2 * Copyright (c) 2003,2004,2008 The DragonFly Project. All rights reserved.
3 * Copyright (c) 2008 Jordan Gordeev.
5 * This code is derived from software contributed to The DragonFly Project
6 * by Matthew Dillon <dillon@backplane.com>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * Copyright (c) 1990 The Regents of the University of California.
36 * All rights reserved.
38 * This code is derived from software contributed to Berkeley by
41 * Redistribution and use in source and binary forms, with or without
42 * modification, are permitted provided that the following conditions
44 * 1. Redistributions of source code must retain the above copyright
45 * notice, this list of conditions and the following disclaimer.
46 * 2. Redistributions in binary form must reproduce the above copyright
47 * notice, this list of conditions and the following disclaimer in the
48 * documentation and/or other materials provided with the distribution.
49 * 3. All advertising materials mentioning features or use of this software
50 * must display the following acknowledgement:
51 * This product includes software developed by the University of
52 * California, Berkeley and its contributors.
53 * 4. Neither the name of the University nor the names of its contributors
54 * may be used to endorse or promote products derived from this software
55 * without specific prior written permission.
57 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
58 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
59 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
60 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
61 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
62 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
63 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
64 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
65 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
66 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
69 * $FreeBSD: src/sys/i386/i386/swtch.s,v 1.89.2.10 2003/01/23 03:36:24 ps Exp $
72 //#include "use_npx.h"
74 #include <sys/rtprio.h>
76 #include <machine/asmacros.h>
77 #include <machine/segments.h>
79 #include <machine/pmap.h>
81 #include <machine_base/apic/apicreg.h>
83 #include <machine/lock.h>
88 #define MPLOCKED lock ;
96 .globl lwkt_switch_return
98 #if defined(SWTCH_OPTIM_STATS)
99 .globl swtch_optim_stats, tlb_flush_count
100 swtch_optim_stats: .long 0 /* number of _swtch_optims */
101 tlb_flush_count: .long 0
108 * cpu_heavy_switch(struct thread *next_thread)
110 * Switch from the current thread to a new thread. This entry
111 * is normally called via the thread->td_switch function, and will
112 * only be called when the current thread is a heavy weight process.
114 * Some instructions have been reordered to reduce pipeline stalls.
116 * YYY disable interrupts once giant is removed.
118 ENTRY(cpu_heavy_switch)
120 * Save RIP, RSP and callee-saved registers (RBX, RBP, R12-R15).
122 movq PCPU(curthread),%rcx
123 /* On top of the stack is the return adress. */
124 movq (%rsp),%rax /* (reorder optimization) */
125 movq TD_PCB(%rcx),%rdx /* RDX = PCB */
126 movq %rax,PCB_RIP(%rdx) /* return PC may be modified */
127 movq %rbx,PCB_RBX(%rdx)
128 movq %rsp,PCB_RSP(%rdx)
129 movq %rbp,PCB_RBP(%rdx)
130 movq %r12,PCB_R12(%rdx)
131 movq %r13,PCB_R13(%rdx)
132 movq %r14,PCB_R14(%rdx)
133 movq %r15,PCB_R15(%rdx)
136 * Clear the cpu bit in the pmap active mask. The restore
137 * function will set the bit in the pmap active mask.
139 * Special case: when switching between threads sharing the
140 * same vmspace if we avoid clearing the bit we do not have
141 * to reload %cr3 (if we clear the bit we could race page
142 * table ops done by other threads and would have to reload
143 * %cr3, because those ops will not know to IPI us).
145 movq %rcx,%rbx /* RBX = oldthread */
146 movq TD_LWP(%rcx),%rcx /* RCX = oldlwp */
147 movq TD_LWP(%rdi),%r13 /* R13 = newlwp */
148 movq LWP_VMSPACE(%rcx), %rcx /* RCX = oldvmspace */
149 testq %r13,%r13 /* might not be a heavy */
151 cmpq LWP_VMSPACE(%r13),%rcx /* same vmspace? */
154 movslq PCPU(cpuid), %rax
155 MPLOCKED btrq %rax, VM_PMAP+PM_ACTIVE(%rcx)
159 * Push the LWKT switch restore function, which resumes a heavy
160 * weight process. Note that the LWKT switcher is based on
161 * TD_SP, while the heavy weight process switcher is based on
162 * PCB_RSP. TD_SP is usually two ints pushed relative to
163 * PCB_RSP. We push the flags for later restore by cpu_heavy_restore.
167 movq $cpu_heavy_restore, %rax
169 movq %rsp,TD_SP(%rbx)
172 * Save debug regs if necessary
174 movq PCB_FLAGS(%rdx),%rax
175 andq $PCB_DBREGS,%rax
176 jz 1f /* no, skip over */
177 movq %dr7,%rax /* yes, do the save */
178 movq %rax,PCB_DR7(%rdx)
179 /* JG correct value? */
180 andq $0x0000fc00, %rax /* disable all watchpoints */
183 movq %rax,PCB_DR6(%rdx)
185 movq %rax,PCB_DR3(%rdx)
187 movq %rax,PCB_DR2(%rdx)
189 movq %rax,PCB_DR1(%rdx)
191 movq %rax,PCB_DR0(%rdx)
196 * Save the FP state if we have used the FP. Note that calling
197 * npxsave will NULL out PCPU(npxthread).
199 cmpq %rbx,PCPU(npxthread)
201 movq %rdi,%r12 /* save %rdi. %r12 is callee-saved */
202 movq TD_SAVEFPU(%rbx),%rdi
203 call npxsave /* do it in a big C function */
204 movq %r12,%rdi /* restore %rdi */
209 * Switch to the next thread, which was passed as an argument
210 * to cpu_heavy_switch(). The argument is in %rdi.
211 * Set the current thread, load the stack pointer,
212 * and 'ret' into the switch-restore function.
214 * The switch restore function expects the new thread to be in %rax
215 * and the old one to be in %rbx.
217 * There is a one-instruction window where curthread is the new
218 * thread but %rsp still points to the old thread's stack, but
219 * we are protected by a critical section so it is ok.
221 movq %rdi,%rax /* RAX = newtd, RBX = oldtd */
222 movq %rax,PCPU(curthread)
223 movq TD_SP(%rax),%rsp
227 * cpu_exit_switch(struct thread *next)
229 * The switch function is changed to this when a thread is going away
230 * for good. We have to ensure that the MMU state is not cached, and
231 * we don't bother saving the existing thread state before switching.
233 * At this point we are in a critical section and this cpu owns the
234 * thread's token, which serves as an interlock until the switchout is
237 ENTRY(cpu_exit_switch)
239 * Get us out of the vmspace
247 /* JG no increment of statistics counters? see cpu_heavy_restore */
250 movq PCPU(curthread),%rbx
253 * If this is a process/lwp, deactivate the pmap after we've
256 movq TD_LWP(%rbx),%rcx
259 movslq PCPU(cpuid), %rax
260 movq LWP_VMSPACE(%rcx), %rcx /* RCX = vmspace */
261 MPLOCKED btrq %rax, VM_PMAP+PM_ACTIVE(%rcx)
264 * Switch to the next thread. RET into the restore function, which
265 * expects the new thread in RAX and the old in RBX.
267 * There is a one-instruction window where curthread is the new
268 * thread but %rsp still points to the old thread's stack, but
269 * we are protected by a critical section so it is ok.
273 movq %rax,PCPU(curthread)
274 movq TD_SP(%rax),%rsp
278 * cpu_heavy_restore() (current thread in %rax on entry, old thread in %rbx)
280 * Restore the thread after an LWKT switch. This entry is normally
281 * called via the LWKT switch restore function, which was pulled
282 * off the thread stack and jumped to.
284 * This entry is only called if the thread was previously saved
285 * using cpu_heavy_switch() (the heavy weight process thread switcher),
286 * or when a new process is initially scheduled.
288 * NOTE: The lwp may be in any state, not necessarily LSRUN, because
289 * a preemption switch may interrupt the process and then return via
292 * YYY theoretically we do not have to restore everything here, a lot
293 * of this junk can wait until we return to usermode. But for now
294 * we restore everything.
296 * YYY the PCB crap is really crap, it makes startup a bitch because
297 * we can't switch away.
299 * YYY note: spl check is done in mi_switch when it splx()'s.
302 ENTRY(cpu_heavy_restore)
303 movq TD_PCB(%rax),%rdx /* RDX = PCB */
304 movq %rdx, PCPU(common_tss) + TSS_RSP0
307 #if defined(SWTCH_OPTIM_STATS)
308 incl _swtch_optim_stats
311 * Tell the pmap that our cpu is using the VMSPACE now. We cannot
312 * safely test/reload %cr3 until after we have set the bit in the
315 * We must do an interlocked test of the CPUMASK_BIT at the same
316 * time. If found to be set we will have to wait for it to clear
317 * and then do a forced reload of %cr3 (even if the value matches).
319 * XXX When switching between two LWPs sharing the same vmspace
320 * the cpu_heavy_switch() code currently avoids clearing the
321 * cpu bit in PM_ACTIVE. So if the bit is already set we can
322 * avoid checking for the interlock via CPUMASK_BIT. We currently
323 * do not perform this optimization.
325 * %rax is needed for the cmpxchgl so store newthread in %r12
328 movq TD_LWP(%rax),%rcx
329 movq LWP_VMSPACE(%rcx),%rcx /* RCX = vmspace */
331 movq %rax,%r12 /* save newthread ptr */
333 movq VM_PMAP+PM_ACTIVE(%rcx),%rax /* old contents */
334 movq PCPU(cpumask),%rsi /* new contents */
336 MPLOCKED cmpxchgq %rsi,VM_PMAP+PM_ACTIVE(%rcx)
342 btq $CPUMASK_BIT,%rax /* test interlock */
346 movq TD_PCB(%r12),%rdx /* XXX debugging unconditional */
347 movq PCB_CR3(%rdx),%rdx /* reloading of %cr3 */
351 movq %rcx,%rdi /* (found to be set) */
352 call pmap_interlock_wait /* pmap_interlock_wait(%rdi:vm) */
355 * Need unconditional load cr3
358 movq TD_PCB(%rax),%rdx /* RDX = PCB */
359 movq PCB_CR3(%rdx),%rcx /* RCX = desired CR3 */
360 jmp 2f /* unconditional reload */
362 movq %r12,%rax /* restore RAX = newthread */
364 movq PCPU(cpumask),%rsi
365 orq %rsi,VM_PMAP+PM_ACTIVE(%rcx)
368 * Restore the MMU address space. If it is the same as the last
369 * thread we don't have to invalidate the tlb (i.e. reload cr3).
370 * YYY which naturally also means that the PM_ACTIVE bit had better
371 * already have been set before we set it above, check? YYY
373 movq TD_PCB(%rax),%rdx /* RDX = PCB */
374 movq %cr3,%rsi /* RSI = current CR3 */
375 movq PCB_CR3(%rdx),%rcx /* RCX = desired CR3 */
379 #if defined(SWTCH_OPTIM_STATS)
380 decl _swtch_optim_stats
381 incl _tlb_flush_count
387 * NOTE: %rbx is the previous thread and %rax is the new thread.
388 * %rbx is retained throughout so we can return it.
390 * lwkt_switch[_return] is responsible for handling TDF_RUNNING.
394 * Deal with the PCB extension, restore the private tss
396 movq PCB_EXT(%rdx),%rdi /* check for a PCB extension */
397 movq $1,%rcx /* maybe mark use of a private tss */
404 * Going back to the common_tss. We may need to update TSS_RSP0
405 * which sets the top of the supervisor stack when entering from
406 * usermode. The PCB is at the top of the stack but we need another
407 * 16 bytes to take vm86 into account.
410 /*leaq -TF_SIZE(%rdx),%rcx*/
411 movq %rcx, PCPU(common_tss) + TSS_RSP0
414 cmpl $0,PCPU(private_tss) /* don't have to reload if */
415 je 3f /* already using the common TSS */
418 subq %rcx,%rcx /* unmark use of private tss */
421 * Get the address of the common TSS descriptor for the ltr.
422 * There is no way to get the address of a segment-accessed variable
423 * so we store a self-referential pointer at the base of the per-cpu
424 * data area and add the appropriate offset.
427 movq $gd_common_tssd, %rdi
428 /* JG name for "%gs:0"? */
432 * Move the correct TSS descriptor into the GDT slot, then reload
437 movl %rcx,PCPU(private_tss) /* mark/unmark private tss */
438 movq PCPU(tss_gdt), %rbx /* entry in GDT */
441 movl $GPROC0_SEL*8, %esi /* GSEL(entry, SEL_KPL) */
447 * Restore the user %gs and %fs
449 movq PCB_FSBASE(%rdx),%r9
450 cmpq PCPU(user_fs),%r9
453 movq %r9,PCPU(user_fs)
454 movl $MSR_FSBASE,%ecx
455 movl PCB_FSBASE(%r10),%eax
456 movl PCB_FSBASE+4(%r10),%edx
460 movq PCB_GSBASE(%rdx),%r9
461 cmpq PCPU(user_gs),%r9
464 movq %r9,PCPU(user_gs)
465 movl $MSR_KGSBASE,%ecx /* later swapgs moves it to GSBASE */
466 movl PCB_GSBASE(%r10),%eax
467 movl PCB_GSBASE+4(%r10),%edx
473 * Restore general registers. %rbx is restored later.
475 movq PCB_RSP(%rdx), %rsp
476 movq PCB_RBP(%rdx), %rbp
477 movq PCB_R12(%rdx), %r12
478 movq PCB_R13(%rdx), %r13
479 movq PCB_R14(%rdx), %r14
480 movq PCB_R15(%rdx), %r15
481 movq PCB_RIP(%rdx), %rax
488 * Restore the user LDT if we have one
490 cmpl $0, PCB_USERLDT(%edx)
492 movl _default_ldt,%eax
493 cmpl PCPU(currentldt),%eax
496 movl %eax,PCPU(currentldt)
505 * Restore the user TLS if we have one
513 * Restore the DEBUG register state if necessary.
515 movq PCB_FLAGS(%rdx),%rax
516 andq $PCB_DBREGS,%rax
517 jz 1f /* no, skip over */
518 movq PCB_DR6(%rdx),%rax /* yes, do the restore */
520 movq PCB_DR3(%rdx),%rax
522 movq PCB_DR2(%rdx),%rax
524 movq PCB_DR1(%rdx),%rax
526 movq PCB_DR0(%rdx),%rax
528 movq %dr7,%rax /* load dr7 so as not to disturb */
529 /* JG correct value? */
530 andq $0x0000fc00,%rax /* reserved bits */
531 /* JG we've got more registers on x86_64 */
532 movq PCB_DR7(%rdx),%rcx
533 /* JG correct value? */
534 andq $~0x0000fc00,%rcx
539 * Clear the QUICKRET flag when restoring a user process context
540 * so we don't try to do a quick syscall return.
543 andl $~RQF_QUICKRET,PCPU(reqflags)
545 movq PCB_RBX(%rdx),%rbx
549 * savectx(struct pcb *pcb)
551 * Update pcb, saving current processor state.
555 /* JG use %rdi instead of %rcx everywhere? */
558 /* caller's return address - child won't execute this routine */
560 movq %rax,PCB_RIP(%rcx)
563 movq %rax,PCB_CR3(%rcx)
565 movq %rbx,PCB_RBX(%rcx)
566 movq %rsp,PCB_RSP(%rcx)
567 movq %rbp,PCB_RBP(%rcx)
568 movq %r12,PCB_R12(%rcx)
569 movq %r13,PCB_R13(%rcx)
570 movq %r14,PCB_R14(%rcx)
571 movq %r15,PCB_R15(%rcx)
575 * If npxthread == NULL, then the npx h/w state is irrelevant and the
576 * state had better already be in the pcb. This is true for forks
577 * but not for dumps (the old book-keeping with FP flags in the pcb
578 * always lost for dumps because the dump pcb has 0 flags).
580 * If npxthread != NULL, then we have to save the npx h/w state to
581 * npxthread's pcb and copy it to the requested pcb, or save to the
582 * requested pcb and reload. Copying is easier because we would
583 * have to handle h/w bugs for reloading. We used to lose the
584 * parent's npx state for forks by forgetting to reload.
586 movq PCPU(npxthread),%rax
590 pushq %rcx /* target pcb */
591 movq TD_SAVEFPU(%rax),%rax /* originating savefpu area */
600 movq $PCB_SAVEFPU_SIZE,%rdx
601 leaq PCB_SAVEFPU(%rcx),%rcx
611 * cpu_idle_restore() (current thread in %rax on entry) (one-time execution)
613 * Don't bother setting up any regs other than %rbp so backtraces
614 * don't die. This restore function is used to bootstrap into the
615 * cpu_idle() LWKT only, after that cpu_lwkt_*() will be used for
618 * Clear TDF_RUNNING in old thread only after we've cleaned up %cr3.
619 * This only occurs during system boot so no special handling is
620 * required for migration.
622 * If we are an AP we have to call ap_init() before jumping to
623 * cpu_idle(). ap_init() will synchronize with the BP and finish
624 * setting up various ncpu-dependant globaldata fields. This may
625 * happen on UP as well as SMP if we happen to be simulating multiple
628 ENTRY(cpu_idle_restore)
636 andl $~TDF_RUNNING,TD_FLAGS(%rbx)
637 orl $TDF_RUNNING,TD_FLAGS(%rax) /* manual, no switch_return */
645 * ap_init can decide to enable interrupts early, but otherwise, or if
646 * we are UP, do it here.
652 * cpu_kthread_restore() (current thread is %rax on entry, previous is %rbx)
653 * (one-time execution)
655 * Don't bother setting up any regs other then %rbp so backtraces
656 * don't die. This restore function is used to bootstrap into an
657 * LWKT based kernel thread only. cpu_lwkt_switch() will be used
660 * Because this switch target does not 'return' to lwkt_switch()
661 * we have to call lwkt_switch_return(otd) to clean up otd.
664 * Since all of our context is on the stack we are reentrant and
665 * we can release our critical section and enable interrupts early.
667 ENTRY(cpu_kthread_restore)
670 movq TD_PCB(%rax),%r13
675 * rax and rbx come from the switchout code. Call
676 * lwkt_switch_return(otd).
678 * NOTE: unlike i386, %rsi and %rdi are not call-saved regs.
682 call lwkt_switch_return
684 decl TD_CRITCOUNT(%rax)
685 movq PCB_R12(%r13),%rdi /* argument to RBX function */
686 movq PCB_RBX(%r13),%rax /* thread function */
687 /* note: top of stack return address inherited by function */
691 * cpu_lwkt_switch(struct thread *)
693 * Standard LWKT switching function. Only non-scratch registers are
694 * saved and we don't bother with the MMU state or anything else.
696 * This function is always called while in a critical section.
698 * There is a one-instruction window where curthread is the new
699 * thread but %rsp still points to the old thread's stack, but
700 * we are protected by a critical section so it is ok.
702 ENTRY(cpu_lwkt_switch)
703 pushq %rbp /* JG note: GDB hacked to locate ebp rel to td_sp */
705 movq PCPU(curthread),%rbx /* becomes old thread in restore */
715 * Save the FP state if we have used the FP. Note that calling
716 * npxsave will NULL out PCPU(npxthread).
718 * We have to deal with the FP state for LWKT threads in case they
719 * happen to get preempted or block while doing an optimized
720 * bzero/bcopy/memcpy.
722 cmpq %rbx,PCPU(npxthread)
724 movq %rdi,%r12 /* save %rdi. %r12 is callee-saved */
725 movq TD_SAVEFPU(%rbx),%rdi
726 call npxsave /* do it in a big C function */
727 movq %r12,%rdi /* restore %rdi */
731 movq %rdi,%rax /* switch to this thread */
732 pushq $cpu_lwkt_restore
733 movq %rsp,TD_SP(%rbx)
735 * %rax contains new thread, %rbx contains old thread.
737 movq %rax,PCPU(curthread)
738 movq TD_SP(%rax),%rsp
742 * cpu_lwkt_restore() (current thread in %rax on entry)
744 * Standard LWKT restore function. This function is always called
745 * while in a critical section.
747 * Warning: due to preemption the restore function can be used to
748 * 'return' to the original thread. Interrupt disablement must be
749 * protected through the switch so we cannot run splz here.
751 * YYY we theoretically do not need to load KPML4phys into cr3, but if
752 * so we need a way to detect when the PTD we are using is being
753 * deleted due to a process exiting.
755 ENTRY(cpu_lwkt_restore)
756 movq KPML4phys,%rcx /* YYY borrow but beware desched/cpuchg/exit */
765 * Safety, clear RSP0 in the tss so it isn't pointing at the
766 * previous thread's kstack (if a heavy weight user thread).
767 * RSP0 should only be used in ring 3 transitions and kernel
768 * threads run in ring 0 so there should be none.
771 movq %rdx, PCPU(common_tss) + TSS_RSP0
774 * NOTE: %rbx is the previous thread and %rax is the new thread.
775 * %rbx is retained throughout so we can return it.
777 * lwkt_switch[_return] is responsible for handling TDF_RUNNING.