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>
48 #include <sys/kernel.h>
51 #include <sys/malloc.h>
52 #include <sys/memrange.h>
53 #include <sys/mutex.h>
56 #include <sys/sched.h>
58 #include <sys/sysctl.h>
61 #include <vm/vm_param.h>
63 #include <vm/vm_kern.h>
64 #include <vm/vm_extern.h>
66 #include <x86/apicreg.h>
67 #include <machine/clock.h>
68 #include <machine/cputypes.h>
70 #include <machine/md_var.h>
71 #include <machine/pcb.h>
72 #include <machine/psl.h>
73 #include <machine/smp.h>
74 #include <machine/specialreg.h>
75 #include <machine/cpu.h>
77 #define WARMBOOT_TARGET 0
78 #define WARMBOOT_OFF (KERNBASE + 0x0467)
79 #define WARMBOOT_SEG (KERNBASE + 0x0469)
81 #define CMOS_REG (0x70)
82 #define CMOS_DATA (0x71)
83 #define BIOS_RESET (0x0f)
84 #define BIOS_WARM (0x0a)
86 /* lock region used by kernel profiling */
89 int mp_naps; /* # of Applications processors */
90 int boot_cpu_id = -1; /* designated BSP */
92 extern struct pcpu __pcpu[];
94 /* AP uses this during bootstrap. Do not staticize. */
98 /* Free these after use */
99 void *bootstacks[MAXCPU];
102 struct pcb stoppcbs[MAXCPU];
103 struct susppcb **susppcbs;
106 /* Interrupt counts. */
107 static u_long *ipi_preempt_counts[MAXCPU];
108 static u_long *ipi_ast_counts[MAXCPU];
109 u_long *ipi_invltlb_counts[MAXCPU];
110 u_long *ipi_invlrng_counts[MAXCPU];
111 u_long *ipi_invlpg_counts[MAXCPU];
112 u_long *ipi_invlcache_counts[MAXCPU];
113 u_long *ipi_rendezvous_counts[MAXCPU];
114 static u_long *ipi_hardclock_counts[MAXCPU];
117 /* Default cpu_ops implementation. */
118 struct cpu_ops cpu_ops;
121 * Local data and functions.
124 static volatile cpuset_t ipi_stop_nmi_pending;
126 /* used to hold the AP's until we are ready to release them */
127 struct mtx ap_boot_mtx;
129 /* Set to 1 once we're ready to let the APs out of the pen. */
130 volatile int aps_ready = 0;
133 * Store data from cpu_add() until later in the boot when we actually setup
136 struct cpu_info cpu_info[MAX_APIC_ID + 1];
137 int apic_cpuids[MAX_APIC_ID + 1];
138 int cpu_apic_ids[MAXCPU];
140 /* Holds pending bitmap based IPIs per CPU */
141 volatile u_int cpu_ipi_pending[MAXCPU];
143 static void release_aps(void *dummy);
144 static void cpustop_handler_post(u_int cpu);
146 static int hyperthreading_allowed = 1;
147 SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_allowed, CTLFLAG_RDTUN,
148 &hyperthreading_allowed, 0, "Use Intel HTT logical CPUs");
150 static struct topo_node topo_root;
152 static int pkg_id_shift;
153 static int core_id_shift;
154 static int disabled_cpus;
159 } static caches[MAX_CACHE_LEVELS];
162 mem_range_AP_init(void)
165 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
166 mem_range_softc.mr_op->initAP(&mem_range_softc);
170 * Round up to the next power of two, if necessary, and then
172 * Returns -1 if argument is zero.
178 return (fls(x << (1 - powerof2(x))) - 1);
182 * Add a cache level to the cache topology description.
185 add_deterministic_cache(int type, int level, int share_count)
191 printf("unexpected cache type %d\n", type);
194 if (type == 2) /* ignore instruction cache */
196 if (level == 0 || level > MAX_CACHE_LEVELS) {
197 printf("unexpected cache level %d\n", type);
201 if (caches[level - 1].present) {
202 printf("WARNING: multiple entries for L%u data cache\n", level);
203 printf("%u => %u\n", caches[level - 1].id_shift,
204 mask_width(share_count));
206 caches[level - 1].id_shift = mask_width(share_count);
207 caches[level - 1].present = 1;
209 if (caches[level - 1].id_shift > pkg_id_shift) {
210 printf("WARNING: L%u data cache covers more "
211 "APIC IDs than a package\n", level);
212 printf("%u > %u\n", caches[level - 1].id_shift, pkg_id_shift);
213 caches[level - 1].id_shift = pkg_id_shift;
215 if (caches[level - 1].id_shift < core_id_shift) {
216 printf("WARNING: L%u data cache covers less "
217 "APIC IDs than a core\n", level);
218 printf("%u < %u\n", caches[level - 1].id_shift, core_id_shift);
219 caches[level - 1].id_shift = core_id_shift;
226 * Determine topology of processing units and caches for AMD CPUs.
228 * - AMD CPUID Specification (Publication # 25481)
229 * - BKDG for AMD NPT Family 0Fh Processors (Publication # 32559)
230 * - BKDG For AMD Family 10h Processors (Publication # 31116)
231 * - BKDG For AMD Family 15h Models 00h-0Fh Processors (Publication # 42301)
232 * - BKDG For AMD Family 16h Models 00h-0Fh Processors (Publication # 48751)
240 int nodes_per_socket;
245 /* No multi-core capability. */
246 if ((amd_feature2 & AMDID2_CMP) == 0)
249 /* For families 10h and newer. */
250 pkg_id_shift = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
251 AMDID_COREID_SIZE_SHIFT;
253 /* For 0Fh family. */
254 if (pkg_id_shift == 0)
256 mask_width((cpu_procinfo2 & AMDID_CMP_CORES) + 1);
259 * Families prior to 16h define the following value as
260 * cores per compute unit and we don't really care about the AMD
261 * compute units at the moment. Perhaps we should treat them as
262 * cores and cores within the compute units as hardware threads,
263 * but that's up for debate.
264 * Later families define the value as threads per compute unit,
265 * so we are following AMD's nomenclature here.
267 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0 &&
268 CPUID_TO_FAMILY(cpu_id) >= 0x16) {
269 cpuid_count(0x8000001e, 0, p);
270 share_count = ((p[1] >> 8) & 0xff) + 1;
271 core_id_shift = mask_width(share_count);
274 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0) {
276 cpuid_count(0x8000001d, i, p);
278 level = (p[0] >> 5) & 0x7;
279 share_count = 1 + ((p[0] >> 14) & 0xfff);
281 if (!add_deterministic_cache(type, level, share_count))
285 if (cpu_exthigh >= 0x80000005) {
286 cpuid_count(0x80000005, 0, p);
287 if (((p[2] >> 24) & 0xff) != 0) {
288 caches[0].id_shift = 0;
289 caches[0].present = 1;
292 if (cpu_exthigh >= 0x80000006) {
293 cpuid_count(0x80000006, 0, p);
294 if (((p[2] >> 16) & 0xffff) != 0) {
295 caches[1].id_shift = 0;
296 caches[1].present = 1;
298 if (((p[3] >> 18) & 0x3fff) != 0) {
299 nodes_per_socket = 1;
300 if ((amd_feature2 & AMDID2_NODE_ID) != 0) {
302 * Handle multi-node processors that
303 * have multiple chips, each with its
304 * own L3 cache, on the same die.
306 v = rdmsr(0xc001100c);
307 nodes_per_socket = 1 + ((v >> 3) & 0x7);
310 pkg_id_shift - mask_width(nodes_per_socket);
311 caches[2].present = 1;
318 * Determine topology of processing units for Intel CPUs
319 * using CPUID Leaf 1 and Leaf 4, if supported.
321 * - Intel 64 Architecture Processor Topology Enumeration
322 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
323 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
324 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
327 topo_probe_intel_0x4(void)
333 /* Both zero and one here mean one logical processor per package. */
334 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
335 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
336 if (max_logical <= 1)
339 if (cpu_high >= 0x4) {
340 cpuid_count(0x04, 0, p);
341 max_cores = ((p[0] >> 26) & 0x3f) + 1;
345 core_id_shift = mask_width(max_logical/max_cores);
346 KASSERT(core_id_shift >= 0,
347 ("intel topo: max_cores > max_logical\n"));
348 pkg_id_shift = core_id_shift + mask_width(max_cores);
352 * Determine topology of processing units for Intel CPUs
353 * using CPUID Leaf 11, if supported.
355 * - Intel 64 Architecture Processor Topology Enumeration
356 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
357 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
358 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
361 topo_probe_intel_0xb(void)
368 /* Fall back if CPU leaf 11 doesn't really exist. */
369 cpuid_count(0x0b, 0, p);
371 topo_probe_intel_0x4();
375 /* We only support three levels for now. */
377 cpuid_count(0x0b, i, p);
380 type = (p[2] >> 8) & 0xff;
385 /* TODO: check for duplicate (re-)assignment */
386 if (type == CPUID_TYPE_SMT)
387 core_id_shift = bits;
388 else if (type == CPUID_TYPE_CORE)
391 printf("unknown CPU level type %d\n", type);
394 if (pkg_id_shift < core_id_shift) {
395 printf("WARNING: core covers more APIC IDs than a package\n");
396 core_id_shift = pkg_id_shift;
401 * Determine topology of caches for Intel CPUs.
403 * - Intel 64 Architecture Processor Topology Enumeration
404 * - Intel 64 and IA-32 Architectures Software Developer’s Manual
405 * Volume 2A: Instruction Set Reference, A-M,
409 topo_probe_intel_caches(void)
417 if (cpu_high < 0x4) {
419 * Available cache level and sizes can be determined
420 * via CPUID leaf 2, but that requires a huge table of hardcoded
421 * values, so for now just assume L1 and L2 caches potentially
422 * shared only by HTT processing units, if HTT is present.
424 caches[0].id_shift = pkg_id_shift;
425 caches[0].present = 1;
426 caches[1].id_shift = pkg_id_shift;
427 caches[1].present = 1;
432 cpuid_count(0x4, i, p);
434 level = (p[0] >> 5) & 0x7;
435 share_count = 1 + ((p[0] >> 14) & 0xfff);
437 if (!add_deterministic_cache(type, level, share_count))
443 * Determine topology of processing units and caches for Intel CPUs.
445 * - Intel 64 Architecture Processor Topology Enumeration
448 topo_probe_intel(void)
452 * Note that 0x1 <= cpu_high < 4 case should be
453 * compatible with topo_probe_intel_0x4() logic when
454 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
455 * or it should trigger the fallback otherwise.
458 topo_probe_intel_0xb();
459 else if (cpu_high >= 0x1)
460 topo_probe_intel_0x4();
462 topo_probe_intel_caches();
466 * Topology information is queried only on BSP, on which this
467 * code runs and for which it can query CPUID information.
468 * Then topology is extrapolated on all packages using an
469 * assumption that APIC ID to hardware component ID mapping is
471 * That doesn't necesserily imply that the topology is uniform.
476 static int cpu_topo_probed = 0;
477 struct x86_topo_layer {
481 } topo_layers[MAX_CACHE_LEVELS + 3];
482 struct topo_node *parent;
483 struct topo_node *node;
492 CPU_ZERO(&logical_cpus_mask);
496 else if (cpu_vendor_id == CPU_VENDOR_AMD)
498 else if (cpu_vendor_id == CPU_VENDOR_INTEL)
501 KASSERT(pkg_id_shift >= core_id_shift,
502 ("bug in APIC topology discovery"));
505 bzero(topo_layers, sizeof(topo_layers));
507 topo_layers[nlayers].type = TOPO_TYPE_PKG;
508 topo_layers[nlayers].id_shift = pkg_id_shift;
510 printf("Package ID shift: %u\n", topo_layers[nlayers].id_shift);
514 * Consider all caches to be within a package/chip
515 * and "in front" of all sub-components like
516 * cores and hardware threads.
518 for (i = MAX_CACHE_LEVELS - 1; i >= 0; --i) {
519 if (caches[i].present) {
520 KASSERT(caches[i].id_shift <= pkg_id_shift,
521 ("bug in APIC topology discovery"));
522 KASSERT(caches[i].id_shift >= core_id_shift,
523 ("bug in APIC topology discovery"));
525 topo_layers[nlayers].type = TOPO_TYPE_CACHE;
526 topo_layers[nlayers].subtype = i + 1;
527 topo_layers[nlayers].id_shift = caches[i].id_shift;
529 printf("L%u cache ID shift: %u\n",
530 topo_layers[nlayers].subtype,
531 topo_layers[nlayers].id_shift);
536 if (pkg_id_shift > core_id_shift) {
537 topo_layers[nlayers].type = TOPO_TYPE_CORE;
538 topo_layers[nlayers].id_shift = core_id_shift;
540 printf("Core ID shift: %u\n",
541 topo_layers[nlayers].id_shift);
545 topo_layers[nlayers].type = TOPO_TYPE_PU;
546 topo_layers[nlayers].id_shift = 0;
549 topo_init_root(&topo_root);
550 for (i = 0; i <= MAX_APIC_ID; ++i) {
551 if (!cpu_info[i].cpu_present)
555 for (layer = 0; layer < nlayers; ++layer) {
556 node_id = i >> topo_layers[layer].id_shift;
557 parent = topo_add_node_by_hwid(parent, node_id,
558 topo_layers[layer].type,
559 topo_layers[layer].subtype);
564 for (layer = 0; layer < nlayers; ++layer) {
565 node_id = boot_cpu_id >> topo_layers[layer].id_shift;
566 node = topo_find_node_by_hwid(parent, node_id,
567 topo_layers[layer].type,
568 topo_layers[layer].subtype);
569 topo_promote_child(node);
577 * Assign logical CPU IDs to local APICs.
582 struct topo_node *node;
585 smt_mask = (1u << core_id_shift) - 1;
588 * Assign CPU IDs to local APIC IDs and disable any CPUs
589 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
592 TOPO_FOREACH(node, &topo_root) {
593 if (node->type != TOPO_TYPE_PU)
596 if ((node->hwid & smt_mask) != (boot_cpu_id & smt_mask))
597 cpu_info[node->hwid].cpu_hyperthread = 1;
599 if (resource_disabled("lapic", node->hwid)) {
600 if (node->hwid != boot_cpu_id)
601 cpu_info[node->hwid].cpu_disabled = 1;
603 printf("Cannot disable BSP, APIC ID = %d\n",
607 if (!hyperthreading_allowed &&
608 cpu_info[node->hwid].cpu_hyperthread)
609 cpu_info[node->hwid].cpu_disabled = 1;
611 if (mp_ncpus >= MAXCPU)
612 cpu_info[node->hwid].cpu_disabled = 1;
614 if (cpu_info[node->hwid].cpu_disabled) {
619 cpu_apic_ids[mp_ncpus] = node->hwid;
620 apic_cpuids[node->hwid] = mp_ncpus;
621 topo_set_pu_id(node, mp_ncpus);
625 KASSERT(mp_maxid >= mp_ncpus - 1,
626 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
631 * Print various information about the SMP system hardware and setup.
634 cpu_mp_announce(void)
636 struct topo_node *node;
637 const char *hyperthread;
642 printf("FreeBSD/SMP: ");
643 if (topo_analyze(&topo_root, 1, &pkg_count,
644 &cores_per_pkg, &thrs_per_core)) {
645 printf("%d package(s)", pkg_count);
646 if (cores_per_pkg > 0)
647 printf(" x %d core(s)", cores_per_pkg);
648 if (thrs_per_core > 1)
649 printf(" x %d hardware threads", thrs_per_core);
651 printf("Non-uniform topology");
656 printf("FreeBSD/SMP Online: ");
657 if (topo_analyze(&topo_root, 0, &pkg_count,
658 &cores_per_pkg, &thrs_per_core)) {
659 printf("%d package(s)", pkg_count);
660 if (cores_per_pkg > 0)
661 printf(" x %d core(s)", cores_per_pkg);
662 if (thrs_per_core > 1)
663 printf(" x %d hardware threads", thrs_per_core);
665 printf("Non-uniform topology");
673 TOPO_FOREACH(node, &topo_root) {
674 switch (node->type) {
676 printf("Package HW ID = %u (%#x)\n",
677 node->hwid, node->hwid);
680 printf("\tCore HW ID = %u (%#x)\n",
681 node->hwid, node->hwid);
684 if (cpu_info[node->hwid].cpu_hyperthread)
689 if (node->subtype == 0)
690 printf("\t\tCPU (AP%s): APIC ID: %u (%#x)"
691 "(disabled)\n", hyperthread, node->hwid,
693 else if (node->id == 0)
694 printf("\t\tCPU0 (BSP): APIC ID: %u (%#x)\n",
695 node->hwid, node->hwid);
697 printf("\t\tCPU%u (AP%s): APIC ID: %u (%#x)\n",
698 node->id, hyperthread, node->hwid,
709 * Add a scheduling group, a group of logical processors sharing
710 * a particular cache (and, thus having an affinity), to the scheduling
712 * This function recursively works on lower level caches.
715 x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
717 struct topo_node *node;
722 KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE,
723 ("x86topo_add_sched_group: bad type: %u", root->type));
724 CPU_COPY(&root->cpuset, &cg_root->cg_mask);
725 cg_root->cg_count = root->cpu_count;
726 if (root->type == TOPO_TYPE_SYSTEM)
727 cg_root->cg_level = CG_SHARE_NONE;
729 cg_root->cg_level = root->subtype;
732 * Check how many core nodes we have under the given root node.
733 * If we have multiple logical processors, but not multiple
734 * cores, then those processors must be hardware threads.
738 while (node != NULL) {
739 if (node->type != TOPO_TYPE_CORE) {
740 node = topo_next_node(root, node);
745 node = topo_next_nonchild_node(root, node);
748 if (cg_root->cg_level != CG_SHARE_NONE &&
749 root->cpu_count > 1 && ncores < 2)
750 cg_root->cg_flags = CG_FLAG_SMT;
753 * Find out how many cache nodes we have under the given root node.
754 * We ignore cache nodes that cover all the same processors as the
755 * root node. Also, we do not descend below found cache nodes.
756 * That is, we count top-level "non-redundant" caches under the root
761 while (node != NULL) {
762 if (node->type != TOPO_TYPE_CACHE ||
763 (root->type != TOPO_TYPE_SYSTEM &&
764 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
765 node = topo_next_node(root, node);
769 node = topo_next_nonchild_node(root, node);
772 cg_root->cg_child = smp_topo_alloc(nchildren);
773 cg_root->cg_children = nchildren;
776 * Now find again the same cache nodes as above and recursively
777 * build scheduling topologies for them.
781 while (node != NULL) {
782 if (node->type != TOPO_TYPE_CACHE ||
783 (root->type != TOPO_TYPE_SYSTEM &&
784 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
785 node = topo_next_node(root, node);
788 cg_root->cg_child[i].cg_parent = cg_root;
789 x86topo_add_sched_group(node, &cg_root->cg_child[i]);
791 node = topo_next_nonchild_node(root, node);
796 * Build the MI scheduling topology from the discovered hardware topology.
801 struct cpu_group *cg_root;
804 return (smp_topo_none());
806 cg_root = smp_topo_alloc(1);
807 x86topo_add_sched_group(&topo_root, cg_root);
813 * Add a logical CPU to the topology.
816 cpu_add(u_int apic_id, char boot_cpu)
819 if (apic_id > MAX_APIC_ID) {
820 panic("SMP: APIC ID %d too high", apic_id);
823 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %d added twice",
825 cpu_info[apic_id].cpu_present = 1;
827 KASSERT(boot_cpu_id == -1,
828 ("CPU %d claims to be BSP, but CPU %d already is", apic_id,
830 boot_cpu_id = apic_id;
831 cpu_info[apic_id].cpu_bsp = 1;
833 if (mp_ncpus < MAXCPU) {
835 mp_maxid = mp_ncpus - 1;
838 printf("SMP: Added CPU %d (%s)\n", apic_id, boot_cpu ? "BSP" :
843 cpu_mp_setmaxid(void)
847 * mp_ncpus and mp_maxid should be already set by calls to cpu_add().
848 * If there were no calls to cpu_add() assume this is a UP system.
859 * Always record BSP in CPU map so that the mbuf init code works
862 CPU_SETOF(0, &all_cpus);
863 return (mp_ncpus > 1);
867 * AP CPU's call this to initialize themselves.
870 init_secondary_tail(void)
875 * On real hardware, switch to x2apic mode if possible. Do it
876 * after aps_ready was signalled, to avoid manipulating the
877 * mode while BSP might still want to send some IPI to us
878 * (second startup IPI is ignored on modern hardware etc).
882 /* Initialize the PAT MSR. */
885 /* set up CPU registers and state */
891 /* set up FPU state on the AP */
898 if (cpu_ops.cpu_init)
901 /* A quick check from sanity claus */
902 cpuid = PCPU_GET(cpuid);
903 if (PCPU_GET(apic_id) != lapic_id()) {
904 printf("SMP: cpuid = %d\n", cpuid);
905 printf("SMP: actual apic_id = %d\n", lapic_id());
906 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
907 panic("cpuid mismatch! boom!!");
910 /* Initialize curthread. */
911 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
912 PCPU_SET(curthread, PCPU_GET(idlethread));
916 mtx_lock_spin(&ap_boot_mtx);
918 /* Init local apic for irq's */
921 /* Set memory range attributes for this CPU to match the BSP */
926 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
927 printf("SMP: AP CPU #%d Launched!\n", cpuid);
929 /* Determine if we are a logical CPU. */
930 if (cpu_info[PCPU_GET(apic_id)].cpu_hyperthread)
931 CPU_SET(cpuid, &logical_cpus_mask);
936 if (smp_cpus == mp_ncpus) {
937 /* enable IPI's, tlb shootdown, freezes etc */
938 atomic_store_rel_int(&smp_started, 1);
943 * Enable global pages TLB extension
944 * This also implicitly flushes the TLB
946 load_cr4(rcr4() | CR4_PGE);
947 if (pmap_pcid_enabled)
948 load_cr4(rcr4() | CR4_PCIDE);
954 mtx_unlock_spin(&ap_boot_mtx);
956 /* Wait until all the AP's are up. */
957 while (atomic_load_acq_int(&smp_started) == 0)
960 #ifndef EARLY_AP_STARTUP
961 /* Start per-CPU event timers. */
967 panic("scheduler returned us to %s", __func__);
971 /*******************************************************************
972 * local functions and data
976 * We tell the I/O APIC code about all the CPUs we want to receive
977 * interrupts. If we don't want certain CPUs to receive IRQs we
978 * can simply not tell the I/O APIC code about them in this function.
979 * We also do not tell it about the BSP since it tells itself about
980 * the BSP internally to work with UP kernels and on UP machines.
983 set_interrupt_apic_ids(void)
987 for (i = 0; i < MAXCPU; i++) {
988 apic_id = cpu_apic_ids[i];
991 if (cpu_info[apic_id].cpu_bsp)
993 if (cpu_info[apic_id].cpu_disabled)
996 /* Don't let hyperthreads service interrupts. */
997 if (cpu_info[apic_id].cpu_hyperthread)
1005 #ifdef COUNT_XINVLTLB_HITS
1006 u_int xhits_gbl[MAXCPU];
1007 u_int xhits_pg[MAXCPU];
1008 u_int xhits_rng[MAXCPU];
1009 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1010 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1011 sizeof(xhits_gbl), "IU", "");
1012 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1013 sizeof(xhits_pg), "IU", "");
1014 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1015 sizeof(xhits_rng), "IU", "");
1020 u_int ipi_range_size;
1021 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1022 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1023 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1024 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1026 #endif /* COUNT_XINVLTLB_HITS */
1029 * Init and startup IPI.
1032 ipi_startup(int apic_id, int vector)
1036 * This attempts to follow the algorithm described in the
1037 * Intel Multiprocessor Specification v1.4 in section B.4.
1038 * For each IPI, we allow the local APIC ~20us to deliver the
1039 * IPI. If that times out, we panic.
1043 * first we do an INIT IPI: this INIT IPI might be run, resetting
1044 * and running the target CPU. OR this INIT IPI might be latched (P5
1045 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
1048 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1049 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
1050 lapic_ipi_wait(100);
1052 /* Explicitly deassert the INIT IPI. */
1053 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1054 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT,
1057 DELAY(10000); /* wait ~10mS */
1060 * next we do a STARTUP IPI: the previous INIT IPI might still be
1061 * latched, (P5 bug) this 1st STARTUP would then terminate
1062 * immediately, and the previously started INIT IPI would continue. OR
1063 * the previous INIT IPI has already run. and this STARTUP IPI will
1064 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1067 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1068 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1070 if (!lapic_ipi_wait(100))
1071 panic("Failed to deliver first STARTUP IPI to APIC %d",
1073 DELAY(200); /* wait ~200uS */
1076 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1077 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1078 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1079 * recognized after hardware RESET or INIT IPI.
1081 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1082 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1084 if (!lapic_ipi_wait(100))
1085 panic("Failed to deliver second STARTUP IPI to APIC %d",
1088 DELAY(200); /* wait ~200uS */
1092 * Send an IPI to specified CPU handling the bitmap logic.
1095 ipi_send_cpu(int cpu, u_int ipi)
1097 u_int bitmap, old_pending, new_pending;
1099 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1101 if (IPI_IS_BITMAPED(ipi)) {
1103 ipi = IPI_BITMAP_VECTOR;
1105 old_pending = cpu_ipi_pending[cpu];
1106 new_pending = old_pending | bitmap;
1107 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],
1108 old_pending, new_pending));
1112 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1116 ipi_bitmap_handler(struct trapframe frame)
1118 struct trapframe *oldframe;
1120 int cpu = PCPU_GET(cpuid);
1125 td->td_intr_nesting_level++;
1126 oldframe = td->td_intr_frame;
1127 td->td_intr_frame = &frame;
1128 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1129 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1131 (*ipi_preempt_counts[cpu])++;
1135 if (ipi_bitmap & (1 << IPI_AST)) {
1137 (*ipi_ast_counts[cpu])++;
1139 /* Nothing to do for AST */
1141 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1143 (*ipi_hardclock_counts[cpu])++;
1147 td->td_intr_frame = oldframe;
1148 td->td_intr_nesting_level--;
1153 * send an IPI to a set of cpus.
1156 ipi_selected(cpuset_t cpus, u_int ipi)
1161 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1162 * of help in order to understand what is the source.
1163 * Set the mask of receiving CPUs for this purpose.
1165 if (ipi == IPI_STOP_HARD)
1166 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &cpus);
1168 while ((cpu = CPU_FFS(&cpus)) != 0) {
1170 CPU_CLR(cpu, &cpus);
1171 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1172 ipi_send_cpu(cpu, ipi);
1177 * send an IPI to a specific CPU.
1180 ipi_cpu(int cpu, u_int ipi)
1184 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1185 * of help in order to understand what is the source.
1186 * Set the mask of receiving CPUs for this purpose.
1188 if (ipi == IPI_STOP_HARD)
1189 CPU_SET_ATOMIC(cpu, &ipi_stop_nmi_pending);
1191 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1192 ipi_send_cpu(cpu, ipi);
1196 * send an IPI to all CPUs EXCEPT myself
1199 ipi_all_but_self(u_int ipi)
1201 cpuset_t other_cpus;
1203 other_cpus = all_cpus;
1204 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1205 if (IPI_IS_BITMAPED(ipi)) {
1206 ipi_selected(other_cpus, ipi);
1211 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1212 * of help in order to understand what is the source.
1213 * Set the mask of receiving CPUs for this purpose.
1215 if (ipi == IPI_STOP_HARD)
1216 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &other_cpus);
1218 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1219 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1223 ipi_nmi_handler(void)
1228 * As long as there is not a simple way to know about a NMI's
1229 * source, if the bitmask for the current CPU is present in
1230 * the global pending bitword an IPI_STOP_HARD has been issued
1231 * and should be handled.
1233 cpuid = PCPU_GET(cpuid);
1234 if (!CPU_ISSET(cpuid, &ipi_stop_nmi_pending))
1237 CPU_CLR_ATOMIC(cpuid, &ipi_stop_nmi_pending);
1246 nmi_call_kdb_smp(u_int type, struct trapframe *frame)
1251 cpu = PCPU_GET(cpuid);
1252 if (atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) {
1253 nmi_call_kdb(cpu, type, frame);
1256 savectx(&stoppcbs[cpu]);
1257 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1258 while (!atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1))
1262 atomic_store_rel_int(&nmi_kdb_lock, 0);
1264 cpustop_handler_post(cpu);
1269 * Handle an IPI_STOP by saving our current context and spinning until we
1273 cpustop_handler(void)
1277 cpu = PCPU_GET(cpuid);
1279 savectx(&stoppcbs[cpu]);
1281 /* Indicate that we are stopped */
1282 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1284 /* Wait for restart */
1285 while (!CPU_ISSET(cpu, &started_cpus))
1288 cpustop_handler_post(cpu);
1292 cpustop_handler_post(u_int cpu)
1295 CPU_CLR_ATOMIC(cpu, &started_cpus);
1296 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1298 #if defined(__amd64__) && defined(DDB)
1299 amd64_db_resume_dbreg();
1302 if (cpu == 0 && cpustop_restartfunc != NULL) {
1303 cpustop_restartfunc();
1304 cpustop_restartfunc = NULL;
1309 * Handle an IPI_SUSPEND by saving our current context and spinning until we
1313 cpususpend_handler(void)
1317 mtx_assert(&smp_ipi_mtx, MA_NOTOWNED);
1319 cpu = PCPU_GET(cpuid);
1320 if (savectx(&susppcbs[cpu]->sp_pcb)) {
1322 fpususpend(susppcbs[cpu]->sp_fpususpend);
1324 npxsuspend(susppcbs[cpu]->sp_fpususpend);
1327 CPU_SET_ATOMIC(cpu, &suspended_cpus);
1330 fpuresume(susppcbs[cpu]->sp_fpususpend);
1332 npxresume(susppcbs[cpu]->sp_fpususpend);
1336 PCPU_SET(switchtime, 0);
1337 PCPU_SET(switchticks, ticks);
1339 /* Indicate that we are resumed */
1340 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1343 /* Wait for resume */
1344 while (!CPU_ISSET(cpu, &started_cpus))
1347 if (cpu_ops.cpu_resume)
1348 cpu_ops.cpu_resume();
1354 /* Resume MCA and local APIC */
1359 /* Indicate that we are resumed */
1360 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1361 CPU_CLR_ATOMIC(cpu, &started_cpus);
1366 invlcache_handler(void)
1368 uint32_t generation;
1371 (*ipi_invlcache_counts[PCPU_GET(cpuid)])++;
1372 #endif /* COUNT_IPIS */
1375 * Reading the generation here allows greater parallelism
1376 * since wbinvd is a serializing instruction. Without the
1377 * temporary, we'd wait for wbinvd to complete, then the read
1378 * would execute, then the dependent write, which must then
1379 * complete before return from interrupt.
1381 generation = smp_tlb_generation;
1383 PCPU_SET(smp_tlb_done, generation);
1387 * This is called once the rest of the system is up and running and we're
1388 * ready to let the AP's out of the pen.
1391 release_aps(void *dummy __unused)
1396 atomic_store_rel_int(&aps_ready, 1);
1397 while (smp_started == 0)
1400 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1404 * Setup interrupt counters for IPI handlers.
1407 mp_ipi_intrcnt(void *dummy)
1413 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1414 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1415 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1416 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1417 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1418 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1419 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1420 intrcnt_add(buf, &ipi_invlcache_counts[i]);
1421 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1422 intrcnt_add(buf, &ipi_preempt_counts[i]);
1423 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1424 intrcnt_add(buf, &ipi_ast_counts[i]);
1425 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1426 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1427 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1428 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1431 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);
1435 * Flush the TLB on other CPU's
1438 /* Variables needed for SMP tlb shootdown. */
1439 vm_offset_t smp_tlb_addr1, smp_tlb_addr2;
1440 pmap_t smp_tlb_pmap;
1441 volatile uint32_t smp_tlb_generation;
1444 #define read_eflags() read_rflags()
1448 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap,
1449 vm_offset_t addr1, vm_offset_t addr2)
1451 cpuset_t other_cpus;
1452 volatile uint32_t *p_cpudone;
1453 uint32_t generation;
1457 * Check for other cpus. Return if none.
1459 if (CPU_ISFULLSET(&mask)) {
1463 CPU_CLR(PCPU_GET(cpuid), &mask);
1464 if (CPU_EMPTY(&mask))
1468 if (!(read_eflags() & PSL_I))
1469 panic("%s: interrupts disabled", __func__);
1470 mtx_lock_spin(&smp_ipi_mtx);
1471 smp_tlb_addr1 = addr1;
1472 smp_tlb_addr2 = addr2;
1473 smp_tlb_pmap = pmap;
1474 generation = ++smp_tlb_generation;
1475 if (CPU_ISFULLSET(&mask)) {
1476 ipi_all_but_self(vector);
1477 other_cpus = all_cpus;
1478 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1481 while ((cpu = CPU_FFS(&mask)) != 0) {
1483 CPU_CLR(cpu, &mask);
1484 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__,
1486 ipi_send_cpu(cpu, vector);
1489 while ((cpu = CPU_FFS(&other_cpus)) != 0) {
1491 CPU_CLR(cpu, &other_cpus);
1492 p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done;
1493 while (*p_cpudone != generation)
1496 mtx_unlock_spin(&smp_ipi_mtx);
1500 smp_masked_invltlb(cpuset_t mask, pmap_t pmap)
1504 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0);
1505 #ifdef COUNT_XINVLTLB_HITS
1512 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap)
1516 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0);
1517 #ifdef COUNT_XINVLTLB_HITS
1524 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2,
1529 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap,
1531 #ifdef COUNT_XINVLTLB_HITS
1533 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1539 smp_cache_flush(void)
1543 smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL,
1549 * Handlers for TLB related IPIs
1552 invltlb_handler(void)
1554 uint32_t generation;
1556 #ifdef COUNT_XINVLTLB_HITS
1557 xhits_gbl[PCPU_GET(cpuid)]++;
1558 #endif /* COUNT_XINVLTLB_HITS */
1560 (*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
1561 #endif /* COUNT_IPIS */
1564 * Reading the generation here allows greater parallelism
1565 * since invalidating the TLB is a serializing operation.
1567 generation = smp_tlb_generation;
1568 if (smp_tlb_pmap == kernel_pmap)
1572 PCPU_SET(smp_tlb_done, generation);
1576 invlpg_handler(void)
1578 uint32_t generation;
1580 #ifdef COUNT_XINVLTLB_HITS
1581 xhits_pg[PCPU_GET(cpuid)]++;
1582 #endif /* COUNT_XINVLTLB_HITS */
1584 (*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
1585 #endif /* COUNT_IPIS */
1587 generation = smp_tlb_generation; /* Overlap with serialization */
1588 invlpg(smp_tlb_addr1);
1589 PCPU_SET(smp_tlb_done, generation);
1593 invlrng_handler(void)
1595 vm_offset_t addr, addr2;
1596 uint32_t generation;
1598 #ifdef COUNT_XINVLTLB_HITS
1599 xhits_rng[PCPU_GET(cpuid)]++;
1600 #endif /* COUNT_XINVLTLB_HITS */
1602 (*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
1603 #endif /* COUNT_IPIS */
1605 addr = smp_tlb_addr1;
1606 addr2 = smp_tlb_addr2;
1607 generation = smp_tlb_generation; /* Overlap with serialization */
1611 } while (addr < addr2);
1613 PCPU_SET(smp_tlb_done, generation);