2.6 mca_asm.S VA to PA mappings
2.6.0-test5-ia64-030908, arch/ia64/kernel/mca_asm.S, ia64_os_mca_done_dump sets r2 to ia64_mca_bspstore then does DATA_VA_TO_PA(r2) which uses tpa. This is before MCA has checked for a tlb error. Is it safe to use tpa before checking that the tlb data is valid or should it be using LOAD_PHYSICAL?
Before calling the C code, MCA uses VIRTUAL_MODE_ENTER which in turn calls DATA_PA_TO_VA on ar.bspstore. DATA_PA_TO_VA sets bits 61:63 to 1, which assumes that ar.bspstore is identity mapped. On a non-identity mapped 2.4 kernel I have seen ar.bspstore go
ia64_mca_bspstore (V=0xe000000004d5a6d0)DATA_VA_TO_PA (P=0x0000013014d5a6d0)DATA_PA_TO_VA (V=0xe000013014d5a6d0)
With the 2.4 alt.dtlb handler that still works, although it is confusing to C code which needs to access the RSE data in ia64_mca_bspstore. I doubt that DATA_PA_TO_VA will work if the kernel is in region 5.
Quoted 6 lines
2.6.0-test5-ia64-030908,
arch/ia64/kernel/mca_asm.S,ia64_os_mca_done_dumpsets r2 toia64_mca_bspstore
then doesDATA_VA_TO_PA(r2)> which uses tpa.
This is before MCA has checked for a tlb error. Is it safe to use tpa before checking that the tlb data is valid or should it be usingLOAD_PHYSICAL?
You are right ... if there is a TLB error, then the "tpa" may fail, so LOAD_PHYSICAL should be used here.
Quoted 5 lines
Before calling the C code, MCA uses
VIRTUAL_MODE_ENTERwhich in turn callsDATA_PA_TO_VAon ar.bspstore.DATA_PA_TO_VAsets bits 61:63 to 1, which assumes that ar.bspstore is identity mapped. On a non-identity mapped 2.4 kernel I have seen ar.bspstore go
Quoted 8 lines
ia64_mca_bspstore(V=0xe000000004d5a6d0)DATA_VA_TO_PA(P=0x0000013014d5a6d0)DATA_PA_TO_VA(V=0xe000013014d5a6d0)With the 2.4 alt.dtlb handler that still works, although it is confusing to C code which needs to access the RSE data in
ia64_mca_bspstore. I doubt thatDATA_PA_TO_VAwill work if the kernel is in region 5.
Same thing, we need LOAD_PHYSICAL there too. By the time C-code is called, the TLB will have been repaired (all ITR/DTR entries reloaded) so the C-code can happily use the ia64_mca_bspstore virtual address.
The new version of the MCA TLB fixup code should handle both of these ... if I can just get it working ... I upgraded my test machine on Friday, and now it generates MCAs all the time, not just when I inject them :-(
-Tony
Quoted 10 lines · Tony
Quoted 8 lines
ia64_mca_bspstore(V=0xe000000004d5a6d0)DATA_VA_TO_PA(P=0x0000013014d5a6d0)DATA_PA_TO_VA(V=0xe000013014d5a6d0)With the 2.4 alt.dtlb handler that still works, although it is confusing to C code which needs to access the RSE data in
ia64_mca_bspstore. I doubt thatDATA_PA_TO_VAwill work if the kernel is in region 5.Same thing, we need
LOAD_PHYSICALthere too.
Slight confusion. DATA_VA_TO_PA needs to be replaced by LOAD_PHYSICAL, but what replaces DATA_PA_TO_VA in a non-identity mapped kernel?
bspstore needs a valid virtual address before calling the C code. If the asm code was told the original virtual address (ia64_mca_bspstore) then DATA_PA_TO_VA could convert ar.bspstore from P to V. Without that hint, I see no way of mapping P to V for an arbitrary address held in a register.
Quoted 3 lines
Slight confusion.
DATA_VA_TO_PAneeds to be replaced byLOAD_PHYSICAL, but what replacesDATA_PA_TO_VAin a non-identity mapped kernel?
Good question! I'll start looking at that.
-Tony
Quoted 4 lines · Tony
The new version of the MCA TLB fixup code should handle both of these ... if I can just get it working ... I upgraded my test machine on Friday, and now it generates MCAs all the time, not just when I inject them :-(
If you are working on MCA on 2.4 and using kdb, apply this patch. Some MCA registers were not where the kernel unwind code expected them to be.
ia64/kernel/mca.c renamed from ia64/kernel/mca.c-0 +80 -2
@@ -2380,8 +2380,8 @@ ia64_log_print(int sal_info_type, prfunc/* This bit is tricky. The main MCA handler (but not the MCA rendezvous * handler) has its own stack and bspstore, it does not use the current task * area. The monarch INIT handler has its own stack but uses the current -* bspstore. The slave INIT handlers share a dedicated stack and bspstore, -* single threading through it. +* bspstore. The slave INIT handlers share a dedicated stack but use the +* current bspstore, single threading through the shared stack. * * For all of the MCA and INIT handlers, r13 points to the current task. r12 is * not pointing to the current stack, except in the MCA rendezvous handler.
@@ -2467,6 +2467,13 @@ kdba_release_init_slave_stack(struct pt_* built but they are on the interrupt handler's stack, not on current. Copy * them across to current and adjust b0, bspstore, etc. to suit. Update * kdb_running_process to point to the copies. Finally we can enter kdb. +* +* Assumption: unw_init_running() does DO_SAVE_SWITCH_STACK which calls +* save_switch_stack() which does flushrs. Therefore all registers +* prior to br.call save_switch_stack have been written to backing +* store. +* +* data->bspstore must contain ar.bsp at the time of MCA/INIT. */ static void
@@ -2496,6 +2503,74 @@ kdba_mca_init_handler2(struct kdba_mca_i KDBA_UNWIND_HANDLER(kdba_mca_init_handler, struct kdba_mca_init_data, 0,kdba_mca_init_handler2 (data)); +/* The MCA handler does not use backing store in the process stack, it uses its +* own backing store, ia64_mca_bspstore. How many registers are saved in the +* process stack and how many in ia64_mca_bspstore is timing dependent, RSE +* runs asynchronously. The unwind code requires that all registers be in the +* process stack, so copy any registers from ia64_mca_bspstore to the process +* stack. +* +* Registers from ar.bspstore through ar.bsp+sof at the time of the MCA are +* really in ia64_mca_bspstore, copy them back to the process stack. The copy +* must be done register by register because the process stack and +* ia64_mca_bspstore have different alignments, which means that the saved RNAT +* data occurs at different places. +* +* FIXME: The code assumes that all registers are valid and sets 0 RNaT words +* when copying back to the original stack. +*/ + +static void +kdba_mca_bspstore_fixup (const sal_processor_static_info_t *s) +{ +u64 *old_bspstore, *old_bsp; +u64 *new_bspstore, *new_bsp; +u64 new_bsp_pa, ia64_mca_bspstore_pa; +u64 sof, slots; + + asmvolatile (";;flushrs;; mov %0=ar.bsp;;" :"=r" (new_bsp)); + +/* WAR for inconsistent V->P->V mappings in mca_asm.S for non-identity +* mapped kernels. We can end up with a virtual address in ar.bspstore +* that is not the same as ia64_mca_bspstore but it still points to the +* same physical page as ia64_mca_bspstore. Check the physical address +* instead of the virtual one. +*/ + + new_bsp_pa =ia64_tpa ((u64 )new_bsp); + ia64_mca_bspstore_pa =ia64_tpa ((u64)&ia64_mca_bspstore[0 ]); +if (new_bsp_pa < ia64_mca_bspstore_pa || + new_bsp_pa >= ia64_mca_bspstore_pa +sizeof (ia64_mca_bspstore)) { +kdb_printf ("%s: MCA is not using ia64_mca_bspstore, no fixup done [0x%p]\n" , + __FUNCTION__, new_bsp); +return ; + } + + old_bspstore = (u64 *)(s->ar[18 ]); + old_bsp = (u64 *)(s->ar[17 ]); + sof = s->ar[64 ] &0x7f ;/* from ar.pfs at time of MCA */ + slots =ia64_rse_num_regs (old_bspstore, old_bsp) + sof; + new_bspstore = ia64_mca_bspstore; + new_bsp =ia64_rse_skip_regs (new_bspstore, slots); + +kdb_printf ("DEBUG: %s: old_bspstore 0x%p old_bsp 0x%p sof %ld new_bspstore 0x%p new_bsp 0x%p slots %ld %ld\n" , + __FUNCTION__, old_bspstore, old_bsp, sof, new_bspstore, new_bsp, slots,ia64_rse_num_regs (new_bspstore, new_bsp)); + +while (old_bspstore < old_bsp && new_bspstore < new_bsp) { +if (ia64_rse_is_rnat_slot (new_bspstore)) { + ++new_bspstore; +continue ; + } +if (ia64_rse_is_rnat_slot (old_bspstore)) { + *old_bspstore++ =0 ;/* assume that all registers are valid */ +continue ; + } + *old_bspstore++ = *new_bspstore++; + } +if (ia64_rse_is_rnat_slot (old_bspstore)) + *old_bspstore++ =0 ; +} +static void kdba_mca_init (int sal_info_type) {
@@ -2619,9 +2694,12 @@ kdba_mca_init(int sal_info_type) } } +/* Set up the data required by kdba_mca_init_handler2() */ data.regs = ®s; data.r12 = regs.r12; data.bspstore = s->ar[17 ]; +if (sal_info_type =SAL_INFO_TYPE_MCA ) +kdba_mca_bspstore_fixup (s);unw_init_running (kdba_mca_init_handler, &data); }