On 3/30/2022 12:03 PM, Jarkko Sakkinen wrote:
On Wed, 2022-03-30 at 10:40 -0700, Reinette Chatre wrote:
Could you please elaborate how the compiler will fix it up?
Sure.
Here's the disassembly of the RBX version:
[0x000021a9]> pi 1 lea rax, [rbx + loc.encl_stack]
Here's the same with s/RBX/RIP/:
[0x000021a9]> pi 5 lea rax, loc.encl_stack
Compiler will substitute correct offset relative to the RIP, well, because it can and it makes sense.
It does not make sense to me because, as proven with my test, the two threads end up sharing the same stack memory.
Reinette
On Wed, 2022-03-30 at 13:05 -0700, Reinette Chatre wrote:
On 3/30/2022 12:03 PM, Jarkko Sakkinen wrote:
On Wed, 2022-03-30 at 10:40 -0700, Reinette Chatre wrote:
Could you please elaborate how the compiler will fix it up?
Sure.
Here's the disassembly of the RBX version:
[0x000021a9]> pi 1 lea rax, [rbx + loc.encl_stack]
Here's the same with s/RBX/RIP/:
[0x000021a9]> pi 5 lea rax, loc.encl_stack
Compiler will substitute correct offset relative to the RIP, well, because it can and it makes sense.
It does not make sense to me because, as proven with my test, the two threads end up sharing the same stack memory.
I see, I need to correct my patch, thanks!
RBX gives correct results because of the binary organization, i.e. TCS's are placed to zero offset and forward, and unrelocated symbol is just compiled in as an untranslated offset.
RPI is given correct results but how the semantics work right now is incompatible.
Still, even for kselftest, I would consider a switch because that way:
1. You can layout binary however you wan and things won't break. 2. You can point to any symbol not just stack, if ever need.
I admit it works semantically but it just super unrobust.
BR, Jarkko
Hi Jarkko,
On 3/30/2022 1:40 PM, Jarkko Sakkinen wrote:
On Wed, 2022-03-30 at 13:05 -0700, Reinette Chatre wrote:
On 3/30/2022 12:03 PM, Jarkko Sakkinen wrote:
On Wed, 2022-03-30 at 10:40 -0700, Reinette Chatre wrote:
Could you please elaborate how the compiler will fix it up?
Sure.
Here's the disassembly of the RBX version:
[0x000021a9]> pi 1 lea rax, [rbx + loc.encl_stack]
Here's the same with s/RBX/RIP/:
[0x000021a9]> pi 5 lea rax, loc.encl_stack
Compiler will substitute correct offset relative to the RIP, well, because it can and it makes sense.
It does not make sense to me because, as proven with my test, the two threads end up sharing the same stack memory.
I see, I need to correct my patch, thanks!
RBX gives correct results because of the binary organization, i.e. TCS's are placed to zero offset and forward, and unrelocated symbol is just compiled in as an untranslated offset.
RPI is given correct results but how the semantics work right now is incompatible.
Still, even for kselftest, I would consider a switch because that way:
- You can layout binary however you wan and things won't break.
- You can point to any symbol not just stack, if ever need.
I admit it works semantically but it just super unrobust.
I do not think that we need an exceptionally flexible runtime as part of the SGX selftests but instead something that is easy(*) to understand while also sufficient to support the tests.
Reinette
* I do not actually consider the existing enclave test binary easy to understand (this thread is proof) but keeping its complexity to be minimal would benefit folks needing to ramp up on SGX and/or debug kselftest failures.
On Wed, 2022-03-30 at 14:29 -0700, Reinette Chatre wrote:
Hi Jarkko,
On 3/30/2022 1:40 PM, Jarkko Sakkinen wrote:
On Wed, 2022-03-30 at 13:05 -0700, Reinette Chatre wrote:
On 3/30/2022 12:03 PM, Jarkko Sakkinen wrote:
On Wed, 2022-03-30 at 10:40 -0700, Reinette Chatre wrote:
Could you please elaborate how the compiler will fix it up?
Sure.
Here's the disassembly of the RBX version:
[0x000021a9]> pi 1 lea rax, [rbx + loc.encl_stack]
Here's the same with s/RBX/RIP/:
[0x000021a9]> pi 5 lea rax, loc.encl_stack
Compiler will substitute correct offset relative to the RIP, well, because it can and it makes sense.
It does not make sense to me because, as proven with my test, the two threads end up sharing the same stack memory.
I see, I need to correct my patch, thanks!
RBX gives correct results because of the binary organization, i.e. TCS's are placed to zero offset and forward, and unrelocated symbol is just compiled in as an untranslated offset.
RPI is given correct results but how the semantics work right now is incompatible.
Still, even for kselftest, I would consider a switch because that way:
- You can layout binary however you wan and things
won't break. 2. You can point to any symbol not just stack, if ever need. I admit it works semantically but it just super unrobust.
I do not think that we need an exceptionally flexible runtime as part of the SGX selftests but instead something that is easy(*) to understand while also sufficient to support the tests.
Reinette
- I do not actually consider the existing enclave test binary
easy to understand (this thread is proof) but keeping its complexity to be minimal would benefit folks needing to ramp up on SGX and/or debug kselftest failures.
Based on you feedback I refined the patch:
https://lore.kernel.org/linux-sgx/20220330222834.139769-1-jarkko@kernel.org/...
BR, Jarkko
BR, Jarkko
linux-kselftest-mirror@lists.linaro.org
-
Jarkko Sakkinen -
Reinette Chatre