On Wed, Sep 27, 2006 at 03:53:55PM -0400, Dave Johnson wrote: > I'm running into an odd problem with the DBE exception handler. There is a fundamental problem with the way unmaskable exceptions other than cache errors and NMI are handled. This is the disassembly of the kernel's exception entry path starting at the general exception vector: a8000000003bae20 <except_vec3_generic>: a8000000003bae20: 401b6800 mfc0 k1,$13 a8000000003bae24: 337b007c andi k1,k1,0x7c a8000000003bae28: 001bd878 dsll k1,k1,0x1 a8000000003bae2c: 3c1a003f lui k0,0x3f a8000000003bae30: 035bd02d daddu k0,k0,k1 a8000000003bae34: df5ad2d8 ld k0,-11560(k0) a8000000003bae38: 03400008 jr k0 a8000000003bae3c: 00000000 nop [...] A few types of exceptions will be handled by just using $k0 and $k1; must will save the registers right away: [...] a800000000020440: 401a6000 mfc0 k0,$12 a800000000020444: 001ad0c0 sll k0,k0,0x3 a800000000020448: 0740000a bltz k0,0xa800000000020474 a80000000002044c: 03a0d82d move k1,sp a800000000020450: 403b2000 dmfc0 k1,$4 a800000000020454: 3c1aa800 lui k0,0xa800 a800000000020458: 001bddfa dsrl k1,k1,0x17 a80000000002045c: 675a0000 daddiu k0,k0,0 a800000000020460: 001ad438 dsll k0,k0,0x10 a800000000020464: 675a0043 daddiu k0,k0,67 a800000000020468: 001ad438 dsll k0,k0,0x10 a80000000002046c: 037ad82d daddu k1,k1,k0 a800000000020470: df7bf008 ld k1,-4088(k1) a800000000020474: 03a0d02d move k0,sp a800000000020478: 677dfed0 daddiu sp,k1,-304 a80000000002047c: ffba00e8 sd k0,232(sp) (c0_status.exl is cleared a mile further down) [...] If we take a DBE exception in this code we're in trouble and I've seen systems delivering DBEs highly asynchronously. Afar the Broadcom SOCs fall into that class. So the interesting part is if we take a data bus exception between the stack pointer adjustment and and before EXL is cleared. We're taking a nested exception so c0_epc and c0_cause.bd will not be updated. So when the bus error handler will save the $sp value it saw on entry but will return to the EPC of the first exception, that is only one stack frame will be popped. Whops ... Ralf
