xref: /freebsd-13-stable/sys/i386/i386/mp_machdep.c (revision 3bc80996974a61a4223eae4c1ccd47b6ee32a48a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 1996, by Steve Passe
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. The name of the developer may NOT be used to endorse or promote products
13  *    derived from this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 #include "opt_acpi.h"
30 #include "opt_apic.h"
31 #include "opt_cpu.h"
32 #include "opt_kstack_pages.h"
33 #include "opt_pmap.h"
34 #include "opt_sched.h"
35 #include "opt_smp.h"
36 
37 #if !defined(lint)
38 #if !defined(SMP)
39 #error How did you get here?
40 #endif
41 
42 #ifndef DEV_APIC
43 #error The apic device is required for SMP, add "device apic" to your config file.
44 #endif
45 #endif /* not lint */
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bus.h>
50 #include <sys/cons.h>	/* cngetc() */
51 #include <sys/cpuset.h>
52 #ifdef GPROF
53 #include <sys/gmon.h>
54 #endif
55 #include <sys/kdb.h>
56 #include <sys/kernel.h>
57 #include <sys/ktr.h>
58 #include <sys/lock.h>
59 #include <sys/malloc.h>
60 #include <sys/memrange.h>
61 #include <sys/mutex.h>
62 #include <sys/pcpu.h>
63 #include <sys/proc.h>
64 #include <sys/sched.h>
65 #include <sys/smp.h>
66 #include <sys/sysctl.h>
67 
68 #include <vm/vm.h>
69 #include <vm/vm_param.h>
70 #include <vm/pmap.h>
71 #include <vm/vm_kern.h>
72 #include <vm/vm_extern.h>
73 
74 #include <x86/apicreg.h>
75 #include <machine/clock.h>
76 #include <machine/cpu.h>
77 #include <machine/cputypes.h>
78 #include <x86/mca.h>
79 #include <machine/md_var.h>
80 #include <machine/pcb.h>
81 #include <machine/psl.h>
82 #include <machine/smp.h>
83 #include <machine/specialreg.h>
84 #include <x86/ucode.h>
85 
86 #ifdef DEV_ACPI
87 #include <contrib/dev/acpica/include/acpi.h>
88 #include <dev/acpica/acpivar.h>
89 #endif
90 
91 #define WARMBOOT_TARGET		0
92 #define WARMBOOT_OFF		(PMAP_MAP_LOW + 0x0467)
93 #define WARMBOOT_SEG		(PMAP_MAP_LOW + 0x0469)
94 
95 #define CMOS_REG		(0x70)
96 #define CMOS_DATA		(0x71)
97 #define BIOS_RESET		(0x0f)
98 #define BIOS_WARM		(0x0a)
99 
100 /*
101  * this code MUST be enabled here and in mpboot.s.
102  * it follows the very early stages of AP boot by placing values in CMOS ram.
103  * it NORMALLY will never be needed and thus the primitive method for enabling.
104  *
105 #define CHECK_POINTS
106  */
107 
108 #if defined(CHECK_POINTS)
109 #define CHECK_READ(A)	 (outb(CMOS_REG, (A)), inb(CMOS_DATA))
110 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
111 
112 #define CHECK_INIT(D);				\
113 	CHECK_WRITE(0x34, (D));			\
114 	CHECK_WRITE(0x35, (D));			\
115 	CHECK_WRITE(0x36, (D));			\
116 	CHECK_WRITE(0x37, (D));			\
117 	CHECK_WRITE(0x38, (D));			\
118 	CHECK_WRITE(0x39, (D));
119 
120 #define CHECK_PRINT(S);				\
121 	printf("%s: %d, %d, %d, %d, %d, %d\n",	\
122 	   (S),					\
123 	   CHECK_READ(0x34),			\
124 	   CHECK_READ(0x35),			\
125 	   CHECK_READ(0x36),			\
126 	   CHECK_READ(0x37),			\
127 	   CHECK_READ(0x38),			\
128 	   CHECK_READ(0x39));
129 
130 #else				/* CHECK_POINTS */
131 
132 #define CHECK_INIT(D)
133 #define CHECK_PRINT(S)
134 #define CHECK_WRITE(A, D)
135 
136 #endif				/* CHECK_POINTS */
137 
138 /*
139  * Local data and functions.
140  */
141 
142 static void	install_ap_tramp(void);
143 static int	start_all_aps(void);
144 static int	start_ap(int apic_id);
145 
146 static char *ap_copyout_buf;
147 static char *ap_tramp_stack_base;
148 
149 unsigned int boot_address;
150 
151 #define MiB(v)	(v ## ULL << 20)
152 
153 /* Allocate memory for the AP trampoline. */
154 void
alloc_ap_trampoline(vm_paddr_t * physmap,unsigned int * physmap_idx)155 alloc_ap_trampoline(vm_paddr_t *physmap, unsigned int *physmap_idx)
156 {
157 	unsigned int i;
158 	bool allocated;
159 
160 	allocated = false;
161 	for (i = *physmap_idx; i <= *physmap_idx; i -= 2) {
162 		/*
163 		 * Find a memory region big enough and below the 1MB boundary
164 		 * for the trampoline code.
165 		 * NB: needs to be page aligned.
166 		 */
167 		if (physmap[i] >= MiB(1) ||
168 		    (trunc_page(physmap[i + 1]) - round_page(physmap[i])) <
169 		    round_page(bootMP_size))
170 			continue;
171 
172 		allocated = true;
173 		/*
174 		 * Try to steal from the end of the region to mimic previous
175 		 * behaviour, else fallback to steal from the start.
176 		 */
177 		if (physmap[i + 1] < MiB(1)) {
178 			boot_address = trunc_page(physmap[i + 1]);
179 			if ((physmap[i + 1] - boot_address) < bootMP_size)
180 				boot_address -= round_page(bootMP_size);
181 			physmap[i + 1] = boot_address;
182 		} else {
183 			boot_address = round_page(physmap[i]);
184 			physmap[i] = boot_address + round_page(bootMP_size);
185 		}
186 		if (physmap[i] == physmap[i + 1] && *physmap_idx != 0) {
187 			memmove(&physmap[i], &physmap[i + 2],
188 			    sizeof(*physmap) * (*physmap_idx - i + 2));
189 			*physmap_idx -= 2;
190 		}
191 		break;
192 	}
193 
194 	if (!allocated) {
195 		boot_address = basemem * 1024 - bootMP_size;
196 		if (bootverbose)
197 			printf(
198 "Cannot find enough space for the boot trampoline, placing it at %#x",
199 			    boot_address);
200 	}
201 }
202 
203 /*
204  * Initialize the IPI handlers and start up the AP's.
205  */
206 void
cpu_mp_start(void)207 cpu_mp_start(void)
208 {
209 	int i;
210 
211 	/* Initialize the logical ID to APIC ID table. */
212 	for (i = 0; i < MAXCPU; i++) {
213 		cpu_apic_ids[i] = -1;
214 	}
215 
216 	/* Install an inter-CPU IPI for TLB invalidation */
217 	setidt(IPI_INVLTLB, IDTVEC(invltlb),
218 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
219 	setidt(IPI_INVLPG, IDTVEC(invlpg),
220 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
221 	setidt(IPI_INVLRNG, IDTVEC(invlrng),
222 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
223 
224 	/* Install an inter-CPU IPI for cache invalidation. */
225 	setidt(IPI_INVLCACHE, IDTVEC(invlcache),
226 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
227 
228 	/* Install an inter-CPU IPI for all-CPU rendezvous */
229 	setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous),
230 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
231 
232 	/* Install generic inter-CPU IPI handler */
233 	setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler),
234 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
235 
236 	/* Install an inter-CPU IPI for CPU stop/restart */
237 	setidt(IPI_STOP, IDTVEC(cpustop),
238 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
239 
240 	/* Install an inter-CPU IPI for CPU suspend/resume */
241 	setidt(IPI_SUSPEND, IDTVEC(cpususpend),
242 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
243 
244 	/* Install an IPI for calling delayed SWI */
245 	setidt(IPI_SWI, IDTVEC(ipi_swi),
246 	       SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
247 
248 	/* Set boot_cpu_id if needed. */
249 	if (boot_cpu_id == -1) {
250 		boot_cpu_id = PCPU_GET(apic_id);
251 		cpu_info[boot_cpu_id].cpu_bsp = 1;
252 	} else
253 		KASSERT(boot_cpu_id == PCPU_GET(apic_id),
254 		    ("BSP's APIC ID doesn't match boot_cpu_id"));
255 
256 	/* Probe logical/physical core configuration. */
257 	topo_probe();
258 
259 	assign_cpu_ids();
260 
261 	/* Start each Application Processor */
262 	start_all_aps();
263 
264 	set_interrupt_apic_ids();
265 
266 #if defined(DEV_ACPI) && MAXMEMDOM > 1
267 	acpi_pxm_set_cpu_locality();
268 #endif
269 }
270 
271 /*
272  * AP CPU's call this to initialize themselves.
273  */
274 void
init_secondary(void)275 init_secondary(void)
276 {
277 	struct pcpu *pc;
278 	struct i386tss *common_tssp;
279 	struct region_descriptor r_gdt, r_idt;
280 	int gsel_tss, myid, x;
281 	u_int cr0;
282 
283 	/* bootAP is set in start_ap() to our ID. */
284 	myid = bootAP;
285 
286 	/* Update microcode before doing anything else. */
287 	ucode_load_ap(myid);
288 
289 	/* Get per-cpu data */
290 	pc = &__pcpu[myid];
291 
292 	/* prime data page for it to use */
293 	pcpu_init(pc, myid, sizeof(struct pcpu));
294 	dpcpu_init(dpcpu, myid);
295 	pc->pc_apic_id = cpu_apic_ids[myid];
296 	pc->pc_prvspace = pc;
297 	pc->pc_curthread = 0;
298 	pc->pc_common_tssp = common_tssp = &(__pcpu[0].pc_common_tssp)[myid];
299 
300 	fix_cpuid();
301 
302 	gdt_segs[GPRIV_SEL].ssd_base = (int)pc;
303 	gdt_segs[GPROC0_SEL].ssd_base = (int)common_tssp;
304 	gdt_segs[GLDT_SEL].ssd_base = (int)ldt;
305 
306 	for (x = 0; x < NGDT; x++) {
307 		ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x].sd);
308 	}
309 
310 	r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
311 	r_gdt.rd_base = (int) &gdt[myid * NGDT];
312 	lgdt(&r_gdt);			/* does magic intra-segment return */
313 
314 	r_idt.rd_limit = sizeof(struct gate_descriptor) * NIDT - 1;
315 	r_idt.rd_base = (int)idt;
316 	lidt(&r_idt);
317 
318 	lldt(_default_ldt);
319 	PCPU_SET(currentldt, _default_ldt);
320 
321 	PCPU_SET(trampstk, (uintptr_t)ap_tramp_stack_base + TRAMP_STACK_SZ -
322 	    VM86_STACK_SPACE);
323 
324 	gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
325 	gdt[myid * NGDT + GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;
326 	common_tssp->tss_esp0 = PCPU_GET(trampstk);
327 	common_tssp->tss_ss0 = GSEL(GDATA_SEL, SEL_KPL);
328 	common_tssp->tss_ioopt = sizeof(struct i386tss) << 16;
329 	PCPU_SET(tss_gdt, &gdt[myid * NGDT + GPROC0_SEL].sd);
330 	PCPU_SET(common_tssd, *PCPU_GET(tss_gdt));
331 	ltr(gsel_tss);
332 
333 	PCPU_SET(fsgs_gdt, &gdt[myid * NGDT + GUFS_SEL].sd);
334 	PCPU_SET(copyout_buf, ap_copyout_buf);
335 
336 	/*
337 	 * Set to a known state:
338 	 * Set by mpboot.s: CR0_PG, CR0_PE
339 	 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
340 	 */
341 	cr0 = rcr0();
342 	cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
343 	load_cr0(cr0);
344 	CHECK_WRITE(0x38, 5);
345 
346 	/* signal our startup to the BSP. */
347 	mp_naps++;
348 	CHECK_WRITE(0x39, 6);
349 
350 	/* Spin until the BSP releases the AP's. */
351 	while (atomic_load_acq_int(&aps_ready) == 0)
352 		ia32_pause();
353 
354 	/* BSP may have changed PTD while we were waiting */
355 	invltlb();
356 
357 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
358 	lidt(&r_idt);
359 #endif
360 
361 	init_secondary_tail();
362 }
363 
364 /*
365  * start each AP in our list
366  */
367 #define TMPMAP_START 1
368 static int
start_all_aps(void)369 start_all_aps(void)
370 {
371 	u_char mpbiosreason;
372 	u_int32_t mpbioswarmvec;
373 	int apic_id, cpu;
374 
375 	mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
376 
377 	pmap_remap_lower(true);
378 
379 	/* install the AP 1st level boot code */
380 	install_ap_tramp();
381 
382 	/* save the current value of the warm-start vector */
383 	mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
384 	outb(CMOS_REG, BIOS_RESET);
385 	mpbiosreason = inb(CMOS_DATA);
386 
387 	/* take advantage of the P==V mapping for PTD[0] for AP boot */
388 
389 	/* start each AP */
390 	for (cpu = 1; cpu < mp_ncpus; cpu++) {
391 		apic_id = cpu_apic_ids[cpu];
392 
393 		/* allocate and set up a boot stack data page */
394 		bootstacks[cpu] = (char *)kmem_malloc(kstack_pages * PAGE_SIZE,
395 		    M_WAITOK | M_ZERO);
396 		dpcpu = (void *)kmem_malloc(DPCPU_SIZE, M_WAITOK | M_ZERO);
397 		/* setup a vector to our boot code */
398 		*((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
399 		*((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
400 		outb(CMOS_REG, BIOS_RESET);
401 		outb(CMOS_DATA, BIOS_WARM);	/* 'warm-start' */
402 
403 		bootSTK = (char *)bootstacks[cpu] + kstack_pages *
404 		    PAGE_SIZE - 4;
405 		bootAP = cpu;
406 
407 		ap_tramp_stack_base = pmap_trm_alloc(TRAMP_STACK_SZ, M_NOWAIT);
408 		ap_copyout_buf = pmap_trm_alloc(TRAMP_COPYOUT_SZ, M_NOWAIT);
409 
410 		/* attempt to start the Application Processor */
411 		CHECK_INIT(99);	/* setup checkpoints */
412 		if (!start_ap(apic_id)) {
413 			printf("AP #%d (PHY# %d) failed!\n", cpu, apic_id);
414 			CHECK_PRINT("trace");	/* show checkpoints */
415 			/* better panic as the AP may be running loose */
416 			printf("panic y/n? [y] ");
417 			if (cngetc() != 'n')
418 				panic("bye-bye");
419 		}
420 		CHECK_PRINT("trace");		/* show checkpoints */
421 
422 		CPU_SET(cpu, &all_cpus);	/* record AP in CPU map */
423 	}
424 
425 	pmap_remap_lower(false);
426 
427 	/* restore the warmstart vector */
428 	*(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
429 
430 	outb(CMOS_REG, BIOS_RESET);
431 	outb(CMOS_DATA, mpbiosreason);
432 
433 	/* number of APs actually started */
434 	return mp_naps;
435 }
436 
437 /*
438  * load the 1st level AP boot code into base memory.
439  */
440 
441 /* targets for relocation */
442 extern void bigJump(void);
443 extern void bootCodeSeg(void);
444 extern void bootDataSeg(void);
445 extern void MPentry(void);
446 extern u_int MP_GDT;
447 extern u_int mp_gdtbase;
448 
449 static void
install_ap_tramp(void)450 install_ap_tramp(void)
451 {
452 	int     x;
453 	int     size = *(int *) ((u_long) & bootMP_size);
454 	vm_offset_t va = boot_address;
455 	u_char *src = (u_char *) ((u_long) bootMP);
456 	u_char *dst = (u_char *) va;
457 	u_int   boot_base = (u_int) bootMP;
458 	u_int8_t *dst8;
459 	u_int16_t *dst16;
460 	u_int32_t *dst32;
461 
462 	KASSERT (size <= PAGE_SIZE,
463 	    ("'size' do not fit into PAGE_SIZE, as expected."));
464 	pmap_kenter(va, boot_address);
465 	pmap_invalidate_page (kernel_pmap, va);
466 	for (x = 0; x < size; ++x)
467 		*dst++ = *src++;
468 
469 	/*
470 	 * modify addresses in code we just moved to basemem. unfortunately we
471 	 * need fairly detailed info about mpboot.s for this to work.  changes
472 	 * to mpboot.s might require changes here.
473 	 */
474 
475 	/* boot code is located in KERNEL space */
476 	dst = (u_char *) va;
477 
478 	/* modify the lgdt arg */
479 	dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
480 	*dst32 = boot_address + ((u_int) & MP_GDT - boot_base);
481 
482 	/* modify the ljmp target for MPentry() */
483 	dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
484 	*dst32 = (u_int)MPentry;
485 
486 	/* modify the target for boot code segment */
487 	dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
488 	dst8 = (u_int8_t *) (dst16 + 1);
489 	*dst16 = (u_int) boot_address & 0xffff;
490 	*dst8 = ((u_int) boot_address >> 16) & 0xff;
491 
492 	/* modify the target for boot data segment */
493 	dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
494 	dst8 = (u_int8_t *) (dst16 + 1);
495 	*dst16 = (u_int) boot_address & 0xffff;
496 	*dst8 = ((u_int) boot_address >> 16) & 0xff;
497 }
498 
499 /*
500  * This function starts the AP (application processor) identified
501  * by the APIC ID 'physicalCpu'.  It does quite a "song and dance"
502  * to accomplish this.  This is necessary because of the nuances
503  * of the different hardware we might encounter.  It isn't pretty,
504  * but it seems to work.
505  */
506 static int
start_ap(int apic_id)507 start_ap(int apic_id)
508 {
509 	int vector, ms;
510 	int cpus;
511 
512 	/* calculate the vector */
513 	vector = (boot_address >> 12) & 0xff;
514 
515 	/* used as a watchpoint to signal AP startup */
516 	cpus = mp_naps;
517 
518 	ipi_startup(apic_id, vector);
519 
520 	/* Wait up to 5 seconds for it to start. */
521 	for (ms = 0; ms < 5000; ms++) {
522 		if (mp_naps > cpus)
523 			return 1;	/* return SUCCESS */
524 		DELAY(1000);
525 	}
526 	return 0;		/* return FAILURE */
527 }
528 
529 /*
530  * Flush the TLB on other CPU's
531  */
532 
533 /* Variables needed for SMP tlb shootdown. */
534 vm_offset_t smp_tlb_addr1, smp_tlb_addr2;
535 pmap_t smp_tlb_pmap;
536 volatile uint32_t smp_tlb_generation;
537 
538 /*
539  * Used by pmap to request cache or TLB invalidation on local and
540  * remote processors.  Mask provides the set of remote CPUs which are
541  * to be signalled with the invalidation IPI.  Vector specifies which
542  * invalidation IPI is used.  As an optimization, the curcpu_cb
543  * callback is invoked on the calling CPU while waiting for remote
544  * CPUs to complete the operation.
545  *
546  * The callback function is called unconditionally on the caller's
547  * underlying processor, even when this processor is not set in the
548  * mask.  So, the callback function must be prepared to handle such
549  * spurious invocations.
550  */
551 static void
smp_targeted_tlb_shootdown(cpuset_t mask,u_int vector,pmap_t pmap,vm_offset_t addr1,vm_offset_t addr2,smp_invl_cb_t curcpu_cb)552 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap,
553     vm_offset_t addr1, vm_offset_t addr2, smp_invl_cb_t curcpu_cb)
554 {
555 	cpuset_t other_cpus;
556 	volatile uint32_t *p_cpudone;
557 	uint32_t generation;
558 	int cpu;
559 
560 	/*
561 	 * It is not necessary to signal other CPUs while booting or
562 	 * when in the debugger.
563 	 */
564 	if (kdb_active || KERNEL_PANICKED() || !smp_started) {
565 		curcpu_cb(pmap, addr1, addr2);
566 		return;
567 	}
568 
569 	sched_pin();
570 
571 	/*
572 	 * Check for other cpus.  Return if none.
573 	 */
574 	if (CPU_ISFULLSET(&mask)) {
575 		if (mp_ncpus <= 1)
576 			goto nospinexit;
577 	} else {
578 		CPU_CLR(PCPU_GET(cpuid), &mask);
579 		if (CPU_EMPTY(&mask))
580 			goto nospinexit;
581 	}
582 
583 	KASSERT((read_eflags() & PSL_I) != 0,
584 	    ("smp_targeted_tlb_shootdown: interrupts disabled"));
585 	mtx_lock_spin(&smp_ipi_mtx);
586 	smp_tlb_addr1 = addr1;
587 	smp_tlb_addr2 = addr2;
588 	smp_tlb_pmap = pmap;
589 	generation = ++smp_tlb_generation;
590 	if (CPU_ISFULLSET(&mask)) {
591 		ipi_all_but_self(vector);
592 		other_cpus = all_cpus;
593 		CPU_CLR(PCPU_GET(cpuid), &other_cpus);
594 	} else {
595 		other_cpus = mask;
596 		ipi_selected(mask, vector);
597 	}
598 	curcpu_cb(pmap, addr1, addr2);
599 	CPU_FOREACH_ISSET(cpu, &other_cpus) {
600 		p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done;
601 		while (*p_cpudone != generation)
602 			ia32_pause();
603 	}
604 	mtx_unlock_spin(&smp_ipi_mtx);
605 	sched_unpin();
606 	return;
607 
608 nospinexit:
609 	curcpu_cb(pmap, addr1, addr2);
610 	sched_unpin();
611 }
612 
613 void
smp_masked_invltlb(cpuset_t mask,pmap_t pmap,smp_invl_cb_t curcpu_cb)614 smp_masked_invltlb(cpuset_t mask, pmap_t pmap, smp_invl_cb_t curcpu_cb)
615 {
616 	smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0, curcpu_cb);
617 #ifdef COUNT_XINVLTLB_HITS
618 	ipi_global++;
619 #endif
620 }
621 
622 void
smp_masked_invlpg(cpuset_t mask,vm_offset_t addr,pmap_t pmap,smp_invl_cb_t curcpu_cb)623 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap,
624     smp_invl_cb_t curcpu_cb)
625 {
626 	smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0, curcpu_cb);
627 #ifdef COUNT_XINVLTLB_HITS
628 	ipi_page++;
629 #endif
630 }
631 
632 void
smp_masked_invlpg_range(cpuset_t mask,vm_offset_t addr1,vm_offset_t addr2,pmap_t pmap,smp_invl_cb_t curcpu_cb)633 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2,
634     pmap_t pmap, smp_invl_cb_t curcpu_cb)
635 {
636 	smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap, addr1, addr2,
637 	    curcpu_cb);
638 #ifdef COUNT_XINVLTLB_HITS
639 	ipi_range++;
640 	ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
641 #endif
642 }
643 
644 void
smp_cache_flush(smp_invl_cb_t curcpu_cb)645 smp_cache_flush(smp_invl_cb_t curcpu_cb)
646 {
647 	smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL, 0, 0,
648 	    curcpu_cb);
649 }
650 
651 /*
652  * Handlers for TLB related IPIs
653  */
654 void
invltlb_handler(void)655 invltlb_handler(void)
656 {
657 	uint32_t generation;
658 
659 	trap_check_kstack();
660 #ifdef COUNT_XINVLTLB_HITS
661 	xhits_gbl[PCPU_GET(cpuid)]++;
662 #endif /* COUNT_XINVLTLB_HITS */
663 #ifdef COUNT_IPIS
664 	(*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
665 #endif /* COUNT_IPIS */
666 
667 	/*
668 	 * Reading the generation here allows greater parallelism
669 	 * since invalidating the TLB is a serializing operation.
670 	 */
671 	generation = smp_tlb_generation;
672 	if (smp_tlb_pmap == kernel_pmap)
673 		invltlb_glob();
674 	PCPU_SET(smp_tlb_done, generation);
675 }
676 
677 void
invlpg_handler(void)678 invlpg_handler(void)
679 {
680 	uint32_t generation;
681 
682 	trap_check_kstack();
683 #ifdef COUNT_XINVLTLB_HITS
684 	xhits_pg[PCPU_GET(cpuid)]++;
685 #endif /* COUNT_XINVLTLB_HITS */
686 #ifdef COUNT_IPIS
687 	(*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
688 #endif /* COUNT_IPIS */
689 
690 	generation = smp_tlb_generation;	/* Overlap with serialization */
691 	if (smp_tlb_pmap == kernel_pmap)
692 		invlpg(smp_tlb_addr1);
693 	PCPU_SET(smp_tlb_done, generation);
694 }
695 
696 void
invlrng_handler(void)697 invlrng_handler(void)
698 {
699 	vm_offset_t addr, addr2;
700 	uint32_t generation;
701 
702 	trap_check_kstack();
703 #ifdef COUNT_XINVLTLB_HITS
704 	xhits_rng[PCPU_GET(cpuid)]++;
705 #endif /* COUNT_XINVLTLB_HITS */
706 #ifdef COUNT_IPIS
707 	(*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
708 #endif /* COUNT_IPIS */
709 
710 	addr = smp_tlb_addr1;
711 	addr2 = smp_tlb_addr2;
712 	generation = smp_tlb_generation;	/* Overlap with serialization */
713 	if (smp_tlb_pmap == kernel_pmap) {
714 		do {
715 			invlpg(addr);
716 			addr += PAGE_SIZE;
717 		} while (addr < addr2);
718 	}
719 
720 	PCPU_SET(smp_tlb_done, generation);
721 }
722 
723 void
invlcache_handler(void)724 invlcache_handler(void)
725 {
726 	uint32_t generation;
727 
728 	trap_check_kstack();
729 #ifdef COUNT_IPIS
730 	(*ipi_invlcache_counts[PCPU_GET(cpuid)])++;
731 #endif /* COUNT_IPIS */
732 
733 	/*
734 	 * Reading the generation here allows greater parallelism
735 	 * since wbinvd is a serializing instruction.  Without the
736 	 * temporary, we'd wait for wbinvd to complete, then the read
737 	 * would execute, then the dependent write, which must then
738 	 * complete before return from interrupt.
739 	 */
740 	generation = smp_tlb_generation;
741 	wbinvd();
742 	PCPU_SET(smp_tlb_done, generation);
743 }
744