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
11 #ifndef _SYS_THREAD_H_
12 #define _SYS_THREAD_H_
14 #ifndef _SYS_STDINT_H_
15 #include <sys/stdint.h> /* __int types */
18 #include <sys/param.h> /* MAXCOMLEN */
21 #include <sys/queue.h> /* TAILQ_* macros */
23 #ifndef _SYS_MSGPORT_H_
24 #include <sys/msgport.h> /* lwkt_port */
27 #include <sys/time.h> /* struct timeval */
32 #ifndef _SYS_SPINLOCK_H_
33 #include <sys/spinlock.h>
35 #ifndef _SYS_IOSCHED_H_
36 #include <sys/iosched.h>
38 #include <machine/thread.h>
55 typedef struct lwkt_queue *lwkt_queue_t;
56 typedef struct lwkt_token *lwkt_token_t;
57 typedef struct lwkt_tokref *lwkt_tokref_t;
58 typedef struct lwkt_cpu_msg *lwkt_cpu_msg_t;
59 typedef struct lwkt_cpu_port *lwkt_cpu_port_t;
60 typedef struct lwkt_ipiq *lwkt_ipiq_t;
61 typedef struct lwkt_cpusync *lwkt_cpusync_t;
62 typedef struct thread *thread_t;
64 typedef TAILQ_HEAD(lwkt_queue, thread) lwkt_queue;
67 * Differentiation between kernel threads and user threads. Userland
68 * programs which want to access to kernel structures have to define
69 * _KERNEL_STRUCTURES. This is a kinda safety valve to prevent badly
70 * written user programs from getting an LWKT thread that is neither the
71 * kernel nor the user version.
73 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
74 #ifndef _MACHINE_THREAD_H_
75 #include <machine/thread.h> /* md_thread */
77 #ifndef _MACHINE_FRAME_H_
78 #include <machine/frame.h>
85 * Tokens are used to serialize access to information. They are 'soft'
86 * serialization entities that only stay in effect while a thread is
87 * running. If the thread blocks, other threads can run holding the same
88 * token(s). The tokens are reacquired when the original thread resumes.
90 * A thread can depend on its serialization remaining intact through a
91 * preemption. An interrupt which attempts to use the same token as the
92 * thread being preempted will reschedule itself for non-preemptive
93 * operation, so the new token code is capable of interlocking against
94 * interrupts as well as other cpus. This means that your token can only
95 * be (temporarily) lost if you *explicitly* block.
97 * Tokens are managed through a helper reference structure, lwkt_tokref. Each
98 * thread has a stack of tokref's to keep track of acquired tokens. Multiple
99 * tokref's may reference the same token.
102 typedef struct lwkt_token {
103 struct lwkt_tokref *t_ref; /* Owning ref or NULL */
104 intptr_t t_flags; /* MP lock required */
105 long t_collisions; /* Collision counter */
106 cpumask_t t_collmask; /* Collision cpu mask for resched */
107 const char *t_desc; /* Descriptive name */
111 * Static initialization for a lwkt_token.
113 #define LWKT_TOKEN_INITIALIZER(name) \
123 * Assert that a particular token is held
125 #define LWKT_TOKEN_HELD(tok) _lwkt_token_held(tok, curthread)
127 #define ASSERT_LWKT_TOKEN_HELD(tok) \
128 KKASSERT(LWKT_TOKEN_HELD(tok))
130 #define ASSERT_NO_TOKENS_HELD(td) \
131 KKASSERT((td)->td_toks_stop == &td->td_toks_array[0])
134 * Assert that a particular token is held and we are in a hard
135 * code execution section (interrupt, ipi, or hard code section).
136 * Hard code sections are not allowed to block or potentially block.
137 * e.g. lwkt_gettoken() would only be ok if the token were already
140 #define ASSERT_LWKT_TOKEN_HARD(tok) \
142 globaldata_t zgd __debugvar = mycpu; \
143 KKASSERT((tok)->t_ref && \
144 (tok)->t_ref->tr_owner == zgd->gd_curthread && \
145 zgd->gd_intr_nesting_level > 0); \
149 * Assert that a particular token is held and we are in a normal
150 * critical section. Critical sections will not be preempted but
151 * can explicitly block (tsleep, lwkt_gettoken, etc).
153 #define ASSERT_LWKT_TOKEN_CRIT(tok) \
155 globaldata_t zgd __debugvar = mycpu; \
156 KKASSERT((tok)->t_ref && \
157 (tok)->t_ref->tr_owner == zgd->gd_curthread && \
158 zgd->gd_curthread->td_critcount > 0); \
162 lwkt_token_t tr_tok; /* token in question */
163 struct thread *tr_owner; /* me */
164 intptr_t tr_flags; /* copy of t_flags */
165 const void *tr_stallpc; /* stalled at pc */
168 #define MAXCPUFIFO 16 /* power of 2 */
169 #define MAXCPUFIFO_MASK (MAXCPUFIFO - 1)
170 #define LWKT_MAXTOKENS 32 /* max tokens beneficially held by thread */
173 * Always cast to ipifunc_t when registering an ipi. The actual ipi function
174 * is called with both the data and an interrupt frame, but the ipi function
175 * that is registered might only declare a data argument.
177 typedef void (*ipifunc1_t)(void *arg);
178 typedef void (*ipifunc2_t)(void *arg, int arg2);
179 typedef void (*ipifunc3_t)(void *arg, int arg2, struct intrframe *frame);
181 typedef struct lwkt_ipiq {
182 int ip_rindex; /* only written by target cpu */
183 int ip_xindex; /* written by target, indicates completion */
184 int ip_windex; /* only written by source cpu */
185 ipifunc3_t ip_func[MAXCPUFIFO];
186 void *ip_arg1[MAXCPUFIFO];
187 int ip_arg2[MAXCPUFIFO];
188 u_int ip_npoll; /* synchronization to avoid excess IPIs */
192 * CPU Synchronization structure. See lwkt_cpusync_start() and
193 * lwkt_cpusync_finish() for more information.
195 typedef void (*cpusync_func_t)(void *arg);
197 struct lwkt_cpusync {
198 cpumask_t cs_mask; /* cpus running the sync */
199 cpumask_t cs_mack; /* mask acknowledge */
200 cpusync_func_t cs_func; /* function to execute */
201 void *cs_data; /* function data */
205 * The standard message and queue structure used for communications between
206 * cpus. Messages are typically queued via a machine-specific non-linked
207 * FIFO matrix allowing any cpu to send a message to any other cpu without
210 typedef struct lwkt_cpu_msg {
211 void (*cm_func)(lwkt_cpu_msg_t msg); /* primary dispatch function */
212 int cm_code; /* request code if applicable */
213 int cm_cpu; /* reply to cpu */
214 thread_t cm_originator; /* originating thread for wakeup */
218 * Thread structure. Note that ownership of a thread structure is special
219 * cased and there is no 'token'. A thread is always owned by the cpu
220 * represented by td_gd, any manipulation of the thread by some other cpu
221 * must be done through cpu_*msg() functions. e.g. you could request
222 * ownership of a thread that way, or hand a thread off to another cpu.
224 * NOTE: td_ucred is synchronized from the p_ucred on user->kernel syscall,
225 * trap, and AST/signal transitions to provide a stable ucred for
226 * (primarily) system calls. This field will be NULL for pure kernel
233 TAILQ_ENTRY(thread) td_threadq;
234 TAILQ_ENTRY(thread) td_allq;
235 TAILQ_ENTRY(thread) td_sleepq;
236 lwkt_port td_msgport; /* built-in message port for replies */
237 struct lwp *td_lwp; /* (optional) associated lwp */
238 struct proc *td_proc; /* (optional) associated process */
239 struct pcb *td_pcb; /* points to pcb and top of kstack */
240 struct globaldata *td_gd; /* associated with this cpu */
241 const char *td_wmesg; /* string name for blockage */
242 const volatile void *td_wchan; /* waiting on channel */
243 int td_pri; /* 0-31, 31=highest priority (note 1) */
244 int td_critcount; /* critical section priority */
245 int td_flags; /* TDF flags */
246 int td_wdomain; /* domain for wchan address (typ 0) */
247 void (*td_preemptable)(struct thread *td, int critcount);
248 void (*td_release)(struct thread *td);
249 char *td_kstack; /* kernel stack */
250 int td_kstack_size; /* size of kernel stack */
251 char *td_sp; /* kernel stack pointer for LWKT restore */
252 void (*td_switch)(struct thread *ntd);
253 __uint64_t td_uticks; /* Statclock hits in user mode (uS) */
254 __uint64_t td_sticks; /* Statclock hits in system mode (uS) */
255 __uint64_t td_iticks; /* Statclock hits processing intr (uS) */
256 int td_locks; /* lockmgr lock debugging */
257 void *td_dsched_priv1; /* priv data for I/O schedulers */
258 int td_refs; /* hold position in gd_tdallq / hold free */
259 int td_nest_count; /* prevent splz nesting */
260 int td_unused01[2]; /* for future fields */
262 int td_cscount; /* cpu synchronization master */
264 int td_cscount_unused;
266 int td_unused02[4]; /* for future fields */
267 int td_unused03[4]; /* for future fields */
268 struct iosched_data td_iosdata; /* Dynamic I/O scheduling data */
269 struct timeval td_start; /* start time for a thread/process */
270 char td_comm[MAXCOMLEN+1]; /* typ 16+1 bytes */
271 struct thread *td_preempted; /* we preempted this thread */
272 struct ucred *td_ucred; /* synchronized from p_ucred */
273 struct caps_kinfo *td_caps; /* list of client and server registrations */
274 lwkt_tokref_t td_toks_stop;
275 struct lwkt_tokref td_toks_array[LWKT_MAXTOKENS];
276 int td_fairq_lticks; /* fairq wakeup accumulator reset */
277 int td_fairq_accum; /* fairq priority accumulator */
278 const void *td_mplock_stallpc; /* last mplock stall address */
279 #ifdef DEBUG_CRIT_SECTIONS
280 #define CRIT_DEBUG_ARRAY_SIZE 32
281 #define CRIT_DEBUG_ARRAY_MASK (CRIT_DEBUG_ARRAY_SIZE - 1)
282 const char *td_crit_debug_array[CRIT_DEBUG_ARRAY_SIZE];
283 int td_crit_debug_index;
284 int td_in_crit_report;
286 struct md_thread td_mach;
288 #define SPINLOCK_DEBUG_ARRAY_SIZE 32
289 int td_spinlock_stack_id[SPINLOCK_DEBUG_ARRAY_SIZE];
290 struct spinlock *td_spinlock_stack[SPINLOCK_DEBUG_ARRAY_SIZE];
291 void *td_spinlock_caller_pc[SPINLOCK_DEBUG_ARRAY_SIZE];
294 * Track lockmgr locks held; lk->lk_filename:lk->lk_lineno is the holder
296 #define LOCKMGR_DEBUG_ARRAY_SIZE 8
297 int td_lockmgr_stack_id[LOCKMGR_DEBUG_ARRAY_SIZE];
298 struct lock *td_lockmgr_stack[LOCKMGR_DEBUG_ARRAY_SIZE];
302 #define td_toks_base td_toks_array[0]
303 #define td_toks_end td_toks_array[LWKT_MAXTOKENS]
305 #define TD_TOKS_HELD(td) ((td)->td_toks_stop != &(td)->td_toks_base)
306 #define TD_TOKS_NOT_HELD(td) ((td)->td_toks_stop == &(td)->td_toks_base)
309 * Thread flags. Note that TDF_RUNNING is cleared on the old thread after
310 * we switch to the new one, which is necessary because LWKTs don't need
311 * to hold the BGL. This flag is used by the exit code and the managed
312 * thread migration code. Note in addition that preemption will cause
313 * TDF_RUNNING to be cleared temporarily, so any code checking TDF_RUNNING
314 * must also check TDF_PREEMPT_LOCK.
316 * LWKT threads stay on their (per-cpu) run queue while running, not to
317 * be confused with user processes which are removed from the user scheduling
318 * run queue while actually running.
320 * td_threadq can represent the thread on one of three queues... the LWKT
321 * run queue, a tsleep queue, or an lwkt blocking queue. The LWKT subsystem
322 * does not allow a thread to be scheduled if it already resides on some
325 #define TDF_RUNNING 0x0001 /* thread still active */
326 #define TDF_RUNQ 0x0002 /* on an LWKT run queue */
327 #define TDF_PREEMPT_LOCK 0x0004 /* I have been preempted */
328 #define TDF_PREEMPT_DONE 0x0008 /* acknowledge preemption complete */
329 #define TDF_UNUSED00000010 0x0010
330 #define TDF_MIGRATING 0x0020 /* thread is being migrated */
331 #define TDF_SINTR 0x0040 /* interruptability hint for 'ps' */
332 #define TDF_TSLEEPQ 0x0080 /* on a tsleep wait queue */
334 #define TDF_SYSTHREAD 0x0100 /* allocations may use reserve */
335 #define TDF_ALLOCATED_THREAD 0x0200 /* objcache allocated thread */
336 #define TDF_ALLOCATED_STACK 0x0400 /* objcache allocated stack */
337 #define TDF_VERBOSE 0x0800 /* verbose on exit */
338 #define TDF_DEADLKTREAT 0x1000 /* special lockmgr deadlock treatment */
339 #define TDF_STOPREQ 0x2000 /* suspend_kproc */
340 #define TDF_WAKEREQ 0x4000 /* resume_kproc */
341 #define TDF_TIMEOUT 0x8000 /* tsleep timeout */
342 #define TDF_INTTHREAD 0x00010000 /* interrupt thread */
343 #define TDF_TSLEEP_DESCHEDULED 0x00020000 /* tsleep core deschedule */
344 #define TDF_BLOCKED 0x00040000 /* Thread is blocked */
345 #define TDF_PANICWARN 0x00080000 /* panic warning in switch */
346 #define TDF_BLOCKQ 0x00100000 /* on block queue */
347 #define TDF_UNUSED00200000 0x00200000
348 #define TDF_EXITING 0x00400000 /* thread exiting */
349 #define TDF_USINGFP 0x00800000 /* thread using fp coproc */
350 #define TDF_KERNELFP 0x01000000 /* kernel using fp coproc */
351 #define TDF_UNUSED02000000 0x02000000
352 #define TDF_CRYPTO 0x04000000 /* crypto thread */
353 #define TDF_MARKER 0x80000000 /* fairq marker 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 /* scheduler helper for userland sch */
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_EXITING 19 /* exiting thread */
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 */
383 * Scale is the approximate number of ticks for which we desire the
384 * entire gd_tdrunq to get service. With hz = 100 a scale of 8 is 80ms.
386 * Setting this value too small will result in inefficient switching
389 #define TDFAIRQ_SCALE 8
390 #define TDFAIRQ_MAX(gd) ((gd)->gd_fairq_total_pri * TDFAIRQ_SCALE)
392 #define LWKT_THREAD_STACK (UPAGES * PAGE_SIZE)
394 #define CACHE_NTHREADS 6
396 #define IN_CRITICAL_SECT(td) ((td)->td_critcount)
403 extern struct lwkt_token mp_token;
404 extern struct lwkt_token pmap_token;
405 extern struct lwkt_token dev_token;
406 extern struct lwkt_token vm_token;
407 extern struct lwkt_token vmspace_token;
408 extern struct lwkt_token kvm_token;
409 extern struct lwkt_token proc_token;
410 extern struct lwkt_token tty_token;
411 extern struct lwkt_token vnode_token;
412 extern struct lwkt_token vmobj_token;
417 extern void lwkt_init(void);
418 extern struct thread *lwkt_alloc_thread(struct thread *, int, int, int);
419 extern void lwkt_init_thread(struct thread *, void *, int, int,
420 struct globaldata *);
421 extern void lwkt_set_comm(thread_t, const char *, ...) __printflike(2, 3);
422 extern void lwkt_wait_free(struct thread *);
423 extern void lwkt_free_thread(struct thread *);
424 extern void lwkt_gdinit(struct globaldata *);
425 extern void lwkt_switch(void);
426 extern void lwkt_preempt(thread_t, int);
427 extern void lwkt_schedule(thread_t);
428 extern void lwkt_schedule_noresched(thread_t);
429 extern void lwkt_schedule_self(thread_t);
430 extern void lwkt_deschedule(thread_t);
431 extern void lwkt_deschedule_self(thread_t);
432 extern void lwkt_yield(void);
433 extern void lwkt_user_yield(void);
434 extern void lwkt_token_wait(void);
435 extern void lwkt_hold(thread_t);
436 extern void lwkt_rele(thread_t);
437 extern void lwkt_passive_release(thread_t);
438 extern void lwkt_maybe_splz(thread_t);
440 extern void lwkt_gettoken(lwkt_token_t);
441 extern void lwkt_gettoken_hard(lwkt_token_t);
442 extern int lwkt_trytoken(lwkt_token_t);
443 extern void lwkt_reltoken(lwkt_token_t);
444 extern void lwkt_reltoken_hard(lwkt_token_t);
445 extern int lwkt_cnttoken(lwkt_token_t, thread_t);
446 extern int lwkt_getalltokens(thread_t);
447 extern void lwkt_relalltokens(thread_t);
448 extern void lwkt_drain_token_requests(void);
449 extern void lwkt_token_init(lwkt_token_t, const char *);
450 extern void lwkt_token_uninit(lwkt_token_t);
452 extern void lwkt_token_pool_init(void);
453 extern lwkt_token_t lwkt_token_pool_lookup(void *);
454 extern lwkt_token_t lwkt_getpooltoken(void *);
455 extern void lwkt_relpooltoken(void *);
457 extern void lwkt_setpri(thread_t, int);
458 extern void lwkt_setpri_initial(thread_t, int);
459 extern void lwkt_setpri_self(int);
460 extern void lwkt_fairq_schedulerclock(thread_t td);
461 extern void lwkt_fairq_setpri_self(int pri);
462 extern int lwkt_fairq_push(int pri);
463 extern void lwkt_fairq_pop(int pri);
464 extern void lwkt_fairq_yield(void);
465 extern void lwkt_setcpu_self(struct globaldata *);
466 extern void lwkt_migratecpu(int);
470 extern void lwkt_giveaway(struct thread *);
471 extern void lwkt_acquire(struct thread *);
472 extern int lwkt_send_ipiq3(struct globaldata *, ipifunc3_t, void *, int);
473 extern int lwkt_send_ipiq3_passive(struct globaldata *, ipifunc3_t,
475 extern int lwkt_send_ipiq3_nowait(struct globaldata *, ipifunc3_t,
477 extern int lwkt_send_ipiq3_bycpu(int, ipifunc3_t, void *, int);
478 extern int lwkt_send_ipiq3_mask(cpumask_t, ipifunc3_t, void *, int);
479 extern void lwkt_wait_ipiq(struct globaldata *, int);
480 extern int lwkt_seq_ipiq(struct globaldata *);
481 extern void lwkt_process_ipiq(void);
482 extern void lwkt_process_ipiq_frame(struct intrframe *);
483 extern void lwkt_smp_stopped(void);
484 extern void lwkt_synchronize_ipiqs(const char *);
488 /* lwkt_cpusync_init() - inline function in sys/thread2.h */
489 extern void lwkt_cpusync_simple(cpumask_t, cpusync_func_t, void *);
490 extern void lwkt_cpusync_interlock(lwkt_cpusync_t);
491 extern void lwkt_cpusync_deinterlock(lwkt_cpusync_t);
493 extern void crit_panic(void) __dead2;
494 extern struct lwp *lwkt_preempted_proc(void);
496 extern int lwkt_create (void (*func)(void *), void *, struct thread **,
497 struct thread *, int, int,
498 const char *, ...) __printflike(7, 8);
499 extern void lwkt_exit (void) __dead2;
500 extern void lwkt_remove_tdallq (struct thread *);