2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/rtprio.h>
47 #include <sys/kinfo.h>
48 #include <sys/queue.h>
49 #include <sys/sysctl.h>
50 #include <sys/kthread.h>
51 #include <machine/cpu.h>
54 #include <sys/spinlock.h>
57 #include <sys/thread2.h>
58 #include <sys/spinlock2.h>
59 #include <sys/mplock2.h>
61 #include <sys/dsched.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_extern.h>
72 #include <machine/stdarg.h>
73 #include <machine/smp.h>
75 #if !defined(KTR_CTXSW)
76 #define KTR_CTXSW KTR_ALL
78 KTR_INFO_MASTER(ctxsw);
79 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
87 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
90 static int panic_on_cscount = 0;
92 static __int64_t switch_count = 0;
93 static __int64_t preempt_hit = 0;
94 static __int64_t preempt_miss = 0;
95 static __int64_t preempt_weird = 0;
96 static __int64_t token_contention_count[TDPRI_MAX+1] __debugvar;
97 static int lwkt_use_spin_port;
98 static struct objcache *thread_cache;
101 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
102 static void lwkt_setcpu_remote(void *arg);
105 extern void cpu_heavy_restore(void);
106 extern void cpu_lwkt_restore(void);
107 extern void cpu_kthread_restore(void);
108 extern void cpu_idle_restore(void);
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
114 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
117 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118 "Panic if attempting to switch lwkt's while mastering cpusync");
120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
123 "Successful preemption events");
124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
125 "Failed preemption events");
126 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
129 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_00, CTLFLAG_RW,
130 &token_contention_count[0], 0, "spinning due to token contention");
131 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_01, CTLFLAG_RW,
132 &token_contention_count[1], 0, "spinning due to token contention");
133 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_02, CTLFLAG_RW,
134 &token_contention_count[2], 0, "spinning due to token contention");
135 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_03, CTLFLAG_RW,
136 &token_contention_count[3], 0, "spinning due to token contention");
137 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_04, CTLFLAG_RW,
138 &token_contention_count[4], 0, "spinning due to token contention");
139 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_05, CTLFLAG_RW,
140 &token_contention_count[5], 0, "spinning due to token contention");
141 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_06, CTLFLAG_RW,
142 &token_contention_count[6], 0, "spinning due to token contention");
143 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_07, CTLFLAG_RW,
144 &token_contention_count[7], 0, "spinning due to token contention");
145 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_08, CTLFLAG_RW,
146 &token_contention_count[8], 0, "spinning due to token contention");
147 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_09, CTLFLAG_RW,
148 &token_contention_count[9], 0, "spinning due to token contention");
149 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_10, CTLFLAG_RW,
150 &token_contention_count[10], 0, "spinning due to token contention");
151 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_11, CTLFLAG_RW,
152 &token_contention_count[11], 0, "spinning due to token contention");
153 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_12, CTLFLAG_RW,
154 &token_contention_count[12], 0, "spinning due to token contention");
155 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_13, CTLFLAG_RW,
156 &token_contention_count[13], 0, "spinning due to token contention");
157 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_14, CTLFLAG_RW,
158 &token_contention_count[14], 0, "spinning due to token contention");
159 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_15, CTLFLAG_RW,
160 &token_contention_count[15], 0, "spinning due to token contention");
161 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_16, CTLFLAG_RW,
162 &token_contention_count[16], 0, "spinning due to token contention");
163 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_17, CTLFLAG_RW,
164 &token_contention_count[17], 0, "spinning due to token contention");
165 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_18, CTLFLAG_RW,
166 &token_contention_count[18], 0, "spinning due to token contention");
167 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_19, CTLFLAG_RW,
168 &token_contention_count[19], 0, "spinning due to token contention");
169 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_20, CTLFLAG_RW,
170 &token_contention_count[20], 0, "spinning due to token contention");
171 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_21, CTLFLAG_RW,
172 &token_contention_count[21], 0, "spinning due to token contention");
173 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_22, CTLFLAG_RW,
174 &token_contention_count[22], 0, "spinning due to token contention");
175 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_23, CTLFLAG_RW,
176 &token_contention_count[23], 0, "spinning due to token contention");
177 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_24, CTLFLAG_RW,
178 &token_contention_count[24], 0, "spinning due to token contention");
179 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_25, CTLFLAG_RW,
180 &token_contention_count[25], 0, "spinning due to token contention");
181 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_26, CTLFLAG_RW,
182 &token_contention_count[26], 0, "spinning due to token contention");
183 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_27, CTLFLAG_RW,
184 &token_contention_count[27], 0, "spinning due to token contention");
185 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_28, CTLFLAG_RW,
186 &token_contention_count[28], 0, "spinning due to token contention");
187 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_29, CTLFLAG_RW,
188 &token_contention_count[29], 0, "spinning due to token contention");
189 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_30, CTLFLAG_RW,
190 &token_contention_count[30], 0, "spinning due to token contention");
191 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_31, CTLFLAG_RW,
192 &token_contention_count[31], 0, "spinning due to token contention");
194 static int fairq_enable = 0;
195 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
196 &fairq_enable, 0, "Turn on fairq priority accumulators");
197 static int fairq_bypass = -1;
198 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
199 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
200 extern int lwkt_sched_debug;
201 int lwkt_sched_debug = 0;
202 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
203 &lwkt_sched_debug, 0, "Scheduler debug");
204 static int lwkt_spin_loops = 10;
205 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
206 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
207 static int lwkt_spin_reseq = 0;
208 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
209 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
210 static int lwkt_spin_monitor = 0;
211 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
212 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
213 static int lwkt_spin_fatal = 0; /* disabled */
214 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
215 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
216 static int preempt_enable = 1;
217 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
218 &preempt_enable, 0, "Enable preemption");
219 static int lwkt_cache_threads = 32;
220 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
221 &lwkt_cache_threads, 0, "thread+kstack cache");
223 static __cachealign int lwkt_cseq_rindex;
224 static __cachealign int lwkt_cseq_windex;
227 * These helper procedures handle the runq, they can only be called from
228 * within a critical section.
230 * WARNING! Prior to SMP being brought up it is possible to enqueue and
231 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
232 * instead of 'mycpu' when referencing the globaldata structure. Once
233 * SMP live enqueuing and dequeueing only occurs on the current cpu.
237 _lwkt_dequeue(thread_t td)
239 if (td->td_flags & TDF_RUNQ) {
240 struct globaldata *gd = td->td_gd;
242 td->td_flags &= ~TDF_RUNQ;
243 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
244 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
245 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
252 * NOTE: There are a limited number of lwkt threads runnable since user
253 * processes only schedule one at a time per cpu.
257 _lwkt_enqueue(thread_t td)
261 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
262 struct globaldata *gd = td->td_gd;
264 td->td_flags |= TDF_RUNQ;
265 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
267 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
268 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
270 while (xtd && xtd->td_pri >= td->td_pri)
271 xtd = TAILQ_NEXT(xtd, td_threadq);
273 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
275 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
279 * Request a LWKT reschedule if we are now at the head of the queue.
281 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
287 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
289 struct thread *td = (struct thread *)obj;
291 td->td_kstack = NULL;
292 td->td_kstack_size = 0;
293 td->td_flags = TDF_ALLOCATED_THREAD;
298 _lwkt_thread_dtor(void *obj, void *privdata)
300 struct thread *td = (struct thread *)obj;
302 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
303 ("_lwkt_thread_dtor: not allocated from objcache"));
304 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
305 td->td_kstack_size > 0,
306 ("_lwkt_thread_dtor: corrupted stack"));
307 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
311 * Initialize the lwkt s/system.
313 * Nominally cache up to 32 thread + kstack structures.
318 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
319 thread_cache = objcache_create_mbacked(
320 M_THREAD, sizeof(struct thread),
321 NULL, lwkt_cache_threads,
322 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
326 * Schedule a thread to run. As the current thread we can always safely
327 * schedule ourselves, and a shortcut procedure is provided for that
330 * (non-blocking, self contained on a per cpu basis)
333 lwkt_schedule_self(thread_t td)
335 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
336 crit_enter_quick(td);
337 KASSERT(td != &td->td_gd->gd_idlethread,
338 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
339 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
345 * Deschedule a thread.
347 * (non-blocking, self contained on a per cpu basis)
350 lwkt_deschedule_self(thread_t td)
352 crit_enter_quick(td);
358 * LWKTs operate on a per-cpu basis
360 * WARNING! Called from early boot, 'mycpu' may not work yet.
363 lwkt_gdinit(struct globaldata *gd)
365 TAILQ_INIT(&gd->gd_tdrunq);
366 TAILQ_INIT(&gd->gd_tdallq);
370 * Create a new thread. The thread must be associated with a process context
371 * or LWKT start address before it can be scheduled. If the target cpu is
372 * -1 the thread will be created on the current cpu.
374 * If you intend to create a thread without a process context this function
375 * does everything except load the startup and switcher function.
378 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
380 static int cpu_rotator;
381 globaldata_t gd = mycpu;
385 * If static thread storage is not supplied allocate a thread. Reuse
386 * a cached free thread if possible. gd_freetd is used to keep an exiting
387 * thread intact through the exit.
391 if ((td = gd->gd_freetd) != NULL) {
392 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
394 gd->gd_freetd = NULL;
396 td = objcache_get(thread_cache, M_WAITOK);
397 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
401 KASSERT((td->td_flags &
402 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
403 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
404 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
408 * Try to reuse cached stack.
410 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
411 if (flags & TDF_ALLOCATED_STACK) {
412 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
417 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
418 flags |= TDF_ALLOCATED_STACK;
425 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
430 * Initialize a preexisting thread structure. This function is used by
431 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
433 * All threads start out in a critical section at a priority of
434 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
435 * appropriate. This function may send an IPI message when the
436 * requested cpu is not the current cpu and consequently gd_tdallq may
437 * not be initialized synchronously from the point of view of the originating
440 * NOTE! we have to be careful in regards to creating threads for other cpus
441 * if SMP has not yet been activated.
446 lwkt_init_thread_remote(void *arg)
451 * Protected by critical section held by IPI dispatch
453 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
459 * lwkt core thread structural initialization.
461 * NOTE: All threads are initialized as mpsafe threads.
464 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
465 struct globaldata *gd)
467 globaldata_t mygd = mycpu;
469 bzero(td, sizeof(struct thread));
470 td->td_kstack = stack;
471 td->td_kstack_size = stksize;
472 td->td_flags = flags;
474 td->td_pri = TDPRI_KERN_DAEMON;
475 td->td_critcount = 1;
476 td->td_toks_stop = &td->td_toks_base;
477 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
478 lwkt_initport_spin(&td->td_msgport);
480 lwkt_initport_thread(&td->td_msgport, td);
481 pmap_init_thread(td);
484 * Normally initializing a thread for a remote cpu requires sending an
485 * IPI. However, the idlethread is setup before the other cpus are
486 * activated so we have to treat it as a special case. XXX manipulation
487 * of gd_tdallq requires the BGL.
489 if (gd == mygd || td == &gd->gd_idlethread) {
491 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
494 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
498 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
502 dsched_new_thread(td);
506 lwkt_set_comm(thread_t td, const char *ctl, ...)
511 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
513 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
517 lwkt_hold(thread_t td)
519 atomic_add_int(&td->td_refs, 1);
523 lwkt_rele(thread_t td)
525 KKASSERT(td->td_refs > 0);
526 atomic_add_int(&td->td_refs, -1);
530 lwkt_wait_free(thread_t td)
533 tsleep(td, 0, "tdreap", hz);
537 lwkt_free_thread(thread_t td)
539 KKASSERT(td->td_refs == 0);
540 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
541 if (td->td_flags & TDF_ALLOCATED_THREAD) {
542 objcache_put(thread_cache, td);
543 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
544 /* client-allocated struct with internally allocated stack */
545 KASSERT(td->td_kstack && td->td_kstack_size > 0,
546 ("lwkt_free_thread: corrupted stack"));
547 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
548 td->td_kstack = NULL;
549 td->td_kstack_size = 0;
551 KTR_LOG(ctxsw_deadtd, td);
556 * Switch to the next runnable lwkt. If no LWKTs are runnable then
557 * switch to the idlethread. Switching must occur within a critical
558 * section to avoid races with the scheduling queue.
560 * We always have full control over our cpu's run queue. Other cpus
561 * that wish to manipulate our queue must use the cpu_*msg() calls to
562 * talk to our cpu, so a critical section is all that is needed and
563 * the result is very, very fast thread switching.
565 * The LWKT scheduler uses a fixed priority model and round-robins at
566 * each priority level. User process scheduling is a totally
567 * different beast and LWKT priorities should not be confused with
568 * user process priorities.
570 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
571 * is not called by the current thread in the preemption case, only when
572 * the preempting thread blocks (in order to return to the original thread).
574 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
575 * migration and tsleep deschedule the current lwkt thread and call
576 * lwkt_switch(). In particular, the target cpu of the migration fully
577 * expects the thread to become non-runnable and can deadlock against
578 * cpusync operations if we run any IPIs prior to switching the thread out.
580 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
581 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
586 globaldata_t gd = mycpu;
587 thread_t td = gd->gd_curthread;
592 KKASSERT(gd->gd_processing_ipiq == 0);
595 * Switching from within a 'fast' (non thread switched) interrupt or IPI
596 * is illegal. However, we may have to do it anyway if we hit a fatal
597 * kernel trap or we have paniced.
599 * If this case occurs save and restore the interrupt nesting level.
601 if (gd->gd_intr_nesting_level) {
605 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
606 panic("lwkt_switch: Attempt to switch from a "
607 "a fast interrupt, ipi, or hard code section, "
611 savegdnest = gd->gd_intr_nesting_level;
612 savegdtrap = gd->gd_trap_nesting_level;
613 gd->gd_intr_nesting_level = 0;
614 gd->gd_trap_nesting_level = 0;
615 if ((td->td_flags & TDF_PANICWARN) == 0) {
616 td->td_flags |= TDF_PANICWARN;
617 kprintf("Warning: thread switch from interrupt, IPI, "
618 "or hard code section.\n"
619 "thread %p (%s)\n", td, td->td_comm);
623 gd->gd_intr_nesting_level = savegdnest;
624 gd->gd_trap_nesting_level = savegdtrap;
630 * Release our current user process designation if we are blocking
631 * or if a user reschedule was requested.
633 * NOTE: This function is NOT called if we are switching into or
634 * returning from a preemption.
636 * NOTE: Releasing our current user process designation may cause
637 * it to be assigned to another thread, which in turn will
638 * cause us to block in the usched acquire code when we attempt
639 * to return to userland.
641 * NOTE: On SMP systems this can be very nasty when heavy token
642 * contention is present so we want to be careful not to
643 * release the designation gratuitously.
645 if (td->td_release &&
646 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
654 if (TD_TOKS_HELD(td))
655 lwkt_relalltokens(td);
658 * We had better not be holding any spin locks, but don't get into an
659 * endless panic loop.
661 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
662 ("lwkt_switch: still holding %d exclusive spinlocks!",
663 gd->gd_spinlocks_wr));
668 if (td->td_cscount) {
669 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
671 if (panic_on_cscount)
672 panic("switching while mastering cpusync");
678 * If we had preempted another thread on this cpu, resume the preempted
679 * thread. This occurs transparently, whether the preempted thread
680 * was scheduled or not (it may have been preempted after descheduling
683 * We have to setup the MP lock for the original thread after backing
684 * out the adjustment that was made to curthread when the original
687 if ((ntd = td->td_preempted) != NULL) {
688 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
689 ntd->td_flags |= TDF_PREEMPT_DONE;
692 * The interrupt may have woken a thread up, we need to properly
693 * set the reschedule flag if the originally interrupted thread is
694 * at a lower priority.
696 * The interrupt may not have descheduled.
698 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
700 goto havethread_preempted;
704 * If we cannot obtain ownership of the tokens we cannot immediately
705 * schedule the target thread.
707 * Reminder: Again, we cannot afford to run any IPIs in this path if
708 * the current thread has been descheduled.
711 clear_lwkt_resched();
714 * Hotpath - pull the head of the run queue and attempt to schedule
718 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
722 * Runq is empty, switch to idle to allow it to halt.
724 ntd = &gd->gd_idlethread;
726 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
727 ASSERT_NO_TOKENS_HELD(ntd);
729 cpu_time.cp_msg[0] = 0;
730 cpu_time.cp_stallpc = 0;
737 * Hotpath - schedule ntd.
739 * NOTE: For UP there is no mplock and lwkt_getalltokens()
742 if (TD_TOKS_NOT_HELD(ntd) ||
743 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
749 * Coldpath (SMP only since tokens always succeed on UP)
751 * We had some contention on the thread we wanted to schedule.
752 * What we do now is try to find a thread that we can schedule
755 * The coldpath scan does NOT rearrange threads in the run list.
756 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
757 * the next tick whenever the current head is not the current thread.
760 ++token_contention_count[ntd->td_pri];
764 if (fairq_bypass > 0)
768 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
770 * Never schedule threads returning to userland or the
771 * user thread scheduler helper thread when higher priority
772 * threads are present.
774 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
782 if (TD_TOKS_NOT_HELD(ntd) ||
783 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
787 ++token_contention_count[ntd->td_pri];
794 * We exhausted the run list, meaning that all runnable threads
798 ntd = &gd->gd_idlethread;
800 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
801 ASSERT_NO_TOKENS_HELD(ntd);
802 /* contention case, do not clear contention mask */
806 * We are going to have to retry but if the current thread is not
807 * on the runq we instead switch through the idle thread to get away
808 * from the current thread. We have to flag for lwkt reschedule
809 * to prevent the idle thread from halting.
811 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
812 * instruct it to deal with the potential for deadlocks by
813 * ordering the tokens by address.
815 if ((td->td_flags & TDF_RUNQ) == 0) {
816 need_lwkt_resched(); /* prevent hlt */
819 #if defined(INVARIANTS) && defined(__amd64__)
820 if ((read_rflags() & PSL_I) == 0) {
822 panic("lwkt_switch() called with interrupts disabled");
827 * Number iterations so far. After a certain point we switch to
828 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
830 if (spinning < 0x7FFFFFFF)
835 * lwkt_getalltokens() failed in sorted token mode, we can use
836 * monitor/mwait in this case.
838 if (spinning >= lwkt_spin_loops &&
839 (cpu_mi_feature & CPU_MI_MONITOR) &&
842 cpu_mmw_pause_int(&gd->gd_reqflags,
843 (gd->gd_reqflags | RQF_SPINNING) &
844 ~RQF_IDLECHECK_WK_MASK);
849 * We already checked that td is still scheduled so this should be
855 * This experimental resequencer is used as a fall-back to reduce
856 * hw cache line contention by placing each core's scheduler into a
857 * time-domain-multplexed slot.
859 * The resequencer is disabled by default. It's functionality has
860 * largely been superceeded by the token algorithm which limits races
861 * to a subset of cores.
863 * The resequencer algorithm tends to break down when more than
864 * 20 cores are contending. What appears to happen is that new
865 * tokens can be obtained out of address-sorted order by new cores
866 * while existing cores languish in long delays between retries and
867 * wind up being starved-out of the token acquisition.
869 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
870 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
873 while ((oseq = lwkt_cseq_rindex) != cseq) {
876 if (cpu_mi_feature & CPU_MI_MONITOR) {
877 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
887 atomic_add_int(&lwkt_cseq_rindex, 1);
889 /* highest level for(;;) loop */
894 * If the thread we came up with is a higher or equal priority verses
895 * the thread at the head of the queue we move our thread to the
896 * front. This way we can always check the front of the queue.
898 * Clear gd_idle_repeat when doing a normal switch to a non-idle
901 ntd->td_wmesg = NULL;
902 ++gd->gd_cnt.v_swtch;
904 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
905 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
906 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
907 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
910 gd->gd_idle_repeat = 0;
912 havethread_preempted:
914 * If the new target does not need the MP lock and we are holding it,
915 * release the MP lock. If the new target requires the MP lock we have
916 * already acquired it for the target.
920 KASSERT(ntd->td_critcount,
921 ("priority problem in lwkt_switch %d %d",
922 td->td_critcount, ntd->td_critcount));
926 * Execute the actual thread switch operation. This function
927 * returns to the current thread and returns the previous thread
928 * (which may be different from the thread we switched to).
930 * We are responsible for marking ntd as TDF_RUNNING.
933 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
934 ntd->td_flags |= TDF_RUNNING;
935 lwkt_switch_return(td->td_switch(ntd));
936 /* ntd invalid, td_switch() can return a different thread_t */
945 /* NOTE: current cpu may have changed after switch */
950 * Called by assembly in the td_switch (thread restore path) for thread
951 * bootstrap cases which do not 'return' to lwkt_switch().
954 lwkt_switch_return(thread_t otd)
960 * Check if otd was migrating. Now that we are on ntd we can finish
961 * up the migration. This is a bit messy but it is the only place
962 * where td is known to be fully descheduled.
964 * We can only activate the migration if otd was migrating but not
965 * held on the cpu due to a preemption chain. We still have to
966 * clear TDF_RUNNING on the old thread either way.
968 * We are responsible for clearing the previously running thread's
971 if ((rgd = otd->td_migrate_gd) != NULL &&
972 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
973 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
974 (TDF_MIGRATING | TDF_RUNNING));
975 otd->td_migrate_gd = NULL;
976 otd->td_flags &= ~TDF_RUNNING;
977 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
979 otd->td_flags &= ~TDF_RUNNING;
982 otd->td_flags &= ~TDF_RUNNING;
987 * Request that the target thread preempt the current thread. Preemption
988 * only works under a specific set of conditions:
990 * - We are not preempting ourselves
991 * - The target thread is owned by the current cpu
992 * - We are not currently being preempted
993 * - The target is not currently being preempted
994 * - We are not holding any spin locks
995 * - The target thread is not holding any tokens
996 * - We are able to satisfy the target's MP lock requirements (if any).
998 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
999 * this is called via lwkt_schedule() through the td_preemptable callback.
1000 * critcount is the managed critical priority that we should ignore in order
1001 * to determine whether preemption is possible (aka usually just the crit
1002 * priority of lwkt_schedule() itself).
1004 * XXX at the moment we run the target thread in a critical section during
1005 * the preemption in order to prevent the target from taking interrupts
1006 * that *WE* can't. Preemption is strictly limited to interrupt threads
1007 * and interrupt-like threads, outside of a critical section, and the
1008 * preempted source thread will be resumed the instant the target blocks
1009 * whether or not the source is scheduled (i.e. preemption is supposed to
1010 * be as transparent as possible).
1013 lwkt_preempt(thread_t ntd, int critcount)
1015 struct globaldata *gd = mycpu;
1018 int save_gd_intr_nesting_level;
1021 * The caller has put us in a critical section. We can only preempt
1022 * if the caller of the caller was not in a critical section (basically
1023 * a local interrupt), as determined by the 'critcount' parameter. We
1024 * also can't preempt if the caller is holding any spinlocks (even if
1025 * he isn't in a critical section). This also handles the tokens test.
1027 * YYY The target thread must be in a critical section (else it must
1028 * inherit our critical section? I dunno yet).
1030 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
1032 td = gd->gd_curthread;
1033 if (preempt_enable == 0) {
1035 if (ntd->td_pri > td->td_pri)
1036 need_lwkt_resched();
1041 if (ntd->td_pri <= td->td_pri) {
1045 if (td->td_critcount > critcount) {
1048 need_lwkt_resched();
1053 if (ntd->td_gd != gd) {
1056 need_lwkt_resched();
1062 * We don't have to check spinlocks here as they will also bump
1065 * Do not try to preempt if the target thread is holding any tokens.
1066 * We could try to acquire the tokens but this case is so rare there
1067 * is no need to support it.
1069 KKASSERT(gd->gd_spinlocks_wr == 0);
1071 if (TD_TOKS_HELD(ntd)) {
1074 need_lwkt_resched();
1078 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1081 need_lwkt_resched();
1085 if (ntd->td_preempted) {
1088 need_lwkt_resched();
1092 KKASSERT(gd->gd_processing_ipiq == 0);
1095 * Since we are able to preempt the current thread, there is no need to
1096 * call need_lwkt_resched().
1098 * We must temporarily clear gd_intr_nesting_level around the switch
1099 * since switchouts from the target thread are allowed (they will just
1100 * return to our thread), and since the target thread has its own stack.
1102 * A preemption must switch back to the original thread, assert the
1106 ntd->td_preempted = td;
1107 td->td_flags |= TDF_PREEMPT_LOCK;
1108 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1109 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1110 gd->gd_intr_nesting_level = 0;
1111 ntd->td_flags |= TDF_RUNNING;
1112 xtd = td->td_switch(ntd);
1113 KKASSERT(xtd == ntd);
1114 lwkt_switch_return(xtd);
1115 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1117 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1118 ntd->td_preempted = NULL;
1119 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1123 * Conditionally call splz() if gd_reqflags indicates work is pending.
1124 * This will work inside a critical section but not inside a hard code
1127 * (self contained on a per cpu basis)
1132 globaldata_t gd = mycpu;
1133 thread_t td = gd->gd_curthread;
1135 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1136 gd->gd_intr_nesting_level == 0 &&
1137 td->td_nest_count < 2)
1144 * This version is integrated into crit_exit, reqflags has already
1145 * been tested but td_critcount has not.
1147 * We only want to execute the splz() on the 1->0 transition of
1148 * critcount and not in a hard code section or if too deeply nested.
1151 lwkt_maybe_splz(thread_t td)
1153 globaldata_t gd = td->td_gd;
1155 if (td->td_critcount == 0 &&
1156 gd->gd_intr_nesting_level == 0 &&
1157 td->td_nest_count < 2)
1164 * Drivers which set up processing co-threads can call this function to
1165 * run the co-thread at a higher priority and to allow it to preempt
1169 lwkt_set_interrupt_support_thread(void)
1171 thread_t td = curthread;
1173 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1174 td->td_flags |= TDF_INTTHREAD;
1175 td->td_preemptable = lwkt_preempt;
1180 * This function is used to negotiate a passive release of the current
1181 * process/lwp designation with the user scheduler, allowing the user
1182 * scheduler to schedule another user thread. The related kernel thread
1183 * (curthread) continues running in the released state.
1186 lwkt_passive_release(struct thread *td)
1188 struct lwp *lp = td->td_lwp;
1190 td->td_release = NULL;
1191 lwkt_setpri_self(TDPRI_KERN_USER);
1192 lp->lwp_proc->p_usched->release_curproc(lp);
1197 * This implements a LWKT yield, allowing a kernel thread to yield to other
1198 * kernel threads at the same or higher priority. This function can be
1199 * called in a tight loop and will typically only yield once per tick.
1201 * Most kernel threads run at the same priority in order to allow equal
1204 * (self contained on a per cpu basis)
1209 globaldata_t gd = mycpu;
1210 thread_t td = gd->gd_curthread;
1212 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1214 if (lwkt_resched_wanted()) {
1215 lwkt_schedule_self(curthread);
1221 * This yield is designed for kernel threads with a user context.
1223 * The kernel acting on behalf of the user is potentially cpu-bound,
1224 * this function will efficiently allow other threads to run and also
1225 * switch to other processes by releasing.
1227 * The lwkt_user_yield() function is designed to have very low overhead
1228 * if no yield is determined to be needed.
1231 lwkt_user_yield(void)
1233 globaldata_t gd = mycpu;
1234 thread_t td = gd->gd_curthread;
1237 * Always run any pending interrupts in case we are in a critical
1240 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1244 * Switch (which forces a release) if another kernel thread needs
1245 * the cpu, if userland wants us to resched, or if our kernel
1246 * quantum has run out.
1248 if (lwkt_resched_wanted() ||
1249 user_resched_wanted())
1256 * Reacquire the current process if we are released.
1258 * XXX not implemented atm. The kernel may be holding locks and such,
1259 * so we want the thread to continue to receive cpu.
1261 if (td->td_release == NULL && lp) {
1262 lp->lwp_proc->p_usched->acquire_curproc(lp);
1263 td->td_release = lwkt_passive_release;
1264 lwkt_setpri_self(TDPRI_USER_NORM);
1270 * Generic schedule. Possibly schedule threads belonging to other cpus and
1271 * deal with threads that might be blocked on a wait queue.
1273 * We have a little helper inline function which does additional work after
1274 * the thread has been enqueued, including dealing with preemption and
1275 * setting need_lwkt_resched() (which prevents the kernel from returning
1276 * to userland until it has processed higher priority threads).
1278 * It is possible for this routine to be called after a failed _enqueue
1279 * (due to the target thread migrating, sleeping, or otherwise blocked).
1280 * We have to check that the thread is actually on the run queue!
1284 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1286 if (ntd->td_flags & TDF_RUNQ) {
1287 if (ntd->td_preemptable) {
1288 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1295 _lwkt_schedule(thread_t td)
1297 globaldata_t mygd = mycpu;
1299 KASSERT(td != &td->td_gd->gd_idlethread,
1300 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1301 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1302 crit_enter_gd(mygd);
1303 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1304 if (td == mygd->gd_curthread) {
1308 * If we own the thread, there is no race (since we are in a
1309 * critical section). If we do not own the thread there might
1310 * be a race but the target cpu will deal with it.
1313 if (td->td_gd == mygd) {
1315 _lwkt_schedule_post(mygd, td, 1);
1317 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1321 _lwkt_schedule_post(mygd, td, 1);
1328 lwkt_schedule(thread_t td)
1334 lwkt_schedule_noresched(thread_t td) /* XXX not impl */
1342 * When scheduled remotely if frame != NULL the IPIQ is being
1343 * run via doreti or an interrupt then preemption can be allowed.
1345 * To allow preemption we have to drop the critical section so only
1346 * one is present in _lwkt_schedule_post.
1349 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1351 thread_t td = curthread;
1354 if (frame && ntd->td_preemptable) {
1355 crit_exit_noyield(td);
1356 _lwkt_schedule(ntd);
1357 crit_enter_quick(td);
1359 _lwkt_schedule(ntd);
1364 * Thread migration using a 'Pull' method. The thread may or may not be
1365 * the current thread. It MUST be descheduled and in a stable state.
1366 * lwkt_giveaway() must be called on the cpu owning the thread.
1368 * At any point after lwkt_giveaway() is called, the target cpu may
1369 * 'pull' the thread by calling lwkt_acquire().
1371 * We have to make sure the thread is not sitting on a per-cpu tsleep
1372 * queue or it will blow up when it moves to another cpu.
1374 * MPSAFE - must be called under very specific conditions.
1377 lwkt_giveaway(thread_t td)
1379 globaldata_t gd = mycpu;
1382 if (td->td_flags & TDF_TSLEEPQ)
1384 KKASSERT(td->td_gd == gd);
1385 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1386 td->td_flags |= TDF_MIGRATING;
1391 lwkt_acquire(thread_t td)
1395 int retry = 10000000;
1397 KKASSERT(td->td_flags & TDF_MIGRATING);
1402 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1403 crit_enter_gd(mygd);
1404 DEBUG_PUSH_INFO("lwkt_acquire");
1405 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1407 lwkt_process_ipiq();
1411 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1419 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1420 td->td_flags &= ~TDF_MIGRATING;
1423 crit_enter_gd(mygd);
1424 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1425 td->td_flags &= ~TDF_MIGRATING;
1433 * Generic deschedule. Descheduling threads other then your own should be
1434 * done only in carefully controlled circumstances. Descheduling is
1437 * This function may block if the cpu has run out of messages.
1440 lwkt_deschedule(thread_t td)
1444 if (td == curthread) {
1447 if (td->td_gd == mycpu) {
1450 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1460 * Set the target thread's priority. This routine does not automatically
1461 * switch to a higher priority thread, LWKT threads are not designed for
1462 * continuous priority changes. Yield if you want to switch.
1465 lwkt_setpri(thread_t td, int pri)
1467 if (td->td_pri != pri) {
1470 if (td->td_flags & TDF_RUNQ) {
1471 KKASSERT(td->td_gd == mycpu);
1483 * Set the initial priority for a thread prior to it being scheduled for
1484 * the first time. The thread MUST NOT be scheduled before or during
1485 * this call. The thread may be assigned to a cpu other then the current
1488 * Typically used after a thread has been created with TDF_STOPPREQ,
1489 * and before the thread is initially scheduled.
1492 lwkt_setpri_initial(thread_t td, int pri)
1495 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1500 lwkt_setpri_self(int pri)
1502 thread_t td = curthread;
1504 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1506 if (td->td_flags & TDF_RUNQ) {
1517 * hz tick scheduler clock for LWKT threads
1520 lwkt_schedulerclock(thread_t td)
1522 globaldata_t gd = td->td_gd;
1525 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1527 * If the current thread is at the head of the runq shift it to the
1528 * end of any equal-priority threads and request a LWKT reschedule
1531 xtd = TAILQ_NEXT(td, td_threadq);
1532 if (xtd && xtd->td_pri == td->td_pri) {
1533 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1534 while (xtd && xtd->td_pri == td->td_pri)
1535 xtd = TAILQ_NEXT(xtd, td_threadq);
1537 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1539 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1540 need_lwkt_resched();
1544 * If we scheduled a thread other than the one at the head of the
1545 * queue always request a reschedule every tick.
1547 need_lwkt_resched();
1552 * Migrate the current thread to the specified cpu.
1554 * This is accomplished by descheduling ourselves from the current cpu
1555 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1556 * 'old' thread wants to migrate after it has been completely switched out
1557 * and will complete the migration.
1559 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1561 * We must be sure to release our current process designation (if a user
1562 * process) before clearing out any tsleepq we are on because the release
1563 * code may re-add us.
1565 * We must be sure to remove ourselves from the current cpu's tsleepq
1566 * before potentially moving to another queue. The thread can be on
1567 * a tsleepq due to a left-over tsleep_interlock().
1571 lwkt_setcpu_self(globaldata_t rgd)
1574 thread_t td = curthread;
1576 if (td->td_gd != rgd) {
1577 crit_enter_quick(td);
1581 if (td->td_flags & TDF_TSLEEPQ)
1585 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1586 * trying to deschedule ourselves and switch away, then deschedule
1587 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1588 * call lwkt_switch() to complete the operation.
1590 td->td_flags |= TDF_MIGRATING;
1591 lwkt_deschedule_self(td);
1592 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1593 td->td_migrate_gd = rgd;
1597 * We are now on the target cpu
1599 KKASSERT(rgd == mycpu);
1600 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1601 crit_exit_quick(td);
1607 lwkt_migratecpu(int cpuid)
1612 rgd = globaldata_find(cpuid);
1613 lwkt_setcpu_self(rgd);
1619 * Remote IPI for cpu migration (called while in a critical section so we
1620 * do not have to enter another one).
1622 * The thread (td) has already been completely descheduled from the
1623 * originating cpu and we can simply assert the case. The thread is
1624 * assigned to the new cpu and enqueued.
1626 * The thread will re-add itself to tdallq when it resumes execution.
1629 lwkt_setcpu_remote(void *arg)
1632 globaldata_t gd = mycpu;
1634 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1637 td->td_flags &= ~TDF_MIGRATING;
1638 KKASSERT(td->td_migrate_gd == NULL);
1639 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1645 lwkt_preempted_proc(void)
1647 thread_t td = curthread;
1648 while (td->td_preempted)
1649 td = td->td_preempted;
1654 * Create a kernel process/thread/whatever. It shares it's address space
1655 * with proc0 - ie: kernel only.
1657 * If the cpu is not specified one will be selected. In the future
1658 * specifying a cpu of -1 will enable kernel thread migration between
1662 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1663 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1668 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1672 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1675 * Set up arg0 for 'ps' etc
1677 __va_start(ap, fmt);
1678 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1682 * Schedule the thread to run
1684 if ((td->td_flags & TDF_STOPREQ) == 0)
1687 td->td_flags &= ~TDF_STOPREQ;
1692 * Destroy an LWKT thread. Warning! This function is not called when
1693 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1694 * uses a different reaping mechanism.
1699 thread_t td = curthread;
1704 * Do any cleanup that might block here
1706 if (td->td_flags & TDF_VERBOSE)
1707 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1710 dsched_exit_thread(td);
1713 * Get us into a critical section to interlock gd_freetd and loop
1714 * until we can get it freed.
1716 * We have to cache the current td in gd_freetd because objcache_put()ing
1717 * it would rip it out from under us while our thread is still active.
1720 crit_enter_quick(td);
1721 while ((std = gd->gd_freetd) != NULL) {
1722 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1723 gd->gd_freetd = NULL;
1724 objcache_put(thread_cache, std);
1728 * Remove thread resources from kernel lists and deschedule us for
1729 * the last time. We cannot block after this point or we may end
1730 * up with a stale td on the tsleepq.
1732 if (td->td_flags & TDF_TSLEEPQ)
1734 lwkt_deschedule_self(td);
1735 lwkt_remove_tdallq(td);
1736 KKASSERT(td->td_refs == 0);
1741 KKASSERT(gd->gd_freetd == NULL);
1742 if (td->td_flags & TDF_ALLOCATED_THREAD)
1748 lwkt_remove_tdallq(thread_t td)
1750 KKASSERT(td->td_gd == mycpu);
1751 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1755 * Code reduction and branch prediction improvements. Call/return
1756 * overhead on modern cpus often degenerates into 0 cycles due to
1757 * the cpu's branch prediction hardware and return pc cache. We
1758 * can take advantage of this by not inlining medium-complexity
1759 * functions and we can also reduce the branch prediction impact
1760 * by collapsing perfectly predictable branches into a single
1761 * procedure instead of duplicating it.
1763 * Is any of this noticeable? Probably not, so I'll take the
1764 * smaller code size.
1767 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1769 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1775 thread_t td = curthread;
1776 int lcrit = td->td_critcount;
1778 td->td_critcount = 0;
1779 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1786 * Called from debugger/panic on cpus which have been stopped. We must still
1787 * process the IPIQ while stopped, even if we were stopped while in a critical
1790 * If we are dumping also try to process any pending interrupts. This may
1791 * or may not work depending on the state of the cpu at the point it was
1795 lwkt_smp_stopped(void)
1797 globaldata_t gd = mycpu;
1801 lwkt_process_ipiq();
1804 lwkt_process_ipiq();