2017-04-15 14:20:00 +00:00
|
|
|
/* Copyright (C) 2017 the mpv developers
|
|
|
|
*
|
|
|
|
* Permission to use, copy, modify, and/or distribute this software for any
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
* purpose with or without fee is hereby granted, provided that the above
|
|
|
|
* copyright notice and this permission notice appear in all copies.
|
|
|
|
*
|
|
|
|
* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
|
|
|
|
* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
|
|
|
|
* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
|
|
|
|
* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
|
|
|
|
* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
|
|
|
|
* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
|
|
|
|
* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <stdlib.h>
|
|
|
|
#include <string.h>
|
|
|
|
#include <stdio.h>
|
|
|
|
#include <assert.h>
|
|
|
|
|
|
|
|
#define TA_NO_WRAPPERS
|
|
|
|
#include "ta.h"
|
|
|
|
|
|
|
|
// Note: the actual minimum alignment is dictated by malloc(). It doesn't
|
|
|
|
// make sense to set this value higher than malloc's alignment.
|
|
|
|
#define MIN_ALIGN 16
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
#define TA_MEMORY_DEBUGGING
|
|
|
|
#endif
|
|
|
|
|
|
|
|
struct ta_header {
|
|
|
|
size_t size; // size of the user allocation
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
// Invariant: parent!=NULL => prev==NULL
|
|
|
|
struct ta_header *prev; // siblings list (by destructor order)
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
struct ta_header *next;
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
// Invariant: parent==NULL || parent->child==this
|
|
|
|
struct ta_header *child; // points to first child
|
|
|
|
struct ta_header *parent; // set for _first_ child only, NULL otherwise
|
|
|
|
void (*destructor)(void *);
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
#ifdef TA_MEMORY_DEBUGGING
|
|
|
|
unsigned int canary;
|
|
|
|
struct ta_header *leak_next;
|
|
|
|
struct ta_header *leak_prev;
|
|
|
|
const char *name;
|
|
|
|
#endif
|
|
|
|
};
|
|
|
|
|
|
|
|
#define CANARY 0xD3ADB3EF
|
|
|
|
|
|
|
|
union aligned_header {
|
|
|
|
struct ta_header ta;
|
|
|
|
// Make sure to satisfy typical alignment requirements
|
|
|
|
void *align_ptr;
|
|
|
|
int align_int;
|
|
|
|
double align_d;
|
|
|
|
long long align_ll;
|
|
|
|
char align_min[(sizeof(struct ta_header) + MIN_ALIGN - 1) & ~(MIN_ALIGN - 1)];
|
|
|
|
};
|
|
|
|
|
|
|
|
#define PTR_TO_HEADER(ptr) (&((union aligned_header *)(ptr) - 1)->ta)
|
|
|
|
#define PTR_FROM_HEADER(h) ((void *)((union aligned_header *)(h) + 1))
|
|
|
|
|
|
|
|
#define MAX_ALLOC (((size_t)-1) - sizeof(union aligned_header))
|
|
|
|
|
|
|
|
static void ta_dbg_add(struct ta_header *h);
|
|
|
|
static void ta_dbg_check_header(struct ta_header *h);
|
|
|
|
static void ta_dbg_remove(struct ta_header *h);
|
|
|
|
|
|
|
|
static struct ta_header *get_header(void *ptr)
|
|
|
|
{
|
|
|
|
struct ta_header *h = ptr ? PTR_TO_HEADER(ptr) : NULL;
|
|
|
|
ta_dbg_check_header(h);
|
|
|
|
return h;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Set the parent allocation of ptr. If parent==NULL, remove the parent.
|
2020-02-23 16:45:05 +00:00
|
|
|
* Setting parent==NULL (with ptr!=NULL) unsets the parent of ptr.
|
|
|
|
* With ptr==NULL, the function does nothing.
|
2014-02-23 15:50:51 +00:00
|
|
|
*
|
|
|
|
* Warning: if ta_parent is a direct or indirect child of ptr, things will go
|
|
|
|
* wrong. The function will apparently succeed, but creates circular
|
|
|
|
* parent links, which are not allowed.
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
*/
|
2020-02-23 16:45:05 +00:00
|
|
|
void ta_set_parent(void *ptr, void *ta_parent)
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
{
|
|
|
|
struct ta_header *ch = get_header(ptr);
|
|
|
|
if (!ch)
|
2020-02-23 16:45:05 +00:00
|
|
|
return;
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
struct ta_header *new_parent = get_header(ta_parent);
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
// Unlink from previous parent
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
if (ch->prev)
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
ch->prev->next = ch->next;
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
if (ch->next)
|
|
|
|
ch->next->prev = ch->prev;
|
|
|
|
// If ch was the firs child, change child link of old parent
|
|
|
|
if (ch->parent) {
|
|
|
|
assert(ch->parent->child == ch);
|
|
|
|
ch->parent->child = ch->next;
|
|
|
|
if (ch->parent->child) {
|
|
|
|
assert(!ch->parent->child->parent);
|
|
|
|
ch->parent->child->parent = ch->parent;
|
|
|
|
}
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
}
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
ch->next = ch->prev = ch->parent = NULL;
|
|
|
|
// Link to new parent - insert at start of list (LIFO destructor order)
|
|
|
|
if (new_parent) {
|
|
|
|
ch->next = new_parent->child;
|
|
|
|
if (ch->next) {
|
|
|
|
ch->next->prev = ch;
|
|
|
|
ch->next->parent = NULL;
|
|
|
|
}
|
|
|
|
new_parent->child = ch;
|
|
|
|
ch->parent = new_parent;
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Allocate size bytes of memory. If ta_parent is not NULL, this is used as
|
|
|
|
* parent allocation (if ta_parent is freed, this allocation is automatically
|
|
|
|
* freed as well). size==0 allocates a block of size 0 (i.e. returns non-NULL).
|
|
|
|
* Returns NULL on OOM.
|
|
|
|
*/
|
|
|
|
void *ta_alloc_size(void *ta_parent, size_t size)
|
|
|
|
{
|
|
|
|
if (size >= MAX_ALLOC)
|
|
|
|
return NULL;
|
|
|
|
struct ta_header *h = malloc(sizeof(union aligned_header) + size);
|
|
|
|
if (!h)
|
|
|
|
return NULL;
|
|
|
|
*h = (struct ta_header) {.size = size};
|
|
|
|
ta_dbg_add(h);
|
|
|
|
void *ptr = PTR_FROM_HEADER(h);
|
2020-02-23 16:45:05 +00:00
|
|
|
ta_set_parent(ptr, ta_parent);
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Exactly the same as ta_alloc_size(), but the returned memory block is
|
|
|
|
* initialized to 0.
|
|
|
|
*/
|
|
|
|
void *ta_zalloc_size(void *ta_parent, size_t size)
|
|
|
|
{
|
|
|
|
if (size >= MAX_ALLOC)
|
|
|
|
return NULL;
|
|
|
|
struct ta_header *h = calloc(1, sizeof(union aligned_header) + size);
|
|
|
|
if (!h)
|
|
|
|
return NULL;
|
|
|
|
*h = (struct ta_header) {.size = size};
|
|
|
|
ta_dbg_add(h);
|
|
|
|
void *ptr = PTR_FROM_HEADER(h);
|
2020-02-23 16:45:05 +00:00
|
|
|
ta_set_parent(ptr, ta_parent);
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Reallocate the allocation given by ptr and return a new pointer. Much like
|
|
|
|
* realloc(), the returned pointer can be different, and on OOM, NULL is
|
|
|
|
* returned.
|
|
|
|
*
|
|
|
|
* size==0 is equivalent to ta_free(ptr).
|
|
|
|
* ptr==NULL is equivalent to ta_alloc_size(ta_parent, size).
|
|
|
|
*
|
|
|
|
* ta_parent is used only in the ptr==NULL case.
|
|
|
|
*
|
|
|
|
* Returns NULL if the operation failed.
|
|
|
|
* NULL is also returned if size==0.
|
|
|
|
*/
|
|
|
|
void *ta_realloc_size(void *ta_parent, void *ptr, size_t size)
|
|
|
|
{
|
|
|
|
if (size >= MAX_ALLOC)
|
|
|
|
return NULL;
|
|
|
|
if (!size) {
|
|
|
|
ta_free(ptr);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
if (!ptr)
|
|
|
|
return ta_alloc_size(ta_parent, size);
|
|
|
|
struct ta_header *h = get_header(ptr);
|
|
|
|
struct ta_header *old_h = h;
|
|
|
|
if (h->size == size)
|
|
|
|
return ptr;
|
|
|
|
ta_dbg_remove(h);
|
|
|
|
h = realloc(h, sizeof(union aligned_header) + size);
|
|
|
|
ta_dbg_add(h ? h : old_h);
|
|
|
|
if (!h)
|
|
|
|
return NULL;
|
|
|
|
h->size = size;
|
|
|
|
if (h != old_h) {
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
// Relink parent
|
2020-02-23 18:45:25 +00:00
|
|
|
if (h->parent)
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
h->parent->child = h;
|
|
|
|
// Relink siblings
|
|
|
|
if (h->next)
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
h->next->prev = h;
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
if (h->prev)
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
h->prev->next = h;
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
// Relink children
|
|
|
|
if (h->child)
|
|
|
|
h->child->parent = h;
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
}
|
|
|
|
return PTR_FROM_HEADER(h);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Return the allocated size of ptr. This returns the size parameter of the
|
|
|
|
* most recent ta_alloc.../ta_realloc... call.
|
|
|
|
* If ptr==NULL, return 0.
|
|
|
|
*/
|
|
|
|
size_t ta_get_size(void *ptr)
|
|
|
|
{
|
|
|
|
struct ta_header *h = get_header(ptr);
|
|
|
|
return h ? h->size : 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Free all allocations that (recursively) have ptr as parent allocation, but
|
|
|
|
* do not free ptr itself.
|
|
|
|
*/
|
|
|
|
void ta_free_children(void *ptr)
|
|
|
|
{
|
|
|
|
struct ta_header *h = get_header(ptr);
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
while (h && h->child)
|
|
|
|
ta_free(PTR_FROM_HEADER(h->child));
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Free the given allocation, and all of its direct and indirect children.
|
|
|
|
*/
|
|
|
|
void ta_free(void *ptr)
|
|
|
|
{
|
|
|
|
struct ta_header *h = get_header(ptr);
|
|
|
|
if (!h)
|
|
|
|
return;
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
if (h->destructor)
|
|
|
|
h->destructor(ptr);
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
ta_free_children(ptr);
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
ta_set_parent(ptr, NULL);
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
ta_dbg_remove(h);
|
|
|
|
free(h);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Set a destructor that is to be called when the given allocation is freed.
|
|
|
|
* (Whether the allocation is directly freed with ta_free() or indirectly by
|
|
|
|
* freeing its parent does not matter.) There is only one destructor. If an
|
|
|
|
* destructor was already set, it's overwritten.
|
|
|
|
*
|
|
|
|
* The destructor will be called with ptr as argument. The destructor can do
|
|
|
|
* almost anything, but it must not attempt to free or realloc ptr. The
|
|
|
|
* destructor is run before the allocation's children are freed (also, before
|
|
|
|
* their destructors are run).
|
|
|
|
*/
|
2020-02-23 16:45:05 +00:00
|
|
|
void ta_set_destructor(void *ptr, void (*destructor)(void *))
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
{
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
struct ta_header *h = get_header(ptr);
|
2020-02-23 16:45:05 +00:00
|
|
|
if (h)
|
|
|
|
h->destructor = destructor;
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef TA_MEMORY_DEBUGGING
|
|
|
|
|
|
|
|
#include <pthread.h>
|
|
|
|
|
|
|
|
static pthread_mutex_t ta_dbg_mutex = PTHREAD_MUTEX_INITIALIZER;
|
|
|
|
static bool enable_leak_check; // pretty much constant
|
|
|
|
static struct ta_header leak_node;
|
|
|
|
static char allocation_is_string;
|
|
|
|
|
|
|
|
static void ta_dbg_add(struct ta_header *h)
|
|
|
|
{
|
|
|
|
h->canary = CANARY;
|
|
|
|
if (enable_leak_check) {
|
|
|
|
pthread_mutex_lock(&ta_dbg_mutex);
|
|
|
|
h->leak_next = &leak_node;
|
|
|
|
h->leak_prev = leak_node.leak_prev;
|
|
|
|
leak_node.leak_prev->leak_next = h;
|
|
|
|
leak_node.leak_prev = h;
|
|
|
|
pthread_mutex_unlock(&ta_dbg_mutex);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ta_dbg_check_header(struct ta_header *h)
|
|
|
|
{
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
if (h) {
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
assert(h->canary == CANARY);
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
if (h->parent) {
|
|
|
|
assert(!h->prev);
|
|
|
|
assert(h->parent->child == h);
|
|
|
|
}
|
|
|
|
}
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ta_dbg_remove(struct ta_header *h)
|
|
|
|
{
|
|
|
|
ta_dbg_check_header(h);
|
|
|
|
if (h->leak_next) { // assume checking for !=NULL invariant ok without lock
|
|
|
|
pthread_mutex_lock(&ta_dbg_mutex);
|
|
|
|
h->leak_next->leak_prev = h->leak_prev;
|
|
|
|
h->leak_prev->leak_next = h->leak_next;
|
|
|
|
pthread_mutex_unlock(&ta_dbg_mutex);
|
|
|
|
h->leak_next = h->leak_prev = NULL;
|
|
|
|
}
|
|
|
|
h->canary = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static size_t get_children_size(struct ta_header *h)
|
|
|
|
{
|
|
|
|
size_t size = 0;
|
2020-02-29 19:04:29 +00:00
|
|
|
for (struct ta_header *s = h->child; s; s = s->next)
|
ta: remove seperate internal "ext" header
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
2020-02-23 16:36:19 +00:00
|
|
|
size += s->size + get_children_size(s);
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
return size;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void print_leak_report(void)
|
|
|
|
{
|
|
|
|
pthread_mutex_lock(&ta_dbg_mutex);
|
|
|
|
if (leak_node.leak_next && leak_node.leak_next != &leak_node) {
|
|
|
|
size_t size = 0;
|
|
|
|
size_t num_blocks = 0;
|
|
|
|
fprintf(stderr, "Blocks not freed:\n");
|
|
|
|
fprintf(stderr, " %-20s %10s %10s %s\n",
|
|
|
|
"Ptr", "Bytes", "C. Bytes", "Name");
|
|
|
|
while (leak_node.leak_next != &leak_node) {
|
|
|
|
struct ta_header *cur = leak_node.leak_next;
|
|
|
|
// Don't list those with parent; logically, only parents are listed
|
|
|
|
if (!cur->next) {
|
|
|
|
size_t c_size = get_children_size(cur);
|
|
|
|
char name[30] = {0};
|
|
|
|
if (cur->name)
|
|
|
|
snprintf(name, sizeof(name), "%s", cur->name);
|
|
|
|
if (cur->name == &allocation_is_string) {
|
|
|
|
snprintf(name, sizeof(name), "'%.*s'",
|
|
|
|
(int)cur->size, (char *)PTR_FROM_HEADER(cur));
|
|
|
|
}
|
|
|
|
for (int n = 0; n < sizeof(name); n++) {
|
|
|
|
if (name[n] && name[n] < 0x20)
|
|
|
|
name[n] = '.';
|
|
|
|
}
|
|
|
|
fprintf(stderr, " %-20p %10zu %10zu %s\n",
|
|
|
|
cur, cur->size, c_size, name);
|
|
|
|
}
|
|
|
|
size += cur->size;
|
|
|
|
num_blocks += 1;
|
|
|
|
// Unlink, and don't confuse valgrind by leaving live pointers.
|
|
|
|
cur->leak_next->leak_prev = cur->leak_prev;
|
|
|
|
cur->leak_prev->leak_next = cur->leak_next;
|
|
|
|
cur->leak_next = cur->leak_prev = NULL;
|
|
|
|
}
|
|
|
|
fprintf(stderr, "%zu bytes in %zu blocks.\n", size, num_blocks);
|
|
|
|
}
|
|
|
|
pthread_mutex_unlock(&ta_dbg_mutex);
|
|
|
|
}
|
|
|
|
|
|
|
|
void ta_enable_leak_report(void)
|
|
|
|
{
|
|
|
|
pthread_mutex_lock(&ta_dbg_mutex);
|
|
|
|
enable_leak_check = true;
|
|
|
|
if (!leak_node.leak_prev && !leak_node.leak_next) {
|
|
|
|
leak_node.leak_prev = &leak_node;
|
|
|
|
leak_node.leak_next = &leak_node;
|
|
|
|
atexit(print_leak_report);
|
|
|
|
}
|
|
|
|
pthread_mutex_unlock(&ta_dbg_mutex);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Set a (static) string that will be printed if the memory allocation in ptr
|
|
|
|
* shows up on the leak report. The string must stay valid until ptr is freed.
|
|
|
|
* Calling it on ptr==NULL does nothing.
|
|
|
|
* Typically used to set location info.
|
|
|
|
* Always returns ptr (useful for chaining function calls).
|
|
|
|
*/
|
|
|
|
void *ta_dbg_set_loc(void *ptr, const char *loc)
|
|
|
|
{
|
|
|
|
struct ta_header *h = get_header(ptr);
|
|
|
|
if (h)
|
|
|
|
h->name = loc;
|
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Mark the allocation as string. The leak report will print it literally.
|
|
|
|
*/
|
|
|
|
void *ta_dbg_mark_as_string(void *ptr)
|
|
|
|
{
|
|
|
|
// Specially handled by leak report code.
|
|
|
|
return ta_dbg_set_loc(ptr, &allocation_is_string);
|
|
|
|
}
|
|
|
|
|
|
|
|
#else
|
|
|
|
|
|
|
|
static void ta_dbg_add(struct ta_header *h){}
|
|
|
|
static void ta_dbg_check_header(struct ta_header *h){}
|
|
|
|
static void ta_dbg_remove(struct ta_header *h){}
|
|
|
|
|
|
|
|
void ta_enable_leak_report(void){}
|
2013-12-19 20:27:02 +00:00
|
|
|
void *ta_dbg_set_loc(void *ptr, const char *loc){return ptr;}
|
|
|
|
void *ta_dbg_mark_as_string(void *ptr){return ptr;}
|
Replace talloc
There are multiple reasons to do this. One big reason is the license:
talloc is LGPLv3+, which forces mpv to be licensed as GPLv3+.
Another one is that our talloc copy contains modifications, which makes
it essentially incompatible with upstream talloc (in particular, our
version aborts on out of memory conditions - well, it wasn't my idea).
Updating from upstream is also a bit involved - the talloc source is
not really organized in a way to allow copying it into projects (and
this isn't an intended use-case).
Finally, talloc is kind of big and bloated. The replacement halves the
amount of code - mainly because we didn't use all talloc features. It's
even more extreme if you compare upstream talloc (~4700 lines) and the
new allocator without talloc compat (~900 lines).
The replacement provides all features we need. It also doesn't clash
with talloc. (The talloc compatibility wrapper uses macros to avoid
introducing linker-level symbols which could clash with libtalloc.)
It also tries to lower the overhead (only 4 words opposed to 10 words
in talloc for leaf nodes in release mode). Debugging features like leak
reporting can be enabled at compile time and add somewhat more overhead.
Though I'm not sure whether the overhead reduction was actually
successful: allocations with children need an "extra" header, which adds
plenty of overhead, and it turns out that almost half of all allocations
have children. Maybe the implementation could be simplified and the
extra header removed - even then, overhead would be lower than talloc's.
Currently, debugging features can be entirely deactivated by defining
NDEBUG - I'm not sure if anything defines this directly yet, though.
Unlike in talloc, the leak reporting stuff is thread-safe. (That's also
why it's far less elegant, and requires extra list pointers.)
Comes with a compatibility layer, so no changes to mpv source code
are needed. The idea is that we will pretend to be using talloc for
a while, so that we can revert to our old talloc implementation at
any time for debugging purposes.
Some inspiration was taken from Mesa's ralloc:
http://cgit.freedesktop.org/mesa/mesa/tree/src/glsl/ralloc.h
This is another talloc replacement, but lacks some features we need
(getting size of an allocation, debugging features, being able to
access children in the dtor).
There's some information in ta/README what will happen next and how the
transition is expected to progress.
2013-10-12 23:17:45 +00:00
|
|
|
|
|
|
|
#endif
|