Use a doubly-linked list for _IO_list_all (bug 27777)

This patch fixes BZ #27777 "fclose does a linear search, takes ages when
many FILE* are opened".  Simply put, the master list of opened (FILE*),
namely _IO_list_all, is a singly-linked list.  As a consequence, the
removal of a single element is in O(N), which cripples the performance
of fclose().  The patch switches to a doubly-linked list, yielding O(1)
removal.  The one padding field in struct _IO_FILE, __pad5, is renamed
to _prevchain for a doubly-linked list.  Since fields in struct _IO_FILE
after the _lock field are internal to glibc and opaque to applications.
We can change them as long as the size of struct _IO_FILE is unchanged,
which is checked as the part of glibc ABI with sizes of _IO_2_1_stdin_,
_IO_2_1_stdout_ and _IO_2_1_stderr_.

NB: When _IO_vtable_offset (fp) == 0, copy relocation will cover the
whole struct _IO_FILE.  Otherwise, only fields up to the _lock field
will be copied to applications at run-time.  It is used to check if
the _prevchain field can be safely accessed.

After opening 2 million (FILE*), the fclose() of 100 of them takes quite
a few seconds without the patch, and under 2 seconds with it on a loaded
machine.

No test is added since there are no functional changes.

Co-Authored-By: H.J. Lu <hjl.tools@gmail.com>
Signed-off-by: Alexandre Ferrieux <alexandre.ferrieux@orange.com>
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
This commit is contained in:
Alexandre Ferrieux 2024-05-16 05:54:30 -07:00 committed by H.J. Lu
parent a81cdde1cb
commit 2a99e2398d
3 changed files with 43 additions and 2 deletions

View File

@ -92,10 +92,10 @@ struct _IO_FILE_complete
struct _IO_wide_data *_wide_data;
struct _IO_FILE *_freeres_list;
void *_freeres_buf;
size_t __pad5;
struct _IO_FILE **_prevchain;
int _mode;
/* Make sure we don't get into trouble again. */
char _unused2[15 * sizeof (int) - 4 * sizeof (void *) - sizeof (size_t)];
char _unused2[15 * sizeof (int) - 5 * sizeof (void *)];
};
/* These macros are used by bits/stdio.h and internal headers. */

View File

@ -48,6 +48,19 @@ flush_cleanup (void *not_used)
}
#endif
/* Fields in struct _IO_FILE after the _lock field are internal to
glibc and opaque to applications. We can change them as long as
the size of struct _IO_FILE is unchanged, which is checked as the
part of glibc ABI with sizes of _IO_2_1_stdin_, _IO_2_1_stdout_
and _IO_2_1_stderr_.
NB: When _IO_vtable_offset (fp) == 0, copy relocation will cover the
whole struct _IO_FILE. Otherwise, only fields up to the _lock field
will be copied. */
_Static_assert (offsetof (struct _IO_FILE, _prevchain)
> offsetof (struct _IO_FILE, _lock),
"offset of _prevchain > offset of _lock");
void
_IO_un_link (struct _IO_FILE_plus *fp)
{
@ -62,6 +75,14 @@ _IO_un_link (struct _IO_FILE_plus *fp)
#endif
if (_IO_list_all == NULL)
;
else if (_IO_vtable_offset ((FILE *) fp) == 0)
{
FILE **pr = fp->file._prevchain;
FILE *nx = fp->file._chain;
*pr = nx;
if (nx != NULL)
nx->_prevchain = pr;
}
else if (fp == _IO_list_all)
_IO_list_all = (struct _IO_FILE_plus *) _IO_list_all->file._chain;
else
@ -95,6 +116,11 @@ _IO_link_in (struct _IO_FILE_plus *fp)
_IO_flockfile ((FILE *) fp);
#endif
fp->file._chain = (FILE *) _IO_list_all;
if (_IO_vtable_offset ((FILE *) fp) == 0)
{
fp->file._prevchain = (FILE **) &_IO_list_all;
_IO_list_all->file._prevchain = &fp->file._chain;
}
_IO_list_all = fp;
#ifdef _IO_MTSAFE_IO
_IO_funlockfile ((FILE *) fp);

View File

@ -54,4 +54,19 @@ DEF_STDFILE(_IO_2_1_stdout_, 1, &_IO_2_1_stdin_, _IO_NO_READS);
DEF_STDFILE(_IO_2_1_stderr_, 2, &_IO_2_1_stdout_, _IO_NO_READS+_IO_UNBUFFERED);
struct _IO_FILE_plus *_IO_list_all = &_IO_2_1_stderr_;
/* Finish double-linking stdin, stdout, and stderr in a constructor.
Static initialization cannot complete the _prevchain setup. */
__THROW __attribute__ ((constructor))
static void
_IO_stdfiles_init (void)
{
struct _IO_FILE **f;
for (f = (struct _IO_FILE **) &_IO_list_all;
*f != NULL;
f = &(*f)->_chain)
(*f)->_prevchain = f;
}
libc_hidden_data_def (_IO_list_all)