On Mon 15 Jul 2002 at 13:09:43 +0200, Jaromir Dolecek wrote:
> BTW, can you try -current (1.6D) kernel without your stack size
> changes? There was a fix to kernel stack usage for IPsec committed
> not too long ago, and the fix has not been pulled up to 1.6 branch
> yet.

I tried it just now (source supped 15 july). The results are better than
with the unmodified 1.5ZC kernel: ESP ping traffic works now without
overflowing the interrupt stack. The NFS mount however still overflows
the kernel stack with the following backtrace:

bash-2.04# mount /vol1
panic: kernel stack invalid
Stopped in pid 170 (mount_nfs) at       trap+0x174:     tstl    64(r8)
db> tr
panic: kernel stack invalid
Stack traceback :
0x80336b34: trap+0x174(0x80336bb4)
0x80336bb4: trap type=0xf code=0x0 pc=0x8012a96c psl=0xcc0008
0x80336b80: SHA1Transform+0x15a(0x8aae94f4,0x8aae9510)
0x8aae9470: SHA1Update+0x66(0x8aae94f4,0x801904ac,0x200)
0x8aae94a8: rndpool_extract_data+0x5a(0x80190484,0x886822f0,0x8,0)
0x8aae9550: rnd_extract_data+0x23(0x886822f0,0x8,0)
0x8aae9590: key_randomfill+0x1a(0x886822f0,0x8)
0x8aae95bc: key_sa_stir_iv+0x25(0x88694380)
0x8aae95e8: esp_cbc_encrypt+0x457(0x80dc2200,0x14,0x68,0x88694380,0x801648a0,0x8
)
0x8aae9648: esp_output+0x60b(0x80dc2200,0x80dc22f5,0x81ecd100,0x88602a00,0x2)
0x8aae96c0: esp4_output+0x42(0x80dc2200,0x88602a00)
0x8aae9704: ipsec4_output+0x22f(0x8aae97a4,0x8868b240,0)
0x8aae9728: ip_output+0x637(0x81ecd100,0,0x81372384,0,0)
0x8aae97c8: udp_output+0x19a(0x80dc2900,0x81372364)
0x8aae980c: udp_usrreq+0x1ca(0x8139967c,0x9,0x80dc2900,0x805d3500,0,0x81f7871c)
0x8aae9844: sosend+0x4b3(0x8139967c,0x805d3500,0,0x80dc2900,0,0)
0x8aae98a8: nfs_send+0x87(0x8139967c,0x805d3500,0x80dc2900,0x88696300)
0x8aae98f0: nfs_request+0x2a6(0x80d10034,0x80dc2a00,0x13,0x81f7871c,0x8867ca00,0
x8aae99d4,0x8aae99d8,0x8aae99dc)
0x8aae997c: nfs_fsinfo+0x1e3(0x88602200,0x80d10034,0x8867ca00,0x81f7871c)
0x8aae99e8: nfs_bioread+0x80(0x80d10034,0x8aae9bb0,0,0x8867ca00,0x1)
0x8aae9acc: nfs_readdir+0x65(0x8aae9b50)
0x8aae9b1c: VOP_READDIR+0x4f(0x80d10034,0x8aae9bb0,0x8867ca00,0x8aae9b9c,0x8aae9
ba0,0x8aae9ba4)
0x8aae9b6c: nfs_cookieheuristic+0x6f(0x80d10034,0x886023c0,0x81f7871c,0x8867ca00
)
0x8aae9bd0: mountnfs+0x1f8(0x8aae9dc8,0x88685000,0x805d3500,0x8aae9d6c,0x8aae9d1
0,0x8aae9cc4,0x81f7871c)
0x8aae9c78: nfs_mount+0x135(0x88685000,0x7ffffca9,0x7ffffacc,0x8aae9e58,0x81f787
1c)
0x8aae9e10: sys_mount+0x2dc(0x81f7871c,0x8aae9f60,0x8aae9f58)
0x8aae9f14: syscall+0x10f(0x8aae9fb4)
db> 

I wonder how close the interrupt stack comes to overflowing - I don't
suppose there is an easy way to determine this. Perhaps I can pre-fill
it with known data and then examine it periodically. How do I break into
ddb from the serial console on a VAX? I find code in vax/vsa/lkc.c for a
graphics console (ctrl-alt-esc) - perhaps I'll need to add that too.

> Jaromir
-Olaf.
-- 
___ Olaf 'Rhialto' Seibert - rhialto@       -- Woe betide the one who feels
\X/ polderland.nl  -- remorse without sin - Tom Poes, "Het boze oog", 4444.