Red Hat Bugzilla – Bug 53647
Asynchronous reading/writing of files on an SMP system
Last modified: 2007-04-18 12:37:06 EDT
From Bugzilla Helper:
User-Agent: Mozilla/4.76 [en] (X11; U; Linux 2.2.16-22enterprise i686)
Description of problem:
One process is writing text lines to a file. Another process is reading
those lines as they are written. Using 2.2.x kernels and/or 2.4.x
uniprocessor kernels, the reading program always reads complete lines.
Using a 2.4.x SMP kernel, the reading process sometimes reads partial
This happens regardless of the i/o type (stream or binary) used., and
mandatory file locking does not appear to help.
Version-Release number of selected component (if applicable):
Steps to Reproduce:
1. The writing program loops: opens the file for append, writes a single
complete line of text, and then closes the file. The reading program: opens
the file for input, then loops reading from the file. The lines are a
fixed size (80 chars) terminated with a newline.
2. A modified version of the writing program tries to use mandatory file
locking to protect the write operation, without success.
Actual Results: Under 2.2 kernels, or 2.4.x uniprocessor, the reading
program always reads a complete line.
Under an SMP 2.4.x kernel, the reading program usually reads a complete
line, but occasionally (depending on the speed of both programs) reads only
a partial line.
Expected Results: The expected behavior without mandatory file locking
does not appear to be documented, but it should be consistent between uni-
and multi-processor kernels.
Using mandatory file locking in the writing program should insure that the
reading program always reads complete lines.
This has been tried with the version 2.4.2-2 kernel on the retail 7.1 CDs,
and also with the updated 2.4.3-12 kernel.
Test programs for both the reading and writing side are attached (simread.c
and simtst.c). In actual use, the reading program is a large commercial
application ; we have no access to the sources. The writing program was
written locally; we have the sources and can make any useful changes to it.
Created attachment 31776 [details]
reader program (self explanatory)
Created attachment 31777 [details]
writer program (self explanatory)
I hope you still have warranty on the big program; read() and write() are ALWAYS
allowed to make partial progress. and always have been. (posix specification)
Jakub: just to make sure, could you check the program as well ? The program
doesn't check read bytes so that looks buggy to me.
The behaviour on SMP _is_ the same as on UP. Neither UP nor SMP makes the
gaurantee that writes are atomic with respect to reads. If you are running on
an otherwise quiescent system so that the write system call happens to be able
to proceed without having to do any IO or hitting a kernel lock which is already
taken, then on UP you'll be lucky and finish the write before the read syscall
can be scheduled. However, on a busy system, even UP will show non-atomic writes.
This is correct behaviour: neither POSIX nor the SingleUnix specification
require writes to be atomic, except for the special case of small writes to pipes.
Mandatory locking will not fix your problem. Taking a mandatory lock in the
write path won't necessarily wait for an already-running read to complete: it
only prevents new reads from beginning. You'd need cooperative locking between
both processes to make the required atomicity guarantees.
If you have no access to the source for the reading application, then the only
solution I can think of is to write a small library stub wrapper around the read
syscall and LD_PRELOAD that.