diff options
author | Timo Teräs <timo.teras@iki.fi> | 2010-05-12 10:32:48 +0300 |
---|---|---|
committer | Denys Vlasenko <vda.linux@googlemail.com> | 2010-08-28 22:15:56 +0200 |
commit | d6d1bd2bbd2d4445b8e3cb1cb9d8d1aad817ba7f (patch) | |
tree | caf7ff18f20131ed87c6ebf2a0d98ad705353d34 /libc/sysdeps/linux/common/sched_yield.c | |
parent | 5a24d88fbb6ff820ac0d0ccd316efe02ef268c9d (diff) |
Fix resolver broken in NPTL build
The combination of
commit aab4df0fb51660300559f5f29290709db2f7bfee
resolv.c: add support for per thread res_state
commit cca45baf8353d1e338d232f5bdb2d1d6b357f1da
/etc/resolv.conf: support "timeout:n" and "attempts:n" options
.. and NPTL results in broken resolver in very annoying ways.
Now, it seems that most of the uclibc code does not work well
if res_state is TLS variable. Technically, this is the correct thing
to do since this gives proper per-thread resolving behavior, and
it also makes the config options overridable per thread. This
probably what apps expect as glibc does it too.
But alas, most places use _res to sync up static global variables
which results in breakage. It gets more or less randomly selected
which threads options get applied. Also in case of multiple servers
it looks like the retry logic is shared between all threads, e.g.
two concurrent resolutions can make other resolvers skip nameservers
due to shared "last_ns_num".
And finally the timeout/attempts commit breaks the accumulated stuff
horribly. What happens is:
1. multithreaded application startups, initializes resolver,
resolves things just fine
2. resolv.conf gets changed, application calls res_init
after res_init uclibc will call res_sync on all resolver functions
to refresh globals from the TLS variable _res
3. res_init was called only in one thread, so other thread's
_res contains all zeroes (yes, this is correct app usage:
res_init should be called only from one thread)
4. threads not calling res_init get broken resolver due to timeout
being set to zero
Now, one proper solution would be to:
1. make __open_nameservers return the configuration options
2. pass the config options struct to res_sync_func so it can do
the proper overrides from per-thread _res
3. remove the related globals and use locally config options from
__open_nameservers
But technically, the correct thing (as in glibc does this) is:
- res_init increases global "res_init timestamp"
- use _res (or pointer within there) for resolver retries etc.
get proper per-thread behavior
- resolvers functions call "maybe_init" which reinitialize the
TLS'ed resolver state (e.g. reload resolv.conf) if their res_init
timestamp is out dated
As an immediate emergency kludge fix, the following might do:
resolv: fix options handling for TLS _res
If _res is in TLS (NPTL), it might not get initialized. Assume
zeroes mean default values.
Signed-off-by: Timo Teräs <timo.teras@iki.fi>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Diffstat (limited to 'libc/sysdeps/linux/common/sched_yield.c')
0 files changed, 0 insertions, 0 deletions