On 2021-05-12 07:50:30, Jens Wiklander wrote:
On Wed, May 12, 2021 at 2:23 AM Tyler Hicks tyhicks@linux.microsoft.com wrote:
On 2021-05-10 09:31:51, Jens Wiklander wrote:
On Fri, May 7, 2021 at 3:17 PM Tyler Hicks tyhicks@linux.microsoft.com wrote:
On 2021-05-07 11:23:17, Jens Wiklander wrote:
On Fri, May 7, 2021 at 9:00 AM Allen Pais apais@linux.microsoft.com wrote:
> On 07-May-2021, at 9:28 AM, Tyler Hicks tyhicks@linux.microsoft.com wrote: > > The .shutdown hook is not called after a kernel crash when a kdump > kernel is pre-loaded. A kexec into the kdump kernel takes place as > quickly as possible without allowing drivers to clean up. > > That means that the OP-TEE shared memory cache, which was initialized by > the kernel that crashed, is still in place when the kdump kernel is > booted. As the kdump kernel is shutdown, the .shutdown hook is called, > which calls optee_disable_shm_cache(), and OP-TEE's > OPTEE_SMC_DISABLE_SHM_CACHE API returns virtual addresses that are not > mapped for the kdump kernel since the cache was set up by the previous > kernel. Trying to dereference the tee_shm pointer or otherwise translate > the address results in a fault that cannot be handled: > > Unable to handle kernel paging request at virtual address ffff4317b9c09744 > Mem abort info: > ESR = 0x96000004 > EC = 0x25: DABT (current EL), IL = 32 bits > SET = 0, FnV = 0 > EA = 0, S1PTW = 0 > Data abort info: > ISV = 0, ISS = 0x00000004 > CM = 0, WnR = 0 > swapper pgtable: 4k pages, 48-bit VAs, pgdp=0000000970b1e000 > [ffff4317b9c09744] pgd=0000000000000000, p4d=0000000000000000 > Internal error: Oops: 96000004 [#1] SMP > Modules linked in: bnxt_en pcie_iproc_platform pcie_iproc diagbe(O) > CPU: 4 PID: 1 Comm: systemd-shutdow Tainted: G O 5.10.19.8 #1 > Hardware name: Redacted (DT) > pstate: 60400005 (nZCv daif +PAN -UAO -TCO BTYPE=--) > pc : tee_shm_free (/usr/src/kernel/drivers/tee/tee_shm.c:363) > lr : optee_disable_shm_cache (/usr/src/kernel/drivers/tee/optee/call.c:441) > sp : ffff80001005bb70 > x29: ffff80001005bb70 x28: ffff608e74648e00 > x27: ffff80001005bb98 x26: dead000000000100 > x25: ffff80001005bbb8 x24: aaaaaaaaaaaaaaaa > x23: ffff608e74cf8818 x22: ffff608e738be600 > x21: ffff80001005bbc8 x20: ffff608e738be638 > x19: ffff4317b9c09700 x18: ffffffffffffffff > x17: 0000000000000041 x16: ffffba61b5171764 > x15: 0000000000000004 x14: 0000000000000fff > x13: ffffba61b5c9dfc8 x12: 0000000000000003 > x11: 0000000000000000 x10: 0000000000000000 > x9 : ffffba61b5413824 x8 : 00000000ffff4317 > x7 : 0000000000000000 x6 : 0000000000000000 > x5 : 0000000000000000 x4 : 0000000000000000 > x3 : 0000000000000000 x2 : ffff4317b9c09700 > x1 : 00000000ffff4317 x0 : ffff4317b9c09700 > Call trace: > tee_shm_free (/usr/src/kernel/drivers/tee/tee_shm.c:363) > optee_disable_shm_cache (/usr/src/kernel/drivers/tee/optee/call.c:441) > optee_shutdown (/usr/src/kernel/drivers/tee/optee/core.c:636) > platform_drv_shutdown (/usr/src/kernel/drivers/base/platform.c:800) > device_shutdown (/usr/src/kernel/include/linux/device.h:758 /usr/src/kernel/drivers/base/core.c:4078) > kernel_restart (/usr/src/kernel/kernel/reboot.c:221 /usr/src/kernel/kernel/reboot.c:248) > __arm64_sys_reboot (/usr/src/kernel/kernel/reboot.c:349 /usr/src/kernel/kernel/reboot.c:312 /usr/src/kernel/kernel/reboot.c:312) > do_el0_svc (/usr/src/kernel/arch/arm64/kernel/syscall.c:56 /usr/src/kernel/arch/arm64/kernel/syscall.c:158 /usr/src/kernel/arch/arm64/kernel/syscall.c:197) > el0_svc (/usr/src/kernel/arch/arm64/kernel/entry-common.c:368) > el0_sync_handler (/usr/src/kernel/arch/arm64/kernel/entry-common.c:428) > el0_sync (/usr/src/kernel/arch/arm64/kernel/entry.S:671) > Code: aa0003f3 b5000060 12800003 14000002 (b9404663) > > When booting the kdump kernel, drain the shared memory cache while being > careful to not translate the addresses returned from > OPTEE_SMC_DISABLE_SHM_CACHE. Once the invalid cache objects are drained > and the cache is disabled, proceed with re-enabling the cache so that we > aren't dealing with invalid addresses while shutting down the kdump > kernel. > > Signed-off-by: Tyler Hicks tyhicks@linux.microsoft.com > --- > > This patch fixes a crash introduced by "optee: fix tee out of memory > failure seen during kexec reboot"[1]. However, I don't think that the > original two patch series[2] plus this patch is the full solution to > properly handling OP-TEE shared memory across kexec. > > While testing this fix, I did about 10 kexec reboots and then triggered > a kernel crash by writing 'c' to /proc/sysrq-trigger. The kdump kernel > became unresponsive during boot while steadily streaming the following > errors to the serial console: > > arm-smmu 64000000.mmu: Blocked unknown Stream ID 0x2000; boot with "arm-smmu.disable_bypass=0" to allow, but this may have security implications > arm-smmu 64000000.mmu: GFSR 0x00000002, GFSYNR0 0x00000002, GFSYNR1 0x00002000, GFSYNR2 0x00000000 > > I suspect that this is related to the problems of OP-TEE shared memory > handling across kexec. My current hunch is that while we've disabled the > shared memory cache with this patch, we haven't unregistered all of the > addresses that the previous kernel (which crashed) had registered with > OP-TEE and that perhaps OP-TEE OS is still trying to make use those > addresses?
@Jens did you have any thoughts on what could be happening here with the arm-smmu errors? Do I need to try to unregister the cached shared memory addresses when booting the kdump kernel, rather than just disabling the caches?
No idea. There's no support for SMMU in upstream OP-TEE. Just disabling the caches should be good enough. You could try to never enable the cache so see if it makes any difference.
I think this is unrelated to OP-TEE and more to do with ongoing DMA activity when the kernel has crashed and we've done an emergency kexec into the kdump kernel which didn't shutdown the SMMU. The SoC I'm using has a v2 SMMU and I think something similar to commit 3f54c447df34 ("iommu/arm-smmu-v3: Don't disable SMMU in kdump kernel") is needed for the v1/v2 SMMU driver. I've prototyped a patch that disables the SMMU interrupts (GFIE and GCFGFIE) in the kdump kernel and testing has looked good so far. I'll send that out as a separate patch after a little more testing.
However, with that change and my earlier change to disable the shm cache during boot, I'm periodically seeing a different issue while the kdump kernel is coming up. I'm pretty certain it was already there before but I wasn't seeing it as often since the SMMU warnings were so "loud".
The kernel waits indefinitely for a secure world thread and boot hangs completely:
[ 243.359489] INFO: task swapper/0:1 blocked for more than 120 seconds. [ 243.366141] Not tainted 5.4.83-microsoft-standard #1 [ 243.371802] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 243.379882] swapper/0 D 0 1 0 0x00000028 [ 243.385543] Call trace: [ 243.388080] __switch_to+0xc8/0x118 [ 243.391683] __schedule+0x2e0/0x700 [ 243.395280] schedule+0x38/0xb8 [ 243.398522] schedule_timeout+0x258/0x388 [ 243.402659] wait_for_completion+0x16c/0x4b8 [ 243.407067] optee_cq_wait_for_completion+0x28/0xa8 [ 243.412100] optee_disable_shm_cache+0xb8/0xf8 [ 243.416685] optee_probe+0x560/0x61c [ 243.420375] platform_drv_probe+0x58/0xa8 [ 243.424512] really_probe+0xe0/0x338 [ 243.428202] driver_probe_device+0x5c/0xf0 [ 243.432427] device_driver_attach+0x74/0x80 [ 243.436744] __driver_attach+0x64/0xe0 [ 243.440611] bus_for_each_dev+0x84/0xd8 [ 243.444570] driver_attach+0x30/0x40 [ 243.448258] bus_add_driver+0x188/0x1e8 [ 243.452215] driver_register+0x64/0x110 [ 243.456172] __platform_driver_register+0x54/0x60 [ 243.461027] optee_driver_init+0x20/0x28 [ 243.465075] do_one_initcall+0x54/0x24c [ 243.469034] kernel_init_freeable+0x1e8/0x2c0 [ 243.473529] kernel_init+0x18/0x118 [ 243.477128] ret_from_fork+0x10/0x18
I'm unable to trigger a sysrq over the serial console of this remote machine so I don't yet know what the other threads on the system are doing during this time. I'll hack something together tomorrow to get a better idea.
The blocked task warning reminded me of when you said this earlier:
Bear in mind that there are other times where we can't recover from a kernel crash. For instance if a thread is executing in OP-TEE in secure world.
I suspect that it is related to what I'm seeing with this blocked task. Can you expand on why we can't recover from a kernel crash if a thread is executing in the secure world?
Threads in OP-TEE are scheduled by Linux so if a thread is executing it may be preempted. In OP-TEE that's a suspended thread waiting to be resumed. If the kernel restarts at this moment that thread will be lost in a suspended state. It may actually explain what you're seeing above. optee_disable_shm_cache() is supposed to try until all threads in OP-TEE are free, that means no suspended threads either.
I think everything is alright when the shutdown path is able to call optee_disable_shm_cache() because we know that there are no suspended threads hanging around. This is the case on the normal reboot and shutdown paths but not the case after a panic with an emergency reboot into the kdump kernel. I verified that I'm seeing OPTEE_SMC_RETURN_ETHREAD_LIMIT returned from the secure world during these hangs.
These suspended threads are a bit dangerous to a restarted kernel in case they are resumed as they may very well be using some old shared memory objects where the physical memory now is used for some other purpose. Cleaning out those threads might be tricky since we can't just reset the secure world state, instead I believe that they will need to be given enough CPU time to eventually complete. However, this is a case which we haven't tested in OP-TEE so there's a risk of running into some not so well tested error paths.
The kdump kernel runs from a pre-reserved area of memory. Therefore, I don't think that there's a chance of the secure world touching physical memory that's being used by the kdump kernel. The problem is that the kdump kernel doesn't have access to the optee_wait_queue of the kernel that crashed. If I understand the RPC scheduling logic correctly, that means that the kdump kernel cannot schedule those suspended threads during boot. I think the only safe option is going to be to bail out of optee_probe(), with -ENODEV, if is_kdump_kernel() returns true.
I tried to skip setting up the shm cache when booting the kdump kernel but saw the same hang in an optee_open_session() -> optee_do_call_with_arg() calling sequence.
Tyler
Cheers, Jens