CN106354572A - Multi-thread data transmission method - Google Patents

Abstract

The invention provides a multi-thread data transmission method. Two integer values including write Pos and read Pos are defined, wherein the write Pos refers to the current position of a data writing thread for writing data into an array space, and the read Pos refers to the current position of a data reading thread for reading the data from the array space. The write Pos value is automatically increased when the data are successfully written, and the read Pos value is automatically increased when the data are successfully read. An integer mask Pos Mask is defined, and the value of the integer mask Pos Mask is N-1. The position of the actual array space for currently writing and reading the data is determined by AND operation of the value of the Pos Mask and values of the read-write positions, namely, the current writing position is data PtrArray [write Pos &Pos Mask], and the reading position is data PtrArray [read Pos &Pos Mask]. Data transmission between two threads can be realized, namely, a queue container as a critical resource is accessed in an unlocked manner, and performance is greatly improved.

Description

A kind of multi-thread data transmission method

Technical field

The present invention relates to a kind of multi-thread data transmission method, more particularly to one kind is be applied to multi-threading correspondence, many The multi-thread data transmission method of thread accesses.

Background technology

Typically in multi-threading correspondence, can there are some visits to critical resource (multiple threads all may be accessed) simultaneously Ask, because critical resource is public, if read and write by multiple threads, critical resource can become a kind of unknown shape simultaneously State, is unsafe, and program can be led to produce baffled mistake, so program often passes through thread when using critical resource Lock, to ensure that the access of critical resource is safely controllable, ensures have any time to only have one thread of meeting to grasp by thread lock Make critical resource；But this mechanism can lead to other thread process waiting states, thus reduce system process performance.

Content of the invention

The technical problem to be solved in the present invention is to provide a kind of multi-thread data transmission method, can solve the problem that between two threads Data transfer, realizes the access of no lock, performance is greatly improved to this critical resource of queue container.

The technical solution used in the present invention is as follows: a kind of multi-thread data transmission method, method particularly includes:

During initialization system, apply for a contiguous memory space that can accommodate n pointer；Described n=2^m, m is more than or equal to 1 Integer；Each pointer position is a unit, and n pointer position forms an array space dataptrarray；Described finger Pin points to user data address；

Define two integer value: writepos and readpos；Wherein, writepos represents that data write thread to be currently ready writing Enter the position in described array space；Readpos represents that digital independent thread is current and to be ready reading data from described array space Position；Data writes successfully, then writepos value is from increasing；Digital independent success, then the increasing certainly of readpos value, will be successfully read The pointer value of array locus reset, after clearing, write thread finds that this position is zero again, can be normally written new User pointer data；

When but dataptrarray writes full data do not walked by timely reading, failure can be write；When in dataptrarray When all values are all zero, failure can be read；

Define an integer mask posmask, its value is n-1；Done and computing by the value of the value of posmask and read and write position, To determine the position in the current real array space being about to and writing and read, that is, presently written position should be Dataptrarray [writepos＆posmask], the position of reading is dataptrarray [readpos＆posmask].

Declare into continuous memory array, the subunit sequence value that can quickly pass through array quickly reads and write number According to pointer；But when data writes and reads array afterbody, gesture is desired by the judgement of program, and this judges multi-thread Cheng Zhonghui causes performance cost, Gu present invention uses one kind can no performance cost, and data reading and write data After tail data, leap to data header；Done and computing by the value of the value of posmask and read and write position, to determine current The real position data pointer in data set space to be ready writing and read reads in dataptrarray and achieves tail head phase Connect, and no any unlocking, there is high performance.

Methods described also includes: during write data, using the lock bussing technique of intel, in the way of atom, judges simultaneously Whether writing position is empty, if it is empty, then writes, and if not empty then it represents that queue has been write full, now write thread is absorbed in Wait；When reading data, using the lock bussing technique of intel, in an atomic manner judge whether reading position is empty, if not Sky, then read and return, and reading position is reset, if it is empty then it represents that whole queue has all been read, now read line Journey is absorbed in wait；Reading thread-data is notified to can read when one data pointer of successful write, now, if reading thread For resting state, then will be waken up；Notify write thread-data writable when one data pointer of successful reading, now, If write thread is resting state, will be waken up.

Because reading and write thread being asynchronous, readpos and writepos there may be the situation of overlap (i.e. When data space be fully written or be space-time), now read and write be same position, be a critical zone, this feelings Condition must make data be protected, and compare traditional thread lock and have very high cost, and this method is with a kind of high side of performance Formula achieves the efficient read-write of critical zone.

In described memory headroom, with every 8 bytes for a unit.Memory headroom is with 8 bytes, i.e. a uint64_t For a unit, the value of the sizeof (void*) under any system can be deposited, that is, store a value (pointer of a pointer Up to one uint64_t).

Compared with prior art, the invention has the beneficial effects as follows: can solve the problem that data transfer between two threads, that is, to queue This critical resource of container realizes the access of no lock, and performance is greatly improved；Achieve critical in the way of a kind of performance is high The efficient read-write in area.

Specific embodiment

In order that the objects, technical solutions and advantages of the present invention become more apparent, with reference to embodiments, to the present invention It is further elaborated.It should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not used to limit Determine the present invention.

Any feature disclosed in this specification (including summary), unless specifically stated otherwise, all can be by other equivalent or tools The alternative features having similar purpose are replaced.I.e., unless specifically stated otherwise, each feature is a series of equivalent or similar characteristics One of example.

Specific embodiment 1

A kind of multi-thread data transmission method, method particularly includes:

During initialization system, apply for a contiguous memory space that can accommodate n pointer；Described n=2^m, m is more than or equal to 1 Integer；Each pointer position is a unit, and n pointer position forms an array space dataptrarray；Described finger Pin points to user data address；

Define two integer value: writepos and readpos；Wherein, writepos represents that data write thread to be currently ready writing Enter the position in described array space；Readpos represents that digital independent thread is current and to be ready reading data from described array space Position；Data writes successfully, then writepos value is from increasing；Digital independent success, then the increasing certainly of readpos value, will be successfully read The pointer value of array locus reset, after clearing, write thread finds that this position is zero again, can be normally written new User pointer data；

When but dataptrarray writes full data do not walked by timely reading, failure can be write；When in dataptrarray When all values are all zero, failure can be read；

Define an integer mask posmask, its value is n-1；Done and computing by the value of the value of posmask and read and write position, To determine the position in the current real array space being about to and writing and read, that is, presently written position should be Dataptrarray [writepos＆posmask], the position of reading is dataptrarray [readpos＆posmask].

Specific embodiment 2

On the basis of specific embodiment 1, methods described also includes: during write data, using the lock bussing technique of intel, with The mode of atom, judges whether writing position is empty simultaneously, if it is empty, then writes, if not empty then it represents that queue has been write completely , now write thread be absorbed in wait:

Void* dataptr is a data pointer needing to be written into queue,

do{

uint32_t writelocation = _sync_fetch_and_add(&writepos,1);

uint64_t* writeptr = (uint64_t*)dataptrarray[posmask&writelocation];

if(__sync_bool_compare_and_swap(writeptr,0,(uint64_t)dataptr))

{

notifyread();

return true;// write successfully

}

else

{

wait();

}

}

When reading data, using the lock bussing technique of intel, in an atomic manner judge whether reading position is empty, if not Sky, then read and return, and reading position is reset, if it is empty then it represents that whole queue has all been read, now read line Journey is absorbed in wait:

Uint64_t readptrvalue=0, is the value reading from queue, is returned to user,

uint32_t readlocaltion = __sync_fetch_and_add(&readpos,1);

uint64_t* readptr = (uint64_t*)dataptrarray[posmask&readlocaltion];

if((readptrvalue =__sync_lock_test_and_set(readptr,0)) != 0 )

{

notifywrite();

return readptrvalue;// read successfully

}

else

{

wait();

}

Reading thread-data is notified to can read when one data pointer of successful write, now, if reading thread is dormancy State, then will be waken up；Notify write thread-data writable when one data pointer of successful reading, now, if write Entering thread is resting state, then will be waken up.

Specific embodiment 3

On the basis of specific embodiment 1 or 2, in described memory headroom, with every 8 bytes for a unit.Memory headroom is with 8 Individual byte, that is, a uint64_t is a unit, can deposit the value of the sizeof (void*) under any system, that is, store The value (up to one uint64_t of a pointer) of one pointer.

Claims (3)

1. a kind of multi-thread data transmission method, method particularly includes:

During initialization system, apply for a contiguous memory space that can accommodate n pointer；Described n=2^m, m is more than or equal to 1 Integer；Each pointer position is a unit, and n pointer position forms an array space dataptrarray；Described pointer Point to user data address；

Define two integer value: writepos and readpos；Wherein, writepos represents that data write thread to be currently ready writing Enter the position in described array space；Readpos represents that digital independent thread is current and to be ready reading data from described array space Position；Data writes successfully, then writepos value is from increasing；Digital independent success, then the increasing certainly of readpos value, will be successfully read Array locus pointer value reset；

Define an integer mask posmask, its value is n-1；Done and computing by the value of the value of posmask and read and write position, To determine the position in the current real array space being about to and writing and read, that is, presently written position should be Dataptrarray [writepos＆posmask], the position of reading is dataptrarray [readpos＆posmask].

2. multi-thread data transmission method according to claim 1, methods described also includes: during write data, adopts The lock bussing technique of intel, in the way of atom, judges whether writing position is empty simultaneously, if it is empty, then writes, if not Empty then it represents that queue has been write full, now write thread is absorbed in wait；When reading data, using the lock bus skill of intel Art, in an atomic manner judge that whether reading position is empty, if not empty, then read and return, and reading position is reset, if For empty then it represents that whole queue has all been read, now read thread and be absorbed in wait；Write a data pointer when successful Shi Tongzhi reads thread-data and can read；Notify write thread-data writable when one data pointer of successful reading.

3. multi-thread data transmission method according to claim 1 and 2, in described memory headroom, with every 8 bytes for one Individual unit.