aboutsummaryrefslogtreecommitdiffstats
path: root/litmus
diff options
context:
space:
mode:
authorGlenn Elliott <gelliott@cs.unc.edu>2012-12-11 22:01:01 -0500
committerGlenn Elliott <gelliott@cs.unc.edu>2012-12-12 14:14:41 -0500
commitc8483ef0959672310bf4ebb72e1a308b00543f74 (patch)
tree9cb306009b01c5226178f69172738026431d37f2 /litmus
parentfbd9574e298157b54c38f82f536e5cea8f766dff (diff)
make klmirqd work like aux tasks. checkpoint.
this code is untested!
Diffstat (limited to 'litmus')
-rw-r--r--litmus/Kconfig20
-rw-r--r--litmus/aux_tasks.c40
-rw-r--r--litmus/edf_common.c60
-rw-r--r--litmus/ikglp_lock.c12
-rw-r--r--litmus/jobs.c2
-rw-r--r--litmus/kfmlp_lock.c11
-rw-r--r--litmus/litmus.c48
-rw-r--r--litmus/litmus_softirq.c1460
-rw-r--r--litmus/locking.c56
-rw-r--r--litmus/nvidia_info.c743
-rw-r--r--litmus/sched_cedf.c165
-rw-r--r--litmus/sched_gsn_edf.c136
-rw-r--r--litmus/sched_plugin.c20
13 files changed, 1301 insertions, 1472 deletions
diff --git a/litmus/Kconfig b/litmus/Kconfig
index f2434b87239b..9aeae659ae32 100644
--- a/litmus/Kconfig
+++ b/litmus/Kconfig
@@ -382,7 +382,7 @@ menu "Interrupt Handling"
382choice 382choice
383 prompt "Scheduling of interrupt bottom-halves in Litmus." 383 prompt "Scheduling of interrupt bottom-halves in Litmus."
384 default LITMUS_SOFTIRQD_NONE 384 default LITMUS_SOFTIRQD_NONE
385 depends on LITMUS_LOCKING && !LITMUS_THREAD_ALL_SOFTIRQ 385 depends on LITMUS_LOCKING
386 help 386 help
387 Schedule tasklets with known priorities in Litmus. 387 Schedule tasklets with known priorities in Litmus.
388 388
@@ -398,7 +398,7 @@ config LITMUS_SOFTIRQD
398 specifically dispatched to these workers. (Softirqs for 398 specifically dispatched to these workers. (Softirqs for
399 Litmus tasks are not magically redirected to klmirqd.) 399 Litmus tasks are not magically redirected to klmirqd.)
400 400
401 G-EDF/RM, C-EDF/RM ONLY for now! 401 G-EDF, C-EDF ONLY for now!
402 402
403 403
404config LITMUS_PAI_SOFTIRQD 404config LITMUS_PAI_SOFTIRQD
@@ -409,19 +409,11 @@ config LITMUS_PAI_SOFTIRQD
409 at the cost of non-preemptive durations of bottom half 409 at the cost of non-preemptive durations of bottom half
410 processing. 410 processing.
411 411
412 G-EDF/RM, C-EDF/RM ONLY for now! 412 G-EDF, C-EDF ONLY for now!
413 413
414endchoice 414endchoice
415 415
416 416
417config NR_LITMUS_SOFTIRQD
418 int "Number of klmirqd."
419 depends on LITMUS_SOFTIRQD
420 range 1 4096
421 default "1"
422 help
423 Should be <= to the number of CPUs in your system.
424
425config LITMUS_NVIDIA 417config LITMUS_NVIDIA
426 bool "Litmus handling of NVIDIA interrupts." 418 bool "Litmus handling of NVIDIA interrupts."
427 default n 419 default n
@@ -445,7 +437,7 @@ config LITMUS_AFFINITY_AWARE_GPU_ASSINGMENT
445config NV_DEVICE_NUM 437config NV_DEVICE_NUM
446 int "Number of NVIDIA GPUs." 438 int "Number of NVIDIA GPUs."
447 depends on LITMUS_SOFTIRQD || LITMUS_PAI_SOFTIRQD 439 depends on LITMUS_SOFTIRQD || LITMUS_PAI_SOFTIRQD
448 range 1 4096 440 range 1 16
449 default "1" 441 default "1"
450 help 442 help
451 Should be (<= to the number of CPUs) and 443 Should be (<= to the number of CPUs) and
@@ -453,11 +445,11 @@ config NV_DEVICE_NUM
453 445
454config NV_MAX_SIMULT_USERS 446config NV_MAX_SIMULT_USERS
455 int "Maximum number of threads sharing a GPU simultanously" 447 int "Maximum number of threads sharing a GPU simultanously"
456 depends on LITMUS_SOFTIRQD || LITMUS_PAI_SOFTIRQD 448 depends on LITMUS_NVIDIA
457 range 1 3 449 range 1 3
458 default "2" 450 default "2"
459 help 451 help
460 Should be equal to the #copy_engines + #execution_engines 452 Should be at least equal to the #copy_engines + #execution_engines
461 of the GPUs in your system. 453 of the GPUs in your system.
462 454
463 Scientific/Professional GPUs = 3 (ex. M2070, Quadro 6000?) 455 Scientific/Professional GPUs = 3 (ex. M2070, Quadro 6000?)
diff --git a/litmus/aux_tasks.c b/litmus/aux_tasks.c
index 20f477f6e3bc..ef26bba3be77 100644
--- a/litmus/aux_tasks.c
+++ b/litmus/aux_tasks.c
@@ -54,7 +54,7 @@ int exit_aux_task(struct task_struct *t)
54 TRACE_CUR("Aux task %s/%d is exiting from %s/%d.\n", t->comm, t->pid, t->group_leader->comm, t->group_leader->pid); 54 TRACE_CUR("Aux task %s/%d is exiting from %s/%d.\n", t->comm, t->pid, t->group_leader->comm, t->group_leader->pid);
55 55
56 tsk_rt(t)->is_aux_task = 0; 56 tsk_rt(t)->is_aux_task = 0;
57 57
58#ifdef CONFIG_REALTIME_AUX_TASK_PRIORITY_INHERITANCE 58#ifdef CONFIG_REALTIME_AUX_TASK_PRIORITY_INHERITANCE
59 list_del(&tsk_rt(t)->aux_task_node); 59 list_del(&tsk_rt(t)->aux_task_node);
60 if (tsk_rt(t)->inh_task) { 60 if (tsk_rt(t)->inh_task) {
@@ -218,36 +218,36 @@ int make_aux_task_if_required(struct task_struct *t)
218{ 218{
219 struct task_struct *leader; 219 struct task_struct *leader;
220 int retval = 0; 220 int retval = 0;
221 221
222 read_lock_irq(&tasklist_lock); 222 read_lock_irq(&tasklist_lock);
223 223
224 leader = t->group_leader; 224 leader = t->group_leader;
225 225
226 if(!tsk_aux(leader)->initialized || !tsk_aux(leader)->aux_future) { 226 if(!tsk_aux(leader)->initialized || !tsk_aux(leader)->aux_future) {
227 goto out; 227 goto out;
228 } 228 }
229 229
230 TRACE_CUR("Making %s/%d in %s/%d an aux thread.\n", t->comm, t->pid, leader->comm, leader->pid); 230 TRACE_CUR("Making %s/%d in %s/%d an aux thread.\n", t->comm, t->pid, leader->comm, leader->pid);
231 231
232 INIT_LIST_HEAD(&tsk_rt(t)->aux_task_node); 232 INIT_LIST_HEAD(&tsk_rt(t)->aux_task_node);
233 INIT_BINHEAP_NODE(&tsk_rt(t)->aux_task_owner_node); 233 INIT_BINHEAP_NODE(&tsk_rt(t)->aux_task_owner_node);
234 234
235 retval = admit_aux_task(t); 235 retval = admit_aux_task(t);
236 if (retval == 0) { 236 if (retval == 0) {
237 tsk_rt(t)->is_aux_task = 1; 237 tsk_rt(t)->is_aux_task = 1;
238 238
239#ifdef CONFIG_REALTIME_AUX_TASK_PRIORITY_INHERITANCE 239#ifdef CONFIG_REALTIME_AUX_TASK_PRIORITY_INHERITANCE
240 list_add_tail(&tsk_rt(t)->aux_task_node, &tsk_aux(leader)->aux_tasks); 240 list_add_tail(&tsk_rt(t)->aux_task_node, &tsk_aux(leader)->aux_tasks);
241 241
242 if (!binheap_empty(&tsk_aux(leader)->aux_task_owners)) { 242 if (!binheap_empty(&tsk_aux(leader)->aux_task_owners)) {
243 struct task_struct *hp = 243 struct task_struct *hp =
244 container_of(binheap_top_entry(&tsk_aux(leader)->aux_task_owners, struct rt_param, aux_task_owner_node), 244 container_of(binheap_top_entry(&tsk_aux(leader)->aux_task_owners, struct rt_param, aux_task_owner_node),
245 struct task_struct, rt_param); 245 struct task_struct, rt_param);
246 246
247 TRACE_CUR("hp in group: %s/%d\n", hp->comm, hp->pid); 247 TRACE_CUR("hp in group: %s/%d\n", hp->comm, hp->pid);
248 248
249 retval = litmus->__increase_prio(t, (tsk_rt(hp)->inh_task)? tsk_rt(hp)->inh_task : hp); 249 retval = litmus->__increase_prio(t, (tsk_rt(hp)->inh_task)? tsk_rt(hp)->inh_task : hp);
250 250
251 if (retval != 0) { 251 if (retval != 0) {
252 /* don't know how to recover from bugs with prio inheritance. better just crash. */ 252 /* don't know how to recover from bugs with prio inheritance. better just crash. */
253 read_unlock_irq(&tasklist_lock); 253 read_unlock_irq(&tasklist_lock);
@@ -256,7 +256,7 @@ int make_aux_task_if_required(struct task_struct *t)
256 } 256 }
257#endif 257#endif
258 } 258 }
259 259
260out: 260out:
261 read_unlock_irq(&tasklist_lock); 261 read_unlock_irq(&tasklist_lock);
262 262
@@ -385,7 +385,7 @@ static long __do_enable_aux_tasks(int flags)
385 if (flags & AUX_FUTURE) { 385 if (flags & AUX_FUTURE) {
386 tsk_aux(leader)->aux_future = 1; 386 tsk_aux(leader)->aux_future = 1;
387 } 387 }
388 388
389 t = leader; 389 t = leader;
390 do { 390 do {
391 if (!tsk_rt(t)->has_aux_tasks && !tsk_rt(t)->is_aux_task) { 391 if (!tsk_rt(t)->has_aux_tasks && !tsk_rt(t)->is_aux_task) {
@@ -398,22 +398,22 @@ static long __do_enable_aux_tasks(int flags)
398 TRACE_CUR("Checking task in %s/%d: %s/%d = (p = %llu):\n", 398 TRACE_CUR("Checking task in %s/%d: %s/%d = (p = %llu):\n",
399 leader->comm, leader->pid, t->comm, t->pid, 399 leader->comm, leader->pid, t->comm, t->pid,
400 tsk_rt(t)->task_params.period); 400 tsk_rt(t)->task_params.period);
401 401
402 /* inspect period to see if it is an rt task */ 402 /* inspect period to see if it is an rt task */
403 if (tsk_rt(t)->task_params.period == 0) { 403 if (tsk_rt(t)->task_params.period == 0) {
404 if (flags && AUX_CURRENT) { 404 if (flags && AUX_CURRENT) {
405 if (!tsk_rt(t)->is_aux_task) { 405 if (!tsk_rt(t)->is_aux_task) {
406 int admit_ret; 406 int admit_ret;
407 407
408 TRACE_CUR("AUX task in %s/%d: %s/%d:\n", leader->comm, leader->pid, t->comm, t->pid); 408 TRACE_CUR("AUX task in %s/%d: %s/%d:\n", leader->comm, leader->pid, t->comm, t->pid);
409 409
410 admit_ret = admit_aux_task(t); 410 admit_ret = admit_aux_task(t);
411 411
412 if (admit_ret == 0) { 412 if (admit_ret == 0) {
413 /* hasn't been aux_tasks_increase_priorityted into rt. make it a aux. */ 413 /* hasn't been aux_tasks_increase_priorityted into rt. make it a aux. */
414 tsk_rt(t)->is_aux_task = 1; 414 tsk_rt(t)->is_aux_task = 1;
415 aux_tasks_added = 1; 415 aux_tasks_added = 1;
416 416
417#ifdef CONFIG_REALTIME_AUX_TASK_PRIORITY_INHERITANCE 417#ifdef CONFIG_REALTIME_AUX_TASK_PRIORITY_INHERITANCE
418 list_add_tail(&tsk_rt(t)->aux_task_node, &tsk_aux(leader)->aux_tasks); 418 list_add_tail(&tsk_rt(t)->aux_task_node, &tsk_aux(leader)->aux_tasks);
419#endif 419#endif
@@ -464,7 +464,7 @@ static long __do_disable_aux_tasks(int flags)
464 if (flags & AUX_FUTURE) { 464 if (flags & AUX_FUTURE) {
465 tsk_aux(leader)->aux_future = 0; 465 tsk_aux(leader)->aux_future = 0;
466 } 466 }
467 467
468 if (flags & AUX_CURRENT) { 468 if (flags & AUX_CURRENT) {
469 t = leader; 469 t = leader;
470 do { 470 do {
diff --git a/litmus/edf_common.c b/litmus/edf_common.c
index c279bf12a7f5..27b728a55669 100644
--- a/litmus/edf_common.c
+++ b/litmus/edf_common.c
@@ -73,6 +73,22 @@ int edf_higher_prio(struct task_struct* first, struct task_struct* second)
73 return first && !second; 73 return first && !second;
74 } 74 }
75 75
76 /* There is some goofy stuff in this code here. There are three subclasses
77 * within the SCHED_LITMUS scheduling class:
78 * 1) Auxiliary tasks: COTS helper threads from the application level that
79 * are forced to be real-time.
80 * 2) klmirqd interrupt threads: Litmus threaded interrupt handlers.
81 * 3) Normal Litmus tasks.
82 *
83 * At their base priorities, #3 > #2 > #1. However, #1 and #2 threads might
84 * inherit a priority from a task of #3.
85 *
86 * The code proceeds in the following manner:
87 * 1) Make aux and klmirqd threads with base-priorities have low priorities.
88 * 2) Determine effective priorities.
89 * 3) Perform priority comparison. Favor #3 over #1 and #2 in case of tie.
90 */
91
76 92
77#if defined(CONFIG_REALTIME_AUX_TASK_PRIORITY_BOOSTED) 93#if defined(CONFIG_REALTIME_AUX_TASK_PRIORITY_BOOSTED)
78 /* run aux tasks at max priority */ 94 /* run aux tasks at max priority */
@@ -109,7 +125,7 @@ int edf_higher_prio(struct task_struct* first, struct task_struct* second)
109 return temp; 125 return temp;
110 } 126 }
111 } 127 }
112 128
113 if (first->rt_param.is_aux_task && second->rt_param.is_aux_task && 129 if (first->rt_param.is_aux_task && second->rt_param.is_aux_task &&
114 first->rt_param.inh_task == second->rt_param.inh_task) { // inh_task is !NULL for both tasks since neither was a lo_aux task 130 first->rt_param.inh_task == second->rt_param.inh_task) { // inh_task is !NULL for both tasks since neither was a lo_aux task
115 // Both aux tasks inherit from the same task, so tie-break 131 // Both aux tasks inherit from the same task, so tie-break
@@ -120,6 +136,36 @@ int edf_higher_prio(struct task_struct* first, struct task_struct* second)
120 } 136 }
121#endif 137#endif
122 138
139#ifdef CONFIG_LITMUS_SOFTIRQD
140 {
141 int first_lo_klmirqd = first->rt_param.is_interrupt_thread && !first->rt_param.inh_task;
142 int second_lo_klmirqd = second->rt_param.is_interrupt_thread && !second->rt_param.inh_task;
143
144 /* prioritize aux tasks without inheritance below real-time tasks */
145 if (first_lo_klmirqd || second_lo_klmirqd) {
146 // one of these is an klmirqd thread without inheritance.
147 if(first_lo_klmirqd && second_lo_klmirqd) {
148 TRACE_CUR("klmirqd tie break!\n"); // tie-break by BASE priority of the aux tasks
149 goto klmirqd_tie_break;
150 }
151 else {
152 // make the klmirqd thread (second) lowest priority real-time task
153 int temp = (first_lo_klmirqd) ? !is_realtime(second) : !is_realtime(first);
154 TRACE_CUR("%s/%d >> %s/%d --- %d\n", first->comm, first->pid, second->comm, second->pid, temp);
155 return temp;
156 }
157 }
158
159 if (first->rt_param.is_interrupt_thread && second->rt_param.is_interrupt_thread &&
160 first->rt_param.inh_task == second->rt_param.inh_task) { // inh_task is !NULL for both tasks since neither was a lo_klmirqd task
161 // Both klmirqd tasks inherit from the same task, so tie-break
162 // by base priority of the klmirqd tasks.
163 TRACE_CUR("klmirqd tie break!\n");
164 goto klmirqd_tie_break;
165 }
166 }
167#endif
168
123 169
124#ifdef CONFIG_LITMUS_LOCKING 170#ifdef CONFIG_LITMUS_LOCKING
125 /* Check for EFFECTIVE priorities. Change task 171 /* Check for EFFECTIVE priorities. Change task
@@ -161,7 +207,8 @@ int edf_higher_prio(struct task_struct* first, struct task_struct* second)
161#endif 207#endif
162 208
163aux_tie_break: 209aux_tie_break:
164 210klmirqd_tie_break:
211
165 if (!is_realtime(second_task)) { 212 if (!is_realtime(second_task)) {
166 return 1; 213 return 1;
167 } 214 }
@@ -230,15 +277,13 @@ aux_tie_break:
230 } 277 }
231 else if (first_task->pid == second_task->pid) { 278 else if (first_task->pid == second_task->pid) {
232#ifdef CONFIG_LITMUS_SOFTIRQD 279#ifdef CONFIG_LITMUS_SOFTIRQD
233 if (first_task->rt_param.is_proxy_thread < 280 if (first_task->rt_param.is_interrupt_thread < second_task->rt_param.is_interrupt_thread) {
234 second_task->rt_param.is_proxy_thread) {
235 return 1; 281 return 1;
236 } 282 }
237 else if (first_task->rt_param.is_proxy_thread == second_task->rt_param.is_proxy_thread) { 283 else if (first_task->rt_param.is_interrupt_thread == second_task->rt_param.is_interrupt_thread) {
238#endif 284#endif
239 285
240#if defined(CONFIG_REALTIME_AUX_TASK_PRIORITY_INHERITANCE) 286#if defined(CONFIG_REALTIME_AUX_TASK_PRIORITY_INHERITANCE)
241 /* is this dead code? */
242 if (tsk_rt(first)->is_aux_task < tsk_rt(second)->is_aux_task) { 287 if (tsk_rt(first)->is_aux_task < tsk_rt(second)->is_aux_task) {
243 return 1; 288 return 1;
244 } 289 }
@@ -246,8 +291,7 @@ aux_tie_break:
246#endif 291#endif
247 292
248 /* Something could be wrong if you get this far. */ 293 /* Something could be wrong if you get this far. */
249 if (unlikely(first->rt_param.inh_task == 294 if (unlikely(first->rt_param.inh_task == second->rt_param.inh_task)) {
250 second->rt_param.inh_task)) {
251 /* Both tasks have the same inherited priority. 295 /* Both tasks have the same inherited priority.
252 * Likely in a bug-condition. 296 * Likely in a bug-condition.
253 */ 297 */
diff --git a/litmus/ikglp_lock.c b/litmus/ikglp_lock.c
index b29828344dd1..a4ae74331782 100644
--- a/litmus/ikglp_lock.c
+++ b/litmus/ikglp_lock.c
@@ -1960,11 +1960,11 @@ static struct affinity_observer* ikglp_aff_obs_new(struct affinity_observer_ops*
1960 return(NULL); 1960 return(NULL);
1961 } 1961 }
1962 1962
1963 if(aff_args.nr_simult_users > NV_MAX_SIMULT_USERS) { 1963// if(aff_args.nr_simult_users > NV_MAX_SIMULT_USERS) {
1964 TRACE_CUR("System does not support #simult_users > %d. %d requested.\n", 1964// TRACE_CUR("System does not support #simult_users > %d. %d requested.\n",
1965 NV_MAX_SIMULT_USERS, aff_args.nr_simult_users); 1965// NV_MAX_SIMULT_USERS, aff_args.nr_simult_users);
1966// return(NULL); 1966//// return(NULL);
1967 } 1967// }
1968 1968
1969 ikglp_aff = kmalloc(sizeof(*ikglp_aff), GFP_KERNEL); 1969 ikglp_aff = kmalloc(sizeof(*ikglp_aff), GFP_KERNEL);
1970 if(!ikglp_aff) { 1970 if(!ikglp_aff) {
@@ -2124,7 +2124,7 @@ struct fifo_queue* gpu_ikglp_advise_enqueue(struct ikglp_affinity* aff, struct t
2124 if(aff->q_info[i].q->count < max_fifo_len) { 2124 if(aff->q_info[i].q->count < max_fifo_len) {
2125 int want = 0; 2125 int want = 0;
2126 2126
2127 lt_t migration = 2127 lt_t migration =
2128 get_gpu_estimate(t, 2128 get_gpu_estimate(t,
2129 gpu_migration_distance(tsk_rt(t)->last_gpu, 2129 gpu_migration_distance(tsk_rt(t)->last_gpu,
2130 replica_to_gpu(aff, i))); 2130 replica_to_gpu(aff, i)));
diff --git a/litmus/jobs.c b/litmus/jobs.c
index 9fe4eb1fa168..8593a8d2f107 100644
--- a/litmus/jobs.c
+++ b/litmus/jobs.c
@@ -30,7 +30,7 @@ void prepare_for_next_period(struct task_struct *t)
30 * release and deadline. Lateness may be negative. 30 * release and deadline. Lateness may be negative.
31 */ 31 */
32 t->rt_param.job_params.lateness = 32 t->rt_param.job_params.lateness =
33 (long long)litmus_clock() - 33 (long long)litmus_clock() -
34 (long long)t->rt_param.job_params.deadline; 34 (long long)t->rt_param.job_params.deadline;
35 35
36 setup_release(t, get_release(t) + get_rt_period(t)); 36 setup_release(t, get_release(t) + get_rt_period(t));
diff --git a/litmus/kfmlp_lock.c b/litmus/kfmlp_lock.c
index ab472330095d..785a095275e6 100644
--- a/litmus/kfmlp_lock.c
+++ b/litmus/kfmlp_lock.c
@@ -587,11 +587,11 @@ static struct affinity_observer* kfmlp_aff_obs_new(struct affinity_observer_ops*
587 return(NULL); 587 return(NULL);
588 } 588 }
589 589
590 if(aff_args.nr_simult_users > NV_MAX_SIMULT_USERS) { 590// if(aff_args.nr_simult_users > NV_MAX_SIMULT_USERS) {
591 TRACE_CUR("System does not support #simult_users > %d. %d requested.\n", 591// TRACE_CUR("System does not support #simult_users > %d. %d requested.\n",
592 NV_MAX_SIMULT_USERS, aff_args.nr_simult_users); 592// NV_MAX_SIMULT_USERS, aff_args.nr_simult_users);
593// return(NULL); 593//// return(NULL);
594 } 594// }
595 595
596 kfmlp_aff = kmalloc(sizeof(*kfmlp_aff), GFP_KERNEL); 596 kfmlp_aff = kmalloc(sizeof(*kfmlp_aff), GFP_KERNEL);
597 if(!kfmlp_aff) { 597 if(!kfmlp_aff) {
@@ -829,6 +829,7 @@ void gpu_kfmlp_notify_acquired(struct kfmlp_affinity* aff, struct kfmlp_queue* f
829 829
830 reg_nv_device(gpu, 1, t); // register 830 reg_nv_device(gpu, 1, t); // register
831 831
832
832 tsk_rt(t)->suspend_gpu_tracker_on_block = 0; 833 tsk_rt(t)->suspend_gpu_tracker_on_block = 0;
833 reset_gpu_tracker(t); 834 reset_gpu_tracker(t);
834 start_gpu_tracker(t); 835 start_gpu_tracker(t);
diff --git a/litmus/litmus.c b/litmus/litmus.c
index 3b8017397e80..fa244ba53e22 100644
--- a/litmus/litmus.c
+++ b/litmus/litmus.c
@@ -60,28 +60,6 @@ void bheap_node_free(struct bheap_node* hn)
60struct release_heap* release_heap_alloc(int gfp_flags); 60struct release_heap* release_heap_alloc(int gfp_flags);
61void release_heap_free(struct release_heap* rh); 61void release_heap_free(struct release_heap* rh);
62 62
63#ifdef CONFIG_LITMUS_NVIDIA
64/*
65 * sys_register_nv_device
66 * @nv_device_id: The Nvidia device id that the task want to register
67 * @reg_action: set to '1' to register the specified device. zero otherwise.
68 * Syscall for register task's designated nvidia device into NV_DEVICE_REG array
69 * Returns EFAULT if nv_device_id is out of range.
70 * 0 if success
71 */
72asmlinkage long sys_register_nv_device(int nv_device_id, int reg_action)
73{
74 /* register the device to caller (aka 'current') */
75 return(reg_nv_device(nv_device_id, reg_action, current));
76}
77#else
78asmlinkage long sys_register_nv_device(int nv_device_id, int reg_action)
79{
80 return(-EINVAL);
81}
82#endif
83
84
85/* 63/*
86 * sys_set_task_rt_param 64 * sys_set_task_rt_param
87 * @pid: Pid of the task which scheduling parameters must be changed 65 * @pid: Pid of the task which scheduling parameters must be changed
@@ -393,22 +371,11 @@ static void reinit_litmus_state(struct task_struct* p, int restore)
393// WARN_ON(!binheap_empty(&p->rt_param.hp_blocked_tasks)); 371// WARN_ON(!binheap_empty(&p->rt_param.hp_blocked_tasks));
394#endif 372#endif
395 373
396#ifdef CONFIG_LITMUS_SOFTIRQD
397 /* We probably should not have any tasklets executing for
398 * us at this time.
399 */
400 WARN_ON(p->rt_param.cur_klmirqd);
401 WARN_ON(atomic_read(&p->rt_param.klmirqd_sem_stat) == HELD);
402
403 if(p->rt_param.cur_klmirqd)
404 flush_pending(p->rt_param.cur_klmirqd, p);
405
406 if(atomic_read(&p->rt_param.klmirqd_sem_stat) == HELD)
407 up_and_set_stat(p, NOT_HELD, &p->rt_param.klmirqd_sem);
408#endif
409 374
410#ifdef CONFIG_LITMUS_NVIDIA 375#ifdef CONFIG_LITMUS_NVIDIA
411 WARN_ON(p->rt_param.held_gpus != 0); 376 WARN_ON(p->rt_param.held_gpus != 0);
377
378 INIT_BINHEAP_NODE(&p->rt_param.gpu_owner_node);
412#endif 379#endif
413 380
414 /* Cleanup everything else. */ 381 /* Cleanup everything else. */
@@ -477,11 +444,9 @@ long __litmus_admit_task(struct task_struct* tsk)
477 //INIT_BINHEAP_HANDLE(&tsk_rt(tsk)->hp_blocked_tasks, prio_order); // done by scheduler 444 //INIT_BINHEAP_HANDLE(&tsk_rt(tsk)->hp_blocked_tasks, prio_order); // done by scheduler
478#endif 445#endif
479#ifdef CONFIG_LITMUS_SOFTIRQD 446#ifdef CONFIG_LITMUS_SOFTIRQD
480 /* proxy thread off by default */ 447 /* not an interrupt thread by default */
481 tsk_rt(tsk)is_proxy_thread = 0; 448 tsk_rt(tsk)->is_interrupt_thread = 0;
482 tsk_rt(tsk)cur_klmirqd = NULL; 449 tsk_rt(tsk)->klmirqd_info = NULL;
483 mutex_init(&tsk_rt(tsk)->klmirqd_sem);
484 atomic_set(&tsk_rt(tsk)->klmirqd_sem_stat, NOT_HELD);
485#endif 450#endif
486 451
487 retval = litmus->admit_task(tsk); 452 retval = litmus->admit_task(tsk);
@@ -580,8 +545,7 @@ int switch_sched_plugin(struct sched_plugin* plugin)
580 cpu_relax(); 545 cpu_relax();
581 546
582#ifdef CONFIG_LITMUS_SOFTIRQD 547#ifdef CONFIG_LITMUS_SOFTIRQD
583 if(!klmirqd_is_dead()) 548 if (!klmirqd_is_dead()) {
584 {
585 kill_klmirqd(); 549 kill_klmirqd();
586 } 550 }
587#endif 551#endif
diff --git a/litmus/litmus_softirq.c b/litmus/litmus_softirq.c
index 73a3053e662b..44e2d38ad982 100644
--- a/litmus/litmus_softirq.c
+++ b/litmus/litmus_softirq.c
@@ -18,10 +18,6 @@
18 18
19/* TODO: Remove unneeded mb() and other barriers. */ 19/* TODO: Remove unneeded mb() and other barriers. */
20 20
21
22/* counts number of daemons ready to handle litmus irqs. */
23static atomic_t num_ready_klmirqds = ATOMIC_INIT(0);
24
25enum pending_flags 21enum pending_flags
26{ 22{
27 LIT_TASKLET_LOW = 0x1, 23 LIT_TASKLET_LOW = 0x1,
@@ -29,35 +25,313 @@ enum pending_flags
29 LIT_WORK = LIT_TASKLET_HI<<1 25 LIT_WORK = LIT_TASKLET_HI<<1
30}; 26};
31 27
32/* only support tasklet processing for now. */ 28struct klmirqd_registration
33struct tasklet_head
34{ 29{
35 struct tasklet_struct *head; 30 raw_spinlock_t lock;
36 struct tasklet_struct **tail; 31 u32 nr_threads;
32 unsigned int initialized:1;
33 unsigned int shuttingdown:1;
34 struct list_head threads;
37}; 35};
38 36
39struct klmirqd_info 37static atomic_t klmirqd_id_gen = ATOMIC_INIT(0);
38
39static struct klmirqd_registration klmirqd_state;
40
41
42
43void init_klmirqd(void)
44{
45 raw_spin_lock_init(&klmirqd_state.lock);
46
47 klmirqd_state.nr_threads = 0;
48 klmirqd_state.initialized = 1;
49 klmirqd_state.shuttingdown = 0;
50 INIT_LIST_HEAD(&klmirqd_state.threads);
51}
52
53static int __klmirqd_is_ready(void)
54{
55 return (klmirqd_state.initialized == 1 && klmirqd_state.shuttingdown == 0);
56}
57
58int klmirqd_is_ready(void)
59{
60 unsigned long flags;
61 int ret;
62
63 raw_spin_lock_irqsave(&klmirqd_state.lock, flags);
64 ret = __klmirqd_is_ready();
65 raw_spin_unlock_irqrestore(&klmirqd_state.lock, flags);
66
67 return ret;
68}
69
70int klmirqd_is_dead(void)
71{
72 return(!klmirqd_is_ready());
73}
74
75
76void kill_klmirqd(void)
77{
78 if(!klmirqd_is_dead())
79 {
80 unsigned long flags;
81 struct list_head *pos;
82
83 raw_spin_lock_irqsave(&klmirqd_state.lock, flags);
84
85 TRACE("%s: Killing all klmirqd threads! (%d of them)\n", __FUNCTION__, klmirqd_state.nr_threads);
86
87 klmirqd_state.shuttingdown = 1;
88
89 list_for_each(pos, &klmirqd_state.threads) {
90 struct klmirqd_info* info = list_entry(pos, struct klmirqd_info, klmirqd_reg);
91
92 if(info->terminating != 1)
93 {
94 info->terminating = 1;
95 mb(); /* just to be sure? */
96 flush_pending(info->klmirqd);
97
98 /* signal termination */
99 kthread_stop(info->klmirqd);
100 }
101 }
102
103 raw_spin_unlock_irqrestore(&klmirqd_state.lock, flags);
104 }
105}
106
107
108
109void kill_klmirqd_thread(struct task_struct* klmirqd_thread)
40{ 110{
41 struct task_struct* klmirqd; 111 unsigned long flags;
42 struct task_struct* current_owner; 112 struct klmirqd_info* info;
43 int terminating;
44 113
114 if (!tsk_rt(klmirqd_thread)->is_interrupt_thread) {
115 TRACE("%s/%d is not a klmirqd thread\n", klmirqd_thread->comm, klmirqd_thread->pid);
116 return;
117 }
118
119 TRACE("%s: Killing klmirqd thread %s/%d\n", __FUNCTION__, klmirqd_thread->comm, klmirqd_thread->pid);
45 120
46 raw_spinlock_t lock; 121 raw_spin_lock_irqsave(&klmirqd_state.lock, flags);
47 122
48 u32 pending; 123 info = tsk_rt(klmirqd_thread)->klmirqd_info;
49 atomic_t num_hi_pending; 124
50 atomic_t num_low_pending; 125 if(info->terminating != 1) {
51 atomic_t num_work_pending; 126 info->terminating = 1;
127 mb();
128
129 flush_pending(klmirqd_thread);
130 kthread_stop(klmirqd_thread);
131 }
132
133 raw_spin_unlock_irqrestore(&klmirqd_state.lock, flags);
134}
52 135
53 /* in order of priority */ 136
54 struct tasklet_head pending_tasklets_hi; 137
55 struct tasklet_head pending_tasklets; 138struct klmirqd_launch_data
56 struct list_head worklist; 139{
140 int cpu_affinity;
141 klmirqd_callback_t* cb;
142 struct work_struct work;
57}; 143};
58 144
59/* one list for each klmirqd */ 145static int run_klmirqd(void* callback);
60static struct klmirqd_info klmirqds[NR_LITMUS_SOFTIRQD]; 146
147
148/* executed by a kworker from workqueues */
149static void __launch_klmirqd_thread(struct work_struct *work)
150{
151 int id;
152 struct task_struct* thread = NULL;
153 struct klmirqd_launch_data* launch_data =
154 container_of(work, struct klmirqd_launch_data, work);
155
156 TRACE("%s: Creating klmirqd thread\n", __FUNCTION__);
157
158 id = atomic_inc_return(&klmirqd_id_gen);
159
160 if (launch_data->cpu_affinity != -1) {
161 thread = kthread_create(
162 run_klmirqd,
163 /* treat the affinity as a pointer, we'll cast it back later */
164 (void*)launch_data->cb,
165 "klmirqd_th%d/%d",
166 id,
167 launch_data->cpu_affinity);
168
169 /* litmus will put is in the right cluster. */
170 kthread_bind(thread, launch_data->cpu_affinity);
171
172 TRACE("%s: Launching klmirqd_th%d/%d\n", __FUNCTION__, id, launch_data->cpu_affinity);
173 }
174 else {
175 thread = kthread_create(
176 run_klmirqd,
177 /* treat the affinity as a pointer, we'll cast it back later */
178 (void*)launch_data->cb,
179 "klmirqd_th%d",
180 id);
181
182 TRACE("%s: Launching klmirqd_th%d\n", __FUNCTION__, id);
183 }
184
185 if (thread) {
186 wake_up_process(thread);
187 }
188 else {
189 TRACE("Could not create klmirqd/%d thread!\n", id);
190 }
191
192 kfree(launch_data);
193}
194
195
196int launch_klmirqd_thread(int cpu, klmirqd_callback_t* cb)
197{
198 struct klmirqd_launch_data* delayed_launch;
199
200 if (!klmirqd_is_ready()) {
201 TRACE("klmirqd is not ready. Check that it was initialized!\n");
202 return -1;
203 }
204
205 /* tell a work queue to launch the threads. we can't make scheduling
206 calls since we're in an atomic state. */
207 delayed_launch = kmalloc(sizeof(struct klmirqd_launch_data), GFP_ATOMIC);
208 delayed_launch->cpu_affinity = cpu;
209 delayed_launch->cb = cb;
210 INIT_WORK(&delayed_launch->work, __launch_klmirqd_thread);
211 schedule_work(&delayed_launch->work);
212
213 return 0;
214}
215
216
217
218
219#define KLMIRQD_SLICE_NR_JIFFIES 1
220#define KLMIRQD_SLICE_NS ((NSEC_PER_SEC / HZ) * KLMIRQD_SLICE_NR_JIFFIES)
221
222static int set_litmus_daemon_sched(struct task_struct* tsk)
223{
224 int ret = 0;
225
226 struct rt_task tp = {
227 .period = KLMIRQD_SLICE_NS, /* dummy 1 second period */
228 .relative_deadline = KLMIRQD_SLICE_NS,
229 .exec_cost = KLMIRQD_SLICE_NS,
230 .phase = 0,
231 .cpu = task_cpu(current),
232 .budget_policy = NO_ENFORCEMENT,
233 .budget_signal_policy = NO_SIGNALS,
234 .cls = RT_CLASS_BEST_EFFORT
235 };
236
237 struct sched_param param = { .sched_priority = 0};
238
239 TRACE_CUR("Setting %s/%d as daemon thread.\n", tsk->comm, tsk->pid);
240
241 /* set task params */
242 tsk_rt(tsk)->task_params = tp;
243 tsk_rt(tsk)->is_interrupt_thread = 1;
244
245 /* inform the OS we're SCHED_LITMUS --
246 sched_setscheduler_nocheck() calls litmus_admit_task(). */
247 sched_setscheduler_nocheck(tsk, SCHED_LITMUS, &param);
248
249 return ret;
250}
251
252static int register_klmirqd(struct task_struct* tsk)
253{
254 int retval = 0;
255 unsigned long flags;
256 struct klmirqd_info *info = NULL;
257
258 if (!tsk_rt(tsk)->is_interrupt_thread) {
259 TRACE("Only proxy threads already running in Litmus may become klmirqd threads!\n");
260 WARN_ON(1);
261 retval = -1;
262 goto out;
263 }
264
265 raw_spin_lock_irqsave(&klmirqd_state.lock, flags);
266
267 if (!__klmirqd_is_ready()) {
268 TRACE("klmirqd is not ready! Did you forget to initialize it?\n");
269 WARN_ON(1);
270 retval = -1;
271 goto out_unlock;
272 }
273
274 /* allocate and initialize klmirqd data for the thread */
275 info = kmalloc(sizeof(struct klmirqd_info), GFP_KERNEL);
276 if (!info) {
277 TRACE("Failed to allocate klmirqd_info struct!\n");
278 retval = -1; /* todo: pick better code */
279 goto out_unlock;
280 }
281 memset(info, 0, sizeof(struct klmirqd_info));
282 info->klmirqd = tsk;
283 info->pending_tasklets_hi.tail = &info->pending_tasklets_hi.head;
284 info->pending_tasklets.tail = &info->pending_tasklets.head;
285 INIT_LIST_HEAD(&info->worklist);
286 INIT_LIST_HEAD(&info->klmirqd_reg);
287 raw_spin_lock_init(&info->lock);
288
289
290 /* now register with klmirqd */
291 list_add_tail(&info->klmirqd_reg, &klmirqd_state.threads);
292 ++klmirqd_state.nr_threads;
293
294 /* update the task struct to point to klmirqd info */
295 tsk_rt(tsk)->klmirqd_info = info;
296
297out_unlock:
298 raw_spin_unlock_irqrestore(&klmirqd_state.lock, flags);
299
300out:
301 return retval;
302}
303
304static int unregister_klmirqd(struct task_struct* tsk)
305{
306 int retval = 0;
307 unsigned long flags;
308 struct klmirqd_info *info = tsk_rt(tsk)->klmirqd_info;
309
310 if (!tsk_rt(tsk)->is_interrupt_thread || !info) {
311 TRACE("%s/%d is not a klmirqd thread!\n", tsk->comm, tsk->pid);
312 WARN_ON(1);
313 retval = -1;
314 goto out;
315 }
316
317 raw_spin_lock_irqsave(&klmirqd_state.lock, flags);
318
319 /* remove the entry in the klmirqd thread list */
320 list_del(&info->klmirqd_reg);
321 --klmirqd_state.nr_threads;
322
323 /* remove link to klmirqd info from thread */
324 tsk_rt(tsk)->klmirqd_info = NULL;
325
326 /* clean up memory */
327 kfree(info);
328
329 raw_spin_unlock_irqrestore(&klmirqd_state.lock, flags);
330
331out:
332 return retval;
333}
334
61 335
62 336
63 337
@@ -67,35 +341,50 @@ int proc_read_klmirqd_stats(char *page, char **start,
67 off_t off, int count, 341 off_t off, int count,
68 int *eof, void *data) 342 int *eof, void *data)
69{ 343{
70 int len = snprintf(page, PAGE_SIZE, 344 unsigned long flags;
71 "num ready klmirqds: %d\n\n", 345 int len;
72 atomic_read(&num_ready_klmirqds)); 346
73 347 raw_spin_lock_irqsave(&klmirqd_state.lock, flags);
74 if(klmirqd_is_ready()) 348
75 { 349 if (klmirqd_state.initialized) {
76 int i; 350 if (!klmirqd_state.shuttingdown) {
77 for(i = 0; i < NR_LITMUS_SOFTIRQD; ++i) 351 struct list_head *pos;
78 { 352
79 len += 353 len = snprintf(page, PAGE_SIZE,
80 snprintf(page + len - 1, PAGE_SIZE, /* -1 to strip off \0 */ 354 "num ready klmirqds: %d\n\n",
81 "klmirqd_th%d: %s/%d\n" 355 klmirqd_state.nr_threads);
82 "\tcurrent_owner: %s/%d\n" 356
83 "\tpending: %x\n" 357 list_for_each(pos, &klmirqd_state.threads) {
84 "\tnum hi: %d\n" 358 struct klmirqd_info* info = list_entry(pos, struct klmirqd_info, klmirqd_reg);
85 "\tnum low: %d\n" 359
86 "\tnum work: %d\n\n", 360 len +=
87 i, 361 snprintf(page + len - 1, PAGE_SIZE, /* -1 to strip off \0 */
88 klmirqds[i].klmirqd->comm, klmirqds[i].klmirqd->pid, 362 "klmirqd_thread: %s/%d\n"
89 (klmirqds[i].current_owner != NULL) ? 363 "\tcurrent_owner: %s/%d\n"
90 klmirqds[i].current_owner->comm : "(null)", 364 "\tpending: %x\n"
91 (klmirqds[i].current_owner != NULL) ? 365 "\tnum hi: %d\n"
92 klmirqds[i].current_owner->pid : 0, 366 "\tnum low: %d\n"
93 klmirqds[i].pending, 367 "\tnum work: %d\n\n",
94 atomic_read(&klmirqds[i].num_hi_pending), 368 info->klmirqd->comm, info->klmirqd->pid,
95 atomic_read(&klmirqds[i].num_low_pending), 369 (info->current_owner != NULL) ?
96 atomic_read(&klmirqds[i].num_work_pending)); 370 info->current_owner->comm : "(null)",
371 (info->current_owner != NULL) ?
372 info->current_owner->pid : 0,
373 info->pending,
374 atomic_read(&info->num_hi_pending),
375 atomic_read(&info->num_low_pending),
376 atomic_read(&info->num_work_pending));
377 }
378 }
379 else {
380 len = snprintf(page, PAGE_SIZE, "klmirqd is shutting down\n");
97 } 381 }
98 } 382 }
383 else {
384 len = snprintf(page, PAGE_SIZE, "klmirqd is not initialized!\n");
385 }
386
387 raw_spin_unlock_irqrestore(&klmirqd_state.lock, flags);
99 388
100 return(len); 389 return(len);
101} 390}
@@ -162,6 +451,15 @@ static void dump_state(struct klmirqd_info* which, const char* caller)
162#endif 451#endif
163 452
164 453
454
455
456
457
458
459
460
461
462
165/* forward declarations */ 463/* forward declarations */
166static void ___litmus_tasklet_schedule(struct tasklet_struct *t, 464static void ___litmus_tasklet_schedule(struct tasklet_struct *t,
167 struct klmirqd_info *which, 465 struct klmirqd_info *which,
@@ -174,24 +472,6 @@ static void ___litmus_schedule_work(struct work_struct *w,
174 int wakeup); 472 int wakeup);
175 473
176 474
177
178inline unsigned int klmirqd_id(struct task_struct* tsk)
179{
180 int i;
181 for(i = 0; i < NR_LITMUS_SOFTIRQD; ++i)
182 {
183 if(klmirqds[i].klmirqd == tsk)
184 {
185 return i;
186 }
187 }
188
189 BUG();
190
191 return 0;
192}
193
194
195inline static u32 litirq_pending_hi_irqoff(struct klmirqd_info* which) 475inline static u32 litirq_pending_hi_irqoff(struct klmirqd_info* which)
196{ 476{
197 return (which->pending & LIT_TASKLET_HI); 477 return (which->pending & LIT_TASKLET_HI);
@@ -225,200 +505,11 @@ inline static u32 litirq_pending(struct klmirqd_info* which)
225 return pending; 505 return pending;
226}; 506};
227 507
228inline static u32 litirq_pending_with_owner(struct klmirqd_info* which, struct task_struct* owner)
229{
230 unsigned long flags;
231 u32 pending;
232
233 raw_spin_lock_irqsave(&which->lock, flags);
234 pending = litirq_pending_irqoff(which);
235 if(pending)
236 {
237 if(which->current_owner != owner)
238 {
239 pending = 0; // owner switch!
240 }
241 }
242 raw_spin_unlock_irqrestore(&which->lock, flags);
243
244 return pending;
245}
246
247
248inline static u32 litirq_pending_and_sem_and_owner(struct klmirqd_info* which,
249 struct mutex** sem,
250 struct task_struct** t)
251{
252 unsigned long flags;
253 u32 pending;
254
255 /* init values */
256 *sem = NULL;
257 *t = NULL;
258
259 raw_spin_lock_irqsave(&which->lock, flags);
260
261 pending = litirq_pending_irqoff(which);
262 if(pending)
263 {
264 if(which->current_owner != NULL)
265 {
266 *t = which->current_owner;
267 *sem = &tsk_rt(which->current_owner)->klmirqd_sem;
268 }
269 else
270 {
271 BUG();
272 }
273 }
274 raw_spin_unlock_irqrestore(&which->lock, flags);
275
276 if(likely(*sem))
277 {
278 return pending;
279 }
280 else
281 {
282 return 0;
283 }
284}
285
286/* returns true if the next piece of work to do is from a different owner.
287 */
288static int tasklet_ownership_change(
289 struct klmirqd_info* which,
290 enum pending_flags taskletQ)
291{
292 /* this function doesn't have to look at work objects since they have
293 priority below tasklets. */
294
295 unsigned long flags;
296 int ret = 0;
297
298 raw_spin_lock_irqsave(&which->lock, flags);
299
300 switch(taskletQ)
301 {
302 case LIT_TASKLET_HI:
303 if(litirq_pending_hi_irqoff(which))
304 {
305 ret = (which->pending_tasklets_hi.head->owner !=
306 which->current_owner);
307 }
308 break;
309 case LIT_TASKLET_LOW:
310 if(litirq_pending_low_irqoff(which))
311 {
312 ret = (which->pending_tasklets.head->owner !=
313 which->current_owner);
314 }
315 break;
316 default:
317 break;
318 }
319
320 raw_spin_unlock_irqrestore(&which->lock, flags);
321
322 TRACE_TASK(which->klmirqd, "ownership change needed: %d\n", ret);
323
324 return ret;
325}
326
327
328static void __reeval_prio(struct klmirqd_info* which)
329{
330 struct task_struct* next_owner = NULL;
331 struct task_struct* klmirqd = which->klmirqd;
332
333 /* Check in prio-order */
334 u32 pending = litirq_pending_irqoff(which);
335
336 //__dump_state(which, "__reeval_prio: before");
337
338 if(pending)
339 {
340 if(pending & LIT_TASKLET_HI)
341 {
342 next_owner = which->pending_tasklets_hi.head->owner;
343 }
344 else if(pending & LIT_TASKLET_LOW)
345 {
346 next_owner = which->pending_tasklets.head->owner;
347 }
348 else if(pending & LIT_WORK)
349 {
350 struct work_struct* work =
351 list_first_entry(&which->worklist, struct work_struct, entry);
352 next_owner = work->owner;
353 }
354 }
355
356 if(next_owner != which->current_owner)
357 {
358 struct task_struct* old_owner = which->current_owner;
359
360 /* bind the next owner. */
361 which->current_owner = next_owner;
362 mb();
363
364 if(next_owner != NULL)
365 {
366 if(!in_interrupt())
367 {
368 TRACE_CUR("%s: Ownership change: %s/%d to %s/%d\n", __FUNCTION__,
369 ((tsk_rt(klmirqd)->inh_task) ? tsk_rt(klmirqd)->inh_task : klmirqd)->comm,
370 ((tsk_rt(klmirqd)->inh_task) ? tsk_rt(klmirqd)->inh_task : klmirqd)->pid,
371 next_owner->comm, next_owner->pid);
372 }
373 else
374 {
375 TRACE("%s: Ownership change: %s/%d to %s/%d\n", __FUNCTION__,
376 ((tsk_rt(klmirqd)->inh_task) ? tsk_rt(klmirqd)->inh_task : klmirqd)->comm,
377 ((tsk_rt(klmirqd)->inh_task) ? tsk_rt(klmirqd)->inh_task : klmirqd)->pid,
378 next_owner->comm, next_owner->pid);
379 }
380
381 litmus->increase_prio_inheritance_klmirqd(klmirqd, old_owner, next_owner);
382 }
383 else
384 {
385 if(likely(!in_interrupt()))
386 {
387 TRACE_CUR("%s: Ownership change: %s/%d to NULL (reverting)\n",
388 __FUNCTION__, klmirqd->comm, klmirqd->pid);
389 }
390 else
391 {
392 // is this a bug?
393 TRACE("%s: Ownership change: %s/%d to NULL (reverting)\n",
394 __FUNCTION__, klmirqd->comm, klmirqd->pid);
395 }
396
397 BUG_ON(pending != 0);
398 litmus->decrease_prio_inheritance_klmirqd(klmirqd, old_owner, NULL);
399 }
400 }
401
402 //__dump_state(which, "__reeval_prio: after");
403}
404
405static void reeval_prio(struct klmirqd_info* which)
406{
407 unsigned long flags;
408
409 raw_spin_lock_irqsave(&which->lock, flags);
410 __reeval_prio(which);
411 raw_spin_unlock_irqrestore(&which->lock, flags);
412}
413
414
415static void wakeup_litirqd_locked(struct klmirqd_info* which) 508static void wakeup_litirqd_locked(struct klmirqd_info* which)
416{ 509{
417 /* Interrupts are disabled: no need to stop preemption */ 510 /* Interrupts are disabled: no need to stop preemption */
418 if (which && which->klmirqd) 511 if (which && which->klmirqd)
419 { 512 {
420 __reeval_prio(which); /* configure the proper priority */
421
422 if(which->klmirqd->state != TASK_RUNNING) 513 if(which->klmirqd->state != TASK_RUNNING)
423 { 514 {
424 TRACE("%s: Waking up klmirqd: %s/%d\n", __FUNCTION__, 515 TRACE("%s: Waking up klmirqd: %s/%d\n", __FUNCTION__,
@@ -468,7 +559,7 @@ static void do_lit_tasklet(struct klmirqd_info* which,
468 list = list->next; 559 list = list->next;
469 560
470 /* execute tasklet if it has my priority and is free */ 561 /* execute tasklet if it has my priority and is free */
471 if ((t->owner == which->current_owner) && tasklet_trylock(t)) { 562 if (tasklet_trylock(t)) {
472 if (!atomic_read(&t->count)) { 563 if (!atomic_read(&t->count)) {
473 564
474 sched_trace_tasklet_begin(t->owner); 565 sched_trace_tasklet_begin(t->owner);
@@ -503,15 +594,14 @@ static void do_lit_tasklet(struct klmirqd_info* which,
503 594
504// returns 1 if priorities need to be changed to continue processing 595// returns 1 if priorities need to be changed to continue processing
505// pending tasklets. 596// pending tasklets.
506static int do_litirq(struct klmirqd_info* which) 597static void do_litirq(struct klmirqd_info* which)
507{ 598{
508 u32 pending; 599 u32 pending;
509 int resched = 0;
510 600
511 if(in_interrupt()) 601 if(in_interrupt())
512 { 602 {
513 TRACE("%s: exiting early: in interrupt context!\n", __FUNCTION__); 603 TRACE("%s: exiting early: in interrupt context!\n", __FUNCTION__);
514 return(0); 604 return;
515 } 605 }
516 606
517 if(which->klmirqd != current) 607 if(which->klmirqd != current)
@@ -519,59 +609,40 @@ static int do_litirq(struct klmirqd_info* which)
519 TRACE_CUR("%s: exiting early: thread/info mismatch! Running %s/%d but given %s/%d.\n", 609 TRACE_CUR("%s: exiting early: thread/info mismatch! Running %s/%d but given %s/%d.\n",
520 __FUNCTION__, current->comm, current->pid, 610 __FUNCTION__, current->comm, current->pid,
521 which->klmirqd->comm, which->klmirqd->pid); 611 which->klmirqd->comm, which->klmirqd->pid);
522 return(0); 612 return;
523 } 613 }
524 614
525 if(!is_realtime(current)) 615 if(!is_realtime(current))
526 { 616 {
527 TRACE_CUR("%s: exiting early: klmirqd is not real-time. Sched Policy = %d\n", 617 TRACE_CUR("%s: exiting early: klmirqd is not real-time. Sched Policy = %d\n",
528 __FUNCTION__, current->policy); 618 __FUNCTION__, current->policy);
529 return(0); 619 return;
530 } 620 }
531 621
532 622
533 /* We only handle tasklets & work objects, no need for RCU triggers? */ 623 /* We only handle tasklets & work objects, no need for RCU triggers? */
534 624
535 pending = litirq_pending(which); 625 pending = litirq_pending(which);
536 if(pending) 626 if(pending) {
537 {
538 /* extract the work to do and do it! */ 627 /* extract the work to do and do it! */
539 if(pending & LIT_TASKLET_HI) 628 if(pending & LIT_TASKLET_HI) {
540 {
541 TRACE_CUR("%s: Invoking HI tasklets.\n", __FUNCTION__); 629 TRACE_CUR("%s: Invoking HI tasklets.\n", __FUNCTION__);
542 do_lit_tasklet(which, &which->pending_tasklets_hi); 630 do_lit_tasklet(which, &which->pending_tasklets_hi);
543 resched = tasklet_ownership_change(which, LIT_TASKLET_HI);
544
545 if(resched)
546 {
547 TRACE_CUR("%s: HI tasklets of another owner remain. "
548 "Skipping any LOW tasklets.\n", __FUNCTION__);
549 }
550 } 631 }
551 632
552 if(!resched && (pending & LIT_TASKLET_LOW)) 633 if(pending & LIT_TASKLET_LOW) {
553 {
554 TRACE_CUR("%s: Invoking LOW tasklets.\n", __FUNCTION__); 634 TRACE_CUR("%s: Invoking LOW tasklets.\n", __FUNCTION__);
555 do_lit_tasklet(which, &which->pending_tasklets); 635 do_lit_tasklet(which, &which->pending_tasklets);
556 resched = tasklet_ownership_change(which, LIT_TASKLET_LOW);
557
558 if(resched)
559 {
560 TRACE_CUR("%s: LOW tasklets of another owner remain. "
561 "Skipping any work objects.\n", __FUNCTION__);
562 }
563 } 636 }
564 } 637 }
565
566 return(resched);
567} 638}
568 639
569 640
570static void do_work(struct klmirqd_info* which) 641static void do_work(struct klmirqd_info* which)
571{ 642{
572 unsigned long flags; 643 unsigned long flags;
573 work_func_t f;
574 struct work_struct* work; 644 struct work_struct* work;
645 work_func_t f;
575 646
576 // only execute one work-queue item to yield to tasklets. 647 // only execute one work-queue item to yield to tasklets.
577 // ...is this a good idea, or should we just batch them? 648 // ...is this a good idea, or should we just batch them?
@@ -594,125 +665,58 @@ static void do_work(struct klmirqd_info* which)
594 raw_spin_unlock_irqrestore(&which->lock, flags); 665 raw_spin_unlock_irqrestore(&which->lock, flags);
595 666
596 667
668 TRACE_CUR("%s: Invoking work object.\n", __FUNCTION__);
669 // do the work!
670 work_clear_pending(work);
671 f = work->func;
672 f(work); /* can't touch 'work' after this point,
673 the user may have freed it. */
597 674
598 /* safe to read current_owner outside of lock since only this thread 675 atomic_dec(&which->num_work_pending);
599 may write to the pointer. */
600 if(work->owner == which->current_owner)
601 {
602 TRACE_CUR("%s: Invoking work object.\n", __FUNCTION__);
603 // do the work!
604 work_clear_pending(work);
605 f = work->func;
606 f(work); /* can't touch 'work' after this point,
607 the user may have freed it. */
608
609 atomic_dec(&which->num_work_pending);
610 }
611 else
612 {
613 TRACE_CUR("%s: Could not invoke work object. Requeuing.\n",
614 __FUNCTION__);
615 ___litmus_schedule_work(work, which, 0);
616 }
617 676
618no_work: 677no_work:
619 return; 678 return;
620} 679}
621 680
622 681
623static int set_litmus_daemon_sched(void)
624{
625 /* set up a daemon job that will never complete.
626 it should only ever run on behalf of another
627 real-time task.
628
629 TODO: Transition to a new job whenever a
630 new tasklet is handled */
631
632 int ret = 0;
633
634 struct rt_task tp = {
635 .exec_cost = 0,
636 .period = 1000000000, /* dummy 1 second period */
637 .phase = 0,
638 .cpu = task_cpu(current),
639 .budget_policy = NO_ENFORCEMENT,
640 .cls = RT_CLASS_BEST_EFFORT
641 };
642
643 struct sched_param param = { .sched_priority = 0};
644
645
646 /* set task params, mark as proxy thread, and init other data */
647 tsk_rt(current)->task_params = tp;
648 tsk_rt(current)->is_proxy_thread = 1;
649 tsk_rt(current)->cur_klmirqd = NULL;
650 mutex_init(&tsk_rt(current)->klmirqd_sem);
651 atomic_set(&tsk_rt(current)->klmirqd_sem_stat, NOT_HELD);
652
653 /* inform the OS we're SCHED_LITMUS --
654 sched_setscheduler_nocheck() calls litmus_admit_task(). */
655 sched_setscheduler_nocheck(current, SCHED_LITMUS, &param);
656
657 return ret;
658}
659
660static void enter_execution_phase(struct klmirqd_info* which,
661 struct mutex* sem,
662 struct task_struct* t)
663{
664 TRACE_CUR("%s: Trying to enter execution phase. "
665 "Acquiring semaphore of %s/%d\n", __FUNCTION__,
666 t->comm, t->pid);
667 down_and_set_stat(current, HELD, sem);
668 TRACE_CUR("%s: Execution phase entered! "
669 "Acquired semaphore of %s/%d\n", __FUNCTION__,
670 t->comm, t->pid);
671}
672
673static void exit_execution_phase(struct klmirqd_info* which,
674 struct mutex* sem,
675 struct task_struct* t)
676{
677 TRACE_CUR("%s: Exiting execution phase. "
678 "Releasing semaphore of %s/%d\n", __FUNCTION__,
679 t->comm, t->pid);
680 if(atomic_read(&tsk_rt(current)->klmirqd_sem_stat) == HELD)
681 {
682 up_and_set_stat(current, NOT_HELD, sem);
683 TRACE_CUR("%s: Execution phase exited! "
684 "Released semaphore of %s/%d\n", __FUNCTION__,
685 t->comm, t->pid);
686 }
687 else
688 {
689 TRACE_CUR("%s: COULDN'T RELEASE SEMAPHORE BECAUSE ONE IS NOT HELD!\n", __FUNCTION__);
690 }
691}
692 682
693/* main loop for klitsoftirqd */ 683/* main loop for klitsoftirqd */
694static int run_klmirqd(void* unused) 684static int run_klmirqd(void* callback)
695{ 685{
696 struct klmirqd_info* which = &klmirqds[klmirqd_id(current)]; 686 int retval = 0;
697 struct mutex* sem; 687 struct klmirqd_info* info = NULL;
698 struct task_struct* owner; 688 klmirqd_callback_t* cb = (klmirqd_callback_t*)(callback);
699 689
700 int rt_status = set_litmus_daemon_sched(); 690 retval = set_litmus_daemon_sched(current);
701 691 if (retval != 0) {
702 if(rt_status != 0)
703 {
704 TRACE_CUR("%s: Failed to transition to rt-task.\n", __FUNCTION__); 692 TRACE_CUR("%s: Failed to transition to rt-task.\n", __FUNCTION__);
705 goto rt_failed; 693 goto failed;
706 } 694 }
707 695
708 atomic_inc(&num_ready_klmirqds); 696 retval = register_klmirqd(current);
697 if (retval != 0) {
698 TRACE_CUR("%s: Failed to become a klmirqd thread.\n", __FUNCTION__);
699 goto failed;
700 }
701
702 if (cb && cb->func) {
703 retval = cb->func(cb->arg);
704 if (retval != 0) {
705 TRACE_CUR("%s: klmirqd callback reported failure. retval = %d\n", __FUNCTION__, retval);
706 goto failed_unregister;
707 }
708 }
709
710 /* enter the interrupt handling workloop */
711
712 info = tsk_rt(current)->klmirqd_info;
709 713
710 set_current_state(TASK_INTERRUPTIBLE); 714 set_current_state(TASK_INTERRUPTIBLE);
711 715
712 while (!kthread_should_stop()) 716 while (!kthread_should_stop())
713 { 717 {
714 preempt_disable(); 718 preempt_disable();
715 if (!litirq_pending(which)) 719 if (!litirq_pending(info))
716 { 720 {
717 /* sleep for work */ 721 /* sleep for work */
718 TRACE_CUR("%s: No more tasklets or work objects. Going to sleep.\n", 722 TRACE_CUR("%s: No more tasklets or work objects. Going to sleep.\n",
@@ -731,17 +735,10 @@ static int run_klmirqd(void* unused)
731 735
732 __set_current_state(TASK_RUNNING); 736 __set_current_state(TASK_RUNNING);
733 737
734 while (litirq_pending_and_sem_and_owner(which, &sem, &owner)) 738 while (litirq_pending(info))
735 { 739 {
736 int needs_resched = 0;
737
738 preempt_enable_no_resched(); 740 preempt_enable_no_resched();
739 741
740 BUG_ON(sem == NULL);
741
742 // wait to enter execution phase; wait for 'current_owner' to block.
743 enter_execution_phase(which, sem, owner);
744
745 if(kthread_should_stop()) 742 if(kthread_should_stop())
746 { 743 {
747 TRACE_CUR("%s:%d: Signaled to terminate.\n", __FUNCTION__, __LINE__); 744 TRACE_CUR("%s:%d: Signaled to terminate.\n", __FUNCTION__, __LINE__);
@@ -753,36 +750,23 @@ static int run_klmirqd(void* unused)
753 /* Double check that there's still pending work and the owner hasn't 750 /* Double check that there's still pending work and the owner hasn't
754 * changed. Pending items may have been flushed while we were sleeping. 751 * changed. Pending items may have been flushed while we were sleeping.
755 */ 752 */
756 if(litirq_pending_with_owner(which, owner)) 753 if(litirq_pending(info))
757 { 754 {
758 TRACE_CUR("%s: Executing tasklets and/or work objects.\n", 755 TRACE_CUR("%s: Executing tasklets and/or work objects.\n",
759 __FUNCTION__); 756 __FUNCTION__);
760 757
761 needs_resched = do_litirq(which); 758 do_litirq(info);
762 759
763 preempt_enable_no_resched(); 760 preempt_enable_no_resched();
764 761
765 // work objects are preemptible. 762 // work objects are preemptible.
766 if(!needs_resched) 763 do_work(info);
767 {
768 do_work(which);
769 }
770
771 // exit execution phase.
772 exit_execution_phase(which, sem, owner);
773
774 TRACE_CUR("%s: Setting up next priority.\n", __FUNCTION__);
775 reeval_prio(which); /* check if we need to change priority here */
776 } 764 }
777 else 765 else
778 { 766 {
779 TRACE_CUR("%s: Pending work was flushed! Prev owner was %s/%d\n", 767 TRACE_CUR("%s: Pending work was flushed!\n", __FUNCTION__);
780 __FUNCTION__,
781 owner->comm, owner->pid);
782 preempt_enable_no_resched();
783 768
784 // exit execution phase. 769 preempt_enable_no_resched();
785 exit_execution_phase(which, sem, owner);
786 } 770 }
787 771
788 cond_resched(); 772 cond_resched();
@@ -793,183 +777,39 @@ static int run_klmirqd(void* unused)
793 } 777 }
794 __set_current_state(TASK_RUNNING); 778 __set_current_state(TASK_RUNNING);
795 779
796 atomic_dec(&num_ready_klmirqds); 780failed_unregister:
781 /* remove our registration from klmirqd */
782 unregister_klmirqd(current);
797 783
798rt_failed: 784failed:
799 litmus_exit_task(current); 785 litmus_exit_task(current);
800 786
801 return rt_status; 787 return retval;
802} 788}
803 789
804 790
805struct klmirqd_launch_data 791void flush_pending(struct task_struct* tsk)
806{
807 int* cpu_affinity;
808 struct work_struct work;
809};
810
811/* executed by a kworker from workqueues */
812static void launch_klmirqd(struct work_struct *work)
813{ 792{
814 int i; 793 unsigned long flags;
815 794 struct tasklet_struct *list;
816 struct klmirqd_launch_data* launch_data = 795 u32 work_flushed = 0;
817 container_of(work, struct klmirqd_launch_data, work);
818
819 TRACE("%s: Creating %d klmirqds\n", __FUNCTION__, NR_LITMUS_SOFTIRQD);
820
821 /* create the daemon threads */
822 for(i = 0; i < NR_LITMUS_SOFTIRQD; ++i)
823 {
824 if(launch_data->cpu_affinity)
825 {
826 klmirqds[i].klmirqd =
827 kthread_create(
828 run_klmirqd,
829 /* treat the affinity as a pointer, we'll cast it back later */
830 (void*)(long long)launch_data->cpu_affinity[i],
831 "klmirqd_th%d/%d",
832 i,
833 launch_data->cpu_affinity[i]);
834
835 /* litmus will put is in the right cluster. */
836 kthread_bind(klmirqds[i].klmirqd, launch_data->cpu_affinity[i]);
837 }
838 else
839 {
840 klmirqds[i].klmirqd =
841 kthread_create(
842 run_klmirqd,
843 /* treat the affinity as a pointer, we'll cast it back later */
844 (void*)(long long)(-1),
845 "klmirqd_th%d",
846 i);
847 }
848 }
849
850 TRACE("%s: Launching %d klmirqds\n", __FUNCTION__, NR_LITMUS_SOFTIRQD);
851
852 /* unleash the daemons */
853 for(i = 0; i < NR_LITMUS_SOFTIRQD; ++i)
854 {
855 wake_up_process(klmirqds[i].klmirqd);
856 }
857
858 if(launch_data->cpu_affinity)
859 kfree(launch_data->cpu_affinity);
860 kfree(launch_data);
861}
862 796
797 struct klmirqd_info *which;
863 798
864void spawn_klmirqd(int* affinity) 799 if (!tsk_rt(tsk)->is_interrupt_thread) {
865{ 800 TRACE("%s/%d is not a proxy thread\n", tsk->comm, tsk->pid);
866 int i; 801 WARN_ON(1);
867 struct klmirqd_launch_data* delayed_launch;
868
869 if(atomic_read(&num_ready_klmirqds) != 0)
870 {
871 TRACE("%s: At least one klmirqd is already running! Need to call kill_klmirqd()?\n");
872 return; 802 return;
873 } 803 }
874 804
875 /* init the tasklet & work queues */ 805 which = tsk_rt(tsk)->klmirqd_info;
876 for(i = 0; i < NR_LITMUS_SOFTIRQD; ++i) 806 if (!which) {
877 { 807 TRACE("%s/%d is not a klmirqd thread!\n", tsk->comm, tsk->pid);
878 klmirqds[i].terminating = 0; 808 WARN_ON(1);
879 klmirqds[i].pending = 0; 809 return;
880
881 klmirqds[i].num_hi_pending.counter = 0;
882 klmirqds[i].num_low_pending.counter = 0;
883 klmirqds[i].num_work_pending.counter = 0;
884
885 klmirqds[i].pending_tasklets_hi.head = NULL;
886 klmirqds[i].pending_tasklets_hi.tail = &klmirqds[i].pending_tasklets_hi.head;
887
888 klmirqds[i].pending_tasklets.head = NULL;
889 klmirqds[i].pending_tasklets.tail = &klmirqds[i].pending_tasklets.head;
890
891 INIT_LIST_HEAD(&klmirqds[i].worklist);
892
893 raw_spin_lock_init(&klmirqds[i].lock);
894 }
895
896 /* wait to flush the initializations to memory since other threads
897 will access it. */
898 mb();
899
900 /* tell a work queue to launch the threads. we can't make scheduling
901 calls since we're in an atomic state. */
902 TRACE("%s: Setting callback up to launch klmirqds\n", __FUNCTION__);
903 delayed_launch = kmalloc(sizeof(struct klmirqd_launch_data), GFP_ATOMIC);
904 if(affinity)
905 {
906 delayed_launch->cpu_affinity =
907 kmalloc(sizeof(int)*NR_LITMUS_SOFTIRQD, GFP_ATOMIC);
908
909 memcpy(delayed_launch->cpu_affinity, affinity,
910 sizeof(int)*NR_LITMUS_SOFTIRQD);
911 }
912 else
913 {
914 delayed_launch->cpu_affinity = NULL;
915 }
916 INIT_WORK(&delayed_launch->work, launch_klmirqd);
917 schedule_work(&delayed_launch->work);
918}
919
920
921void kill_klmirqd(void)
922{
923 if(!klmirqd_is_dead())
924 {
925 int i;
926
927 TRACE("%s: Killing %d klmirqds\n", __FUNCTION__, NR_LITMUS_SOFTIRQD);
928
929 for(i = 0; i < NR_LITMUS_SOFTIRQD; ++i)
930 {
931 if(klmirqds[i].terminating != 1)
932 {
933 klmirqds[i].terminating = 1;
934 mb(); /* just to be sure? */
935 flush_pending(klmirqds[i].klmirqd, NULL);
936
937 /* signal termination */
938 kthread_stop(klmirqds[i].klmirqd);
939 }
940 }
941 } 810 }
942}
943 811
944 812
945int klmirqd_is_ready(void)
946{
947 return(atomic_read(&num_ready_klmirqds) == NR_LITMUS_SOFTIRQD);
948}
949
950int klmirqd_is_dead(void)
951{
952 return(atomic_read(&num_ready_klmirqds) == 0);
953}
954
955
956struct task_struct* get_klmirqd(unsigned int k_id)
957{
958 return(klmirqds[k_id].klmirqd);
959}
960
961
962void flush_pending(struct task_struct* klmirqd_thread,
963 struct task_struct* owner)
964{
965 unsigned int k_id = klmirqd_id(klmirqd_thread);
966 struct klmirqd_info *which = &klmirqds[k_id];
967
968 unsigned long flags;
969 struct tasklet_struct *list;
970
971 u32 work_flushed = 0;
972
973 raw_spin_lock_irqsave(&which->lock, flags); 813 raw_spin_lock_irqsave(&which->lock, flags);
974 814
975 //__dump_state(which, "flush_pending: before"); 815 //__dump_state(which, "flush_pending: before");
@@ -990,35 +830,27 @@ void flush_pending(struct task_struct* klmirqd_thread,
990 struct tasklet_struct *t = list; 830 struct tasklet_struct *t = list;
991 list = list->next; 831 list = list->next;
992 832
993 if(likely((t->owner == owner) || (owner == NULL))) 833 if(unlikely(!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state)))
994 { 834 {
995 if(unlikely(!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))) 835 BUG();
996 { 836 }
997 BUG();
998 }
999 837
1000 work_flushed |= LIT_TASKLET_HI; 838 work_flushed |= LIT_TASKLET_HI;
1001 839
1002 t->owner = NULL; 840 t->owner = NULL;
1003 841
1004 // WTF? 842 // WTF?
1005 if(!test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) 843 if(!test_and_set_bit(TASKLET_STATE_SCHED, &t->state))
1006 { 844 {
1007 atomic_dec(&which->num_hi_pending); 845 atomic_dec(&which->num_hi_pending);
1008 ___tasklet_hi_schedule(t); 846 ___tasklet_hi_schedule(t);
1009 }
1010 else
1011 {
1012 TRACE("%s: dropped hi tasklet??\n", __FUNCTION__);
1013 BUG();
1014 }
1015 } 847 }
1016 else 848 else
1017 { 849 {
1018 TRACE("%s: Could not flush a HI tasklet.\n", __FUNCTION__); 850 TRACE("%s: dropped hi tasklet??\n", __FUNCTION__);
1019 // put back on queue. 851 BUG();
1020 ___litmus_tasklet_hi_schedule(t, which, 0);
1021 } 852 }
853
1022 } 854 }
1023 } 855 }
1024 856
@@ -1038,34 +870,25 @@ void flush_pending(struct task_struct* klmirqd_thread,
1038 struct tasklet_struct *t = list; 870 struct tasklet_struct *t = list;
1039 list = list->next; 871 list = list->next;
1040 872
1041 if(likely((t->owner == owner) || (owner == NULL))) 873 if(unlikely(!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state)))
1042 { 874 {
1043 if(unlikely(!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))) 875 BUG();
1044 { 876 }
1045 BUG();
1046 }
1047 877
1048 work_flushed |= LIT_TASKLET_LOW; 878 work_flushed |= LIT_TASKLET_LOW;
1049 879
1050 t->owner = NULL; 880 t->owner = NULL;
1051 sched_trace_tasklet_end(owner, 1ul); 881// sched_trace_tasklet_end(owner, 1ul);
1052 882
1053 if(!test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) 883 if(!test_and_set_bit(TASKLET_STATE_SCHED, &t->state))
1054 { 884 {
1055 atomic_dec(&which->num_low_pending); 885 atomic_dec(&which->num_low_pending);
1056 ___tasklet_schedule(t); 886 ___tasklet_schedule(t);
1057 }
1058 else
1059 {
1060 TRACE("%s: dropped tasklet??\n", __FUNCTION__);
1061 BUG();
1062 }
1063 } 887 }
1064 else 888 else
1065 { 889 {
1066 TRACE("%s: Could not flush a LOW tasklet.\n", __FUNCTION__); 890 TRACE("%s: dropped tasklet??\n", __FUNCTION__);
1067 // put back on queue 891 BUG();
1068 ___litmus_tasklet_schedule(t, which, 0);
1069 } 892 }
1070 } 893 }
1071 } 894 }
@@ -1083,21 +906,12 @@ void flush_pending(struct task_struct* klmirqd_thread,
1083 list_first_entry(&which->worklist, struct work_struct, entry); 906 list_first_entry(&which->worklist, struct work_struct, entry);
1084 list_del_init(&work->entry); 907 list_del_init(&work->entry);
1085 908
1086 if(likely((work->owner == owner) || (owner == NULL))) 909 work_flushed |= LIT_WORK;
1087 { 910 atomic_dec(&which->num_work_pending);
1088 work_flushed |= LIT_WORK;
1089 atomic_dec(&which->num_work_pending);
1090 911
1091 work->owner = NULL; 912 work->owner = NULL;
1092 sched_trace_work_end(owner, current, 1ul); 913// sched_trace_work_end(owner, current, 1ul);
1093 __schedule_work(work); 914 __schedule_work(work);
1094 }
1095 else
1096 {
1097 TRACE("%s: Could not flush a work object.\n", __FUNCTION__);
1098 // put back on queue
1099 ___litmus_schedule_work(work, which, 0);
1100 }
1101 } 915 }
1102 } 916 }
1103 917
@@ -1106,22 +920,6 @@ void flush_pending(struct task_struct* klmirqd_thread,
1106 920
1107 mb(); /* commit changes to pending flags */ 921 mb(); /* commit changes to pending flags */
1108 922
1109 /* reset the scheduling priority */
1110 if(work_flushed)
1111 {
1112 __reeval_prio(which);
1113
1114 /* Try to offload flushed tasklets to Linux's ksoftirqd. */
1115 if(work_flushed & (LIT_TASKLET_LOW | LIT_TASKLET_HI))
1116 {
1117 wakeup_softirqd();
1118 }
1119 }
1120 else
1121 {
1122 TRACE_CUR("%s: no work flushed, so __reeval_prio() skipped\n", __FUNCTION__);
1123 }
1124
1125 raw_spin_unlock_irqrestore(&which->lock, flags); 923 raw_spin_unlock_irqrestore(&which->lock, flags);
1126} 924}
1127 925
@@ -1161,39 +959,27 @@ static void ___litmus_tasklet_schedule(struct tasklet_struct *t,
1161 raw_spin_unlock_irqrestore(&which->lock, flags); 959 raw_spin_unlock_irqrestore(&which->lock, flags);
1162} 960}
1163 961
1164int __litmus_tasklet_schedule(struct tasklet_struct *t, unsigned int k_id) 962
963int __litmus_tasklet_schedule(struct tasklet_struct *t, struct task_struct* klmirqd_thread)
1165{ 964{
1166 int ret = 0; /* assume failure */ 965 int ret = 0; /* assume failure */
1167 if(unlikely((t->owner == NULL) || !is_realtime(t->owner))) 966 struct klmirqd_info* info;
1168 {
1169 TRACE("%s: No owner associated with this tasklet!\n", __FUNCTION__);
1170 BUG();
1171 }
1172 967
1173 if(unlikely(k_id >= NR_LITMUS_SOFTIRQD)) 968 if (unlikely(!is_realtime(klmirqd_thread) ||
1174 { 969 !tsk_rt(klmirqd_thread)->is_interrupt_thread ||
1175 TRACE("%s: No klmirqd_th%d!\n", __FUNCTION__, k_id); 970 !tsk_rt(klmirqd_thread)->klmirqd_info)) {
1176 BUG(); 971 TRACE("%s: %s/%d can't handle tasklets\n", klmirqd_thread->comm, klmirqd_thread->pid);
1177 } 972 return ret;
973 }
1178 974
1179 if(likely(!klmirqds[k_id].terminating)) 975 info = tsk_rt(klmirqd_thread)->klmirqd_info;
1180 {
1181 /* Can't accept tasklets while we're processing a workqueue
1182 because they're handled by the same thread. This case is
1183 very RARE.
1184 976
1185 TODO: Use a separate thread for work objects!!!!!! 977 if (likely(!info->terminating)) {
1186 */ 978 ret = 1;
1187 if(likely(atomic_read(&klmirqds[k_id].num_work_pending) == 0)) 979 ___litmus_tasklet_schedule(t, info, 1);
1188 { 980 }
1189 ret = 1; 981 else {
1190 ___litmus_tasklet_schedule(t, &klmirqds[k_id], 1); 982 TRACE("%s: Tasklet rejected because %s/%d is terminating\n", klmirqd_thread->comm, klmirqd_thread->pid);
1191 }
1192 else
1193 {
1194 TRACE("%s: rejected tasklet because of pending work.\n",
1195 __FUNCTION__);
1196 }
1197 } 983 }
1198 return(ret); 984 return(ret);
1199} 985}
@@ -1230,100 +1016,77 @@ static void ___litmus_tasklet_hi_schedule(struct tasklet_struct *t,
1230 raw_spin_unlock_irqrestore(&which->lock, flags); 1016 raw_spin_unlock_irqrestore(&which->lock, flags);
1231} 1017}
1232 1018
1233int __litmus_tasklet_hi_schedule(struct tasklet_struct *t, unsigned int k_id) 1019int __litmus_tasklet_hi_schedule(struct tasklet_struct *t, struct task_struct* klmirqd_thread)
1234{ 1020{
1235 int ret = 0; /* assume failure */ 1021 int ret = 0; /* assume failure */
1236 if(unlikely((t->owner == NULL) || !is_realtime(t->owner))) 1022 struct klmirqd_info* info;
1237 {
1238 TRACE("%s: No owner associated with this tasklet!\n", __FUNCTION__);
1239 BUG();
1240 }
1241 1023
1242 if(unlikely(k_id >= NR_LITMUS_SOFTIRQD)) 1024 if (unlikely(!is_realtime(klmirqd_thread) ||
1243 { 1025 !tsk_rt(klmirqd_thread)->is_interrupt_thread ||
1244 TRACE("%s: No klmirqd_th%d!\n", __FUNCTION__, k_id); 1026 !tsk_rt(klmirqd_thread)->klmirqd_info)) {
1245 BUG(); 1027 TRACE("%s: %s/%d can't handle tasklets\n", klmirqd_thread->comm, klmirqd_thread->pid);
1246 } 1028 return ret;
1029 }
1247 1030
1248 if(unlikely(!klmirqd_is_ready())) 1031 info = tsk_rt(klmirqd_thread)->klmirqd_info;
1249 {
1250 TRACE("%s: klmirqd is not ready!\n", __FUNCTION__, k_id);
1251 BUG();
1252 }
1253 1032
1254 if(likely(!klmirqds[k_id].terminating)) 1033 if (likely(!info->terminating)) {
1255 { 1034 ret = 1;
1256 if(likely(atomic_read(&klmirqds[k_id].num_work_pending) == 0)) 1035 ___litmus_tasklet_hi_schedule(t, info, 1);
1257 {
1258 ret = 1;
1259 ___litmus_tasklet_hi_schedule(t, &klmirqds[k_id], 1);
1260 }
1261 else
1262 {
1263 TRACE("%s: rejected tasklet because of pending work.\n",
1264 __FUNCTION__);
1265 }
1266 } 1036 }
1037 else {
1038 TRACE("%s: Tasklet rejected because %s/%d is terminating\n", klmirqd_thread->comm, klmirqd_thread->pid);
1039 }
1040
1267 return(ret); 1041 return(ret);
1268} 1042}
1269 1043
1270EXPORT_SYMBOL(__litmus_tasklet_hi_schedule); 1044EXPORT_SYMBOL(__litmus_tasklet_hi_schedule);
1271 1045
1272 1046
1273int __litmus_tasklet_hi_schedule_first(struct tasklet_struct *t, unsigned int k_id) 1047int __litmus_tasklet_hi_schedule_first(struct tasklet_struct *t, struct task_struct* klmirqd_thread)
1274{ 1048{
1275 int ret = 0; /* assume failure */ 1049 int ret = 0; /* assume failure */
1276 u32 old_pending; 1050 u32 old_pending;
1051 struct klmirqd_info* info;
1277 1052
1278 BUG_ON(!irqs_disabled()); 1053 BUG_ON(!irqs_disabled());
1279 1054
1280 if(unlikely((t->owner == NULL) || !is_realtime(t->owner))) 1055 if (unlikely(!is_realtime(klmirqd_thread) ||
1281 { 1056 !tsk_rt(klmirqd_thread)->is_interrupt_thread ||
1282 TRACE("%s: No owner associated with this tasklet!\n", __FUNCTION__); 1057 !tsk_rt(klmirqd_thread)->klmirqd_info)) {
1283 BUG(); 1058 TRACE("%s: %s/%d can't handle tasklets\n", klmirqd_thread->comm, klmirqd_thread->pid);
1284 } 1059 return ret;
1060 }
1285 1061
1286 if(unlikely(k_id >= NR_LITMUS_SOFTIRQD)) 1062 info = tsk_rt(klmirqd_thread)->klmirqd_info;
1287 {
1288 TRACE("%s: No klmirqd_th%u!\n", __FUNCTION__, k_id);
1289 BUG();
1290 }
1291 1063
1292 if(unlikely(!klmirqd_is_ready())) 1064 if (likely(!info->terminating)) {
1293 {
1294 TRACE("%s: klmirqd is not ready!\n", __FUNCTION__, k_id);
1295 BUG();
1296 }
1297 1065
1298 if(likely(!klmirqds[k_id].terminating)) 1066 raw_spin_lock(&info->lock);
1299 {
1300 raw_spin_lock(&klmirqds[k_id].lock);
1301 1067
1302 if(likely(atomic_read(&klmirqds[k_id].num_work_pending) == 0)) 1068 ret = 1; // success!
1303 {
1304 ret = 1; // success!
1305 1069
1306 t->next = klmirqds[k_id].pending_tasklets_hi.head; 1070 t->next = info->pending_tasklets_hi.head;
1307 klmirqds[k_id].pending_tasklets_hi.head = t; 1071 info->pending_tasklets_hi.head = t;
1308 1072
1309 old_pending = klmirqds[k_id].pending; 1073 old_pending = info->pending;
1310 klmirqds[k_id].pending |= LIT_TASKLET_HI; 1074 info->pending |= LIT_TASKLET_HI;
1311 1075
1312 atomic_inc(&klmirqds[k_id].num_hi_pending); 1076 atomic_inc(&info->num_hi_pending);
1313 1077
1314 mb(); 1078 mb();
1315 1079
1316 if(!old_pending) 1080 if(!old_pending) {
1317 wakeup_litirqd_locked(&klmirqds[k_id]); /* wake up the klmirqd */ 1081 wakeup_litirqd_locked(info); /* wake up the klmirqd */
1318 }
1319 else
1320 {
1321 TRACE("%s: rejected tasklet because of pending work.\n",
1322 __FUNCTION__);
1323 } 1082 }
1324 1083
1325 raw_spin_unlock(&klmirqds[k_id].lock); 1084 raw_spin_unlock(&info->lock);
1326 } 1085 }
1086 else {
1087 TRACE("%s: Tasklet rejected because %s/%d is terminating\n", klmirqd_thread->comm, klmirqd_thread->pid);
1088 }
1089
1327 return(ret); 1090 return(ret);
1328} 1091}
1329 1092
@@ -1358,225 +1121,30 @@ static void ___litmus_schedule_work(struct work_struct *w,
1358 raw_spin_unlock_irqrestore(&which->lock, flags); 1121 raw_spin_unlock_irqrestore(&which->lock, flags);
1359} 1122}
1360 1123
1361int __litmus_schedule_work(struct work_struct *w, unsigned int k_id) 1124int __litmus_schedule_work(struct work_struct *w, struct task_struct* klmirqd_thread)
1362{ 1125{
1363 int ret = 1; /* assume success */ 1126 int ret = 1; /* assume success */
1364 if(unlikely(w->owner == NULL) || !is_realtime(w->owner)) 1127 struct klmirqd_info* info;
1365 {
1366 TRACE("%s: No owner associated with this work object!\n", __FUNCTION__);
1367 BUG();
1368 }
1369
1370 if(unlikely(k_id >= NR_LITMUS_SOFTIRQD))
1371 {
1372 TRACE("%s: No klmirqd_th%u!\n", k_id);
1373 BUG();
1374 }
1375
1376 if(unlikely(!klmirqd_is_ready()))
1377 {
1378 TRACE("%s: klmirqd is not ready!\n", __FUNCTION__, k_id);
1379 BUG();
1380 }
1381
1382 if(likely(!klmirqds[k_id].terminating))
1383 ___litmus_schedule_work(w, &klmirqds[k_id], 1);
1384 else
1385 ret = 0;
1386 return(ret);
1387}
1388EXPORT_SYMBOL(__litmus_schedule_work);
1389
1390
1391static int set_klmirqd_sem_status(unsigned long stat)
1392{
1393 TRACE_CUR("SETTING STATUS FROM %d TO %d\n",
1394 atomic_read(&tsk_rt(current)->klmirqd_sem_stat),
1395 stat);
1396 atomic_set(&tsk_rt(current)->klmirqd_sem_stat, stat);
1397 //mb();
1398
1399 return(0);
1400}
1401
1402static int set_klmirqd_sem_status_if_not_held(unsigned long stat)
1403{
1404 if(atomic_read(&tsk_rt(current)->klmirqd_sem_stat) != HELD)
1405 {
1406 return(set_klmirqd_sem_status(stat));
1407 }
1408 return(-1);
1409}
1410
1411
1412void __down_and_reset_and_set_stat(struct task_struct* t,
1413 enum klmirqd_sem_status to_reset,
1414 enum klmirqd_sem_status to_set,
1415 struct mutex* sem)
1416{
1417#if 0
1418 struct rt_param* param = container_of(sem, struct rt_param, klmirqd_sem);
1419 struct task_struct* task = container_of(param, struct task_struct, rt_param);
1420
1421 TRACE_CUR("%s: entered. Locking semaphore of %s/%d\n",
1422 __FUNCTION__, task->comm, task->pid);
1423#endif
1424
1425 mutex_lock_sfx(sem,
1426 set_klmirqd_sem_status_if_not_held, to_reset,
1427 set_klmirqd_sem_status, to_set);
1428#if 0
1429 TRACE_CUR("%s: exiting. Have semaphore of %s/%d\n",
1430 __FUNCTION__, task->comm, task->pid);
1431#endif
1432}
1433
1434void down_and_set_stat(struct task_struct* t,
1435 enum klmirqd_sem_status to_set,
1436 struct mutex* sem)
1437{
1438#if 0
1439 struct rt_param* param = container_of(sem, struct rt_param, klmirqd_sem);
1440 struct task_struct* task = container_of(param, struct task_struct, rt_param);
1441
1442 TRACE_CUR("%s: entered. Locking semaphore of %s/%d\n",
1443 __FUNCTION__, task->comm, task->pid);
1444#endif
1445
1446 mutex_lock_sfx(sem,
1447 NULL, 0,
1448 set_klmirqd_sem_status, to_set);
1449
1450#if 0
1451 TRACE_CUR("%s: exiting. Have semaphore of %s/%d\n",
1452 __FUNCTION__, task->comm, task->pid);
1453#endif
1454}
1455
1456
1457void up_and_set_stat(struct task_struct* t,
1458 enum klmirqd_sem_status to_set,
1459 struct mutex* sem)
1460{
1461#if 0
1462 struct rt_param* param = container_of(sem, struct rt_param, klmirqd_sem);
1463 struct task_struct* task = container_of(param, struct task_struct, rt_param);
1464
1465 TRACE_CUR("%s: entered. Unlocking semaphore of %s/%d\n",
1466 __FUNCTION__,
1467 task->comm, task->pid);
1468#endif
1469
1470 mutex_unlock_sfx(sem, NULL, 0,
1471 set_klmirqd_sem_status, to_set);
1472
1473#if 0
1474 TRACE_CUR("%s: exiting. Unlocked semaphore of %s/%d\n",
1475 __FUNCTION__,
1476 task->comm, task->pid);
1477#endif
1478}
1479
1480
1481
1482void release_klmirqd_lock(struct task_struct* t)
1483{
1484 if(is_realtime(t) && (atomic_read(&tsk_rt(t)->klmirqd_sem_stat) == HELD))
1485 {
1486 struct mutex* sem;
1487 struct task_struct* owner = t;
1488
1489 if(t->state == TASK_RUNNING)
1490 {
1491 TRACE_TASK(t, "NOT giving up klmirqd_sem because we're not blocked!\n");
1492 return;
1493 }
1494 1128
1495 if(likely(!tsk_rt(t)->is_proxy_thread)) 1129 if (unlikely(!is_realtime(klmirqd_thread) ||
1496 { 1130 !tsk_rt(klmirqd_thread)->is_interrupt_thread ||
1497 sem = &tsk_rt(t)->klmirqd_sem; 1131 !tsk_rt(klmirqd_thread)->klmirqd_info)) {
1498 } 1132 TRACE("%s: %s/%d can't handle work items\n", klmirqd_thread->comm, klmirqd_thread->pid);
1499 else 1133 return ret;
1500 {
1501 unsigned int k_id = klmirqd_id(t);
1502 owner = klmirqds[k_id].current_owner;
1503
1504 BUG_ON(t != klmirqds[k_id].klmirqd);
1505
1506 if(likely(owner))
1507 {
1508 sem = &tsk_rt(owner)->klmirqd_sem;
1509 }
1510 else
1511 {
1512 BUG();
1513
1514 // We had the rug pulled out from under us. Abort attempt
1515 // to reacquire the lock since our client no longer needs us.
1516 TRACE_CUR("HUH?! How did this happen?\n");
1517 atomic_set(&tsk_rt(t)->klmirqd_sem_stat, NOT_HELD);
1518 return;
1519 }
1520 }
1521
1522 //TRACE_CUR("Releasing semaphore of %s/%d...\n", owner->comm, owner->pid);
1523 up_and_set_stat(t, NEED_TO_REACQUIRE, sem);
1524 //TRACE_CUR("Semaphore of %s/%d released!\n", owner->comm, owner->pid);
1525 }
1526 /*
1527 else if(is_realtime(t))
1528 {
1529 TRACE_CUR("%s: Nothing to do. Stat = %d\n", __FUNCTION__, tsk_rt(t)->klmirqd_sem_stat);
1530 } 1134 }
1531 */
1532}
1533 1135
1534int reacquire_klmirqd_lock(struct task_struct* t) 1136 info = tsk_rt(klmirqd_thread)->klmirqd_info;
1535{
1536 int ret = 0;
1537
1538 if(is_realtime(t) && (atomic_read(&tsk_rt(t)->klmirqd_sem_stat) == NEED_TO_REACQUIRE))
1539 {
1540 struct mutex* sem;
1541 struct task_struct* owner = t;
1542
1543 if(likely(!tsk_rt(t)->is_proxy_thread))
1544 {
1545 sem = &tsk_rt(t)->klmirqd_sem;
1546 }
1547 else
1548 {
1549 unsigned int k_id = klmirqd_id(t);
1550 //struct task_struct* owner = klmirqds[k_id].current_owner;
1551 owner = klmirqds[k_id].current_owner;
1552
1553 BUG_ON(t != klmirqds[k_id].klmirqd);
1554 1137
1555 if(likely(owner))
1556 {
1557 sem = &tsk_rt(owner)->klmirqd_sem;
1558 }
1559 else
1560 {
1561 // We had the rug pulled out from under us. Abort attempt
1562 // to reacquire the lock since our client no longer needs us.
1563 TRACE_CUR("No longer needs to reacquire klmirqd_sem!\n");
1564 atomic_set(&tsk_rt(t)->klmirqd_sem_stat, NOT_HELD);
1565 return(0);
1566 }
1567 }
1568 1138
1569 //TRACE_CUR("Trying to reacquire semaphore of %s/%d\n", owner->comm, owner->pid); 1139 if (likely(!info->terminating)) {
1570 __down_and_reset_and_set_stat(t, REACQUIRING, HELD, sem); 1140 ___litmus_schedule_work(w, info, 1);
1571 //TRACE_CUR("Reacquired semaphore %s/%d\n", owner->comm, owner->pid);
1572 } 1141 }
1573 /* 1142 else {
1574 else if(is_realtime(t)) 1143 TRACE("%s: Work rejected because %s/%d is terminating\n", klmirqd_thread->comm, klmirqd_thread->pid);
1575 { 1144 ret = 0;
1576 TRACE_CUR("%s: Nothing to do. Stat = %d\n", __FUNCTION__, tsk_rt(t)->klmirqd_sem_stat);
1577 } 1145 }
1578 */
1579 1146
1580 return(ret); 1147 return(ret);
1581} 1148}
1149EXPORT_SYMBOL(__litmus_schedule_work);
1582 1150
diff --git a/litmus/locking.c b/litmus/locking.c
index 22f46df4308a..7af1dd69a079 100644
--- a/litmus/locking.c
+++ b/litmus/locking.c
@@ -543,32 +543,54 @@ out:
543 543
544void suspend_for_lock(void) 544void suspend_for_lock(void)
545{ 545{
546#ifdef CONFIG_REALTIME_AUX_TASKS 546#if defined(CONFIG_REALTIME_AUX_TASKS) || defined(CONFIG_LITMUS_NVIDIA)
547#if 0
548 unsigned int restore = 0;
549 struct task_struct *t = current; 547 struct task_struct *t = current;
550 unsigned int hide;
551
552 if (tsk_rt(t)->has_aux_tasks) {
553 /* hide from aux tasks so they can't inherit our priority when we block
554 * for a litmus lock. inheritance is already going to a litmus lock
555 * holder. */
556 hide = tsk_rt(t)->hide_from_aux_tasks;
557 restore = 1;
558 tsk_rt(t)->hide_from_aux_tasks = 1;
559 }
560#endif 548#endif
549
550#ifdef CONFIG_REALTIME_AUX_TASKS
551 unsigned int aux_restore = 0;
552 unsigned int aux_hide;
553#endif
554
555#ifdef CONFIG_LITMUS_NVIDIA
556 unsigned int gpu_restore = 0;
557 unsigned int gpu_hide;
558#endif
559
560
561//#ifdef CONFIG_REALTIME_AUX_TASKS
562// if (tsk_rt(t)->has_aux_tasks) {
563// /* hide from aux tasks so they can't inherit our priority when we block
564// * for a litmus lock. inheritance is already going to a litmus lock
565// * holder. */
566// aux_hide = tsk_rt(t)->hide_from_aux_tasks;
567// aux_restore = 1;
568// tsk_rt(t)->hide_from_aux_tasks = 1;
569// }
570//#endif
571
572#ifdef CONFIG_LITMUS_NVIDIA
573 if (tsk_rt(t)->held_gpus) {
574 gpu_hide = tsk_rt(t)->hide_from_gpu;
575 gpu_restore = 1;
576 tsk_rt(t)->hide_from_gpu = 1;
577 }
561#endif 578#endif
562 579
563 schedule(); 580 schedule();
564 581
565#ifdef CONFIG_REALTIME_AUX_TASKS 582#ifdef CONFIG_LITMUS_NVIDIA
566#if 0 583 if (gpu_restore) {
567 if (restore) {
568 /* restore our state */ 584 /* restore our state */
569 tsk_rt(t)->hide_from_aux_tasks = hide; 585 tsk_rt(t)->hide_from_gpu = gpu_hide;
570 } 586 }
571#endif 587#endif
588
589#ifdef CONFIG_REALTIME_AUX_TASKS
590 if (aux_restore) {
591 /* restore our state */
592 tsk_rt(t)->hide_from_aux_tasks = aux_hide;
593 }
572#endif 594#endif
573} 595}
574 596
diff --git a/litmus/nvidia_info.c b/litmus/nvidia_info.c
index 22586cde8255..b29f4d3f0dac 100644
--- a/litmus/nvidia_info.c
+++ b/litmus/nvidia_info.c
@@ -10,6 +10,10 @@
10 10
11#include <litmus/binheap.h> 11#include <litmus/binheap.h>
12 12
13#ifdef CONFIG_LITMUS_SOFTIRQD
14#include <litmus/litmus_softirq.h>
15#endif
16
13typedef unsigned char NvV8; /* "void": enumerated or multiple fields */ 17typedef unsigned char NvV8; /* "void": enumerated or multiple fields */
14typedef unsigned short NvV16; /* "void": enumerated or multiple fields */ 18typedef unsigned short NvV16; /* "void": enumerated or multiple fields */
15typedef unsigned char NvU8; /* 0 to 255 */ 19typedef unsigned char NvU8; /* 0 to 255 */
@@ -296,9 +300,14 @@ static struct notifier_block nvidia_going = {
296}; 300};
297#endif 301#endif
298 302
303
304
305static int init_nv_device_reg(void);
306static int shutdown_nv_device_reg(void);
307
308
299int init_nvidia_info(void) 309int init_nvidia_info(void)
300{ 310{
301#if 1
302 mutex_lock(&module_mutex); 311 mutex_lock(&module_mutex);
303 nvidia_mod = find_module("nvidia"); 312 nvidia_mod = find_module("nvidia");
304 mutex_unlock(&module_mutex); 313 mutex_unlock(&module_mutex);
@@ -315,13 +324,14 @@ int init_nvidia_info(void)
315 TRACE("%s : Could not find NVIDIA module! Loaded?\n", __FUNCTION__); 324 TRACE("%s : Could not find NVIDIA module! Loaded?\n", __FUNCTION__);
316 return(-1); 325 return(-1);
317 } 326 }
318#endif
319} 327}
320 328
321void shutdown_nvidia_info(void) 329void shutdown_nvidia_info(void)
322{ 330{
323 nvidia_mod = NULL; 331 nvidia_mod = NULL;
324 mb(); 332 mb();
333
334 shutdown_nv_device_reg();
325} 335}
326 336
327/* works with pointers to static data inside the module too. */ 337/* works with pointers to static data inside the module too. */
@@ -351,20 +361,6 @@ u32 get_tasklet_nv_device_num(const struct tasklet_struct *t)
351 BUG_ON(linuxstate->device_num >= NV_DEVICE_NUM); 361 BUG_ON(linuxstate->device_num >= NV_DEVICE_NUM);
352 362
353 return(linuxstate->device_num); 363 return(linuxstate->device_num);
354
355 //int DEVICE_NUM_OFFSET = (void*)(&(linuxstate->device_num)) - (void*)(nvstate);
356
357#if 0
358 // offset determined though observed behavior of the NV driver.
359 //const int DEVICE_NUM_OFFSET = 0x480; // CUDA 4.0 RC1
360 //const int DEVICE_NUM_OFFSET = 0x510; // CUDA 4.0 RC2
361
362 void* state = (void*)(t->data);
363 void* device_num_ptr = state + DEVICE_NUM_OFFSET;
364
365 //dump_nvidia_info(t);
366 return(*((u32*)device_num_ptr));
367#endif
368} 364}
369 365
370u32 get_work_nv_device_num(const struct work_struct *t) 366u32 get_work_nv_device_num(const struct work_struct *t)
@@ -377,203 +373,452 @@ u32 get_work_nv_device_num(const struct work_struct *t)
377} 373}
378 374
379 375
376///////////////////////////////////////////////////////////////////////////////
377///////////////////////////////////////////////////////////////////////////////
378///////////////////////////////////////////////////////////////////////////////
379
380
380typedef struct { 381typedef struct {
381 raw_spinlock_t lock; 382 raw_spinlock_t lock; /* not needed if GPU not shared between scheudling domains */
382 int nr_owners; 383 struct binheap owners;
383 struct task_struct* max_prio_owner; 384
384 struct task_struct* owners[NV_MAX_SIMULT_USERS]; 385#ifdef CONFIG_LITMUS_SOFTIRQD
386 klmirqd_callback_t callback;
387 struct task_struct* thread;
388 int ready:1; /* todo: make threads check for the ready flag */
389#endif
385}nv_device_registry_t; 390}nv_device_registry_t;
386 391
392
387static nv_device_registry_t NV_DEVICE_REG[NV_DEVICE_NUM]; 393static nv_device_registry_t NV_DEVICE_REG[NV_DEVICE_NUM];
388 394
389int init_nv_device_reg(void) 395
396
397
398
399#ifdef CONFIG_LITMUS_SOFTIRQD
400static int nvidia_klmirqd_cb(void *arg)
390{ 401{
391 int i; 402 unsigned long flags;
403 int reg_device_id = (int)(long long)(arg);
404 nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id];
392 405
393 memset(NV_DEVICE_REG, 0, sizeof(NV_DEVICE_REG)); 406 TRACE("nv klmirqd callback for GPU %d\n", reg_device_id);
394 407
395 for(i = 0; i < NV_DEVICE_NUM; ++i) 408 raw_spin_lock_irqsave(&reg->lock, flags);
396 { 409 reg->thread = current;
397 raw_spin_lock_init(&NV_DEVICE_REG[i].lock); 410 reg->ready = 1;
398 } 411 raw_spin_unlock_irqrestore(&reg->lock, flags);
399 412
400 return(1); 413 return 0;
401} 414}
415#endif
402 416
403/* use to get nv_device_id by given owner. 417
404 (if return -1, can't get the assocaite device id)*/ 418static int gpu_owner_max_priority_order(struct binheap_node *a,
405/* 419 struct binheap_node *b)
406int get_nv_device_id(struct task_struct* owner)
407{ 420{
408 int i; 421 struct task_struct *d_a = container_of(binheap_entry(a, struct rt_param, gpu_owner_node),
409 if(!owner) 422 struct task_struct, rt_param);
410 { 423 struct task_struct *d_b = container_of(binheap_entry(b, struct rt_param, gpu_owner_node),
411 return(-1); 424 struct task_struct, rt_param);
412 } 425
413 for(i = 0; i < NV_DEVICE_NUM; ++i) 426 BUG_ON(!d_a);
414 { 427 BUG_ON(!d_b);
415 if(NV_DEVICE_REG[i].device_owner == owner) 428
416 return(i); 429 return litmus->compare(d_a, d_b);
417 }
418 return(-1);
419} 430}
420*/
421 431
422static struct task_struct* find_hp_owner(nv_device_registry_t *reg, struct task_struct *skip) { 432static int init_nv_device_reg(void)
433{
423 int i; 434 int i;
424 struct task_struct *found = NULL; 435
425 for(i = 0; i < reg->nr_owners; ++i) { 436#ifdef CONFIG_LITMUS_SOFTIRQD
426 if(reg->owners[i] && reg->owners[i] != skip && litmus->compare(reg->owners[i], found)) { 437 if (!klmirqd_is_ready()) {
427 found = reg->owners[i]; 438 TRACE("klmirqd is not ready!\n");
439 return 0;
440 }
441#endif
442
443 memset(NV_DEVICE_REG, 0, sizeof(NV_DEVICE_REG));
444 mb();
445
446
447 for(i = 0; i < NV_DEVICE_NUM; ++i) {
448 raw_spin_lock_init(&NV_DEVICE_REG[i].lock);
449 INIT_BINHEAP_HANDLE(&NV_DEVICE_REG[i].owners, gpu_owner_max_priority_order);
450
451#ifdef CONFIG_LITMUS_SOFTIRQD
452 // TODO: Make thread spawning this a litmus plugin call.
453 NV_DEVICE_REG[i].callback.func = nvidia_klmirqd_cb;
454 NV_DEVICE_REG[i].callback.arg = (void*)(long long)(i);
455 mb();
456
457 if(launch_klmirqd_thread(0, &NV_DEVICE_REG[i].callback) != 0) {
458 TRACE("Failed to create klmirqd thread for GPU %d\n", i);
428 } 459 }
460#endif
429 } 461 }
430 return found; 462
463 return(1);
431} 464}
432 465
433#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 466
434void pai_check_priority_increase(struct task_struct *t, int reg_device_id) 467/* The following code is full of nasty race conditions... */
468/* spawning of klimirqd threads can race with init_nv_device_reg()!!!! */
469static int shutdown_nv_device_reg(void)
435{ 470{
436 unsigned long flags; 471 TRACE("Shutting down nv device registration.\n");
437 nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id]; 472
473#ifdef CONFIG_LITMUS_SOFTIRQD
474 {
475 int i;
476 nv_device_registry_t *reg;
438 477
439 if(reg->max_prio_owner != t) { 478 for (i = 0; i < NV_DEVICE_NUM; ++i) {
440 479
441 raw_spin_lock_irqsave(&reg->lock, flags); 480 TRACE("Shutting down GPU %d.\n", i);
442 481
443 if(reg->max_prio_owner != t) { 482 reg = &NV_DEVICE_REG[i];
444 if(litmus->compare(t, reg->max_prio_owner)) { 483
445 litmus->change_prio_pai_tasklet(reg->max_prio_owner, t); 484 if (reg->thread && reg->ready) {
446 reg->max_prio_owner = t; 485 kill_klmirqd_thread(reg->thread);
486
487 /* assume that all goes according to plan... */
488 reg->thread = NULL;
489 reg->ready = 0;
447 } 490 }
448 }
449 491
450 raw_spin_unlock_irqrestore(&reg->lock, flags); 492 while (!binheap_empty(&reg->owners)) {
493 binheap_delete_root(&reg->owners, struct rt_param, gpu_owner_node);
494 }
495 }
451 } 496 }
497#endif
498
499 return(1);
452} 500}
453 501
454 502
455void pai_check_priority_decrease(struct task_struct *t, int reg_device_id) 503/* use to get the owner of nv_device_id. */
504struct task_struct* get_nv_max_device_owner(u32 target_device_id)
456{ 505{
457 unsigned long flags; 506 struct task_struct *owner = NULL;
458 nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id]; 507 nv_device_registry_t *reg;
459 508
460 if(reg->max_prio_owner == t) { 509 BUG_ON(target_device_id >= NV_DEVICE_NUM);
461 510
462 raw_spin_lock_irqsave(&reg->lock, flags); 511 reg = &NV_DEVICE_REG[target_device_id];
463 512
464 if(reg->max_prio_owner == t) { 513 if (!binheap_empty(&reg->owners)) {
465 reg->max_prio_owner = find_hp_owner(reg, NULL); 514 struct task_struct *hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
466 if(reg->max_prio_owner != t) { 515 struct task_struct, rt_param);
467 litmus->change_prio_pai_tasklet(t, reg->max_prio_owner); 516 TRACE_CUR("hp: %s/%d\n", hp->comm, hp->pid);
468 } 517 }
469 }
470 518
471 raw_spin_unlock_irqrestore(&reg->lock, flags); 519 return(owner);
520}
521
522#ifdef CONFIG_LITMUS_SOFTIRQD
523struct task_struct* get_nv_klmirqd_thread(u32 target_device_id)
524{
525 struct task_struct *klmirqd = NULL;
526 nv_device_registry_t *reg;
527
528 BUG_ON(target_device_id >= NV_DEVICE_NUM);
529
530 reg = &NV_DEVICE_REG[target_device_id];
531
532 if(likely(reg->ready)) {
533 klmirqd = reg->thread;
472 } 534 }
535
536 return klmirqd;
473} 537}
474#endif 538#endif
475 539
476static int __reg_nv_device(int reg_device_id, struct task_struct *t) 540
541
542
543
544#ifdef CONFIG_LITMUS_SOFTIRQD
545static int gpu_klmirqd_increase_priority(struct task_struct *klmirqd, struct task_struct *hp)
477{ 546{
478 int ret = 0; 547 int retval = 0;
479 int i;
480 struct task_struct *old_max = NULL;
481 unsigned long flags;
482 nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id];
483 548
484 if(test_bit(reg_device_id, &tsk_rt(t)->held_gpus)) { 549 TRACE_CUR("Increasing priority of nv klmirqd: %s/%d.\n", klmirqd->comm, klmirqd->pid);
485 // TODO: check if taks is already registered.
486 return ret; // assume already registered.
487 }
488 550
551 /* the klmirqd thread should never attempt to hold a litmus-level real-time
552 * so nested support is not required */
553 retval = litmus->__increase_prio(klmirqd, hp);
489 554
490 raw_spin_lock_irqsave(&reg->lock, flags); 555 return retval;
556}
557
558static int gpu_klmirqd_decrease_priority(struct task_struct *klmirqd, struct task_struct *hp)
559{
560 int retval = 0;
491 561
492 if(reg->nr_owners < NV_MAX_SIMULT_USERS) { 562 TRACE_CUR("Decreasing priority of nv klmirqd: %s/%d.\n", klmirqd->comm, klmirqd->pid);
493 TRACE_TASK(t, "registers GPU %d\n", reg_device_id);
494 for(i = 0; i < NV_MAX_SIMULT_USERS; ++i) {
495 if(reg->owners[i] == NULL) {
496 reg->owners[i] = t;
497 563
498 //if(edf_higher_prio(t, reg->max_prio_owner)) { 564 /* the klmirqd thread should never attempt to hold a litmus-level real-time
499 if(litmus->compare(t, reg->max_prio_owner)) { 565 * so nested support is not required */
500 old_max = reg->max_prio_owner; 566 retval = litmus->__decrease_prio(klmirqd, hp);
501 reg->max_prio_owner = t;
502 567
503#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 568 return retval;
504 litmus->change_prio_pai_tasklet(old_max, t); 569}
505#endif 570#endif
506 } 571
572
573
574
575/* call when an aux_owner becomes real-time */
576long enable_gpu_owner(struct task_struct *t)
577{
578 long retval = 0;
579// unsigned long flags;
580 int gpu;
581 nv_device_registry_t *reg;
507 582
508#ifdef CONFIG_LITMUS_SOFTIRQD 583#ifdef CONFIG_LITMUS_SOFTIRQD
509 down_and_set_stat(t, HELD, &tsk_rt(t)->klmirqd_sem); 584 struct task_struct *hp;
510#endif 585#endif
511 ++(reg->nr_owners);
512 586
513 break; 587 if (!tsk_rt(t)->held_gpus) {
514 } 588 TRACE_CUR("task %s/%d does not hold any GPUs\n", t->comm, t->pid);
515 } 589 return -1;
516 } 590 }
517 else 591
518 { 592 BUG_ON(!is_realtime(t));
519 TRACE_CUR("%s: device %d is already in use!\n", __FUNCTION__, reg_device_id); 593
520 //ret = -EBUSY; 594 gpu = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
595
596 if (binheap_is_in_heap(&tsk_rt(t)->gpu_owner_node)) {
597 TRACE_CUR("task %s/%d is already active on GPU %d\n", t->comm, t->pid, gpu);
598 goto out;
521 } 599 }
522 600
523 raw_spin_unlock_irqrestore(&reg->lock, flags); 601 /* update the registration (and maybe klmirqd) */
602 reg = &NV_DEVICE_REG[gpu];
524 603
525 __set_bit(reg_device_id, &tsk_rt(t)->held_gpus); 604// raw_spin_lock_irqsave(&reg->lock, flags);
526 605
527 return(ret); 606 binheap_add(&tsk_rt(t)->gpu_owner_node, &reg->owners,
607 struct rt_param, gpu_owner_node);
608
609
610#ifdef CONFIG_LITMUS_SOFTIRQD
611 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
612 struct task_struct, rt_param);
613
614 if (hp == t) {
615 /* we're the new hp */
616 TRACE_CUR("%s/%d is new hp on GPU %d.\n", t->comm, t->pid, gpu);
617
618 retval = gpu_klmirqd_increase_priority(reg->thread, (tsk_rt(hp)->inh_task)? tsk_rt(hp)->inh_task : hp);
619 }
620#endif
621
622// raw_spin_unlock_irqsave(&reg->lock, flags);
623
624out:
625 return retval;
528} 626}
529 627
530static int __clear_reg_nv_device(int de_reg_device_id, struct task_struct *t) 628/* call when an aux_owner exits real-time */
629long disable_gpu_owner(struct task_struct *t)
531{ 630{
532 int ret = 0; 631 long retval = 0;
533 int i; 632// unsigned long flags;
534 unsigned long flags; 633 int gpu;
535 nv_device_registry_t *reg = &NV_DEVICE_REG[de_reg_device_id]; 634 nv_device_registry_t *reg;
536 635
537#ifdef CONFIG_LITMUS_SOFTIRQD 636#ifdef CONFIG_LITMUS_SOFTIRQD
538 struct task_struct* klmirqd_th = get_klmirqd(de_reg_device_id); 637 struct task_struct *hp;
638 struct task_struct *new_hp = NULL;
539#endif 639#endif
540 640
541 if(!test_bit(de_reg_device_id, &tsk_rt(t)->held_gpus)) { 641 if (!tsk_rt(t)->held_gpus) {
542 return ret; 642 TRACE_CUR("task %s/%d does not hold any GPUs\n", t->comm, t->pid);
643 return -1;
543 } 644 }
544 645
545 raw_spin_lock_irqsave(&reg->lock, flags); 646 BUG_ON(!is_realtime(t));
546 647
547 TRACE_TASK(t, "unregisters GPU %d\n", de_reg_device_id); 648 gpu = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
649
650 if (!binheap_is_in_heap(&tsk_rt(t)->gpu_owner_node)) {
651 TRACE_CUR("task %s/%d is not active on GPU %d\n", t->comm, t->pid, gpu);
652 goto out;
653 }
654
655 TRACE_CUR("task %s/%d exiting from GPU %d.\n", t->comm, t->pid, gpu);
656
657
658 reg = &NV_DEVICE_REG[gpu];
659
660// raw_spin_lock_irqsave(&reg->lock, flags);
548 661
549 for(i = 0; i < NV_MAX_SIMULT_USERS; ++i) {
550 if(reg->owners[i] == t) {
551#ifdef CONFIG_LITMUS_SOFTIRQD
552 flush_pending(klmirqd_th, t);
553#endif
554 if(reg->max_prio_owner == t) {
555 reg->max_prio_owner = find_hp_owner(reg, t);
556#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
557 litmus->change_prio_pai_tasklet(t, reg->max_prio_owner);
558#endif
559 }
560 662
561#ifdef CONFIG_LITMUS_SOFTIRQD 663#ifdef CONFIG_LITMUS_SOFTIRQD
562 up_and_set_stat(t, NOT_HELD, &tsk_rt(t)->klmirqd_sem); 664 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
665 struct task_struct, rt_param);
666
667 binheap_delete(&tsk_rt(t)->gpu_owner_node, &reg->owners);
668
669
670 if (!binheap_empty(&reg->owners)) {
671 new_hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
672 struct task_struct, rt_param);
673 }
674
675 if (hp == t && new_hp != t) {
676 struct task_struct *to_inh = NULL;
677
678 TRACE_CUR("%s/%d is no longer hp on GPU %d.\n", t->comm, t->pid, gpu);
679
680 if (new_hp) {
681 to_inh = (tsk_rt(new_hp)->inh_task) ? tsk_rt(new_hp)->inh_task : new_hp;
682 }
683
684 retval = gpu_klmirqd_decrease_priority(reg->thread, to_inh);
685 }
686#else
687 binheap_delete(&tsk_rt(t)->gpu_owner_node, &reg->owners);
563#endif 688#endif
564 689
565 reg->owners[i] = NULL; 690// raw_spin_unlock_irqsave(&reg->lock, flags);
566 --(reg->nr_owners); 691
692
693out:
694 return retval;
695}
696
697
698
699
700
701
702
703
704
705
706int gpu_owner_increase_priority(struct task_struct *t)
707{
708 int retval = 0;
709 int gpu;
710 nv_device_registry_t *reg;
711
712 struct task_struct *hp = NULL;
713 struct task_struct *hp_eff = NULL;
714
715 BUG_ON(!is_realtime(t));
716 BUG_ON(!tsk_rt(t)->held_gpus);
717
718 gpu = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
719
720 if (!binheap_is_in_heap(&tsk_rt(t)->gpu_owner_node)) {
721 WARN_ON(!is_running(t));
722 TRACE_CUR("gpu klmirqd may not inherit from %s/%d on GPU %d\n",
723 t->comm, t->pid, gpu);
724 goto out;
725 }
726
727
728
729
730 TRACE_CUR("task %s/%d on GPU %d increasing priority.\n", t->comm, t->pid, gpu);
731 reg = &NV_DEVICE_REG[gpu];
732
733 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
734 struct task_struct, rt_param);
735 hp_eff = effective_priority(hp);
736
737 if (hp != t) { /* our position in the heap may have changed. hp is already at the root. */
738 binheap_decrease(&tsk_rt(t)->gpu_owner_node, &reg->owners);
739 }
740
741 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
742 struct task_struct, rt_param);
743
744 if (effective_priority(hp) != hp_eff) { /* the eff. prio. of hp has changed */
745 hp_eff = effective_priority(hp);
746 TRACE_CUR("%s/%d is new hp on GPU %d.\n", t->comm, t->pid, gpu);
747
748 retval = gpu_klmirqd_increase_priority(reg->thread, hp_eff);
749 }
750
751out:
752 return retval;
753}
754
755
756int gpu_owner_decrease_priority(struct task_struct *t)
757{
758 int retval = 0;
759 int gpu;
760 nv_device_registry_t *reg;
761
762 struct task_struct *hp = NULL;
763 struct task_struct *hp_eff = NULL;
567 764
568 break; 765 BUG_ON(!is_realtime(t));
766 BUG_ON(!tsk_rt(t)->held_gpus);
767
768 gpu = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
769
770 if (!binheap_is_in_heap(&tsk_rt(t)->aux_task_owner_node)) {
771 WARN_ON(!is_running(t));
772 TRACE_CUR("aux tasks may not inherit from %s/%d on GPU %d\n",
773 t->comm, t->pid, gpu);
774 goto out;
775 }
776
777 TRACE_CUR("task %s/%d on GPU %d decresing priority.\n", t->comm, t->pid, gpu);
778 reg = &NV_DEVICE_REG[gpu];
779
780 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
781 struct task_struct, rt_param);
782 hp_eff = effective_priority(hp);
783 binheap_delete(&tsk_rt(t)->gpu_owner_node, &reg->owners);
784 binheap_add(&tsk_rt(t)->gpu_owner_node, &reg->owners,
785 struct rt_param, gpu_owner_node);
786
787 if (hp == t) { /* t was originally the hp */
788 struct task_struct *new_hp =
789 container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
790 struct task_struct, rt_param);
791 if (effective_priority(new_hp) != hp_eff) { /* eff prio. of hp has changed */
792 hp_eff = effective_priority(new_hp);
793 TRACE_CUR("%s/%d is no longer hp on GPU %d.\n", t->comm, t->pid, gpu);
794 retval = gpu_klmirqd_decrease_priority(reg->thread, hp_eff);
569 } 795 }
570 } 796 }
571 797
572 raw_spin_unlock_irqrestore(&reg->lock, flags); 798out:
799 return retval;
800}
801
802
803
804
805
806
807
808
809
810static int __reg_nv_device(int reg_device_id, struct task_struct *t)
811{
812 __set_bit(reg_device_id, &tsk_rt(t)->held_gpus);
813
814 return(0);
815}
573 816
817static int __clear_reg_nv_device(int de_reg_device_id, struct task_struct *t)
818{
574 __clear_bit(de_reg_device_id, &tsk_rt(t)->held_gpus); 819 __clear_bit(de_reg_device_id, &tsk_rt(t)->held_gpus);
575 820
576 return(ret); 821 return(0);
577} 822}
578 823
579 824
@@ -596,55 +841,213 @@ int reg_nv_device(int reg_device_id, int reg_action, struct task_struct *t)
596 return(ret); 841 return(ret);
597} 842}
598 843
599/* use to get the owner of nv_device_id. */
600struct task_struct* get_nv_max_device_owner(u32 target_device_id)
601{
602 struct task_struct *owner = NULL;
603 BUG_ON(target_device_id >= NV_DEVICE_NUM);
604 owner = NV_DEVICE_REG[target_device_id].max_prio_owner;
605 return(owner);
606}
607 844
608void lock_nv_registry(u32 target_device_id, unsigned long* flags)
609{
610 BUG_ON(target_device_id >= NV_DEVICE_NUM);
611 845
612 if(in_interrupt())
613 TRACE("Locking registry for %d.\n", target_device_id);
614 else
615 TRACE_CUR("Locking registry for %d.\n", target_device_id);
616 846
617 raw_spin_lock_irqsave(&NV_DEVICE_REG[target_device_id].lock, *flags);
618}
619 847
620void unlock_nv_registry(u32 target_device_id, unsigned long* flags)
621{
622 BUG_ON(target_device_id >= NV_DEVICE_NUM);
623 848
624 if(in_interrupt())
625 TRACE("Unlocking registry for %d.\n", target_device_id);
626 else
627 TRACE_CUR("Unlocking registry for %d.\n", target_device_id);
628 849
629 raw_spin_unlock_irqrestore(&NV_DEVICE_REG[target_device_id].lock, *flags);
630}
631 850
632 851
633//void increment_nv_int_count(u32 device) 852
853
854
855
856
857
858
859
860
861
862#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
863//void pai_check_priority_increase(struct task_struct *t, int reg_device_id)
864//{
865// unsigned long flags;
866// nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id];
867//
868//
869//
870// if(reg->max_prio_owner != t) {
871//
872// raw_spin_lock_irqsave(&reg->lock, flags);
873//
874// if(reg->max_prio_owner != t) {
875// if(litmus->compare(t, reg->max_prio_owner)) {
876// litmus->change_prio_pai_tasklet(reg->max_prio_owner, t);
877// reg->max_prio_owner = t;
878// }
879// }
880//
881// raw_spin_unlock_irqrestore(&reg->lock, flags);
882// }
883//}
884//
885//
886//void pai_check_priority_decrease(struct task_struct *t, int reg_device_id)
634//{ 887//{
635// unsigned long flags; 888// unsigned long flags;
636// struct task_struct* owner; 889// nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id];
637// 890//
638// lock_nv_registry(device, &flags); 891// if(reg->max_prio_owner == t) {
639// 892//
640// owner = NV_DEVICE_REG[device].device_owner; 893// raw_spin_lock_irqsave(&reg->lock, flags);
641// if(owner) 894//
895// if(reg->max_prio_owner == t) {
896// reg->max_prio_owner = find_hp_owner(reg, NULL);
897// if(reg->max_prio_owner != t) {
898// litmus->change_prio_pai_tasklet(t, reg->max_prio_owner);
899// }
900// }
901//
902// raw_spin_unlock_irqrestore(&reg->lock, flags);
903// }
904//}
905#endif
906
907
908
909
910
911//static int __reg_nv_device(int reg_device_id, struct task_struct *t)
912//{
913// int ret = 0;
914// int i;
915// struct task_struct *old_max = NULL;
916//
917//
918// raw_spin_lock_irqsave(&reg->lock, flags);
919//
920// if(reg->nr_owners < NV_MAX_SIMULT_USERS) {
921// TRACE_TASK(t, "registers GPU %d\n", reg_device_id);
922// for(i = 0; i < NV_MAX_SIMULT_USERS; ++i) {
923// if(reg->owners[i] == NULL) {
924// reg->owners[i] = t;
925//
926// //if(edf_higher_prio(t, reg->max_prio_owner)) {
927// if(litmus->compare(t, reg->max_prio_owner)) {
928// old_max = reg->max_prio_owner;
929// reg->max_prio_owner = t;
930//
931//#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
932// litmus->change_prio_pai_tasklet(old_max, t);
933//#endif
934// }
935//
936//#ifdef CONFIG_LITMUS_SOFTIRQD
937// down_and_set_stat(t, HELD, &tsk_rt(t)->klmirqd_sem);
938//#endif
939// ++(reg->nr_owners);
940//
941// break;
942// }
943// }
944// }
945// else
642// { 946// {
643// atomic_inc(&tsk_rt(owner)->nv_int_count); 947// TRACE_CUR("%s: device %d is already in use!\n", __FUNCTION__, reg_device_id);
948// //ret = -EBUSY;
644// } 949// }
645// 950//
646// unlock_nv_registry(device, &flags); 951// raw_spin_unlock_irqrestore(&reg->lock, flags);
952//
953// __set_bit(reg_device_id, &tsk_rt(t)->held_gpus);
954//
955// return(ret);
956//}
957//
958//static int __clear_reg_nv_device(int de_reg_device_id, struct task_struct *t)
959//{
960// int ret = 0;
961// int i;
962// unsigned long flags;
963// nv_device_registry_t *reg = &NV_DEVICE_REG[de_reg_device_id];
964//
965//#ifdef CONFIG_LITMUS_SOFTIRQD
966// struct task_struct* klmirqd_th = get_klmirqd(de_reg_device_id);
967//#endif
968//
969// if(!test_bit(de_reg_device_id, &tsk_rt(t)->held_gpus)) {
970// return ret;
971// }
972//
973// raw_spin_lock_irqsave(&reg->lock, flags);
974//
975// TRACE_TASK(t, "unregisters GPU %d\n", de_reg_device_id);
976//
977// for(i = 0; i < NV_MAX_SIMULT_USERS; ++i) {
978// if(reg->owners[i] == t) {
979//#ifdef CONFIG_LITMUS_SOFTIRQD
980// flush_pending(klmirqd_th, t);
981//#endif
982// if(reg->max_prio_owner == t) {
983// reg->max_prio_owner = find_hp_owner(reg, t);
984//#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
985// litmus->change_prio_pai_tasklet(t, reg->max_prio_owner);
986//#endif
987// }
988//
989//#ifdef CONFIG_LITMUS_SOFTIRQD
990// up_and_set_stat(t, NOT_HELD, &tsk_rt(t)->klmirqd_sem);
991//#endif
992//
993// reg->owners[i] = NULL;
994// --(reg->nr_owners);
995//
996// break;
997// }
998// }
999//
1000// raw_spin_unlock_irqrestore(&reg->lock, flags);
1001//
1002// __clear_bit(de_reg_device_id, &tsk_rt(t)->held_gpus);
1003//
1004// return(ret);
1005//}
1006//
1007//
1008//int reg_nv_device(int reg_device_id, int reg_action, struct task_struct *t)
1009//{
1010// int ret;
1011//
1012// if((reg_device_id < NV_DEVICE_NUM) && (reg_device_id >= 0))
1013// {
1014// if(reg_action)
1015// ret = __reg_nv_device(reg_device_id, t);
1016// else
1017// ret = __clear_reg_nv_device(reg_device_id, t);
1018// }
1019// else
1020// {
1021// ret = -ENODEV;
1022// }
1023//
1024// return(ret);
1025//}
1026
1027
1028
1029//void lock_nv_registry(u32 target_device_id, unsigned long* flags)
1030//{
1031// BUG_ON(target_device_id >= NV_DEVICE_NUM);
1032//
1033// if(in_interrupt())
1034// TRACE("Locking registry for %d.\n", target_device_id);
1035// else
1036// TRACE_CUR("Locking registry for %d.\n", target_device_id);
1037//
1038// raw_spin_lock_irqsave(&NV_DEVICE_REG[target_device_id].lock, *flags);
1039//}
1040//
1041//void unlock_nv_registry(u32 target_device_id, unsigned long* flags)
1042//{
1043// BUG_ON(target_device_id >= NV_DEVICE_NUM);
1044//
1045// if(in_interrupt())
1046// TRACE("Unlocking registry for %d.\n", target_device_id);
1047// else
1048// TRACE_CUR("Unlocking registry for %d.\n", target_device_id);
1049//
1050// raw_spin_unlock_irqrestore(&NV_DEVICE_REG[target_device_id].lock, *flags);
647//} 1051//}
648//EXPORT_SYMBOL(increment_nv_int_count);
649 1052
650 1053
diff --git a/litmus/sched_cedf.c b/litmus/sched_cedf.c
index 44c8336c5061..84aafca78cde 100644
--- a/litmus/sched_cedf.c
+++ b/litmus/sched_cedf.c
@@ -44,6 +44,7 @@
44 44
45#include <litmus/bheap.h> 45#include <litmus/bheap.h>
46#include <litmus/binheap.h> 46#include <litmus/binheap.h>
47#include <litmus/trace.h>
47 48
48#ifdef CONFIG_LITMUS_LOCKING 49#ifdef CONFIG_LITMUS_LOCKING
49#include <litmus/kfmlp_lock.h> 50#include <litmus/kfmlp_lock.h>
@@ -75,7 +76,6 @@
75 76
76#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 77#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
77#include <linux/interrupt.h> 78#include <linux/interrupt.h>
78#include <litmus/trace.h>
79#endif 79#endif
80 80
81#ifdef CONFIG_LITMUS_NVIDIA 81#ifdef CONFIG_LITMUS_NVIDIA
@@ -118,14 +118,6 @@ DEFINE_PER_CPU(cpu_entry_t, cedf_cpu_entries);
118#define test_will_schedule(cpu) \ 118#define test_will_schedule(cpu) \
119 (atomic_read(&per_cpu(cedf_cpu_entries, cpu).will_schedule)) 119 (atomic_read(&per_cpu(cedf_cpu_entries, cpu).will_schedule))
120 120
121#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
122struct tasklet_head
123{
124 struct tasklet_struct *head;
125 struct tasklet_struct **tail;
126};
127#endif
128
129/* 121/*
130 * In C-EDF there is a cedf domain _per_ cluster 122 * In C-EDF there is a cedf domain _per_ cluster
131 * The number of clusters is dynamically determined accordingly to the 123 * The number of clusters is dynamically determined accordingly to the
@@ -1038,6 +1030,13 @@ static void cedf_task_wake_up(struct task_struct *task)
1038 } 1030 }
1039#endif 1031#endif
1040 1032
1033#ifdef CONFIG_LITMUS_NVIDIA
1034 if (tsk_rt(task)->held_gpus && !tsk_rt(task)->hide_from_gpu) {
1035 TRACE_CUR("%s/%d is ready so gpu klmirqd tasks may not inherit.\n", task->comm, task->pid);
1036 disable_gpu_owner(task);
1037 }
1038#endif
1039
1041 cedf_job_arrival(task); 1040 cedf_job_arrival(task);
1042 raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags); 1041 raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);
1043} 1042}
@@ -1064,6 +1063,14 @@ static void cedf_task_block(struct task_struct *t)
1064 } 1063 }
1065#endif 1064#endif
1066 1065
1066#ifdef CONFIG_LITMUS_NVIDIA
1067 if (tsk_rt(t)->held_gpus && !tsk_rt(t)->hide_from_gpu) {
1068
1069 TRACE_CUR("%s/%d is blocked so aux tasks may inherit.\n", t->comm, t->pid);
1070 enable_gpu_owner(t);
1071 }
1072#endif
1073
1067 raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags); 1074 raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);
1068 1075
1069 BUG_ON(!is_realtime(t)); 1076 BUG_ON(!is_realtime(t));
@@ -1092,6 +1099,13 @@ static void cedf_task_exit(struct task_struct * t)
1092 } 1099 }
1093#endif 1100#endif
1094 1101
1102#ifdef CONFIG_LITMUS_NVIDIA
1103 /* make sure we clean up on our way out */
1104 if(tsk_rt(t)->held_gpus) {
1105 disable_gpu_owner(t);
1106 }
1107#endif
1108
1095 unlink(t); 1109 unlink(t);
1096 if (tsk_rt(t)->scheduled_on != NO_CPU) { 1110 if (tsk_rt(t)->scheduled_on != NO_CPU) {
1097 cpu_entry_t *cpu; 1111 cpu_entry_t *cpu;
@@ -1208,6 +1222,13 @@ static int __increase_priority_inheritance(struct task_struct* t,
1208 aux_task_owner_increase_priority(t); 1222 aux_task_owner_increase_priority(t);
1209 } 1223 }
1210#endif 1224#endif
1225
1226#ifdef CONFIG_LITMUS_NVIDIA
1227 /* propagate to gpu klmirqd */
1228 if (tsk_rt(t)->held_gpus) {
1229 gpu_owner_increase_priority(t);
1230 }
1231#endif
1211 } 1232 }
1212#ifdef CONFIG_LITMUS_NESTED_LOCKING 1233#ifdef CONFIG_LITMUS_NESTED_LOCKING
1213 } 1234 }
@@ -1237,16 +1258,6 @@ static void increase_priority_inheritance(struct task_struct* t, struct task_str
1237 1258
1238 __increase_priority_inheritance(t, prio_inh); 1259 __increase_priority_inheritance(t, prio_inh);
1239 1260
1240#ifdef CONFIG_LITMUS_SOFTIRQD
1241 if(tsk_rt(t)->cur_klmirqd != NULL)
1242 {
1243 TRACE_TASK(t, "%s/%d inherits a new priority!\n",
1244 tsk_rt(t)->cur_klmirqd->comm, tsk_rt(t)->cur_klmirqd->pid);
1245
1246 __increase_priority_inheritance(tsk_rt(t)->cur_klmirqd, prio_inh);
1247 }
1248#endif
1249
1250 raw_spin_unlock(&cluster->cluster_lock); 1261 raw_spin_unlock(&cluster->cluster_lock);
1251 1262
1252#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA) 1263#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA)
@@ -1320,6 +1331,13 @@ static int __decrease_priority_inheritance(struct task_struct* t,
1320 } 1331 }
1321#endif 1332#endif
1322 1333
1334#ifdef CONFIG_LITMUS_NVIDIA
1335 /* propagate to gpu */
1336 if (tsk_rt(t)->held_gpus) {
1337 gpu_owner_decrease_priority(t);
1338 }
1339#endif
1340
1323#ifdef CONFIG_LITMUS_NESTED_LOCKING 1341#ifdef CONFIG_LITMUS_NESTED_LOCKING
1324 } 1342 }
1325 else { 1343 else {
@@ -1346,16 +1364,6 @@ static void decrease_priority_inheritance(struct task_struct* t,
1346 raw_spin_lock(&cluster->cluster_lock); 1364 raw_spin_lock(&cluster->cluster_lock);
1347 __decrease_priority_inheritance(t, prio_inh); 1365 __decrease_priority_inheritance(t, prio_inh);
1348 1366
1349#ifdef CONFIG_LITMUS_SOFTIRQD
1350 if(tsk_rt(t)->cur_klmirqd != NULL)
1351 {
1352 TRACE_TASK(t, "%s/%d decreases in priority!\n",
1353 tsk_rt(t)->cur_klmirqd->comm, tsk_rt(t)->cur_klmirqd->pid);
1354
1355 __decrease_priority_inheritance(tsk_rt(t)->cur_klmirqd, prio_inh);
1356 }
1357#endif
1358
1359 raw_spin_unlock(&cluster->cluster_lock); 1367 raw_spin_unlock(&cluster->cluster_lock);
1360 1368
1361#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA) 1369#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA)
@@ -1371,73 +1379,6 @@ static void decrease_priority_inheritance(struct task_struct* t,
1371} 1379}
1372 1380
1373 1381
1374
1375
1376
1377#ifdef CONFIG_LITMUS_SOFTIRQD
1378/* called with IRQs off */
1379static void increase_priority_inheritance_klmirqd(struct task_struct* klmirqd,
1380 struct task_struct* old_owner,
1381 struct task_struct* new_owner)
1382{
1383 cedf_domain_t* cluster = task_cpu_cluster(klmirqd);
1384
1385 BUG_ON(!(tsk_rt(klmirqd)->is_proxy_thread));
1386
1387 raw_spin_lock(&cluster->cluster_lock);
1388
1389 if(old_owner != new_owner)
1390 {
1391 if(old_owner)
1392 {
1393 // unreachable?
1394 tsk_rt(old_owner)->cur_klmirqd = NULL;
1395 }
1396
1397 TRACE_TASK(klmirqd, "giving ownership to %s/%d.\n",
1398 new_owner->comm, new_owner->pid);
1399
1400 tsk_rt(new_owner)->cur_klmirqd = klmirqd;
1401 }
1402
1403 __decrease_priority_inheritance(klmirqd, NULL); // kludge to clear out cur prio.
1404
1405 __increase_priority_inheritance(klmirqd,
1406 (tsk_rt(new_owner)->inh_task == NULL) ?
1407 new_owner :
1408 tsk_rt(new_owner)->inh_task);
1409
1410 raw_spin_unlock(&cluster->cluster_lock);
1411}
1412
1413
1414/* called with IRQs off */
1415static void decrease_priority_inheritance_klmirqd(struct task_struct* klmirqd,
1416 struct task_struct* old_owner,
1417 struct task_struct* new_owner)
1418{
1419 cedf_domain_t* cluster = task_cpu_cluster(klmirqd);
1420
1421 BUG_ON(!(tsk_rt(klmirqd)->is_proxy_thread));
1422
1423 raw_spin_lock(&cluster->cluster_lock);
1424
1425 TRACE_TASK(klmirqd, "priority restored\n");
1426
1427 __decrease_priority_inheritance(klmirqd, new_owner);
1428
1429 tsk_rt(old_owner)->cur_klmirqd = NULL;
1430
1431 raw_spin_unlock(&cluster->cluster_lock);
1432}
1433#endif // CONFIG_LITMUS_SOFTIRQD
1434
1435
1436
1437
1438
1439
1440
1441#ifdef CONFIG_LITMUS_NESTED_LOCKING 1382#ifdef CONFIG_LITMUS_NESTED_LOCKING
1442 1383
1443/* called with IRQs off */ 1384/* called with IRQs off */
@@ -1836,33 +1777,7 @@ static long cedf_activate_plugin(void)
1836 } 1777 }
1837 1778
1838#ifdef CONFIG_LITMUS_SOFTIRQD 1779#ifdef CONFIG_LITMUS_SOFTIRQD
1839 { 1780 init_klmirqd();
1840 /* distribute the daemons evenly across the clusters. */
1841 int* affinity = kmalloc(NR_LITMUS_SOFTIRQD * sizeof(int), GFP_ATOMIC);
1842 int num_daemons_per_cluster = NR_LITMUS_SOFTIRQD / num_clusters;
1843 int left_over = NR_LITMUS_SOFTIRQD % num_clusters;
1844
1845 int daemon = 0;
1846 for(i = 0; i < num_clusters; ++i)
1847 {
1848 int num_on_this_cluster = num_daemons_per_cluster;
1849 if(left_over)
1850 {
1851 ++num_on_this_cluster;
1852 --left_over;
1853 }
1854
1855 for(j = 0; j < num_on_this_cluster; ++j)
1856 {
1857 // first CPU of this cluster
1858 affinity[daemon++] = i*cluster_size;
1859 }
1860 }
1861
1862 spawn_klmirqd(affinity);
1863
1864 kfree(affinity);
1865 }
1866#endif 1781#endif
1867 1782
1868#ifdef CONFIG_LITMUS_NVIDIA 1783#ifdef CONFIG_LITMUS_NVIDIA
@@ -1906,10 +1821,6 @@ static struct sched_plugin cedf_plugin __cacheline_aligned_in_smp = {
1906#ifdef CONFIG_LITMUS_AFFINITY_LOCKING 1821#ifdef CONFIG_LITMUS_AFFINITY_LOCKING
1907 .allocate_aff_obs = cedf_allocate_affinity_observer, 1822 .allocate_aff_obs = cedf_allocate_affinity_observer,
1908#endif 1823#endif
1909#ifdef CONFIG_LITMUS_SOFTIRQD
1910 .increase_prio_klmirqd = increase_priority_inheritance_klmirqd,
1911 .decrease_prio_klmirqd = decrease_priority_inheritance_klmirqd,
1912#endif
1913#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 1824#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
1914 .enqueue_pai_tasklet = cedf_enqueue_pai_tasklet, 1825 .enqueue_pai_tasklet = cedf_enqueue_pai_tasklet,
1915 .change_prio_pai_tasklet = cedf_change_prio_pai_tasklet, 1826 .change_prio_pai_tasklet = cedf_change_prio_pai_tasklet,
diff --git a/litmus/sched_gsn_edf.c b/litmus/sched_gsn_edf.c
index d52be9325044..f27c104ea027 100644
--- a/litmus/sched_gsn_edf.c
+++ b/litmus/sched_gsn_edf.c
@@ -26,6 +26,7 @@
26 26
27#include <litmus/bheap.h> 27#include <litmus/bheap.h>
28#include <litmus/binheap.h> 28#include <litmus/binheap.h>
29#include <litmus/trace.h>
29 30
30#ifdef CONFIG_LITMUS_LOCKING 31#ifdef CONFIG_LITMUS_LOCKING
31#include <litmus/kfmlp_lock.h> 32#include <litmus/kfmlp_lock.h>
@@ -50,7 +51,6 @@
50 51
51#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 52#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
52#include <linux/interrupt.h> 53#include <linux/interrupt.h>
53#include <litmus/trace.h>
54#endif 54#endif
55 55
56#ifdef CONFIG_LITMUS_NVIDIA 56#ifdef CONFIG_LITMUS_NVIDIA
@@ -156,12 +156,6 @@ static raw_spinlock_t* gsnedf_get_dgl_spinlock(struct task_struct *t)
156#endif 156#endif
157 157
158#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 158#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
159struct tasklet_head
160{
161 struct tasklet_struct *head;
162 struct tasklet_struct **tail;
163};
164
165struct tasklet_head gsnedf_pending_tasklets; 159struct tasklet_head gsnedf_pending_tasklets;
166#endif 160#endif
167 161
@@ -938,13 +932,6 @@ static struct task_struct* gsnedf_schedule(struct task_struct * prev)
938 932
939 sched_state_task_picked(); 933 sched_state_task_picked();
940 934
941#if 0
942 if (next && is_realtime(next) && tsk_rt(next)->is_aux_task && !tsk_rt(next)->inh_task) {
943 TRACE_TASK(next, "is aux with no inheritance. preventing it from actually running.\n");
944 next = NULL;
945 }
946#endif
947
948 raw_spin_unlock(&gsnedf_lock); 935 raw_spin_unlock(&gsnedf_lock);
949 936
950#ifdef WANT_ALL_SCHED_EVENTS 937#ifdef WANT_ALL_SCHED_EVENTS
@@ -1056,6 +1043,13 @@ static void gsnedf_task_wake_up(struct task_struct *task)
1056 } 1043 }
1057#endif 1044#endif
1058 1045
1046#ifdef CONFIG_LITMUS_NVIDIA
1047 if (tsk_rt(task)->held_gpus && !tsk_rt(task)->hide_from_gpu) {
1048 TRACE_CUR("%s/%d is ready so gpu klmirqd tasks may not inherit.\n", task->comm, task->pid);
1049 disable_gpu_owner(task);
1050 }
1051#endif
1052
1059 gsnedf_job_arrival(task); 1053 gsnedf_job_arrival(task);
1060 raw_spin_unlock_irqrestore(&gsnedf_lock, flags); 1054 raw_spin_unlock_irqrestore(&gsnedf_lock, flags);
1061} 1055}
@@ -1079,6 +1073,14 @@ static void gsnedf_task_block(struct task_struct *t)
1079 } 1073 }
1080#endif 1074#endif
1081 1075
1076#ifdef CONFIG_LITMUS_NVIDIA
1077 if (tsk_rt(t)->held_gpus && !tsk_rt(t)->hide_from_gpu) {
1078
1079 TRACE_CUR("%s/%d is blocked so aux tasks may inherit.\n", t->comm, t->pid);
1080 enable_gpu_owner(t);
1081 }
1082#endif
1083
1082 raw_spin_unlock_irqrestore(&gsnedf_lock, flags); 1084 raw_spin_unlock_irqrestore(&gsnedf_lock, flags);
1083 1085
1084 BUG_ON(!is_realtime(t)); 1086 BUG_ON(!is_realtime(t));
@@ -1106,6 +1108,13 @@ static void gsnedf_task_exit(struct task_struct * t)
1106 } 1108 }
1107#endif 1109#endif
1108 1110
1111#ifdef CONFIG_LITMUS_NVIDIA
1112 /* make sure we clean up on our way out */
1113 if(tsk_rt(t)->held_gpus) {
1114 disable_gpu_owner(t);
1115 }
1116#endif
1117
1109 unlink(t); 1118 unlink(t);
1110 if (tsk_rt(t)->scheduled_on != NO_CPU) { 1119 if (tsk_rt(t)->scheduled_on != NO_CPU) {
1111 gsnedf_cpus[tsk_rt(t)->scheduled_on]->scheduled = NULL; 1120 gsnedf_cpus[tsk_rt(t)->scheduled_on]->scheduled = NULL;
@@ -1154,7 +1163,6 @@ static int __increase_priority_inheritance(struct task_struct* t,
1154 1163
1155#ifdef CONFIG_LITMUS_NESTED_LOCKING 1164#ifdef CONFIG_LITMUS_NESTED_LOCKING
1156 /* this sanity check allows for weaker locking in protocols */ 1165 /* this sanity check allows for weaker locking in protocols */
1157 /* TODO (klmirqd): Skip this check if 't' is a proxy thread (???) */
1158 if(__edf_higher_prio(prio_inh, BASE, t, EFFECTIVE)) { 1166 if(__edf_higher_prio(prio_inh, BASE, t, EFFECTIVE)) {
1159#endif 1167#endif
1160 TRACE_TASK(t, "inherits priority from %s/%d\n", 1168 TRACE_TASK(t, "inherits priority from %s/%d\n",
@@ -1218,6 +1226,14 @@ static int __increase_priority_inheritance(struct task_struct* t,
1218 aux_task_owner_increase_priority(t); 1226 aux_task_owner_increase_priority(t);
1219 } 1227 }
1220#endif 1228#endif
1229
1230#ifdef CONFIG_LITMUS_NVIDIA
1231 /* propagate to gpu klmirqd */
1232 if (tsk_rt(t)->held_gpus) {
1233 gpu_owner_increase_priority(t);
1234 }
1235#endif
1236
1221 } 1237 }
1222#ifdef CONFIG_LITMUS_NESTED_LOCKING 1238#ifdef CONFIG_LITMUS_NESTED_LOCKING
1223 } 1239 }
@@ -1247,16 +1263,6 @@ static void increase_priority_inheritance(struct task_struct* t, struct task_str
1247 1263
1248 success = __increase_priority_inheritance(t, prio_inh); 1264 success = __increase_priority_inheritance(t, prio_inh);
1249 1265
1250#ifdef CONFIG_LITMUS_SOFTIRQD
1251 if(tsk_rt(t)->cur_klmirqd != NULL)
1252 {
1253 TRACE_TASK(t, "%s/%d inherits a new priority!\n",
1254 tsk_rt(t)->cur_klmirqd->comm, tsk_rt(t)->cur_klmirqd->pid);
1255
1256 __increase_priority_inheritance(tsk_rt(t)->cur_klmirqd, prio_inh);
1257 }
1258#endif
1259
1260 raw_spin_unlock(&gsnedf_lock); 1266 raw_spin_unlock(&gsnedf_lock);
1261 1267
1262#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA) 1268#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA)
@@ -1330,6 +1336,14 @@ static int __decrease_priority_inheritance(struct task_struct* t,
1330 } 1336 }
1331#endif 1337#endif
1332 1338
1339#ifdef CONFIG_LITMUS_NVIDIA
1340 /* propagate to gpu */
1341 if (tsk_rt(t)->held_gpus) {
1342 gpu_owner_decrease_priority(t);
1343 }
1344#endif
1345
1346
1333#ifdef CONFIG_LITMUS_NESTED_LOCKING 1347#ifdef CONFIG_LITMUS_NESTED_LOCKING
1334 } 1348 }
1335 else { 1349 else {
@@ -1357,16 +1371,6 @@ static void decrease_priority_inheritance(struct task_struct* t,
1357 1371
1358 success = __decrease_priority_inheritance(t, prio_inh); 1372 success = __decrease_priority_inheritance(t, prio_inh);
1359 1373
1360#ifdef CONFIG_LITMUS_SOFTIRQD
1361 if(tsk_rt(t)->cur_klmirqd != NULL)
1362 {
1363 TRACE_TASK(t, "%s/%d decreases in priority!\n",
1364 tsk_rt(t)->cur_klmirqd->comm, tsk_rt(t)->cur_klmirqd->pid);
1365
1366 __decrease_priority_inheritance(tsk_rt(t)->cur_klmirqd, prio_inh);
1367 }
1368#endif
1369
1370 raw_spin_unlock(&gsnedf_lock); 1374 raw_spin_unlock(&gsnedf_lock);
1371 1375
1372#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA) 1376#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA)
@@ -1382,62 +1386,6 @@ static void decrease_priority_inheritance(struct task_struct* t,
1382} 1386}
1383 1387
1384 1388
1385#ifdef CONFIG_LITMUS_SOFTIRQD
1386/* called with IRQs off */
1387static void increase_priority_inheritance_klmirqd(struct task_struct* klmirqd,
1388 struct task_struct* old_owner,
1389 struct task_struct* new_owner)
1390{
1391 BUG_ON(!(tsk_rt(klmirqd)->is_proxy_thread));
1392
1393 raw_spin_lock(&gsnedf_lock);
1394
1395 if(old_owner != new_owner)
1396 {
1397 if(old_owner)
1398 {
1399 // unreachable?
1400 tsk_rt(old_owner)->cur_klmirqd = NULL;
1401 }
1402
1403 TRACE_TASK(klmirqd, "giving ownership to %s/%d.\n",
1404 new_owner->comm, new_owner->pid);
1405
1406 tsk_rt(new_owner)->cur_klmirqd = klmirqd;
1407 }
1408
1409 __decrease_priority_inheritance(klmirqd, NULL); // kludge to clear out cur prio.
1410
1411 __increase_priority_inheritance(klmirqd,
1412 (tsk_rt(new_owner)->inh_task == NULL) ?
1413 new_owner :
1414 tsk_rt(new_owner)->inh_task);
1415
1416 raw_spin_unlock(&gsnedf_lock);
1417}
1418
1419
1420/* called with IRQs off */
1421static void decrease_priority_inheritance_klmirqd(struct task_struct* klmirqd,
1422 struct task_struct* old_owner,
1423 struct task_struct* new_owner)
1424{
1425 BUG_ON(!(tsk_rt(klmirqd)->is_proxy_thread));
1426
1427 raw_spin_lock(&gsnedf_lock);
1428
1429 TRACE_TASK(klmirqd, "priority restored\n");
1430
1431 __decrease_priority_inheritance(klmirqd, new_owner);
1432
1433 tsk_rt(old_owner)->cur_klmirqd = NULL;
1434
1435 raw_spin_unlock(&gsnedf_lock);
1436}
1437#endif
1438
1439
1440
1441 1389
1442#ifdef CONFIG_LITMUS_NESTED_LOCKING 1390#ifdef CONFIG_LITMUS_NESTED_LOCKING
1443 1391
@@ -1923,7 +1871,7 @@ static long gsnedf_activate_plugin(void)
1923#endif 1871#endif
1924 1872
1925#ifdef CONFIG_LITMUS_SOFTIRQD 1873#ifdef CONFIG_LITMUS_SOFTIRQD
1926 spawn_klmirqd(NULL); 1874 init_klmirqd();
1927#endif 1875#endif
1928 1876
1929#ifdef CONFIG_LITMUS_NVIDIA 1877#ifdef CONFIG_LITMUS_NVIDIA
@@ -1965,10 +1913,6 @@ static struct sched_plugin gsn_edf_plugin __cacheline_aligned_in_smp = {
1965#ifdef CONFIG_LITMUS_AFFINITY_LOCKING 1913#ifdef CONFIG_LITMUS_AFFINITY_LOCKING
1966 .allocate_aff_obs = gsnedf_allocate_affinity_observer, 1914 .allocate_aff_obs = gsnedf_allocate_affinity_observer,
1967#endif 1915#endif
1968#ifdef CONFIG_LITMUS_SOFTIRQD
1969 .increase_prio_klmirqd = increase_priority_inheritance_klmirqd,
1970 .decrease_prio_klmirqd = decrease_priority_inheritance_klmirqd,
1971#endif
1972#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 1916#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
1973 .enqueue_pai_tasklet = gsnedf_enqueue_pai_tasklet, 1917 .enqueue_pai_tasklet = gsnedf_enqueue_pai_tasklet,
1974 .change_prio_pai_tasklet = gsnedf_change_prio_pai_tasklet, 1918 .change_prio_pai_tasklet = gsnedf_change_prio_pai_tasklet,
diff --git a/litmus/sched_plugin.c b/litmus/sched_plugin.c
index cda67e0f6bc8..30c216fd6fdc 100644
--- a/litmus/sched_plugin.c
+++ b/litmus/sched_plugin.c
@@ -167,18 +167,6 @@ static int litmus_dummy___decrease_prio(struct task_struct* t, struct task_struc
167} 167}
168#endif 168#endif
169 169
170#ifdef CONFIG_LITMUS_SOFTIRQD
171static void litmus_dummy_increase_prio_klmirqd(struct task_struct* klmirqd,
172 struct task_struct* old_owner,
173 struct task_struct* new_owner)
174{
175}
176
177static void litmus_dummy_decrease_prio_klmirqd(struct task_struct* klmirqd,
178 struct task_struct* old_owner)
179{
180}
181#endif
182 170
183#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 171#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
184static int litmus_dummy_enqueue_pai_tasklet(struct tasklet_struct* t) 172static int litmus_dummy_enqueue_pai_tasklet(struct tasklet_struct* t)
@@ -263,10 +251,6 @@ struct sched_plugin linux_sched_plugin = {
263 .nested_decrease_prio = litmus_dummy_nested_decrease_prio, 251 .nested_decrease_prio = litmus_dummy_nested_decrease_prio,
264 .__compare = litmus_dummy___compare, 252 .__compare = litmus_dummy___compare,
265#endif 253#endif
266#ifdef CONFIG_LITMUS_SOFTIRQD
267 .increase_prio_klmirqd = litmus_dummy_increase_prio_klmirqd,
268 .decrease_prio_klmirqd = litmus_dummy_decrease_prio_klmirqd,
269#endif
270#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 254#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
271 .enqueue_pai_tasklet = litmus_dummy_enqueue_pai_tasklet, 255 .enqueue_pai_tasklet = litmus_dummy_enqueue_pai_tasklet,
272 .change_prio_pai_tasklet = litmus_dummy_change_prio_pai_tasklet, 256 .change_prio_pai_tasklet = litmus_dummy_change_prio_pai_tasklet,
@@ -327,10 +311,6 @@ int register_sched_plugin(struct sched_plugin* plugin)
327 CHECK(nested_decrease_prio); 311 CHECK(nested_decrease_prio);
328 CHECK(__compare); 312 CHECK(__compare);
329#endif 313#endif
330#ifdef CONFIG_LITMUS_SOFTIRQD
331 CHECK(increase_prio_klmirqd);
332 CHECK(decrease_prio_klmirqd);
333#endif
334#ifdef CONFIG_LITMUS_PAI_SOFTIRQD 314#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
335 CHECK(enqueue_pai_tasklet); 315 CHECK(enqueue_pai_tasklet);
336 CHECK(change_prio_pai_tasklet); 316 CHECK(change_prio_pai_tasklet);