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. Each 99 * thread has a stack of tokref's to keep track of acquired tokens. Multiple 100 * tokref's may reference the same token. 101 */ 102 103 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 const char *t_desc; /* Descriptive name */ 108 } lwkt_token; 109 110 #define LWKT_TOKEN_MPSAFE 0x0001 111 112 /* 113 * Static initialization for a lwkt_token. 114 * UP - Not MPSAFE (full MP lock will also be acquired) 115 * MP - Is MPSAFE (only the token will be acquired) 116 */ 117 #define LWKT_TOKEN_UP_INITIALIZER(name) \ 118 { \ 119 .t_ref = NULL, \ 120 .t_flags = 0, \ 121 .t_collisions = 0, \ 122 .t_desc = #name \ 123 } 124 125 #define LWKT_TOKEN_MP_INITIALIZER(name) \ 126 { \ 127 .t_ref = NULL, \ 128 .t_flags = LWKT_TOKEN_MPSAFE, \ 129 .t_collisions = 0, \ 130 .t_desc = #name \ 131 } 132 133 /* 134 * Assert that a particular token is held 135 */ 136 #define ASSERT_LWKT_TOKEN_HELD(tok) \ 137 KKASSERT((tok)->t_ref && (tok)->t_ref->tr_owner == curthread) 138 #define ASSERT_NO_TOKENS_HELD(td) \ 139 KKASSERT((td)->td_toks_stop == &td->toks_array[0]) 140 141 /* 142 * Assert that a particular token is held and we are in a hard 143 * code execution section (interrupt, ipi, or hard code section). 144 * Hard code sections are not allowed to block or potentially block. 145 * e.g. lwkt_gettoken() would only be ok if the token were already 146 * held. 147 */ 148 #define ASSERT_LWKT_TOKEN_HARD(tok) \ 149 do { \ 150 globaldata_t zgd __debugvar = mycpu; \ 151 KKASSERT((tok)->t_ref && \ 152 (tok)->t_ref->tr_owner == zgd->gd_curthread && \ 153 zgd->gd_intr_nesting_level > 0); \ 154 } while(0) 155 156 /* 157 * Assert that a particular token is held and we are in a normal 158 * critical section. Critical sections will not be preempted but 159 * can explicitly block (tsleep, lwkt_gettoken, etc). 160 */ 161 #define ASSERT_LWKT_TOKEN_CRIT(tok) \ 162 do { \ 163 globaldata_t zgd __debugvar = mycpu; \ 164 KKASSERT((tok)->t_ref && \ 165 (tok)->t_ref->tr_owner == zgd->gd_curthread && \ 166 zgd->gd_curthread->td_critcount > 0); \ 167 } while(0) 168 169 struct lwkt_tokref { 170 lwkt_token_t tr_tok; /* token in question */ 171 struct thread *tr_owner; /* me */ 172 intptr_t tr_flags; /* copy of t_flags */ 173 const void *tr_stallpc; /* stalled at pc */ 174 }; 175 176 #define MAXCPUFIFO 16 /* power of 2 */ 177 #define MAXCPUFIFO_MASK (MAXCPUFIFO - 1) 178 #define LWKT_MAXTOKENS 32 /* max tokens beneficially held by thread */ 179 180 /* 181 * Always cast to ipifunc_t when registering an ipi. The actual ipi function 182 * is called with both the data and an interrupt frame, but the ipi function 183 * that is registered might only declare a data argument. 184 */ 185 typedef void (*ipifunc1_t)(void *arg); 186 typedef void (*ipifunc2_t)(void *arg, int arg2); 187 typedef void (*ipifunc3_t)(void *arg, int arg2, struct intrframe *frame); 188 189 typedef struct lwkt_ipiq { 190 int ip_rindex; /* only written by target cpu */ 191 int ip_xindex; /* written by target, indicates completion */ 192 int ip_windex; /* only written by source cpu */ 193 ipifunc3_t ip_func[MAXCPUFIFO]; 194 void *ip_arg1[MAXCPUFIFO]; 195 int ip_arg2[MAXCPUFIFO]; 196 u_int ip_npoll; /* synchronization to avoid excess IPIs */ 197 } lwkt_ipiq; 198 199 /* 200 * CPU Synchronization structure. See lwkt_cpusync_start() and 201 * lwkt_cpusync_finish() for more information. 202 */ 203 typedef void (*cpusync_func_t)(lwkt_cpusync_t poll); 204 typedef void (*cpusync_func2_t)(void *data); 205 206 struct lwkt_cpusync { 207 cpusync_func_t cs_run_func; /* run (tandem w/ acquire) */ 208 cpusync_func_t cs_fin1_func; /* fin1 (synchronized) */ 209 cpusync_func2_t cs_fin2_func; /* fin2 (tandem w/ release) */ 210 void *cs_data; 211 int cs_maxcount; 212 volatile int cs_count; 213 cpumask_t cs_mask; 214 }; 215 216 /* 217 * The standard message and queue structure used for communications between 218 * cpus. Messages are typically queued via a machine-specific non-linked 219 * FIFO matrix allowing any cpu to send a message to any other cpu without 220 * blocking. 221 */ 222 typedef struct lwkt_cpu_msg { 223 void (*cm_func)(lwkt_cpu_msg_t msg); /* primary dispatch function */ 224 int cm_code; /* request code if applicable */ 225 int cm_cpu; /* reply to cpu */ 226 thread_t cm_originator; /* originating thread for wakeup */ 227 } lwkt_cpu_msg; 228 229 /* 230 * Thread structure. Note that ownership of a thread structure is special 231 * cased and there is no 'token'. A thread is always owned by the cpu 232 * represented by td_gd, any manipulation of the thread by some other cpu 233 * must be done through cpu_*msg() functions. e.g. you could request 234 * ownership of a thread that way, or hand a thread off to another cpu. 235 * 236 * NOTE: td_ucred is synchronized from the p_ucred on user->kernel syscall, 237 * trap, and AST/signal transitions to provide a stable ucred for 238 * (primarily) system calls. This field will be NULL for pure kernel 239 * threads. 240 */ 241 struct md_intr_info; 242 struct caps_kinfo; 243 244 struct thread { 245 TAILQ_ENTRY(thread) td_threadq; 246 TAILQ_ENTRY(thread) td_allq; 247 TAILQ_ENTRY(thread) td_sleepq; 248 lwkt_port td_msgport; /* built-in message port for replies */ 249 struct lwp *td_lwp; /* (optional) associated lwp */ 250 struct proc *td_proc; /* (optional) associated process */ 251 struct pcb *td_pcb; /* points to pcb and top of kstack */ 252 struct globaldata *td_gd; /* associated with this cpu */ 253 const char *td_wmesg; /* string name for blockage */ 254 const volatile void *td_wchan; /* waiting on channel */ 255 int td_pri; /* 0-31, 31=highest priority (note 1) */ 256 int td_critcount; /* critical section priority */ 257 int td_flags; /* TDF flags */ 258 int td_wdomain; /* domain for wchan address (typ 0) */ 259 void (*td_preemptable)(struct thread *td, int critcount); 260 void (*td_release)(struct thread *td); 261 char *td_kstack; /* kernel stack */ 262 int td_kstack_size; /* size of kernel stack */ 263 char *td_sp; /* kernel stack pointer for LWKT restore */ 264 void (*td_switch)(struct thread *ntd); 265 __uint64_t td_uticks; /* Statclock hits in user mode (uS) */ 266 __uint64_t td_sticks; /* Statclock hits in system mode (uS) */ 267 __uint64_t td_iticks; /* Statclock hits processing intr (uS) */ 268 int td_locks; /* lockmgr lock debugging */ 269 void *td_dsched_priv1; /* priv data for I/O schedulers */ 270 int td_refs; /* hold position in gd_tdallq / hold free */ 271 int td_nest_count; /* prevent splz nesting */ 272 #ifdef SMP 273 int td_mpcount; /* MP lock held (count) */ 274 int td_xpcount; /* MP lock held inherited (count) */ 275 int td_cscount; /* cpu synchronization master */ 276 int td_unused02[4]; /* for future fields */ 277 #else 278 int td_mpcount_unused; /* filler so size matches */ 279 int td_xpcount_unused; 280 int td_cscount_unused; 281 int td_unused02[4]; 282 #endif 283 int td_unused03[4]; /* for future fields */ 284 struct iosched_data td_iosdata; /* Dynamic I/O scheduling data */ 285 struct timeval td_start; /* start time for a thread/process */ 286 char td_comm[MAXCOMLEN+1]; /* typ 16+1 bytes */ 287 struct thread *td_preempted; /* we preempted this thread */ 288 struct ucred *td_ucred; /* synchronized from p_ucred */ 289 struct caps_kinfo *td_caps; /* list of client and server registrations */ 290 lwkt_tokref_t td_toks_stop; 291 struct lwkt_tokref td_toks_array[LWKT_MAXTOKENS]; 292 int td_fairq_lticks; /* fairq wakeup accumulator reset */ 293 int td_fairq_accum; /* fairq priority accumulator */ 294 const void *td_mplock_stallpc; /* last mplock stall address */ 295 #ifdef DEBUG_CRIT_SECTIONS 296 #define CRIT_DEBUG_ARRAY_SIZE 32 297 #define CRIT_DEBUG_ARRAY_MASK (CRIT_DEBUG_ARRAY_SIZE - 1) 298 const char *td_crit_debug_array[CRIT_DEBUG_ARRAY_SIZE]; 299 int td_crit_debug_index; 300 int td_in_crit_report; 301 #endif 302 struct md_thread td_mach; 303 }; 304 305 #define td_toks_base td_toks_array[0] 306 #define td_toks_end td_toks_array[LWKT_MAXTOKENS] 307 308 #define TD_TOKS_HELD(td) ((td)->td_toks_stop != &(td)->td_toks_base) 309 #define TD_TOKS_NOT_HELD(td) ((td)->td_toks_stop == &(td)->td_toks_base) 310 311 /* 312 * Thread flags. Note that TDF_RUNNING is cleared on the old thread after 313 * we switch to the new one, which is necessary because LWKTs don't need 314 * to hold the BGL. This flag is used by the exit code and the managed 315 * thread migration code. Note in addition that preemption will cause 316 * TDF_RUNNING to be cleared temporarily, so any code checking TDF_RUNNING 317 * must also check TDF_PREEMPT_LOCK. 318 * 319 * LWKT threads stay on their (per-cpu) run queue while running, not to 320 * be confused with user processes which are removed from the user scheduling 321 * run queue while actually running. 322 * 323 * td_threadq can represent the thread on one of three queues... the LWKT 324 * run queue, a tsleep queue, or an lwkt blocking queue. The LWKT subsystem 325 * does not allow a thread to be scheduled if it already resides on some 326 * queue. 327 */ 328 #define TDF_RUNNING 0x0001 /* thread still active */ 329 #define TDF_RUNQ 0x0002 /* on an LWKT run queue */ 330 #define TDF_PREEMPT_LOCK 0x0004 /* I have been preempted */ 331 #define TDF_PREEMPT_DONE 0x0008 /* acknowledge preemption complete */ 332 #define TDF_IDLE_NOHLT 0x0010 /* we need to spin */ 333 #define TDF_MIGRATING 0x0020 /* thread is being migrated */ 334 #define TDF_SINTR 0x0040 /* interruptability hint for 'ps' */ 335 #define TDF_TSLEEPQ 0x0080 /* on a tsleep wait queue */ 336 337 #define TDF_SYSTHREAD 0x0100 /* allocations may use reserve */ 338 #define TDF_ALLOCATED_THREAD 0x0200 /* objcache allocated thread */ 339 #define TDF_ALLOCATED_STACK 0x0400 /* objcache allocated stack */ 340 #define TDF_VERBOSE 0x0800 /* verbose on exit */ 341 #define TDF_DEADLKTREAT 0x1000 /* special lockmgr deadlock treatment */ 342 #define TDF_STOPREQ 0x2000 /* suspend_kproc */ 343 #define TDF_WAKEREQ 0x4000 /* resume_kproc */ 344 #define TDF_TIMEOUT 0x8000 /* tsleep timeout */ 345 #define TDF_INTTHREAD 0x00010000 /* interrupt thread */ 346 #define TDF_TSLEEP_DESCHEDULED 0x00020000 /* tsleep core deschedule */ 347 #define TDF_BLOCKED 0x00040000 /* Thread is blocked */ 348 #define TDF_PANICWARN 0x00080000 /* panic warning in switch */ 349 #define TDF_BLOCKQ 0x00100000 /* on block queue */ 350 #define TDF_UNUSED00200000 0x00200000 351 #define TDF_EXITING 0x00400000 /* thread exiting */ 352 #define TDF_USINGFP 0x00800000 /* thread using fp coproc */ 353 #define TDF_KERNELFP 0x01000000 /* kernel using fp coproc */ 354 #define TDF_UNUSED02000000 0x02000000 355 #define TDF_CRYPTO 0x04000000 /* crypto thread */ 356 #define TDF_MARKER 0x80000000 /* fairq marker thread */ 357 358 /* 359 * Thread priorities. Typically only one thread from any given 360 * user process scheduling queue is on the LWKT run queue at a time. 361 * Remember that there is one LWKT run queue per cpu. 362 * 363 * Critical sections are handled by bumping td_pri above TDPRI_MAX, which 364 * causes interrupts to be masked as they occur. When this occurs a 365 * rollup flag will be set in mycpu->gd_reqflags. 366 */ 367 #define TDPRI_IDLE_THREAD 0 /* the idle thread */ 368 #define TDPRI_IDLE_WORK 1 /* idle work (page zero, etc) */ 369 #define TDPRI_USER_SCHEDULER 2 /* user scheduler helper */ 370 #define TDPRI_USER_IDLE 4 /* user scheduler idle */ 371 #define TDPRI_USER_NORM 6 /* user scheduler normal */ 372 #define TDPRI_USER_REAL 8 /* user scheduler real time */ 373 #define TDPRI_KERN_LPSCHED 9 /* scheduler helper for userland sch */ 374 #define TDPRI_KERN_USER 10 /* kernel / block in syscall */ 375 #define TDPRI_KERN_DAEMON 12 /* kernel daemon (pageout, etc) */ 376 #define TDPRI_SOFT_NORM 14 /* kernel / normal */ 377 #define TDPRI_SOFT_TIMER 16 /* kernel / timer */ 378 #define TDPRI_EXITING 19 /* exiting thread */ 379 #define TDPRI_INT_SUPPORT 20 /* kernel / high priority support */ 380 #define TDPRI_INT_LOW 27 /* low priority interrupt */ 381 #define TDPRI_INT_MED 28 /* medium priority interrupt */ 382 #define TDPRI_INT_HIGH 29 /* high priority interrupt */ 383 #define TDPRI_MAX 31 384 385 /* 386 * Scale is the approximate number of ticks for which we desire the 387 * entire gd_tdrunq to get service. With hz = 100 a scale of 8 is 80ms. 388 * 389 * Setting this value too small will result in inefficient switching 390 * rates. 391 */ 392 #define TDFAIRQ_SCALE 8 393 #define TDFAIRQ_MAX(gd) ((gd)->gd_fairq_total_pri * TDFAIRQ_SCALE) 394 395 #define LWKT_THREAD_STACK (UPAGES * PAGE_SIZE) 396 397 #define CACHE_NTHREADS 6 398 399 #define IN_CRITICAL_SECT(td) ((td)->td_critcount) 400 401 #ifdef _KERNEL 402 403 /* 404 * Global tokens 405 */ 406 extern struct lwkt_token pmap_token; 407 extern struct lwkt_token dev_token; 408 extern struct lwkt_token vm_token; 409 extern struct lwkt_token vmspace_token; 410 extern struct lwkt_token kvm_token; 411 extern struct lwkt_token proc_token; 412 extern struct lwkt_token tty_token; 413 extern struct lwkt_token vnode_token; 414 extern struct lwkt_token vmobj_token; 415 416 /* 417 * Procedures 418 */ 419 extern void lwkt_init(void); 420 extern struct thread *lwkt_alloc_thread(struct thread *, int, int, int); 421 extern void lwkt_init_thread(struct thread *, void *, int, int, 422 struct globaldata *); 423 extern void lwkt_set_comm(thread_t, const char *, ...) __printflike(2, 3); 424 extern void lwkt_wait_free(struct thread *); 425 extern void lwkt_free_thread(struct thread *); 426 extern void lwkt_gdinit(struct globaldata *); 427 extern void lwkt_switch(void); 428 extern void lwkt_preempt(thread_t, int); 429 extern void lwkt_schedule(thread_t); 430 extern void lwkt_schedule_noresched(thread_t); 431 extern void lwkt_schedule_self(thread_t); 432 extern void lwkt_deschedule(thread_t); 433 extern void lwkt_deschedule_self(thread_t); 434 extern void lwkt_yield(void); 435 extern void lwkt_user_yield(void); 436 extern void lwkt_token_wait(void); 437 extern void lwkt_hold(thread_t); 438 extern void lwkt_rele(thread_t); 439 extern void lwkt_passive_release(thread_t); 440 extern void lwkt_maybe_splz(thread_t); 441 442 extern void lwkt_gettoken(lwkt_token_t); 443 extern void lwkt_gettoken_hard(lwkt_token_t); 444 extern int lwkt_trytoken(lwkt_token_t); 445 extern void lwkt_reltoken(lwkt_token_t); 446 extern void lwkt_reltoken_hard(lwkt_token_t); 447 extern int lwkt_getalltokens(thread_t, const char **, const void **); 448 extern void lwkt_relalltokens(thread_t); 449 extern void lwkt_drain_token_requests(void); 450 extern void lwkt_token_init(lwkt_token_t, int, const char *); 451 extern void lwkt_token_uninit(lwkt_token_t); 452 453 extern void lwkt_token_pool_init(void); 454 extern lwkt_token_t lwkt_token_pool_lookup(void *); 455 extern lwkt_token_t lwkt_getpooltoken(void *); 456 extern void lwkt_relpooltoken(void *); 457 458 extern void lwkt_setpri(thread_t, int); 459 extern void lwkt_setpri_initial(thread_t, int); 460 extern void lwkt_setpri_self(int); 461 extern void lwkt_fairq_schedulerclock(thread_t td); 462 extern void lwkt_fairq_setpri_self(int pri); 463 extern int lwkt_fairq_push(int pri); 464 extern void lwkt_fairq_pop(int pri); 465 extern void lwkt_fairq_yield(void); 466 extern void lwkt_setcpu_self(struct globaldata *); 467 extern void lwkt_migratecpu(int); 468 469 #ifdef SMP 470 471 extern void lwkt_giveaway(struct thread *); 472 extern void lwkt_acquire(struct thread *); 473 extern int lwkt_send_ipiq3(struct globaldata *, ipifunc3_t, void *, int); 474 extern int lwkt_send_ipiq3_passive(struct globaldata *, ipifunc3_t, 475 void *, int); 476 extern int lwkt_send_ipiq3_nowait(struct globaldata *, ipifunc3_t, 477 void *, int); 478 extern int lwkt_send_ipiq3_bycpu(int, ipifunc3_t, void *, int); 479 extern int lwkt_send_ipiq3_mask(cpumask_t, ipifunc3_t, void *, int); 480 extern void lwkt_wait_ipiq(struct globaldata *, int); 481 extern int lwkt_seq_ipiq(struct globaldata *); 482 extern void lwkt_process_ipiq(void); 483 extern void lwkt_process_ipiq_frame(struct intrframe *); 484 extern void lwkt_smp_stopped(void); 485 extern void lwkt_synchronize_ipiqs(const char *); 486 487 #endif /* SMP */ 488 489 extern void lwkt_cpusync_simple(cpumask_t, cpusync_func_t, void *); 490 extern void lwkt_cpusync_fastdata(cpumask_t, cpusync_func2_t, void *); 491 extern void lwkt_cpusync_start(cpumask_t, lwkt_cpusync_t); 492 extern void lwkt_cpusync_add(cpumask_t, lwkt_cpusync_t); 493 extern void lwkt_cpusync_finish(lwkt_cpusync_t); 494 495 extern void crit_panic(void) __dead2; 496 extern struct lwp *lwkt_preempted_proc(void); 497 498 extern int lwkt_create (void (*func)(void *), void *, struct thread **, 499 struct thread *, int, int, 500 const char *, ...) __printflike(7, 8); 501 extern void lwkt_exit (void) __dead2; 502 extern void lwkt_remove_tdallq (struct thread *); 503 504 #endif 505 506 #endif 507 508