xref: /dragonfly/sys/sys/thread.h (revision 872b5d49)
1 /*
2  * SYS/THREAD.H
3  *
4  *	Implements the architecture independant portion of the LWKT
5  *	subsystem.
6  *
7  * Types which must already be defined when this header is included by
8  * userland:	struct md_thread
9  *
10  * $DragonFly: src/sys/sys/thread.h,v 1.97 2008/09/20 04:31:02 sephe Exp $
11  */
12 
13 #ifndef _SYS_THREAD_H_
14 #define _SYS_THREAD_H_
15 
16 #ifndef _SYS_STDINT_H_
17 #include <sys/stdint.h>		/* __int types */
18 #endif
19 #ifndef _SYS_PARAM_H_
20 #include <sys/param.h>		/* MAXCOMLEN */
21 #endif
22 #ifndef _SYS_QUEUE_H_
23 #include <sys/queue.h>		/* TAILQ_* macros */
24 #endif
25 #ifndef _SYS_MSGPORT_H_
26 #include <sys/msgport.h>	/* lwkt_port */
27 #endif
28 #ifndef _SYS_TIME_H_
29 #include <sys/time.h>   	/* struct timeval */
30 #endif
31 #ifndef _SYS_SPINLOCK_H_
32 #include <sys/spinlock.h>
33 #endif
34 #ifndef _SYS_IOSCHED_H_
35 #include <sys/iosched.h>
36 #endif
37 #ifndef _MACHINE_THREAD_H_
38 #include <machine/thread.h>
39 #endif
40 
41 struct globaldata;
42 struct lwp;
43 struct proc;
44 struct thread;
45 struct lwkt_queue;
46 struct lwkt_token;
47 struct lwkt_tokref;
48 struct lwkt_ipiq;
49 struct lwkt_cpu_msg;
50 struct lwkt_cpu_port;
51 struct lwkt_msg;
52 struct lwkt_port;
53 struct lwkt_cpusync;
54 union sysunion;
55 
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;
64 
65 typedef TAILQ_HEAD(lwkt_queue, thread) lwkt_queue;
66 
67 /*
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.
73  */
74 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
75 #ifndef _MACHINE_THREAD_H_
76 #include <machine/thread.h>		/* md_thread */
77 #endif
78 #ifndef _MACHINE_FRAME_H_
79 #include <machine/frame.h>
80 #endif
81 #else
82 struct intrframe;
83 #endif
84 
85 /*
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.
90  *
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.
97  *
98  * Tokens are managed through a helper reference structure, lwkt_tokref,
99  * which is typically declared on the caller's stack.  Multiple tokref's
100  * may reference the same token.
101  *
102  * It is possible to detect that your token was temporarily lost via
103  * lwkt_token_is_stale(), which uses the t_lastowner field.  This field
104  * does NOT necessarily represent the current owner and can become stale
105  * (not point to a valid structure).  It is used solely to detect
106  * whether the token was temporarily lost to another thread.  The lost
107  * state is cleared by the function.
108  */
109 
110 typedef struct lwkt_token {
111 #ifdef SMP
112     struct spinlock	t_spinlock;	/* Controls access */
113 #else
114     struct spinlock	t_unused01;
115 #endif
116     struct thread	*t_owner;	/* The current owner of the token */
117     int			t_count;	/* Per-thread count */
118     struct thread       *t_lastowner;	/* Last owner that acquired token */
119 } lwkt_token;
120 
121 #ifdef SMP
122 #define LWKT_TOKEN_INITIALIZER(head) \
123 { \
124 	.t_spinlock = SPINLOCK_INITIALIZER(head.t_spinlock), \
125 	.t_owner = NULL, \
126 	.t_lastowner = NULL, \
127 	.t_count = 0 \
128 }
129 #else
130 #define LWKT_TOKEN_INITIALIZER(head) \
131 { \
132 	.t_owner = NULL, \
133 	.t_lastowner = NULL, \
134 	.t_count = 0 \
135 }
136 #endif
137 
138 #define ASSERT_LWKT_TOKEN_HELD(token) \
139 	KKASSERT((token)->t_owner == curthread)
140 
141 typedef struct lwkt_tokref {
142     lwkt_token_t	tr_tok;		/* token in question */
143     lwkt_tokref_t	tr_next;	/* linked list */
144     int			tr_state;	/* 0 = don't have, 1 = have */
145 } lwkt_tokref;
146 
147 #define LWKT_TOKREF_INIT(tok)		\
148 			{ tok, NULL, 0 }
149 #define LWKT_TOKREF_DECLARE(name, tok)	\
150 			lwkt_tokref name = LWKT_TOKREF_INIT(tok)
151 
152 #define MAXCPUFIFO      16	/* power of 2 */
153 #define MAXCPUFIFO_MASK	(MAXCPUFIFO - 1)
154 #define LWKT_MAXTOKENS	16	/* max tokens beneficially held by thread */
155 
156 /*
157  * Always cast to ipifunc_t when registering an ipi.  The actual ipi function
158  * is called with both the data and an interrupt frame, but the ipi function
159  * that is registered might only declare a data argument.
160  */
161 typedef void (*ipifunc1_t)(void *arg);
162 typedef void (*ipifunc2_t)(void *arg, int arg2);
163 typedef void (*ipifunc3_t)(void *arg, int arg2, struct intrframe *frame);
164 
165 typedef struct lwkt_ipiq {
166     int		ip_rindex;      /* only written by target cpu */
167     int		ip_xindex;      /* written by target, indicates completion */
168     int		ip_windex;      /* only written by source cpu */
169     ipifunc3_t	ip_func[MAXCPUFIFO];
170     void	*ip_arg1[MAXCPUFIFO];
171     int		ip_arg2[MAXCPUFIFO];
172     u_int	ip_npoll;	/* synchronization to avoid excess IPIs */
173 } lwkt_ipiq;
174 
175 /*
176  * CPU Synchronization structure.  See lwkt_cpusync_start() and
177  * lwkt_cpusync_finish() for more information.
178  */
179 typedef void (*cpusync_func_t)(lwkt_cpusync_t poll);
180 typedef void (*cpusync_func2_t)(void *data);
181 
182 struct lwkt_cpusync {
183     cpusync_func_t cs_run_func;		/* run (tandem w/ acquire) */
184     cpusync_func_t cs_fin1_func;	/* fin1 (synchronized) */
185     cpusync_func2_t cs_fin2_func;	/* fin2 (tandem w/ release) */
186     void	*cs_data;
187     int		cs_maxcount;
188     volatile int cs_count;
189     cpumask_t	cs_mask;
190 };
191 
192 /*
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.
197  */
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;
204 
205 /*
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.
211  *
212  * NOTE: td_pri is bumped by TDPRI_CRIT when entering a critical section,
213  * but this does not effect how the thread is scheduled by LWKT.
214  */
215 struct md_intr_info;
216 struct caps_kinfo;
217 
218 struct thread {
219     TAILQ_ENTRY(thread) td_threadq;
220     TAILQ_ENTRY(thread) td_allq;
221     TAILQ_ENTRY(thread) td_sleepq;
222     lwkt_port	td_msgport;	/* built-in message port for replies */
223     struct lwp	*td_lwp;	/* (optional) associated lwp */
224     struct proc	*td_proc;	/* (optional) associated process */
225     struct pcb	*td_pcb;	/* points to pcb and top of kstack */
226     struct globaldata *td_gd;	/* associated with this cpu */
227     const char	*td_wmesg;	/* string name for blockage */
228     void	*td_wchan;	/* waiting on channel */
229     int		td_pri;		/* 0-31, 31=highest priority (note 1) */
230     int		td_flags;	/* TDF flags */
231     int		td_wdomain;	/* domain for wchan address (typ 0) */
232     void	(*td_preemptable)(struct thread *td, int critpri);
233     void	(*td_release)(struct thread *td);
234     char	*td_kstack;	/* kernel stack */
235     int		td_kstack_size;	/* size of kernel stack */
236     char	*td_sp;		/* kernel stack pointer for LWKT restore */
237     void	(*td_switch)(struct thread *ntd);
238     __uint64_t	td_uticks;	/* Statclock hits in user mode (uS) */
239     __uint64_t	td_sticks;      /* Statclock hits in system mode (uS) */
240     __uint64_t	td_iticks;	/* Statclock hits processing intr (uS) */
241     int		td_locks;	/* lockmgr lock debugging */
242     int		td_unused01;
243     int		td_refs;	/* hold position in gd_tdallq / hold free */
244     int		td_nest_count;	/* prevent splz nesting */
245 #ifdef SMP
246     int		td_mpcount;	/* MP lock held (count) */
247     int		td_cscount;	/* cpu synchronization master */
248 #else
249     int		td_mpcount_unused;	/* filler so size matches */
250     int		td_cscount_unused;
251 #endif
252     struct iosched_data td_iosdata;	/* Dynamic I/O scheduling data */
253     struct timeval td_start;	/* start time for a thread/process */
254     char	td_comm[MAXCOMLEN+1]; /* typ 16+1 bytes */
255     struct thread *td_preempted; /* we preempted this thread */
256     struct caps_kinfo *td_caps;	/* list of client and server registrations */
257     lwkt_tokref_t td_toks;	/* tokens beneficially held */
258 #ifdef DEBUG_CRIT_SECTIONS
259 #define CRIT_DEBUG_ARRAY_SIZE   32
260 #define CRIT_DEBUG_ARRAY_MASK   (CRIT_DEBUG_ARRAY_SIZE - 1)
261     const char	*td_crit_debug_array[CRIT_DEBUG_ARRAY_SIZE];
262     int		td_crit_debug_index;
263     int		td_in_crit_report;
264 #endif
265     struct md_thread td_mach;
266 };
267 
268 /*
269  * Thread flags.  Note that TDF_RUNNING is cleared on the old thread after
270  * we switch to the new one, which is necessary because LWKTs don't need
271  * to hold the BGL.  This flag is used by the exit code and the managed
272  * thread migration code.  Note in addition that preemption will cause
273  * TDF_RUNNING to be cleared temporarily, so any code checking TDF_RUNNING
274  * must also check TDF_PREEMPT_LOCK.
275  *
276  * LWKT threads stay on their (per-cpu) run queue while running, not to
277  * be confused with user processes which are removed from the user scheduling
278  * run queue while actually running.
279  *
280  * td_threadq can represent the thread on one of three queues... the LWKT
281  * run queue, a tsleep queue, or an lwkt blocking queue.  The LWKT subsystem
282  * does not allow a thread to be scheduled if it already resides on some
283  * queue.
284  */
285 #define TDF_RUNNING		0x0001	/* thread still active */
286 #define TDF_RUNQ		0x0002	/* on an LWKT run queue */
287 #define TDF_PREEMPT_LOCK	0x0004	/* I have been preempted */
288 #define TDF_PREEMPT_DONE	0x0008	/* acknowledge preemption complete */
289 #define TDF_IDLE_NOHLT		0x0010	/* we need to spin */
290 #define TDF_MIGRATING		0x0020	/* thread is being migrated */
291 #define TDF_SINTR		0x0040	/* interruptability hint for 'ps' */
292 #define TDF_TSLEEPQ		0x0080	/* on a tsleep wait queue */
293 
294 #define TDF_SYSTHREAD		0x0100	/* allocations may use reserve */
295 #define TDF_ALLOCATED_THREAD	0x0200	/* objcache allocated thread */
296 #define TDF_ALLOCATED_STACK	0x0400	/* objcache allocated stack */
297 #define TDF_VERBOSE		0x0800	/* verbose on exit */
298 #define TDF_DEADLKTREAT		0x1000	/* special lockmgr deadlock treatment */
299 #define TDF_STOPREQ		0x2000	/* suspend_kproc */
300 #define TDF_WAKEREQ		0x4000	/* resume_kproc */
301 #define TDF_TIMEOUT		0x8000	/* tsleep timeout */
302 #define TDF_INTTHREAD		0x00010000	/* interrupt thread */
303 #define TDF_TSLEEP_DESCHEDULED	0x00020000	/* tsleep core deschedule */
304 #define TDF_BLOCKED		0x00040000	/* Thread is blocked */
305 #define TDF_PANICWARN		0x00080000	/* panic warning in switch */
306 #define TDF_BLOCKQ		0x00100000	/* on block queue */
307 #define TDF_MPSAFE		0x00200000	/* (thread creation) */
308 #define TDF_EXITING		0x00400000	/* thread exiting */
309 #define TDF_USINGFP		0x00800000	/* thread using fp coproc */
310 #define TDF_KERNELFP		0x01000000	/* kernel using fp coproc */
311 #define TDF_NETWORK		0x02000000	/* network proto thread */
312 
313 /*
314  * Thread priorities.  Typically only one thread from any given
315  * user process scheduling queue is on the LWKT run queue at a time.
316  * Remember that there is one LWKT run queue per cpu.
317  *
318  * Critical sections are handled by bumping td_pri above TDPRI_MAX, which
319  * causes interrupts to be masked as they occur.  When this occurs a
320  * rollup flag will be set in mycpu->gd_reqflags.
321  */
322 #define TDPRI_IDLE_THREAD	0	/* the idle thread */
323 #define TDPRI_USER_SCHEDULER	2	/* user scheduler helper */
324 #define TDPRI_USER_IDLE		4	/* user scheduler idle */
325 #define TDPRI_USER_NORM		6	/* user scheduler normal */
326 #define TDPRI_USER_REAL		8	/* user scheduler real time */
327 #define TDPRI_KERN_LPSCHED	9	/* scheduler helper for userland sch */
328 #define TDPRI_KERN_USER		10	/* kernel / block in syscall */
329 #define TDPRI_KERN_DAEMON	12	/* kernel daemon (pageout, etc) */
330 #define TDPRI_SOFT_NORM		14	/* kernel / normal */
331 #define TDPRI_SOFT_TIMER	16	/* kernel / timer */
332 #define TDPRI_EXITING		19	/* exiting thread */
333 #define TDPRI_INT_SUPPORT	20	/* kernel / high priority support */
334 #define TDPRI_INT_LOW		27	/* low priority interrupt */
335 #define TDPRI_INT_MED		28	/* medium priority interrupt */
336 #define TDPRI_INT_HIGH		29	/* high priority interrupt */
337 #define TDPRI_MAX		31
338 
339 #define TDPRI_MASK		31
340 #define TDPRI_CRIT		32	/* high bits of td_pri used for crit */
341 
342 #ifdef _KERNEL
343 #define LWKT_THREAD_STACK	(UPAGES * PAGE_SIZE)
344 #endif
345 
346 #define CACHE_NTHREADS		6
347 
348 #define IN_CRITICAL_SECT(td)	((td)->td_pri >= TDPRI_CRIT)
349 
350 extern void lwkt_init(void);
351 extern struct thread *lwkt_alloc_thread(struct thread *, int, int, int);
352 extern void lwkt_init_thread(struct thread *, void *, int, int,
353 			     struct globaldata *);
354 extern void lwkt_set_comm(thread_t, const char *, ...);
355 extern void lwkt_wait_free(struct thread *);
356 extern void lwkt_free_thread(struct thread *);
357 extern void lwkt_gdinit(struct globaldata *);
358 extern void lwkt_switch(void);
359 extern void lwkt_preempt(thread_t, int);
360 extern void lwkt_schedule(thread_t);
361 extern void lwkt_schedule_noresched(thread_t);
362 extern void lwkt_schedule_self(thread_t);
363 extern void lwkt_deschedule(thread_t);
364 extern void lwkt_deschedule_self(thread_t);
365 extern void lwkt_yield(void);
366 extern void lwkt_user_yield(void);
367 extern void lwkt_token_wait(void);
368 extern void lwkt_hold(thread_t);
369 extern void lwkt_rele(thread_t);
370 extern void lwkt_passive_release(thread_t);
371 
372 extern void lwkt_gettoken(lwkt_tokref_t, lwkt_token_t);
373 extern int lwkt_trytoken(lwkt_tokref_t, lwkt_token_t);
374 extern void lwkt_gettokref(lwkt_tokref_t);
375 extern int  lwkt_trytokref(lwkt_tokref_t);
376 extern void lwkt_reltoken(lwkt_tokref_t);
377 extern int  lwkt_getalltokens(thread_t);
378 extern void lwkt_relalltokens(thread_t);
379 extern void lwkt_drain_token_requests(void);
380 extern void lwkt_token_init(lwkt_token_t);
381 extern void lwkt_token_uninit(lwkt_token_t);
382 extern int  lwkt_token_is_stale(lwkt_tokref_t);
383 
384 extern void lwkt_token_pool_init(void);
385 extern lwkt_token_t lwkt_token_pool_get(void *);
386 
387 extern void lwkt_setpri(thread_t, int);
388 extern void lwkt_setpri_initial(thread_t, int);
389 extern void lwkt_setpri_self(int);
390 extern int lwkt_check_resched(thread_t);
391 extern void lwkt_setcpu_self(struct globaldata *);
392 extern void lwkt_migratecpu(int);
393 
394 #ifdef SMP
395 
396 extern void lwkt_giveaway(struct thread *);
397 extern void lwkt_acquire(struct thread *);
398 extern int  lwkt_send_ipiq3(struct globaldata *, ipifunc3_t, void *, int);
399 extern int  lwkt_send_ipiq3_passive(struct globaldata *, ipifunc3_t,
400 				    void *, int);
401 extern int  lwkt_send_ipiq3_nowait(struct globaldata *, ipifunc3_t,
402 				   void *, int);
403 extern int  lwkt_send_ipiq3_bycpu(int, ipifunc3_t, void *, int);
404 extern int  lwkt_send_ipiq3_mask(cpumask_t, ipifunc3_t, void *, int);
405 extern void lwkt_wait_ipiq(struct globaldata *, int);
406 extern int  lwkt_seq_ipiq(struct globaldata *);
407 extern void lwkt_process_ipiq(void);
408 #ifdef _KERNEL
409 extern void lwkt_process_ipiq_frame(struct intrframe *);
410 #endif
411 extern void lwkt_smp_stopped(void);
412 extern void lwkt_synchronize_ipiqs(const char *);
413 
414 #endif /* SMP */
415 
416 extern void lwkt_cpusync_simple(cpumask_t, cpusync_func_t, void *);
417 extern void lwkt_cpusync_fastdata(cpumask_t, cpusync_func2_t, void *);
418 extern void lwkt_cpusync_start(cpumask_t, lwkt_cpusync_t);
419 extern void lwkt_cpusync_add(cpumask_t, lwkt_cpusync_t);
420 extern void lwkt_cpusync_finish(lwkt_cpusync_t);
421 
422 extern void crit_panic(void);
423 extern struct lwp *lwkt_preempted_proc(void);
424 
425 extern int  lwkt_create (void (*func)(void *), void *, struct thread **,
426 		         struct thread *, int, int, const char *, ...);
427 extern void lwkt_exit (void) __dead2;
428 extern void lwkt_remove_tdallq (struct thread *);
429 extern void lwkt_mp_lock_contested(void);
430 extern void lwkt_mp_lock_uncontested(void);
431 
432 #endif
433 
434