> The flags field in struct nfs_inode is protected by the BKL. The > following two code paths (there may be more, but my test program only > hits these two) modify the flags without obtaining the lock: > > nfs_end_data_update > nfs_release > nfs_file_release > __fput > fput > filp_close > sys_close > syscall_call > > nfs_revalidate_mapping > nfs_file_write > do_sync_write > vfs_write > sys_write > syscall_call > > Running multiple instances of a simple program [1] that opens, writes > to, and closes NFS mounted files eventually results in the programs > hanging on an SMP system (see kernel .config [3]). > > I've been testing this with 100 instances of the program: > $ ./breaknfs 100 & > > Usually within 10 minutes, all instances of breaknfs will hang. They > disappear from the output of 'top' and there is no NFS > activity between > the client and server.
[ sysrq output snipped... ] > I've reproduced this bug on 2.6.11.10, 2.6.12-mm2, and 2.6.13-rc2. > > With my patch against 2.6.13-rc2 below, I ran 100 instances > of breaknfs > with this patch for 14 hours and I was unable to get the > client to hang. i agree this is a problem. but instead of using heavyweight synchronization, why not convert the NFS_INO flags into atomic bitops? i have a patch that does that; would need to be ported to the latest kernels and tested to see if it addresses the problem. nick, are you interested in trying it out? - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
