nfsd open file limits

Recently I came across a user who requested an increase to the ulimit settings for nfsd kernel processes.

root      1122  0.0  0.0      0     0 ?        S    11:43   0:00 [nfsd]

# grep 'open file' /proc/1122/limits
Max open files            1024                 4096                 files

This appears to default to 1024/4096 soft/hard.

As you can see from the brackets surrounding nfsd, this is a kernel process spawned from kthreadd and thus won’t inherit limits from systemd (or limits.conf)

I decided to throw together a quick C++ program proving that these limits do not impact how many open files a client can utilize.

#include <iostream>
#include <fstream>
#include <dirent.h>
#include <chrono>
#include <thread>
#include<unistd.h>

using namespace std;
int main() {
        DIR *dir;
        struct dirent *entry;
        string filename;

        dir = opendir(path);
        std::fstream fs[8194];
        int count = 0;

        chdir("/export");

        while ((entry = readdir(dir)) != NULL) {
          printf("  %s\n", entry->d_name);
          fs[count].open(entry->d_name);
          count++;
        }
        std::this_thread::sleep_for(std::chrono::milliseconds(100000));
        closedir(dir);
        return 0;
}

On the NFS server in question, I created 8192 files.

[root@nfs export]# for x in {1..8192}; do touch $x; done

I also ensured that only 1 [nfsd] thread was running (to rule out the open files being split between multiple nfsd threads).

On the client I made sure the user had appropriate ulimit settings

# ulimit -n
9000

Then I ran the above program to hold open all 8192 files. As you can see below, there was no problem doing so.

# lsof +D /export/ | wc -l
8191

Tested with NFSv3 (with lockd) and NFSv4.

Conclusion: The [nfsd] limits shown in /proc has no impact on the nfs clients.