On Thu, Mar 11, 2021 at 12:36:51PM +0100, Peter Zijlstra wrote: > On Thu, Mar 11, 2021 at 12:09:15PM +0100, Michal Hocko wrote: > > > Sorry for being dense but I do not follow. You have provided the > > following example > > spin_lock(&A); > > <IRQ> > > spin_lock(&A); > > > > if A == hugetlb_lock then we should never reenter with > > free_huge_page > > What I'm saying is that if irq_disabled(), the that interrupt cannot > happen, so the second spin_lock cannot happen, so the deadlock cannot > happen. > > So: '!irqs_disabled() && in_atomic()' is sufficient to avoid the IRQ > recursion deadlock. > > Also, Linus hates constructs like this: > > > https://lkml.kernel.org/r/CAHk-=wht7kaeyr5xew2orj7m0hibvxz3t+2ie8vnhlqfdbn...@mail.gmail.com
To be fair, later in that same thread Linus states that his main concern is not core code, but rather driver code: https://lore.kernel.org/lkml/CAHk-=wjsmycgmhjrcmeetr3r+k5bpsrtmvwfd8iaoqcyd_v...@mail.gmail.com/ Nevertheless, if the job can be done reasonably without checking the preemption/interrupt state, why not? And Mike's patch is still useful for people hitting this bug. Thanx, Paul > > From the code simplicity POV (and hugetlb has grown a lot of complexity) > > it would be really easiest to make sure __free_huge_page to be called > > from a non-atomic process context. There are few ways to do that > > - defer each call to a WQ - user visible which sucks > > - defer from atomic or otherwise non-sleeping contextx - requires > > reliable in_atomic AFAICS > > - defer sleeping operations - makes the code flow more complex and it > > would be again user visible in some cases. > > > > So I would say we are in "pick your own poison" kind of situation. > > Just to be clear: > > NAK on this patch and any and all ductape crap. Fix it properly, make > hugetlb_lock, spool->lock IRQ-safe, move the workqueue into the CMA > thing. > > The code really doesn't look _that_ complicated.
