Re: [PATCH] arm: Handle starting up in secure mode

From: Christopher Covington
Date: Wed Aug 26 2015 - 10:19:48 EST


Hi,

Thank you for the feedback.

On 08/26/2015 06:48 AM, Russell King - ARM Linux wrote:
> On Wed, Aug 26, 2015 at 11:39:42AM +0100, Dave Martin wrote:
>> On Mon, Aug 24, 2015 at 09:55:26AM -0400, Christopher Covington wrote:
>>> ARM Linux appears to have never been made aware of the ARMv7 security
>>> extensions. When CONFIG_ARM_SEC_EXT=y, have it probe for its security
>>> state by checking whether CNTFRQ is writeable and potentially make
>>> mode changes based on the information. The most features are available
>>> from hypervisor (HYP) mode, so switch to it possible. Failing that,
>>> prefer non-secure supervisor (SVC) mode to secure supervisor mode.
>>
>> Up to now we've steered clear of this, since it's a bit of a fig leaf
>> for broken firmware unless Linux actually has some valid use for the
>> Security Extensions itself.
>>
>> Shouldn't the bootloader or firmware be doing this stuff, and if not,
>> why not?

In my opinion, it is sometimes useful, particularly when running on simulators
or performing bring up, to have as little dependency on firmwares and
bootloaders as possible. I want to run Linux out of reset, or get as close to
that as possible. I want an ease of use comparable to QEMU's Linux loader on
other simulators, like ARM's RTSM/AEM/FVPs, without various different,
fragmented model plugins or the overhead and noise of the semihosting
bootwrapper. I have attempted to craft the patches in such a way that the
enjoyment of sophisticated firmwares and bootloaders remains unhindered.

>> Some other things that would need to be considered in any case:
>>
>> * SoC-specific setup of the Non-secure view of the system: This has
>> to happen very early, so making it DT aware is going to be hard --
>> failing that, we are effectively risking bringing back board files.
>> The split in responsibility between firmware/bootloader and kernel
>> needs to be clearly defined and (as far as possible) platform-
>> independent, otherwise we'll have total chaos.

My intention was to only perform architecturally defined initialization. All
implementation defined specifics would remain delegated to firmware, simulator
reset state hacks, or external debuggers, as the case may be. Note that using
this code does not preclude running firmware, just that if you want Linux to
do the exception level/mode transitions, you can omit those instructions from
firmware. I'll have to refresh my memory--it's been a while since I worked on
this, I apologize--but I think I used a modified bootwrapper to still perform
its usual GIC initialization.

>> * Out of reset, generally the CPU state is only fully defined for the
>> highest exception level. You probably need to be doing more setup
>> than you're currently doing.

Agreed. This is just the first step.

>> * SMP, secondary boot and suspend/resume -- again involving board-
>> specific code.

I've only tested single core so far, but I was thinking external debuggers and
simulators may be able to emulate PSCI in a semihosting-like manner, if
merging holding pen code is as frowned upon as I think it is (that may be an
AArch64-specific statement, but it'd probably be nice if similar facilities
were available across A32 and A64).

>> * You need to safely "park" the Secure World before running anything
>> in Non-Secure. As a minimum, you would need to quiesce any
>> Secure interrupt sources, disable all interrupt traps to Monitor
>> mode, and make sure that the Monitor vectors point somewhere
>> real, so that executing SMC doesn't send the CPU off into the
>> long grass...

Thanks. I'll work towards support for these aspects in the next version.

> Another question is: has this been tested with kexec?

I have not tested it with kexec. My testing was mostly of cold boots performed
some time back on ARM's AEM/RSTM/FVP using its configurable parameters and
some simple custom firmwares to simulate entry under all the circumstances I
could think of. I grepped instruction traces to make sure the initial mode
after firmware was as intended and then that AArch32 Linux switched into the
intended mode.

EL3: not present, AArch64, AArch32
EL2: not present, AArch64, AArch32
AArch32 Linux starts in: svc_s, svc_n, hyp_n, mon_s

By my calculations, there were 36 - 17 = 19 valid combinations.

Thanks,
Christopher Covington

--
Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/