nfsd: move init of percpu reply_cache_stats counters back to nfsd_init_net

In the Linux kernel, the following vulnerability has been resolved:

nfsd: move init of percpu reply_cache_stats counters back to nfsd_init_net

Commit f5f9d4a314da ("nfsd: move reply cache initialization into nfsd

startup") moved the initialization of the reply cache into nfsd startup,

but didn't account for the stats counters, which can be accessed before

nfsd is ever started. The result can be a NULL pointer dereference when

someone accesses /proc/fs/nfsd/reply_cache_stats while nfsd is still

shut down.

This is a regression and a user-triggerable oops in the right situation:

- non-x86_64 arch

- /proc/fs/nfsd is mounted in the namespace

- nfsd is not started in the namespace

- unprivileged user calls "cat /proc/fs/nfsd/reply_cache_stats"

Although this is easy to trigger on some arches (like aarch64), on

x86_64, calling this_cpu_ptr(NULL) evidently returns a pointer to the

fixed_percpu_data. That struct looks just enough like a newly

initialized percpu var to allow nfsd_reply_cache_stats_show to access

it without Oopsing.

Move the initialization of the per-net+per-cpu reply-cache counters

back into nfsd_init_net, while leaving the rest of the reply cache

allocations to be done at nfsd startup time.

Kudos to Eirik who did most of the legwork to track this down.