[v5,1/1] memalign: Support scanning for aligned chunks.

Adhemerval Zanella Netto <adhemerval.zanella@linaro.org> writes:
> Hi DJ (I think I got it right now),

Yup!

> patch looks good, some comments below.

v5 attached with changes as noted.

>> +/* Iterates through the tcache linked list.  */
>> +static __always_inline void *
>
> Why not use 'tcache_next *' as return type here?
>
>> +tcache_next (tcache_entry *e)

IIRC I copied tcache_get(), which returns that.
Fixed.

>> +	while (te != NULL && ((intptr_t)te & (alignment - 1)) != 0)
>
> Maybe use '!PTR_IS_ALIGNED (te, alignment)' here?

Yup.

>> +	  {
>> +	    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);
>> @@ -3857,7 +3914,7 @@ _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;
>>  
>
> I think the style chance should be on a different patch.

Perhaps but IIRC I needed those to get the warnings down to zero so I'd
prefer to leave them in.  Too much effort to split them out.

>> +/* 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
>
> Shouldn't it return a size_t here?

Fixed.

>> +chunk_ok_for_memalign (mchunkptr p, size_t alignment, size_t bytes)
>>  _int_memalign (mstate av, size_t alignment, size_t bytes)
>>  {
>> @@ -4945,8 +5039,7 @@ _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;
>> -
>> -
>
> Spurious extra new lines?

The original had three blank lines there for some reason.  I wouldn't
have bothered if I didn't have to add a new decl there anyway.

>> diff --git a/malloc/tst-memalign-2.c b/malloc/tst-memalign-2.c
>> new file mode 100644
>> index 0000000000..ed3660959a
>> --- /dev/null
>> +++ b/malloc/tst-memalign-2.c
>> @@ -0,0 +1,136 @@
>> +/* Test for memalign chunk reuse
>
> Missing period.

Fixed.

>> +  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);
>
> Should we also check for non NULL and return alignment as sanity checks here?

Done.

>> +
>> +      TEST_VERIFY (tcache_allocs[i].ptr1 == tcache_allocs[i].ptr2);
>> +    }
>> +
>> +  /* Test for non-head tcache hits.  */
>> +  for (i = 0; i < 10; ++ i)
>
> Maybe use array_length (ptr) here.

Done.


From e32abda27e5c0aa82f4b736fdca35d56bf665cce Mon Sep 17 00:00:00 2001
From: DJ Delorie via Libc-alpha <libc-alpha@sourceware.org>
Date: Wed, 29 Mar 2023 00:18:40 -0400
Subject: memalign: Support scanning for aligned 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.

On 29/03/23 01:20, DJ Delorie wrote:
> From e32abda27e5c0aa82f4b736fdca35d56bf665cce Mon Sep 17 00:00:00 2001
> From: DJ Delorie via Libc-alpha <libc-alpha@sourceware.org>
> Date: Wed, 29 Mar 2023 00:18:40 -0400
> Subject: memalign: Support scanning for aligned 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.

LGTM, thanks.

Reviewed-by: Adhemerval Zanella  <adhemerval.zanella@linaro.org>

> 
> diff --git a/malloc/Makefile b/malloc/Makefile
> index dfb51d344c..79178c4905 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 76c50e3f58..8ebc4372bc 100644
> --- a/malloc/malloc.c
> +++ b/malloc/malloc.c
> @@ -3162,19 +3162,44 @@ 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 tcache_entry *
> +tcache_next (tcache_entry *e)
> +{
> +  return (tcache_entry *) REVEAL_PTR (e->next);
> +}
> +
>  static void
>  tcache_thread_shutdown (void)
>  {
> @@ -3283,7 +3308,7 @@ __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);
> @@ -3542,6 +3567,38 @@ _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 && !PTR_IS_ALIGNED (te, alignment))
> +	  {
> +	    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);
> @@ -3847,7 +3904,7 @@ _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;
>  
> @@ -3905,7 +3962,7 @@ _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;
>  
> @@ -3967,7 +4024,7 @@ _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;
>  
> @@ -4047,7 +4104,7 @@ _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);
> @@ -4921,6 +4978,43 @@ _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 size_t
> +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)
>  {
> @@ -4934,8 +5028,7 @@ _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)
> @@ -4944,29 +5037,142 @@ _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..4996578e9f
> --- /dev/null
> +++ b/malloc/tst-memalign-2.c
> @@ -0,0 +1,155 @@
> +/* 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 <libc-pointer-arith.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;
> +
> +/* Sanity checks, ancillary to the actual test.  */
> +#define CHECK(p,a) \
> +  if (p == NULL || !PTR_IS_ALIGNED (p, a)) \
> +    FAIL_EXIT1 ("NULL or misaligned memory detected.\n");
> +
> +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);
> +      CHECK (tcache_allocs[i].ptr1, tcache_allocs[i].alignment);
> +      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);
> +      CHECK (tcache_allocs[i].ptr2, tcache_allocs[i].alignment);
> +      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 < array_length (ptr); ++ i)
> +    {
> +      if (i == 4)
> +	{
> +	  ptr[i] = memalign (64, 256);
> +	  CHECK (ptr[i], 64);
> +	}
> +      else
> +	{
> +	  ptr[i] = malloc (256);
> +	  CHECK (ptr[i], 4);
> +	}
> +    }
> +  for (i = 0; i < array_length (ptr); ++ i)
> +    free (ptr[i]);
> +
> +  p = memalign (64, 256);
> +  CHECK (p, 64);
> +
> +  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);
> +      CHECK (large_allocs[i].ptr1, large_allocs[i].alignment);
> +      /* Keep chunks from combining by fragmenting the heap.  */
> +      p = malloc (512);
> +      CHECK (p, 4);
> +    }
> +
> +  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);
> +      CHECK (large_allocs[i].ptr2, large_allocs[i].alignment);
> +    }
> +
> +  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;
> +}
> +
> +#include <support/test-driver.c>
>

Thanks!  Pushed.

On Wed, Mar 29, 2023 at 5:36 PM DJ Delorie via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
>
> Thanks!  Pushed.

Crashes previously working rust code..

for example "ripgrep" (command rg)

#rg FIND_PA
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p
|| chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk
(mem2chunk (p))

Will see if anything else died ;)

[1]    2123 IOT instruction (core dumped)  rg FIND_PA

On 30/03/23 07:04, Cristian Rodríguez wrote:
> On Wed, Mar 29, 2023 at 5:36 PM DJ Delorie via Libc-alpha
> <libc-alpha@sourceware.org> wrote:
>>
>>
>> Thanks!  Pushed.
> 
> Crashes previously working rust code..
> 
> for example "ripgrep" (command rg)
> 
> #rg FIND_PA
> Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p
> || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk
> (mem2chunk (p))
> 
> Will see if anything else died ;)
> 
> [1]    2123 IOT instruction (core dumped)  rg FIND_PA

Do you have any testcase that triggers it?

On Thu, Mar 30, 2023 at 7:50 AM Adhemerval Zanella Netto <
adhemerval.zanella@linaro.org> wrote:

> Do you have any testcase that triggers it?

I'll try.. cannot compile ripgrep with debug info either ..because this
commit also makes rustc crash with the same message.😕

On 29/03/23 01:20, DJ Delorie wrote:
> From e32abda27e5c0aa82f4b736fdca35d56bf665cce Mon Sep 17 00:00:00 2001
> From: DJ Delorie via Libc-alpha <libc-alpha@sourceware.org>
> Date: Wed, 29 Mar 2023 00:18:40 -0400
> Subject: memalign: Support scanning for aligned 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.

So it seems this patch does trigger a regression.  I am seeing on a speccpu2017
benchmark (cam4_s) failure:

****************************************
Contents of cam4_s_base.gcc-64.err
****************************************
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Program received signal SIGABRT: Process abort signal.
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))
Fatal glibc error: malloc.c:3617 (_mid_memalign): assertion failed: !p || chunk_is_mmapped (mem2chunk (p)) || ar_ptr == arena_for_chunk (mem2chunk (p))

I have not yet isolated the malloc calls patterns, but I would like to give you
a heads up that this does seems to be an issue with at least on reproduce.

On 29.03.23 22:36, DJ Delorie via Libc-alpha wrote:
> 
> Thanks!  Pushed.
> 
On s390 (31bit), I see the test fail:
FAIL: malloc/tst-memalign-2-mcheck

After adding those printfs ...:
diff --git a/malloc/tst-memalign-2.c b/malloc/tst-memalign-2.c
index 4996578e9f..adfebf8384 100644
--- a/malloc/tst-memalign-2.c
+++ b/malloc/tst-memalign-2.c
@@ -72,10 +72,12 @@ do_test (void)
     {
       tcache_allocs[i].ptr1 = memalign (tcache_allocs[i].alignment,
tcache_allocs[i].size);
       CHECK (tcache_allocs[i].ptr1, tcache_allocs[i].alignment);
+      printf ("%d# ptr1=%p\n", i, tcache_allocs[i].ptr1);
       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);
       CHECK (tcache_allocs[i].ptr2, tcache_allocs[i].alignment);
+      printf ("%d# ptr2=%p\n", i, tcache_allocs[i].ptr2);
       free (tcache_allocs[i].ptr2);

       TEST_VERIFY (tcache_allocs[i].ptr1 == tcache_allocs[i].ptr2);


... I've got this output:
0# ptr1=0x55bc71b8
0# ptr2=0x55bc71b8
1# ptr1=0x55bc7210
1# ptr2=0x55bc7260
error: tst-memalign-2.c:83: not true: tcache_allocs[i].ptr1 ==
tcache_allocs[i].ptr2
2# ptr1=0x55bc72e0
2# ptr2=0x55bc72e0
error: 1 test failures


malloc/tst-memalign-2 (without mcheck) is passing.
PASS: malloc/tst-memalign-2
original exit status 0
0# ptr1=0x55c0e190
0# ptr2=0x55c0e190
1# ptr1=0x55c0e190
1# ptr2=0x55c0e190
2# ptr1=0x55c0e1c0
2# ptr2=0x55c0e1c0

Can you please help.

Thanks,
Stefan

Stefan Liebler via Libc-alpha <libc-alpha@sourceware.org> writes:
> On s390 (31bit), I see the test fail:
> FAIL: malloc/tst-memalign-2-mcheck

Please try
https://sourceware.org/pipermail/libc-alpha/2023-April/146959.html

I fixed that test there, hopefully you're seeing the same thing I saw ;-)

On 05.04.23 19:58, DJ Delorie wrote:
> Stefan Liebler via Libc-alpha <libc-alpha@sourceware.org> writes:
>> On s390 (31bit), I see the test fail:
>> FAIL: malloc/tst-memalign-2-mcheck
> 
> Please try
> https://sourceware.org/pipermail/libc-alpha/2023-April/146959.html
> 
> I fixed that test there, hopefully you're seeing the same thing I saw ;-)
> 
Hi DJ,

I've applied your patch
"[patch v2] malloc: set NON_MAIN_ARENA flag for reclaimed memalign chunk
(BZ #30101)"
on top of
"hurd: Don't leak __hurd_reply_port0"
commit cd019ddd892e182277fadd6aedccc57fa3923c8d

Now I get those fails:
on s390x (64bit) and x86_64:
- malloc/tst-memalign-2-mcheck.out
error: tst-memalign-2.c:114: not true: count > 0
error: 1 test failures

- malloc/tst-memalign-3-mcheck.out
error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
tcache_allocs[i].ptr2
error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
tcache_allocs[i].ptr2
error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
tcache_allocs[i].ptr2
error: 3 test failures


on s390 (31bit):
- malloc/tst-memalign-2-mcheck.out
error: tst-memalign-2.c:86: not true: tcache_allocs[i].ptr1 ==
tcache_allocs[i].ptr2
error: 1 test failures

- malloc/tst-memalign-3-mcheck.out
error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
tcache_allocs[i].ptr2
error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
tcache_allocs[i].ptr2
error: tst-memalign-3.c:117: not true: count > 0
error: 3 test failures


Do you also see the fails on x86_64?

On 11.04.23 13:40, Stefan Liebler via Libc-alpha wrote:
> On 05.04.23 19:58, DJ Delorie wrote:
>> Stefan Liebler via Libc-alpha <libc-alpha@sourceware.org> writes:
>>> On s390 (31bit), I see the test fail:
>>> FAIL: malloc/tst-memalign-2-mcheck
>>
>> Please try
>> https://sourceware.org/pipermail/libc-alpha/2023-April/146959.html
>>
>> I fixed that test there, hopefully you're seeing the same thing I saw ;-)
>>
> Hi DJ,
> 
> I've applied your patch
> "[patch v2] malloc: set NON_MAIN_ARENA flag for reclaimed memalign chunk
> (BZ #30101)"
> on top of
> "hurd: Don't leak __hurd_reply_port0"
> commit cd019ddd892e182277fadd6aedccc57fa3923c8d
> 
> Now I get those fails:
> on s390x (64bit) and x86_64:
> - malloc/tst-memalign-2-mcheck.out
> error: tst-memalign-2.c:114: not true: count > 0
> error: 1 test failures
> 
> - malloc/tst-memalign-3-mcheck.out
> error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
> tcache_allocs[i].ptr2
> error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
> tcache_allocs[i].ptr2
> error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
> tcache_allocs[i].ptr2
> error: 3 test failures
> 
> 
> on s390 (31bit):
> - malloc/tst-memalign-2-mcheck.out
> error: tst-memalign-2.c:86: not true: tcache_allocs[i].ptr1 ==
> tcache_allocs[i].ptr2
> error: 1 test failures
> 
> - malloc/tst-memalign-3-mcheck.out
> error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
> tcache_allocs[i].ptr2
> error: tst-memalign-3.c:89: not true: tcache_allocs[i].ptr1 ==
> tcache_allocs[i].ptr2
> error: tst-memalign-3.c:117: not true: count > 0
> error: 3 test failures
> 
> 
> Do you also see the fails on x86_64?

Just as information. I also see the same fails as described above with
"[patch v3] malloc: set NON_MAIN_ARENA flag for reclaimed memalign chunk
(BZ #30101)"
https://sourceware.org/pipermail/libc-alpha/2023-April/147181.html

On Thu, 2023-03-30 at 18:43 -0300, Cristian Rodríguez via Libc-alpha
wrote:
> On Thu, Mar 30, 2023 at 7:50 AM Adhemerval Zanella Netto <
> adhemerval.zanella@linaro.org> wrote:
> 
> > Do you have any testcase that triggers it?
> 
> I'll try.. cannot compile ripgrep with debug info either ..because this
> commit also makes rustc crash with the same message.😕

On LoongArch (with a GCC 12.2 but LoongArch backend patched to match GCC
trunk), this causes an almost deterministic failure of malloc/tst-
malloc-thread-fail-malloc-hugetlb2:

FAIL: malloc/tst-malloc-thread-fail-malloc-hugetlb2
original exit status 1
error: exit status 139 from child process

Xi Ruoyao <xry111@xry111.site> writes:
> On LoongArch (with a GCC 12.2 but LoongArch backend patched to match GCC
> trunk), this causes an almost deterministic failure of malloc/tst-
> malloc-thread-fail-malloc-hugetlb2:

(1) What's in malloc/tst-malloc-thread-fail-malloc-hugetlb2.out ?

(2) Have you tried with "export TIMEOUTFACTOR=20" ? some tests fail
    reliably on slower systems due to timeouts.

On Wed, 2023-04-12 at 13:16 -0400, DJ Delorie wrote:
> Xi Ruoyao <xry111@xry111.site> writes:
> > On LoongArch (with a GCC 12.2 but LoongArch backend patched to match
> > GCC
> > trunk), this causes an almost deterministic failure of malloc/tst-
> > malloc-thread-fail-malloc-hugetlb2:
> 
> (1) What's in malloc/tst-malloc-thread-fail-malloc-hugetlb2.out ?
> 
> (2) Have you tried with "export TIMEOUTFACTOR=20" ? some tests fail
>     reliably on slower systems due to timeouts.

No, it's not a timeout, but a segment fault.  The output is just "error:
exit status 139 from child process".

I'm not an expert with ptmalloc, but there is some code like:

  arena_get (ar_ptr, bytes + alignment + MINSIZE);

  p = _int_memalign (ar_ptr, alignment, bytes);
  if (!p && ar_ptr != NULL)
    {
      LIBC_PROBE (memory_memalign_retry, 2, bytes, alignment);
      ar_ptr = arena_get_retry (ar_ptr, bytes);
      p = _int_memalign (ar_ptr, alignment, bytes);
    }

arena_get can set ar_ptr to NULL (at least when the system memory is not
enough).

Then _int_memalign is dereferencing ar_ptr w/o any nullity check (bin_at
seems a "fancy" dereference operation to me).

Then we test ar_ptr != NULL in the if statement.

Now it looks like a notorious "NULL check after dereferencing" pattern.
So I added a nullity check:

diff --git a/malloc/malloc.c b/malloc/malloc.c
index 0315ac5d16..ed10b6b0e3 100644
--- a/malloc/malloc.c
+++ b/malloc/malloc.c
@@ -5025,7 +5025,7 @@ _int_memalign (mstate av, size_t alignment, size_t
bytes)
   mchunkptr victim;
 
   nb = checked_request2size (bytes);
-  if (nb == 0)
+  if (nb == 0 || !av)
     {
       __set_errno (ENOMEM);
       return NULL;

And it indeed fixed the test for me.  But I'm not sure if it's the
correct solution and I've not ran the complete test suite with the
change yet.

On 12/04/23 14:04, Xi Ruoyao wrote:
> On Thu, 2023-03-30 at 18:43 -0300, Cristian Rodríguez via Libc-alpha
> wrote:
>> On Thu, Mar 30, 2023 at 7:50 AM Adhemerval Zanella Netto <
>> adhemerval.zanella@linaro.org> wrote:
>>
>>> Do you have any testcase that triggers it?
>>
>> I'll try.. cannot compile ripgrep with debug info either ..because this
>> commit also makes rustc crash with the same message.😕
> 
> On LoongArch (with a GCC 12.2 but LoongArch backend patched to match GCC
> trunk), this causes an almost deterministic failure of malloc/tst-
> malloc-thread-fail-malloc-hugetlb2:
> 
> FAIL: malloc/tst-malloc-thread-fail-malloc-hugetlb2
> original exit status 1
> error: exit status 139 from child process
> 

I can reproduce on x86_64 as well in a non-deterministic manner (with multiple
'rm malloc/*.out && make malloc/tests -j24'.  The core file shows internal
metadata corruption:

$ gdb elf/ld.so core
[...]
#0  __pthread_kill_implementation (threadid=<optimized out>, signo=signo@entry=6, no_tid=no_tid@entry=0)
    at pthread_kill.c:44
44            return INTERNAL_SYSCALL_ERROR_P (ret) ? INTERNAL_SYSCALL_ERRNO (ret) : 0;
(gdb) bt
#0  __pthread_kill_implementation (threadid=<optimized out>, signo=signo@entry=6, no_tid=no_tid@entry=0)
    at pthread_kill.c:44
#1  0x00007fae15090017 in __pthread_kill_internal (signo=6, threadid=<optimized out>) at pthread_kill.c:78
#2  0x00007fae1503e326 in __GI_raise (sig=sig@entry=6) at ../sysdeps/posix/raise.c:26
#3  0x00007fae150264b1 in __GI_abort () at abort.c:79
#4  0x00007fae150272cc in __libc_message (fmt=fmt@entry=0x7fae151b0632 "%s\n") at ../sysdeps/posix/libc_fatal.c:150
#5  0x00007fae1509b873 in malloc_printerr (str=str@entry=0x7fae151b3458 "free(): double free detected in tcache 2")
    at malloc.c:5847
#6  0x00007fae1509dd5b in _int_free (av=0x7fae1524cac0 <main_arena>, p=0x7fae14e00290, have_lock=0) at malloc.c:4508
#7  0x00007fae150a07bb in __GI___libc_free (mem=<optimized out>) at malloc.c:3386
#8  0x00007fae15445594 in ?? ()
#9  0x00007fae14e002a0 in ?? ()
#10 0x73afe8780d804600 in ?? ()
#11 0x00007ffec090cc68 in ?? ()
#12 0x00007fae1544622b in ?? ()
#13 0x0000000000000000 in ?? ()

Adhemerval Zanella Netto <adhemerval.zanella@linaro.org> writes:
> I can reproduce on x86_64 as well in a non-deterministic manner (with multiple

Also malloc/tst-malloc-thread-fail-malloc-hugetlb2 ?

On 12/04/23 14:40, DJ Delorie wrote:
> Adhemerval Zanella Netto <adhemerval.zanella@linaro.org> writes:
>> I can reproduce on x86_64 as well in a non-deterministic manner (with multiple
> 
> Also malloc/tst-malloc-thread-fail-malloc-hugetlb2 ?
> 

Yes and with  malloc/tst-malloc-thread-fail-malloc-hugetlb1 as well. But 
the hugetlb1 and hugetlb2 just essentially enable hugepages with extra 
mprotect/madvise calls.  I don't think they play a role in this issue
(they won't affect any metadata setting, just the timings on the tests).

Adhemerval Zanella Netto <adhemerval.zanella@linaro.org> writes:
> Yes and with  malloc/tst-malloc-thread-fail-malloc-hugetlb1 as well. But 
> the hugetlb1 and hugetlb2 just essentially enable hugepages with extra 
> mprotect/madvise calls.  I don't think they play a role in this issue
> (they won't affect any metadata setting, just the timings on the tests).

I still can't reproduce this; how much free memory is on that machine?
Have you tested this particular setup without my patch?  I inspected all
the tcache code changes and didn't find anything...

On 12/04/23 22:57, DJ Delorie wrote:
> Adhemerval Zanella Netto <adhemerval.zanella@linaro.org> writes:
>> Yes and with  malloc/tst-malloc-thread-fail-malloc-hugetlb1 as well. But 
>> the hugetlb1 and hugetlb2 just essentially enable hugepages with extra 
>> mprotect/madvise calls.  I don't think they play a role in this issue
>> (they won't affect any metadata setting, just the timings on the tests).
> 
> I still can't reproduce this; how much free memory is on that machine?

About 64GB:

$ free
               total        used        free      shared  buff/cache   available
Mem:        65760296     6847324    41000244       17540    17912728    58182972
Swap:        8388604           0     8388604

> Have you tested this particular setup without my patch?  I inspected all
> the tcache code changes and didn't find anything...
> 

I will check this patch, thanks.