Add fallback if we have make a weird GC decision. (#50682)

If something odd happens during GC (the PC goes to sleep) or a very big
transient the heuristics might make a bad decision. What this PR
implements is if we try to make our target more than double the one we
had before we fallback to a more conservative method. This fixes the new
issue @vtjnash found in #40644
for me.

Author: gbaraldi

Make.inc

@@ -1503,7 +1503,7 @@ endef
 WINE ?= wine
 
 ifeq ($(BINARY),32)
-HEAPLIM := --heap-size-hint=500M
+HEAPLIM := --heap-size-hint=1000M
 else
 HEAPLIM :=
 endif

src/gc-debug.c

@@ -1224,18 +1224,18 @@ void _report_gc_finished(uint64_t pause, uint64_t freed, int full, int recollect
     if (!gc_logging_enabled) {
         return;
     }
-    jl_safe_printf("GC: pause %.2fms. collected %fMB. %s %s\n",
+    jl_safe_printf("\nGC: pause %.2fms. collected %fMB. %s %s\n",
         pause/1e6, freed/(double)(1<<20),
         full ? "full" : "incr",
         recollect ? "recollect" : ""
     );
 
-    jl_safe_printf("Heap stats: bytes_mapped %.2f MB, bytes_resident %.2f MB, heap_size %.2f MB, heap_target %.2f MB, live_bytes %.2f MB\n, Fragmentation %.3f",
+    jl_safe_printf("Heap stats: bytes_mapped %.2f MB, bytes_resident %.2f MB,\nheap_size %.2f MB, heap_target %.2f MB, Fragmentation %.3f\n",
         jl_atomic_load_relaxed(&gc_heap_stats.bytes_mapped)/(double)(1<<20),
         jl_atomic_load_relaxed(&gc_heap_stats.bytes_resident)/(double)(1<<20),
+        // live_bytes/(double)(1<<20), live byes tracking is not accurate.
         jl_atomic_load_relaxed(&gc_heap_stats.heap_size)/(double)(1<<20),
         jl_atomic_load_relaxed(&gc_heap_stats.heap_target)/(double)(1<<20),
-        live_bytes/(double)(1<<20),
         (double)live_bytes/(double)jl_atomic_load_relaxed(&gc_heap_stats.heap_size)
     );
     // Should fragmentation use bytes_resident instead of heap_size?

src/gc.c

@@ -1,6 +1,7 @@
 // This file is a part of Julia. License is MIT: https://julialang.org/license
 
 #include "gc.h"
+#include "julia.h"
 #include "julia_gcext.h"
 #include "julia_assert.h"
 #ifdef __GLIBC__
@@ -696,8 +697,8 @@ static uint64_t old_heap_size = 0;
 static uint64_t old_alloc_diff = 0;
 static uint64_t old_freed_diff = 0;
 static uint64_t gc_end_time = 0;
-
-
+static int thrash_counter = 0;
+static int thrashing = 0;
 // global variables for GC stats
 
 // Resetting the object to a young object, this is used when marking the
@@ -1170,7 +1171,10 @@ static void combine_thread_gc_counts(jl_gc_num_t *dest) JL_NOTSAFEPOINT
             dest->bigalloc += jl_atomic_load_relaxed(&ptls->gc_num.bigalloc);
             uint64_t alloc_acc = jl_atomic_load_relaxed(&ptls->gc_num.alloc_acc);
             uint64_t free_acc = jl_atomic_load_relaxed(&ptls->gc_num.free_acc);
+            dest->freed += jl_atomic_load_relaxed(&ptls->gc_num.free_acc);
             jl_atomic_store_relaxed(&gc_heap_stats.heap_size, alloc_acc - free_acc + jl_atomic_load_relaxed(&gc_heap_stats.heap_size));
+            jl_atomic_store_relaxed(&ptls->gc_num.alloc_acc, 0);
+            jl_atomic_store_relaxed(&ptls->gc_num.free_acc, 0);
         }
     }
 }
@@ -3266,9 +3270,6 @@ static int _jl_gc_collect(jl_ptls_t ptls, jl_gc_collection_t collection)
     // If the live data outgrows the suggested max_total_memory
     // we keep going with minimum intervals and full gcs until
     // we either free some space or get an OOM error.
-    if (live_bytes > max_total_memory) {
-        sweep_full = 1;
-    }
     if (gc_sweep_always_full) {
         sweep_full = 1;
     }
@@ -3320,7 +3321,6 @@ static int _jl_gc_collect(jl_ptls_t ptls, jl_gc_collection_t collection)
         gc_num.last_incremental_sweep = gc_end_time;
     }
 
-    int thrashing = 0; // maybe we should report this to the user or error out?
     size_t heap_size = jl_atomic_load_relaxed(&gc_heap_stats.heap_size);
     double target_allocs = 0.0;
     double min_interval = default_collect_interval;
@@ -3331,24 +3331,32 @@ static int _jl_gc_collect(jl_ptls_t ptls, jl_gc_collection_t collection)
         double collect_smooth_factor = 0.5;
         double tuning_factor = 0.03;
         double alloc_mem = jl_gc_smooth(old_alloc_diff, alloc_diff, alloc_smooth_factor);
-        double alloc_time = jl_gc_smooth(old_mut_time, mutator_time, alloc_smooth_factor);
+        double alloc_time = jl_gc_smooth(old_mut_time, mutator_time + sweep_time, alloc_smooth_factor); // Charge sweeping to the mutator
         double gc_mem = jl_gc_smooth(old_freed_diff, freed_diff, collect_smooth_factor);
-        double gc_time = jl_gc_smooth(old_pause_time, pause, collect_smooth_factor);
+        double gc_time = jl_gc_smooth(old_pause_time, pause - sweep_time, collect_smooth_factor);
         old_alloc_diff = alloc_diff;
         old_mut_time = mutator_time;
         old_freed_diff = freed_diff;
         old_pause_time = pause;
-        old_heap_size = heap_size;
-        thrashing = gc_time > mutator_time * 98 ? 1 : 0;
+        old_heap_size = heap_size; // TODO: Update these values dynamically instead of just during the GC
+        if (gc_time > alloc_time * 95 && !(thrash_counter < 4))
+            thrash_counter += 1;
+        else if (thrash_counter > 0)
+            thrash_counter -= 1;
         if (alloc_mem != 0 && alloc_time != 0 && gc_mem != 0 && gc_time != 0 ) {
             double alloc_rate = alloc_mem/alloc_time;
             double gc_rate = gc_mem/gc_time;
             target_allocs = sqrt(((double)heap_size/min_interval * alloc_rate)/(gc_rate * tuning_factor)); // work on multiples of min interval
         }
     }
-    if (target_allocs == 0.0 || thrashing) // If we are thrashing go back to default
-        target_allocs = 2*sqrt((double)heap_size/min_interval);
+    if (thrashing == 0 && thrash_counter >= 3)
+        thrashing = 1;
+    else if (thrashing == 1 && thrash_counter <= 2)
+        thrashing = 0; // maybe we should report this to the user or error out?
 
+    int bad_result = (target_allocs*min_interval + heap_size) > 2 * jl_atomic_load_relaxed(&gc_heap_stats.heap_target); // Don't follow through on a bad decision
+    if (target_allocs == 0.0 || thrashing || bad_result) // If we are thrashing go back to default
+        target_allocs = 2*sqrt((double)heap_size/min_interval);
     uint64_t target_heap = (uint64_t)target_allocs*min_interval + heap_size;
     if (target_heap > max_total_memory && !thrashing) // Allow it to go over if we are thrashing if we die we die
         target_heap = max_total_memory;
@@ -3612,10 +3620,10 @@ void jl_gc_init(void)
     total_mem = uv_get_total_memory();
     uint64_t constrained_mem = uv_get_constrained_memory();
     if (constrained_mem > 0 && constrained_mem < total_mem)
-        total_mem = constrained_mem;
+        jl_gc_set_max_memory(constrained_mem - 250*1024*1024); // LLVM + other libraries need some amount of memory
 #endif
     if (jl_options.heap_size_hint)
-        jl_gc_set_max_memory(jl_options.heap_size_hint);
+        jl_gc_set_max_memory(jl_options.heap_size_hint - 250*1024*1024);
 
     t_start = jl_hrtime();
 }
@@ -3718,7 +3726,26 @@ JL_DLLEXPORT void *jl_gc_counted_realloc_with_old_size(void *p, size_t old, size
                 jl_atomic_load_relaxed(&ptls->gc_num.allocd) + (sz - old));
         jl_atomic_store_relaxed(&ptls->gc_num.realloc,
             jl_atomic_load_relaxed(&ptls->gc_num.realloc) + 1);
-        jl_atomic_fetch_add_relaxed(&gc_heap_stats.heap_size, sz-old);
+
+        int64_t diff = sz - old;
+        if (diff < 0) {
+            uint64_t free_acc = jl_atomic_load_relaxed(&ptls->gc_num.free_acc);
+            if (free_acc + diff < 16*1024)
+                jl_atomic_store_relaxed(&ptls->gc_num.free_acc, free_acc + (-diff));
+            else {
+                jl_atomic_fetch_add_relaxed(&gc_heap_stats.heap_size, -(free_acc + (-diff)));
+                jl_atomic_store_relaxed(&ptls->gc_num.free_acc, 0);
+            }
+        }
+        else {
+            uint64_t alloc_acc = jl_atomic_load_relaxed(&ptls->gc_num.alloc_acc);
+            if (alloc_acc + diff < 16*1024)
+                jl_atomic_store_relaxed(&ptls->gc_num.alloc_acc, alloc_acc + diff);
+            else {
+                jl_atomic_fetch_add_relaxed(&gc_heap_stats.heap_size, alloc_acc + diff);
+                jl_atomic_store_relaxed(&ptls->gc_num.alloc_acc, 0);
+            }
+        }
     }
     return realloc(p, sz);
 }
@@ -3835,7 +3862,27 @@ static void *gc_managed_realloc_(jl_ptls_t ptls, void *d, size_t sz, size_t olds
             jl_atomic_load_relaxed(&ptls->gc_num.allocd) + (allocsz - oldsz));
     jl_atomic_store_relaxed(&ptls->gc_num.realloc,
         jl_atomic_load_relaxed(&ptls->gc_num.realloc) + 1);
-    jl_atomic_fetch_add_relaxed(&gc_heap_stats.heap_size, allocsz-oldsz);
+
+    int64_t diff = allocsz - oldsz;
+    if (diff < 0) {
+        uint64_t free_acc = jl_atomic_load_relaxed(&ptls->gc_num.free_acc);
+        if (free_acc + diff < 16*1024)
+            jl_atomic_store_relaxed(&ptls->gc_num.free_acc, free_acc + (-diff));
+        else {
+            jl_atomic_fetch_add_relaxed(&gc_heap_stats.heap_size, -(free_acc + (-diff)));
+            jl_atomic_store_relaxed(&ptls->gc_num.free_acc, 0);
+        }
+    }
+    else {
+        uint64_t alloc_acc = jl_atomic_load_relaxed(&ptls->gc_num.alloc_acc);
+        if (alloc_acc + diff < 16*1024)
+            jl_atomic_store_relaxed(&ptls->gc_num.alloc_acc, alloc_acc + diff);
+        else {
+            jl_atomic_fetch_add_relaxed(&gc_heap_stats.heap_size, alloc_acc + diff);
+            jl_atomic_store_relaxed(&ptls->gc_num.alloc_acc, 0);
+        }
+    }
+
     int last_errno = errno;
 #ifdef _OS_WINDOWS_
     DWORD last_error = GetLastError();

test/cmdlineargs.jl

@@ -974,6 +974,6 @@ end
         @test lines[3] == "foo"
         @test lines[4] == "bar"
     end
-#heap-size-hint
-@test readchomp(`$(Base.julia_cmd()) --startup-file=no --heap-size-hint=500M -e "println(@ccall jl_gc_get_max_memory()::UInt64)"`) == "524288000"
+#heap-size-hint, we reserve 250 MB for non GC memory (llvm, etc.)
+@test readchomp(`$(Base.julia_cmd()) --startup-file=no --heap-size-hint=500M -e "println(@ccall jl_gc_get_max_memory()::UInt64)"`) == "$((500-250)*1024*1024)"
 end

test/testenv.jl

@@ -38,12 +38,7 @@ if !@isdefined(testenv_defined)
     function addprocs_with_testenv(X; rr_allowed=true, kwargs...)
         exename = rr_allowed ? `$rr_exename $test_exename` : test_exename
         if X isa Integer
-            if Sys.iswindows()
-                heap_size=round(Int,(Sys.free_memory()/(1024^2)/(X+1)))
-                heap_size -= 300 # I don't know anymore
-            else
-                heap_size=round(Int,(Sys.total_memory()/(1024^2)/(X+1)))
-            end
+            heap_size=round(Int,(Sys.free_memory()/(1024^2)/(X+1)))
             push!(test_exeflags.exec, "--heap-size-hint=$(heap_size)M")
         end
         addprocs(X; exename=exename, exeflags=test_exeflags, kwargs...)