Skip to contentSkip to conversation

optee: Do not send requests to supplicant during shutdown

PATCH In review Lars Persson 4 participants 8 messages Jan 20, 2022 – Jan 27, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 optee: Do not send requests to supplicant during shutdown

The addition of a shutdown hook by commit f25889f93184 ("optee:
fix tee out of memory failure seen during kexec reboot") introduced a kernel shutdown regression that can be triggered after running the xtest suites.

Once the shutdown hook is called it is not possible to communicate any more with the supplicant process because the system is not scheduling task any longer. Thus if the optee driver shutdown path receives a supplicant RPC request from the OP-TEE we will deadlock the kernel's shutdown.

This unexpected event will in fact occur after the xtest suite has been run. It seems some cached SHM kept alive a context object which in turn kept alive a session towards a PTA or TA. Closing the session results in a socket RPC command being sent back from OP-TEE.

This sequence of events is captured by a 5.15 kernel annotated with extra prints:

Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001079380
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001CC5580
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308A80
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308B00
optee: optee_handle_rpc: a0=0XFFFF0000 a1=0XA0 a2=0X0
optee: optee_handle_rpc: a0=0XFFFF0005 a1=0XFFFFFF80 a2=0X61E6500
optee: handle_rpc_func_cmd: cmd = 0XA
optee_supp_thrd_req: func=0XA

Introduce a shutdown state in the optee device object to return an immediate error to all RPC requests in the shutdown path.

Fixes f25889f93184 ("optee: fix tee out of memory failure seen during kexec reboot")
---
 drivers/tee/optee/optee_private.h | 1 +
 drivers/tee/optee/smc_abi.c       | 5 ++++-
 drivers/tee/optee/supp.c          | 8 ++++++++
 3 files changed, 13 insertions(+), 1 deletion(-)
drivers/tee/optee/optee_private.h +1 -0
index 46f74ab07c7e..9eb72931e11f 100644
@@ -162,6 +162,7 @@ struct optee {
 	struct tee_shm_pool *pool;
 	unsigned int rpc_arg_count;
 	bool   scan_bus_done;
+	bool   shutting_down;
 	struct workqueue_struct *scan_bus_wq;
 	struct work_struct scan_bus_work;
 };
drivers/tee/optee/smc_abi.c +4 -1
index 449d6a72d289..10af747da816 100644
@@ -1356,7 +1356,10 @@ static int optee_smc_remove(struct platform_device *pdev)
  */
 static void optee_shutdown(struct platform_device *pdev)
 {
-	optee_disable_shm_cache(platform_get_drvdata(pdev));
+	struct optee *optee = platform_get_drvdata(pdev);
+
+	optee->shutting_down = true;
+	optee_disable_shm_cache(optee);
 }
 
 static int optee_probe(struct platform_device *pdev)
drivers/tee/optee/supp.c +8 -0
index 322a543b8c27..801b4ec659e8 100644
@@ -83,6 +83,14 @@ u32 optee_supp_thrd_req(struct tee_context *ctx, u32 func, size_t num_params,
 	bool interruptable;
 	u32 ret;
 
+	/*
+	 * When the system is shutting down we cannot talk
+	 * to the supplicant anymore even if we seem to have
+	 * an open session to it.
+	 */
+	if (optee->shutting_down)
+		return TEEC_ERROR_COMMUNICATION;
+
 	/*
 	 * Return in case there is no supplicant available and
 	 * non-blocking request.
Signature · 1 line
2.30.2
#1 lore raw
Quoted 36 lines · Lars Persson

The addition of a shutdown hook by commit f25889f93184 ("optee:
fix tee out of memory failure seen during kexec reboot") introduced a kernel shutdown regression that can be triggered after running the xtest suites.

Once the shutdown hook is called it is not possible to communicate any more with the supplicant process because the system is not scheduling task any longer. Thus if the optee driver shutdown path receives a supplicant RPC request from the OP-TEE we will deadlock the kernel's shutdown.

This unexpected event will in fact occur after the xtest suite has been run. It seems some cached SHM kept alive a context object which in turn kept alive a session towards a PTA or TA. Closing the session results in a socket RPC command being sent back from OP-TEE.

This sequence of events is captured by a 5.15 kernel annotated with extra prints:

Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001079380
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001CC5580
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308A80
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308B00
optee: optee_handle_rpc: a0=0XFFFF0000 a1=0XA0 a2=0X0
optee: optee_handle_rpc: a0=0XFFFF0005 a1=0XFFFFFF80 a2=0X61E6500
optee: handle_rpc_func_cmd: cmd = 0XA
optee_supp_thrd_req: func=0XA

Quoted 5 lines

Introduce a shutdown state in the optee device object to return an immediate error to all RPC requests in the shutdown path.

Fixes f25889f93184 ("optee: fix tee out of memory failure seen during kexec reboot")

I think that this change looks alright to fix the deadlock but I don't feel qualified to definitively say that this is the right fix. Something makes me wonder why the session is kept alive and if there's an underlying bug that we're masking over but I don't understand the lifetime of the objects, sessions, etc., when the supplicant is involved. I'm going to defer to Jens and Sumit on the design aspects.

Tyler

Quoted 56 lines
---
 drivers/tee/optee/optee_private.h | 1 +
 drivers/tee/optee/smc_abi.c       | 5 ++++-
 drivers/tee/optee/supp.c          | 8 ++++++++
 3 files changed, 13 insertions(+), 1 deletion(-)
drivers/tee/optee/optee_private.h +1 -0
index 46f74ab07c7e..9eb72931e11f 100644
@@ -162,6 +162,7 @@ struct optee {
 	struct tee_shm_pool *pool;
 	unsigned int rpc_arg_count;
 	bool   scan_bus_done;
+	bool   shutting_down;
 	struct workqueue_struct *scan_bus_wq;
 	struct work_struct scan_bus_work;
 };
drivers/tee/optee/smc_abi.c +4 -1
index 449d6a72d289..10af747da816 100644
@@ -1356,7 +1356,10 @@ static int optee_smc_remove(struct platform_device *pdev)
  */
 static void optee_shutdown(struct platform_device *pdev)
 {
-	optee_disable_shm_cache(platform_get_drvdata(pdev));
+	struct optee *optee = platform_get_drvdata(pdev);
+
+	optee->shutting_down = true;
+	optee_disable_shm_cache(optee);
 }
 
 static int optee_probe(struct platform_device *pdev)
drivers/tee/optee/supp.c +8 -0
index 322a543b8c27..801b4ec659e8 100644
@@ -83,6 +83,14 @@ u32 optee_supp_thrd_req(struct tee_context *ctx, u32 func, size_t num_params,
 	bool interruptable;
 	u32 ret;
 
+	/*
+	 * When the system is shutting down we cannot talk
+	 * to the supplicant anymore even if we seem to have
+	 * an open session to it.
+	 */
+	if (optee->shutting_down)
+		return TEEC_ERROR_COMMUNICATION;
+
 	/*
 	 * Return in case there is no supplicant available and
 	 * non-blocking request.
Signature · 1 line
2.30.2
#2 lore raw
In reply to Lars Persson

Hi Lars,

Quoted 20 lines · Lars Persson

The addition of a shutdown hook by commit f25889f93184 ("optee:
fix tee out of memory failure seen during kexec reboot") introduced a kernel shutdown regression that can be triggered after running the xtest suites.

Once the shutdown hook is called it is not possible to communicate any more with the supplicant process because the system is not scheduling task any longer. Thus if the optee driver shutdown path receives a supplicant RPC request from the OP-TEE we will deadlock the kernel's shutdown.

This unexpected event will in fact occur after the xtest suite has been run. It seems some cached SHM kept alive a context object which in turn kept alive a session towards a PTA or TA. Closing the session results in a socket RPC command being sent back from OP-TEE.

This sequence of events is captured by a 5.15 kernel annotated with extra prints:

Quoted 17 lines

Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001079380
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001CC5580
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308A80
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308B00
optee: optee_handle_rpc: a0=0XFFFF0000 a1=0XA0 a2=0X0
optee: optee_handle_rpc: a0=0XFFFF0005 a1=0XFFFFFF80 a2=0X61E6500
optee: handle_rpc_func_cmd: cmd = 0XA
optee_supp_thrd_req: func=0XA

This looks like another side effect (earlier known one is here [1]) of shared memory cache being allocated via RPC in a particular client's context. There is an appropriate fix from Jens here [2] to rather use driver internal tee_contex for RPC allocations. Can you try that and let us know if you still observe this issue?

[1] https://github.com/OP-TEE/optee_os/issues/1918
[2] https://lore.kernel.org/lkml/20220114150824.3578829-8-jens.wiklander(a)linaro.org/

-Sumit

Quoted 61 lines

Introduce a shutdown state in the optee device object to return an immediate error to all RPC requests in the shutdown path.

Fixes f25889f93184 ("optee: fix tee out of memory failure seen during kexec reboot")
---
 drivers/tee/optee/optee_private.h | 1 +
 drivers/tee/optee/smc_abi.c       | 5 ++++-
 drivers/tee/optee/supp.c          | 8 ++++++++
 3 files changed, 13 insertions(+), 1 deletion(-)
drivers/tee/optee/optee_private.h +1 -0
index 46f74ab07c7e..9eb72931e11f 100644
@@ -162,6 +162,7 @@ struct optee {
        struct tee_shm_pool *pool;
        unsigned int rpc_arg_count;
        bool   scan_bus_done;
+       bool   shutting_down;
        struct workqueue_struct *scan_bus_wq;
        struct work_struct scan_bus_work;
 };
drivers/tee/optee/smc_abi.c +4 -1
index 449d6a72d289..10af747da816 100644
@@ -1356,7 +1356,10 @@ static int optee_smc_remove(struct platform_device *pdev)
  */
 static void optee_shutdown(struct platform_device *pdev)
 {
-       optee_disable_shm_cache(platform_get_drvdata(pdev));
+       struct optee *optee = platform_get_drvdata(pdev);
+
+       optee->shutting_down = true;
+       optee_disable_shm_cache(optee);
 }
 
 static int optee_probe(struct platform_device *pdev)
drivers/tee/optee/supp.c +8 -1
index 322a543b8c27..801b4ec659e8 100644
@@ -83,6 +83,14 @@ u32 optee_supp_thrd_req(struct tee_context *ctx, u32 func, size_t num_params,
        bool interruptable;
        u32 ret;
 
+       /*
+        * When the system is shutting down we cannot talk
+        * to the supplicant anymore even if we seem to have
+        * an open session to it.
+        */
+       if (optee->shutting_down)
+               return TEEC_ERROR_COMMUNICATION;
+
        /*
         * Return in case there is no supplicant available and
         * non-blocking request.
--

2.30.2

#3 lore raw
Quoted 47 lines · Sumit Garg

Hi Lars,

Quoted 20 lines · Lars Persson

The addition of a shutdown hook by commit f25889f93184 ("optee:
fix tee out of memory failure seen during kexec reboot") introduced a kernel shutdown regression that can be triggered after running the xtest suites.

Once the shutdown hook is called it is not possible to communicate any more with the supplicant process because the system is not scheduling task any longer. Thus if the optee driver shutdown path receives a supplicant RPC request from the OP-TEE we will deadlock the kernel's shutdown.

This unexpected event will in fact occur after the xtest suite has been run. It seems some cached SHM kept alive a context object which in turn kept alive a session towards a PTA or TA. Closing the session results in a socket RPC command being sent back from OP-TEE.

This sequence of events is captured by a 5.15 kernel annotated with extra prints:

Quoted 17 lines

  Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
   freeing SHM ptr 0xFFFFFF8001079380
  Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
   freeing SHM ptr 0xFFFFFF8001CC5580
  Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
   freeing SHM ptr 0xFFFFFF8006308A80
  Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
   freeing SHM ptr 0xFFFFFF8006308B00
  optee: optee_handle_rpc: a0=0XFFFF0000 a1=0XA0 a2=0X0
  optee: optee_handle_rpc: a0=0XFFFF0005 a1=0XFFFFFF80 a2=0X61E6500
  optee: handle_rpc_func_cmd: cmd = 0XA
  optee_supp_thrd_req: func=0XA

This looks like another side effect (earlier known one is here [1]) of shared memory cache being allocated via RPC in a particular client's context. There is an appropriate fix from Jens here [2] to rather use driver internal tee_contex for RPC allocations. Can you try that and let us know if you still observe this issue?

Quoted 8 lines

[1] https://github.com/OP-TEE/optee_os/issues/1918
<https://github.com/OP-TEE/optee_os/issues/1918>
[2]
https://lore.kernel.org/lkml/20220114150824.3578829-8-jens.wiklander(a)linaro.org/
<https://lore.kernel.org/lkml/[email protected]/>

-Sumit

Indeed it seems this is the root cause. Backporting the patches to 5.15 also fixes our shutdown deadlock.

Nevertheless I prefer to have this extra guard in the supplicant interface to prevent other shutdown deadlocks in the future.

/Lars

#4 lore raw
Quoted 57 lines · Lars Persson
Quoted 53 lines · Sumit Garg

Hi Lars,

Quoted 20 lines · Lars Persson

The addition of a shutdown hook by commit f25889f93184 ("optee:
fix tee out of memory failure seen during kexec reboot") introduced a kernel shutdown regression that can be triggered after running the xtest suites.

Once the shutdown hook is called it is not possible to communicate any more with the supplicant process because the system is not scheduling task any longer. Thus if the optee driver shutdown path receives a supplicant RPC request from the OP-TEE we will deadlock the kernel's shutdown.

This unexpected event will in fact occur after the xtest suite has been run. It seems some cached SHM kept alive a context object which in turn kept alive a session towards a PTA or TA. Closing the session results in a socket RPC command being sent back from OP-TEE.

This sequence of events is captured by a 5.15 kernel annotated with extra prints:

Quoted 17 lines

Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001079380
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001CC5580
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308A80
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308B00
optee: optee_handle_rpc: a0=0XFFFF0000 a1=0XA0 a2=0X0
optee: optee_handle_rpc: a0=0XFFFF0005 a1=0XFFFFFF80 a2=0X61E6500
optee: handle_rpc_func_cmd: cmd = 0XA
optee_supp_thrd_req: func=0XA

This looks like another side effect (earlier known one is here [1]) of shared memory cache being allocated via RPC in a particular client's context. There is an appropriate fix from Jens here [2] to rather use driver internal tee_contex for RPC allocations. Can you try that and let us know if you still observe this issue?

[1] https://github.com/OP-TEE/optee_os/issues/1918
<https://github.com/OP-TEE/optee_os/issues/1918>
[2]
https://lore.kernel.org/lkml/20220114150824.3578829-8-jens.wiklander(a)linaro.org/
<https://lore.kernel.org/lkml/[email protected]/>

Quoted 5 lines
Quoted 2 lines

-Sumit

Indeed it seems this is the root cause. Backporting the patches to 5.15 also fixes our shutdown deadlock.

That's good to hear.

Quoted 3 lines

Nevertheless I prefer to have this extra guard in the supplicant interface to prevent other shutdown deadlocks in the future.

Yes, I agree, this still makes sense. Perhaps with an updated description to avoid unnecessary confusion when applied on top of my "tee: shared memory updates" patches.

Thanks,
Jens

#5 lore raw
Quoted 68 lines · Jens Wiklander
Quoted 57 lines · Lars Persson
Quoted 53 lines · Sumit Garg

Hi Lars,

Quoted 20 lines · Lars Persson

The addition of a shutdown hook by commit f25889f93184 ("optee:
fix tee out of memory failure seen during kexec reboot") introduced a kernel shutdown regression that can be triggered after running the xtest suites.

Once the shutdown hook is called it is not possible to communicate any more with the supplicant process because the system is not scheduling task any longer. Thus if the optee driver shutdown path receives a supplicant RPC request from the OP-TEE we will deadlock the kernel's shutdown.

This unexpected event will in fact occur after the xtest suite has been run. It seems some cached SHM kept alive a context object which in turn kept alive a session towards a PTA or TA. Closing the session results in a socket RPC command being sent back from OP-TEE.

This sequence of events is captured by a 5.15 kernel annotated with extra prints:

Quoted 17 lines

Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001079380
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001CC5580
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308A80
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308B00
optee: optee_handle_rpc: a0=0XFFFF0000 a1=0XA0 a2=0X0
optee: optee_handle_rpc: a0=0XFFFF0005 a1=0XFFFFFF80 a2=0X61E6500
optee: handle_rpc_func_cmd: cmd = 0XA
optee_supp_thrd_req: func=0XA

This looks like another side effect (earlier known one is here [1]) of shared memory cache being allocated via RPC in a particular client's context. There is an appropriate fix from Jens here [2] to rather use driver internal tee_contex for RPC allocations. Can you try that and let us know if you still observe this issue?

[1] https://github.com/OP-TEE/optee_os/issues/1918
<https://github.com/OP-TEE/optee_os/issues/1918>
[2]
https://lore.kernel.org/lkml/20220114150824.3578829-8-jens.wiklander(a)linaro.org/
<https://lore.kernel.org/lkml/20220114150824.3578829-8-jens.wiklander(a)linaro.org/>

Quoted 5 lines
Quoted 2 lines

-Sumit

Indeed it seems this is the root cause. Backporting the patches to 5.15 also fixes our shutdown deadlock.

That's good to hear.

Quoted 1 line
Quoted 7 lines
Quoted 2 lines

Nevertheless I prefer to have this extra guard in the supplicant interface to prevent other shutdown deadlocks in the future.

Yes, I agree, this still makes sense. Perhaps with an updated description to avoid unnecessary confusion when applied on top of my "tee: shared memory updates" patches.

That's fine with me and we should drop the fixes tag as well.

-Sumit

Quoted 2 lines

Thanks,
Jens

#6 lore raw
Quoted 77 lines · Sumit Garg
Quoted 47 lines · Jens Wiklander
Quoted 45 lines · Lars Persson
Quoted 41 lines · Sumit Garg

Hi Lars,

Quoted 37 lines · Lars Persson

The addition of a shutdown hook by commit f25889f93184 ("optee:
fix tee out of memory failure seen during kexec reboot") introduced a kernel shutdown regression that can be triggered after running the xtest suites.

Once the shutdown hook is called it is not possible to communicate any more with the supplicant process because the system is not scheduling task any longer. Thus if the optee driver shutdown path receives a supplicant RPC request from the OP-TEE we will deadlock the kernel's shutdown.

This unexpected event will in fact occur after the xtest suite has been run. It seems some cached SHM kept alive a context object which in turn kept alive a session towards a PTA or TA. Closing the session results in a socket RPC command being sent back from OP-TEE.

This sequence of events is captured by a 5.15 kernel annotated with extra prints:

  Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
   freeing SHM ptr 0xFFFFFF8001079380
  Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
   freeing SHM ptr 0xFFFFFF8001CC5580
  Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
   freeing SHM ptr 0xFFFFFF8006308A80
  Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
   freeing SHM ptr 0xFFFFFF8006308B00
  optee: optee_handle_rpc: a0=0XFFFF0000 a1=0XA0 a2=0X0
  optee: optee_handle_rpc: a0=0XFFFF0005 a1=0XFFFFFF80 a2=0X61E6500
  optee: handle_rpc_func_cmd: cmd = 0XA
  optee_supp_thrd_req: func=0XA

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

This looks like another side effect (earlier known one is here [1]) of shared memory cache being allocated via RPC in a particular client's context. There is an appropriate fix from Jens here [2] to rather use driver internal tee_contex for RPC allocations. Can you try that and let us know if you still observe this issue?

[1] https://github.com/OP-TEE/optee_os/issues/1918

<https://github.com/OP-TEE/optee_os/issues/1918>

Quoted 1 line
Quoted 1 line
Quoted 1 line

<https://github.com/OP-TEE/optee_os/issues/1918

<https://github.com/OP-TEE/optee_os/issues/1918>>

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

[2]
https://lore.kernel.org/lkml/20220114150824.3578829-8-jens.wiklander(a)linaro.org/

<https://lore.kernel.org/lkml/[email protected]/>

Quoted 1 line
Quoted 1 line
Quoted 1 line

<https://lore.kernel.org/lkml/20220114150824.3578829-8-jens.wiklander(a)linaro.org/

<https://lore.kernel.org/lkml/[email protected]/>>

Quoted 13 lines
Quoted 6 lines
Quoted 3 lines

-Sumit

Indeed it seems this is the root cause. Backporting the patches to 5.15 also fixes our shutdown deadlock.

That's good to hear.

Quoted 3 lines

Nevertheless I prefer to have this extra guard in the supplicant interface to prevent other shutdown deadlocks in the future.

Quoted 7 lines
Quoted 4 lines

Yes, I agree, this still makes sense. Perhaps with an updated description to avoid unnecessary confusion when applied on top of my "tee: shared memory updates" patches.

That's fine with me and we should drop the fixes tag as well.

Should we really drop the Fixes ? Depending on if and when the SHM owning context patch is backported, this fix may still be relevant for the 5.15 LTS branch.

/Lars

#7 lore raw
Quoted 86 lines · Lars Persson
Quoted 82 lines · Sumit Garg
Quoted 47 lines · Jens Wiklander
Quoted 45 lines · Lars Persson
Quoted 41 lines · Sumit Garg

Hi Lars,

Quoted 37 lines · Lars Persson

The addition of a shutdown hook by commit f25889f93184 ("optee:
fix tee out of memory failure seen during kexec reboot") introduced a kernel shutdown regression that can be triggered after running the xtest suites.

Once the shutdown hook is called it is not possible to communicate any more with the supplicant process because the system is not scheduling task any longer. Thus if the optee driver shutdown path receives a supplicant RPC request from the OP-TEE we will deadlock the kernel's shutdown.

This unexpected event will in fact occur after the xtest suite has been run. It seems some cached SHM kept alive a context object which in turn kept alive a session towards a PTA or TA. Closing the session results in a socket RPC command being sent back from OP-TEE.

This sequence of events is captured by a 5.15 kernel annotated with extra prints:

Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001079380
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8001CC5580
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308A80
Calling OPTEE_SMC_DISABLE_SHM_CACHE OPTEE_SMC_DISABLE_SHM_CACHE returned 0
freeing SHM ptr 0xFFFFFF8006308B00
optee: optee_handle_rpc: a0=0XFFFF0000 a1=0XA0 a2=0X0
optee: optee_handle_rpc: a0=0XFFFF0005 a1=0XFFFFFF80 a2=0X61E6500
optee: handle_rpc_func_cmd: cmd = 0XA
optee_supp_thrd_req: func=0XA

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

This looks like another side effect (earlier known one is here [1]) of shared memory cache being allocated via RPC in a particular client's context. There is an appropriate fix from Jens here [2] to rather use driver internal tee_contex for RPC allocations. Can you try that and let us know if you still observe this issue?

[1] https://github.com/OP-TEE/optee_os/issues/1918

<https://github.com/OP-TEE/optee_os/issues/1918>

Quoted 1 line
Quoted 1 line
Quoted 1 line

<https://github.com/OP-TEE/optee_os/issues/1918

<https://github.com/OP-TEE/optee_os/issues/1918>>

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

[2]
https://lore.kernel.org/lkml/20220114150824.3578829-8-jens.wiklander(a)linaro.org/

<https://lore.kernel.org/lkml/[email protected]/>

Quoted 1 line
Quoted 1 line
Quoted 1 line

<https://lore.kernel.org/lkml/[email protected]/

<https://lore.kernel.org/lkml/[email protected]/>>

Quoted 17 lines
Quoted 6 lines
Quoted 3 lines

-Sumit

Indeed it seems this is the root cause. Backporting the patches to 5.15 also fixes our shutdown deadlock.

That's good to hear.

Quoted 3 lines

Nevertheless I prefer to have this extra guard in the supplicant interface to prevent other shutdown deadlocks in the future.

Yes, I agree, this still makes sense. Perhaps with an updated description to avoid unnecessary confusion when applied on top of my "tee: shared memory updates" patches.

Quoted 6 lines
Quoted 2 lines

That's fine with me and we should drop the fixes tag as well.

Should we really drop the Fixes ? Depending on if and when the SHM owning context patch is backported, this fix may still be relevant for the 5.15 LTS branch.

I rather consider this as a robustness patch where the actual problem is being fixed by Jens patch. I hope we should be able to target Jens patch for 5.17 as a fix which should be backported to affected stable releases.

-Sumit

Quoted 4 lines

/Lars