| Message ID | 1329330098-31636-2-git-send-email-dave.martin@linaro.org |
|---|---|
| State | Not Applicable |
| Headers | show |
On Wed, 15 Feb 2012 18:21:37 +0000, Dave Martin <dave.martin@linaro.org> wrote: > The architecture says that registers such as VTTBR and HVBAR are > only accessible from the Secure World if in Monitor mode and with > SCR.NS = 1. > > Myabe some versions of the model don't enforce this, but the one I > have does, and it appears to be an architectural requirement. This > patch avoids the resulting #undefs. Damn, I wish I'd read this before I started debugging! Thanks for doing all the work though. I've pushed your fixes to github for the moment. I'll fold them together into a nice bisectable commit for merge into the linaro boot-wrapper tree. This will be cleaner when I remove the old hvbar trap, but this works. Thanks! Rusty.
On Mon, Feb 20, 2012 at 09:50:03AM +1030, Rusty Russell wrote: > On Wed, 15 Feb 2012 18:21:37 +0000, Dave Martin <dave.martin@linaro.org> wrote: > > The architecture says that registers such as VTTBR and HVBAR are > > only accessible from the Secure World if in Monitor mode and with > > SCR.NS = 1. > > > > Myabe some versions of the model don't enforce this, but the one I > > have does, and it appears to be an architectural requirement. This > > patch avoids the resulting #undefs. > > Damn, I wish I'd read this before I started debugging! Thanks for doing > all the work though. This wansn't too much of an issue, since I've encountered this kind of thing before. It's not that intuitive though... I can kind of see why SCR.NS=1 is needed there, but the hypervisor system registers are not actually banked between the Secure and Normal Worlds, so they could have got away without it. > I've pushed your fixes to github for the moment. I'll fold them > together into a nice bisectable commit for merge into the linaro > boot-wrapper tree. > > This will be cleaner when I remove the old hvbar trap, but this works. My fix is a bodge really -- it could probably be done more neatly. Thanks for merging it in, anyway. Cheers ---Dave
diff --git a/monitor.S b/monitor.S index 052ab1e..334186e 100644 --- a/monitor.S +++ b/monitor.S @@ -25,7 +25,7 @@ @ 1: ldr sp, =_monitor_stack - push {r11, r12} + push {r10-r12,lr} cmp r7, #0xffffffff beq _non_sec @@ -36,16 +36,19 @@ movnes pc, lr and r12, r7, #0xf cmp r12, #0x0 - popgt {r11, r12} - movgts pc, lr + ldmgtfd sp!, {r10-r12,pc}^ @ Check the VMID is 0 + mrc p15, 0, r10, c1, c1, 0 @ SCR + orr r11, r10, #1 @ SCR.NS = 1 + mcr p15, 0, r11, c1, c1, 0 + isb mrrc p15, 6, r12, r11, c2 + mcr p15, 0, r10, c1, c1, 0 @ Restore SCR lsr r11, r11, #16 and r11, r11, #0xff cmp r11, #0 - popne {r11, r12} - movnes pc, lr + ldmnefd sp!, {r10-r12,pc}^ @ Jump to the right function and r12, r7, #0xf @@ -68,16 +71,18 @@ _non_sec: ldr r11, =0x131 orr r12, r12, r11 mcr p15, 0, r12, c1, c1, 0 - pop {r11, r12} - movs pc, lr + ldmfd sp!, {r10-r12,pc}^ @ @ Read/Write HVBAR @ _write_hvbar: + orr r11, r10, #1 @ SCR.NS = 1 (r10 already = SCR) + mcr p15, 0, r11, c1, c1, 0 + isb mcr p15, 4, r0, c12, c0, 0 - pop {r11, r12} - movs pc, lr + mcr p15, 0, r10, c1, c1, 0 @ Restore SCR + ldmfd sp!, {r10-r12,pc}^ .ltorg
The architecture says that registers such as VTTBR and HVBAR are only accessible from the Secure World if in Monitor mode and with SCR.NS = 1. Myabe some versions of the model don't enforce this, but the one I have does, and it appears to be an architectural requirement. This patch avoids the resulting #undefs. Signed-off-by: Dave Martin <dave.martin@linaro.org> --- monitor.S | 23 ++++++++++++++--------- 1 files changed, 14 insertions(+), 9 deletions(-)