1
0
mirror of https://github.com/mpv-player/mpv synced 2024-12-11 17:37:23 +00:00
mpv/ta/ta.c

445 lines
13 KiB
C
Raw Normal View History

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
/* Permission to use, copy, modify, and/or distribute this software for any
* 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
struct ta_header *prev; // ring list containing siblings
struct ta_header *next;
struct ta_ext_header *ext;
#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))
// Needed for non-leaf allocations, or extended features such as destructors.
struct ta_ext_header {
struct ta_header *header; // points back to normal header
struct ta_header children; // list of children, with this as sentinel
void (*destructor)(void *);
};
// ta_ext_header.children.size is set to this
#define CHILDREN_SENTINEL ((size_t)-1)
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;
}
static struct ta_ext_header *get_or_alloc_ext_header(void *ptr)
{
struct ta_header *h = get_header(ptr);
if (!h)
return NULL;
if (!h->ext) {
h->ext = malloc(sizeof(struct ta_ext_header));
if (!h->ext)
return NULL;
*h->ext = (struct ta_ext_header) {
.header = h,
.children = {
.next = &h->ext->children,
.prev = &h->ext->children,
// Needed by ta_find_parent():
.size = CHILDREN_SENTINEL,
.ext = h->ext,
},
};
}
return h->ext;
}
/* Set the parent allocation of ptr. If parent==NULL, remove the parent.
* Setting parent==NULL (with ptr!=NULL) always succeeds, and unsets the
* parent of ptr. Operations ptr==NULL always succeed and do nothing.
* Returns true on success, false on OOM.
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
*/
bool ta_set_parent(void *ptr, void *ta_parent)
{
struct ta_header *ch = get_header(ptr);
if (!ch)
return true;
struct ta_ext_header *parent_eh = get_or_alloc_ext_header(ta_parent);
if (ta_parent && !parent_eh) // do nothing on OOM
return false;
// Unlink from previous parent
if (ch->next) {
ch->next->prev = ch->prev;
ch->prev->next = ch->next;
ch->next = ch->prev = NULL;
}
// Link to new parent - insert at end of list (possibly orders destructors)
if (parent_eh) {
struct ta_header *children = &parent_eh->children;
ch->next = children;
ch->prev = children->prev;
children->prev->next = ch;
children->prev = ch;
}
return true;
}
/* 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);
if (!ta_set_parent(ptr, ta_parent)) {
ta_free(ptr);
return NULL;
}
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);
if (!ta_set_parent(ptr, ta_parent)) {
ta_free(ptr);
return NULL;
}
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) {
if (h->next) {
// Relink siblings
h->next->prev = h;
h->prev->next = h;
}
if (h->ext) {
// Relink children
h->ext->header = h;
h->ext->children.next->prev = &h->ext->children;
h->ext->children.prev->next = &h->ext->children;
}
}
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);
struct ta_ext_header *eh = h ? h->ext : NULL;
if (!eh)
return;
while (eh->children.next != &eh->children)
ta_free(PTR_FROM_HEADER(eh->children.next));
}
/* 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;
if (h->ext && h->ext->destructor)
h->ext->destructor(ptr);
ta_free_children(ptr);
if (h->next) {
// Unlink from sibling list
h->next->prev = h->prev;
h->prev->next = h->next;
}
ta_dbg_remove(h);
free(h->ext);
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).
*
* Returns false if ptr==NULL, or on OOM.
*/
bool ta_set_destructor(void *ptr, void (*destructor)(void *))
{
struct ta_ext_header *eh = get_or_alloc_ext_header(ptr);
if (!eh)
return false;
eh->destructor = destructor;
return true;
}
/* Return the ptr's parent allocation, or NULL if there isn't any.
*
* Warning: this has O(N) runtime complexity with N sibling allocations!
*/
void *ta_find_parent(void *ptr)
{
struct ta_header *h = get_header(ptr);
if (!h || !h->next)
return NULL;
for (struct ta_header *cur = h->next; cur != h; cur = cur->next) {
if (cur->size == CHILDREN_SENTINEL)
return PTR_FROM_HEADER(cur->ext->header);
}
return NULL;
}
#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)
{
if (h)
assert(h->canary == CANARY);
}
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;
if (h->ext) {
struct ta_header *s;
for (s = h->ext->children.next; s != &h->ext->children; s = s->next)
size += s->size + get_children_size(s);
}
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