git: 1ae20f7c70ea - main - kern: malloc: fix panic on M_WAITOK during THREAD_NO_SLEEPING()

[John Baldwin]

On 3/10/21 3:57 AM, Mateusz Guzik wrote:
> There is something very wrong going on here.
> 
> Key thing to note is that malloc is ultimately a wrapper around
> uma_zalloc. Whatever asserts you may want to add to malloc to catch
> problems sooner, should also present in uma.
> 
> uma has the following:
> 
>          if (flags & M_WAITOK) {
>                  WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
>                      "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
>          }

This warns about holding locks, not about TD_NO_SLEEPING.  Witness' role
is to check lock orders and warn about invalid operations when holding
certain locks.  It does not currently verify anything about other inhibitions
like TD_NO_SLEEPING.  See, e.g. the list of assertions in userret() which
includes a WITNESS_WARN in addition to several other checks (including
TD_NO_SLEEPING).  Arguably we should just move the TD_NO_SLEEPING check
from malloc() to here along with the td_intr_nesting_level.  Any case
where you can't use malloc() with M_WAITOK you can't use uma with M_WAITOK
either.

> This code used to execute prior to this commit and fail to catch the
> problems which got reported already.
> 
> In other words:
> - WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK is incomplete in terms of
> internals its internals
> - the above should be in malloc, perhaps after being abstracted into a
> an assert-helper
> - fixing witness deficiency will probably find a slew of new problems
> - this should remain as a warning (maybe rate-limited) for the foreseable future

I don't know that we've had so many issues that we can't just fix them on
HEAD right now vs having to make this a warning instead of keeping it as
a panic.

-- 
John Baldwin