LD_PRELOAD override to add AT_STATX_DONT_SYNC to the statx syscall flags

gistfile1.txt

## Create 10000 files in an nfs directory 
$ dd if=/dev/zero bs=10000 count=1 | split -b1

## in another shell, write some data in the same directory
$ dd if=/dev/zero of=data bs=1M

## run stat with and without --cached option to spot the runtime difference
$ hyperfine --warmup 1 --runs 3 'find . -type f -print0 | xargs -0 stat --cached=always >/dev/null' 'find . -type f -print0 | xargs -0 stat >/dev/null'
Benchmark 1: find . -type f -print0 | xargs -0 stat --cached=always >/dev/null
  Time (mean ± σ):      1.233 s ±  0.119 s    [User: 0.064 s, System: 0.104 s]
  Range (min … max):    1.119 s …  1.357 s    3 runs
 
Benchmark 2: find . -type f -print0 | xargs -0 stat >/dev/null
  Time (mean ± σ):     11.099 s ±  4.458 s    [User: 0.062 s, System: 5.520 s]
  Range (min … max):    7.721 s … 16.152 s    3 runs
 
Summary
  find . -type f -print0 | xargs -0 stat --cached=always >/dev/null ran
    9.00 ± 3.72 times faster than find . -type f -print0 | xargs -0 stat >/dev/null

gistfile2.txt

## without the dd write in the same directory
$ hyperfine --warmup 1 --runs 10 'ls -l > /dev/null' 'LD_PRELOAD=/home/jbdenis/statx_override/statx_dont_sync_override.so ls -l >/dev/null'                                                                                  
Benchmark 1: ls -l > /dev/null
  Time (mean ± σ):     319.2 ms ± 365.9 ms    [User: 17.2 ms, System: 29.8 ms]
  Range (min … max):    35.4 ms … 853.4 ms    10 runs

Benchmark 2: LD_PRELOAD=/home/jbdenis/statx_override/statx_dont_sync_override.so ls -l >/dev/null
  Time (mean ± σ):      34.4 ms ±   0.7 ms    [User: 16.5 ms, System: 17.2 ms]
  Range (min … max):    33.1 ms …  35.7 ms    10 runs

Summary
  LD_PRELOAD=/home/jbdenis/statx_override/statx_dont_sync_override.so ls -l >/dev/null ran
    9.27 ± 10.63 times faster than ls -l > /dev/null

gistfile3.txt

## with the dd write in the same directory
$ hyperfine --warmup 1 --runs 10 'ls -l > /dev/null' 'LD_PRELOAD=/home/jbdenis/statx_override/statx_dont_sync_override.so ls -l >/dev/null'                                                                                  
Benchmark 1: ls -l > /dev/null
  Time (mean ± σ):      3.278 s ±  1.579 s    [User: 0.017 s, System: 1.987 s]
  Range (min … max):    2.169 s …  7.638 s    10 runs

Benchmark 2: LD_PRELOAD=/home/jbdenis/statx_override/statx_dont_sync_override.so ls -l >/dev/null
  Time (mean ± σ):      35.4 ms ±   1.5 ms    [User: 16.6 ms, System: 18.2 ms]
  Range (min … max):    33.9 ms …  38.4 ms    10 runs

Summary
  LD_PRELOAD=/home/jbdenis/statx_override/statx_dont_sync_override.so ls -l >/dev/null ran
   92.50 ± 44.70 times faster than ls -l > /dev/null

Makefile

CC=gcc
CFLAGS=-shared -fPIC -Wall
LIBS=-ldl
TARGET=statx_dont_sync_override.so
SRC=statx_dont_sync_override.c
 
all: $(TARGET)
 
$(TARGET): $(SRC)
    $(CC) $(CFLAGS) -o $(TARGET) $(SRC) $(LIBS)
 
test: $(TARGET)
    @LD_PRELOAD=./$(TARGET) ls -l
 
clean:
    rm -f $(TARGET)

statx_dont_sync_override.c

#define _GNU_SOURCE
#include <dlfcn.h>
#include <fcntl.h>
#include <sys/stat.h>
 
// Define the type of the original function
typedef int (*orig_statx_type)(int dirfd, const char *pathname, int flags,
                               unsigned int mask, struct statx *statxbuf);
 
 
int statx(int dirfd, const char *pathname, int flags, unsigned int mask,
          struct statx *statxbuf) {
    // Get a reference to the original function
    orig_statx_type orig_statx;
    orig_statx = dlsym(RTLD_NEXT, "statx");
 
    // Set STATX_DONT_SYNC flag and unset STATX_FORCE_SYNC
    // flags &= ~AT_STATX_FORCE_SYNC;
    flags |= AT_STATX_DONT_SYNC;
 
    // Call the original function with the modified flags
    return orig_statx(dirfd, pathname, flags, mask, statxbuf);
}