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 static MALLOC_DEFINE(M_CPUS, "cpus", "CPU items");
80 /* lock region used by kernel profiling */
83 int mp_naps; /* # of Applications processors */
84 int boot_cpu_id = -1; /* designated BSP */
86 /* AP uses this during bootstrap. Do not staticize. */
90 /* Free these after use */
91 void *bootstacks[MAXCPU];
94 struct pcb stoppcbs[MAXCPU];
95 struct susppcb **susppcbs;
98 /* Interrupt counts. */
99 static u_long *ipi_preempt_counts[MAXCPU];
100 static u_long *ipi_ast_counts[MAXCPU];
101 u_long *ipi_invltlb_counts[MAXCPU];
102 u_long *ipi_invlrng_counts[MAXCPU];
103 u_long *ipi_invlpg_counts[MAXCPU];
104 u_long *ipi_invlcache_counts[MAXCPU];
105 u_long *ipi_rendezvous_counts[MAXCPU];
106 static u_long *ipi_hardclock_counts[MAXCPU];
109 /* Default cpu_ops implementation. */
110 struct cpu_ops cpu_ops;
113 * Local data and functions.
116 static volatile cpuset_t ipi_stop_nmi_pending;
118 volatile cpuset_t resuming_cpus;
119 volatile cpuset_t toresume_cpus;
121 /* used to hold the AP's until we are ready to release them */
122 struct mtx ap_boot_mtx;
124 /* Set to 1 once we're ready to let the APs out of the pen. */
125 volatile int aps_ready = 0;
128 * Store data from cpu_add() until later in the boot when we actually setup
131 struct cpu_info *cpu_info;
133 int cpu_apic_ids[MAXCPU];
134 _Static_assert(MAXCPU <= MAX_APIC_ID,
135 "MAXCPU cannot be larger that MAX_APIC_ID");
136 _Static_assert(xAPIC_MAX_APIC_ID <= MAX_APIC_ID,
137 "xAPIC_MAX_APIC_ID cannot be larger that MAX_APIC_ID");
139 /* Holds pending bitmap based IPIs per CPU */
140 volatile u_int cpu_ipi_pending[MAXCPU];
142 static void release_aps(void *dummy);
143 static void cpustop_handler_post(u_int cpu);
145 static int hyperthreading_allowed = 1;
146 SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_allowed, CTLFLAG_RDTUN,
147 &hyperthreading_allowed, 0, "Use Intel HTT logical CPUs");
149 static struct topo_node topo_root;
151 static int pkg_id_shift;
152 static int node_id_shift;
153 static int core_id_shift;
154 static int disabled_cpus;
159 } static caches[MAX_CACHE_LEVELS];
161 unsigned int boot_address;
163 #define MiB(v) (v ## ULL << 20)
166 mem_range_AP_init(void)
169 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
170 mem_range_softc.mr_op->initAP(&mem_range_softc);
174 * Round up to the next power of two, if necessary, and then
176 * Returns -1 if argument is zero.
182 return (fls(x << (1 - powerof2(x))) - 1);
186 * Add a cache level to the cache topology description.
189 add_deterministic_cache(int type, int level, int share_count)
195 printf("unexpected cache type %d\n", type);
198 if (type == 2) /* ignore instruction cache */
200 if (level == 0 || level > MAX_CACHE_LEVELS) {
201 printf("unexpected cache level %d\n", type);
205 if (caches[level - 1].present) {
206 printf("WARNING: multiple entries for L%u data cache\n", level);
207 printf("%u => %u\n", caches[level - 1].id_shift,
208 mask_width(share_count));
210 caches[level - 1].id_shift = mask_width(share_count);
211 caches[level - 1].present = 1;
213 if (caches[level - 1].id_shift > pkg_id_shift) {
214 printf("WARNING: L%u data cache covers more "
215 "APIC IDs than a package (%u > %u)\n", level,
216 caches[level - 1].id_shift, pkg_id_shift);
217 caches[level - 1].id_shift = pkg_id_shift;
219 if (caches[level - 1].id_shift < core_id_shift) {
220 printf("WARNING: L%u data cache covers fewer "
221 "APIC IDs than a core (%u < %u)\n", level,
222 caches[level - 1].id_shift, core_id_shift);
223 caches[level - 1].id_shift = core_id_shift;
230 * Determine topology of processing units and caches for AMD CPUs.
232 * - AMD CPUID Specification (Publication # 25481)
233 * - BKDG for AMD NPT Family 0Fh Processors (Publication # 32559)
234 * - BKDG For AMD Family 10h Processors (Publication # 31116)
235 * - BKDG For AMD Family 15h Models 00h-0Fh Processors (Publication # 42301)
236 * - BKDG For AMD Family 16h Models 00h-0Fh Processors (Publication # 48751)
244 int nodes_per_socket;
249 /* No multi-core capability. */
250 if ((amd_feature2 & AMDID2_CMP) == 0)
253 /* For families 10h and newer. */
254 pkg_id_shift = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
255 AMDID_COREID_SIZE_SHIFT;
257 /* For 0Fh family. */
258 if (pkg_id_shift == 0)
260 mask_width((cpu_procinfo2 & AMDID_CMP_CORES) + 1);
263 * Families prior to 16h define the following value as
264 * cores per compute unit and we don't really care about the AMD
265 * compute units at the moment. Perhaps we should treat them as
266 * cores and cores within the compute units as hardware threads,
267 * but that's up for debate.
268 * Later families define the value as threads per compute unit,
269 * so we are following AMD's nomenclature here.
271 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0 &&
272 CPUID_TO_FAMILY(cpu_id) >= 0x16) {
273 cpuid_count(0x8000001e, 0, p);
274 share_count = ((p[1] >> 8) & 0xff) + 1;
275 core_id_shift = mask_width(share_count);
278 * For Zen (17h), gather Nodes per Processor. Each node is a
279 * Zeppelin die; TR and EPYC CPUs will have multiple dies per
280 * package. Communication latency between dies is higher than
283 nodes_per_socket = ((p[2] >> 8) & 0x7) + 1;
284 node_id_shift = pkg_id_shift - mask_width(nodes_per_socket);
287 if ((amd_feature2 & AMDID2_TOPOLOGY) != 0) {
289 cpuid_count(0x8000001d, i, p);
291 level = (p[0] >> 5) & 0x7;
292 share_count = 1 + ((p[0] >> 14) & 0xfff);
294 if (!add_deterministic_cache(type, level, share_count))
298 if (cpu_exthigh >= 0x80000005) {
299 cpuid_count(0x80000005, 0, p);
300 if (((p[2] >> 24) & 0xff) != 0) {
301 caches[0].id_shift = 0;
302 caches[0].present = 1;
305 if (cpu_exthigh >= 0x80000006) {
306 cpuid_count(0x80000006, 0, p);
307 if (((p[2] >> 16) & 0xffff) != 0) {
308 caches[1].id_shift = 0;
309 caches[1].present = 1;
311 if (((p[3] >> 18) & 0x3fff) != 0) {
312 nodes_per_socket = 1;
313 if ((amd_feature2 & AMDID2_NODE_ID) != 0) {
315 * Handle multi-node processors that
316 * have multiple chips, each with its
317 * own L3 cache, on the same die.
319 v = rdmsr(0xc001100c);
320 nodes_per_socket = 1 + ((v >> 3) & 0x7);
323 pkg_id_shift - mask_width(nodes_per_socket);
324 caches[2].present = 1;
331 * Determine topology of processing units for Intel CPUs
332 * using CPUID Leaf 1 and Leaf 4, if supported.
334 * - Intel 64 Architecture Processor Topology Enumeration
335 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
336 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
337 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
340 topo_probe_intel_0x4(void)
346 /* Both zero and one here mean one logical processor per package. */
347 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
348 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
349 if (max_logical <= 1)
352 if (cpu_high >= 0x4) {
353 cpuid_count(0x04, 0, p);
354 max_cores = ((p[0] >> 26) & 0x3f) + 1;
358 core_id_shift = mask_width(max_logical/max_cores);
359 KASSERT(core_id_shift >= 0,
360 ("intel topo: max_cores > max_logical\n"));
361 pkg_id_shift = core_id_shift + mask_width(max_cores);
365 * Determine topology of processing units for Intel CPUs
366 * using CPUID Leaf 11, if supported.
368 * - Intel 64 Architecture Processor Topology Enumeration
369 * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual,
370 * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS
371 * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS
374 topo_probe_intel_0xb(void)
381 /* Fall back if CPU leaf 11 doesn't really exist. */
382 cpuid_count(0x0b, 0, p);
384 topo_probe_intel_0x4();
388 /* We only support three levels for now. */
390 cpuid_count(0x0b, i, p);
393 type = (p[2] >> 8) & 0xff;
398 /* TODO: check for duplicate (re-)assignment */
399 if (type == CPUID_TYPE_SMT)
400 core_id_shift = bits;
401 else if (type == CPUID_TYPE_CORE)
404 printf("unknown CPU level type %d\n", type);
407 if (pkg_id_shift < core_id_shift) {
408 printf("WARNING: core covers more APIC IDs than a package\n");
409 core_id_shift = pkg_id_shift;
414 * Determine topology of caches for Intel CPUs.
416 * - Intel 64 Architecture Processor Topology Enumeration
417 * - Intel 64 and IA-32 Architectures Software Developer’s Manual
418 * Volume 2A: Instruction Set Reference, A-M,
422 topo_probe_intel_caches(void)
430 if (cpu_high < 0x4) {
432 * Available cache level and sizes can be determined
433 * via CPUID leaf 2, but that requires a huge table of hardcoded
434 * values, so for now just assume L1 and L2 caches potentially
435 * shared only by HTT processing units, if HTT is present.
437 caches[0].id_shift = pkg_id_shift;
438 caches[0].present = 1;
439 caches[1].id_shift = pkg_id_shift;
440 caches[1].present = 1;
445 cpuid_count(0x4, i, p);
447 level = (p[0] >> 5) & 0x7;
448 share_count = 1 + ((p[0] >> 14) & 0xfff);
450 if (!add_deterministic_cache(type, level, share_count))
456 * Determine topology of processing units and caches for Intel CPUs.
458 * - Intel 64 Architecture Processor Topology Enumeration
461 topo_probe_intel(void)
465 * Note that 0x1 <= cpu_high < 4 case should be
466 * compatible with topo_probe_intel_0x4() logic when
467 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
468 * or it should trigger the fallback otherwise.
471 topo_probe_intel_0xb();
472 else if (cpu_high >= 0x1)
473 topo_probe_intel_0x4();
475 topo_probe_intel_caches();
479 * Topology information is queried only on BSP, on which this
480 * code runs and for which it can query CPUID information.
481 * Then topology is extrapolated on all packages using an
482 * assumption that APIC ID to hardware component ID mapping is
484 * That doesn't necesserily imply that the topology is uniform.
489 static int cpu_topo_probed = 0;
490 struct x86_topo_layer {
494 } topo_layers[MAX_CACHE_LEVELS + 4];
495 struct topo_node *parent;
496 struct topo_node *node;
505 CPU_ZERO(&logical_cpus_mask);
509 else if (cpu_vendor_id == CPU_VENDOR_AMD)
511 else if (cpu_vendor_id == CPU_VENDOR_INTEL)
514 KASSERT(pkg_id_shift >= core_id_shift,
515 ("bug in APIC topology discovery"));
518 bzero(topo_layers, sizeof(topo_layers));
520 topo_layers[nlayers].type = TOPO_TYPE_PKG;
521 topo_layers[nlayers].id_shift = pkg_id_shift;
523 printf("Package ID shift: %u\n", topo_layers[nlayers].id_shift);
526 if (pkg_id_shift > node_id_shift && node_id_shift != 0) {
527 topo_layers[nlayers].type = TOPO_TYPE_GROUP;
528 topo_layers[nlayers].id_shift = node_id_shift;
530 printf("Node ID shift: %u\n",
531 topo_layers[nlayers].id_shift);
536 * Consider all caches to be within a package/chip
537 * and "in front" of all sub-components like
538 * cores and hardware threads.
540 for (i = MAX_CACHE_LEVELS - 1; i >= 0; --i) {
541 if (caches[i].present) {
542 if (node_id_shift != 0)
543 KASSERT(caches[i].id_shift <= node_id_shift,
544 ("bug in APIC topology discovery"));
545 KASSERT(caches[i].id_shift <= pkg_id_shift,
546 ("bug in APIC topology discovery"));
547 KASSERT(caches[i].id_shift >= core_id_shift,
548 ("bug in APIC topology discovery"));
550 topo_layers[nlayers].type = TOPO_TYPE_CACHE;
551 topo_layers[nlayers].subtype = i + 1;
552 topo_layers[nlayers].id_shift = caches[i].id_shift;
554 printf("L%u cache ID shift: %u\n",
555 topo_layers[nlayers].subtype,
556 topo_layers[nlayers].id_shift);
561 if (pkg_id_shift > core_id_shift) {
562 topo_layers[nlayers].type = TOPO_TYPE_CORE;
563 topo_layers[nlayers].id_shift = core_id_shift;
565 printf("Core ID shift: %u\n",
566 topo_layers[nlayers].id_shift);
570 topo_layers[nlayers].type = TOPO_TYPE_PU;
571 topo_layers[nlayers].id_shift = 0;
574 topo_init_root(&topo_root);
575 for (i = 0; i <= max_apic_id; ++i) {
576 if (!cpu_info[i].cpu_present)
580 for (layer = 0; layer < nlayers; ++layer) {
581 node_id = i >> topo_layers[layer].id_shift;
582 parent = topo_add_node_by_hwid(parent, node_id,
583 topo_layers[layer].type,
584 topo_layers[layer].subtype);
589 for (layer = 0; layer < nlayers; ++layer) {
590 node_id = boot_cpu_id >> topo_layers[layer].id_shift;
591 node = topo_find_node_by_hwid(parent, node_id,
592 topo_layers[layer].type,
593 topo_layers[layer].subtype);
594 topo_promote_child(node);
602 * Assign logical CPU IDs to local APICs.
607 struct topo_node *node;
610 smt_mask = (1u << core_id_shift) - 1;
613 * Assign CPU IDs to local APIC IDs and disable any CPUs
614 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
617 TOPO_FOREACH(node, &topo_root) {
618 if (node->type != TOPO_TYPE_PU)
621 if ((node->hwid & smt_mask) != (boot_cpu_id & smt_mask))
622 cpu_info[node->hwid].cpu_hyperthread = 1;
624 if (resource_disabled("lapic", node->hwid)) {
625 if (node->hwid != boot_cpu_id)
626 cpu_info[node->hwid].cpu_disabled = 1;
628 printf("Cannot disable BSP, APIC ID = %d\n",
632 if (!hyperthreading_allowed &&
633 cpu_info[node->hwid].cpu_hyperthread)
634 cpu_info[node->hwid].cpu_disabled = 1;
636 if (mp_ncpus >= MAXCPU)
637 cpu_info[node->hwid].cpu_disabled = 1;
639 if (cpu_info[node->hwid].cpu_disabled) {
644 cpu_apic_ids[mp_ncpus] = node->hwid;
645 apic_cpuids[node->hwid] = mp_ncpus;
646 topo_set_pu_id(node, mp_ncpus);
650 KASSERT(mp_maxid >= mp_ncpus - 1,
651 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
656 * Print various information about the SMP system hardware and setup.
659 cpu_mp_announce(void)
661 struct topo_node *node;
662 const char *hyperthread;
663 struct topo_analysis topology;
665 printf("FreeBSD/SMP: ");
666 if (topo_analyze(&topo_root, 1, &topology)) {
667 printf("%d package(s)", topology.entities[TOPO_LEVEL_PKG]);
668 if (topology.entities[TOPO_LEVEL_GROUP] > 1)
669 printf(" x %d groups",
670 topology.entities[TOPO_LEVEL_GROUP]);
671 if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1)
672 printf(" x %d cache groups",
673 topology.entities[TOPO_LEVEL_CACHEGROUP]);
674 if (topology.entities[TOPO_LEVEL_CORE] > 0)
675 printf(" x %d core(s)",
676 topology.entities[TOPO_LEVEL_CORE]);
677 if (topology.entities[TOPO_LEVEL_THREAD] > 1)
678 printf(" x %d hardware threads",
679 topology.entities[TOPO_LEVEL_THREAD]);
681 printf("Non-uniform topology");
686 printf("FreeBSD/SMP Online: ");
687 if (topo_analyze(&topo_root, 0, &topology)) {
688 printf("%d package(s)",
689 topology.entities[TOPO_LEVEL_PKG]);
690 if (topology.entities[TOPO_LEVEL_GROUP] > 1)
691 printf(" x %d groups",
692 topology.entities[TOPO_LEVEL_GROUP]);
693 if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1)
694 printf(" x %d cache groups",
695 topology.entities[TOPO_LEVEL_CACHEGROUP]);
696 if (topology.entities[TOPO_LEVEL_CORE] > 0)
697 printf(" x %d core(s)",
698 topology.entities[TOPO_LEVEL_CORE]);
699 if (topology.entities[TOPO_LEVEL_THREAD] > 1)
700 printf(" x %d hardware threads",
701 topology.entities[TOPO_LEVEL_THREAD]);
703 printf("Non-uniform topology");
711 TOPO_FOREACH(node, &topo_root) {
712 switch (node->type) {
714 printf("Package HW ID = %u\n", node->hwid);
717 printf("\tCore HW ID = %u\n", node->hwid);
720 if (cpu_info[node->hwid].cpu_hyperthread)
725 if (node->subtype == 0)
726 printf("\t\tCPU (AP%s): APIC ID: %u"
727 "(disabled)\n", hyperthread, node->hwid);
728 else if (node->id == 0)
729 printf("\t\tCPU0 (BSP): APIC ID: %u\n",
732 printf("\t\tCPU%u (AP%s): APIC ID: %u\n",
733 node->id, hyperthread, node->hwid);
743 * Add a scheduling group, a group of logical processors sharing
744 * a particular cache (and, thus having an affinity), to the scheduling
746 * This function recursively works on lower level caches.
749 x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root)
751 struct topo_node *node;
756 KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE ||
757 root->type == TOPO_TYPE_GROUP,
758 ("x86topo_add_sched_group: bad type: %u", root->type));
759 CPU_COPY(&root->cpuset, &cg_root->cg_mask);
760 cg_root->cg_count = root->cpu_count;
761 if (root->type == TOPO_TYPE_SYSTEM)
762 cg_root->cg_level = CG_SHARE_NONE;
764 cg_root->cg_level = root->subtype;
767 * Check how many core nodes we have under the given root node.
768 * If we have multiple logical processors, but not multiple
769 * cores, then those processors must be hardware threads.
773 while (node != NULL) {
774 if (node->type != TOPO_TYPE_CORE) {
775 node = topo_next_node(root, node);
780 node = topo_next_nonchild_node(root, node);
783 if (cg_root->cg_level != CG_SHARE_NONE &&
784 root->cpu_count > 1 && ncores < 2)
785 cg_root->cg_flags = CG_FLAG_SMT;
788 * Find out how many cache nodes we have under the given root node.
789 * We ignore cache nodes that cover all the same processors as the
790 * root node. Also, we do not descend below found cache nodes.
791 * That is, we count top-level "non-redundant" caches under the root
796 while (node != NULL) {
797 if ((node->type != TOPO_TYPE_GROUP &&
798 node->type != TOPO_TYPE_CACHE) ||
799 (root->type != TOPO_TYPE_SYSTEM &&
800 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
801 node = topo_next_node(root, node);
805 node = topo_next_nonchild_node(root, node);
808 cg_root->cg_child = smp_topo_alloc(nchildren);
809 cg_root->cg_children = nchildren;
812 * Now find again the same cache nodes as above and recursively
813 * build scheduling topologies for them.
817 while (node != NULL) {
818 if ((node->type != TOPO_TYPE_GROUP &&
819 node->type != TOPO_TYPE_CACHE) ||
820 (root->type != TOPO_TYPE_SYSTEM &&
821 CPU_CMP(&node->cpuset, &root->cpuset) == 0)) {
822 node = topo_next_node(root, node);
825 cg_root->cg_child[i].cg_parent = cg_root;
826 x86topo_add_sched_group(node, &cg_root->cg_child[i]);
828 node = topo_next_nonchild_node(root, node);
833 * Build the MI scheduling topology from the discovered hardware topology.
838 struct cpu_group *cg_root;
841 return (smp_topo_none());
843 cg_root = smp_topo_alloc(1);
844 x86topo_add_sched_group(&topo_root, cg_root);
849 cpu_alloc(void *dummy __unused)
852 * Dynamically allocate the arrays that depend on the
855 cpu_info = malloc(sizeof(*cpu_info) * (max_apic_id + 1), M_CPUS,
857 apic_cpuids = malloc(sizeof(*apic_cpuids) * (max_apic_id + 1), M_CPUS,
860 SYSINIT(cpu_alloc, SI_SUB_CPU, SI_ORDER_FIRST, cpu_alloc, NULL);
863 * Add a logical CPU to the topology.
866 cpu_add(u_int apic_id, char boot_cpu)
869 if (apic_id > max_apic_id) {
870 panic("SMP: APIC ID %d too high", apic_id);
873 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %u added twice",
875 cpu_info[apic_id].cpu_present = 1;
877 KASSERT(boot_cpu_id == -1,
878 ("CPU %u claims to be BSP, but CPU %u already is", apic_id,
880 boot_cpu_id = apic_id;
881 cpu_info[apic_id].cpu_bsp = 1;
884 printf("SMP: Added CPU %u (%s)\n", apic_id, boot_cpu ? "BSP" :
889 cpu_mp_setmaxid(void)
893 * mp_ncpus and mp_maxid should be already set by calls to cpu_add().
894 * If there were no calls to cpu_add() assume this is a UP system.
905 * Always record BSP in CPU map so that the mbuf init code works
908 CPU_SETOF(0, &all_cpus);
909 return (mp_ncpus > 1);
912 /* Allocate memory for the AP trampoline. */
914 alloc_ap_trampoline(vm_paddr_t *physmap, unsigned int *physmap_idx)
920 for (i = *physmap_idx; i <= *physmap_idx; i -= 2) {
922 * Find a memory region big enough and below the 1MB boundary
923 * for the trampoline code.
924 * NB: needs to be page aligned.
926 if (physmap[i] >= MiB(1) ||
927 (trunc_page(physmap[i + 1]) - round_page(physmap[i])) <
928 round_page(bootMP_size))
933 * Try to steal from the end of the region to mimic previous
934 * behaviour, else fallback to steal from the start.
936 if (physmap[i + 1] < MiB(1)) {
937 boot_address = trunc_page(physmap[i + 1]);
938 if ((physmap[i + 1] - boot_address) < bootMP_size)
939 boot_address -= round_page(bootMP_size);
940 physmap[i + 1] = boot_address;
942 boot_address = round_page(physmap[i]);
943 physmap[i] = boot_address + round_page(bootMP_size);
945 if (physmap[i] == physmap[i + 1] && *physmap_idx != 0) {
946 memmove(&physmap[i], &physmap[i + 2],
947 sizeof(*physmap) * (*physmap_idx - i + 2));
954 boot_address = basemem * 1024 - bootMP_size;
957 "Cannot find enough space for the boot trampoline, placing it at %#x",
963 * AP CPU's call this to initialize themselves.
966 init_secondary_tail(void)
971 * On real hardware, switch to x2apic mode if possible. Do it
972 * after aps_ready was signalled, to avoid manipulating the
973 * mode while BSP might still want to send some IPI to us
974 * (second startup IPI is ignored on modern hardware etc).
978 /* Initialize the PAT MSR. */
981 /* set up CPU registers and state */
987 /* set up FPU state on the AP */
994 if (cpu_ops.cpu_init)
997 /* A quick check from sanity claus */
998 cpuid = PCPU_GET(cpuid);
999 if (PCPU_GET(apic_id) != lapic_id()) {
1000 printf("SMP: cpuid = %d\n", cpuid);
1001 printf("SMP: actual apic_id = %d\n", lapic_id());
1002 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
1003 panic("cpuid mismatch! boom!!");
1006 /* Initialize curthread. */
1007 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
1008 PCPU_SET(curthread, PCPU_GET(idlethread));
1010 mtx_lock_spin(&ap_boot_mtx);
1014 /* Init local apic for irq's */
1017 /* Set memory range attributes for this CPU to match the BSP */
1018 mem_range_AP_init();
1022 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
1024 printf("SMP: AP CPU #%d Launched!\n", cpuid);
1026 printf("%s%d%s", smp_cpus == 2 ? "Launching APs: " : "",
1027 cpuid, smp_cpus == mp_ncpus ? "\n" : " ");
1029 /* Determine if we are a logical CPU. */
1030 if (cpu_info[PCPU_GET(apic_id)].cpu_hyperthread)
1031 CPU_SET(cpuid, &logical_cpus_mask);
1036 if (smp_cpus == mp_ncpus) {
1037 /* enable IPI's, tlb shootdown, freezes etc */
1038 atomic_store_rel_int(&smp_started, 1);
1043 * Enable global pages TLB extension
1044 * This also implicitly flushes the TLB
1046 load_cr4(rcr4() | CR4_PGE);
1047 if (pmap_pcid_enabled)
1048 load_cr4(rcr4() | CR4_PCIDE);
1054 mtx_unlock_spin(&ap_boot_mtx);
1056 /* Wait until all the AP's are up. */
1057 while (atomic_load_acq_int(&smp_started) == 0)
1060 #ifndef EARLY_AP_STARTUP
1061 /* Start per-CPU event timers. */
1062 cpu_initclocks_ap();
1067 panic("scheduler returned us to %s", __func__);
1071 /*******************************************************************
1072 * local functions and data
1076 * We tell the I/O APIC code about all the CPUs we want to receive
1077 * interrupts. If we don't want certain CPUs to receive IRQs we
1078 * can simply not tell the I/O APIC code about them in this function.
1079 * We also do not tell it about the BSP since it tells itself about
1080 * the BSP internally to work with UP kernels and on UP machines.
1083 set_interrupt_apic_ids(void)
1087 for (i = 0; i < MAXCPU; i++) {
1088 apic_id = cpu_apic_ids[i];
1091 if (cpu_info[apic_id].cpu_bsp)
1093 if (cpu_info[apic_id].cpu_disabled)
1096 /* Don't let hyperthreads service interrupts. */
1097 if (cpu_info[apic_id].cpu_hyperthread)
1105 #ifdef COUNT_XINVLTLB_HITS
1106 u_int xhits_gbl[MAXCPU];
1107 u_int xhits_pg[MAXCPU];
1108 u_int xhits_rng[MAXCPU];
1109 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1110 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1111 sizeof(xhits_gbl), "IU", "");
1112 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1113 sizeof(xhits_pg), "IU", "");
1114 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1115 sizeof(xhits_rng), "IU", "");
1120 u_int ipi_range_size;
1121 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1122 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1123 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1124 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1126 #endif /* COUNT_XINVLTLB_HITS */
1129 * Init and startup IPI.
1132 ipi_startup(int apic_id, int vector)
1136 * This attempts to follow the algorithm described in the
1137 * Intel Multiprocessor Specification v1.4 in section B.4.
1138 * For each IPI, we allow the local APIC ~20us to deliver the
1139 * IPI. If that times out, we panic.
1143 * first we do an INIT IPI: this INIT IPI might be run, resetting
1144 * and running the target CPU. OR this INIT IPI might be latched (P5
1145 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
1148 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1149 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
1150 lapic_ipi_wait(100);
1152 /* Explicitly deassert the INIT IPI. */
1153 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL |
1154 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT,
1157 DELAY(10000); /* wait ~10mS */
1160 * next we do a STARTUP IPI: the previous INIT IPI might still be
1161 * latched, (P5 bug) this 1st STARTUP would then terminate
1162 * immediately, and the previously started INIT IPI would continue. OR
1163 * the previous INIT IPI has already run. and this STARTUP IPI will
1164 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1167 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1168 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1170 if (!lapic_ipi_wait(100))
1171 panic("Failed to deliver first STARTUP IPI to APIC %d",
1173 DELAY(200); /* wait ~200uS */
1176 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1177 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1178 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1179 * recognized after hardware RESET or INIT IPI.
1181 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1182 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1184 if (!lapic_ipi_wait(100))
1185 panic("Failed to deliver second STARTUP IPI to APIC %d",
1188 DELAY(200); /* wait ~200uS */
1192 * Send an IPI to specified CPU handling the bitmap logic.
1195 ipi_send_cpu(int cpu, u_int ipi)
1197 u_int bitmap, old_pending, new_pending;
1199 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1201 if (IPI_IS_BITMAPED(ipi)) {
1203 ipi = IPI_BITMAP_VECTOR;
1205 old_pending = cpu_ipi_pending[cpu];
1206 new_pending = old_pending | bitmap;
1207 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],
1208 old_pending, new_pending));
1212 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1216 ipi_bitmap_handler(struct trapframe frame)
1218 struct trapframe *oldframe;
1220 int cpu = PCPU_GET(cpuid);
1225 td->td_intr_nesting_level++;
1226 oldframe = td->td_intr_frame;
1227 td->td_intr_frame = &frame;
1228 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1229 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1231 (*ipi_preempt_counts[cpu])++;
1235 if (ipi_bitmap & (1 << IPI_AST)) {
1237 (*ipi_ast_counts[cpu])++;
1239 /* Nothing to do for AST */
1241 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1243 (*ipi_hardclock_counts[cpu])++;
1247 td->td_intr_frame = oldframe;
1248 td->td_intr_nesting_level--;
1253 * send an IPI to a set of cpus.
1256 ipi_selected(cpuset_t cpus, u_int ipi)
1261 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1262 * of help in order to understand what is the source.
1263 * Set the mask of receiving CPUs for this purpose.
1265 if (ipi == IPI_STOP_HARD)
1266 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &cpus);
1268 while ((cpu = CPU_FFS(&cpus)) != 0) {
1270 CPU_CLR(cpu, &cpus);
1271 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1272 ipi_send_cpu(cpu, ipi);
1277 * send an IPI to a specific CPU.
1280 ipi_cpu(int cpu, u_int ipi)
1284 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1285 * of help in order to understand what is the source.
1286 * Set the mask of receiving CPUs for this purpose.
1288 if (ipi == IPI_STOP_HARD)
1289 CPU_SET_ATOMIC(cpu, &ipi_stop_nmi_pending);
1291 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1292 ipi_send_cpu(cpu, ipi);
1296 * send an IPI to all CPUs EXCEPT myself
1299 ipi_all_but_self(u_int ipi)
1301 cpuset_t other_cpus;
1303 other_cpus = all_cpus;
1304 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1305 if (IPI_IS_BITMAPED(ipi)) {
1306 ipi_selected(other_cpus, ipi);
1311 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1312 * of help in order to understand what is the source.
1313 * Set the mask of receiving CPUs for this purpose.
1315 if (ipi == IPI_STOP_HARD)
1316 CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &other_cpus);
1318 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1319 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1323 ipi_nmi_handler(void)
1328 * As long as there is not a simple way to know about a NMI's
1329 * source, if the bitmask for the current CPU is present in
1330 * the global pending bitword an IPI_STOP_HARD has been issued
1331 * and should be handled.
1333 cpuid = PCPU_GET(cpuid);
1334 if (!CPU_ISSET(cpuid, &ipi_stop_nmi_pending))
1337 CPU_CLR_ATOMIC(cpuid, &ipi_stop_nmi_pending);
1346 nmi_call_kdb_smp(u_int type, struct trapframe *frame)
1351 cpu = PCPU_GET(cpuid);
1352 if (atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) {
1353 nmi_call_kdb(cpu, type, frame);
1356 savectx(&stoppcbs[cpu]);
1357 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1358 while (!atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1))
1362 atomic_store_rel_int(&nmi_kdb_lock, 0);
1364 cpustop_handler_post(cpu);
1369 * Handle an IPI_STOP by saving our current context and spinning until we
1373 cpustop_handler(void)
1377 cpu = PCPU_GET(cpuid);
1379 savectx(&stoppcbs[cpu]);
1381 /* Indicate that we are stopped */
1382 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1384 /* Wait for restart */
1385 while (!CPU_ISSET(cpu, &started_cpus))
1388 cpustop_handler_post(cpu);
1392 cpustop_handler_post(u_int cpu)
1395 CPU_CLR_ATOMIC(cpu, &started_cpus);
1396 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1399 * We don't broadcast TLB invalidations to other CPUs when they are
1400 * stopped. Hence, we clear the TLB before resuming.
1404 #if defined(__amd64__) && defined(DDB)
1405 amd64_db_resume_dbreg();
1408 if (cpu == 0 && cpustop_restartfunc != NULL) {
1409 cpustop_restartfunc();
1410 cpustop_restartfunc = NULL;
1415 * Handle an IPI_SUSPEND by saving our current context and spinning until we
1419 cpususpend_handler(void)
1423 mtx_assert(&smp_ipi_mtx, MA_NOTOWNED);
1425 cpu = PCPU_GET(cpuid);
1426 if (savectx(&susppcbs[cpu]->sp_pcb)) {
1428 fpususpend(susppcbs[cpu]->sp_fpususpend);
1430 npxsuspend(susppcbs[cpu]->sp_fpususpend);
1433 * suspended_cpus is cleared shortly after each AP is restarted
1434 * by a Startup IPI, so that the BSP can proceed to restarting
1437 * resuming_cpus gets cleared when the AP completes
1438 * initialization after having been released by the BSP.
1439 * resuming_cpus is probably not the best name for the
1440 * variable, because it is actually a set of processors that
1441 * haven't resumed yet and haven't necessarily started resuming.
1443 * Note that suspended_cpus is meaningful only for ACPI suspend
1444 * as it's not really used for Xen suspend since the APs are
1445 * automatically restored to the running state and the correct
1446 * context. For the same reason resumectx is never called in
1449 CPU_SET_ATOMIC(cpu, &suspended_cpus);
1450 CPU_SET_ATOMIC(cpu, &resuming_cpus);
1453 * Invalidate the cache after setting the global status bits.
1454 * The last AP to set its bit may end up being an Owner of the
1455 * corresponding cache line in MOESI protocol. The AP may be
1456 * stopped before the cache line is written to the main memory.
1461 fpuresume(susppcbs[cpu]->sp_fpususpend);
1463 npxresume(susppcbs[cpu]->sp_fpususpend);
1467 PCPU_SET(switchtime, 0);
1468 PCPU_SET(switchticks, ticks);
1470 /* Indicate that we have restarted and restored the context. */
1471 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1474 /* Wait for resume directive */
1475 while (!CPU_ISSET(cpu, &toresume_cpus))
1479 /* Finish removing the identity mapping of low memory for this AP. */
1483 if (cpu_ops.cpu_resume)
1484 cpu_ops.cpu_resume();
1490 /* Resume MCA and local APIC */
1495 /* Indicate that we are resumed */
1496 CPU_CLR_ATOMIC(cpu, &resuming_cpus);
1497 CPU_CLR_ATOMIC(cpu, &suspended_cpus);
1498 CPU_CLR_ATOMIC(cpu, &toresume_cpus);
1503 invlcache_handler(void)
1505 uint32_t generation;
1508 (*ipi_invlcache_counts[PCPU_GET(cpuid)])++;
1509 #endif /* COUNT_IPIS */
1512 * Reading the generation here allows greater parallelism
1513 * since wbinvd is a serializing instruction. Without the
1514 * temporary, we'd wait for wbinvd to complete, then the read
1515 * would execute, then the dependent write, which must then
1516 * complete before return from interrupt.
1518 generation = smp_tlb_generation;
1520 PCPU_SET(smp_tlb_done, generation);
1524 * This is called once the rest of the system is up and running and we're
1525 * ready to let the AP's out of the pen.
1528 release_aps(void *dummy __unused)
1533 atomic_store_rel_int(&aps_ready, 1);
1534 while (smp_started == 0)
1537 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1541 * Setup interrupt counters for IPI handlers.
1544 mp_ipi_intrcnt(void *dummy)
1550 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1551 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1552 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1553 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1554 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1555 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1556 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1557 intrcnt_add(buf, &ipi_invlcache_counts[i]);
1558 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1559 intrcnt_add(buf, &ipi_preempt_counts[i]);
1560 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1561 intrcnt_add(buf, &ipi_ast_counts[i]);
1562 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1563 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1564 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1565 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1568 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);
1572 * Flush the TLB on other CPU's
1575 /* Variables needed for SMP tlb shootdown. */
1576 vm_offset_t smp_tlb_addr1, smp_tlb_addr2;
1577 pmap_t smp_tlb_pmap;
1578 volatile uint32_t smp_tlb_generation;
1581 #define read_eflags() read_rflags()
1585 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap,
1586 vm_offset_t addr1, vm_offset_t addr2)
1588 cpuset_t other_cpus;
1589 volatile uint32_t *p_cpudone;
1590 uint32_t generation;
1593 /* It is not necessary to signal other CPUs while in the debugger. */
1594 if (kdb_active || panicstr != NULL)
1598 * Check for other cpus. Return if none.
1600 if (CPU_ISFULLSET(&mask)) {
1604 CPU_CLR(PCPU_GET(cpuid), &mask);
1605 if (CPU_EMPTY(&mask))
1609 if (!(read_eflags() & PSL_I))
1610 panic("%s: interrupts disabled", __func__);
1611 mtx_lock_spin(&smp_ipi_mtx);
1612 smp_tlb_addr1 = addr1;
1613 smp_tlb_addr2 = addr2;
1614 smp_tlb_pmap = pmap;
1615 generation = ++smp_tlb_generation;
1616 if (CPU_ISFULLSET(&mask)) {
1617 ipi_all_but_self(vector);
1618 other_cpus = all_cpus;
1619 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1622 while ((cpu = CPU_FFS(&mask)) != 0) {
1624 CPU_CLR(cpu, &mask);
1625 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__,
1627 ipi_send_cpu(cpu, vector);
1630 while ((cpu = CPU_FFS(&other_cpus)) != 0) {
1632 CPU_CLR(cpu, &other_cpus);
1633 p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done;
1634 while (*p_cpudone != generation)
1637 mtx_unlock_spin(&smp_ipi_mtx);
1641 smp_masked_invltlb(cpuset_t mask, pmap_t pmap)
1645 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0);
1646 #ifdef COUNT_XINVLTLB_HITS
1653 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap)
1657 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0);
1658 #ifdef COUNT_XINVLTLB_HITS
1665 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2,
1670 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap,
1672 #ifdef COUNT_XINVLTLB_HITS
1674 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1680 smp_cache_flush(void)
1684 smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL,
1690 * Handlers for TLB related IPIs
1693 invltlb_handler(void)
1695 uint32_t generation;
1697 #ifdef COUNT_XINVLTLB_HITS
1698 xhits_gbl[PCPU_GET(cpuid)]++;
1699 #endif /* COUNT_XINVLTLB_HITS */
1701 (*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
1702 #endif /* COUNT_IPIS */
1705 * Reading the generation here allows greater parallelism
1706 * since invalidating the TLB is a serializing operation.
1708 generation = smp_tlb_generation;
1709 if (smp_tlb_pmap == kernel_pmap)
1715 PCPU_SET(smp_tlb_done, generation);
1719 invlpg_handler(void)
1721 uint32_t generation;
1723 #ifdef COUNT_XINVLTLB_HITS
1724 xhits_pg[PCPU_GET(cpuid)]++;
1725 #endif /* COUNT_XINVLTLB_HITS */
1727 (*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
1728 #endif /* COUNT_IPIS */
1730 generation = smp_tlb_generation; /* Overlap with serialization */
1732 if (smp_tlb_pmap == kernel_pmap)
1734 invlpg(smp_tlb_addr1);
1735 PCPU_SET(smp_tlb_done, generation);
1739 invlrng_handler(void)
1741 vm_offset_t addr, addr2;
1742 uint32_t generation;
1744 #ifdef COUNT_XINVLTLB_HITS
1745 xhits_rng[PCPU_GET(cpuid)]++;
1746 #endif /* COUNT_XINVLTLB_HITS */
1748 (*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
1749 #endif /* COUNT_IPIS */
1751 addr = smp_tlb_addr1;
1752 addr2 = smp_tlb_addr2;
1753 generation = smp_tlb_generation; /* Overlap with serialization */
1755 if (smp_tlb_pmap == kernel_pmap)
1760 } while (addr < addr2);
1762 PCPU_SET(smp_tlb_done, generation);