Messages in this thread | | | Date | Wed, 22 Sep 2010 15:19:36 -0400 | From | Mathieu Desnoyers <> | Subject | Re: [PATCH 2/2] Rewrite jump_label.c to use binary search |
| |
* Jason Baron (jbaron@redhat.com) wrote: > On Wed, Sep 22, 2010 at 11:02:50AM -0400, Mathieu Desnoyers wrote: > > * Andi Kleen (andi@firstfloor.org) wrote: > > > > > > > > > > >>> + for (; entry < stop && entry->key == key; entry++) > > > >>> + if (kernel_text_address(entry->code)) > > > >> > > > >> This does not work for modules I'm afraid, only for the core kernel. You > > > >> should test for __module_text_address() somewhere. > > > > > > > > I thought it was shared now, but ok. > > > > > > Double checked. This is ok because kernel_text_address() > > > already checks for modules. You were probably thinking > > > of __kernel_text_address() > > > > Ah right, > > > > Although we have another problem: > > > > __module_text_address() includes module init text, which defeats the > > purpose of the check put in there by Jason. > > > > So the check works for the core kernel, but not for modules. > > > > Mathieu > > > > it works for modules too...it does: > > struct module *__module_text_address(unsigned long addr) > { > struct module *mod = __module_address(addr); > if (mod) { > /* Make sure it's within the text section. */ > if (!within(addr, mod->module_init, mod->init_text_size) > && !within(addr, mod->module_core, > mod->core_text_size)) > mod = NULL; > } > return mod; > } > > and then in kernel/module.c we have : > > > module_free(mod, mod->module_init); > mod->module_init = NULL; > > > So, I was relying on the fact module_init gets set to NULL after the > free happens. However, there a small race there in that the vfree() > happens before module_init() is set to NULL. So that is probably most > easily fixed be wrapping those two lines with the jump_label_mutex.
It's both module_init = NULL _and_ init_text_size = 0 that make sure the test "within(addr, mod->module_init, mod->init_text_size)" is valid. Just the "module_init = NULL" can cause problems with addresses in the low range of kernel addresses. With a long enough module init section, the offset from NULL can end up (temporarily) in the kernel address range.
But this is all wrong: __module_text_address is relying on preempt_disable() to ensure coherency of this test is just racy, as you point out above. So we either do the RCU synchronization properly, or hold the module_mutex around the module text address test _and_ actual access to the module init section.
Thanks,
Mathieu
> > thanks, > > -Jason > >
-- Mathieu Desnoyers Operating System Efficiency R&D Consultant EfficiOS Inc. http://www.efficios.com
| |