On 4/24/18 2:17 AM, Alexey Dobriyan wrote:
> On Mon, Apr 23, 2018 at 10:21:01PM -0400, jeffm@xxxxxxxx wrote:
>> Memory pressure isn't really an issue on this machine, so we
>> end up using well over 100GB for proc files.
>
> Text files at scale!
>
>> With these patches applied, running the same testcase, the proc_inode
>> cache only gets to about 600k objects, which is about 99.7% fewer. I
>> get that procfs isn't supposed to be scalable, but this is kind of
>> extreme. :)>
> Easy stuff:
> * all ->get_link hooks are broken in RCU lookup (use GFP_KERNEL),
It's a pretty common pattern in the kernel, but it's just as easy to set
inode->i_link during instantiation and keep RCU lookup. There aren't so
many of these to make it a real burden on memory.
> * "%.*s" for dentry names is probably unnecessary,
> they're always NUL terminated
Ack.
> * kasprintf() does printing twice, since we're kind of care about /proc
> performance, allocate for the worst case.
Ack, integrated with ->get_link fix.
> * "int nlinks = nlink_tgid;"
> Unsigned police.
Ack. nlink_t{,g}id are both u8, but it's easy to make it consistent.
> * (inode->i_mode & S_IFLNK)
> this is sketchy, S_ISLNK exists.
>
Ack.
Notes of my own:
proc_task_count_links also had the logic backward. It would add an
extra link to the count for the symlink rather than the dir.
proc_pid_files_revalidate only needs to check if the tasks share files
since it won't be called if it's not a symlink.
Thanks for the review,
-Jeff
--
Jeff Mahoney
SUSE Labs
