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)
....@@ -52,9 +52,6 @@
5252 #define NMI_OFFSET (1UL << NMI_SHIFT)
5353
5454 #define SOFTIRQ_DISABLE_OFFSET (2 * SOFTIRQ_OFFSET)
55
-
56
-/* We use the MSB mostly because its available */
57
-#define PREEMPT_NEED_RESCHED 0x80000000
5855
5956 #define PREEMPT_DISABLED (PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED)
6057
....@@ -80,31 +77,37 @@
8077 /* preempt_count() and related functions, depends on PREEMPT_NEED_RESCHED */
8178 #include <asm/preempt.h>
8279
80
+#define nmi_count() (preempt_count() & NMI_MASK)
8381 #define hardirq_count() (preempt_count() & HARDIRQ_MASK)
84
-#define softirq_count() (preempt_count() & SOFTIRQ_MASK)
85
-#define irq_count() (preempt_count() & (HARDIRQ_MASK | SOFTIRQ_MASK \
86
- | NMI_MASK))
82
+#ifdef CONFIG_PREEMPT_RT
83
+# define softirq_count() (current->softirq_disable_cnt & SOFTIRQ_MASK)
84
+#else
85
+# define softirq_count() (preempt_count() & SOFTIRQ_MASK)
86
+#endif
87
+#define irq_count() (nmi_count() | hardirq_count() | softirq_count())
8788
8889 /*
89
- * Are we doing bottom half or hardware interrupt processing?
90
+ * Macros to retrieve the current execution context:
9091 *
91
- * 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()
92105 * in_softirq() - We have BH disabled, or are processing softirqs
93106 * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled
94
- * in_serving_softirq() - We're in softirq context
95
- * in_nmi() - We're in NMI context
96
- * in_task() - We're in task context
97
- *
98
- * Note: due to the BH disabled confusion: in_softirq(),in_interrupt() really
99
- * should not be used in new code.
100107 */
101108 #define in_irq() (hardirq_count())
102109 #define in_softirq() (softirq_count())
103110 #define in_interrupt() (irq_count())
104
-#define in_serving_softirq() (softirq_count() & SOFTIRQ_OFFSET)
105
-#define in_nmi() (preempt_count() & NMI_MASK)
106
-#define in_task() (!(preempt_count() & \
107
- (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
108111
109112 /*
110113 * The preempt_count offset after preempt_disable();
....@@ -118,7 +121,11 @@
118121 /*
119122 * The preempt_count offset after spin_lock()
120123 */
124
+#if !defined(CONFIG_PREEMPT_RT)
121125 #define PREEMPT_LOCK_OFFSET PREEMPT_DISABLE_OFFSET
126
+#else
127
+#define PREEMPT_LOCK_OFFSET 0
128
+#endif
122129
123130 /*
124131 * The preempt_count offset needed for things like:
....@@ -167,11 +174,31 @@
167174 #define preempt_count_inc() preempt_count_add(1)
168175 #define preempt_count_dec() preempt_count_sub(1)
169176
177
+#ifdef CONFIG_PREEMPT_LAZY
178
+#define add_preempt_lazy_count(val) do { preempt_lazy_count() += (val); } while (0)
179
+#define sub_preempt_lazy_count(val) do { preempt_lazy_count() -= (val); } while (0)
180
+#define inc_preempt_lazy_count() add_preempt_lazy_count(1)
181
+#define dec_preempt_lazy_count() sub_preempt_lazy_count(1)
182
+#define preempt_lazy_count() (current_thread_info()->preempt_lazy_count)
183
+#else
184
+#define add_preempt_lazy_count(val) do { } while (0)
185
+#define sub_preempt_lazy_count(val) do { } while (0)
186
+#define inc_preempt_lazy_count() do { } while (0)
187
+#define dec_preempt_lazy_count() do { } while (0)
188
+#define preempt_lazy_count() (0)
189
+#endif
190
+
170191 #ifdef CONFIG_PREEMPT_COUNT
171192
172193 #define preempt_disable() \
173194 do { \
174195 preempt_count_inc(); \
196
+ barrier(); \
197
+} while (0)
198
+
199
+#define preempt_lazy_disable() \
200
+do { \
201
+ inc_preempt_lazy_count(); \
175202 barrier(); \
176203 } while (0)
177204
....@@ -181,11 +208,17 @@
181208 preempt_count_dec(); \
182209 } while (0)
183210
184
-#define preempt_enable_no_resched() sched_preempt_enable_no_resched()
211
+#ifndef CONFIG_PREEMPT_RT
212
+# define preempt_enable_no_resched() sched_preempt_enable_no_resched()
213
+# define preempt_check_resched_rt() barrier();
214
+#else
215
+# define preempt_enable_no_resched() preempt_enable()
216
+# define preempt_check_resched_rt() preempt_check_resched()
217
+#endif
185218
186219 #define preemptible() (preempt_count() == 0 && !irqs_disabled())
187220
188
-#ifdef CONFIG_PREEMPT
221
+#ifdef CONFIG_PREEMPTION
189222 #define preempt_enable() \
190223 do { \
191224 barrier(); \
....@@ -206,11 +239,29 @@
206239 __preempt_schedule(); \
207240 } while (0)
208241
209
-#else /* !CONFIG_PREEMPT */
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
+ */
246
+#define preempt_lazy_enable() \
247
+do { \
248
+ dec_preempt_lazy_count(); \
249
+ barrier(); \
250
+ if (should_resched(0)) \
251
+ __preempt_schedule(); \
252
+} while (0)
253
+
254
+#else /* !CONFIG_PREEMPTION */
210255 #define preempt_enable() \
211256 do { \
212257 barrier(); \
213258 preempt_count_dec(); \
259
+} while (0)
260
+
261
+#define preempt_lazy_enable() \
262
+do { \
263
+ dec_preempt_lazy_count(); \
264
+ barrier(); \
214265 } while (0)
215266
216267 #define preempt_enable_notrace() \
....@@ -220,7 +271,7 @@
220271 } while (0)
221272
222273 #define preempt_check_resched() do { } while (0)
223
-#endif /* CONFIG_PREEMPT */
274
+#endif /* CONFIG_PREEMPTION */
224275
225276 #define preempt_disable_notrace() \
226277 do { \
....@@ -251,7 +302,11 @@
251302 #define preempt_disable_notrace() barrier()
252303 #define preempt_enable_no_resched_notrace() barrier()
253304 #define preempt_enable_notrace() barrier()
305
+#define preempt_check_resched_rt() barrier()
254306 #define preemptible() 0
307
+
308
+#define preempt_lazy_disable() barrier()
309
+#define preempt_lazy_enable() barrier()
255310
256311 #endif /* CONFIG_PREEMPT_COUNT */
257312
....@@ -271,9 +326,21 @@
271326 } while (0)
272327 #define preempt_fold_need_resched() \
273328 do { \
274
- if (tif_need_resched()) \
329
+ if (tif_need_resched_now()) \
275330 set_preempt_need_resched(); \
276331 } while (0)
332
+
333
+#ifdef CONFIG_PREEMPT_RT
334
+# define preempt_disable_rt() preempt_disable()
335
+# define preempt_enable_rt() preempt_enable()
336
+# define preempt_disable_nort() barrier()
337
+# define preempt_enable_nort() barrier()
338
+#else
339
+# define preempt_disable_rt() barrier()
340
+# define preempt_enable_rt() barrier()
341
+# define preempt_disable_nort() preempt_disable()
342
+# define preempt_enable_nort() preempt_enable()
343
+#endif
277344
278345 #ifdef CONFIG_PREEMPT_NOTIFIERS
279346
....@@ -325,4 +392,78 @@
325392
326393 #endif
327394
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
+
328469 #endif /* __LINUX_PREEMPT_H */