NFS client regression, simple test program

[Jakob Oestergaard <jakob@unthought.net>]

[-- Attachment #1: Type: text/plain, Size: 455 bytes --]


Attached is a small test program that exposes the regression.

Kernel LEADING_EMPTY_SPACE wall-clock time to run
2.6.15    0                0.2s
2.6.15    1                29.3s
2.6.14.7  0                0.2s
2.6.14.7  1                0.2s

So it seems that if a file is repeatedly read from and written to, it
matters a lot for later kernels whether or not the blocks that are
worked on, are offset a single byte or not...

Any ideas?

-- 

 / jakob


[-- Attachment #2: nfsbench.c --]
[-- Type: text/x-csrc, Size: 1227 bytes --]

/*
 * Tiny benchmark for Linux NFS client problem 2.6.14.7->2.6.15
 *
 * By Jakob Oestergaard, joe@unthought.net / joe@evalesco.com
 *
 * Compile as:
 *  gcc -o nfsbench -O3 -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 nfsbench.c
 */

#include <unistd.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <fcntl.h>
#include <stdint.h>

#define MYBLOCKSIZE 32768
#define MYSUBSIZE 1024

/* Set this define to non-zero to expose the regression */
#define LEADING_EMPTY_SPACE 1

int main(int argc, char **argv)
{
  int testfd = open("testfile", O_RDWR|O_CREAT|O_TRUNC, 0666);
  off_t blockoff;
  uint8_t garbage[MYBLOCKSIZE];

  memset(garbage, 42, sizeof garbage);

  for (blockoff = LEADING_EMPTY_SPACE;
       blockoff < 10 * 1024 * 1024;
       blockoff += MYBLOCKSIZE) {
    off_t inblockoff;
    lseek(testfd, blockoff, SEEK_SET);
    write(testfd, garbage, MYBLOCKSIZE);

    for (inblockoff = MYBLOCKSIZE / 4;
	 inblockoff < MYBLOCKSIZE * 3 / 4;
	 inblockoff += MYSUBSIZE * 2) {
      uint8_t gdata[MYBLOCKSIZE];
      lseek(testfd, blockoff, SEEK_SET);
      read(testfd, gdata, MYBLOCKSIZE);
      lseek(testfd, -inblockoff, SEEK_CUR);
      write(testfd, gdata + inblockoff, MYSUBSIZE);
    }
  }

  close(testfd);
}