hc
2023-12-11 d2ccde1c8e90d38cee87a1b0309ad2827f3fd30d
kernel/include/linux/preempt.h
....@@ -26,13 +26,13 @@
2626 * PREEMPT_MASK: 0x000000ff
2727 * SOFTIRQ_MASK: 0x0000ff00
2828 * HARDIRQ_MASK: 0x000f0000
29
- * NMI_MASK: 0x00100000
29
+ * NMI_MASK: 0x00f00000
3030 * PREEMPT_NEED_RESCHED: 0x80000000
3131 */
3232 #define PREEMPT_BITS 8
3333 #define SOFTIRQ_BITS 8
3434 #define HARDIRQ_BITS 4
35
-#define NMI_BITS 1
35
+#define NMI_BITS 4
3636
3737 #define PREEMPT_SHIFT 0
3838 #define SOFTIRQ_SHIFT (PREEMPT_SHIFT + PREEMPT_BITS)
....@@ -51,14 +51,7 @@
5151 #define HARDIRQ_OFFSET (1UL << HARDIRQ_SHIFT)
5252 #define NMI_OFFSET (1UL << NMI_SHIFT)
5353
54
-#ifndef CONFIG_PREEMPT_RT_FULL
55
-# define SOFTIRQ_DISABLE_OFFSET (2 * SOFTIRQ_OFFSET)
56
-#else
57
-# define SOFTIRQ_DISABLE_OFFSET (0)
58
-#endif
59
-
60
-/* We use the MSB mostly because its available */
61
-#define PREEMPT_NEED_RESCHED 0x80000000
54
+#define SOFTIRQ_DISABLE_OFFSET (2 * SOFTIRQ_OFFSET)
6255
6356 #define PREEMPT_DISABLED (PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
6457
....@@ -84,36 +77,37 @@
8477 /* preempt_count() and related functions, depends on PREEMPT_NEED_RESCHED */
8578 #include <asm/preempt.h>
8679
80
+#define nmi_count() (preempt_count() & NMI_MASK)
8781 #define hardirq_count() (preempt_count() & HARDIRQ_MASK)
88
-#define irq_count() (preempt_count() & (HARDIRQ_MASK | SOFTIRQ_MASK \
89
- | NMI_MASK))
90
-#ifndef CONFIG_PREEMPT_RT_FULL
91
-# define softirq_count() (preempt_count() & SOFTIRQ_MASK)
92
-# define in_serving_softirq() (softirq_count() & SOFTIRQ_OFFSET)
82
+#ifdef CONFIG_PREEMPT_RT
83
+# define softirq_count() (current->softirq_disable_cnt & SOFTIRQ_MASK)
9384 #else
94
-# define softirq_count() ((unsigned long)current->softirq_nestcnt)
95
-extern int in_serving_softirq(void);
85
+# define softirq_count() (preempt_count() & SOFTIRQ_MASK)
9686 #endif
87
+#define irq_count() (nmi_count() | hardirq_count() | softirq_count())
9788
9889 /*
99
- * Are we doing bottom half or hardware interrupt processing?
90
+ * Macros to retrieve the current execution context:
10091 *
101
- * in_irq() - We're in (hard) IRQ context
92
+ * in_nmi() - We're in NMI context
93
+ * in_hardirq() - We're in hard IRQ context
94
+ * in_serving_softirq() - We're in softirq context
95
+ * in_task() - We're in task context
96
+ */
97
+#define in_nmi() (nmi_count())
98
+#define in_hardirq() (hardirq_count())
99
+#define in_serving_softirq() (softirq_count() & SOFTIRQ_OFFSET)
100
+#define in_task() (!(in_nmi() | in_hardirq() | in_serving_softirq()))
101
+
102
+/*
103
+ * The following macros are deprecated and should not be used in new code:
104
+ * in_irq() - Obsolete version of in_hardirq()
102105 * in_softirq() - We have BH disabled, or are processing softirqs
103106 * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
104
- * in_serving_softirq() - We're in softirq context
105
- * in_nmi() - We're in NMI context
106
- * in_task() - We're in task context
107
- *
108
- * Note: due to the BH disabled confusion: in_softirq(),in_interrupt() really
109
- * should not be used in new code.
110107 */
111108 #define in_irq() (hardirq_count())
112109 #define in_softirq() (softirq_count())
113110 #define in_interrupt() (irq_count())
114
-#define in_nmi() (preempt_count() & NMI_MASK)
115
-#define in_task() (!(preempt_count() & \
116
- (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
117111
118112 /*
119113 * The preempt_count offset after preempt_disable();
....@@ -127,7 +121,7 @@
127121 /*
128122 * The preempt_count offset after spin_lock()
129123 */
130
-#if !defined(CONFIG_PREEMPT_RT_FULL)
124
+#if !defined(CONFIG_PREEMPT_RT)
131125 #define PREEMPT_LOCK_OFFSET PREEMPT_DISABLE_OFFSET
132126 #else
133127 #define PREEMPT_LOCK_OFFSET 0
....@@ -214,42 +208,17 @@
214208 preempt_count_dec(); \
215209 } while (0)
216210
217
-#ifdef CONFIG_PREEMPT_RT_BASE
211
+#ifndef CONFIG_PREEMPT_RT
218212 # define preempt_enable_no_resched() sched_preempt_enable_no_resched()
219
-# define preempt_check_resched_rt() preempt_check_resched()
213
+# define preempt_check_resched_rt() barrier();
220214 #else
221215 # define preempt_enable_no_resched() preempt_enable()
222
-# define preempt_check_resched_rt() barrier();
216
+# define preempt_check_resched_rt() preempt_check_resched()
223217 #endif
224218
225219 #define preemptible() (preempt_count() == 0 && !irqs_disabled())
226220
227
-#if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT_RT_BASE)
228
-
229
-extern void migrate_disable(void);
230
-extern void migrate_enable(void);
231
-
232
-int __migrate_disabled(struct task_struct *p);
233
-
234
-#elif !defined(CONFIG_SMP) && defined(CONFIG_PREEMPT_RT_BASE)
235
-
236
-extern void migrate_disable(void);
237
-extern void migrate_enable(void);
238
-static inline int __migrate_disabled(struct task_struct *p)
239
-{
240
- return 0;
241
-}
242
-
243
-#else
244
-#define migrate_disable() preempt_disable()
245
-#define migrate_enable() preempt_enable()
246
-static inline int __migrate_disabled(struct task_struct *p)
247
-{
248
- return 0;
249
-}
250
-#endif
251
-
252
-#ifdef CONFIG_PREEMPT
221
+#ifdef CONFIG_PREEMPTION
253222 #define preempt_enable() \
254223 do { \
255224 barrier(); \
....@@ -270,14 +239,19 @@
270239 __preempt_schedule(); \
271240 } while (0)
272241
242
+/*
243
+ * open code preempt_check_resched() because it is not exported to modules and
244
+ * used by local_unlock() or bpf_enable_instrumentation().
245
+ */
273246 #define preempt_lazy_enable() \
274247 do { \
275248 dec_preempt_lazy_count(); \
276249 barrier(); \
277
- preempt_check_resched(); \
250
+ if (should_resched(0)) \
251
+ __preempt_schedule(); \
278252 } while (0)
279253
280
-#else /* !CONFIG_PREEMPT */
254
+#else /* !CONFIG_PREEMPTION */
281255 #define preempt_enable() \
282256 do { \
283257 barrier(); \
....@@ -297,7 +271,7 @@
297271 } while (0)
298272
299273 #define preempt_check_resched() do { } while (0)
300
-#endif /* CONFIG_PREEMPT */
274
+#endif /* CONFIG_PREEMPTION */
301275
302276 #define preempt_disable_notrace() \
303277 do { \
....@@ -331,13 +305,9 @@
331305 #define preempt_check_resched_rt() barrier()
332306 #define preemptible() 0
333307
334
-#define migrate_disable() barrier()
335
-#define migrate_enable() barrier()
308
+#define preempt_lazy_disable() barrier()
309
+#define preempt_lazy_enable() barrier()
336310
337
-static inline int __migrate_disabled(struct task_struct *p)
338
-{
339
- return 0;
340
-}
341311 #endif /* CONFIG_PREEMPT_COUNT */
342312
343313 #ifdef MODULE
....@@ -360,7 +330,7 @@
360330 set_preempt_need_resched(); \
361331 } while (0)
362332
363
-#ifdef CONFIG_PREEMPT_RT_FULL
333
+#ifdef CONFIG_PREEMPT_RT
364334 # define preempt_disable_rt() preempt_disable()
365335 # define preempt_enable_rt() preempt_enable()
366336 # define preempt_disable_nort() barrier()
....@@ -422,4 +392,78 @@
422392
423393 #endif
424394
395
+#ifdef CONFIG_SMP
396
+
397
+/*
398
+ * Migrate-Disable and why it is undesired.
399
+ *
400
+ * When a preempted task becomes elegible to run under the ideal model (IOW it
401
+ * becomes one of the M highest priority tasks), it might still have to wait
402
+ * for the preemptee's migrate_disable() section to complete. Thereby suffering
403
+ * a reduction in bandwidth in the exact duration of the migrate_disable()
404
+ * section.
405
+ *
406
+ * Per this argument, the change from preempt_disable() to migrate_disable()
407
+ * gets us:
408
+ *
409
+ * - a higher priority tasks gains reduced wake-up latency; with preempt_disable()
410
+ * it would have had to wait for the lower priority task.
411
+ *
412
+ * - a lower priority tasks; which under preempt_disable() could've instantly
413
+ * migrated away when another CPU becomes available, is now constrained
414
+ * by the ability to push the higher priority task away, which might itself be
415
+ * in a migrate_disable() section, reducing it's available bandwidth.
416
+ *
417
+ * IOW it trades latency / moves the interference term, but it stays in the
418
+ * system, and as long as it remains unbounded, the system is not fully
419
+ * deterministic.
420
+ *
421
+ *
422
+ * The reason we have it anyway.
423
+ *
424
+ * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a
425
+ * number of primitives into becoming preemptible, they would also allow
426
+ * migration. This turns out to break a bunch of per-cpu usage. To this end,
427
+ * all these primitives employ migirate_disable() to restore this implicit
428
+ * assumption.
429
+ *
430
+ * This is a 'temporary' work-around at best. The correct solution is getting
431
+ * rid of the above assumptions and reworking the code to employ explicit
432
+ * per-cpu locking or short preempt-disable regions.
433
+ *
434
+ * The end goal must be to get rid of migrate_disable(), alternatively we need
435
+ * a schedulability theory that does not depend on abritrary migration.
436
+ *
437
+ *
438
+ * Notes on the implementation.
439
+ *
440
+ * The implementation is particularly tricky since existing code patterns
441
+ * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
442
+ * This means that it cannot use cpus_read_lock() to serialize against hotplug,
443
+ * nor can it easily migrate itself into a pending affinity mask change on
444
+ * migrate_enable().
445
+ *
446
+ *
447
+ * Note: even non-work-conserving schedulers like semi-partitioned depends on
448
+ * migration, so migrate_disable() is not only a problem for
449
+ * work-conserving schedulers.
450
+ *
451
+ */
452
+extern void migrate_disable(void);
453
+extern void migrate_enable(void);
454
+
455
+#else
456
+
457
+static inline void migrate_disable(void)
458
+{
459
+ preempt_lazy_disable();
460
+}
461
+
462
+static inline void migrate_enable(void)
463
+{
464
+ preempt_lazy_enable();
465
+}
466
+
467
+#endif /* CONFIG_SMP */
468
+
425469 #endif /* __LINUX_PREEMPT_H */