usched: Allow process to change self cpu affinity
[dragonfly.git] / sys / sys / thread.h
blob052373106b26eeb9ad99159cd1202fa951397a07
1 /*
2 * SYS/THREAD.H
4 * Implements the architecture independant portion of the LWKT
5 * subsystem.
7 * Types which must already be defined when this header is included by
8 * userland: struct md_thread
9 */
11 #ifndef _SYS_THREAD_H_
12 #define _SYS_THREAD_H_
14 #ifndef _SYS_STDINT_H_
15 #include <sys/stdint.h> /* __int types */
16 #endif
17 #ifndef _SYS_PARAM_H_
18 #include <sys/param.h> /* MAXCOMLEN */
19 #endif
20 #ifndef _SYS_QUEUE_H_
21 #include <sys/queue.h> /* TAILQ_* macros */
22 #endif
23 #ifndef _SYS_MSGPORT_H_
24 #include <sys/msgport.h> /* lwkt_port */
25 #endif
26 #ifndef _SYS_TIME_H_
27 #include <sys/time.h> /* struct timeval */
28 #endif
29 #ifndef _SYS_LOCK_H
30 #include <sys/lock.h>
31 #endif
32 #ifndef _SYS_SPINLOCK_H_
33 #include <sys/spinlock.h>
34 #endif
35 #ifndef _SYS_IOSCHED_H_
36 #include <sys/iosched.h>
37 #endif
38 #include <machine/thread.h>
40 struct globaldata;
41 struct lwp;
42 struct proc;
43 struct thread;
44 struct lwkt_queue;
45 struct lwkt_token;
46 struct lwkt_tokref;
47 struct lwkt_ipiq;
48 struct lwkt_cpu_msg;
49 struct lwkt_cpu_port;
50 struct lwkt_cpusync;
51 union sysunion;
53 typedef struct lwkt_queue *lwkt_queue_t;
54 typedef struct lwkt_token *lwkt_token_t;
55 typedef struct lwkt_tokref *lwkt_tokref_t;
56 typedef struct lwkt_cpu_msg *lwkt_cpu_msg_t;
57 typedef struct lwkt_cpu_port *lwkt_cpu_port_t;
58 typedef struct lwkt_ipiq *lwkt_ipiq_t;
59 typedef struct lwkt_cpusync *lwkt_cpusync_t;
60 typedef struct thread *thread_t;
62 typedef TAILQ_HEAD(lwkt_queue, thread) lwkt_queue;
65 * Differentiation between kernel threads and user threads. Userland
66 * programs which want to access to kernel structures have to define
67 * _KERNEL_STRUCTURES. This is a kinda safety valve to prevent badly
68 * written user programs from getting an LWKT thread that is neither the
69 * kernel nor the user version.
71 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
72 #ifndef _CPU_FRAME_H_
73 #include <machine/frame.h>
74 #endif
75 #else
76 struct intrframe;
77 #endif
80 * Tokens are used to serialize access to information. They are 'soft'
81 * serialization entities that only stay in effect while a thread is
82 * running. If the thread blocks, other threads can run holding the same
83 * token(s). The tokens are reacquired when the original thread resumes.
85 * A thread can depend on its serialization remaining intact through a
86 * preemption. An interrupt which attempts to use the same token as the
87 * thread being preempted will reschedule itself for non-preemptive
88 * operation, so the new token code is capable of interlocking against
89 * interrupts as well as other cpus. This means that your token can only
90 * be (temporarily) lost if you *explicitly* block.
92 * Tokens are managed through a helper reference structure, lwkt_tokref. Each
93 * thread has a stack of tokref's to keep track of acquired tokens. Multiple
94 * tokref's may reference the same token.
96 * Tokens can be held shared or exclusive. An exclusive holder is able
97 * to set the TOK_EXCLUSIVE bit in t_count as long as no bit in the count
98 * mask is set. If unable to accomplish this TOK_EXCLREQ can be set instead
99 * which prevents any new shared acquisitions while the exclusive requestor
100 * spins in the scheduler. A shared holder can bump t_count by the increment
101 * value as long as neither TOK_EXCLUSIVE or TOK_EXCLREQ is set, else spin
102 * in the scheduler.
104 * Multiple exclusive tokens are handled by treating the additional tokens
105 * as a special case of the shared token, incrementing the count value. This
106 * reduces the complexity of the token release code.
109 typedef struct lwkt_token {
110 long t_count; /* Shared/exclreq/exclusive access */
111 struct lwkt_tokref *t_ref; /* Exclusive ref */
112 long t_collisions; /* Collision counter */
113 const char *t_desc; /* Descriptive name */
114 } lwkt_token;
116 #define TOK_EXCLUSIVE 0x00000001 /* Exclusive lock held */
117 #define TOK_EXCLREQ 0x00000002 /* Exclusive request pending */
118 #define TOK_INCR 4 /* Shared count increment */
119 #define TOK_COUNTMASK (~(long)(TOK_EXCLUSIVE|TOK_EXCLREQ))
122 * Static initialization for a lwkt_token.
124 #define LWKT_TOKEN_INITIALIZER(name) \
126 .t_count = 0, \
127 .t_ref = NULL, \
128 .t_collisions = 0, \
129 .t_desc = #name \
133 * Assert that a particular token is held
135 #define LWKT_TOKEN_HELD_ANY(tok) _lwkt_token_held_any(tok, curthread)
136 #define LWKT_TOKEN_HELD_EXCL(tok) _lwkt_token_held_excl(tok, curthread)
138 #define ASSERT_LWKT_TOKEN_HELD(tok) \
139 KKASSERT(LWKT_TOKEN_HELD_ANY(tok))
141 #define ASSERT_LWKT_TOKEN_HELD_EXCL(tok) \
142 KKASSERT(LWKT_TOKEN_HELD_EXCL(tok))
144 #define ASSERT_NO_TOKENS_HELD(td) \
145 KKASSERT((td)->td_toks_stop == &td->td_toks_array[0])
147 struct lwkt_tokref {
148 lwkt_token_t tr_tok; /* token in question */
149 long tr_count; /* TOK_EXCLUSIVE|TOK_EXCLREQ or 0 */
150 struct thread *tr_owner; /* me */
153 #define MAXCPUFIFO 32 /* power of 2 */
154 #define MAXCPUFIFO_MASK (MAXCPUFIFO - 1)
155 #define LWKT_MAXTOKENS 32 /* max tokens beneficially held by thread */
158 * Always cast to ipifunc_t when registering an ipi. The actual ipi function
159 * is called with both the data and an interrupt frame, but the ipi function
160 * that is registered might only declare a data argument.
162 typedef void (*ipifunc1_t)(void *arg);
163 typedef void (*ipifunc2_t)(void *arg, int arg2);
164 typedef void (*ipifunc3_t)(void *arg, int arg2, struct intrframe *frame);
166 struct lwkt_ipiq {
167 int ip_rindex; /* only written by target cpu */
168 int ip_xindex; /* written by target, indicates completion */
169 int ip_windex; /* only written by source cpu */
170 int ip_drain; /* drain source limit */
171 struct {
172 ipifunc3_t func;
173 void *arg1;
174 int arg2;
175 char filler[32 - sizeof(int) - sizeof(void *) * 2];
176 } ip_info[MAXCPUFIFO];
180 * CPU Synchronization structure. See lwkt_cpusync_start() and
181 * lwkt_cpusync_finish() for more information.
183 typedef void (*cpusync_func_t)(void *arg);
185 struct lwkt_cpusync {
186 cpumask_t cs_mask; /* cpus running the sync */
187 cpumask_t cs_mack; /* mask acknowledge */
188 cpusync_func_t cs_func; /* function to execute */
189 void *cs_data; /* function data */
193 * The standard message and queue structure used for communications between
194 * cpus. Messages are typically queued via a machine-specific non-linked
195 * FIFO matrix allowing any cpu to send a message to any other cpu without
196 * blocking.
198 typedef struct lwkt_cpu_msg {
199 void (*cm_func)(lwkt_cpu_msg_t msg); /* primary dispatch function */
200 int cm_code; /* request code if applicable */
201 int cm_cpu; /* reply to cpu */
202 thread_t cm_originator; /* originating thread for wakeup */
203 } lwkt_cpu_msg;
206 * Thread structure. Note that ownership of a thread structure is special
207 * cased and there is no 'token'. A thread is always owned by the cpu
208 * represented by td_gd, any manipulation of the thread by some other cpu
209 * must be done through cpu_*msg() functions. e.g. you could request
210 * ownership of a thread that way, or hand a thread off to another cpu.
212 * NOTE: td_ucred is synchronized from the p_ucred on user->kernel syscall,
213 * trap, and AST/signal transitions to provide a stable ucred for
214 * (primarily) system calls. This field will be NULL for pure kernel
215 * threads.
217 struct md_intr_info;
219 struct thread {
220 TAILQ_ENTRY(thread) td_threadq;
221 TAILQ_ENTRY(thread) td_allq;
222 TAILQ_ENTRY(thread) td_sleepq;
223 lwkt_port td_msgport; /* built-in message port for replies */
224 struct lwp *td_lwp; /* (optional) associated lwp */
225 struct proc *td_proc; /* (optional) associated process */
226 struct pcb *td_pcb; /* points to pcb and top of kstack */
227 struct globaldata *td_gd; /* associated with this cpu */
228 const char *td_wmesg; /* string name for blockage */
229 const volatile void *td_wchan; /* waiting on channel */
230 int td_pri; /* 0-31, 31=highest priority (note 1) */
231 int td_critcount; /* critical section priority */
232 u_int td_flags; /* TDF flags */
233 int td_wdomain; /* domain for wchan address (typ 0) */
234 void (*td_preemptable)(struct thread *td, int critcount);
235 void (*td_release)(struct thread *td);
236 char *td_kstack; /* kernel stack */
237 int td_kstack_size; /* size of kernel stack */
238 char *td_sp; /* kernel stack pointer for LWKT restore */
239 thread_t (*td_switch)(struct thread *ntd);
240 __uint64_t td_uticks; /* Statclock hits in user mode (uS) */
241 __uint64_t td_sticks; /* Statclock hits in system mode (uS) */
242 __uint64_t td_iticks; /* Statclock hits processing intr (uS) */
243 int td_locks; /* lockmgr lock debugging */
244 void *td_unused01; /* (future I/O scheduler heuristic) */
245 int td_refs; /* hold position in gd_tdallq / hold free */
246 int td_nest_count; /* prevent splz nesting */
247 u_int td_contended; /* token contention count */
248 u_int td_mpflags; /* flags can be set by foreign cpus */
249 int td_cscount; /* cpu synchronization master */
250 int td_wakefromcpu; /* who woke me up? */
251 int td_upri; /* user priority (sub-priority under td_pri) */
252 int td_type; /* thread type, TD_TYPE_ */
253 int td_tracker; /* for callers to debug lock counts */
254 int td_unused03[4]; /* for future fields */
255 struct iosched_data td_iosdata; /* Dynamic I/O scheduling data */
256 struct timeval td_start; /* start time for a thread/process */
257 char td_comm[MAXCOMLEN+1]; /* typ 16+1 bytes */
258 struct thread *td_preempted; /* we preempted this thread */
259 struct ucred *td_ucred; /* synchronized from p_ucred */
260 void *td_vmm; /* vmm private data */
261 lwkt_tokref_t td_toks_have; /* tokens we own */
262 lwkt_tokref_t td_toks_stop; /* tokens we want */
263 struct lwkt_tokref td_toks_array[LWKT_MAXTOKENS];
264 int td_fairq_load; /* fairq */
265 int td_fairq_count; /* fairq */
266 struct globaldata *td_migrate_gd; /* target gd for thread migration */
267 #ifdef DEBUG_CRIT_SECTIONS
268 #define CRIT_DEBUG_ARRAY_SIZE 32
269 #define CRIT_DEBUG_ARRAY_MASK (CRIT_DEBUG_ARRAY_SIZE - 1)
270 const char *td_crit_debug_array[CRIT_DEBUG_ARRAY_SIZE];
271 int td_crit_debug_index;
272 int td_in_crit_report;
273 #endif
274 struct md_thread td_mach;
275 #ifdef DEBUG_LOCKS
276 #define SPINLOCK_DEBUG_ARRAY_SIZE 32
277 int td_spinlock_stack_id[SPINLOCK_DEBUG_ARRAY_SIZE];
278 struct spinlock *td_spinlock_stack[SPINLOCK_DEBUG_ARRAY_SIZE];
279 void *td_spinlock_caller_pc[SPINLOCK_DEBUG_ARRAY_SIZE];
282 * Track lockmgr locks held; lk->lk_filename:lk->lk_lineno is the holder
284 #define LOCKMGR_DEBUG_ARRAY_SIZE 8
285 int td_lockmgr_stack_id[LOCKMGR_DEBUG_ARRAY_SIZE];
286 struct lock *td_lockmgr_stack[LOCKMGR_DEBUG_ARRAY_SIZE];
287 #endif
290 #define td_toks_base td_toks_array[0]
291 #define td_toks_end td_toks_array[LWKT_MAXTOKENS]
293 #define TD_TOKS_HELD(td) ((td)->td_toks_stop != &(td)->td_toks_base)
294 #define TD_TOKS_NOT_HELD(td) ((td)->td_toks_stop == &(td)->td_toks_base)
297 * Thread flags. Note that TDF_RUNNING is cleared on the old thread after
298 * we switch to the new one, which is necessary because LWKTs don't need
299 * to hold the BGL. This flag is used by the exit code and the managed
300 * thread migration code. Note in addition that preemption will cause
301 * TDF_RUNNING to be cleared temporarily, so any code checking TDF_RUNNING
302 * must also check TDF_PREEMPT_LOCK.
304 * LWKT threads stay on their (per-cpu) run queue while running, not to
305 * be confused with user processes which are removed from the user scheduling
306 * run queue while actually running.
308 * td_threadq can represent the thread on one of three queues... the LWKT
309 * run queue, a tsleep queue, or an lwkt blocking queue. The LWKT subsystem
310 * does not allow a thread to be scheduled if it already resides on some
311 * queue.
313 #define TDF_RUNNING 0x00000001 /* thread still active */
314 #define TDF_RUNQ 0x00000002 /* on an LWKT run queue */
315 #define TDF_PREEMPT_LOCK 0x00000004 /* I have been preempted */
316 #define TDF_PREEMPT_DONE 0x00000008 /* ac preemption complete */
317 #define TDF_NOSTART 0x00000010 /* do not schedule on create */
318 #define TDF_MIGRATING 0x00000020 /* thread is being migrated */
319 #define TDF_SINTR 0x00000040 /* interruptability for 'ps' */
320 #define TDF_TSLEEPQ 0x00000080 /* on a tsleep wait queue */
322 #define TDF_SYSTHREAD 0x00000100 /* reserve memory may be used */
323 #define TDF_ALLOCATED_THREAD 0x00000200 /* objcache allocated thread */
324 #define TDF_ALLOCATED_STACK 0x00000400 /* objcache allocated stack */
325 #define TDF_VERBOSE 0x00000800 /* verbose on exit */
326 #define TDF_DEADLKTREAT 0x00001000 /* special lockmgr treatment */
327 #define TDF_MARKER 0x00002000 /* tdallq list scan marker */
328 #define TDF_TIMEOUT_RUNNING 0x00004000 /* tsleep timeout race */
329 #define TDF_TIMEOUT 0x00008000 /* tsleep timeout */
330 #define TDF_INTTHREAD 0x00010000 /* interrupt thread */
331 #define TDF_TSLEEP_DESCHEDULED 0x00020000 /* tsleep core deschedule */
332 #define TDF_BLOCKED 0x00040000 /* Thread is blocked */
333 #define TDF_PANICWARN 0x00080000 /* panic warning in switch */
334 #define TDF_BLOCKQ 0x00100000 /* on block queue */
335 #define TDF_FORCE_SPINPORT 0x00200000
336 #define TDF_EXITING 0x00400000 /* thread exiting */
337 #define TDF_USINGFP 0x00800000 /* thread using fp coproc */
338 #define TDF_KERNELFP 0x01000000 /* kernel using fp coproc */
339 #define TDF_DELAYED_WAKEUP 0x02000000
340 #define TDF_FIXEDCPU 0x04000000 /* running cpu is fixed */
341 #define TDF_USERMODE 0x08000000 /* in or entering user mode */
342 #define TDF_NOFAULT 0x10000000 /* force onfault on fault */
344 #define TDF_MP_STOPREQ 0x00000001 /* suspend_kproc */
345 #define TDF_MP_WAKEREQ 0x00000002 /* resume_kproc */
346 #define TDF_MP_EXITWAIT 0x00000004 /* reaper, see lwp_wait() */
347 #define TDF_MP_EXITSIG 0x00000008 /* reaper, see lwp_wait() */
348 #define TDF_MP_BATCH_DEMARC 0x00000010 /* batch mode handling */
349 #define TDF_MP_DIDYIELD 0x00000020 /* effects scheduling */
351 #define TD_TYPE_GENERIC 0 /* generic thread */
352 #define TD_TYPE_CRYPTO 1 /* crypto thread */
353 #define TD_TYPE_NETISR 2 /* netisr thread */
356 * Thread priorities. Typically only one thread from any given
357 * user process scheduling queue is on the LWKT run queue at a time.
358 * Remember that there is one LWKT run queue per cpu.
360 * Critical sections are handled by bumping td_pri above TDPRI_MAX, which
361 * causes interrupts to be masked as they occur. When this occurs a
362 * rollup flag will be set in mycpu->gd_reqflags.
364 #define TDPRI_IDLE_THREAD 0 /* the idle thread */
365 #define TDPRI_IDLE_WORK 1 /* idle work (page zero, etc) */
366 #define TDPRI_USER_SCHEDULER 2 /* user scheduler helper */
367 #define TDPRI_USER_IDLE 4 /* user scheduler idle */
368 #define TDPRI_USER_NORM 6 /* user scheduler normal */
369 #define TDPRI_USER_REAL 8 /* user scheduler real time */
370 #define TDPRI_KERN_LPSCHED 9 /* (comparison point only) */
371 #define TDPRI_KERN_USER 10 /* kernel / block in syscall */
372 #define TDPRI_KERN_DAEMON 12 /* kernel daemon (pageout, etc) */
373 #define TDPRI_SOFT_NORM 14 /* kernel / normal */
374 #define TDPRI_SOFT_TIMER 16 /* kernel / timer */
375 #define TDPRI_UNUSED19 19
376 #define TDPRI_INT_SUPPORT 20 /* kernel / high priority support */
377 #define TDPRI_INT_LOW 27 /* low priority interrupt */
378 #define TDPRI_INT_MED 28 /* medium priority interrupt */
379 #define TDPRI_INT_HIGH 29 /* high priority interrupt */
380 #define TDPRI_MAX 31
382 #define LWKT_THREAD_STACK (UPAGES * PAGE_SIZE)
384 #define IN_CRITICAL_SECT(td) ((td)->td_critcount)
386 #ifdef _KERNEL
389 * Global tokens
391 extern struct lwkt_token mp_token;
392 extern struct lwkt_token pmap_token;
393 extern struct lwkt_token dev_token;
394 extern struct lwkt_token vm_token;
395 extern struct lwkt_token vmspace_token;
396 extern struct lwkt_token kvm_token;
397 extern struct lwkt_token sigio_token;
398 extern struct lwkt_token tty_token;
399 extern struct lwkt_token vnode_token;
400 extern struct lwkt_token revoke_token;
403 * Procedures
405 extern struct thread *lwkt_alloc_thread(struct thread *, int, int, int);
406 extern void lwkt_init_thread(struct thread *, void *, int, int,
407 struct globaldata *);
408 extern void lwkt_set_interrupt_support_thread(void);
409 extern void lwkt_set_comm(thread_t, const char *, ...) __printflike(2, 3);
410 extern void lwkt_free_thread(struct thread *);
411 extern void lwkt_gdinit(struct globaldata *);
412 extern void lwkt_switch(void);
413 extern void lwkt_switch_return(struct thread *);
414 extern void lwkt_preempt(thread_t, int);
415 extern void lwkt_schedule(thread_t);
416 extern void lwkt_schedule_noresched(thread_t);
417 extern void lwkt_schedule_self(thread_t);
418 extern void lwkt_deschedule(thread_t);
419 extern void lwkt_deschedule_self(thread_t);
420 extern void lwkt_yield(void);
421 extern void lwkt_yield_quick(void);
422 extern void lwkt_user_yield(void);
423 extern void lwkt_hold(thread_t);
424 extern void lwkt_rele(thread_t);
425 extern void lwkt_passive_release(thread_t);
426 extern void lwkt_maybe_splz(thread_t);
428 extern void lwkt_gettoken(lwkt_token_t);
429 extern void lwkt_gettoken_shared(lwkt_token_t);
430 extern int lwkt_trytoken(lwkt_token_t);
431 extern void lwkt_reltoken(lwkt_token_t);
432 extern int lwkt_cnttoken(lwkt_token_t, thread_t);
433 extern int lwkt_getalltokens(thread_t, int);
434 extern void lwkt_relalltokens(thread_t);
435 extern void lwkt_token_init(lwkt_token_t, const char *);
436 extern void lwkt_token_uninit(lwkt_token_t);
438 extern void lwkt_token_pool_init(void);
439 extern lwkt_token_t lwkt_token_pool_lookup(void *);
440 extern lwkt_token_t lwkt_getpooltoken(void *);
441 extern void lwkt_relpooltoken(void *);
443 extern void lwkt_token_swap(void);
445 extern void lwkt_setpri(thread_t, int);
446 extern void lwkt_setpri_initial(thread_t, int);
447 extern void lwkt_setpri_self(int);
448 extern void lwkt_schedulerclock(thread_t td);
449 extern void lwkt_setcpu_self(struct globaldata *);
450 extern void lwkt_migratecpu(int);
452 extern void lwkt_giveaway(struct thread *);
453 extern void lwkt_acquire(struct thread *);
454 extern int lwkt_send_ipiq3(struct globaldata *, ipifunc3_t, void *, int);
455 extern int lwkt_send_ipiq3_passive(struct globaldata *, ipifunc3_t,
456 void *, int);
457 extern int lwkt_send_ipiq3_bycpu(int, ipifunc3_t, void *, int);
458 extern int lwkt_send_ipiq3_mask(cpumask_t, ipifunc3_t, void *, int);
459 extern void lwkt_wait_ipiq(struct globaldata *, int);
460 extern void lwkt_process_ipiq(void);
461 extern void lwkt_process_ipiq_frame(struct intrframe *);
462 extern void lwkt_smp_stopped(void);
463 extern void lwkt_synchronize_ipiqs(const char *);
465 /* lwkt_cpusync_init() - inline function in sys/thread2.h */
466 extern void lwkt_cpusync_simple(cpumask_t, cpusync_func_t, void *);
467 extern void lwkt_cpusync_interlock(lwkt_cpusync_t);
468 extern void lwkt_cpusync_deinterlock(lwkt_cpusync_t);
469 extern void lwkt_cpusync_quick(lwkt_cpusync_t);
471 extern void crit_panic(void) __dead2;
472 extern struct lwp *lwkt_preempted_proc(void);
474 extern int lwkt_create (void (*func)(void *), void *, struct thread **,
475 struct thread *, int, int,
476 const char *, ...) __printflike(7, 8);
477 extern void lwkt_exit (void) __dead2;
478 extern void lwkt_remove_tdallq (struct thread *);
480 #endif
482 #endif