From mboxrd@z Thu Jan 1 00:00:00 1970 From: James Simmons Date: Thu, 27 Feb 2020 16:17:52 -0500 Subject: [lustre-devel] [PATCH 604/622] lnet: fix small race in unloading klnd modules. In-Reply-To: <1582838290-17243-1-git-send-email-jsimmons@infradead.org> References: <1582838290-17243-1-git-send-email-jsimmons@infradead.org> Message-ID: <1582838290-17243-605-git-send-email-jsimmons@infradead.org> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: lustre-devel@lists.lustre.org From: Mr NeilBrown Reference counting of klnd modules is handled by the module itself. Currently, it is possible for a module to be completely unloaded between the time when the module called module_put(), and when it subsequently returns from the function that makes that call. During this time there may be one or two instructions to execute, and if the module is unmapped before they are executed, an exception will result. The module unload will call lnet_unregister_lnd() which takes the_lnet.ln_lnd_mutex, so module unload cannot complete while that is held. lnd_startup is called with this mutex held to avoid any races, but lnd_shutdown is not. Adding that protection will close the race. WC-bug-id: https://jira.whamcloud.com/browse/LU-12678 Lustre-commit: c087091cd901 ("LU-12678 lnet: fix small race in unloading klnd modules.") Signed-off-by: Mr NeilBrown Reviewed-on: https://review.whamcloud.com/36853 Reviewed-by: Serguei Smirnov Reviewed-by: James Simmons Reviewed-by: Oleg Drokin Signed-off-by: James Simmons --- net/lnet/lnet/api-ni.c | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/net/lnet/lnet/api-ni.c b/net/lnet/lnet/api-ni.c index 0ca8bef..5df39aa 100644 --- a/net/lnet/lnet/api-ni.c +++ b/net/lnet/lnet/api-ni.c @@ -1983,7 +1983,14 @@ static void lnet_push_target_fini(void) islo = ni->ni_net->net_lnd->lnd_type == LOLND; LASSERT(!in_interrupt()); + /* Holding the mutex makes it safe for lnd_shutdown + * to call module_put(). Module unload cannot finish + * until lnet_unregister_lnd() completes, and that + * requires the mutex. + */ + mutex_lock(&the_lnet.ln_lnd_mutex); net->net_lnd->lnd_shutdown(ni); + mutex_unlock(&the_lnet.ln_lnd_mutex); if (!islo) CDEBUG(D_LNI, "Removed LNI %s\n", -- 1.8.3.1