Re: git: 56ac3351c61f - main - witness: fix operation without WITNESS_SKIPSPIN

On Fri, Jul 17, 2026 at 09:02:36AM -0400, Mark Johnston wrote:
M> On Fri, Jul 17, 2026 at 01:40:13AM +0000, Gleb Smirnoff wrote:
M> > The branch main has been updated by glebius:
M> > 
M> > URL: https://cgit.FreeBSD.org/src/commit/?id=56ac3351c61f285ee5b8c9bd4411eeaeb20efe26
M> > 
M> > commit 56ac3351c61f285ee5b8c9bd4411eeaeb20efe26
M> > Author:     Gleb Smirnoff <[email protected]>
M> > AuthorDate: 2026-07-17 01:37:03 +0000
M> > Commit:     Gleb Smirnoff <[email protected]>
M> > CommitDate: 2026-07-17 01:37:03 +0000
M> > 
M> >     witness: fix operation without WITNESS_SKIPSPIN
M> >     
M> >     Since malloc(9) even with M_NOWAIT is forbidden when we hold a spinlock,
M> >     we can't print detailed lock tree as the operation tries to allocate
M> >     memory.
M> 
M> I think malloc() actually can't be used at all here.  It's possible to
M> recurse into the vmem boundary tag allocator; syzkaller triggered the
M> panic below once for me.
M> 
M> Can we preallocate a single tracker at boot time, protected by w_mtx?

I'd leave this to Jonathan's decision.

My personal take is that WITNESS is a helper for developers, not a production
feature.  If the code gives us a hint on lock order in 99% cases and recurses
deep in malloc() in 1% times it is good enough.  Of course if recursion can be
avioded easily, this should be done.

-- 
Gleb Smirnoff