 identify_secondary_cpu(c
);
399 topology_same_node(struct cpuinfo_x86
*c
, struct cpuinfo_x86
*o
)
401 int cpu1
= c
->cpu_index
, cpu2
= o
->cpu_index
;
403 return (cpu_to_node(cpu1
) == cpu_to_node(cpu2
));
407 topology_sane(struct cpuinfo_x86
*c
, struct cpuinfo_x86
*o
, const char *name
)
409 int cpu1
= c
->cpu_index
, cpu2
= o
->cpu_index
;
411 return !WARN_ONCE(!topology_same_node(c
, o
),
412 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
413 "[node: %d != %d]. Ignoring dependency.\n",
414 cpu1
, name
, cpu2
, cpu_to_node(cpu1
), cpu_to_node(cpu2
));
417 #define link_mask(mfunc, c1, c2) \
419 cpumask_set_cpu((c1), mfunc(c2)); \
420 cpumask_set_cpu((c2), mfunc(c1)); \
423 static bool match_smt(struct cpuinfo_x86
*c
, struct cpuinfo_x86
*o
)
425 if (boot_cpu_has(X86_FEATURE_TOPOEXT
)) {
426 int cpu1
= c
->cpu_index
, cpu2
= o
->cpu_index
;
428 if (c
->phys_proc_id
== o
->phys_proc_id
&&
429 per_cpu(cpu_llc_id
, cpu1
) == per_cpu(cpu_llc_id
, cpu2
) &&
430 c
->cpu_core_id
== o
->cpu_core_id
)
431 return topology_sane(c
, o
, "smt");
433 } else if (c
->phys_proc_id
== o
->phys_proc_id
&&
434 c
->cpu_core_id
== o
->cpu_core_id
) {
435 return topology_sane(c
, o
, "smt");
441 static bool match_llc(struct cpuinfo_x86
*c
, struct cpuinfo_x86
*o
)
443 int cpu1
= c
->cpu_index
, cpu2
= o
->cpu_index
;
445 if (per_cpu(cpu_llc_id
, cpu1
) != BAD_APICID
&&
446 per_cpu(cpu_llc_id
, cpu1
) == per_cpu(cpu_llc_id
, cpu2
))
447 return topology_sane(c
, o
, "llc");
453 * Unlike the other levels, we do not enforce keeping a
454 * multicore group inside a NUMA node. If this happens, we will
455 * discard the MC level of the topology later.
457 static bool match_die(struct cpuinfo_x86
*c
, struct cpuinfo_x86
*o
)
459 if (c
->phys_proc_id
== o
->phys_proc_id
)
464 static struct sched_domain_topology_level numa_inside_package_topology
[] = {
465 #ifdef CONFIG_SCHED_SMT
466 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
468 #ifdef CONFIG_SCHED_MC
469 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
474 * set_sched_topology() sets the topology internal to a CPU. The
475 * NUMA topologies are layered on top of it to build the full
478 * If NUMA nodes are observed to occur within a CPU package, this
479 * function should be called. It forces the sched domain code to
480 * only use the SMT level for the CPU portion of the topology.
481 * This essentially falls back to relying on NUMA information
482 * from the SRAT table to describe the entire system topology
483 * (except for hyperthreads).
485 static void primarily_use_numa_for_topology(void)
487 set_sched_topology(numa_inside_package_topology
);
490 void set_cpu_sibling_map(int cpu
)
492 bool has_smt
= smp_num_siblings
> 1;
493 bool has_mp
= has_smt
|| boot_cpu_data
.x86_max_cores
> 1;
494 struct cpuinfo_x86
*c
= &cpu_data(cpu
);
495 struct cpuinfo_x86
*o
;
498 cpumask_set_cpu(cpu
, cpu_sibling_setup_mask
);
501 cpumask_set_cpu(cpu
, topology_sibling_cpumask(cpu
));
502 cpumask_set_cpu(cpu
, cpu_llc_shared_mask(cpu
));
503 cpumask_set_cpu(cpu
, topology_core_cpumask(cpu
));
508 for_each_cpu(i
, cpu_sibling_setup_mask
) {
511 if ((i
== cpu
) || (has_smt
&& match_smt(c
, o
)))
512 link_mask(topology_sibling_cpumask
, cpu
, i
);
514 if ((i
== cpu
) || (has_mp
&& match_llc(c
, o
)))
515 link_mask(cpu_llc_shared_mask
, cpu
, i
);
520 * This needs a separate iteration over the cpus because we rely on all
521 * topology_sibling_cpumask links to be set-up.
523 for_each_cpu(i
, cpu_sibling_setup_mask
) {
526 if ((i
== cpu
) || (has_mp
&& match_die(c
, o
))) {
527 link_mask(topology_core_cpumask
, cpu
, i
);
530 * Does this new cpu bringup a new core?
533 topology_sibling_cpumask(cpu
)) == 1) {
535 * for each core in package, increment
536 * the booted_cores for this new cpu
539 topology_sibling_cpumask(i
)) == i
)
542 * increment the core count for all
543 * the other cpus in this package
546 cpu_data(i
).booted_cores
++;
547 } else if (i
!= cpu
&& !c
->booted_cores
)
548 c
->booted_cores
= cpu_data(i
).booted_cores
;
550 if (match_die(c
, o
) && !topology_same_node(c
, o
))
551 primarily_use_numa_for_topology();
555 /* maps the cpu to the sched domain representing multi-core */
556 const struct cpumask
*cpu_coregroup_mask(int cpu
)
558 return cpu_llc_shared_mask(cpu
);
561 static void impress_friends(void)
564 unsigned long bogosum
= 0;
566 * Allow the user to impress friends.
568 pr_debug("Before bogomips\n");
569 for_each_possible_cpu(cpu
)
570 if (cpumask_test_cpu(cpu
, cpu_callout_mask
))
571 bogosum
+= cpu_data(cpu
).loops_per_jiffy
;
572 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
575 (bogosum
/(5000/HZ
))%100);
577 pr_debug("Before bogocount - setting activated=1\n");
580 void __inquire_remote_apic(int apicid
)
582 unsigned i
, regs
[] = { APIC_ID
>> 4, APIC_LVR
>> 4, APIC_SPIV
>> 4 };
583 const char * const names
[] = { "ID", "VERSION", "SPIV" };
587 pr_info("Inquiring remote APIC 0x%x...\n", apicid
);
589 for (i
= 0; i
< ARRAY_SIZE(regs
); i
++) {
590 pr_info("... APIC 0x%x %s: ", apicid
, names
[i
]);
595 status
= safe_apic_wait_icr_idle();
597 pr_cont("a previous APIC delivery may have failed\n");
599 apic_icr_write(APIC_DM_REMRD
| regs
[i
], apicid
);
604 status
= apic_read(APIC_ICR
) & APIC_ICR_RR_MASK
;
605 } while (status
== APIC_ICR_RR_INPROG
&& timeout
++ < 1000);
608 case APIC_ICR_RR_VALID
:
609 status
= apic_read(APIC_RRR
);
610 pr_cont("%08x\n", status
);
619 * The Multiprocessor Specification 1.4 (1997) example code suggests
620 * that there should be a 10ms delay between the BSP asserting INIT
621 * and de-asserting INIT, when starting a remote processor.
622 * But that slows boot and resume on modern processors, which include
623 * many cores and don't require that delay.
625 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
626 * Modern processor families are quirked to remove the delay entirely.
628 #define UDELAY_10MS_DEFAULT 10000
630 static unsigned int init_udelay
= UINT_MAX
;
632 static int __init
cpu_init_udelay(char *str
)
634 get_option(&str
, &init_udelay
);
638 early_param("cpu_init_udelay", cpu_init_udelay
);
640 static void __init
smp_quirk_init_udelay(void)
642 /* if cmdline changed it from default, leave it alone */
643 if (init_udelay
!= UINT_MAX
)
646 /* if modern processor, use no delay */
647 if (((boot_cpu_data
.x86_vendor
== X86_VENDOR_INTEL
) && (boot_cpu_data
.x86
== 6)) ||
648 ((boot_cpu_data
.x86_vendor
== X86_VENDOR_AMD
) && (boot_cpu_data
.x86
>= 0xF))) {
652 /* else, use legacy delay */
653 init_udelay
= UDELAY_10MS_DEFAULT
;
657 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
658 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
659 * won't ... remember to clear down the APIC, etc later.
662 wakeup_secondary_cpu_via_nmi(int apicid
, unsigned long start_eip
)
664 unsigned long send_status
, accept_status
= 0;
668 /* Boot on the stack */
669 /* Kick the second */
670 apic_icr_write(APIC_DM_NMI
| apic
->dest_logical
, apicid
);
672 pr_debug("Waiting for send to finish...\n");
673 send_status
= safe_apic_wait_icr_idle();
676 * Give the other CPU some time to accept the IPI.
679 if (APIC_INTEGRATED(apic_version
[boot_cpu_physical_apicid
])) {
680 maxlvt
= lapic_get_maxlvt();
681 if (maxlvt
> 3) /* Due to the Pentium erratum 3AP. */
682 apic_write(APIC_ESR
, 0);
683 accept_status
= (apic_read(APIC_ESR
) & 0xEF);
685 pr_debug("NMI sent\n");
688 pr_err("APIC never delivered???\n");
690 pr_err("APIC delivery error (%lx)\n", accept_status
);
692 return (send_status
| accept_status
);
696 wakeup_secondary_cpu_via_init(int phys_apicid
, unsigned long start_eip
)
698 unsigned long send_status
= 0, accept_status
= 0;
699 int maxlvt
, num_starts
, j
;
701 maxlvt
= lapic_get_maxlvt();
704 * Be paranoid about clearing APIC errors.
706 if (APIC_INTEGRATED(apic_version
[phys_apicid
])) {
707 if (maxlvt
> 3) /* Due to the Pentium erratum 3AP. */
708 apic_write(APIC_ESR
, 0);
712 pr_debug("Asserting INIT\n");
715 * Turn INIT on target chip
720 apic_icr_write(APIC_INT_LEVELTRIG
| APIC_INT_ASSERT
| APIC_DM_INIT
,
723 pr_debug("Waiting for send to finish...\n");
724 send_status
= safe_apic_wait_icr_idle();
728 pr_debug("Deasserting INIT\n");
732 apic_icr_write(APIC_INT_LEVELTRIG
| APIC_DM_INIT
, phys_apicid
);
734 pr_debug("Waiting for send to finish...\n");
735 send_status
= safe_apic_wait_icr_idle();
740 * Should we send STARTUP IPIs ?
742 * Determine this based on the APIC version.
743 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
745 if (APIC_INTEGRATED(apic_version
[phys_apicid
]))
751 * Run STARTUP IPI loop.
753 pr_debug("#startup loops: %d\n", num_starts
);
755 for (j
= 1; j
<= num_starts
; j
++) {
756 pr_debug("Sending STARTUP #%d\n", j
);
757 if (maxlvt
> 3) /* Due to the Pentium erratum 3AP. */
758 apic_write(APIC_ESR
, 0);
760 pr_debug("After apic_write\n");
767 /* Boot on the stack */
768 /* Kick the second */
769 apic_icr_write(APIC_DM_STARTUP
| (start_eip
>> 12),
773 * Give the other CPU some time to accept the IPI.
775 if (init_udelay
== 0)
780 pr_debug("Startup point 1\n");
782 pr_debug("Waiting for send to finish...\n");
783 send_status
= safe_apic_wait_icr_idle();
786 * Give the other CPU some time to accept the IPI.
788 if (init_udelay
== 0)
793 if (maxlvt
> 3) /* Due to the Pentium erratum 3AP. */
794 apic_write(APIC_ESR
, 0);
795 accept_status
= (apic_read(APIC_ESR
) & 0xEF);
796 if (send_status
|| accept_status
)
799 pr_debug("After Startup\n");
802 pr_err("APIC never delivered???\n");
804 pr_err("APIC delivery error (%lx)\n", accept_status
);
806 return (send_status
| accept_status
);
809 void smp_announce(void)
811 int num_nodes
= num_online_nodes();
813 printk(KERN_INFO
"x86: Booted up %d node%s, %d CPUs\n",
814 num_nodes
, (num_nodes
> 1 ? "s" : ""), num_online_cpus());
817 /* reduce the number of lines printed when booting a large cpu count system */
818 static void announce_cpu(int cpu
, int apicid
)
820 static int current_node
= -1;
821 int node
= early_cpu_to_node(cpu
);
822 static int width
, node_width
;
825 width
= num_digits(num_possible_cpus()) + 1; /* + '#' sign */
828 node_width
= num_digits(num_possible_nodes()) + 1; /* + '#' */
831 printk(KERN_INFO
"x86: Booting SMP configuration:\n");
833 if (system_state
== SYSTEM_BOOTING
) {
834 if (node
!= current_node
) {
835 if (current_node
> (-1))
839 printk(KERN_INFO
".... node %*s#%d, CPUs: ",
840 node_width
- num_digits(node
), " ", node
);
843 /* Add padding for the BSP */
845 pr_cont("%*s", width
+ 1, " ");
847 pr_cont("%*s#%d", width
- num_digits(cpu
), " ", cpu
);
850 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
854 static int wakeup_cpu0_nmi(unsigned int cmd
, struct pt_regs
*regs
)
858 cpu
= smp_processor_id();
859 if (cpu
== 0 && !cpu_online(cpu
) && enable_start_cpu0
)
866 * Wake up AP by INIT, INIT, STARTUP sequence.
868 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
869 * boot-strap code which is not a desired behavior for waking up BSP. To
870 * void the boot-strap code, wake up CPU0 by NMI instead.
872 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
873 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
874 * We'll change this code in the future to wake up hard offlined CPU0 if
875 * real platform and request are available.
878 wakeup_cpu_via_init_nmi(int cpu
, unsigned long start_ip
, int apicid
,
879 int *cpu0_nmi_registered
)
887 * Wake up AP by INIT, INIT, STARTUP sequence.
890 boot_error
= wakeup_secondary_cpu_via_init(apicid
, start_ip
);
895 * Wake up BSP by nmi.
897 * Register a NMI handler to help wake up CPU0.
899 boot_error
= register_nmi_handler(NMI_LOCAL
,
900 wakeup_cpu0_nmi
, 0, "wake_cpu0");
903 enable_start_cpu0
= 1;
904 *cpu0_nmi_registered
= 1;
905 if (apic
->dest_logical
== APIC_DEST_LOGICAL
)
906 id
= cpu0_logical_apicid
;
909 boot_error
= wakeup_secondary_cpu_via_nmi(id
, start_ip
);
918 void common_cpu_up(unsigned int cpu
, struct task_struct
*idle
)
920 /* Just in case we booted with a single CPU. */
921 alternatives_enable_smp();
923 per_cpu(current_task
, cpu
) = idle
;
926 /* Stack for startup_32 can be just as for start_secondary onwards */
928 per_cpu(cpu_current_top_of_stack
, cpu
) =
929 (unsigned long)task_stack_page(idle
) + THREAD_SIZE
;
931 clear_tsk_thread_flag(idle
, TIF_FORK
);
932 initial_gs
= per_cpu_offset(cpu
);
937 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
938 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
939 * Returns zero if CPU booted OK, else error code from
940 * ->wakeup_secondary_cpu.
942 static int do_boot_cpu(int apicid
, int cpu
, struct task_struct
*idle
)
944 volatile u32
*trampoline_status
=
945 (volatile u32
*) __va(real_mode_header
->trampoline_status
);
946 /* start_ip had better be page-aligned! */
947 unsigned long start_ip
= real_mode_header
->trampoline_start
;
949 unsigned long boot_error
= 0;
950 int cpu0_nmi_registered
= 0;
951 unsigned long timeout
;
953 idle
->thread
.sp
= (unsigned long) (((struct pt_regs
*)
954 (THREAD_SIZE
+ task_stack_page(idle
))) - 1);
956 early_gdt_descr
.address
= (unsigned long)get_cpu_gdt_table(cpu
);
957 initial_code
= (unsigned long)start_secondary
;
958 stack_start
= idle
->thread
.sp
;
961 * Enable the espfix hack for this CPU
963 #ifdef CONFIG_X86_ESPFIX64
967 /* So we see what's up */
968 announce_cpu(cpu
, apicid
);
971 * This grunge runs the startup process for
972 * the targeted processor.
975 if (get_uv_system_type() != UV_NON_UNIQUE_APIC
) {
977 pr_debug("Setting warm reset code and vector.\n");
979 smpboot_setup_warm_reset_vector(start_ip
);
981 * Be paranoid about clearing APIC errors.
983 if (APIC_INTEGRATED(apic_version
[boot_cpu_physical_apicid
])) {
984 apic_write(APIC_ESR
, 0);
990 * AP might wait on cpu_callout_mask in cpu_init() with
991 * cpu_initialized_mask set if previous attempt to online
992 * it timed-out. Clear cpu_initialized_mask so that after
993 * INIT/SIPI it could start with a clean state.
995 cpumask_clear_cpu(cpu
, cpu_initialized_mask
);
999 * Wake up a CPU in difference cases:
1000 * - Use the method in the APIC driver if it's defined
1002 * - Use an INIT boot APIC message for APs or NMI for BSP.
1004 if (apic
->wakeup_secondary_cpu
)
1005 boot_error
= apic
->wakeup_secondary_cpu(apicid
, start_ip
);
1007 boot_error
= wakeup_cpu_via_init_nmi(cpu
, start_ip
, apicid
,
1008 &cpu0_nmi_registered
);
1012 * Wait 10s total for first sign of life from AP
1015 timeout
= jiffies
+ 10*HZ
;
1016 while (time_before(jiffies
, timeout
)) {
1017 if (cpumask_test_cpu(cpu
, cpu_initialized_mask
)) {
1019 * Tell AP to proceed with initialization
1021 cpumask_set_cpu(cpu
, cpu_callout_mask
);
1031 * Wait till AP completes initial initialization
1033 while (!cpumask_test_cpu(cpu
, cpu_callin_mask
)) {
1035 * Allow other tasks to run while we wait for the
1036 * AP to come online. This also gives a chance
1037 * for the MTRR work(triggered by the AP coming online)
1038 * to be completed in the stop machine context.
1044 /* mark "stuck" area as not stuck */
1045 *trampoline_status
= 0;
1047 if (get_uv_system_type() != UV_NON_UNIQUE_APIC
) {
1049 * Cleanup possible dangling ends...
1051 smpboot_restore_warm_reset_vector();
1054 * Clean up the nmi handler. Do this after the callin and callout sync
1055 * to avoid impact of possible long unregister time.
1057 if (cpu0_nmi_registered
)
1058 unregister_nmi_handler(NMI_LOCAL
, "wake_cpu0");
1063 int native_cpu_up(unsigned int cpu
, struct task_struct
*tidle
)
1065 int apicid
= apic
->cpu_present_to_apicid(cpu
);
1066 unsigned long flags
;
1069 WARN_ON(irqs_disabled());
1071 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu
);
1073 if (apicid
== BAD_APICID
||
1074 !physid_isset(apicid
, phys_cpu_present_map
) ||
1075 !apic
->apic_id_valid(apicid
)) {
1076 pr_err("%s: bad cpu %d\n", __func__
, cpu
);
1081 * Already booted CPU?
1083 if (cpumask_test_cpu(cpu
, cpu_callin_mask
)) {
1084 pr_debug("do_boot_cpu %d Already started\n", cpu
);
1089 * Save current MTRR state in case it was changed since early boot
1090 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1094 /* x86 CPUs take themselves offline, so delayed offline is OK. */
1095 err
= cpu_check_up_prepare(cpu
);
1096 if (err
&& err
!= -EBUSY
)
1099 /* the FPU context is blank, nobody can own it */
1100 __cpu_disable_lazy_restore(cpu
);
1102 common_cpu_up(cpu
, tidle
);
1105 * We have to walk the irq descriptors to setup the vector
1106 * space for the cpu which comes online. Prevent irq
1107 * alloc/free across the bringup.
1111 err
= do_boot_cpu(apicid
, cpu
, tidle
);
1114 irq_unlock_sparse();
1115 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err
, cpu
);
1120 * Check TSC synchronization with the AP (keep irqs disabled
1123 local_irq_save(flags
);
1124 check_tsc_sync_source(cpu
);
1125 local_irq_restore(flags
);
1127 while (!cpu_online(cpu
)) {
1129 touch_nmi_watchdog();
1132 irq_unlock_sparse();
1138 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1140 void arch_disable_smp_support(void)
1142 disable_ioapic_support();
1146 * Fall back to non SMP mode after errors.
1148 * RED-PEN audit/test this more. I bet there is more state messed up here.
1150 static __init
void disable_smp(void)
1152 pr_info("SMP disabled\n");
1154 disable_ioapic_support();
1156 init_cpu_present(cpumask_of(0));
1157 init_cpu_possible(cpumask_of(0));
1159 if (smp_found_config
)
1160 physid_set_mask_of_physid(boot_cpu_physical_apicid
, &phys_cpu_present_map
);
1162 physid_set_mask_of_physid(0, &phys_cpu_present_map
);
1163 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1164 cpumask_set_cpu(0, topology_core_cpumask(0));
1175 * Various sanity checks.
1177 static int __init
smp_sanity_check(unsigned max_cpus
)
1181 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1182 if (def_to_bigsmp
&& nr_cpu_ids
> 8) {
1186 pr_warn("More than 8 CPUs detected - skipping them\n"
1187 "Use CONFIG_X86_BIGSMP\n");
1190 for_each_present_cpu(cpu
) {
1192 set_cpu_present(cpu
, false);
1197 for_each_possible_cpu(cpu
) {
1199 set_cpu_possible(cpu
, false);
1207 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map
)) {
1208 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1209 hard_smp_processor_id());
1211 physid_set(hard_smp_processor_id(), phys_cpu_present_map
);
1215 * If we couldn't find an SMP configuration at boot time,
1216 * get out of here now!
1218 if (!smp_found_config
&& !acpi_lapic
) {
1220 pr_notice("SMP motherboard not detected\n");
1221 return SMP_NO_CONFIG
;
1225 * Should not be necessary because the MP table should list the boot
1226 * CPU too, but we do it for the sake of robustness anyway.
1228 if (!apic
->check_phys_apicid_present(boot_cpu_physical_apicid
)) {
1229 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1230 boot_cpu_physical_apicid
);
1231 physid_set(hard_smp_processor_id(), phys_cpu_present_map
);
1236 * If we couldn't find a local APIC, then get out of here now!
1238 if (APIC_INTEGRATED(apic_version
[boot_cpu_physical_apicid
]) &&
1239 !boot_cpu_has(X86_FEATURE_APIC
)) {
1240 if (!disable_apic
) {
1241 pr_err("BIOS bug, local APIC #%d not detected!...\n",
1242 boot_cpu_physical_apicid
);
1243 pr_err("... forcing use of dummy APIC emulation (tell your hw vendor)\n");
1249 * If SMP should be disabled, then really disable it!
1252 pr_info("SMP mode deactivated\n");
1253 return SMP_FORCE_UP
;
1259 static void __init
smp_cpu_index_default(void)
1262 struct cpuinfo_x86
*c
;
1264 for_each_possible_cpu(i
) {
1266 /* mark all to hotplug */
1267 c
->cpu_index
= nr_cpu_ids
;
1272 * Prepare for SMP bootup. The MP table or ACPI has been read
1273 * earlier. Just do some sanity checking here and enable APIC mode.
1275 void __init
native_smp_prepare_cpus(unsigned int max_cpus
)
1279 smp_cpu_index_default();
1282 * Setup boot CPU information
1284 smp_store_boot_cpu_info(); /* Final full version of the data */
1285 cpumask_copy(cpu_callin_mask
, cpumask_of(0));
1288 current_thread_info()->cpu
= 0; /* needed? */
1289 for_each_possible_cpu(i
) {
1290 zalloc_cpumask_var(&per_cpu(cpu_sibling_map
, i
), GFP_KERNEL
);
1291 zalloc_cpumask_var(&per_cpu(cpu_core_map
, i
), GFP_KERNEL
);
1292 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map
, i
), GFP_KERNEL
);
1294 set_cpu_sibling_map(0);
1296 switch (smp_sanity_check(max_cpus
)) {
1299 if (APIC_init_uniprocessor())
1300 pr_notice("Local APIC not detected. Using dummy APIC emulation.\n");
1307 apic_bsp_setup(false);
1313 default_setup_apic_routing();
1315 if (read_apic_id() != boot_cpu_physical_apicid
) {
1316 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1317 read_apic_id(), boot_cpu_physical_apicid
);
1318 /* Or can we switch back to PIC here? */
1321 cpu0_logical_apicid
= apic_bsp_setup(false);
1323 pr_info("CPU%d: ", 0);
1324 print_cpu_info(&cpu_data(0));
1329 set_mtrr_aps_delayed_init();
1331 smp_quirk_init_udelay();
1334 void arch_enable_nonboot_cpus_begin(void)
1336 set_mtrr_aps_delayed_init();
1339 void arch_enable_nonboot_cpus_end(void)
1345 * Early setup to make printk work.
1347 void __init
native_smp_prepare_boot_cpu(void)
1349 int me
= smp_processor_id();
1350 switch_to_new_gdt(me
);
1351 /* already set me in cpu_online_mask in boot_cpu_init() */
1352 cpumask_set_cpu(me
, cpu_callout_mask
);
1353 cpu_set_state_online(me
);
1356 void __init
native_smp_cpus_done(unsigned int max_cpus
)
1358 pr_debug("Boot done\n");
1362 setup_ioapic_dest();
1366 static int __initdata setup_possible_cpus
= -1;
1367 static int __init
_setup_possible_cpus(char *str
)
1369 get_option(&str
, &setup_possible_cpus
);
1372 early_param("possible_cpus", _setup_possible_cpus
);
1376 * cpu_possible_mask should be static, it cannot change as cpu's
1377 * are onlined, or offlined. The reason is per-cpu data-structures
1378 * are allocated by some modules at init time, and dont expect to
1379 * do this dynamically on cpu arrival/departure.
1380 * cpu_present_mask on the other hand can change dynamically.
1381 * In case when cpu_hotplug is not compiled, then we resort to current
1382 * behaviour, which is cpu_possible == cpu_present.
1385 * Three ways to find out the number of additional hotplug CPUs:
1386 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1387 * - The user can overwrite it with possible_cpus=NUM
1388 * - Otherwise don't reserve additional CPUs.
1389 * We do this because additional CPUs waste a lot of memory.
1392 __init
void prefill_possible_map(void)
1396 /* no processor from mptable or madt */
1397 if (!num_processors
)
1400 i
= setup_max_cpus
?: 1;
1401 if (setup_possible_cpus
== -1) {
1402 possible
= num_processors
;
1403 #ifdef CONFIG_HOTPLUG_CPU
1405 possible
+= disabled_cpus
;
1411 possible
= setup_possible_cpus
;
1413 total_cpus
= max_t(int, possible
, num_processors
+ disabled_cpus
);
1415 /* nr_cpu_ids could be reduced via nr_cpus= */
1416 if (possible
> nr_cpu_ids
) {
1417 pr_warn("%d Processors exceeds NR_CPUS limit of %d\n",
1418 possible
, nr_cpu_ids
);
1419 possible
= nr_cpu_ids
;
1422 #ifdef CONFIG_HOTPLUG_CPU
1423 if (!setup_max_cpus
)
1426 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1427 possible
, setup_max_cpus
);
1431 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1432 possible
, max_t(int, possible
- num_processors
, 0));
1434 for (i
= 0; i
< possible
; i
++)
1435 set_cpu_possible(i
, true);
1436 for (; i
< NR_CPUS
; i
++)
1437 set_cpu_possible(i
, false);
1439 nr_cpu_ids
= possible
;
1442 #ifdef CONFIG_HOTPLUG_CPU
1444 static void remove_siblinginfo(int cpu
)
1447 struct cpuinfo_x86
*c
= &cpu_data(cpu
);
1449 for_each_cpu(sibling
, topology_core_cpumask(cpu
)) {
1450 cpumask_clear_cpu(cpu
, topology_core_cpumask(sibling
));
1452 * last thread sibling in this cpu core going down
1454 if (cpumask_weight(topology_sibling_cpumask(cpu
)) == 1)
1455 cpu_data(sibling
).booted_cores
--;
1458 for_each_cpu(sibling
, topology_sibling_cpumask(cpu
))
1459 cpumask_clear_cpu(cpu
, topology_sibling_cpumask(sibling
));
1460 for_each_cpu(sibling
, cpu_llc_shared_mask(cpu
))
1461 cpumask_clear_cpu(cpu
, cpu_llc_shared_mask(sibling
));
1462 cpumask_clear(cpu_llc_shared_mask(cpu
));
1463 cpumask_clear(topology_sibling_cpumask(cpu
));
1464 cpumask_clear(topology_core_cpumask(cpu
));
1465 c
->phys_proc_id
= 0;
1467 cpumask_clear_cpu(cpu
, cpu_sibling_setup_mask
);
1470 static void remove_cpu_from_maps(int cpu
)
1472 set_cpu_online(cpu
, false);
1473 cpumask_clear_cpu(cpu
, cpu_callout_mask
);
1474 cpumask_clear_cpu(cpu
, cpu_callin_mask
);
1475 /* was set by cpu_init() */
1476 cpumask_clear_cpu(cpu
, cpu_initialized_mask
);
1477 numa_remove_cpu(cpu
);
1480 void cpu_disable_common(void)
1482 int cpu
= smp_processor_id();
1484 remove_siblinginfo(cpu
);
1486 /* It's now safe to remove this processor from the online map */
1488 remove_cpu_from_maps(cpu
);
1489 unlock_vector_lock();
1493 int native_cpu_disable(void)
1497 ret
= check_irq_vectors_for_cpu_disable();
1502 cpu_disable_common();
1507 int common_cpu_die(unsigned int cpu
)
1511 /* We don't do anything here: idle task is faking death itself. */
1513 /* They ack this in play_dead() by setting CPU_DEAD */
1514 if (cpu_wait_death(cpu
, 5)) {
1515 if (system_state
== SYSTEM_RUNNING
)
1516 pr_info("CPU %u is now offline\n", cpu
);
1518 pr_err("CPU %u didn't die...\n", cpu
);
1525 void native_cpu_die(unsigned int cpu
)
1527 common_cpu_die(cpu
);
1530 void play_dead_common(void)
1533 reset_lazy_tlbstate();
1534 amd_e400_remove_cpu(raw_smp_processor_id());
1537 (void)cpu_report_death();
1540 * With physical CPU hotplug, we should halt the cpu
1542 local_irq_disable();
1545 static bool wakeup_cpu0(void)
1547 if (smp_processor_id() == 0 && enable_start_cpu0
)
1554 * We need to flush the caches before going to sleep, lest we have
1555 * dirty data in our caches when we come back up.
1557 static inline void mwait_play_dead(void)
1559 unsigned int eax
, ebx
, ecx
, edx
;
1560 unsigned int highest_cstate
= 0;
1561 unsigned int highest_subcstate
= 0;
1565 if (!this_cpu_has(X86_FEATURE_MWAIT
))
1567 if (!this_cpu_has(X86_FEATURE_CLFLUSH
))
1569 if (__this_cpu_read(cpu_info
.cpuid_level
) < CPUID_MWAIT_LEAF
)
1572 eax
= CPUID_MWAIT_LEAF
;
1574 native_cpuid(&eax
, &ebx
, &ecx
, &edx
);
1577 * eax will be 0 if EDX enumeration is not valid.
1578 * Initialized below to cstate, sub_cstate value when EDX is valid.
1580 if (!(ecx
& CPUID5_ECX_EXTENSIONS_SUPPORTED
)) {
1583 edx
>>= MWAIT_SUBSTATE_SIZE
;
1584 for (i
= 0; i
< 7 && edx
; i
++, edx
>>= MWAIT_SUBSTATE_SIZE
) {
1585 if (edx
& MWAIT_SUBSTATE_MASK
) {
1587 highest_subcstate
= edx
& MWAIT_SUBSTATE_MASK
;
1590 eax
= (highest_cstate
<< MWAIT_SUBSTATE_SIZE
) |
1591 (highest_subcstate
- 1);
1595 * This should be a memory location in a cache line which is
1596 * unlikely to be touched by other processors. The actual
1597 * content is immaterial as it is not actually modified in any way.
1599 mwait_ptr
= ¤t_thread_info()->flags
;
1605 * The CLFLUSH is a workaround for erratum AAI65 for
1606 * the Xeon 7400 series. It's not clear it is actually
1607 * needed, but it should be harmless in either case.
1608 * The WBINVD is insufficient due to the spurious-wakeup
1609 * case where we return around the loop.
1614 __monitor(mwait_ptr
, 0, 0);
1618 * If NMI wants to wake up CPU0, start CPU0.
1625 static inline void hlt_play_dead(void)
1627 if (__this_cpu_read(cpu_info
.x86
) >= 4)
1633 * If NMI wants to wake up CPU0, start CPU0.
1640 void native_play_dead(void)
1643 tboot_shutdown(TB_SHUTDOWN_WFS
);
1645 mwait_play_dead(); /* Only returns on failure */
1646 if (cpuidle_play_dead())
1650 #else /* ... !CONFIG_HOTPLUG_CPU */
1651 int native_cpu_disable(void)
1656 void native_cpu_die(unsigned int cpu
)
1658 /* We said "no" in __cpu_disable */
1662 void native_play_dead(void)