PUFFS deadlocks when memory is low (was: Re: [PATCH] Re: zero-filed page on VOP_PUTPAGES)

[manu%netbsd.org@localhost (Emmanuel Dreyfus)]

Emmanuel Dreyfus <manu%netbsd.org@localhost> wrote:

> Keeping MAX(rvap->va_size, vp->v_size) in puffs_vnop_getattr() is simple
> but it has a drawback with distributed filesystems: a client will not
> notify that an open file was shrunk by another client. I think this is
> better than deadlocks and data corruption. 
> 
> If nobody has a better idea, I will got with this simple fix.

Again, I spoke too fast: even with this simplier fix, it still crashes
under low memory.

The reason is ioflush sleeping on a VOP_FSYNC to a PUFFS filesystem.
Here is its backtrace:

sleepq_block
cv_wait_sig
puffs_msg_wait
puffs_msg_wait2
dosetattr
flushvncache
puffs_vnop_fsync
VOP_FSYNC
pageflush
puffs_vfsop_sync
sync_fsync
VOP_FSYNC
sched_sync

At the same time, the filesystem server (perfused) awaits memory to be
freed, and here is our deadlock.
sleepq_block
mtsleep
uvm_wait
uvm_km_alloc
kmem_backend_alloc
vmem_xalloc
vmem_alloc
kmem_alloc
putter_fop_write
dofilewrite
sys_write

So we have a completely different bug that the story of zeroed chunks of
data: a PUFFS filesystem will cause system deadlock when memory is low,
on netbsd-5. Before I start running tests, anyone knows if improvements
have been done in -current on that front?

-- 
Emmanuel Dreyfus
http://hcpnet.free.fr/pubz
manu%netbsd.org@localhost