4 * Implements the architecture independant portion of the LWKT
7 * Types which must already be defined when this header is included by
8 * userland: struct md_thread
10 * $DragonFly: src/sys/sys/thread.h,v 1.97 2008/09/20 04:31:02 sephe Exp $
13 #ifndef _SYS_THREAD_H_
14 #define _SYS_THREAD_H_
16 #ifndef _SYS_STDINT_H_
17 #include <sys/stdint.h> /* __int types */
20 #include <sys/param.h> /* MAXCOMLEN */
23 #include <sys/queue.h> /* TAILQ_* macros */
25 #ifndef _SYS_MSGPORT_H_
26 #include <sys/msgport.h> /* lwkt_port */
29 #include <sys/time.h> /* struct timeval */
31 #ifndef _SYS_SPINLOCK_H_
32 #include <sys/spinlock.h>
34 #ifndef _SYS_IOSCHED_H_
35 #include <sys/iosched.h>
37 #ifndef _MACHINE_THREAD_H_
38 #include <machine/thread.h>
56 typedef struct lwkt_queue *lwkt_queue_t;
57 typedef struct lwkt_token *lwkt_token_t;
58 typedef struct lwkt_tokref *lwkt_tokref_t;
59 typedef struct lwkt_cpu_msg *lwkt_cpu_msg_t;
60 typedef struct lwkt_cpu_port *lwkt_cpu_port_t;
61 typedef struct lwkt_ipiq *lwkt_ipiq_t;
62 typedef struct lwkt_cpusync *lwkt_cpusync_t;
63 typedef struct thread *thread_t;
65 typedef TAILQ_HEAD(lwkt_queue, thread) lwkt_queue;
68 * Differentiation between kernel threads and user threads. Userland
69 * programs which want to access to kernel structures have to define
70 * _KERNEL_STRUCTURES. This is a kinda safety valve to prevent badly
71 * written user programs from getting an LWKT thread that is neither the
72 * kernel nor the user version.
74 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
75 #ifndef _MACHINE_THREAD_H_
76 #include <machine/thread.h> /* md_thread */
78 #ifndef _MACHINE_FRAME_H_
79 #include <machine/frame.h>
86 * Tokens are used to serialize access to information. They are 'soft'
87 * serialization entities that only stay in effect while a thread is
88 * running. If the thread blocks, other threads can run holding the same
89 * token(s). The tokens are reacquired when the original thread resumes.
91 * A thread can depend on its serialization remaining intact through a
92 * preemption. An interrupt which attempts to use the same token as the
93 * thread being preempted will reschedule itself for non-preemptive
94 * operation, so the new token code is capable of interlocking against
95 * interrupts as well as other cpus. This means that your token can only
96 * be (temporarily) lost if you *explicitly* block.
98 * Tokens are managed through a helper reference structure, lwkt_tokref. Each
99 * thread has a stack of tokref's to keep track of acquired tokens. Multiple
100 * tokref's may reference the same token.
103 typedef struct lwkt_token {
104 struct lwkt_tokref *t_ref; /* Owning ref or NULL */
105 intptr_t t_flags; /* MP lock required */
106 long t_collisions; /* Collision counter */
107 cpumask_t t_collmask; /* Collision cpu mask for resched */
108 const char *t_desc; /* Descriptive name */
111 #define LWKT_TOKEN_MPSAFE 0x0001
114 * Static initialization for a lwkt_token.
115 * UP - Not MPSAFE (full MP lock will also be acquired)
116 * MP - Is MPSAFE (only the token will be acquired)
118 #define LWKT_TOKEN_UP_INITIALIZER(name) \
127 #define LWKT_TOKEN_MP_INITIALIZER(name) \
130 .t_flags = LWKT_TOKEN_MPSAFE, \
137 * Assert that a particular token is held
139 #define LWKT_TOKEN_HELD(tok) _lwkt_token_held(tok, curthread)
141 #define ASSERT_LWKT_TOKEN_HELD(tok) \
142 KKASSERT(LWKT_TOKEN_HELD(tok))
144 #define ASSERT_NO_TOKENS_HELD(td) \
145 KKASSERT((td)->td_toks_stop == &td->td_toks_array[0])
148 * Assert that a particular token is held and we are in a hard
149 * code execution section (interrupt, ipi, or hard code section).
150 * Hard code sections are not allowed to block or potentially block.
151 * e.g. lwkt_gettoken() would only be ok if the token were already
154 #define ASSERT_LWKT_TOKEN_HARD(tok) \
156 globaldata_t zgd __debugvar = mycpu; \
157 KKASSERT((tok)->t_ref && \
158 (tok)->t_ref->tr_owner == zgd->gd_curthread && \
159 zgd->gd_intr_nesting_level > 0); \
163 * Assert that a particular token is held and we are in a normal
164 * critical section. Critical sections will not be preempted but
165 * can explicitly block (tsleep, lwkt_gettoken, etc).
167 #define ASSERT_LWKT_TOKEN_CRIT(tok) \
169 globaldata_t zgd __debugvar = mycpu; \
170 KKASSERT((tok)->t_ref && \
171 (tok)->t_ref->tr_owner == zgd->gd_curthread && \
172 zgd->gd_curthread->td_critcount > 0); \
176 lwkt_token_t tr_tok; /* token in question */
177 struct thread *tr_owner; /* me */
178 intptr_t tr_flags; /* copy of t_flags */
179 const void *tr_stallpc; /* stalled at pc */
182 #define MAXCPUFIFO 16 /* power of 2 */
183 #define MAXCPUFIFO_MASK (MAXCPUFIFO - 1)
184 #define LWKT_MAXTOKENS 32 /* max tokens beneficially held by thread */
187 * Always cast to ipifunc_t when registering an ipi. The actual ipi function
188 * is called with both the data and an interrupt frame, but the ipi function
189 * that is registered might only declare a data argument.
191 typedef void (*ipifunc1_t)(void *arg);
192 typedef void (*ipifunc2_t)(void *arg, int arg2);
193 typedef void (*ipifunc3_t)(void *arg, int arg2, struct intrframe *frame);
195 typedef struct lwkt_ipiq {
196 int ip_rindex; /* only written by target cpu */
197 int ip_xindex; /* written by target, indicates completion */
198 int ip_windex; /* only written by source cpu */
199 ipifunc3_t ip_func[MAXCPUFIFO];
200 void *ip_arg1[MAXCPUFIFO];
201 int ip_arg2[MAXCPUFIFO];
202 u_int ip_npoll; /* synchronization to avoid excess IPIs */
206 * CPU Synchronization structure. See lwkt_cpusync_start() and
207 * lwkt_cpusync_finish() for more information.
209 typedef void (*cpusync_func_t)(void *arg);
211 struct lwkt_cpusync {
212 cpumask_t cs_mask; /* cpus running the sync */
213 cpumask_t cs_mack; /* mask acknowledge */
214 cpusync_func_t cs_func; /* function to execute */
215 void *cs_data; /* function data */
219 * The standard message and queue structure used for communications between
220 * cpus. Messages are typically queued via a machine-specific non-linked
221 * FIFO matrix allowing any cpu to send a message to any other cpu without
224 typedef struct lwkt_cpu_msg {
225 void (*cm_func)(lwkt_cpu_msg_t msg); /* primary dispatch function */
226 int cm_code; /* request code if applicable */
227 int cm_cpu; /* reply to cpu */
228 thread_t cm_originator; /* originating thread for wakeup */
232 * Thread structure. Note that ownership of a thread structure is special
233 * cased and there is no 'token'. A thread is always owned by the cpu
234 * represented by td_gd, any manipulation of the thread by some other cpu
235 * must be done through cpu_*msg() functions. e.g. you could request
236 * ownership of a thread that way, or hand a thread off to another cpu.
238 * NOTE: td_ucred is synchronized from the p_ucred on user->kernel syscall,
239 * trap, and AST/signal transitions to provide a stable ucred for
240 * (primarily) system calls. This field will be NULL for pure kernel
247 TAILQ_ENTRY(thread) td_threadq;
248 TAILQ_ENTRY(thread) td_allq;
249 TAILQ_ENTRY(thread) td_sleepq;
250 lwkt_port td_msgport; /* built-in message port for replies */
251 struct lwp *td_lwp; /* (optional) associated lwp */
252 struct proc *td_proc; /* (optional) associated process */
253 struct pcb *td_pcb; /* points to pcb and top of kstack */
254 struct globaldata *td_gd; /* associated with this cpu */
255 const char *td_wmesg; /* string name for blockage */
256 const volatile void *td_wchan; /* waiting on channel */
257 int td_pri; /* 0-31, 31=highest priority (note 1) */
258 int td_critcount; /* critical section priority */
259 int td_flags; /* TDF flags */
260 int td_wdomain; /* domain for wchan address (typ 0) */
261 void (*td_preemptable)(struct thread *td, int critcount);
262 void (*td_release)(struct thread *td);
263 char *td_kstack; /* kernel stack */
264 int td_kstack_size; /* size of kernel stack */
265 char *td_sp; /* kernel stack pointer for LWKT restore */
266 void (*td_switch)(struct thread *ntd);
267 __uint64_t td_uticks; /* Statclock hits in user mode (uS) */
268 __uint64_t td_sticks; /* Statclock hits in system mode (uS) */
269 __uint64_t td_iticks; /* Statclock hits processing intr (uS) */
270 int td_locks; /* lockmgr lock debugging */
271 void *td_dsched_priv1; /* priv data for I/O schedulers */
272 int td_refs; /* hold position in gd_tdallq / hold free */
273 int td_nest_count; /* prevent splz nesting */
274 int td_unused01[2]; /* for future fields */
276 int td_cscount; /* cpu synchronization master */
278 int td_cscount_unused;
280 int td_unused02[4]; /* for future fields */
281 int td_unused03[4]; /* for future fields */
282 struct iosched_data td_iosdata; /* Dynamic I/O scheduling data */
283 struct timeval td_start; /* start time for a thread/process */
284 char td_comm[MAXCOMLEN+1]; /* typ 16+1 bytes */
285 struct thread *td_preempted; /* we preempted this thread */
286 struct ucred *td_ucred; /* synchronized from p_ucred */
287 struct caps_kinfo *td_caps; /* list of client and server registrations */
288 lwkt_tokref_t td_toks_stop;
289 struct lwkt_tokref td_toks_array[LWKT_MAXTOKENS];
290 int td_fairq_lticks; /* fairq wakeup accumulator reset */
291 int td_fairq_accum; /* fairq priority accumulator */
292 const void *td_mplock_stallpc; /* last mplock stall address */
293 #ifdef DEBUG_CRIT_SECTIONS
294 #define CRIT_DEBUG_ARRAY_SIZE 32
295 #define CRIT_DEBUG_ARRAY_MASK (CRIT_DEBUG_ARRAY_SIZE - 1)
296 const char *td_crit_debug_array[CRIT_DEBUG_ARRAY_SIZE];
297 int td_crit_debug_index;
298 int td_in_crit_report;
300 struct md_thread td_mach;
302 #define SPINLOCK_DEBUG_ARRAY_SIZE 32
303 int td_spinlock_stack_id[SPINLOCK_DEBUG_ARRAY_SIZE];
304 struct spinlock *td_spinlock_stack[SPINLOCK_DEBUG_ARRAY_SIZE];
305 void *td_spinlock_caller_pc[SPINLOCK_DEBUG_ARRAY_SIZE];
309 #define td_toks_base td_toks_array[0]
310 #define td_toks_end td_toks_array[LWKT_MAXTOKENS]
312 #define TD_TOKS_HELD(td) ((td)->td_toks_stop != &(td)->td_toks_base)
313 #define TD_TOKS_NOT_HELD(td) ((td)->td_toks_stop == &(td)->td_toks_base)
316 * Thread flags. Note that TDF_RUNNING is cleared on the old thread after
317 * we switch to the new one, which is necessary because LWKTs don't need
318 * to hold the BGL. This flag is used by the exit code and the managed
319 * thread migration code. Note in addition that preemption will cause
320 * TDF_RUNNING to be cleared temporarily, so any code checking TDF_RUNNING
321 * must also check TDF_PREEMPT_LOCK.
323 * LWKT threads stay on their (per-cpu) run queue while running, not to
324 * be confused with user processes which are removed from the user scheduling
325 * run queue while actually running.
327 * td_threadq can represent the thread on one of three queues... the LWKT
328 * run queue, a tsleep queue, or an lwkt blocking queue. The LWKT subsystem
329 * does not allow a thread to be scheduled if it already resides on some
332 #define TDF_RUNNING 0x0001 /* thread still active */
333 #define TDF_RUNQ 0x0002 /* on an LWKT run queue */
334 #define TDF_PREEMPT_LOCK 0x0004 /* I have been preempted */
335 #define TDF_PREEMPT_DONE 0x0008 /* acknowledge preemption complete */
336 #define TDF_UNUSED00000010 0x0010
337 #define TDF_MIGRATING 0x0020 /* thread is being migrated */
338 #define TDF_SINTR 0x0040 /* interruptability hint for 'ps' */
339 #define TDF_TSLEEPQ 0x0080 /* on a tsleep wait queue */
341 #define TDF_SYSTHREAD 0x0100 /* allocations may use reserve */
342 #define TDF_ALLOCATED_THREAD 0x0200 /* objcache allocated thread */
343 #define TDF_ALLOCATED_STACK 0x0400 /* objcache allocated stack */
344 #define TDF_VERBOSE 0x0800 /* verbose on exit */
345 #define TDF_DEADLKTREAT 0x1000 /* special lockmgr deadlock treatment */
346 #define TDF_STOPREQ 0x2000 /* suspend_kproc */
347 #define TDF_WAKEREQ 0x4000 /* resume_kproc */
348 #define TDF_TIMEOUT 0x8000 /* tsleep timeout */
349 #define TDF_INTTHREAD 0x00010000 /* interrupt thread */
350 #define TDF_TSLEEP_DESCHEDULED 0x00020000 /* tsleep core deschedule */
351 #define TDF_BLOCKED 0x00040000 /* Thread is blocked */
352 #define TDF_PANICWARN 0x00080000 /* panic warning in switch */
353 #define TDF_BLOCKQ 0x00100000 /* on block queue */
354 #define TDF_UNUSED00200000 0x00200000
355 #define TDF_EXITING 0x00400000 /* thread exiting */
356 #define TDF_USINGFP 0x00800000 /* thread using fp coproc */
357 #define TDF_KERNELFP 0x01000000 /* kernel using fp coproc */
358 #define TDF_UNUSED02000000 0x02000000
359 #define TDF_CRYPTO 0x04000000 /* crypto thread */
360 #define TDF_MARKER 0x80000000 /* fairq marker thread */
363 * Thread priorities. Typically only one thread from any given
364 * user process scheduling queue is on the LWKT run queue at a time.
365 * Remember that there is one LWKT run queue per cpu.
367 * Critical sections are handled by bumping td_pri above TDPRI_MAX, which
368 * causes interrupts to be masked as they occur. When this occurs a
369 * rollup flag will be set in mycpu->gd_reqflags.
371 #define TDPRI_IDLE_THREAD 0 /* the idle thread */
372 #define TDPRI_IDLE_WORK 1 /* idle work (page zero, etc) */
373 #define TDPRI_USER_SCHEDULER 2 /* user scheduler helper */
374 #define TDPRI_USER_IDLE 4 /* user scheduler idle */
375 #define TDPRI_USER_NORM 6 /* user scheduler normal */
376 #define TDPRI_USER_REAL 8 /* user scheduler real time */
377 #define TDPRI_KERN_LPSCHED 9 /* scheduler helper for userland sch */
378 #define TDPRI_KERN_USER 10 /* kernel / block in syscall */
379 #define TDPRI_KERN_DAEMON 12 /* kernel daemon (pageout, etc) */
380 #define TDPRI_SOFT_NORM 14 /* kernel / normal */
381 #define TDPRI_SOFT_TIMER 16 /* kernel / timer */
382 #define TDPRI_EXITING 19 /* exiting thread */
383 #define TDPRI_INT_SUPPORT 20 /* kernel / high priority support */
384 #define TDPRI_INT_LOW 27 /* low priority interrupt */
385 #define TDPRI_INT_MED 28 /* medium priority interrupt */
386 #define TDPRI_INT_HIGH 29 /* high priority interrupt */
390 * Scale is the approximate number of ticks for which we desire the
391 * entire gd_tdrunq to get service. With hz = 100 a scale of 8 is 80ms.
393 * Setting this value too small will result in inefficient switching
396 #define TDFAIRQ_SCALE 8
397 #define TDFAIRQ_MAX(gd) ((gd)->gd_fairq_total_pri * TDFAIRQ_SCALE)
399 #define LWKT_THREAD_STACK (UPAGES * PAGE_SIZE)
401 #define CACHE_NTHREADS 6
403 #define IN_CRITICAL_SECT(td) ((td)->td_critcount)
410 extern struct lwkt_token mp_token;
411 extern struct lwkt_token pmap_token;
412 extern struct lwkt_token dev_token;
413 extern struct lwkt_token vm_token;
414 extern struct lwkt_token vmspace_token;
415 extern struct lwkt_token kvm_token;
416 extern struct lwkt_token proc_token;
417 extern struct lwkt_token tty_token;
418 extern struct lwkt_token vnode_token;
419 extern struct lwkt_token vmobj_token;
424 extern void lwkt_init(void);
425 extern struct thread *lwkt_alloc_thread(struct thread *, int, int, int);
426 extern void lwkt_init_thread(struct thread *, void *, int, int,
427 struct globaldata *);
428 extern void lwkt_set_comm(thread_t, const char *, ...) __printflike(2, 3);
429 extern void lwkt_wait_free(struct thread *);
430 extern void lwkt_free_thread(struct thread *);
431 extern void lwkt_gdinit(struct globaldata *);
432 extern void lwkt_switch(void);
433 extern void lwkt_preempt(thread_t, int);
434 extern void lwkt_schedule(thread_t);
435 extern void lwkt_schedule_noresched(thread_t);
436 extern void lwkt_schedule_self(thread_t);
437 extern void lwkt_deschedule(thread_t);
438 extern void lwkt_deschedule_self(thread_t);
439 extern void lwkt_yield(void);
440 extern void lwkt_user_yield(void);
441 extern void lwkt_token_wait(void);
442 extern void lwkt_hold(thread_t);
443 extern void lwkt_rele(thread_t);
444 extern void lwkt_passive_release(thread_t);
445 extern void lwkt_maybe_splz(thread_t);
447 extern void lwkt_gettoken(lwkt_token_t);
448 extern void lwkt_gettoken_hard(lwkt_token_t);
449 extern int lwkt_trytoken(lwkt_token_t);
450 extern void lwkt_reltoken(lwkt_token_t);
451 extern void lwkt_reltoken_hard(lwkt_token_t);
452 extern int lwkt_cnttoken(lwkt_token_t, thread_t);
453 extern int lwkt_getalltokens(thread_t);
454 extern void lwkt_relalltokens(thread_t);
455 extern void lwkt_drain_token_requests(void);
456 extern void lwkt_token_init(lwkt_token_t, int, const char *);
457 extern void lwkt_token_uninit(lwkt_token_t);
459 extern void lwkt_token_pool_init(void);
460 extern lwkt_token_t lwkt_token_pool_lookup(void *);
461 extern lwkt_token_t lwkt_getpooltoken(void *);
462 extern void lwkt_relpooltoken(void *);
464 extern void lwkt_setpri(thread_t, int);
465 extern void lwkt_setpri_initial(thread_t, int);
466 extern void lwkt_setpri_self(int);
467 extern void lwkt_fairq_schedulerclock(thread_t td);
468 extern void lwkt_fairq_setpri_self(int pri);
469 extern int lwkt_fairq_push(int pri);
470 extern void lwkt_fairq_pop(int pri);
471 extern void lwkt_fairq_yield(void);
472 extern void lwkt_setcpu_self(struct globaldata *);
473 extern void lwkt_migratecpu(int);
477 extern void lwkt_giveaway(struct thread *);
478 extern void lwkt_acquire(struct thread *);
479 extern int lwkt_send_ipiq3(struct globaldata *, ipifunc3_t, void *, int);
480 extern int lwkt_send_ipiq3_passive(struct globaldata *, ipifunc3_t,
482 extern int lwkt_send_ipiq3_nowait(struct globaldata *, ipifunc3_t,
484 extern int lwkt_send_ipiq3_bycpu(int, ipifunc3_t, void *, int);
485 extern int lwkt_send_ipiq3_mask(cpumask_t, ipifunc3_t, void *, int);
486 extern void lwkt_wait_ipiq(struct globaldata *, int);
487 extern int lwkt_seq_ipiq(struct globaldata *);
488 extern void lwkt_process_ipiq(void);
489 extern void lwkt_process_ipiq_frame(struct intrframe *);
490 extern void lwkt_smp_stopped(void);
491 extern void lwkt_synchronize_ipiqs(const char *);
495 /* lwkt_cpusync_init() - inline function in sys/thread2.h */
496 extern void lwkt_cpusync_simple(cpumask_t, cpusync_func_t, void *);
497 extern void lwkt_cpusync_interlock(lwkt_cpusync_t);
498 extern void lwkt_cpusync_deinterlock(lwkt_cpusync_t);
500 extern void crit_panic(void) __dead2;
501 extern struct lwp *lwkt_preempted_proc(void);
503 extern int lwkt_create (void (*func)(void *), void *, struct thread **,
504 struct thread *, int, int,
505 const char *, ...) __printflike(7, 8);
506 extern void lwkt_exit (void) __dead2;
507 extern void lwkt_remove_tdallq (struct thread *);