Namespaces
Enumerations
enum SyncOpt| Name | Value |
|---|---|
| AcqRelOps | 0 |
| Fence | 1 |
| None | 2 |
An enumeration of various synchronization options to use when performing
memory transfer operations with WellDefinedCopy(To|From).
:: AcqRelOps ::
Use either memory_order_acquire (CopyFrom) or memory_order_release (CopyTo)
on every atomic load/store operation during the transfer to/from the shared
buffer.
:: Fence ::
Use either a memory_order_acquire thread fence (CopyFrom) after the transfer
operation, or a memory_order_release (CopyTo) thread fence before the
operation, and memory_order_relaxed for each of the atomic load/store
operations during the transfer.
:: None ::
Simply use memory_order_relaxed for each of the atomic load/store
operations during the transfer. Do not actually introduce any
explicit synchronization behavior.
WARNING: Use cases for this transfer mode tend to unusual. Users will almost
always want some form of synchronization to take place during their
transfers. One example of where it may be appropriate to use SyncOpt::None
might be a situation where users are attempting to observe the state of more
than one object while inside of a sequence lock read transaction, and the
user has decided that it is better to use a thread fence than to use acquire
semantics on each element transferred. Such a sequence might look something
like this:
Foo foo1, foo2;
Bar bar1, bar2;
...
WellDefinedCopyFrom
<SyncOpt
::None, alignof(Foo)>(
&foo
1,
&src
_foo1, sizeof(foo1));
WellDefinedCopyFrom
<SyncOpt
::None, alignof(Foo)>(
&foo
2,
&src
_foo2, sizeof(foo2));
WellDefinedCopyFrom
<SyncOpt
::None, alignof(Foo)>(
&bar
1,
&src
_bar1, sizeof(bar1));
WellDefinedCopyFrom
<SyncOpt
::Fence, alignof(Foo)>(
&bar
2,
&src
_bar2, sizeof(bar2));
Note that it is the _last_ transfer operation which includes the fence. In
the case of a CopyTo operation (when publishing data) it would be the _first_
operation which included the fence, not the last.
Defined at line 54 of file ../../zircon/system/ulib/concurrent/include/lib/concurrent/common.h
Records
Functions
-
template <typename Transaction>ChainLockGuard<Transaction> <deduction guide for ChainLockGuard> (ChainLock<Transaction> & )Deduction guides that allow the Transaction type to be deduced from the ChainLock instantiation
passed to the constructor.
Example:
using MyChainLock = concurrent::ChainLock
<MyTransaction
>;
using concurrent::ChainLockGuard;
MyChainLock lock;
ChainLockGuard guard{lock}; // Deduces ChainLockGuard
<MyTransaction
> from the lock argument.
-
template <typename Transaction>ChainLockGuard<Transaction> <deduction guide for ChainLockGuard> (ChainLock<Transaction> & lock, typename ChainLockGuard<Transaction>::AdoptTag ) -
template <typename Transaction>ChainLockGuard<Transaction> <deduction guide for ChainLockGuard> (ChainLock<Transaction> & lock, typename ChainLockGuard<Transaction>::DeferTag ) -
template <typename Transaction>ChainLockGuard<Transaction> <deduction guide for ChainLockGuard> (ChainLock<Transaction> & lock, typename ChainLockGuard<Transaction>::TakeTag ) -
template <SyncOpt kSyncOpt = SyncOpt::AcqRelOps, size_t kWorstCaseAlignment = 1>void WellDefinedCopyTo (void * dst, const void * src, size_t size_bytes)Copy |size_bytes| bytes from |src| to |dst| using atomic store operations to move
the element into |dst| so that the behavior of the system is always well
defined, even if there is a WellDefinedCopyFrom operation reading from the
memory pointed to by |dst| concurrent with this CopyTo operation.
WellDefinedCopyTo has memcpy semantics, not memmove semantics. In other
words, it is illegal for |src| or |dst| to overlap in any way.
While it is not required, by default, that |src| and |dst| have any specific
alignment, both |src| and |dst| *must* have the _same_ alignment.
IOW: (|src|
&
0x7) *must* equal (|dst|
&
0x7)
:: Template Args ::
|kSyncOpt|
Controls the options for memory order synchronization. See the comments in
lib/concurrent/common.h for details.
|kWorstCaseAlignment|
An explicit guarantee of the worst case alignment that |src|/|dst| will obey.
When this alignment guarantee is greater than or equal to the maximum
internal transfer granularity of 64 bits, the initial explicit alignment step
of the operation can be optimized away for a minor performance gain.
Defined at line 47 of file ../../zircon/system/ulib/concurrent/include/lib/concurrent/copy.h
-
template <SyncOpt kSyncOpt = SyncOpt::AcqRelOps, size_t kWorstCaseAlignment = 1>void WellDefinedCopyFrom (void * dst, const void * src, size_t size_bytes)Copy |size_bytes| bytes from |src| to |dst| using atomic load operations to load the
element from |src| so that the behavior of the system is always well defined,
even if there is a WellDefinedCopyTo operation writing to the memory pointed
to by |src| concurrent with this CopyFrom operation.
WellDefinedCopyFrom has memcpy semantics, not memmove semantics. In other
words, it is illegal for |src| or |dst| to overlap in any way.
While it is not required, by default, that |src| and |dst| have any specific
alignment, both |src| and |dst| *must* have the _same_ alignment.
IOW: (|src|
&
0x7) *must* equal (|dst|
&
0x7)
:: Template Args ::
|kSyncOpt|
Controls the options for memory order synchronization. See the comments in
lib/concurrent/common.h for details.
|kWorstCaseAlignment|
An explicit guarantee of the worst case alignment that |src|/|dst| will obey.
When this alignment guarantee is greater than or equal to the maximum
internal transfer granularity of 64 bits, the initial explicit alignment step
of the operation can be optimized away for a minor performance gain.
Defined at line 87 of file ../../zircon/system/ulib/concurrent/include/lib/concurrent/copy.h
-
template <typename Transaction>bool AcquireBothOrBackoff (ChainLock<Transaction> & firstChainLock<Transaction> & second)Utility to simplify expressions where two chain locks need to be acquired with backoff. Returns
true if both locks where acquired or false if the current transaction needs to back off.
Defined at line 405 of file ../../zircon/system/ulib/concurrent/include/lib/concurrent/chainlock.h