| PMEMOBJ_TX_BEGIN(3) | PMDK Programmer's Manual | PMEMOBJ_TX_BEGIN(3) |
pmemobj_tx_stage(),
pmemobj_tx_begin(), pmemobj_tx_lock(), pmemobj_tx_xlock(), pmemobj_tx_abort(), pmemobj_tx_commit(), pmemobj_tx_end(), pmemobj_tx_errno(), pmemobj_tx_process(),
TX_BEGIN_PARAM(), TX_BEGIN_CB(), TX_BEGIN(), TX_ONABORT, TX_ONCOMMIT, TX_FINALLY, TX_END,
pmemobj_tx_log_append_buffer(), pmemobj_tx_xlog_append_buffer(), pmemobj_tx_log_auto_alloc(), pmemobj_tx_log_snapshots_max_size(), pmemobj_tx_log_intents_max_size(),
pmemobj_tx_set_user_data(), pmemobj_tx_get_user_data(),
pmemobj_tx_set_failure_behavior(), pmemobj_tx_get_failure_behavior() - transactional object manipulation
#include <libpmemobj.h>
enum pobj_tx_stage pmemobj_tx_stage(void);
int pmemobj_tx_begin(PMEMobjpool *pop, jmp_buf *env, enum pobj_tx_param, ...);
int pmemobj_tx_lock(enum tx_lock lock_type, void *lockp);
int pmemobj_tx_xlock(enum tx_lock lock_type, void *lockp, uint64_t flags);
void pmemobj_tx_abort(int errnum);
void pmemobj_tx_commit(void);
int pmemobj_tx_end(void);
int pmemobj_tx_errno(void);
void pmemobj_tx_process(void);
TX_BEGIN_PARAM(PMEMobjpool *pop, ...)
TX_BEGIN_CB(PMEMobjpool *pop, cb, arg, ...)
TX_BEGIN(PMEMobjpool *pop)
TX_ONABORT
TX_ONCOMMIT
TX_FINALLY
TX_END
int pmemobj_tx_log_append_buffer(enum pobj_log_type type, void *addr, size_t size);
int pmemobj_tx_xlog_append_buffer(enum pobj_log_type type, void *addr, size_t size, uint64_t flags);
int pmemobj_tx_log_auto_alloc(enum pobj_log_type type, int on_off);
size_t pmemobj_tx_log_snapshots_max_size(size_t *sizes, size_t nsizes);
size_t pmemobj_tx_log_intents_max_size(size_t nintents);
void pmemobj_tx_set_user_data(void *data);
void *pmemobj_tx_get_user_data(void);
void pmemobj_tx_set_failure_behavior(enum pobj_tx_failure_behavior behavior);
enum pobj_tx_failure_behavior pmemobj_tx_get_failure_behavior(void);
The non-transactional functions and macros described in pmemobj_alloc(3), pmemobj_list_insert(3) and POBJ_LIST_HEAD(3) only guarantee the atomicity of a single operation on an object. In case of more complex changes involving multiple operations on an object, or allocation and modification of multiple objects, data consistency and fail-safety may be provided only by using atomic transactions.
A transaction is defined as series of operations on persistent memory objects that either all occur, or nothing occurs. In particular, if the execution of a transaction is interrupted by a power failure or a system crash, it is guaranteed that after system restart, all the changes made as a part of the uncompleted transaction will be rolled back, restoring the consistent state of the memory pool from the moment when the transaction was started.
Note that transactions do not provide atomicity with respect to other threads. All the modifications performed within the transactions are immediately visible to other threads. Therefore it is the responsibility of the application to implement a proper thread synchronization mechanism.
Each thread may have only one transaction open at a time, but that transaction may be nested. Nested transactions are flattened. Committing the nested transaction does not commit the outer transaction; however, errors in the nested transaction are propagated up to the outermost level, resulting in the interruption of the entire transaction.
Each transaction is visible only for the thread that started it. No other threads can add operations, commit or abort the transaction initiated by another thread. Multiple threads may have transactions open on a given memory pool at the same time.
Please see the CAVEATS section below for known limitations of the transactional API.
The pmemobj_tx_stage() function returns the current transaction stage for a thread. Stages are changed only by the pmemobj_tx_*() functions. Transaction stages are defined as follows:
The pmemobj_tx_begin() function starts a new transaction in the current thread. If called within an open transaction, it starts a nested transaction. The caller may use the env argument to provide a pointer to a calling environment to be restored in case of transaction abort. This information must be provided by the caller using the setjmp(3) macro.
A new transaction may be started only if the current stage is TX_STAGE_NONE or TX_STAGE_WORK. If successful, the transaction stage changes to TX_STAGE_WORK. Otherwise, the stage is changed to TX_STAGE_ONABORT.
Optionally, a list of parameters for the transaction may be provided. Each parameter consists of a type followed by a type-specific number of values. Currently there are 4 types:
Using TX_PARAM_MUTEX or TX_PARAM_RWLOCK causes the specified lock to be acquired at the beginning of the transaction. TX_PARAM_RWLOCK acquires the lock for writing. It is guaranteed that pmemobj_tx_begin() will acquire all locks prior to successful completion, and they will be held by the current thread until the outermost transaction is finished. Locks are taken in order from left to right. To avoid deadlocks, the user is responsible for proper lock ordering.
TX_PARAM_CB registers the specified callback function to be executed at each transaction stage. For TX_STAGE_WORK, the callback is executed prior to commit. For all other stages, the callback is executed as the first operation after a stage change. It will also be called after each transaction; in this case the stage parameter will be set to TX_STAGE_NONE. pmemobj_tx_callback must be compatible with:
void func(PMEMobjpool *pop, enum pobj_tx_stage stage, void *arg)
pop is a pool identifier used in pmemobj_tx_begin(), stage is a current transaction stage and arg is the second parameter of TX_PARAM_CB. Without considering transaction nesting, this mechanism can be considered an alternative method for executing code between stages (instead of TX_ONCOMMIT, TX_ONABORT, etc). However, there are 2 significant differences when nested transactions are used:
Note that TX_PARAM_CB does not replace the TX_ONCOMMIT, TX_ONABORT, etc. macros. They can be used together: the callback will be executed before a TX_ONCOMMIT, TX_ONABORT, etc. section.
TX_PARAM_CB can be used when the code dealing with transaction stage changes is shared between multiple users or when it must be executed only in the outer transaction. For example it can be very useful when the application must synchronize persistent and transient state.
The pmemobj_tx_lock() function acquires the lock lockp of type lock_type and adds it to the current transaction. lock_type may be TX_LOCK_MUTEX or TX_LOCK_RWLOCK; lockp must be of type PMEMmutex or PMEMrwlock, respectively. If lock_type is TX_LOCK_RWLOCK the lock is acquired for writing. If the lock is not successfully acquired, the function returns an error number. This function must be called during TX_STAGE_WORK.
The pmemobj_tx_xlock() function behaves exactly the same as pmemobj_tx_lock() when flags equals POBJ_XLOCK_NO_ABORT. When flags equals 0 and if the lock is not successfully acquired,the transaction is aborted. flags is a bitmask of the following values:
pmemobj_tx_abort() aborts the current transaction and causes a transition to TX_STAGE_ONABORT. If errnum is equal to 0, the transaction error code is set to ECANCELED; otherwise, it is set to errnum. This function must be called during TX_STAGE_WORK.
The pmemobj_tx_commit() function commits the current open transaction and causes a transition to TX_STAGE_ONCOMMIT. If called in the context of the outermost transaction, all the changes may be considered as durably written upon successful completion. This function must be called during TX_STAGE_WORK.
The pmemobj_tx_end() function performs a cleanup of the current transaction. If called in the context of the outermost transaction, it releases all the locks acquired by pmemobj_tx_begin() for outer and nested transactions. If called in the context of a nested transaction, it returns to the context of the outer transaction in TX_STAGE_WORK, without releasing any locks. The pmemobj_tx_end() function can be called during TX_STAGE_NONE if transitioned to this stage using pmemobj_tx_process(). If not already in TX_STAGE_NONE, it causes the transition to TX_STAGE_NONE. pmemobj_tx_end must always be called for each pmemobj_tx_begin(), even if starting the transaction failed. This function must not be called during TX_STAGE_WORK.
The pmemobj_tx_errno() function returns the error code of the last transaction.
The pmemobj_tx_process() function performs the actions associated with the current stage of the transaction, and makes the transition to the next stage. It must be called in a transaction. The current stage must always be obtained by a call to pmemobj_tx_stage(). pmemobj_tx_process() performs the following transitions in the transaction stage flow:
pmemobj_tx_process() must not be called after calling pmemobj_tx_end() for the outermost transaction.
In addition to the above API, libpmemobj(7) offers a more intuitive method of building transactions using the set of macros described below. When using these macros, the complete transaction flow looks like this:
TX_BEGIN(Pop) {
/* the actual transaction code goes here... */
} TX_ONCOMMIT {
/*
* optional - executed only if the above block
* successfully completes
*/
} TX_ONABORT {
/*
* optional - executed only if starting the transaction fails,
* or if transaction is aborted by an error or a call to
* pmemobj_tx_abort()
*/
} TX_FINALLY {
/*
* optional - if exists, it is executed after
* TX_ONCOMMIT or TX_ONABORT block
*/
} TX_END /* mandatory */
TX_BEGIN_PARAM(PMEMobjpool *pop, ...)
TX_BEGIN_CB(PMEMobjpool *pop, cb, arg, ...)
TX_BEGIN(PMEMobjpool *pop)
The TX_BEGIN_PARAM(), TX_BEGIN_CB() and TX_BEGIN() macros start a new transaction in the same way as pmemobj_tx_begin(), except that instead of the environment buffer provided by a caller, they set up the local jmp_buf buffer and use it to catch the transaction abort. The TX_BEGIN() macro starts a transaction without any options. TX_BEGIN_PARAM may be used when there is a need to acquire locks prior to starting a transaction (such as for a multi-threaded program) or set up a transaction stage callback. TX_BEGIN_CB is just a wrapper around TX_BEGIN_PARAM that validates the callback signature. (For compatibility there is also a TX_BEGIN_LOCK macro, which is an alias for TX_BEGIN_PARAM). Each of these macros must be followed by a block of code with all the operations that are to be performed atomically.
The TX_ONABORT macro starts a block of code that will be executed only if starting the transaction fails due to an error in pmemobj_tx_begin(), or if the transaction is aborted. This block is optional, but in practice it should not be omitted. If it is desirable to crash the application when a transaction aborts and there is no TX_ONABORT section, the application can define the POBJ_TX_CRASH_ON_NO_ONABORT macro before inclusion of <libpmemobj.h>. This provides a default TX_ONABORT section which just calls abort(3).
The TX_ONCOMMIT macro starts a block of code that will be executed only if the transaction is successfully committed, which means that the execution of code in the TX_BEGIN() block has not been interrupted by an error or by a call to pmemobj_tx_abort(). This block is optional.
The TX_FINALLY macro starts a block of code that will be executed regardless of whether the transaction is committed or aborted. This block is optional.
The TX_END macro cleans up and closes the transaction started by the TX_BEGIN() / TX_BEGIN_PARAM() / TX_BEGIN_CB() macros. It is mandatory to terminate each transaction with this macro. If the transaction was aborted, errno is set appropriately.
From libpmemobj implementation perspective there are two types of operations in a transaction:
pmemobj_tx_add_range(3) and all its variants belong to the snapshots group.
pmemobj_tx_alloc(3) (with its variants), pmemobj_tx_free(3), pmemobj_tx_realloc(3) (with its variants) and pmemobj_tx_publish(3) belong to the intents group. Even though pmemobj_tx_alloc() allocates memory immediately, it modifies only the runtime state and postpones persistent memory modifications to the commit phase. pmemobj_tx_free(3) cannot free the object immediately, because of possible transaction rollback, so it postpones both the action and persistent memory modifications to the commit phase. pmemobj_tx_realloc(3) is just a combination of those two. pmemobj_tx_publish(3) postpones reservations and deferred frees to the commit phase.
Those two types of operations (snapshots and intents) require that libpmemobj builds a persistent log of operations. Intent log (also known as a “redo log”) is applied on commit and snapshot log (also known as an “undo log”) is applied on abort.
When libpmemobj transaction starts, it’s not possible to predict how much persistent memory space will be needed for those logs. This means that libpmemobj must internally allocate this space whenever it’s needed. This has two downsides:
To solve both of these problems libpmemobj exposes the following functions:
pmemobj_tx_log_append_buffer() appends a given range of memory [addr, addr + size) to the log type of the current transaction. type can be one of the two values (with meanings described above):
The range of memory must belong to the same pool the transaction is on and must not be used by more than one thread at the same time. The latter condition can be verified with tx.debug.verify_user_buffers ctl (see pmemobj_ctl_get(3)).
The pmemobj_tx_xlog_append_buffer() function behaves exactly the same as pmemobj_tx_log_append_buffer() when flags equals zero. flags is a bitmask of the following values:
pmemobj_tx_log_snapshots_max_size calculates the maximum size of a buffer which will be able to hold nsizes snapshots, each of size sizes[i]. Application should not expect this function to return the same value between restarts. In future versions of libpmemobj this function can return smaller (because of better accuracy or space optimizations) or higher (because of higher alignment required for better performance) value. This function is independent of transaction stage and can be called both inside and outside of transaction. If the returned value S is greater than PMEMOBJ_MAX_ALLOC_SIZE, the buffer should be split into N chunks of size PMEMOBJ_MAX_ALLOC_SIZE, where N is equal to (S / PMEMOBJ_MAX_ALLOC_SIZE) (rounded down) and the last chunk of size (S - (N * PMEMOBJ_MAX_ALLOC_SIZE)).
pmemobj_tx_log_intents_max_size calculates the maximum size of a buffer which will be able to hold nintents intents. Just like with pmemobj_tx_log_snapshots_max_size, application should not expect this function to return the same value between restarts, for the same reasons. This function is independent of transaction stage and can be called both inside and outside of transaction.
pmemobj_tx_log_auto_alloc() disables (on_off set to 0) or enables (on_off set to 1) automatic allocation of internal logs of given type. It can be used to verify that the buffer set with pmemobj_tx_log_append_buffer() is big enough to hold the log, without reaching out-of-space scenario.
The pmemobj_tx_set_user_data() function associates custom volatile state, represented by pointer data, with the current transaction. This state can later be retrieved using pmemobj_tx_get_user_data() function. If pmemobj_tx_set_user_data() was not called for a current transaction, pmemobj_tx_get_user_data() will return NULL. These functions must be called during TX_STAGE_WORK or TX_STAGE_ONABORT or TX_STAGE_ONCOMMIT or TX_STAGE_FINALLY.
pmemobj_tx_set_failure_behavior() specifies what should happen in case of an error within the transaction. It only affects functions which take a NO_ABORT flag. If pmemobj_tx_set_failure_behavior() is called with POBJ_TX_FAILURE_RETURN a NO_ABORT flag is implicitly passed to all functions which accept this flag. If called with POBJ_TX_FAILURE_ABORT then all functions abort the transaction (unless NO_ABORT flag is passed explicitly). This setting is inherited by inner transactions. It does not affect any of the outer transactions. Aborting on failure is the default behavior. pmemobj_tx_get_failure_behavior() returns failure behavior for the current transaction. Both pmemobj_tx_set_failure_behavior() and pmemobj_tx_get_failure_behavior() must be called during TX_STAGE_WORK.
The pmemobj_tx_stage() function returns the stage of the current transaction stage for a thread.
On success, pmemobj_tx_begin() returns 0. Otherwise, an error number is returned.
The pmemobj_tx_begin() and pmemobj_tx_lock() functions return zero if lockp is successfully added to the transaction. Otherwise, an error number is returned.
The pmemobj_tx_xlock() function return zero if lockp is successfully added to the transaction. Otherwise, the error number is returned, errno is set and when flags do not contain POBJ_XLOCK_NO_ABORT, the transaction is aborted.
The pmemobj_tx_abort() and pmemobj_tx_commit() functions return no value.
The pmemobj_tx_end() function returns 0 if the transaction was successful. Otherwise it returns the error code set by pmemobj_tx_abort(). Note that pmemobj_tx_abort() can be called internally by the library.
The pmemobj_tx_errno() function returns the error code of the last transaction.
The pmemobj_tx_process() function returns no value.
On success, pmemobj_tx_log_append_buffer() returns 0. Otherwise, the stage is changed to TX_STAGE_ONABORT, errno is set appropriately and transaction is aborted.
On success, pmemobj_tx_xlog_append_buffer() returns 0. Otherwise, the error number is returned, errno is set and when flags do not contain POBJ_XLOG_NO_ABORT, the transaction is aborted.
On success, pmemobj_tx_log_auto_alloc() returns 0. Otherwise, the transaction is aborted and an error number is returned.
On success, pmemobj_tx_log_snapshots_max_size() returns size of the buffer. On failure it returns SIZE_MAX and sets errno appropriately.
On success, pmemobj_tx_log_intents_max_size() returns size of the buffer. On failure it returns SIZE_MAX and sets errno appropriately.
Transaction flow control is governed by the setjmp(3) and longjmp(3) macros, and they are used in both the macro and function flavors of the API. The transaction will longjmp on transaction abort. This has one major drawback, which is described in the ISO C standard subsection 7.13.2.1. It says that the values of objects of automatic storage duration that are local to the function containing the setjmp invocation that do not have volatile-qualified type and have been changed between the setjmp invocation and longjmp call are indeterminate.
The following example illustrates the issue described above.
int *bad_example_1 = (int *)0xBAADF00D;
int *bad_example_2 = (int *)0xBAADF00D;
int *bad_example_3 = (int *)0xBAADF00D;
int * volatile good_example = (int *)0xBAADF00D;
TX_BEGIN(pop) {
bad_example_1 = malloc(sizeof(int));
bad_example_2 = malloc(sizeof(int));
bad_example_3 = malloc(sizeof(int));
good_example = malloc(sizeof(int));
/* manual or library abort called here */
pmemobj_tx_abort(EINVAL);
} TX_ONCOMMIT {
/*
* This section is longjmp-safe
*/
} TX_ONABORT {
/*
* This section is not longjmp-safe
*/
free(good_example); /* OK */
free(bad_example_1); /* undefined behavior */
} TX_FINALLY {
/*
* This section is not longjmp-safe on transaction abort only
*/
free(bad_example_2); /* undefined behavior */
} TX_END
free(bad_example_3); /* undefined behavior */
Objects which are not volatile-qualified, are of automatic storage duration and have been changed between the invocations of setjmp(3) and longjmp(3) (that also means within the work section of the transaction after TX_BEGIN()) should not be used after a transaction abort, or should be used with utmost care. This also includes code after the TX_END macro.
libpmemobj(7) is not cancellation-safe. The pool will never be corrupted because of a canceled thread, but other threads may stall waiting on locks taken by that thread. If the application wants to use pthread_cancel(3), it must disable cancellation before calling any libpmemobj(7) APIs (see pthread_setcancelstate(3) with PTHREAD_CANCEL_DISABLE), and re-enable it afterwards. Deferring cancellation (pthread_setcanceltype(3) with PTHREAD_CANCEL_DEFERRED) is not safe enough, because libpmemobj(7) internally may call functions that are specified as cancellation points in POSIX.
libpmemobj(7) relies on the library destructor being called from the main thread. For this reason, all functions that might trigger destruction (e.g. dlclose(3)) should be called in the main thread. Otherwise some of the resources associated with that thread might not be cleaned up properly.
dlclose(3), longjmp(3), pmemobj_tx_add_range(3), pmemobj_tx_alloc(3), pthread_setcancelstate(3), pthread_setcanceltype(3), setjmp(3), libpmemobj(7) and <https://pmem.io>
| 2020-10-28 | PMDK - pmemobj API version 2.3 |