tmpfs under i386 PAE kernel swamps pgdaemon?

["John D. Baker" <jdbaker%consolidated.net@localhost>]

As my usual i386 build hosts have failed in some way (won't power up),
I've arranged to netboot my primary amd64 build host with a netbsd-9/i386
install.  I'm using a GENERIC_PAE kernel (includes GENERIC and enables
options PAE) so I can use all 12GB the machine has.  Mainly, I have a
"/tmp" on tmpfs and put the pkgsrc ${WRKOBJDIR} there.

When running native amd64, if tmpfs /tmp starts getting too full, I get
"no space left on device" errors, like a physical disk.

When running a PAE i386 kernel, the machine just grinds to a halt.  I
can usually interrupt the build and run 'top' a couple of times before
it wedges completely.  When I do, I see the 'pgdaemon' thread of the
"[system]" process using 99%+ of the CPU.

One data point is that on amd64, I can extract "www/firefox" to tmpfs
"/tmp" and build it without issue while on i386 PAE, the extraction
phase hangs the machine as described above, so I had to move that one
to physical disk.

When the machine is hung, I can usually recover by dropping to DDB at
the console and rebooting (if the displayport monitor was turned off,
there will be no signal when turned back on, so I have to do that step
blind).  Sometimes, however, I have to force a power cycle and clean up
(raidframeparity, fsck) when it reboots.

Depending on the time of day (or night), I see lots of 'cron' processes
stuck in "tstile" and/or shell instances in "wait".

(Also, I have to disable i915drmkms to boot a PAE kernel or it halts with
a kernel error...)

Hmm. PAE has 2^40 addressing.  Is that physical or virtual?  Perhaps
I'm running out of virtual space before reaching the physical limits?

-- 
|/"\ John D. Baker, KN5UKS               NetBSD     Darwin/MacOS X
|\ / jdbaker[snail]consolidated[flyspeck]net  OpenBSD            FreeBSD
| X  No HTML/proprietary data in email.   BSD just sits there and works!
|/ \ GPGkeyID:  D703 4A7E 479F 63F8 D3F4  BD99 9572 8F23 E4AD 1645