diff mbox series

[v3,1/1] memalign: Support scanning for aligned chunks.

Message ID xnilnsbp3e.fsf@greed.delorie.com
State New
Headers show
Series [v3,1/1] memalign: Support scanning for aligned chunks. | expand

Commit Message

DJ Delorie July 20, 2022, 12:32 a.m. UTC
Florian Weimer <fweimer@redhat.com> writes:
> MAYBE_INIT_TCACHE does not seem necessary.  If the tcache is empty,
> there's no need to initialize it and scane it.  And why not scan the
> entire chain here?  It's another source of missed chunk reuse.
>
> The tcache null check should be written as tcache != NULL.

[v3: cleaned up all the != NULL cases, added messy logic to remove
non-head tcache chunks]

This patch adds a chunk scanning algorithm to the _int_memalign code
path that reduces heap fragmentation by reusing already aligned chunks
instead of always looking for chunks of larger sizes and splitting
them.  The tcache macros are extended to allow removing a chunk from
the middle of the list.

The goal is to fix the pathological use cases where heaps grow
continuously in workloads that are heavy users of memalign.

Note that tst-memalign-2 checks for tcache operation, which
malloc-check bypasses.

Comments

DJ Delorie July 22, 2022, 8:21 p.m. UTC | #1
Sidebar - I ran the malloc benchtests on before/after builds, and the
variation between the two seemed no more or less than the variation
between just running the "before" tests twice.

The variation, just running the same tests multiple times, seems larger
than the gains/losses we typically get with malloc tweaks, though.
Computers are just too fast and unpredictable these days ;-)

I also have a python script that compares malloc benchtest runs (below).
Run like this:

 $ compare-malloc.py ~/tools/upstream/glibc.pristine.build/benchtests ~/tools/upstream/glibc.memalign.build/benchtests `ls -1d bench-malloc-simple-*.out | nsort` > /tmp/malloc2.csv

If there's interest, I can add this to benchtests/scripts

---------- 8< ----------
#!/usr/bin/python

import sys
import json
import argparse
import pprint
    
def do_keys (pattern, bd, ad):
    avg = 0
    count = 0
    print_it = 0;
    for b in sorted(bd.keys()):
        if pattern in b:
            if print_it:
                print ("%s,%s,%s,%s" % (save_b, bv, av, imp));
            save_b = b;
            bv = bd[b]
            av = ad[b]
            imp = av / bv;
            avg += imp
            count = count + 1
            print_it = 1

    if print_it:
        print ("%s,%s,%s,%s,%s" % (save_b, bv, av, imp, avg/count));

def compare_files(fn1, fn2, t):
    b = open (fn1, 'r')
    before = json.load(b)
    a = open (fn2, 'r')
    after = json.load(a)

    print (t)
    print ("Test,Before,After,Ratio,Average")
    do_keys ('max_rss', before['functions']['malloc'][''], after['functions']['malloc'][''])
    do_keys ('main_arena', before['functions']['malloc'][''], after['functions']['malloc'][''])
    do_keys ('thread_arena', before['functions']['malloc'][''], after['functions']['malloc'][''])
    print ("");

def main(argv):
    before_dir = argv[1];
    after_dir = argv[2];
    files = argv[3:]

    for f in files:
        bf = before_dir + "/" + f;
        af = after_dir + "/" + f;
        compare_files (bf, af, f)

if __name__ == '__main__':
    main(sys.argv)
Joseph Myers July 22, 2022, 8:28 p.m. UTC | #2
On Tue, 19 Jul 2022, DJ Delorie via Libc-alpha wrote:

> +#define TEST_FUNCTION do_test ()
> +#include "../test-skeleton.c"

New tests should be written to use <support/test-driver.c> rather than the 
old test-skeleton.c wrapper round it.
diff mbox series

Patch

diff --git a/malloc/Makefile b/malloc/Makefile
index 4e32de2a0b..084c408ac7 100644
--- a/malloc/Makefile
+++ b/malloc/Makefile
@@ -43,6 +43,7 @@  tests := mallocbug tst-malloc tst-valloc tst-calloc tst-obstack \
 	 tst-tcfree1 tst-tcfree2 tst-tcfree3 \
 	 tst-safe-linking \
 	 tst-mallocalign1 \
+	 tst-memalign-2
 
 tests-static := \
 	 tst-interpose-static-nothread \
@@ -72,7 +73,7 @@  test-srcs = tst-mtrace
 # with MALLOC_CHECK_=3 because they expect a specific failure.
 tests-exclude-malloc-check = tst-malloc-check tst-malloc-usable \
 	tst-mxfast tst-safe-linking \
-	tst-compathooks-off tst-compathooks-on
+	tst-compathooks-off tst-compathooks-on tst-memalign-2
 
 # Run all tests with MALLOC_CHECK_=3
 tests-malloc-check = $(filter-out $(tests-exclude-malloc-check) \
diff --git a/malloc/malloc.c b/malloc/malloc.c
index 12908b8f97..364125e6b4 100644
--- a/malloc/malloc.c
+++ b/malloc/malloc.c
@@ -3185,19 +3185,44 @@  tcache_put (mchunkptr chunk, size_t tc_idx)
 }
 
 /* Caller must ensure that we know tc_idx is valid and there's
-   available chunks to remove.  */
+   available chunks to remove.  Removes chunk from the middle of the
+   list.  */
 static __always_inline void *
-tcache_get (size_t tc_idx)
+tcache_get_n (size_t tc_idx, tcache_entry **ep)
 {
-  tcache_entry *e = tcache->entries[tc_idx];
+  tcache_entry *e;
+  if (ep == &(tcache->entries[tc_idx]))
+    e = *ep;
+  else
+    e = REVEAL_PTR (*ep);
+
   if (__glibc_unlikely (!aligned_OK (e)))
     malloc_printerr ("malloc(): unaligned tcache chunk detected");
-  tcache->entries[tc_idx] = REVEAL_PTR (e->next);
+
+  if (ep == &(tcache->entries[tc_idx]))
+      *ep = REVEAL_PTR (e->next);
+  else
+    *ep = PROTECT_PTR (ep, REVEAL_PTR (e->next));
+
   --(tcache->counts[tc_idx]);
   e->key = 0;
   return (void *) e;
 }
 
+/* Like the above, but removes from the head of the list.  */
+static __always_inline void *
+tcache_get (size_t tc_idx)
+{
+  return tcache_get_n (tc_idx, & tcache->entries[tc_idx]);
+}
+
+/* Iterates through the tcache linked list.  */
+static __always_inline void *
+tcache_next (tcache_entry *e)
+{
+  return (tcache_entry *) REVEAL_PTR (e->next);
+}
+
 static void
 tcache_thread_shutdown (void)
 {
@@ -3306,7 +3331,7 @@  __libc_malloc (size_t bytes)
 
   DIAG_PUSH_NEEDS_COMMENT;
   if (tc_idx < mp_.tcache_bins
-      && tcache
+      && tcache != NULL
       && tcache->counts[tc_idx] > 0)
     {
       victim = tcache_get (tc_idx);
@@ -3557,6 +3582,38 @@  _mid_memalign (size_t alignment, size_t bytes, void *address)
       alignment = a;
     }
 
+#if USE_TCACHE
+  {
+    size_t tbytes;
+    tbytes = checked_request2size (bytes);
+    if (tbytes == 0)
+      {
+	__set_errno (ENOMEM);
+	return NULL;
+      }
+    size_t tc_idx = csize2tidx (tbytes);
+
+    if (tc_idx < mp_.tcache_bins
+	&& tcache != NULL
+	&& tcache->counts[tc_idx] > 0)
+      {
+	/* The tcache itself isn't encoded, but the chain is.  */
+	tcache_entry **tep = & tcache->entries[tc_idx];
+	tcache_entry *te = *tep;
+	while (te != NULL && ((intptr_t)te & (alignment - 1)) != 0)
+	  {
+	    tep = & (te->next);
+	    te = tcache_next (te);
+	  }
+	if (te != NULL)
+	  {
+	    void *victim = tcache_get_n (tc_idx, tep);
+	    return tag_new_usable (victim);
+	  }
+      }
+  }
+#endif
+
   if (SINGLE_THREAD_P)
     {
       p = _int_memalign (&main_arena, alignment, bytes);
@@ -3862,7 +3919,7 @@  _int_malloc (mstate av, size_t bytes)
 	      /* While we're here, if we see other chunks of the same size,
 		 stash them in the tcache.  */
 	      size_t tc_idx = csize2tidx (nb);
-	      if (tcache && tc_idx < mp_.tcache_bins)
+	      if (tcache != NULL && tc_idx < mp_.tcache_bins)
 		{
 		  mchunkptr tc_victim;
 
@@ -3920,7 +3977,7 @@  _int_malloc (mstate av, size_t bytes)
 	  /* While we're here, if we see other chunks of the same size,
 	     stash them in the tcache.  */
 	  size_t tc_idx = csize2tidx (nb);
-	  if (tcache && tc_idx < mp_.tcache_bins)
+	  if (tcache != NULL && tc_idx < mp_.tcache_bins)
 	    {
 	      mchunkptr tc_victim;
 
@@ -3982,7 +4039,7 @@  _int_malloc (mstate av, size_t bytes)
 #if USE_TCACHE
   INTERNAL_SIZE_T tcache_nb = 0;
   size_t tc_idx = csize2tidx (nb);
-  if (tcache && tc_idx < mp_.tcache_bins)
+  if (tcache != NULL && tc_idx < mp_.tcache_bins)
     tcache_nb = nb;
   int return_cached = 0;
 
@@ -4064,7 +4121,7 @@  _int_malloc (mstate av, size_t bytes)
 #if USE_TCACHE
 	      /* Fill cache first, return to user only if cache fills.
 		 We may return one of these chunks later.  */
-	      if (tcache_nb
+	      if (tcache_nb > 0
 		  && tcache->counts[tc_idx] < mp_.tcache_count)
 		{
 		  tcache_put (victim, tc_idx);
@@ -4937,6 +4994,43 @@  _int_realloc (mstate av, mchunkptr oldp, INTERNAL_SIZE_T oldsize,
    ------------------------------ memalign ------------------------------
  */
 
+/* Returns 0 if the chunk is not and does not contain the requested
+   aligned sub-chunk, else returns the amount of "waste" from
+   trimming.  BYTES is the *user* byte size, not the chunk byte
+   size.  */
+static int
+chunk_ok_for_memalign (mchunkptr p, size_t alignment, size_t bytes)
+{
+  void *m = chunk2mem (p);
+  INTERNAL_SIZE_T size = memsize (p);
+  void *aligned_m = m;
+
+  if (__glibc_unlikely (misaligned_chunk (p)))
+    malloc_printerr ("_int_memalign(): unaligned chunk detected");
+
+  aligned_m = PTR_ALIGN_UP (m, alignment);
+
+  INTERNAL_SIZE_T front_extra = (intptr_t) aligned_m - (intptr_t) m;
+
+  /* We can't trim off the front as it's too small.  */
+  if (front_extra > 0 && front_extra < MINSIZE)
+    return 0;
+
+  /* If it's a perfect fit, it's an exception to the return value rule
+     (we would return zero waste, which looks like "not usable"), so
+     handle it here by returning a small non-zero value instead.  */
+  if (size == bytes && front_extra == 0)
+    return 1;
+
+  /* If the block we need fits in the chunk, calculate total waste.  */
+  if (size > bytes + front_extra)
+    return size - bytes;
+
+  /* Can't use this chunk.  */ 
+  return 0;
+}
+
+/* BYTES is user requested bytes, not requested chunksize bytes.  */
 static void *
 _int_memalign (mstate av, size_t alignment, size_t bytes)
 {
@@ -4950,8 +5044,7 @@  _int_memalign (mstate av, size_t alignment, size_t bytes)
   mchunkptr remainder;            /* spare room at end to split off */
   unsigned long remainder_size;   /* its size */
   INTERNAL_SIZE_T size;
-
-
+  mchunkptr victim;
 
   nb = checked_request2size (bytes);
   if (nb == 0)
@@ -4960,29 +5053,142 @@  _int_memalign (mstate av, size_t alignment, size_t bytes)
       return NULL;
     }
 
-  /*
-     Strategy: find a spot within that chunk that meets the alignment
+  /* We can't check tcache here because we hold the arena lock, which
+     tcache doesn't expect.  We expect it has been checked
+     earlier.  */
+
+  /* Strategy: search the bins looking for an existing block that
+     meets our needs.  We scan a range of bins from "exact size" to
+     "just under 2x", spanning the small/large barrier if needed.  If
+     we don't find anything in those bins, the common malloc code will
+     scan starting at 2x.  */
+
+  /* This will be set if we found a candidate chunk.  */
+  victim = NULL;
+
+  /* Fast bins are singly-linked, hard to remove a chunk from the middle
+     and unlikely to meet our alignment requirements.  We have not done
+     any experimentation with searching for aligned fastbins.  */
+
+  int first_bin_index;
+  int first_largebin_index;
+  int last_bin_index;
+
+  if (in_smallbin_range (nb))
+    first_bin_index = smallbin_index (nb);
+  else
+    first_bin_index = largebin_index (nb);
+
+  if (in_smallbin_range (nb * 2))
+    last_bin_index = smallbin_index (nb * 2);
+  else
+    last_bin_index = largebin_index (nb * 2);
+
+  first_largebin_index = largebin_index (MIN_LARGE_SIZE);
+
+  int victim_index;                 /* its bin index */
+
+  for (victim_index = first_bin_index;
+       victim_index < last_bin_index;
+       victim_index ++)
+    {
+      victim = NULL;
+
+      if (victim_index < first_largebin_index)
+    {
+      /* Check small bins.  Small bin chunks are doubly-linked despite
+	 being the same size.  */
+
+      mchunkptr fwd;                    /* misc temp for linking */
+      mchunkptr bck;                    /* misc temp for linking */
+
+      bck = bin_at (av, victim_index);
+      fwd = bck->fd;
+      while (fwd != bck)
+	{
+	  if (chunk_ok_for_memalign (fwd, alignment, bytes) > 0)
+	    {
+	      victim = fwd;
+
+	      /* Unlink it */
+	      victim->fd->bk = victim->bk;
+	      victim->bk->fd = victim->fd;
+	      break;
+	    }
+
+	  fwd = fwd->fd;
+	}
+    }
+  else
+    {
+      /* Check large bins.  */
+      mchunkptr fwd;                    /* misc temp for linking */
+      mchunkptr bck;                    /* misc temp for linking */
+      mchunkptr best = NULL;
+      size_t best_size = 0;
+
+      bck = bin_at (av, victim_index);
+      fwd = bck->fd;
+
+      while (fwd != bck)
+	{
+	  int extra;
+
+	  if (chunksize (fwd) < nb)
+	      break;
+	  extra = chunk_ok_for_memalign (fwd, alignment, bytes);
+	  if (extra > 0
+	      && (extra <= best_size || best == NULL))
+	    {
+	      best = fwd;
+	      best_size = extra;
+	    }
+
+	  fwd = fwd->fd;
+	}
+      victim = best;
+
+      if (victim != NULL)
+	{
+	  unlink_chunk (av, victim);
+	  break;
+	}
+    }
+
+      if (victim != NULL)
+	break;
+    }
+
+  /* Strategy: find a spot within that chunk that meets the alignment
      request, and then possibly free the leading and trailing space.
-   */
+     This strategy is incredibly costly and can lead to external
+     fragmentation if header and footer chunks are unused.  */
 
-  /* Call malloc with worst case padding to hit alignment. */
+  if (victim != NULL)
+    {
+      p = victim;
+      m = chunk2mem (p);
+      set_inuse (p);
+    }
+  else
+    {
+      /* Call malloc with worst case padding to hit alignment. */
 
-  m = (char *) (_int_malloc (av, nb + alignment + MINSIZE));
+      m = (char *) (_int_malloc (av, nb + alignment + MINSIZE));
 
-  if (m == 0)
-    return 0;           /* propagate failure */
+      if (m == 0)
+	return 0;           /* propagate failure */
 
-  p = mem2chunk (m);
+      p = mem2chunk (m);
+    }
 
   if ((((unsigned long) (m)) % alignment) != 0)   /* misaligned */
-
-    { /*
-                Find an aligned spot inside chunk.  Since we need to give back
-                leading space in a chunk of at least MINSIZE, if the first
-                calculation places us at a spot with less than MINSIZE leader,
-                we can move to the next aligned spot -- we've allocated enough
-                total room so that this is always possible.
-                 */
+    {
+      /* Find an aligned spot inside chunk.  Since we need to give back
+         leading space in a chunk of at least MINSIZE, if the first
+         calculation places us at a spot with less than MINSIZE leader,
+         we can move to the next aligned spot -- we've allocated enough
+         total room so that this is always possible.  */
       brk = (char *) mem2chunk (((unsigned long) (m + alignment - 1)) &
                                 - ((signed long) alignment));
       if ((unsigned long) (brk - (char *) (p)) < MINSIZE)
diff --git a/malloc/tst-memalign-2.c b/malloc/tst-memalign-2.c
new file mode 100644
index 0000000000..432a37bd8b
--- /dev/null
+++ b/malloc/tst-memalign-2.c
@@ -0,0 +1,137 @@ 
+/* Test for memalign chunk reuse
+   Copyright (C) 2022 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <errno.h>
+#include <malloc.h>
+#include <stdio.h>
+#include <string.h>
+#include <unistd.h>
+#include <array_length.h>
+
+#include <support/check.h>
+
+typedef struct TestCase {
+  size_t size;
+  size_t alignment;
+  void *ptr1;
+  void *ptr2;
+} TestCase;
+
+static TestCase tcache_allocs[] = {
+  { 24, 8, NULL, NULL },
+  { 24, 16, NULL, NULL },
+  { 128, 32, NULL, NULL }
+};
+#define TN array_length (tcache_allocs)
+
+static TestCase large_allocs[] = {
+  { 23450, 64, NULL, NULL },
+  { 23450, 64, NULL, NULL },
+  { 23550, 64, NULL, NULL },
+  { 23550, 64, NULL, NULL },
+  { 23650, 64, NULL, NULL },
+  { 23650, 64, NULL, NULL },
+  { 33650, 64, NULL, NULL },
+  { 33650, 64, NULL, NULL }
+};
+#define LN array_length (large_allocs)
+
+void *p;
+
+static int
+do_test (void)
+{
+  int i, j;
+  int count;
+  void *ptr[10];
+  void *p;
+
+  /* TCache test.  */
+
+  for (i = 0; i < TN; ++ i)
+    {
+      tcache_allocs[i].ptr1 = memalign (tcache_allocs[i].alignment, tcache_allocs[i].size);
+      free (tcache_allocs[i].ptr1);
+      /* This should return the same chunk as was just free'd.  */
+      tcache_allocs[i].ptr2 = memalign (tcache_allocs[i].alignment, tcache_allocs[i].size);
+      free (tcache_allocs[i].ptr2);
+
+      TEST_VERIFY (tcache_allocs[i].ptr1 == tcache_allocs[i].ptr2);
+    }
+
+  /* Test for non-head tcache hits.  */
+  for (i = 0; i < 10; ++ i)
+    {
+      if (i == 4)
+	ptr[i] = memalign (64, 256);
+      else
+	ptr[i] = malloc (256);
+    }
+  for (i = 0; i < 10; ++ i)
+    free (ptr[i]);
+
+  p = memalign (64, 256);
+
+  count = 0;
+  for (i = 0; i < 10; ++ i)
+    if (ptr[i] == p)
+      ++ count;
+  free (p);
+  TEST_VERIFY (count > 0);
+
+  /* Large bins test.  */
+
+  for (i = 0; i < LN; ++ i)
+    {
+      large_allocs[i].ptr1 = memalign (large_allocs[i].alignment, large_allocs[i].size);
+      /* Keep chunks from combining by fragmenting the heap.  */
+      p = malloc (512);
+    }
+
+  for (i = 0; i < LN; ++ i)
+    free (large_allocs[i].ptr1);
+
+  /* Force the unsorted bins to be scanned and moved to small/large
+     bins.  */
+  p = malloc (60000);
+
+  for (i = 0; i < LN; ++ i)
+    large_allocs[i].ptr2 = memalign (large_allocs[i].alignment, large_allocs[i].size);
+
+  count = 0;
+  for (i = 0; i < LN; ++ i)
+    {
+      int ok = 0;
+      for (j = 0; j < LN; ++ j)
+	if (large_allocs[i].ptr1 == large_allocs[j].ptr2)
+	  ok = 1;
+      if (ok == 1)
+	count ++;
+    }
+
+  /* The allocation algorithm is complicated outside of the memalign
+     logic, so just make sure it's working for most of the
+     allocations.  This avoids possible boundary conditions with
+     empty/full heaps.  */
+  TEST_VERIFY (count > LN / 2);
+
+  return 0;
+}
+
+#define TEST_FUNCTION do_test ()
+#include "../test-skeleton.c"