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("$FreeBSD$");
34 #include "opt_kstack_pages.h"
36 #include "opt_sched.h"
39 #include <sys/param.h>
40 #include <sys/systm.h>
42 #include <sys/cons.h> /* cngetc() */
43 #include <sys/cpuset.h>
47 #include <sys/interrupt.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>
66 #include <vm/vm_map.h>
68 #include <x86/apicreg.h>
69 #include <machine/clock.h>
70 #include <machine/cpu.h>
71 #include <machine/cputypes.h>
73 #include <machine/md_var.h>
74 #include <machine/pcb.h>
75 #include <machine/psl.h>
76 #include <machine/smp.h>
77 #include <machine/specialreg.h>
78 #include <x86/ucode.h>
80 static MALLOC_DEFINE(M_CPUS, "cpus", "CPU items");
82 /* lock region used by kernel profiling */
85 int mp_naps; /* # of Applications processors */
86 int boot_cpu_id = -1; /* designated BSP */
88 /* AP uses this during bootstrap. Do not staticize. */
92 /* Free these after use */
93 void *bootstacks[MAXCPU];
96 struct pcb stoppcbs[MAXCPU];
97 struct susppcb **susppcbs;
100 /* Interrupt counts. */
101 static u_long *ipi_preempt_counts[MAXCPU];
102 static u_long *ipi_ast_counts[MAXCPU];
103 u_long *ipi_invltlb_counts[MAXCPU];
104 u_long *ipi_invlrng_counts[MAXCPU];
105 u_long *ipi_invlpg_counts[MAXCPU];
106 u_long *ipi_invlcache_counts[MAXCPU];
107 u_long *ipi_rendezvous_counts[MAXCPU];
108 static u_long *ipi_hardclock_counts[MAXCPU];
111 /* Default cpu_ops implementation. */
112 struct cpu_ops cpu_ops;
115 * Local data and functions.
118 static volatile cpuset_t ipi_stop_nmi_pending;
120 volatile cpuset_t resuming_cpus;
121 volatile cpuset_t toresume_cpus;
123 /* used to hold the AP's until we are ready to release them */
124 struct mtx ap_boot_mtx;
126 /* Set to 1 once we're ready to let the APs out of the pen. */
127 volatile int aps_ready = 0;
130 * Store data from cpu_add() until later in the boot when we actually setup
133 struct cpu_info *cpu_info;
135 int cpu_apic_ids[MAXCPU];
136 _Static_assert(MAXCPU <= MAX_APIC_ID,
137 "MAXCPU cannot be larger that MAX_APIC_ID");
138 _Static_assert(xAPIC_MAX_APIC_ID <= MAX_APIC_ID,
139 "xAPIC_MAX_APIC_ID cannot be larger that MAX_APIC_ID");
141 /* Holds pending bitmap based IPIs per CPU */
142 volatile u_int cpu_ipi_pending[MAXCPU];
144 static void release_aps(void *dummy);
145 static void cpustop_handler_post(u_int cpu);
147 static int hyperthreading_allowed = 1;
148 SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_allowed, CTLFLAG_RDTUN,
149 &hyperthreading_allowed, 0, "Use Intel HTT logical CPUs");
151 static struct topo_node topo_root;
153 static int pkg_id_shift;
154 static int node_id_shift;
155 static int core_id_shift;
156 static int disabled_cpus;
161 } static caches[MAX_CACHE_LEVELS];
163 unsigned int boot_address;
165 #define MiB(v) (v ## ULL << 20)
168 mem_range_AP_init(void)
171 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
172 mem_range_softc.mr_op->initAP(&mem_range_softc);
176 * Round up to the next power of two, if necessary, and then
178 * Returns -1 if argument is zero.
184 return (fls(x << (1 - powerof2(x))) - 1);
188 * Add a cache level to the cache topology description.
191 add_deterministic_cache(int type, int level, int share_count)
197 printf("unexpected cache type %d\n", type);
200 if (type == 2) /* ignore instruction cache */
202 if (level == 0 || level > MAX_CACHE_LEVELS) {
203 printf("unexpected cache level %d\n", type);
207 if (caches[level - 1].present) {
208 printf("WARNING: multiple entries for L%u data cache\n", level);
209 printf("%u => %u\n", caches[level - 1].id_shift,
210 mask_width(share_count));
212 caches[level - 1].id_shift = mask_width(share_count);
213 caches[level - 1].present = 1;
215 if (caches[level - 1].id_shift > pkg_id_shift) {
216 printf("WARNING: L%u data cache covers more "
217 "APIC IDs than a package (%u > %u)\n", level,
218 caches[level - 1].id_shift, pkg_id_shift);
219 caches[level - 1].id_shift = pkg_id_shift;
221 if (caches[level - 1].id_shift < core_id_shift) {
222 printf("WARNING: L%u data cache covers fewer "
223 "APIC IDs than a core (%u < %u)\n", level,
224 caches[level - 1].id_shift, core_id_shift);
225 caches[level - 1].id_shift = core_id_shift;
232 * Determine topology of processing units and caches for AMD CPUs.
234 * - AMD CPUID Specification (Publication # 25481)
235 * - BKDG for AMD NPT Family 0Fh Processors (Publication # 32559)
236 * - BKDG For AMD Family 10h Processors (Publication # 31116)
237 * - BKDG For AMD Family 15h Models 00h-0Fh Processors (Publication # 42301)
238 * - BKDG For AMD Family 16h Models 00h-0Fh Processors (Publication # 48751)
239 * - PPR For AMD Family 17h Models 00h-0Fh Processors (Publication # 54945)
247 int nodes_per_socket;
252 /* No multi-core capability. */
253 if ((amd_feature2 & AMDID2_CMP) == 0)
256 /* For families 10h and newer. */
257 pkg_id_shift = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
258 AMDID_COREID_SIZE_SHIFT;
260 /* For 0Fh family. */
261 if (pkg_id_shift == 0)
263 mask_width((cpu_procinfo2 & AMDID_CMP_CORES) + 1);
266 * Families prior to 16h define the following value as
267 * cores per compute unit and we don't really care about the AMD
268 * compute units at the moment. Perhaps we should treat them as
269 * cores and cores within the compute units as hardware threads,
270 * but that's up for debate.
271 * Later families define the value as threads per compute unit,
272 * so we are following AMD's nomenclature here.
274 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0 &&
275 CPUID_TO_FAMILY(cpu_id) >= 0x16) {
276 cpuid_count(0x8000001e, 0, p);
277 share_count = ((p[1] >> 8) & 0xff) + 1;
278 core_id_shift = mask_width(share_count);
281 * For Zen (17h), gather Nodes per Processor. Each node is a
282 * Zeppelin die; TR and EPYC CPUs will have multiple dies per
283 * package. Communication latency between dies is higher than
286 nodes_per_socket = ((p[2] >> 8) & 0x7) + 1;
287 node_id_shift = pkg_id_shift - mask_width(nodes_per_socket);
290 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0) {
292 cpuid_count(0x8000001d, i, p);
294 level = (p[0] >> 5) & 0x7;
295 share_count = 1 + ((p[0] >> 14) & 0xfff);
297 if (!add_deterministic_cache(type, level, share_count))
301 if (cpu_exthigh >= 0x80000005) {
302 cpuid_count(0x80000005, 0, p);
303 if (((p[2] >> 24) & 0xff) != 0) {
304 caches[0].id_shift = 0;
305 caches[0].present = 1;
308 if (cpu_exthigh >= 0x80000006) {
309 cpuid_count(0x80000006, 0, p);
310 if (((p[2] >> 16) & 0xffff) != 0) {
311 caches[1].id_shift = 0;
312 caches[1].present = 1;
314 if (((p[3] >> 18) & 0x3fff) != 0) {
315 nodes_per_socket = 1;
316 if ((amd_feature2 & AMDID2_NODE_ID) != 0) {
318 * Handle multi-node processors that
319 * have multiple chips, each with its
320 * own L3 cache, on the same die.
322 v = rdmsr(0xc001100c);
323 nodes_per_socket = 1 + ((v >> 3) & 0x7);
326 pkg_id_shift - mask_width(nodes_per_socket);
327 caches[2].present = 1;
334 * Determine topology of processing units for Intel CPUs
335 * using CPUID Leaf 1 and Leaf 4, if supported.
337 * - Intel 64 Architecture Processor Topology Enumeration
338 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
339 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
340 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
343 topo_probe_intel_0x4(void)
349 /* Both zero and one here mean one logical processor per package. */
350 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
351 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
352 if (max_logical <= 1)
355 if (cpu_high >= 0x4) {
356 cpuid_count(0x04, 0, p);
357 max_cores = ((p[0] >> 26) & 0x3f) + 1;
361 core_id_shift = mask_width(max_logical/max_cores);
362 KASSERT(core_id_shift >= 0,
363 ("intel topo: max_cores > max_logical\n"));
364 pkg_id_shift = core_id_shift + mask_width(max_cores);
368 * Determine topology of processing units for Intel CPUs
369 * using CPUID Leaf 11, if supported.
371 * - Intel 64 Architecture Processor Topology Enumeration
372 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
373 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
374 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
377 topo_probe_intel_0xb(void)
384 /* Fall back if CPU leaf 11 doesn't really exist. */
385 cpuid_count(0x0b, 0, p);
387 topo_probe_intel_0x4();
391 /* We only support three levels for now. */
393 cpuid_count(0x0b, i, p);
396 type = (p[2] >> 8) & 0xff;
401 /* TODO: check for duplicate (re-)assignment */
402 if (type == CPUID_TYPE_SMT)
403 core_id_shift = bits;
404 else if (type == CPUID_TYPE_CORE)
407 printf("unknown CPU level type %d\n", type);
410 if (pkg_id_shift < core_id_shift) {
411 printf("WARNING: core covers more APIC IDs than a package\n");
412 core_id_shift = pkg_id_shift;
417 * Determine topology of caches for Intel CPUs.
419 * - Intel 64 Architecture Processor Topology Enumeration
420 * - Intel 64 and IA-32 Architectures Software Developer’s Manual
421 * Volume 2A: Instruction Set Reference, A-M,
425 topo_probe_intel_caches(void)
433 if (cpu_high < 0x4) {
435 * Available cache level and sizes can be determined
436 * via CPUID leaf 2, but that requires a huge table of hardcoded
437 * values, so for now just assume L1 and L2 caches potentially
438 * shared only by HTT processing units, if HTT is present.
440 caches[0].id_shift = pkg_id_shift;
441 caches[0].present = 1;
442 caches[1].id_shift = pkg_id_shift;
443 caches[1].present = 1;
448 cpuid_count(0x4, i, p);
450 level = (p[0] >> 5) & 0x7;
451 share_count = 1 + ((p[0] >> 14) & 0xfff);
453 if (!add_deterministic_cache(type, level, share_count))
459 * Determine topology of processing units and caches for Intel CPUs.
461 * - Intel 64 Architecture Processor Topology Enumeration
464 topo_probe_intel(void)
468 * Note that 0x1 <= cpu_high < 4 case should be
469 * compatible with topo_probe_intel_0x4() logic when
470 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
471 * or it should trigger the fallback otherwise.
474 topo_probe_intel_0xb();
475 else if (cpu_high >= 0x1)
476 topo_probe_intel_0x4();
478 topo_probe_intel_caches();
482 * Topology information is queried only on BSP, on which this
483 * code runs and for which it can query CPUID information.
484 * Then topology is extrapolated on all packages using an
485 * assumption that APIC ID to hardware component ID mapping is
487 * That doesn't necesserily imply that the topology is uniform.
492 static int cpu_topo_probed = 0;
493 struct x86_topo_layer {
497 } topo_layers[MAX_CACHE_LEVELS + 4];
498 struct topo_node *parent;
499 struct topo_node *node;
508 CPU_ZERO(&logical_cpus_mask);
512 else if (cpu_vendor_id == CPU_VENDOR_AMD ||
513 cpu_vendor_id == CPU_VENDOR_HYGON)
515 else if (cpu_vendor_id == CPU_VENDOR_INTEL)
518 KASSERT(pkg_id_shift >= core_id_shift,
519 ("bug in APIC topology discovery"));
522 bzero(topo_layers, sizeof(topo_layers));
524 topo_layers[nlayers].type = TOPO_TYPE_PKG;
525 topo_layers[nlayers].id_shift = pkg_id_shift;
527 printf("Package ID shift: %u\n", topo_layers[nlayers].id_shift);
530 if (pkg_id_shift > node_id_shift && node_id_shift != 0) {
531 topo_layers[nlayers].type = TOPO_TYPE_GROUP;
532 topo_layers[nlayers].id_shift = node_id_shift;
534 printf("Node ID shift: %u\n",
535 topo_layers[nlayers].id_shift);
540 * Consider all caches to be within a package/chip
541 * and "in front" of all sub-components like
542 * cores and hardware threads.
544 for (i = MAX_CACHE_LEVELS - 1; i >= 0; --i) {
545 if (caches[i].present) {
546 if (node_id_shift != 0)
547 KASSERT(caches[i].id_shift <= node_id_shift,
548 ("bug in APIC topology discovery"));
549 KASSERT(caches[i].id_shift <= pkg_id_shift,
550 ("bug in APIC topology discovery"));
551 KASSERT(caches[i].id_shift >= core_id_shift,
552 ("bug in APIC topology discovery"));
554 topo_layers[nlayers].type = TOPO_TYPE_CACHE;
555 topo_layers[nlayers].subtype = i + 1;
556 topo_layers[nlayers].id_shift = caches[i].id_shift;
558 printf("L%u cache ID shift: %u\n",
559 topo_layers[nlayers].subtype,
560 topo_layers[nlayers].id_shift);
565 if (pkg_id_shift > core_id_shift) {
566 topo_layers[nlayers].type = TOPO_TYPE_CORE;
567 topo_layers[nlayers].id_shift = core_id_shift;
569 printf("Core ID shift: %u\n",
570 topo_layers[nlayers].id_shift);
574 topo_layers[nlayers].type = TOPO_TYPE_PU;
575 topo_layers[nlayers].id_shift = 0;
578 topo_init_root(&topo_root);
579 for (i = 0; i <= max_apic_id; ++i) {
580 if (!cpu_info[i].cpu_present)
584 for (layer = 0; layer < nlayers; ++layer) {
585 node_id = i >> topo_layers[layer].id_shift;
586 parent = topo_add_node_by_hwid(parent, node_id,
587 topo_layers[layer].type,
588 topo_layers[layer].subtype);
593 for (layer = 0; layer < nlayers; ++layer) {
594 node_id = boot_cpu_id >> topo_layers[layer].id_shift;
595 node = topo_find_node_by_hwid(parent, node_id,
596 topo_layers[layer].type,
597 topo_layers[layer].subtype);
598 topo_promote_child(node);
606 * Assign logical CPU IDs to local APICs.
611 struct topo_node *node;
615 smt_mask = (1u << core_id_shift) - 1;
618 * Assign CPU IDs to local APIC IDs and disable any CPUs
619 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
623 TOPO_FOREACH(node, &topo_root) {
624 if (node->type != TOPO_TYPE_PU)
627 if ((node->hwid & smt_mask) != (boot_cpu_id & smt_mask))
628 cpu_info[node->hwid].cpu_hyperthread = 1;
630 if (resource_disabled("lapic", node->hwid)) {
631 if (node->hwid != boot_cpu_id)
632 cpu_info[node->hwid].cpu_disabled = 1;
634 printf("Cannot disable BSP, APIC ID = %d\n",
638 if (!hyperthreading_allowed &&
639 cpu_info[node->hwid].cpu_hyperthread)
640 cpu_info[node->hwid].cpu_disabled = 1;
642 if (mp_ncpus >= MAXCPU)
643 cpu_info[node->hwid].cpu_disabled = 1;
645 if (cpu_info[node->hwid].cpu_disabled) {
650 if (cpu_info[node->hwid].cpu_hyperthread)
653 cpu_apic_ids[mp_ncpus] = node->hwid;
654 apic_cpuids[node->hwid] = mp_ncpus;
655 topo_set_pu_id(node, mp_ncpus);
659 KASSERT(mp_maxid >= mp_ncpus - 1,
660 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
663 mp_ncores = mp_ncpus - nhyper;
664 smp_threads_per_core = mp_ncpus / mp_ncores;
668 * Print various information about the SMP system hardware and setup.
671 cpu_mp_announce(void)
673 struct topo_node *node;
674 const char *hyperthread;
675 struct topo_analysis topology;
677 printf("FreeBSD/SMP: ");
678 if (topo_analyze(&topo_root, 1, &topology)) {
679 printf("%d package(s)", topology.entities[TOPO_LEVEL_PKG]);
680 if (topology.entities[TOPO_LEVEL_GROUP] > 1)
681 printf(" x %d groups",
682 topology.entities[TOPO_LEVEL_GROUP]);
683 if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1)
684 printf(" x %d cache groups",
685 topology.entities[TOPO_LEVEL_CACHEGROUP]);
686 if (topology.entities[TOPO_LEVEL_CORE] > 0)
687 printf(" x %d core(s)",
688 topology.entities[TOPO_LEVEL_CORE]);
689 if (topology.entities[TOPO_LEVEL_THREAD] > 1)
690 printf(" x %d hardware threads",
691 topology.entities[TOPO_LEVEL_THREAD]);
693 printf("Non-uniform topology");
698 printf("FreeBSD/SMP Online: ");
699 if (topo_analyze(&topo_root, 0, &topology)) {
700 printf("%d package(s)",
701 topology.entities[TOPO_LEVEL_PKG]);
702 if (topology.entities[TOPO_LEVEL_GROUP] > 1)
703 printf(" x %d groups",
704 topology.entities[TOPO_LEVEL_GROUP]);
705 if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1)
706 printf(" x %d cache groups",
707 topology.entities[TOPO_LEVEL_CACHEGROUP]);
708 if (topology.entities[TOPO_LEVEL_CORE] > 0)
709 printf(" x %d core(s)",
710 topology.entities[TOPO_LEVEL_CORE]);
711 if (topology.entities[TOPO_LEVEL_THREAD] > 1)
712 printf(" x %d hardware threads",
713 topology.entities[TOPO_LEVEL_THREAD]);
715 printf("Non-uniform topology");
723 TOPO_FOREACH(node, &topo_root) {
724 switch (node->type) {
726 printf("Package HW ID = %u\n", node->hwid);
729 printf("\tCore HW ID = %u\n", node->hwid);
732 if (cpu_info[node->hwid].cpu_hyperthread)
737 if (node->subtype == 0)
738 printf("\t\tCPU (AP%s): APIC ID: %u"
739 "(disabled)\n", hyperthread, node->hwid);
740 else if (node->id == 0)
741 printf("\t\tCPU0 (BSP): APIC ID: %u\n",
744 printf("\t\tCPU%u (AP%s): APIC ID: %u\n",
745 node->id, hyperthread, node->hwid);
755 * Add a scheduling group, a group of logical processors sharing
756 * a particular cache (and, thus having an affinity), to the scheduling
758 * This function recursively works on lower level caches.
761 x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
763 struct topo_node *node;
768 KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE ||
769 root->type == TOPO_TYPE_GROUP,
770 ("x86topo_add_sched_group: bad type: %u", root->type));
771 CPU_COPY(&root->cpuset, &cg_root->cg_mask);
772 cg_root->cg_count = root->cpu_count;
773 if (root->type == TOPO_TYPE_SYSTEM)
774 cg_root->cg_level = CG_SHARE_NONE;
776 cg_root->cg_level = root->subtype;
779 * Check how many core nodes we have under the given root node.
780 * If we have multiple logical processors, but not multiple
781 * cores, then those processors must be hardware threads.
785 while (node != NULL) {
786 if (node->type != TOPO_TYPE_CORE) {
787 node = topo_next_node(root, node);
792 node = topo_next_nonchild_node(root, node);
795 if (cg_root->cg_level != CG_SHARE_NONE &&
796 root->cpu_count > 1 && ncores < 2)
797 cg_root->cg_flags = CG_FLAG_SMT;
800 * Find out how many cache nodes we have under the given root node.
801 * We ignore cache nodes that cover all the same processors as the
802 * root node. Also, we do not descend below found cache nodes.
803 * That is, we count top-level "non-redundant" caches under the root
808 while (node != NULL) {
809 if ((node->type != TOPO_TYPE_GROUP &&
810 node->type != TOPO_TYPE_CACHE) ||
811 (root->type != TOPO_TYPE_SYSTEM &&
812 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
813 node = topo_next_node(root, node);
817 node = topo_next_nonchild_node(root, node);
820 cg_root->cg_child = smp_topo_alloc(nchildren);
821 cg_root->cg_children = nchildren;
824 * Now find again the same cache nodes as above and recursively
825 * build scheduling topologies for them.
829 while (node != NULL) {
830 if ((node->type != TOPO_TYPE_GROUP &&
831 node->type != TOPO_TYPE_CACHE) ||
832 (root->type != TOPO_TYPE_SYSTEM &&
833 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
834 node = topo_next_node(root, node);
837 cg_root->cg_child[i].cg_parent = cg_root;
838 x86topo_add_sched_group(node, &cg_root->cg_child[i]);
840 node = topo_next_nonchild_node(root, node);
845 * Build the MI scheduling topology from the discovered hardware topology.
850 struct cpu_group *cg_root;
853 return (smp_topo_none());
855 cg_root = smp_topo_alloc(1);
856 x86topo_add_sched_group(&topo_root, cg_root);
861 cpu_alloc(void *dummy __unused)
864 * Dynamically allocate the arrays that depend on the
867 cpu_info = malloc(sizeof(*cpu_info) * (max_apic_id + 1), M_CPUS,
869 apic_cpuids = malloc(sizeof(*apic_cpuids) * (max_apic_id + 1), M_CPUS,
872 SYSINIT(cpu_alloc, SI_SUB_CPU, SI_ORDER_FIRST, cpu_alloc, NULL);
875 * Add a logical CPU to the topology.
878 cpu_add(u_int apic_id, char boot_cpu)
881 if (apic_id > max_apic_id) {
882 panic("SMP: APIC ID %d too high", apic_id);
885 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %u added twice",
887 cpu_info[apic_id].cpu_present = 1;
889 KASSERT(boot_cpu_id == -1,
890 ("CPU %u claims to be BSP, but CPU %u already is", apic_id,
892 boot_cpu_id = apic_id;
893 cpu_info[apic_id].cpu_bsp = 1;
896 printf("SMP: Added CPU %u (%s)\n", apic_id, boot_cpu ? "BSP" :
901 cpu_mp_setmaxid(void)
905 * mp_ncpus and mp_maxid should be already set by calls to cpu_add().
906 * If there were no calls to cpu_add() assume this is a UP system.
917 * Always record BSP in CPU map so that the mbuf init code works
920 CPU_SETOF(0, &all_cpus);
921 return (mp_ncpus > 1);
924 /* Allocate memory for the AP trampoline. */
926 alloc_ap_trampoline(vm_paddr_t *physmap, unsigned int *physmap_idx)
932 for (i = *physmap_idx; i <= *physmap_idx; i -= 2) {
934 * Find a memory region big enough and below the 1MB boundary
935 * for the trampoline code.
936 * NB: needs to be page aligned.
938 if (physmap[i] >= MiB(1) ||
939 (trunc_page(physmap[i + 1]) - round_page(physmap[i])) <
940 round_page(bootMP_size))
945 * Try to steal from the end of the region to mimic previous
946 * behaviour, else fallback to steal from the start.
948 if (physmap[i + 1] < MiB(1)) {
949 boot_address = trunc_page(physmap[i + 1]);
950 if ((physmap[i + 1] - boot_address) < bootMP_size)
951 boot_address -= round_page(bootMP_size);
952 physmap[i + 1] = boot_address;
954 boot_address = round_page(physmap[i]);
955 physmap[i] = boot_address + round_page(bootMP_size);
957 if (physmap[i] == physmap[i + 1] && *physmap_idx != 0) {
958 memmove(&physmap[i], &physmap[i + 2],
959 sizeof(*physmap) * (*physmap_idx - i + 2));
966 boot_address = basemem * 1024 - bootMP_size;
969 "Cannot find enough space for the boot trampoline, placing it at %#x",
975 * AP CPU's call this to initialize themselves.
978 init_secondary_tail(void)
982 pmap_activate_boot(vmspace_pmap(proc0.p_vmspace));
985 * On real hardware, switch to x2apic mode if possible. Do it
986 * after aps_ready was signalled, to avoid manipulating the
987 * mode while BSP might still want to send some IPI to us
988 * (second startup IPI is ignored on modern hardware etc).
992 /* Initialize the PAT MSR. */
995 /* set up CPU registers and state */
1001 /* set up FPU state on the AP */
1008 if (cpu_ops.cpu_init)
1011 /* A quick check from sanity claus */
1012 cpuid = PCPU_GET(cpuid);
1013 if (PCPU_GET(apic_id) != lapic_id()) {
1014 printf("SMP: cpuid = %d\n", cpuid);
1015 printf("SMP: actual apic_id = %d\n", lapic_id());
1016 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
1017 panic("cpuid mismatch! boom!!");
1020 /* Initialize curthread. */
1021 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
1022 PCPU_SET(curthread, PCPU_GET(idlethread));
1024 mtx_lock_spin(&ap_boot_mtx);
1028 /* Init local apic for irq's */
1031 /* Set memory range attributes for this CPU to match the BSP */
1032 mem_range_AP_init();
1036 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
1038 printf("SMP: AP CPU #%d Launched!\n", cpuid);
1040 printf("%s%d%s", smp_cpus == 2 ? "Launching APs: " : "",
1041 cpuid, smp_cpus == mp_ncpus ? "\n" : " ");
1043 /* Determine if we are a logical CPU. */
1044 if (cpu_info[PCPU_GET(apic_id)].cpu_hyperthread)
1045 CPU_SET(cpuid, &logical_cpus_mask);
1050 if (smp_cpus == mp_ncpus) {
1051 /* enable IPI's, tlb shootdown, freezes etc */
1052 atomic_store_rel_int(&smp_started, 1);
1057 * Enable global pages TLB extension
1058 * This also implicitly flushes the TLB
1060 load_cr4(rcr4() | CR4_PGE);
1061 if (pmap_pcid_enabled)
1062 load_cr4(rcr4() | CR4_PCIDE);
1068 mtx_unlock_spin(&ap_boot_mtx);
1070 /* Wait until all the AP's are up. */
1071 while (atomic_load_acq_int(&smp_started) == 0)
1074 #ifndef EARLY_AP_STARTUP
1075 /* Start per-CPU event timers. */
1076 cpu_initclocks_ap();
1080 * Assert that smp_after_idle_runnable condition is reasonable.
1082 MPASS(PCPU_GET(curpcb) == NULL);
1086 panic("scheduler returned us to %s", __func__);
1091 smp_after_idle_runnable(void *arg __unused)
1096 for (cpu = 1; cpu < mp_ncpus; cpu++) {
1097 pc = pcpu_find(cpu);
1098 while (atomic_load_ptr(&pc->pc_curpcb) == (uintptr_t)NULL)
1100 kmem_free((vm_offset_t)bootstacks[cpu], kstack_pages *
1104 SYSINIT(smp_after_idle_runnable, SI_SUB_SMP, SI_ORDER_ANY,
1105 smp_after_idle_runnable, NULL);
1108 * We tell the I/O APIC code about all the CPUs we want to receive
1109 * interrupts. If we don't want certain CPUs to receive IRQs we
1110 * can simply not tell the I/O APIC code about them in this function.
1111 * We also do not tell it about the BSP since it tells itself about
1112 * the BSP internally to work with UP kernels and on UP machines.
1115 set_interrupt_apic_ids(void)
1119 for (i = 0; i < MAXCPU; i++) {
1120 apic_id = cpu_apic_ids[i];
1123 if (cpu_info[apic_id].cpu_bsp)
1125 if (cpu_info[apic_id].cpu_disabled)
1128 /* Don't let hyperthreads service interrupts. */
1129 if (cpu_info[apic_id].cpu_hyperthread)
1137 #ifdef COUNT_XINVLTLB_HITS
1138 u_int xhits_gbl[MAXCPU];
1139 u_int xhits_pg[MAXCPU];
1140 u_int xhits_rng[MAXCPU];
1141 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1142 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1143 sizeof(xhits_gbl), "IU", "");
1144 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1145 sizeof(xhits_pg), "IU", "");
1146 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1147 sizeof(xhits_rng), "IU", "");
1152 u_int ipi_range_size;
1153 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1154 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1155 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1156 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1158 #endif /* COUNT_XINVLTLB_HITS */
1161 * Init and startup IPI.
1164 ipi_startup(int apic_id, int vector)
1168 * This attempts to follow the algorithm described in the
1169 * Intel Multiprocessor Specification v1.4 in section B.4.
1170 * For each IPI, we allow the local APIC ~20us to deliver the
1171 * IPI. If that times out, we panic.
1175 * first we do an INIT IPI: this INIT IPI might be run, resetting
1176 * and running the target CPU. OR this INIT IPI might be latched (P5
1177 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
1180 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1181 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
1182 lapic_ipi_wait(100);
1184 /* Explicitly deassert the INIT IPI. */
1185 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1186 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT,
1189 DELAY(10000); /* wait ~10mS */
1192 * next we do a STARTUP IPI: the previous INIT IPI might still be
1193 * latched, (P5 bug) this 1st STARTUP would then terminate
1194 * immediately, and the previously started INIT IPI would continue. OR
1195 * the previous INIT IPI has already run. and this STARTUP IPI will
1196 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1199 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1200 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1202 if (!lapic_ipi_wait(100))
1203 panic("Failed to deliver first STARTUP IPI to APIC %d",
1205 DELAY(200); /* wait ~200uS */
1208 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1209 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1210 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1211 * recognized after hardware RESET or INIT IPI.
1213 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1214 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1216 if (!lapic_ipi_wait(100))
1217 panic("Failed to deliver second STARTUP IPI to APIC %d",
1220 DELAY(200); /* wait ~200uS */
1224 * Send an IPI to specified CPU handling the bitmap logic.
1227 ipi_send_cpu(int cpu, u_int ipi)
1229 u_int bitmap, old_pending, new_pending;
1231 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1233 if (IPI_IS_BITMAPED(ipi)) {
1235 ipi = IPI_BITMAP_VECTOR;
1237 old_pending = cpu_ipi_pending[cpu];
1238 new_pending = old_pending | bitmap;
1239 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],
1240 old_pending, new_pending));
1244 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1248 ipi_bitmap_handler(struct trapframe frame)
1250 struct trapframe *oldframe;
1252 int cpu = PCPU_GET(cpuid);
1257 td->td_intr_nesting_level++;
1258 oldframe = td->td_intr_frame;
1259 td->td_intr_frame = &frame;
1260 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1261 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1263 (*ipi_preempt_counts[cpu])++;
1267 if (ipi_bitmap & (1 << IPI_AST)) {
1269 (*ipi_ast_counts[cpu])++;
1271 /* Nothing to do for AST */
1273 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1275 (*ipi_hardclock_counts[cpu])++;
1279 td->td_intr_frame = oldframe;
1280 td->td_intr_nesting_level--;
1285 * send an IPI to a set of cpus.
1288 ipi_selected(cpuset_t cpus, u_int ipi)
1293 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1294 * of help in order to understand what is the source.
1295 * Set the mask of receiving CPUs for this purpose.
1297 if (ipi == IPI_STOP_HARD)
1298 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &cpus);
1300 while ((cpu = CPU_FFS(&cpus)) != 0) {
1302 CPU_CLR(cpu, &cpus);
1303 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1304 ipi_send_cpu(cpu, ipi);
1309 * send an IPI to a specific CPU.
1312 ipi_cpu(int cpu, u_int ipi)
1316 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1317 * of help in order to understand what is the source.
1318 * Set the mask of receiving CPUs for this purpose.
1320 if (ipi == IPI_STOP_HARD)
1321 CPU_SET_ATOMIC(cpu, &ipi_stop_nmi_pending);
1323 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1324 ipi_send_cpu(cpu, ipi);
1328 * send an IPI to all CPUs EXCEPT myself
1331 ipi_all_but_self(u_int ipi)
1333 cpuset_t other_cpus;
1335 other_cpus = all_cpus;
1336 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1337 if (IPI_IS_BITMAPED(ipi)) {
1338 ipi_selected(other_cpus, ipi);
1343 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1344 * of help in order to understand what is the source.
1345 * Set the mask of receiving CPUs for this purpose.
1347 if (ipi == IPI_STOP_HARD)
1348 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &other_cpus);
1350 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1351 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1355 ipi_self_from_nmi(u_int vector)
1358 lapic_ipi_vectored(vector, APIC_IPI_DEST_SELF);
1360 /* Wait for IPI to finish. */
1361 if (!lapic_ipi_wait(50000)) {
1362 if (panicstr != NULL)
1365 panic("APIC: IPI is stuck");
1370 ipi_nmi_handler(void)
1375 * As long as there is not a simple way to know about a NMI's
1376 * source, if the bitmask for the current CPU is present in
1377 * the global pending bitword an IPI_STOP_HARD has been issued
1378 * and should be handled.
1380 cpuid = PCPU_GET(cpuid);
1381 if (!CPU_ISSET(cpuid, &ipi_stop_nmi_pending))
1384 CPU_CLR_ATOMIC(cpuid, &ipi_stop_nmi_pending);
1392 nmi_call_kdb_smp(u_int type, struct trapframe *frame)
1397 cpu = PCPU_GET(cpuid);
1398 if (atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) {
1399 nmi_call_kdb(cpu, type, frame);
1402 savectx(&stoppcbs[cpu]);
1403 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1404 while (!atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1))
1408 atomic_store_rel_int(&nmi_kdb_lock, 0);
1410 cpustop_handler_post(cpu);
1414 * Handle an IPI_STOP by saving our current context and spinning until we
1418 cpustop_handler(void)
1422 cpu = PCPU_GET(cpuid);
1424 savectx(&stoppcbs[cpu]);
1426 /* Indicate that we are stopped */
1427 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1429 /* Wait for restart */
1430 while (!CPU_ISSET(cpu, &started_cpus))
1433 cpustop_handler_post(cpu);
1437 cpustop_handler_post(u_int cpu)
1440 CPU_CLR_ATOMIC(cpu, &started_cpus);
1441 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1444 * We don't broadcast TLB invalidations to other CPUs when they are
1445 * stopped. Hence, we clear the TLB before resuming.
1449 #if defined(__amd64__) && defined(DDB)
1450 amd64_db_resume_dbreg();
1453 if (cpu == 0 && cpustop_restartfunc != NULL) {
1454 cpustop_restartfunc();
1455 cpustop_restartfunc = NULL;
1460 * Handle an IPI_SUSPEND by saving our current context and spinning until we
1464 cpususpend_handler(void)
1468 mtx_assert(&smp_ipi_mtx, MA_NOTOWNED);
1470 cpu = PCPU_GET(cpuid);
1471 if (savectx(&susppcbs[cpu]->sp_pcb)) {
1473 fpususpend(susppcbs[cpu]->sp_fpususpend);
1475 npxsuspend(susppcbs[cpu]->sp_fpususpend);
1478 * suspended_cpus is cleared shortly after each AP is restarted
1479 * by a Startup IPI, so that the BSP can proceed to restarting
1482 * resuming_cpus gets cleared when the AP completes
1483 * initialization after having been released by the BSP.
1484 * resuming_cpus is probably not the best name for the
1485 * variable, because it is actually a set of processors that
1486 * haven't resumed yet and haven't necessarily started resuming.
1488 * Note that suspended_cpus is meaningful only for ACPI suspend
1489 * as it's not really used for Xen suspend since the APs are
1490 * automatically restored to the running state and the correct
1491 * context. For the same reason resumectx is never called in
1494 CPU_SET_ATOMIC(cpu, &suspended_cpus);
1495 CPU_SET_ATOMIC(cpu, &resuming_cpus);
1498 * Invalidate the cache after setting the global status bits.
1499 * The last AP to set its bit may end up being an Owner of the
1500 * corresponding cache line in MOESI protocol. The AP may be
1501 * stopped before the cache line is written to the main memory.
1506 fpuresume(susppcbs[cpu]->sp_fpususpend);
1508 npxresume(susppcbs[cpu]->sp_fpususpend);
1512 PCPU_SET(switchtime, 0);
1513 PCPU_SET(switchticks, ticks);
1515 /* Indicate that we have restarted and restored the context. */
1516 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1519 /* Wait for resume directive */
1520 while (!CPU_ISSET(cpu, &toresume_cpus))
1523 /* Re-apply microcode updates. */
1527 /* Finish removing the identity mapping of low memory for this AP. */
1531 if (cpu_ops.cpu_resume)
1532 cpu_ops.cpu_resume();
1538 /* Resume MCA and local APIC */
1543 /* Indicate that we are resumed */
1544 CPU_CLR_ATOMIC(cpu, &resuming_cpus);
1545 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1546 CPU_CLR_ATOMIC(cpu, &toresume_cpus);
1551 invlcache_handler(void)
1553 uint32_t generation;
1556 (*ipi_invlcache_counts[PCPU_GET(cpuid)])++;
1557 #endif /* COUNT_IPIS */
1560 * Reading the generation here allows greater parallelism
1561 * since wbinvd is a serializing instruction. Without the
1562 * temporary, we'd wait for wbinvd to complete, then the read
1563 * would execute, then the dependent write, which must then
1564 * complete before return from interrupt.
1566 generation = smp_tlb_generation;
1568 PCPU_SET(smp_tlb_done, generation);
1572 * Handle an IPI_SWI by waking delayed SWI thread.
1575 ipi_swi_handler(struct trapframe frame)
1578 intr_event_handle(clk_intr_event, &frame);
1582 * This is called once the rest of the system is up and running and we're
1583 * ready to let the AP's out of the pen.
1586 release_aps(void *dummy __unused)
1591 atomic_store_rel_int(&aps_ready, 1);
1592 while (smp_started == 0)
1595 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1599 * Setup interrupt counters for IPI handlers.
1602 mp_ipi_intrcnt(void *dummy)
1608 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1609 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1610 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1611 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1612 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1613 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1614 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1615 intrcnt_add(buf, &ipi_invlcache_counts[i]);
1616 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1617 intrcnt_add(buf, &ipi_preempt_counts[i]);
1618 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1619 intrcnt_add(buf, &ipi_ast_counts[i]);
1620 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1621 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1622 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1623 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1626 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);
1630 * Flush the TLB on other CPU's
1633 /* Variables needed for SMP tlb shootdown. */
1634 vm_offset_t smp_tlb_addr1, smp_tlb_addr2;
1635 pmap_t smp_tlb_pmap;
1636 volatile uint32_t smp_tlb_generation;
1639 #define read_eflags() read_rflags()
1643 * Used by pmap to request invalidation of TLB or cache on local and
1644 * remote processors. Mask provides the set of remote CPUs which are
1645 * to be signalled with the IPI specified by vector. The curcpu_cb
1646 * callback is invoked on the calling CPU in a critical section while
1647 * waiting for remote CPUs to complete the operation.
1649 * The callback function is called unconditionally on the caller's
1650 * underlying processor, even when this processor is not set in the
1651 * mask. So, the callback function must be prepared to handle such
1652 * spurious invocations.
1654 * This function must be called with the thread pinned, and it unpins on
1658 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap,
1659 vm_offset_t addr1, vm_offset_t addr2, smp_invl_cb_t curcpu_cb)
1661 cpuset_t other_cpus;
1662 volatile uint32_t *p_cpudone;
1663 uint32_t generation;
1671 * It is not necessary to signal other CPUs while booting or
1672 * when in the debugger.
1674 if (kdb_active || panicstr != NULL || !smp_started)
1677 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
1680 * Check for other cpus. Return if none.
1682 if (CPU_ISFULLSET(&mask)) {
1686 CPU_CLR(PCPU_GET(cpuid), &mask);
1687 if (CPU_EMPTY(&mask))
1691 if (!(read_eflags() & PSL_I))
1692 panic("%s: interrupts disabled", __func__);
1693 mtx_lock_spin(&smp_ipi_mtx);
1694 smp_tlb_addr1 = addr1;
1695 smp_tlb_addr2 = addr2;
1696 smp_tlb_pmap = pmap;
1697 generation = ++smp_tlb_generation;
1698 if (CPU_ISFULLSET(&mask)) {
1699 ipi_all_but_self(vector);
1700 other_cpus = all_cpus;
1701 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1704 while ((cpu = CPU_FFS(&mask)) != 0) {
1706 CPU_CLR(cpu, &mask);
1707 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__,
1709 ipi_send_cpu(cpu, vector);
1712 curcpu_cb(pmap, addr1, addr2);
1713 while ((cpu = CPU_FFS(&other_cpus)) != 0) {
1715 CPU_CLR(cpu, &other_cpus);
1716 p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done;
1717 while (*p_cpudone != generation)
1722 * Unpin before unlocking smp_ipi_mtx. If the thread owes
1723 * preemption, this allows scheduler to select thread on any
1724 * CPU from its cpuset.
1727 mtx_unlock_spin(&smp_ipi_mtx);
1733 curcpu_cb(pmap, addr1, addr2);
1739 smp_masked_invltlb(cpuset_t mask, pmap_t pmap, smp_invl_cb_t curcpu_cb)
1742 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0, curcpu_cb);
1743 #ifdef COUNT_XINVLTLB_HITS
1749 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap,
1750 smp_invl_cb_t curcpu_cb)
1753 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0, curcpu_cb);
1754 #ifdef COUNT_XINVLTLB_HITS
1760 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2,
1761 pmap_t pmap, smp_invl_cb_t curcpu_cb)
1764 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap, addr1, addr2,
1766 #ifdef COUNT_XINVLTLB_HITS
1768 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1773 smp_cache_flush(smp_invl_cb_t curcpu_cb)
1776 smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL, 0, 0,
1781 * Handlers for TLB related IPIs
1784 invltlb_handler(void)
1786 uint32_t generation;
1788 #ifdef COUNT_XINVLTLB_HITS
1789 xhits_gbl[PCPU_GET(cpuid)]++;
1790 #endif /* COUNT_XINVLTLB_HITS */
1792 (*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
1793 #endif /* COUNT_IPIS */
1796 * Reading the generation here allows greater parallelism
1797 * since invalidating the TLB is a serializing operation.
1799 generation = smp_tlb_generation;
1800 if (smp_tlb_pmap == kernel_pmap)
1806 PCPU_SET(smp_tlb_done, generation);
1810 invlpg_handler(void)
1812 uint32_t generation;
1814 #ifdef COUNT_XINVLTLB_HITS
1815 xhits_pg[PCPU_GET(cpuid)]++;
1816 #endif /* COUNT_XINVLTLB_HITS */
1818 (*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
1819 #endif /* COUNT_IPIS */
1821 generation = smp_tlb_generation; /* Overlap with serialization */
1823 if (smp_tlb_pmap == kernel_pmap)
1825 invlpg(smp_tlb_addr1);
1826 PCPU_SET(smp_tlb_done, generation);
1830 invlrng_handler(void)
1832 vm_offset_t addr, addr2;
1833 uint32_t generation;
1835 #ifdef COUNT_XINVLTLB_HITS
1836 xhits_rng[PCPU_GET(cpuid)]++;
1837 #endif /* COUNT_XINVLTLB_HITS */
1839 (*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
1840 #endif /* COUNT_IPIS */
1842 addr = smp_tlb_addr1;
1843 addr2 = smp_tlb_addr2;
1844 generation = smp_tlb_generation; /* Overlap with serialization */
1846 if (smp_tlb_pmap == kernel_pmap)
1851 } while (addr < addr2);
1853 PCPU_SET(smp_tlb_done, generation);