2 * Copyright (c) 1996, by Steve Passe
3 * Copyright (c) 2003, by Peter Wemm
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. The name of the developer may NOT be used to endorse or promote products
12 * derived from this software without specific prior written permission.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
35 #include "opt_kstack_pages.h"
37 #include "opt_sched.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
43 #include <sys/cons.h> /* cngetc() */
44 #include <sys/cpuset.h>
49 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/memrange.h>
54 #include <sys/mutex.h>
57 #include <sys/sched.h>
59 #include <sys/sysctl.h>
62 #include <vm/vm_param.h>
64 #include <vm/vm_kern.h>
65 #include <vm/vm_extern.h>
67 #include <x86/apicreg.h>
68 #include <machine/clock.h>
69 #include <machine/cputypes.h>
71 #include <machine/md_var.h>
72 #include <machine/pcb.h>
73 #include <machine/psl.h>
74 #include <machine/smp.h>
75 #include <machine/specialreg.h>
76 #include <machine/cpu.h>
78 #define WARMBOOT_TARGET 0
79 #define WARMBOOT_OFF (KERNBASE + 0x0467)
80 #define WARMBOOT_SEG (KERNBASE + 0x0469)
82 #define CMOS_REG (0x70)
83 #define CMOS_DATA (0x71)
84 #define BIOS_RESET (0x0f)
85 #define BIOS_WARM (0x0a)
87 static MALLOC_DEFINE(M_CPUS, "cpus", "CPU items");
89 /* lock region used by kernel profiling */
92 int mp_naps; /* # of Applications processors */
93 int boot_cpu_id = -1; /* designated BSP */
95 /* AP uses this during bootstrap. Do not staticize. */
99 /* Free these after use */
100 void *bootstacks[MAXCPU];
103 struct pcb stoppcbs[MAXCPU];
104 struct susppcb **susppcbs;
107 /* Interrupt counts. */
108 static u_long *ipi_preempt_counts[MAXCPU];
109 static u_long *ipi_ast_counts[MAXCPU];
110 u_long *ipi_invltlb_counts[MAXCPU];
111 u_long *ipi_invlrng_counts[MAXCPU];
112 u_long *ipi_invlpg_counts[MAXCPU];
113 u_long *ipi_invlcache_counts[MAXCPU];
114 u_long *ipi_rendezvous_counts[MAXCPU];
115 static u_long *ipi_hardclock_counts[MAXCPU];
118 /* Default cpu_ops implementation. */
119 struct cpu_ops cpu_ops;
122 * Local data and functions.
125 static volatile cpuset_t ipi_stop_nmi_pending;
127 /* used to hold the AP's until we are ready to release them */
128 struct mtx ap_boot_mtx;
130 /* Set to 1 once we're ready to let the APs out of the pen. */
131 volatile int aps_ready = 0;
134 * Store data from cpu_add() until later in the boot when we actually setup
137 struct cpu_info *cpu_info;
139 int cpu_apic_ids[MAXCPU];
140 _Static_assert(MAXCPU <= MAX_APIC_ID,
141 "MAXCPU cannot be larger that MAX_APIC_ID");
142 _Static_assert(xAPIC_MAX_APIC_ID <= MAX_APIC_ID,
143 "xAPIC_MAX_APIC_ID cannot be larger that MAX_APIC_ID");
145 /* Holds pending bitmap based IPIs per CPU */
146 volatile u_int cpu_ipi_pending[MAXCPU];
148 static void release_aps(void *dummy);
149 static void cpustop_handler_post(u_int cpu);
151 static int hyperthreading_allowed = 1;
152 SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_allowed, CTLFLAG_RDTUN,
153 &hyperthreading_allowed, 0, "Use Intel HTT logical CPUs");
155 static struct topo_node topo_root;
157 static int pkg_id_shift;
158 static int core_id_shift;
159 static int disabled_cpus;
164 } static caches[MAX_CACHE_LEVELS];
167 mem_range_AP_init(void)
170 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
171 mem_range_softc.mr_op->initAP(&mem_range_softc);
175 * Round up to the next power of two, if necessary, and then
177 * Returns -1 if argument is zero.
183 return (fls(x << (1 - powerof2(x))) - 1);
187 * Add a cache level to the cache topology description.
190 add_deterministic_cache(int type, int level, int share_count)
196 printf("unexpected cache type %d\n", type);
199 if (type == 2) /* ignore instruction cache */
201 if (level == 0 || level > MAX_CACHE_LEVELS) {
202 printf("unexpected cache level %d\n", type);
206 if (caches[level - 1].present) {
207 printf("WARNING: multiple entries for L%u data cache\n", level);
208 printf("%u => %u\n", caches[level - 1].id_shift,
209 mask_width(share_count));
211 caches[level - 1].id_shift = mask_width(share_count);
212 caches[level - 1].present = 1;
214 if (caches[level - 1].id_shift > pkg_id_shift) {
215 printf("WARNING: L%u data cache covers more "
216 "APIC IDs than a package (%u > %u)\n", level,
217 caches[level - 1].id_shift, pkg_id_shift);
218 caches[level - 1].id_shift = pkg_id_shift;
220 if (caches[level - 1].id_shift < core_id_shift) {
221 printf("WARNING: L%u data cache covers fewer "
222 "APIC IDs than a core (%u < %u)\n", level,
223 caches[level - 1].id_shift, core_id_shift);
224 caches[level - 1].id_shift = core_id_shift;
231 * Determine topology of processing units and caches for AMD CPUs.
233 * - AMD CPUID Specification (Publication # 25481)
234 * - BKDG for AMD NPT Family 0Fh Processors (Publication # 32559)
235 * - BKDG For AMD Family 10h Processors (Publication # 31116)
236 * - BKDG For AMD Family 15h Models 00h-0Fh Processors (Publication # 42301)
237 * - BKDG For AMD Family 16h Models 00h-0Fh Processors (Publication # 48751)
245 int nodes_per_socket;
250 /* No multi-core capability. */
251 if ((amd_feature2 & AMDID2_CMP) == 0)
254 /* For families 10h and newer. */
255 pkg_id_shift = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
256 AMDID_COREID_SIZE_SHIFT;
258 /* For 0Fh family. */
259 if (pkg_id_shift == 0)
261 mask_width((cpu_procinfo2 & AMDID_CMP_CORES) + 1);
264 * Families prior to 16h define the following value as
265 * cores per compute unit and we don't really care about the AMD
266 * compute units at the moment. Perhaps we should treat them as
267 * cores and cores within the compute units as hardware threads,
268 * but that's up for debate.
269 * Later families define the value as threads per compute unit,
270 * so we are following AMD's nomenclature here.
272 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0 &&
273 CPUID_TO_FAMILY(cpu_id) >= 0x16) {
274 cpuid_count(0x8000001e, 0, p);
275 share_count = ((p[1] >> 8) & 0xff) + 1;
276 core_id_shift = mask_width(share_count);
279 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0) {
281 cpuid_count(0x8000001d, i, p);
283 level = (p[0] >> 5) & 0x7;
284 share_count = 1 + ((p[0] >> 14) & 0xfff);
286 if (!add_deterministic_cache(type, level, share_count))
290 if (cpu_exthigh >= 0x80000005) {
291 cpuid_count(0x80000005, 0, p);
292 if (((p[2] >> 24) & 0xff) != 0) {
293 caches[0].id_shift = 0;
294 caches[0].present = 1;
297 if (cpu_exthigh >= 0x80000006) {
298 cpuid_count(0x80000006, 0, p);
299 if (((p[2] >> 16) & 0xffff) != 0) {
300 caches[1].id_shift = 0;
301 caches[1].present = 1;
303 if (((p[3] >> 18) & 0x3fff) != 0) {
304 nodes_per_socket = 1;
305 if ((amd_feature2 & AMDID2_NODE_ID) != 0) {
307 * Handle multi-node processors that
308 * have multiple chips, each with its
309 * own L3 cache, on the same die.
311 v = rdmsr(0xc001100c);
312 nodes_per_socket = 1 + ((v >> 3) & 0x7);
315 pkg_id_shift - mask_width(nodes_per_socket);
316 caches[2].present = 1;
323 * Determine topology of processing units for Intel CPUs
324 * using CPUID Leaf 1 and Leaf 4, if supported.
326 * - Intel 64 Architecture Processor Topology Enumeration
327 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
328 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
329 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
332 topo_probe_intel_0x4(void)
338 /* Both zero and one here mean one logical processor per package. */
339 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
340 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
341 if (max_logical <= 1)
344 if (cpu_high >= 0x4) {
345 cpuid_count(0x04, 0, p);
346 max_cores = ((p[0] >> 26) & 0x3f) + 1;
350 core_id_shift = mask_width(max_logical/max_cores);
351 KASSERT(core_id_shift >= 0,
352 ("intel topo: max_cores > max_logical\n"));
353 pkg_id_shift = core_id_shift + mask_width(max_cores);
357 * Determine topology of processing units for Intel CPUs
358 * using CPUID Leaf 11, if supported.
360 * - Intel 64 Architecture Processor Topology Enumeration
361 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
362 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
363 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
366 topo_probe_intel_0xb(void)
373 /* Fall back if CPU leaf 11 doesn't really exist. */
374 cpuid_count(0x0b, 0, p);
376 topo_probe_intel_0x4();
380 /* We only support three levels for now. */
382 cpuid_count(0x0b, i, p);
385 type = (p[2] >> 8) & 0xff;
390 /* TODO: check for duplicate (re-)assignment */
391 if (type == CPUID_TYPE_SMT)
392 core_id_shift = bits;
393 else if (type == CPUID_TYPE_CORE)
396 printf("unknown CPU level type %d\n", type);
399 if (pkg_id_shift < core_id_shift) {
400 printf("WARNING: core covers more APIC IDs than a package\n");
401 core_id_shift = pkg_id_shift;
406 * Determine topology of caches for Intel CPUs.
408 * - Intel 64 Architecture Processor Topology Enumeration
409 * - Intel 64 and IA-32 Architectures Software Developer’s Manual
410 * Volume 2A: Instruction Set Reference, A-M,
414 topo_probe_intel_caches(void)
422 if (cpu_high < 0x4) {
424 * Available cache level and sizes can be determined
425 * via CPUID leaf 2, but that requires a huge table of hardcoded
426 * values, so for now just assume L1 and L2 caches potentially
427 * shared only by HTT processing units, if HTT is present.
429 caches[0].id_shift = pkg_id_shift;
430 caches[0].present = 1;
431 caches[1].id_shift = pkg_id_shift;
432 caches[1].present = 1;
437 cpuid_count(0x4, i, p);
439 level = (p[0] >> 5) & 0x7;
440 share_count = 1 + ((p[0] >> 14) & 0xfff);
442 if (!add_deterministic_cache(type, level, share_count))
448 * Determine topology of processing units and caches for Intel CPUs.
450 * - Intel 64 Architecture Processor Topology Enumeration
453 topo_probe_intel(void)
457 * Note that 0x1 <= cpu_high < 4 case should be
458 * compatible with topo_probe_intel_0x4() logic when
459 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
460 * or it should trigger the fallback otherwise.
463 topo_probe_intel_0xb();
464 else if (cpu_high >= 0x1)
465 topo_probe_intel_0x4();
467 topo_probe_intel_caches();
471 * Topology information is queried only on BSP, on which this
472 * code runs and for which it can query CPUID information.
473 * Then topology is extrapolated on all packages using an
474 * assumption that APIC ID to hardware component ID mapping is
476 * That doesn't necesserily imply that the topology is uniform.
481 static int cpu_topo_probed = 0;
482 struct x86_topo_layer {
486 } topo_layers[MAX_CACHE_LEVELS + 3];
487 struct topo_node *parent;
488 struct topo_node *node;
497 CPU_ZERO(&logical_cpus_mask);
501 else if (cpu_vendor_id == CPU_VENDOR_AMD)
503 else if (cpu_vendor_id == CPU_VENDOR_INTEL)
506 KASSERT(pkg_id_shift >= core_id_shift,
507 ("bug in APIC topology discovery"));
510 bzero(topo_layers, sizeof(topo_layers));
512 topo_layers[nlayers].type = TOPO_TYPE_PKG;
513 topo_layers[nlayers].id_shift = pkg_id_shift;
515 printf("Package ID shift: %u\n", topo_layers[nlayers].id_shift);
519 * Consider all caches to be within a package/chip
520 * and "in front" of all sub-components like
521 * cores and hardware threads.
523 for (i = MAX_CACHE_LEVELS - 1; i >= 0; --i) {
524 if (caches[i].present) {
525 KASSERT(caches[i].id_shift <= pkg_id_shift,
526 ("bug in APIC topology discovery"));
527 KASSERT(caches[i].id_shift >= core_id_shift,
528 ("bug in APIC topology discovery"));
530 topo_layers[nlayers].type = TOPO_TYPE_CACHE;
531 topo_layers[nlayers].subtype = i + 1;
532 topo_layers[nlayers].id_shift = caches[i].id_shift;
534 printf("L%u cache ID shift: %u\n",
535 topo_layers[nlayers].subtype,
536 topo_layers[nlayers].id_shift);
541 if (pkg_id_shift > core_id_shift) {
542 topo_layers[nlayers].type = TOPO_TYPE_CORE;
543 topo_layers[nlayers].id_shift = core_id_shift;
545 printf("Core ID shift: %u\n",
546 topo_layers[nlayers].id_shift);
550 topo_layers[nlayers].type = TOPO_TYPE_PU;
551 topo_layers[nlayers].id_shift = 0;
554 topo_init_root(&topo_root);
555 for (i = 0; i <= max_apic_id; ++i) {
556 if (!cpu_info[i].cpu_present)
560 for (layer = 0; layer < nlayers; ++layer) {
561 node_id = i >> topo_layers[layer].id_shift;
562 parent = topo_add_node_by_hwid(parent, node_id,
563 topo_layers[layer].type,
564 topo_layers[layer].subtype);
569 for (layer = 0; layer < nlayers; ++layer) {
570 node_id = boot_cpu_id >> topo_layers[layer].id_shift;
571 node = topo_find_node_by_hwid(parent, node_id,
572 topo_layers[layer].type,
573 topo_layers[layer].subtype);
574 topo_promote_child(node);
582 * Assign logical CPU IDs to local APICs.
587 struct topo_node *node;
590 smt_mask = (1u << core_id_shift) - 1;
593 * Assign CPU IDs to local APIC IDs and disable any CPUs
594 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
597 TOPO_FOREACH(node, &topo_root) {
598 if (node->type != TOPO_TYPE_PU)
601 if ((node->hwid & smt_mask) != (boot_cpu_id & smt_mask))
602 cpu_info[node->hwid].cpu_hyperthread = 1;
604 if (resource_disabled("lapic", node->hwid)) {
605 if (node->hwid != boot_cpu_id)
606 cpu_info[node->hwid].cpu_disabled = 1;
608 printf("Cannot disable BSP, APIC ID = %d\n",
612 if (!hyperthreading_allowed &&
613 cpu_info[node->hwid].cpu_hyperthread)
614 cpu_info[node->hwid].cpu_disabled = 1;
616 if (mp_ncpus >= MAXCPU)
617 cpu_info[node->hwid].cpu_disabled = 1;
619 if (cpu_info[node->hwid].cpu_disabled) {
624 cpu_apic_ids[mp_ncpus] = node->hwid;
625 apic_cpuids[node->hwid] = mp_ncpus;
626 topo_set_pu_id(node, mp_ncpus);
630 KASSERT(mp_maxid >= mp_ncpus - 1,
631 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
636 * Print various information about the SMP system hardware and setup.
639 cpu_mp_announce(void)
641 struct topo_node *node;
642 const char *hyperthread;
647 printf("FreeBSD/SMP: ");
648 if (topo_analyze(&topo_root, 1, &pkg_count,
649 &cores_per_pkg, &thrs_per_core)) {
650 printf("%d package(s)", pkg_count);
651 if (cores_per_pkg > 0)
652 printf(" x %d core(s)", cores_per_pkg);
653 if (thrs_per_core > 1)
654 printf(" x %d hardware threads", thrs_per_core);
656 printf("Non-uniform topology");
661 printf("FreeBSD/SMP Online: ");
662 if (topo_analyze(&topo_root, 0, &pkg_count,
663 &cores_per_pkg, &thrs_per_core)) {
664 printf("%d package(s)", pkg_count);
665 if (cores_per_pkg > 0)
666 printf(" x %d core(s)", cores_per_pkg);
667 if (thrs_per_core > 1)
668 printf(" x %d hardware threads", thrs_per_core);
670 printf("Non-uniform topology");
678 TOPO_FOREACH(node, &topo_root) {
679 switch (node->type) {
681 printf("Package HW ID = %u\n", node->hwid);
684 printf("\tCore HW ID = %u\n", node->hwid);
687 if (cpu_info[node->hwid].cpu_hyperthread)
692 if (node->subtype == 0)
693 printf("\t\tCPU (AP%s): APIC ID: %u"
694 "(disabled)\n", hyperthread, node->hwid);
695 else if (node->id == 0)
696 printf("\t\tCPU0 (BSP): APIC ID: %u\n",
699 printf("\t\tCPU%u (AP%s): APIC ID: %u\n",
700 node->id, hyperthread, node->hwid);
710 * Add a scheduling group, a group of logical processors sharing
711 * a particular cache (and, thus having an affinity), to the scheduling
713 * This function recursively works on lower level caches.
716 x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
718 struct topo_node *node;
723 KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE,
724 ("x86topo_add_sched_group: bad type: %u", root->type));
725 CPU_COPY(&root->cpuset, &cg_root->cg_mask);
726 cg_root->cg_count = root->cpu_count;
727 if (root->type == TOPO_TYPE_SYSTEM)
728 cg_root->cg_level = CG_SHARE_NONE;
730 cg_root->cg_level = root->subtype;
733 * Check how many core nodes we have under the given root node.
734 * If we have multiple logical processors, but not multiple
735 * cores, then those processors must be hardware threads.
739 while (node != NULL) {
740 if (node->type != TOPO_TYPE_CORE) {
741 node = topo_next_node(root, node);
746 node = topo_next_nonchild_node(root, node);
749 if (cg_root->cg_level != CG_SHARE_NONE &&
750 root->cpu_count > 1 && ncores < 2)
751 cg_root->cg_flags = CG_FLAG_SMT;
754 * Find out how many cache nodes we have under the given root node.
755 * We ignore cache nodes that cover all the same processors as the
756 * root node. Also, we do not descend below found cache nodes.
757 * That is, we count top-level "non-redundant" caches under the root
762 while (node != NULL) {
763 if (node->type != TOPO_TYPE_CACHE ||
764 (root->type != TOPO_TYPE_SYSTEM &&
765 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
766 node = topo_next_node(root, node);
770 node = topo_next_nonchild_node(root, node);
773 cg_root->cg_child = smp_topo_alloc(nchildren);
774 cg_root->cg_children = nchildren;
777 * Now find again the same cache nodes as above and recursively
778 * build scheduling topologies for them.
782 while (node != NULL) {
783 if (node->type != TOPO_TYPE_CACHE ||
784 (root->type != TOPO_TYPE_SYSTEM &&
785 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
786 node = topo_next_node(root, node);
789 cg_root->cg_child[i].cg_parent = cg_root;
790 x86topo_add_sched_group(node, &cg_root->cg_child[i]);
792 node = topo_next_nonchild_node(root, node);
797 * Build the MI scheduling topology from the discovered hardware topology.
802 struct cpu_group *cg_root;
805 return (smp_topo_none());
807 cg_root = smp_topo_alloc(1);
808 x86topo_add_sched_group(&topo_root, cg_root);
813 cpu_alloc(void *dummy __unused)
816 * Dynamically allocate the arrays that depend on the
819 cpu_info = malloc(sizeof(*cpu_info) * (max_apic_id + 1), M_CPUS,
821 apic_cpuids = malloc(sizeof(*apic_cpuids) * (max_apic_id + 1), M_CPUS,
824 SYSINIT(cpu_alloc, SI_SUB_CPU, SI_ORDER_FIRST, cpu_alloc, NULL);
827 * Add a logical CPU to the topology.
830 cpu_add(u_int apic_id, char boot_cpu)
833 if (apic_id > max_apic_id) {
834 panic("SMP: APIC ID %d too high", apic_id);
837 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %u added twice",
839 cpu_info[apic_id].cpu_present = 1;
841 KASSERT(boot_cpu_id == -1,
842 ("CPU %u claims to be BSP, but CPU %u already is", apic_id,
844 boot_cpu_id = apic_id;
845 cpu_info[apic_id].cpu_bsp = 1;
848 printf("SMP: Added CPU %u (%s)\n", apic_id, boot_cpu ? "BSP" :
853 cpu_mp_setmaxid(void)
857 * mp_ncpus and mp_maxid should be already set by calls to cpu_add().
858 * If there were no calls to cpu_add() assume this is a UP system.
869 * Always record BSP in CPU map so that the mbuf init code works
872 CPU_SETOF(0, &all_cpus);
873 return (mp_ncpus > 1);
877 * AP CPU's call this to initialize themselves.
880 init_secondary_tail(void)
885 * On real hardware, switch to x2apic mode if possible. Do it
886 * after aps_ready was signalled, to avoid manipulating the
887 * mode while BSP might still want to send some IPI to us
888 * (second startup IPI is ignored on modern hardware etc).
892 /* Initialize the PAT MSR. */
895 /* set up CPU registers and state */
901 /* set up FPU state on the AP */
908 if (cpu_ops.cpu_init)
911 /* A quick check from sanity claus */
912 cpuid = PCPU_GET(cpuid);
913 if (PCPU_GET(apic_id) != lapic_id()) {
914 printf("SMP: cpuid = %d\n", cpuid);
915 printf("SMP: actual apic_id = %d\n", lapic_id());
916 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
917 panic("cpuid mismatch! boom!!");
920 /* Initialize curthread. */
921 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
922 PCPU_SET(curthread, PCPU_GET(idlethread));
926 mtx_lock_spin(&ap_boot_mtx);
928 /* Init local apic for irq's */
931 /* Set memory range attributes for this CPU to match the BSP */
936 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
937 printf("SMP: AP CPU #%d Launched!\n", cpuid);
939 /* Determine if we are a logical CPU. */
940 if (cpu_info[PCPU_GET(apic_id)].cpu_hyperthread)
941 CPU_SET(cpuid, &logical_cpus_mask);
946 if (smp_cpus == mp_ncpus) {
947 /* enable IPI's, tlb shootdown, freezes etc */
948 atomic_store_rel_int(&smp_started, 1);
953 * Enable global pages TLB extension
954 * This also implicitly flushes the TLB
956 load_cr4(rcr4() | CR4_PGE);
957 if (pmap_pcid_enabled)
958 load_cr4(rcr4() | CR4_PCIDE);
964 mtx_unlock_spin(&ap_boot_mtx);
966 /* Wait until all the AP's are up. */
967 while (atomic_load_acq_int(&smp_started) == 0)
970 #ifndef EARLY_AP_STARTUP
971 /* Start per-CPU event timers. */
977 panic("scheduler returned us to %s", __func__);
981 /*******************************************************************
982 * local functions and data
986 * We tell the I/O APIC code about all the CPUs we want to receive
987 * interrupts. If we don't want certain CPUs to receive IRQs we
988 * can simply not tell the I/O APIC code about them in this function.
989 * We also do not tell it about the BSP since it tells itself about
990 * the BSP internally to work with UP kernels and on UP machines.
993 set_interrupt_apic_ids(void)
997 for (i = 0; i < MAXCPU; i++) {
998 apic_id = cpu_apic_ids[i];
1001 if (cpu_info[apic_id].cpu_bsp)
1003 if (cpu_info[apic_id].cpu_disabled)
1006 /* Don't let hyperthreads service interrupts. */
1007 if (cpu_info[apic_id].cpu_hyperthread)
1015 #ifdef COUNT_XINVLTLB_HITS
1016 u_int xhits_gbl[MAXCPU];
1017 u_int xhits_pg[MAXCPU];
1018 u_int xhits_rng[MAXCPU];
1019 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1020 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1021 sizeof(xhits_gbl), "IU", "");
1022 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1023 sizeof(xhits_pg), "IU", "");
1024 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1025 sizeof(xhits_rng), "IU", "");
1030 u_int ipi_range_size;
1031 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1032 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1033 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1034 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1036 #endif /* COUNT_XINVLTLB_HITS */
1039 * Init and startup IPI.
1042 ipi_startup(int apic_id, int vector)
1046 * This attempts to follow the algorithm described in the
1047 * Intel Multiprocessor Specification v1.4 in section B.4.
1048 * For each IPI, we allow the local APIC ~20us to deliver the
1049 * IPI. If that times out, we panic.
1053 * first we do an INIT IPI: this INIT IPI might be run, resetting
1054 * and running the target CPU. OR this INIT IPI might be latched (P5
1055 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
1058 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1059 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
1060 lapic_ipi_wait(100);
1062 /* Explicitly deassert the INIT IPI. */
1063 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1064 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT,
1067 DELAY(10000); /* wait ~10mS */
1070 * next we do a STARTUP IPI: the previous INIT IPI might still be
1071 * latched, (P5 bug) this 1st STARTUP would then terminate
1072 * immediately, and the previously started INIT IPI would continue. OR
1073 * the previous INIT IPI has already run. and this STARTUP IPI will
1074 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1077 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1078 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1080 if (!lapic_ipi_wait(100))
1081 panic("Failed to deliver first STARTUP IPI to APIC %d",
1083 DELAY(200); /* wait ~200uS */
1086 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1087 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1088 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1089 * recognized after hardware RESET or INIT IPI.
1091 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1092 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1094 if (!lapic_ipi_wait(100))
1095 panic("Failed to deliver second STARTUP IPI to APIC %d",
1098 DELAY(200); /* wait ~200uS */
1102 * Send an IPI to specified CPU handling the bitmap logic.
1105 ipi_send_cpu(int cpu, u_int ipi)
1107 u_int bitmap, old_pending, new_pending;
1109 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1111 if (IPI_IS_BITMAPED(ipi)) {
1113 ipi = IPI_BITMAP_VECTOR;
1115 old_pending = cpu_ipi_pending[cpu];
1116 new_pending = old_pending | bitmap;
1117 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],
1118 old_pending, new_pending));
1122 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1126 ipi_bitmap_handler(struct trapframe frame)
1128 struct trapframe *oldframe;
1130 int cpu = PCPU_GET(cpuid);
1135 td->td_intr_nesting_level++;
1136 oldframe = td->td_intr_frame;
1137 td->td_intr_frame = &frame;
1138 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1139 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1141 (*ipi_preempt_counts[cpu])++;
1145 if (ipi_bitmap & (1 << IPI_AST)) {
1147 (*ipi_ast_counts[cpu])++;
1149 /* Nothing to do for AST */
1151 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1153 (*ipi_hardclock_counts[cpu])++;
1157 td->td_intr_frame = oldframe;
1158 td->td_intr_nesting_level--;
1163 * send an IPI to a set of cpus.
1166 ipi_selected(cpuset_t cpus, u_int ipi)
1171 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1172 * of help in order to understand what is the source.
1173 * Set the mask of receiving CPUs for this purpose.
1175 if (ipi == IPI_STOP_HARD)
1176 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &cpus);
1178 while ((cpu = CPU_FFS(&cpus)) != 0) {
1180 CPU_CLR(cpu, &cpus);
1181 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1182 ipi_send_cpu(cpu, ipi);
1187 * send an IPI to a specific CPU.
1190 ipi_cpu(int cpu, u_int ipi)
1194 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1195 * of help in order to understand what is the source.
1196 * Set the mask of receiving CPUs for this purpose.
1198 if (ipi == IPI_STOP_HARD)
1199 CPU_SET_ATOMIC(cpu, &ipi_stop_nmi_pending);
1201 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1202 ipi_send_cpu(cpu, ipi);
1206 * send an IPI to all CPUs EXCEPT myself
1209 ipi_all_but_self(u_int ipi)
1211 cpuset_t other_cpus;
1213 other_cpus = all_cpus;
1214 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1215 if (IPI_IS_BITMAPED(ipi)) {
1216 ipi_selected(other_cpus, ipi);
1221 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1222 * of help in order to understand what is the source.
1223 * Set the mask of receiving CPUs for this purpose.
1225 if (ipi == IPI_STOP_HARD)
1226 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &other_cpus);
1228 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1229 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1233 ipi_nmi_handler(void)
1238 * As long as there is not a simple way to know about a NMI's
1239 * source, if the bitmask for the current CPU is present in
1240 * the global pending bitword an IPI_STOP_HARD has been issued
1241 * and should be handled.
1243 cpuid = PCPU_GET(cpuid);
1244 if (!CPU_ISSET(cpuid, &ipi_stop_nmi_pending))
1247 CPU_CLR_ATOMIC(cpuid, &ipi_stop_nmi_pending);
1256 nmi_call_kdb_smp(u_int type, struct trapframe *frame)
1261 cpu = PCPU_GET(cpuid);
1262 if (atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) {
1263 nmi_call_kdb(cpu, type, frame);
1266 savectx(&stoppcbs[cpu]);
1267 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1268 while (!atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1))
1272 atomic_store_rel_int(&nmi_kdb_lock, 0);
1274 cpustop_handler_post(cpu);
1279 * Handle an IPI_STOP by saving our current context and spinning until we
1283 cpustop_handler(void)
1287 cpu = PCPU_GET(cpuid);
1289 savectx(&stoppcbs[cpu]);
1291 /* Indicate that we are stopped */
1292 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1294 /* Wait for restart */
1295 while (!CPU_ISSET(cpu, &started_cpus))
1298 cpustop_handler_post(cpu);
1302 cpustop_handler_post(u_int cpu)
1305 CPU_CLR_ATOMIC(cpu, &started_cpus);
1306 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1309 * We don't broadcast TLB invalidations to other CPUs when they are
1310 * stopped. Hence, we clear the TLB before resuming.
1314 #if defined(__amd64__) && defined(DDB)
1315 amd64_db_resume_dbreg();
1318 if (cpu == 0 && cpustop_restartfunc != NULL) {
1319 cpustop_restartfunc();
1320 cpustop_restartfunc = NULL;
1325 * Handle an IPI_SUSPEND by saving our current context and spinning until we
1329 cpususpend_handler(void)
1333 mtx_assert(&smp_ipi_mtx, MA_NOTOWNED);
1335 cpu = PCPU_GET(cpuid);
1336 if (savectx(&susppcbs[cpu]->sp_pcb)) {
1338 fpususpend(susppcbs[cpu]->sp_fpususpend);
1340 npxsuspend(susppcbs[cpu]->sp_fpususpend);
1343 CPU_SET_ATOMIC(cpu, &suspended_cpus);
1346 fpuresume(susppcbs[cpu]->sp_fpususpend);
1348 npxresume(susppcbs[cpu]->sp_fpususpend);
1352 PCPU_SET(switchtime, 0);
1353 PCPU_SET(switchticks, ticks);
1355 /* Indicate that we are resumed */
1356 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1359 /* Wait for resume */
1360 while (!CPU_ISSET(cpu, &started_cpus))
1363 if (cpu_ops.cpu_resume)
1364 cpu_ops.cpu_resume();
1370 /* Resume MCA and local APIC */
1375 /* Indicate that we are resumed */
1376 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1377 CPU_CLR_ATOMIC(cpu, &started_cpus);
1382 invlcache_handler(void)
1384 uint32_t generation;
1387 (*ipi_invlcache_counts[PCPU_GET(cpuid)])++;
1388 #endif /* COUNT_IPIS */
1391 * Reading the generation here allows greater parallelism
1392 * since wbinvd is a serializing instruction. Without the
1393 * temporary, we'd wait for wbinvd to complete, then the read
1394 * would execute, then the dependent write, which must then
1395 * complete before return from interrupt.
1397 generation = smp_tlb_generation;
1399 PCPU_SET(smp_tlb_done, generation);
1403 * This is called once the rest of the system is up and running and we're
1404 * ready to let the AP's out of the pen.
1407 release_aps(void *dummy __unused)
1412 atomic_store_rel_int(&aps_ready, 1);
1413 while (smp_started == 0)
1416 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1420 * Setup interrupt counters for IPI handlers.
1423 mp_ipi_intrcnt(void *dummy)
1429 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1430 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1431 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1432 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1433 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1434 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1435 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1436 intrcnt_add(buf, &ipi_invlcache_counts[i]);
1437 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1438 intrcnt_add(buf, &ipi_preempt_counts[i]);
1439 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1440 intrcnt_add(buf, &ipi_ast_counts[i]);
1441 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1442 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1443 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1444 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1447 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);
1451 * Flush the TLB on other CPU's
1454 /* Variables needed for SMP tlb shootdown. */
1455 static vm_offset_t smp_tlb_addr1, smp_tlb_addr2;
1456 pmap_t smp_tlb_pmap;
1457 volatile uint32_t smp_tlb_generation;
1460 #define read_eflags() read_rflags()
1464 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap,
1465 vm_offset_t addr1, vm_offset_t addr2)
1467 cpuset_t other_cpus;
1468 volatile uint32_t *p_cpudone;
1469 uint32_t generation;
1472 /* It is not necessary to signal other CPUs while in the debugger. */
1473 if (kdb_active || panicstr != NULL)
1477 * Check for other cpus. Return if none.
1479 if (CPU_ISFULLSET(&mask)) {
1483 CPU_CLR(PCPU_GET(cpuid), &mask);
1484 if (CPU_EMPTY(&mask))
1488 if (!(read_eflags() & PSL_I))
1489 panic("%s: interrupts disabled", __func__);
1490 mtx_lock_spin(&smp_ipi_mtx);
1491 smp_tlb_addr1 = addr1;
1492 smp_tlb_addr2 = addr2;
1493 smp_tlb_pmap = pmap;
1494 generation = ++smp_tlb_generation;
1495 if (CPU_ISFULLSET(&mask)) {
1496 ipi_all_but_self(vector);
1497 other_cpus = all_cpus;
1498 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1501 while ((cpu = CPU_FFS(&mask)) != 0) {
1503 CPU_CLR(cpu, &mask);
1504 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__,
1506 ipi_send_cpu(cpu, vector);
1509 while ((cpu = CPU_FFS(&other_cpus)) != 0) {
1511 CPU_CLR(cpu, &other_cpus);
1512 p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done;
1513 while (*p_cpudone != generation)
1516 mtx_unlock_spin(&smp_ipi_mtx);
1520 smp_masked_invltlb(cpuset_t mask, pmap_t pmap)
1524 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0);
1525 #ifdef COUNT_XINVLTLB_HITS
1532 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr)
1536 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, NULL, addr, 0);
1537 #ifdef COUNT_XINVLTLB_HITS
1544 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2)
1548 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, NULL,
1550 #ifdef COUNT_XINVLTLB_HITS
1552 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1558 smp_cache_flush(void)
1562 smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL,
1568 * Handlers for TLB related IPIs
1571 invltlb_handler(void)
1573 uint32_t generation;
1575 #ifdef COUNT_XINVLTLB_HITS
1576 xhits_gbl[PCPU_GET(cpuid)]++;
1577 #endif /* COUNT_XINVLTLB_HITS */
1579 (*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
1580 #endif /* COUNT_IPIS */
1583 * Reading the generation here allows greater parallelism
1584 * since invalidating the TLB is a serializing operation.
1586 generation = smp_tlb_generation;
1587 if (smp_tlb_pmap == kernel_pmap)
1591 PCPU_SET(smp_tlb_done, generation);
1595 invlpg_handler(void)
1597 uint32_t generation;
1599 #ifdef COUNT_XINVLTLB_HITS
1600 xhits_pg[PCPU_GET(cpuid)]++;
1601 #endif /* COUNT_XINVLTLB_HITS */
1603 (*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
1604 #endif /* COUNT_IPIS */
1606 generation = smp_tlb_generation; /* Overlap with serialization */
1607 invlpg(smp_tlb_addr1);
1608 PCPU_SET(smp_tlb_done, generation);
1612 invlrng_handler(void)
1614 vm_offset_t addr, addr2;
1615 uint32_t generation;
1617 #ifdef COUNT_XINVLTLB_HITS
1618 xhits_rng[PCPU_GET(cpuid)]++;
1619 #endif /* COUNT_XINVLTLB_HITS */
1621 (*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
1622 #endif /* COUNT_IPIS */
1624 addr = smp_tlb_addr1;
1625 addr2 = smp_tlb_addr2;
1626 generation = smp_tlb_generation; /* Overlap with serialization */
1630 } while (addr < addr2);
1632 PCPU_SET(smp_tlb_done, generation);