On Tue, Apr 07, 2020 at 02:01:01PM -0700, Linus Torvalds wrote: > On Tue, Apr 7, 2020 at 1:45 PM Waiman Long <longman@xxxxxxxxxx> wrote: > > > > If the memory is really virtually mapped, the only way to find out the > > size of the object is to use find_vm_area() which can be relatively high > > cost and no simple helper function is available. > > We _could_ just push it down to a "vfree_sensitive()", and do it > inside the vfree logic. That ends up obviously figuring out the size > of the area eventually. > > But since the vmalloc data structures fundamentally aren't irq-safe, > vfree() actually has magical things like "if called in an interrupt, > we'll delay it to work context". > Just some thoughts. Sorry for jumping in. Seems like there is only one place where we can "sleep". I mean when we call vfree(). That is free_vmap_area_noflush() -> try_purge_vmap_area_lazy(). Basically try_purge_vmap_area_lazy() can call the schedule() what is not allowed for IRQs. Instead of inlining the try_purge_vmap_area_lazy() into current context we can schedule_work(). And i think it makes sense from many point of views. Also, we can end up in zeroed non-existance vmap area if we do not find_vmap_area(). Thanks! -- Vlad Rezki
