xnu-4903.231.4.tar.gz

[apple/xnu.git] / osfmk / kern / sched_average.c

diff --git a/osfmk/kern/sched_average.c b/osfmk/kern/sched_average.c
index 411dfb47cb7128c79d6a39b868ff87e0c7bb9009..e7b24bb0d415145621c26f609c5a622b9cad953d 100644 (file)
--- a/osfmk/kern/sched_average.c
+++ b/osfmk/kern/sched_average.c
@@ -77,6 +77,8 @@
 uint32_t       avenrun[3] = {0, 0, 0};
 uint32_t       mach_factor[3] = {0, 0, 0};
 
 uint32_t       avenrun[3] = {0, 0, 0};
 uint32_t       mach_factor[3] = {0, 0, 0};
 

+uint32_t       sched_load_average, sched_mach_factor;
+

 #if defined(CONFIG_SCHED_TIMESHARE_CORE)
 /*
  * Values are scaled by LOAD_SCALE, defined in processor_info.h
 #if defined(CONFIG_SCHED_TIMESHARE_CORE)
 /*
  * Values are scaled by LOAD_SCALE, defined in processor_info.h


@@ -108,8 +110,6 @@ static struct sched_average {
 } sched_average[] = {
        { compute_averunnable, &sched_nrun, 5, 0 },
        { compute_stack_target, NULL, 5, 1 },
 } sched_average[] = {
        { compute_averunnable, &sched_nrun, 5, 0 },
        { compute_stack_target, NULL, 5, 1 },

-       { compute_memory_pressure, NULL, 1, 0 },
-       { compute_zone_gc_throttle, NULL, 60, 0 },

        { compute_pageout_gc_throttle, NULL, 1, 0 },
        { compute_pmap_gc_throttle, NULL, 60, 0 },
 #if CONFIG_TELEMETRY
        { compute_pageout_gc_throttle, NULL, 1, 0 },
        { compute_pmap_gc_throttle, NULL, 60, 0 },
 #if CONFIG_TELEMETRY


@@ -120,127 +120,210 @@ static struct sched_average {
 
 typedef struct sched_average   *sched_average_t;
 
 
 typedef struct sched_average   *sched_average_t;
 

-/* The "stdelta" parameter represents the number of scheduler maintenance
- * "ticks" that have elapsed since the last invocation, subject to
- * integer division imprecision.
+/*
+ * Scheduler load calculation algorithm
+ *
+ * The scheduler load values provide an estimate of the number of runnable 
+ * timeshare threads in the system at various priority bands. The load 
+ * ultimately affects the priority shifts applied to all threads in a band 
+ * causing them to timeshare with other threads in the system. The load is 
+ * maintained in buckets, with each bucket corresponding to a priority band.
+ *
+ * Each runnable thread on the system contributes its load to its priority 
+ * band and to the bands above it. The contribution of a thread to the bands 
+ * above it is not strictly 1:1 and is weighted based on the priority band 
+ * of the thread. The rules of thread load contribution to each of its higher 
+ * bands are as follows:
+ *
+ * - DF threads: Upto (2 * NCPUs) threads
+ * - UT threads: Upto NCPUs threads
+ * - BG threads: Upto 1 thread
+ *
+ * To calculate the load values, the various run buckets are sampled (every 
+ * sched_load_compute_interval_abs) and the weighted contributions of the the
+ * lower bucket threads are added. The resultant value is plugged into an 
+ * exponentially weighted moving average formula: 
+ *     new-load = alpha * old-load + (1 - alpha) * run-bucket-sample-count
+ *     (where, alpha < 1)
+ * The calculations for the scheduler load are done using fixpoint math with 
+ * a scale factor of 16 to avoid expensive divides and floating point 
+ * operations. The final load values are a smooth curve representative of 
+ * the actual number of runnable threads in a priority band.
+ */
+
+/* Maintains the current (scaled for fixpoint) load in various buckets */
+uint32_t sched_load[TH_BUCKET_MAX];
+
+/* 
+ * Alpha factor for the EWMA alogrithm. The current values are chosen as 
+ * 6:10 ("old load":"new samples") to make sure the scheduler reacts fast 
+ * enough to changing system load but does not see too many spikes from bursty 
+ * activity. The current values ensure that the scheduler would converge 
+ * to the latest load in 2-3 sched_load_compute_interval_abs intervals 
+ * (which amounts to ~30-45ms with current values).

  */
  */

+#define SCHED_LOAD_EWMA_ALPHA_OLD      6
+#define SCHED_LOAD_EWMA_ALPHA_NEW      10
+#define SCHED_LOAD_EWMA_ALPHA_SHIFT    4
+static_assert((SCHED_LOAD_EWMA_ALPHA_OLD + SCHED_LOAD_EWMA_ALPHA_NEW) == (1ul << SCHED_LOAD_EWMA_ALPHA_SHIFT));

 
 

+/* For fixpoint EWMA, roundup the load to make it converge */
+#define SCHED_LOAD_EWMA_ROUNDUP(load)  (((load) & (1ul << (SCHED_LOAD_EWMA_ALPHA_SHIFT - 1))) != 0)
+
+/* Macro to convert scaled sched load to a real load value */
+#define SCHED_LOAD_EWMA_UNSCALE(load)  (((load) >> SCHED_LOAD_EWMA_ALPHA_SHIFT) + SCHED_LOAD_EWMA_ROUNDUP(load))
+
+/*
+ * Routine to capture the latest runnable counts and update sched_load */

 void
 void

-compute_averages(uint64_t stdelta)
+compute_sched_load(void)

 {
 {

-       int                     ncpus, nthreads, nshared, nbackground, nshared_non_bg;
-       uint32_t                factor_now, average_now, load_now = 0, background_load_now = 0, combined_fgbg_load_now = 0;
-       sched_average_t         avg;
-       uint64_t                abstime, index;
-

        /*
        /*

-        *      Retrieve counts, ignoring
-        *      the current thread.
+        * Retrieve a snapshot of the current run counts.
+        *
+        * Why not a bcopy()? Because we need atomic word-sized reads of sched_run_buckets,
+        * not byte-by-byte copy.

         */
         */

-       ncpus = processor_avail_count;
-       nthreads = sched_run_count - 1;
-       nshared = sched_share_count;
-       nbackground = sched_background_count;
+       uint32_t ncpus = processor_avail_count;
+       uint32_t load_now[TH_BUCKET_MAX];

 
 

-       /*
-        *      Load average and mach factor calculations for
-        *      those which ask about these things.
-        */
-       average_now = nthreads * LOAD_SCALE;
+       load_now[TH_BUCKET_RUN]      = sched_run_buckets[TH_BUCKET_RUN];
+       load_now[TH_BUCKET_FIXPRI]   = sched_run_buckets[TH_BUCKET_FIXPRI];
+       load_now[TH_BUCKET_SHARE_FG] = sched_run_buckets[TH_BUCKET_SHARE_FG];
+       load_now[TH_BUCKET_SHARE_DF] = sched_run_buckets[TH_BUCKET_SHARE_DF];
+       load_now[TH_BUCKET_SHARE_UT] = sched_run_buckets[TH_BUCKET_SHARE_UT];
+       load_now[TH_BUCKET_SHARE_BG] = sched_run_buckets[TH_BUCKET_SHARE_BG];

 
 

-       if (nthreads > ncpus)
-               factor_now = (ncpus * LOAD_SCALE) / (nthreads + 1);
-       else
-               factor_now = (ncpus - nthreads) * LOAD_SCALE;
+       assert(load_now[TH_BUCKET_RUN] >= 0);
+       assert(load_now[TH_BUCKET_FIXPRI] >= 0);
+
+       uint32_t nthreads = load_now[TH_BUCKET_RUN];
+       uint32_t nfixpri  = load_now[TH_BUCKET_FIXPRI];
+
+       KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+               MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_LOAD) | DBG_FUNC_NONE,
+               load_now[TH_BUCKET_FIXPRI], (load_now[TH_BUCKET_SHARE_FG] + load_now[TH_BUCKET_SHARE_DF]),
+               load_now[TH_BUCKET_SHARE_BG],   load_now[TH_BUCKET_SHARE_UT], 0);

 
 

-       /* For those statistics that formerly relied on being recomputed
-        * on timer ticks, advance by the approximate number of corresponding
-        * elapsed intervals, thus compensating for potential idle intervals.
-        */
-       for (index = 0; index < stdelta; index++) {
-               sched_mach_factor = ((sched_mach_factor << 2) + factor_now) / 5;
-               sched_load_average = ((sched_load_average << 2) + average_now) / 5;
-       }

        /*
         * Compute the timeshare priority conversion factor based on loading.
         * Because our counters may be incremented and accessed
         * concurrently with respect to each other, we may have
        /*
         * Compute the timeshare priority conversion factor based on loading.
         * Because our counters may be incremented and accessed
         * concurrently with respect to each other, we may have

-        * windows where the invariant nthreads >= nshared >= nbackground
+        * windows where the invariant (nthreads - nfixpri) == (fg + df + bg + ut)

         * is broken, so truncate values in these cases.
         */
         * is broken, so truncate values in these cases.
         */

+       uint32_t timeshare_threads = (nthreads - nfixpri);
+       for (uint32_t i = TH_BUCKET_SHARE_FG; i <= TH_BUCKET_SHARE_BG ; i++) {
+               if (load_now[i] > timeshare_threads)
+                       load_now[i] = timeshare_threads;
+       }

 
 

-       if (nshared > nthreads)
-               nshared = nthreads;
-
-       if (nbackground > nshared)
-               nbackground = nshared;
-
-       nshared_non_bg = nshared - nbackground;
-
-       if (nshared_non_bg > ncpus) {
-               if (ncpus > 1)
-                       load_now = nshared_non_bg / ncpus;
-               else
-                       load_now = nshared_non_bg;
+       /* 
+        * Default threads contribute up to (NCPUS * 2) of load to FG threads 
+        */
+       if (load_now[TH_BUCKET_SHARE_DF] <= (ncpus * 2)) {
+               load_now[TH_BUCKET_SHARE_FG] += load_now[TH_BUCKET_SHARE_DF];
+       } else {
+               load_now[TH_BUCKET_SHARE_FG] += (ncpus * 2);
+       }
+       
+       /*
+        * Utility threads contribute up to NCPUS of load to FG & DF threads
+        */
+       if (load_now[TH_BUCKET_SHARE_UT] <= ncpus) {
+               load_now[TH_BUCKET_SHARE_FG] += load_now[TH_BUCKET_SHARE_UT];
+               load_now[TH_BUCKET_SHARE_DF] += load_now[TH_BUCKET_SHARE_UT];
+       } else {
+               load_now[TH_BUCKET_SHARE_FG] += ncpus;
+               load_now[TH_BUCKET_SHARE_DF] += ncpus;
+       }

 
 

-               if (load_now > NRQS - 1)
-                       load_now = NRQS - 1;
+       /*
+        * BG threads contribute up to 1 thread worth of load to FG, DF and UT threads
+        */
+       if (load_now[TH_BUCKET_SHARE_BG] > 0) {
+               load_now[TH_BUCKET_SHARE_FG] += 1;
+               load_now[TH_BUCKET_SHARE_DF] += 1;
+               load_now[TH_BUCKET_SHARE_UT] += 1;

        }
 
        }
 

-       if (nbackground > ncpus) {
-               if (ncpus > 1)
-                       background_load_now = nbackground / ncpus;
-               else
-                       background_load_now = nbackground;
+       /*
+        * The conversion factor consists of two components:
+        * a fixed value based on the absolute time unit (sched_fixed_shift),
+        * and a dynamic portion based on load (sched_load_shifts).
+        *
+        * Zero load results in a out of range shift count.
+        */

 
 

-               if (background_load_now > NRQS - 1)
-                       background_load_now = NRQS - 1;
-       }
+       for (uint32_t i = TH_BUCKET_SHARE_FG; i <= TH_BUCKET_SHARE_BG ; i++) {
+               uint32_t bucket_load = 0;

 
 

-       if (nshared > ncpus) {
-               if (ncpus > 1)
-                       combined_fgbg_load_now = nshared / ncpus;
-               else
-                       combined_fgbg_load_now = nshared;
+               if (load_now[i] > ncpus) {
+                       /* Normalize the load to number of CPUs */
+                       if (ncpus > 1)
+                               bucket_load = load_now[i] / ncpus;
+                       else
+                               bucket_load = load_now[i];

 
 

-               if (combined_fgbg_load_now > NRQS - 1)
-                       combined_fgbg_load_now = NRQS - 1;
+                       if (bucket_load > MAX_LOAD)
+                               bucket_load = MAX_LOAD;
+               }
+               /* Plug the load values into the EWMA algorithm to calculate (scaled for fixpoint) sched_load */
+               sched_load[i] = (sched_load[i] * SCHED_LOAD_EWMA_ALPHA_OLD) + ((bucket_load << SCHED_LOAD_EWMA_ALPHA_SHIFT) * SCHED_LOAD_EWMA_ALPHA_NEW);
+               sched_load[i] = sched_load[i] >> SCHED_LOAD_EWMA_ALPHA_SHIFT;

        }
 
        KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
        }
 
        KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,

-               MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_LOAD) | DBG_FUNC_NONE,
-               (nthreads - nshared), (nshared - nbackground), nbackground, 0, 0);
+               MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_LOAD_EFFECTIVE) | DBG_FUNC_NONE, 
+                       SCHED_LOAD_EWMA_UNSCALE(sched_load[TH_BUCKET_SHARE_FG]), SCHED_LOAD_EWMA_UNSCALE(sched_load[TH_BUCKET_SHARE_DF]),
+                       SCHED_LOAD_EWMA_UNSCALE(sched_load[TH_BUCKET_SHARE_UT]), SCHED_LOAD_EWMA_UNSCALE(sched_load[TH_BUCKET_SHARE_BG]), 0);
+}
+
+void
+compute_averages(uint64_t stdelta)
+{
+       
+       uint32_t nthreads = sched_run_buckets[TH_BUCKET_RUN] - 1;
+       uint32_t ncpus = processor_avail_count;
+       
+       /* Update the global pri_shifts based on the latest values */
+       for (uint32_t i = TH_BUCKET_SHARE_FG; i <= TH_BUCKET_SHARE_BG ; i++) {
+               uint32_t bucket_load = SCHED_LOAD_EWMA_UNSCALE(sched_load[i]);
+               sched_pri_shifts[i] = sched_fixed_shift - sched_load_shifts[bucket_load];
+       }

 
        /*
 
        /*

-        *      Sample total running threads.
+        * Sample total running threads for the load average calculation.

         */
        sched_nrun = nthreads;
         */
        sched_nrun = nthreads;

-       
-#if defined(CONFIG_SCHED_TIMESHARE_CORE)

 
        /*
 
        /*

-        *      The conversion factor consists of
-        *      two components: a fixed value based
-        *      on the absolute time unit, and a
-        *      dynamic portion based on loading.
-        *
-        *      Zero loading results in a out of range
-        *      shift count.  Accumulated usage is ignored
-        *      during conversion and new usage deltas
-        *      are discarded.
+        * Load average and mach factor calculations for
+        * those which ask about these things.

         */
         */

-       sched_pri_shift = sched_fixed_shift - sched_load_shifts[load_now];
-       sched_background_pri_shift = sched_fixed_shift - sched_load_shifts[background_load_now];
-       sched_combined_fgbg_pri_shift = sched_fixed_shift - sched_load_shifts[combined_fgbg_load_now];
+       uint32_t average_now = nthreads * LOAD_SCALE;
+       uint32_t factor_now;
+
+       if (nthreads > ncpus)
+               factor_now = (ncpus * LOAD_SCALE) / (nthreads + 1);
+       else
+               factor_now = (ncpus - nthreads) * LOAD_SCALE;

 
        /*
 
        /*

-        * Compute old-style Mach load averages.
+        * For those statistics that formerly relied on being recomputed
+        * on timer ticks, advance by the approximate number of corresponding
+        * elapsed intervals, thus compensating for potential idle intervals.

         */
         */

+       for (uint32_t index = 0; index < stdelta; index++) {
+               sched_mach_factor = ((sched_mach_factor << 2) + factor_now) / 5;
+               sched_load_average = ((sched_load_average << 2) + average_now) / 5;
+       }

 
 

-       for (index = 0; index < stdelta; index++) {
-               register int            i;
-
-               for (i = 0; i < 3; i++) {
+       /*
+        * Compute old-style Mach load averages.
+        */
+       for (uint32_t index = 0; index < stdelta; index++) {
+               for (uint32_t i = 0; i < 3; i++) {

                        mach_factor[i] = ((mach_factor[i] * fract[i]) +
                                                (factor_now * (LOAD_SCALE - fract[i]))) / LOAD_SCALE;
 
                        mach_factor[i] = ((mach_factor[i] * fract[i]) +
                                                (factor_now * (LOAD_SCALE - fract[i]))) / LOAD_SCALE;
 


@@ -248,13 +331,13 @@ compute_averages(uint64_t stdelta)
                                                (average_now * (LOAD_SCALE - fract[i]))) / LOAD_SCALE;
                }
        }
                                                (average_now * (LOAD_SCALE - fract[i]))) / LOAD_SCALE;
                }
        }

-#endif /* CONFIG_SCHED_TIMESHARE_CORE */

 
        /*
 
        /*

-        *      Compute averages in other components.
+        * Compute averages in other components.

         */
         */

-       abstime = mach_absolute_time();
-       for (avg = sched_average; avg->comp != NULL; ++avg) {
+       uint64_t abstime = mach_absolute_time();
+
+       for (sched_average_t avg = sched_average; avg->comp != NULL; ++avg) {

                if (abstime >= avg->deadline) {
                        uint64_t period_abs = (avg->period * sched_one_second_interval);
                        uint64_t ninvokes = 1;
                if (abstime >= avg->deadline) {
                        uint64_t period_abs = (avg->period * sched_one_second_interval);
                        uint64_t ninvokes = 1;


@@ -262,7 +345,7 @@ compute_averages(uint64_t stdelta)
                        ninvokes += (abstime - avg->deadline) / period_abs;
                        ninvokes = MIN(ninvokes, SCHED_TICK_MAX_DELTA);
 
                        ninvokes += (abstime - avg->deadline) / period_abs;
                        ninvokes = MIN(ninvokes, SCHED_TICK_MAX_DELTA);
 

-                       for (index = 0; index < ninvokes; index++) {
+                       for (uint32_t index = 0; index < ninvokes; index++) {

                                (*avg->comp)(avg->param);
                        }
                        avg->deadline = abstime + period_abs;
                                (*avg->comp)(avg->param);
                        }
                        avg->deadline = abstime + period_abs;