Issue
Consider a file, that is edited with a frequency of tens, hundreds multiple processes per second. Since two or more processes can race for a file access for writing, there need a mechanism to be implemented, to make only one process to access file at one time.
As I understand, calling fopen
(or open
), until the fclose
will do the job - these functions guarantee, that only one process will access the file.
The problem is that the file is needed to be truncated after being opened, because there is a need to read it first and after that rewrite. If there will be two fopen
calls, obviously it will not guarantee cross-process safety.
This answer recommends to use freopen
function after fopen
and before fclose
calls, however, according to the Linux man-pages, freopen
documentation:
If pathname is not a null pointer, freopen() shall close any file descriptor associated with stream.
Currently, the only solution I see is to create an associated file, with the needed to be accessed, and lock it, instead of locking the needed file (although it also be locked for sure), for the period, needed file will be read, closed, truncated, wrote, closed.
Is at least such solution will guarantee safety? Any better solutions?
As users in comments explained me, fopen
does not prevent a file from being opened in another process at all. Instead, flock
should be called additionally.
So the question is can I lock the file, read, truncate, write and then unlock it, or I should use solution with associated file, I described above?
Specific summary question
- As I understand,
flock
accepts file descriptor need to be locked as an argument (and exactly descriptor fromopen
, notfopen
). - But, as I understand, read, truncation and write file requires opening, closing and again opening the file.
- So, according to second fact, if I lock the file before I call
close
, will it be locked afterclose
call? If it automatically unlocks due toclose
it does not make any sense. If it is not unlocks, how to unlock it after all manipulations?
Solution
You write,
Consider a file, that is edited with a frequency of tens, hundreds multiple processes per second. Since two or more processes can race for a file access for writing, there need a mechanism to be implemented, to make only one process to access file at one time.
and you clarify that
It should be strictly
P1read-P1truncate-P1write -> P2read-P2truncate-P2write
, notP1read-P2read-P1truncate-P1write-P2truncate-P2write
That is, you want to serialize units of read(+truncate)+write access to the file.
As I understand, calling
fopen
(oropen
), until thefclose
will do the job - these functions guarantee, that only one process will access the file.
Absolutely not. Neither C nor POSIX makes any such guarantee. In practice, it is not typically true on POSIX systems, and it is not true on Linux in particular. You might see such behavior on Windows, however, which could be where you got the idea.
If you must use a physical file for this purpose then you will need to perform some sort of locking around access to it. Provided that the file is on a local filesystem, the most natural locking mechanism would be flock()
. This implements an advisory locking system, meaning that it applies only to processes that explicitly gate their access via flock()
, but that does not appear to be an issue for your planned use. flock()
has the advantage that if a process dies while holding the lock (with the result that the underlying open file description is closed) then the lock is automatically released. flock()
has the disadvantage that it will not work if the file is replaced -- only if processes always modify it in place.
Another alternative would be a process-shared mutex or semaphore or a named semaphore, which the processes would use (again, cooperatively) to ensure the needed serialization. This has the advantage that it could still work if the file is altogether replaced instead of modified in-place, but the possible disadvantage that if one of the participating processes dies while holding the mutex / semaphore locked then the whole system gets jammed. (But if you use System V semaphores instead of POSIX semaphores then that shortcoming can be addressed.)
But what if you don't use a physical file? What you're describing sounds like it will require moderately high I/O bandwidth, and it will do a great deal of rewriting the same sectors of the underlying physical medium. Why do you want to hit a disk so hard? Or to deal with slow file I/O? You probably would be better off using shared memory (still protected by mutex or semaphore). Or you could implement a data broker service that runs on the system and mediates access to the data.
Addendum - response to additional questions about flock
- As I understand,
flock
accepts file descriptor need to be locked as an argument (and exactly descriptor fromopen
, notfopen
).
Sort of. flock()
requires a file descriptor (from open
) to identify the file to lock. The file is locked, not the file descriptor, but that lock is associated with the file descriptor.
And if you want to use stdio
functions to read and write the file, then you can get a FILE *
from the file descriptor via fdopen()
. The file descriptor is still underneath, so that will not interfere with the locking. If you do this, however, then you should be sure to close via fclose()
ing the stream from fdopen
, not directly close()
ing the file descriptor underneath.
- But, as I understand, read, truncation and write file requires opening, closing and again opening the file.
That is incorrect.
If the data to be written are the same length then you don't need to explicitly truncate at all. After reading the file, rewind()
ing to the beginning or fseek()
ing or lseek()
ing there, then writing, then closing will have the desired effect.
If the data to be written might be shorter then you can truncate()
the file to any length (including 0) without first closing it.
- So, according to second fact, if I lock the file before I call close, will it be locked after close call?
No, and it doesn't need to be (see above). Moreover, it is a convenience that you just need to close the file, not explicitly unlock it.
Outline
// ...
// open the file for reading and writing, creating it if necessary
int fd = open(path_to_file, O_RDWR | O_CREAT, 0600);
// acquire an exclusive lock on the file, blocking if necessary
int result = flock(fd, LOCK_EX);
// Wrap the file descriptor in a stream
FILE *file = fdopen(fd, "r+");
// ... read the file ...
// Rewind to the beginning
rewind(file);
// Truncate if necessary:
result = truncate(fd, 0);
// ... write new data ...
// close the file and underlying file descriptor, flushing any
// buffered output to it. The lock is hereby released.
result = fclose(file);
// done
Note that for clarity and simplicity, I have omitted testing for or handling errors from the various functions. Production code does not have that luxury.
Answered By - John Bollinger Answer Checked By - Candace Johnson (WPSolving Volunteer)