threadLocalVariables.c   [plain text]

/* -*- mode: C++; c-basic-offset: 4; tab-width: 4 -*-
 *
 * Copyright (c) 2010 Apple Inc. All rights reserved.
 *
 * @APPLE_LICENSE_HEADER_START@
 * 
 * This file contains Original Code and/or Modifications of Original Code
 * as defined in and that are subject to the Apple Public Source License
 * Version 2.0 (the 'License'). You may not use this file except in
 * compliance with the License. Please obtain a copy of the License at
 * http://www.opensource.apple.com/apsl/ and read it before using this
 * file.
 * 
 * The Original Code and all software distributed under the License are
 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
 * Please see the License for the specific language governing rights and
 * limitations under the License.
 * 
 * @APPLE_LICENSE_HEADER_END@
 */


#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <stddef.h>
#include <stdio.h>
#include <pthread.h>
#include <Block.h>
#include <malloc/malloc.h>
#include <mach-o/loader.h>
#include <libkern/OSAtomic.h>

#include "dyld_priv.h"


#if __LP64__
	typedef struct mach_header_64		macho_header;
	#define LC_SEGMENT_COMMAND			LC_SEGMENT_64
	typedef struct segment_command_64	macho_segment_command;
	typedef struct section_64			macho_section;
#else
	typedef struct mach_header			macho_header;
	#define LC_SEGMENT_COMMAND			LC_SEGMENT
	typedef struct segment_command		macho_segment_command;
	typedef struct section				macho_section;
#endif

#ifndef S_THREAD_LOCAL_REGULAR
#define S_THREAD_LOCAL_REGULAR                   0x11
#endif

#ifndef S_THREAD_LOCAL_ZEROFILL
#define S_THREAD_LOCAL_ZEROFILL                  0x12
#endif

#ifndef S_THREAD_LOCAL_VARIABLES
#define S_THREAD_LOCAL_VARIABLES                 0x13
#endif

#ifndef S_THREAD_LOCAL_VARIABLE_POINTERS
#define S_THREAD_LOCAL_VARIABLE_POINTERS         0x14
#endif

#ifndef S_THREAD_LOCAL_INIT_FUNCTION_POINTERS
#define S_THREAD_LOCAL_INIT_FUNCTION_POINTERS    0x15
#endif

#ifndef MH_HAS_TLV_DESCRIPTORS
	#define MH_HAS_TLV_DESCRIPTORS 0x800000
#endif


typedef void (*TermFunc)(void*);



#if __has_feature(tls) || __arm64__ || __arm__

typedef struct TLVHandler {
	struct TLVHandler *next;
	dyld_tlv_state_change_handler handler;
	enum dyld_tlv_states state;
} TLVHandler;

// lock-free prepend-only linked list
static TLVHandler * volatile tlv_handlers = NULL;


struct TLVDescriptor
{
	void*			(*thunk)(struct TLVDescriptor*);
	unsigned long	key;
	unsigned long	offset;
};
typedef struct TLVDescriptor  TLVDescriptor;


// implemented in assembly
extern void* tlv_get_addr(TLVDescriptor*);

struct TLVImageInfo
{
	pthread_key_t				key;
	const struct mach_header*	mh;
};
typedef struct TLVImageInfo		TLVImageInfo;

static TLVImageInfo*	tlv_live_images = NULL;
static unsigned int		tlv_live_image_alloc_count = 0;
static unsigned int		tlv_live_image_used_count = 0;
static pthread_mutex_t	tlv_live_image_lock = PTHREAD_MUTEX_INITIALIZER;

static void tlv_set_key_for_image(const struct mach_header* mh, pthread_key_t key)
{
	pthread_mutex_lock(&tlv_live_image_lock);
		if ( tlv_live_image_used_count == tlv_live_image_alloc_count ) {
			unsigned int newCount = (tlv_live_images == NULL) ? 8 : 2*tlv_live_image_alloc_count;
			struct TLVImageInfo* newBuffer = malloc(sizeof(TLVImageInfo)*newCount);
			if ( tlv_live_images != NULL ) {
				memcpy(newBuffer, tlv_live_images, sizeof(TLVImageInfo)*tlv_live_image_used_count);
				free(tlv_live_images);
			}
			tlv_live_images = newBuffer;
			tlv_live_image_alloc_count = newCount;
		}
		tlv_live_images[tlv_live_image_used_count].key = key;
		tlv_live_images[tlv_live_image_used_count].mh = mh;
		++tlv_live_image_used_count;
	pthread_mutex_unlock(&tlv_live_image_lock);
}

static const struct mach_header* tlv_get_image_for_key(pthread_key_t key)
{
	const struct mach_header* result = NULL;
	pthread_mutex_lock(&tlv_live_image_lock);
		for(unsigned int i=0; i < tlv_live_image_used_count; ++i) {
			if ( tlv_live_images[i].key == key ) {
				result = tlv_live_images[i].mh;
				break;
			}
		}
	pthread_mutex_unlock(&tlv_live_image_lock);
	return result;
}


static void
tlv_notify(enum dyld_tlv_states state, void *buffer)
{
	if (!tlv_handlers) return;

	// Always use malloc_size() to ensure allocated and deallocated states 
	// send the same size. tlv_free() doesn't have anything else recorded.
	dyld_tlv_info info = { sizeof(info), buffer, malloc_size(buffer) };
	
	for (TLVHandler *h = tlv_handlers; h != NULL; h = h->next) {
		if (h->state == state  &&  h->handler) {
			h->handler(h->state, &info);
		}
	}
}


// called lazily when TLV is first accessed
__attribute__((visibility("hidden")))
void* tlv_allocate_and_initialize_for_key(pthread_key_t key)
{
	const struct mach_header* mh = tlv_get_image_for_key(key);
	if ( mh == NULL )
		return NULL;	// if data structures are screwed up, don't crash
	
	// first pass, find size and template
	uint8_t*		start = NULL;
	unsigned long	size = 0;
	intptr_t		slide = 0;
	bool			slideComputed = false;
	bool			hasInitializers = false;
	const uint32_t	cmd_count = mh->ncmds;
	const struct load_command* const cmds = (struct load_command*)(((uint8_t*)mh) + sizeof(macho_header));
	const struct load_command* cmd = cmds;
	for (uint32_t i = 0; i < cmd_count; ++i) {
		if ( cmd->cmd == LC_SEGMENT_COMMAND) {
			const macho_segment_command* seg = (macho_segment_command*)cmd;
			if ( !slideComputed && (seg->filesize != 0) ) {
				slide = (uintptr_t)mh - seg->vmaddr;
				slideComputed = true;
			}
			const macho_section* const sectionsStart = (macho_section*)((char*)seg + sizeof(macho_segment_command));
			const macho_section* const sectionsEnd = &sectionsStart[seg->nsects];
			for (const macho_section* sect=sectionsStart; sect < sectionsEnd; ++sect) {
				switch ( sect->flags & SECTION_TYPE ) {
					case S_THREAD_LOCAL_INIT_FUNCTION_POINTERS:
						hasInitializers = true;
						break;
					case S_THREAD_LOCAL_ZEROFILL:
					case S_THREAD_LOCAL_REGULAR:
						if ( start == NULL ) {
							// first of N contiguous TLV template sections, record as if this was only section
							start = (uint8_t*)(sect->addr + slide);
							size = sect->size;
						}
						else {
							// non-first of N contiguous TLV template sections, accumlate values
							const uint8_t* newEnd = (uint8_t*)(sect->addr + slide + sect->size);
							size = newEnd - start;
						}
						break;
				}
			}
		}
		cmd = (const struct load_command*)(((char*)cmd)+cmd->cmdsize);
	}
    // no thread local storage in image: should never happen
    if ( size == 0 )
        return NULL;
	
	// allocate buffer and fill with template
	void* buffer = malloc(size);
	memcpy(buffer, start, size);
	
	// set this thread's value for key to be the new buffer.
	pthread_setspecific(key, buffer);

	// send tlv state notifications
	tlv_notify(dyld_tlv_state_allocated, buffer);
	
	// second pass, run initializers
	if ( hasInitializers ) {
		cmd = cmds;
		for (uint32_t i = 0; i < cmd_count; ++i) {
			if ( cmd->cmd == LC_SEGMENT_COMMAND) {
				const macho_segment_command* seg = (macho_segment_command*)cmd;
				const macho_section* const sectionsStart = (macho_section*)((char*)seg + sizeof(macho_segment_command));
				const macho_section* const sectionsEnd = &sectionsStart[seg->nsects];
				for (const macho_section* sect=sectionsStart; sect < sectionsEnd; ++sect) {
					if ( (sect->flags & SECTION_TYPE) == S_THREAD_LOCAL_INIT_FUNCTION_POINTERS ) {
						typedef void (*InitFunc)(void);
						InitFunc* funcs = (InitFunc*)(sect->addr + slide);
						const size_t count = sect->size / sizeof(uintptr_t);
						for (size_t j=count; j > 0; --j) {
							InitFunc func = funcs[j-1];
							func();
						}
					}
				}
			}
			cmd = (const struct load_command*)(((char*)cmd)+cmd->cmdsize);
		}
	}
	return buffer;
}


// pthread destructor for TLV storage
static void
tlv_free(void *storage)
{
	tlv_notify(dyld_tlv_state_deallocated, storage);
	free(storage);
}


// called when image is loaded
static void tlv_initialize_descriptors(const struct mach_header* mh)
{
	pthread_key_t	key = 0;
	intptr_t		slide = 0;
	bool			slideComputed = false;
	const uint32_t cmd_count = mh->ncmds;
	const struct load_command* const cmds = (struct load_command*)(((uint8_t*)mh) + sizeof(macho_header));
	const struct load_command* cmd = cmds;
	for (uint32_t i = 0; i < cmd_count; ++i) {
		if ( cmd->cmd == LC_SEGMENT_COMMAND) {
			const macho_segment_command* seg = (macho_segment_command*)cmd;
			if ( !slideComputed && (seg->filesize != 0) ) {
				slide = (uintptr_t)mh - seg->vmaddr;
				slideComputed = true;
			}
			const macho_section* const sectionsStart = (macho_section*)((char*)seg + sizeof(macho_segment_command));
			const macho_section* const sectionsEnd = &sectionsStart[seg->nsects];
			for (const macho_section* sect=sectionsStart; sect < sectionsEnd; ++sect) {
				if ( (sect->flags & SECTION_TYPE) == S_THREAD_LOCAL_VARIABLES ) {
					if ( sect->size != 0 ) {
						// allocate pthread key when we first discover this image has TLVs
						if ( key == 0 ) {
							int result = pthread_key_create(&key, &tlv_free);
							if ( result != 0 )
								abort();
							tlv_set_key_for_image(mh, key);
						}
						// initialize each descriptor
						TLVDescriptor* start = (TLVDescriptor*)(sect->addr + slide);
						TLVDescriptor* end = (TLVDescriptor*)(sect->addr + sect->size + slide);
						for (TLVDescriptor* d=start; d < end; ++d) {
							d->thunk = tlv_get_addr;
							d->key = key;
							//d->offset = d->offset;  // offset unchanged
						}
					}
				}
			}
		}
		cmd = (const struct load_command*)(((char*)cmd)+cmd->cmdsize);
	}
}


static void tlv_load_notification(const struct mach_header* mh, intptr_t slide)
{
	// This is called on all images, even those without TLVs. So we want this to be fast.
	// The linker sets MH_HAS_TLV_DESCRIPTORS so we don't have to search images just to find the don't have TLVs.
	if ( mh->flags & MH_HAS_TLV_DESCRIPTORS )
		tlv_initialize_descriptors(mh);
}


void dyld_register_tlv_state_change_handler(enum dyld_tlv_states state, dyld_tlv_state_change_handler handler)
{
	TLVHandler *h = malloc(sizeof(TLVHandler));
	h->state = state;
	h->handler = Block_copy(handler);

	TLVHandler *old;
	do {
		old = tlv_handlers;
		h->next = old;
	} while (! OSAtomicCompareAndSwapPtrBarrier(old, h, (void * volatile *)&tlv_handlers));
}


void dyld_enumerate_tlv_storage(dyld_tlv_state_change_handler handler)
{
	pthread_mutex_lock(&tlv_live_image_lock);
		unsigned int count = tlv_live_image_used_count;
		void *list[count];
		for (unsigned int i = 0; i < count; ++i) {
			list[i] = pthread_getspecific(tlv_live_images[i].key);
		}
	pthread_mutex_unlock(&tlv_live_image_lock);

	for (unsigned int i = 0; i < count; ++i) {
		if (list[i]) {
			dyld_tlv_info info = { sizeof(info), list[i], malloc_size(list[i]) };
			handler(dyld_tlv_state_allocated, &info);
		}
	}
}


//
//  thread_local terminators
//
// C++ 0x allows thread_local C++ objects which have constructors run
// on the thread before any use of the object and the object's destructor
// is run on the thread when the thread terminates.
//
// To support this libdyld gets a pthread key early in process start up and
// uses tlv_finalize and the key's destructor function.  This key must be
// allocated before any thread local variables are instantiated because when
// a thread is terminated, the pthread package runs the destructor function
// on each key's storage values in key allocation order.  Since we want
// C++ objects to be destructred before they are deallocated, we need the 
// destructor key to come before the deallocation key.
//

struct TLVTerminatorListEntry
{
    TermFunc    termFunc;
    void*       objAddr;
};

struct TLVTerminatorList
{
    uint32_t                        allocCount;
    uint32_t                        useCount;
    struct TLVTerminatorListEntry   entries[1];  // variable length
};


static pthread_key_t tlv_terminators_key = 0;

void _tlv_atexit(TermFunc func, void* objAddr)
{
    // NOTE: this does not need locks because it only operates on current thread data
	struct TLVTerminatorList* list = (struct TLVTerminatorList*)pthread_getspecific(tlv_terminators_key);
    if ( list == NULL ) {
        // handle first allocation
        list = (struct TLVTerminatorList*)malloc(offsetof(struct TLVTerminatorList, entries[1]));
        list->allocCount = 1;
        list->useCount = 1;
        list->entries[0].termFunc = func;
        list->entries[0].objAddr = objAddr;
        pthread_setspecific(tlv_terminators_key, list);
    }
    else {
        if ( list->useCount == list->allocCount ) {
            // handle resizing allocation 
            uint32_t newAllocCount = list->allocCount * 2;
            size_t newAllocSize = offsetof(struct TLVTerminatorList, entries[newAllocCount]);
            struct TLVTerminatorList* newlist = (struct TLVTerminatorList*)malloc(newAllocSize);
            newlist->allocCount = newAllocCount;
            newlist->useCount = list->useCount;
            for(uint32_t i=0; i < list->useCount; ++i)
                newlist->entries[i] = list->entries[i];
            pthread_setspecific(tlv_terminators_key, newlist);
            free(list);
            list = newlist;
        }
        // handle appending new entry
        list->entries[list->useCount].termFunc = func;
        list->entries[list->useCount].objAddr = objAddr;
        list->useCount += 1;
    }
}

// called by pthreads when the current thread is going away and 
// _tlv_atexit() has been called on the thread.
static void tlv_finalize(void* storage)
{
    struct TLVTerminatorList* list = (struct TLVTerminatorList*)storage;
    // destroy in reverse order of construction
    for(uint32_t i=list->useCount; i > 0 ; --i) {
        struct TLVTerminatorListEntry* entry = &list->entries[i-1];
        if ( entry->termFunc != NULL ) {
            (*entry->termFunc)(entry->objAddr);
        }
    }
    free(storage);
}

// <rdar://problem/13741816>
// called by exit() before it calls cxa_finalize() so that thread_local
// objects are destroyed before global objects.
void _tlv_exit()
{
	void* termFuncs = pthread_getspecific(tlv_terminators_key);
	if ( termFuncs != NULL )
		tlv_finalize(termFuncs);
}


__attribute__((visibility("hidden")))
void tlv_initializer()
{
    // create pthread key to handle thread_local destructors
    // NOTE: this key must be allocated before any keys for TLV
    // so that _pthread_tsd_cleanup will run destructors before deallocation
    (void)pthread_key_create(&tlv_terminators_key, &tlv_finalize);
       
    // register with dyld for notification when images are loaded
	_dyld_register_func_for_add_image(tlv_load_notification);

}


// linked images with TLV have references to this symbol, but it is never used at runtime
void _tlv_bootstrap()
{
	abort();
}



#else



void dyld_register_tlv_state_change_handler(enum dyld_tlv_states state, dyld_tlv_state_change_handler handler)
{
}

void dyld_enumerate_tlv_storage(dyld_tlv_state_change_handler handler)
{
}

void _tlv_exit()
{
}

void _tlv_atexit(TermFunc func, void* objAddr)
{
}

__attribute__((visibility("hidden")))
void tlv_initializer()
{
}



#endif // __has_feature(tls)