On Fri, Nov 11, 2022 at 4:37 PM Joel Fernandes wrote: > > > > On Nov 11, 2022, at 4:28 PM, Akhil P Oommen > wrote: > > > > On 11/12/2022 1:19 AM, Joel Fernandes (Google) wrote: > >> Even though the GPU is shut down, during kexec reboot we can have > userspace > >> still running. This is especially true if KEXEC_JUMP is not enabled, > because we > >> do not freeze userspace in this case. > >> > >> To prevent crashes, track that the GPU is shutdown and prevent > get_param() from > >> accessing GPU resources if we find it shutdown. > >> > >> This fixes the following crash during kexec reboot on an ARM64 device > with adreno GPU: > >> > >> [ 292.534314] Kernel panic - not syncing: Asynchronous SError Interrupt > >> [ 292.534323] Hardware name: Google Lazor (rev3 - 8) with LTE (DT) > >> [ 292.534326] Call trace: > >> [ 292.534328] dump_backtrace+0x0/0x1d4 > >> [ 292.534337] show_stack+0x20/0x2c > >> [ 292.534342] dump_stack_lvl+0x60/0x78 > >> [ 292.534347] dump_stack+0x18/0x38 > >> [ 292.534352] panic+0x148/0x3b0 > >> [ 292.534357] nmi_panic+0x80/0x94 > >> [ 292.534364] arm64_serror_panic+0x70/0x7c > >> [ 292.534369] do_serror+0x0/0x7c > >> [ 292.534372] do_serror+0x54/0x7c > >> [ 292.534377] el1h_64_error_handler+0x34/0x4c > >> [ 292.534381] el1h_64_error+0x7c/0x80 > >> [ 292.534386] el1_interrupt+0x20/0x58 > >> [ 292.534389] el1h_64_irq_handler+0x18/0x24 > >> [ 292.534395] el1h_64_irq+0x7c/0x80 > >> [ 292.534399] local_daif_inherit+0x10/0x18 > >> [ 292.534405] el1h_64_sync_handler+0x48/0xb4 > >> [ 292.534410] el1h_64_sync+0x7c/0x80 > >> [ 292.534414] a6xx_gmu_set_oob+0xbc/0x1fc > >> [ 292.534422] a6xx_get_timestamp+0x40/0xb4 > >> [ 292.534426] adreno_get_param+0x12c/0x1e0 > >> [ 292.534433] msm_ioctl_get_param+0x64/0x70 > >> [ 292.534440] drm_ioctl_kernel+0xe8/0x158 > >> [ 292.534448] drm_ioctl+0x208/0x320 > >> [ 292.534453] __arm64_sys_ioctl+0x98/0xd0 > >> [ 292.534461] invoke_syscall+0x4c/0x118 > >> [ 292.534467] el0_svc_common+0x98/0x104 > >> [ 292.534473] do_el0_svc+0x30/0x80 > >> [ 292.534478] el0_svc+0x20/0x50 > >> [ 292.534481] el0t_64_sync_handler+0x78/0x108 > >> [ 292.534485] el0t_64_sync+0x1a4/0x1a8 > >> [ 292.534632] Kernel Offset: 0x1a5f800000 from 0xffffffc008000000 > >> [ 292.534635] PHYS_OFFSET: 0x80000000 > >> [ 292.534638] CPU features: 0x40018541,a3300e42 > >> [ 292.534644] Memory Limit: none > >> > >> Cc: Rob Clark > >> Cc: Steven Rostedt > >> Cc: Ricardo Ribalda > >> Cc: Ross Zwisler > >> Signed-off-by: Joel Fernandes (Google) > >> --- > >> drivers/gpu/drm/msm/adreno/adreno_device.c | 1 + > >> drivers/gpu/drm/msm/adreno/adreno_gpu.c | 2 +- > >> drivers/gpu/drm/msm/msm_gpu.h | 3 +++ > >> 3 files changed, 5 insertions(+), 1 deletion(-) > >> > >> diff --git a/drivers/gpu/drm/msm/adreno/adreno_device.c > b/drivers/gpu/drm/msm/adreno/adreno_device.c > >> index f0cff62812c3..03d912dc0130 100644 > >> --- a/drivers/gpu/drm/msm/adreno/adreno_device.c > >> +++ b/drivers/gpu/drm/msm/adreno/adreno_device.c > >> @@ -612,6 +612,7 @@ static void adreno_shutdown(struct platform_device > *pdev) > >> { > >> struct msm_gpu *gpu = dev_to_gpu(&pdev->dev); > >> + gpu->is_shutdown = true; > >> WARN_ON_ONCE(adreno_system_suspend(&pdev->dev)); > >> } > >> diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.c > b/drivers/gpu/drm/msm/adreno/adreno_gpu.c > >> index 382fb7f9e497..6903c6892469 100644 > >> --- a/drivers/gpu/drm/msm/adreno/adreno_gpu.c > >> +++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.c > >> @@ -251,7 +251,7 @@ int adreno_get_param(struct msm_gpu *gpu, struct > msm_file_private *ctx, > >> struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); > >> /* No pointer params yet */ > >> - if (*len != 0) > >> + if (*len != 0 || gpu->is_shutdown) > >> return -EINVAL; > > This will race with shutdown. > > Could you clarify what you mean? At this point in the code, the shutdown > is completed and it crashes here. > Ok so I think you meant that if the shut down happens after we sample the is_shutdown, then we run into the same issue. I can’t reproduce that but I’ll look into that. Another way might be to synchronize using a mutex. Though maybe the shutdown path can wait for active pm_runtime references? Thanks. > > Probably, propagating back the return value of pm_runtime_get() in every > possible ioctl call path is the right thing to do. > > Ok I’ll look into that. But the patch I posted works reliably and fixes > all crashes we could reproduce. > > > I have never thought about this scenario. Do you know why userspace is > not freezed before kexec? > > I am not sure. It depends on how kexec is used. The userspace freeze > happens only when kexec is called to switch back and forth between > different kernels (persistence mode). In such scenario I believe the > userspace has to be frozen and unfrozen. However for normal kexec, that > does not happen. > > Thanks. > > > > > > -Akhil. > >> switch (param) { > >> diff --git a/drivers/gpu/drm/msm/msm_gpu.h > b/drivers/gpu/drm/msm/msm_gpu.h > >> index ff911e7305ce..f18b0a91442b 100644 > >> --- a/drivers/gpu/drm/msm/msm_gpu.h > >> +++ b/drivers/gpu/drm/msm/msm_gpu.h > >> @@ -214,6 +214,9 @@ struct msm_gpu { > >> /* does gpu need hw_init? */ > >> bool needs_hw_init; > >> + /* is the GPU shutdown? */ > >> + bool is_shutdown; > >> + > >> /** > >> * global_faults: number of GPU hangs not attributed to a > particular > >> * address space > > >