ACPICA commit 767ee53354e0c4b7e8e7c57c6dd7bf569f0d52bb
There are issues related to the namespace/interpreter locks, which causes
several ACPI functionalities not specification compliant. The lock issues
were detectec when we were trying to fix the functionalities (please see
Link # [1] for the details).
What's the lock issues? Let's first look into the namespace/interpreter
lock usages inside of the object evaluation and the table loading which are
the key AML interpretion code paths:
Table loading:
acpi_ns_load_table
L(Namespace)
acpi_ns_parse_table
acpi_ns_one_complete_parse(LOAD_PASS1/LOAD_PASS2)
acpi_ds_load1_begion_op
acpi_ds_load1_end_op
acpi_ds_load2_begion_op
acpi_ds_load2_end_op
U(Namespace)
Object evaluation:
acpi_ns_evaluate
L(Interpreter)
acpi_ps_execute_method
acpi_ds_exec_begin_op
acpi_ds_exec_end_op
U(Interpreter)
acpi_ns_load_table
L(Namespace)
U(Namespace)
acpi_ev_initialize_region
L(Namespace)
U(Namespace)
address_space.Setup
address_space.Handler
acpi_os_wait_semaphore
acpi_os_acquire_mutex
acpi_os_sleep
L(Interpreter)
U(Interpreter)
L(Interpreter)
acpi_ex_resolve_node_to_value
U(Interpreter)
acpi_ns_check_return_value
Where:
1. L(Interpreter) means acquire(MTX_INTERPRETER);
2. U(Interpreter) means release(MTX_INTERPRETER);
3. L(Namespace) means acquire(MTX_NAMESPACE);
4. U(Namespace) means release(MTX_NAMESPACE);
We can see that acpi_ns_exec_module_code() (which invokes acpi_ns_evaluate) is
implemented in a deferred way just in order to avoid to reacquire the
namespace lock. This is in fact the root cause of many other ACPICA issues:
1. We now know for sure that the module code should be executed right in
place by the Windows AML interpreter. So in the current design, if
the region initializations/accesses or the table loadings (where the
namespace surely should be locked again) happening during the table
loading period, dead lock could happen because ACPICA never unlocks the
namespace during the AML interpretion.
2. ACPICA interpreter just ensures that all static namespace nodes (named
objects created during the acpi_load_tables()) are created
(acpi_ns_lookup()) with the correct lock held, but doesn't ensure that
the named objects created by the control method are created with the
same correct lock held. It requires the control methods to be executed
in a serial way after "loading a table", that's why ACPICA requires
method auto serialization.
This patch fixes these software design issues by extending interpreter
enter/exit APIs to hold both interpreter/namespace locks to ensure the lock
order correctness, so that we can get these code paths:
Table loading:
acpi_ns_load_table
E(Interpreter)
acpi_ns_parse_table
acpi_ns_one_complete_parse
acpi_ns_execute_table
X(Interpreter)
acpi_ns_load_table
acpi_ev_initialize_region
address_space.Setup
address_space.Handler
acpi_os_wait_semaphore
acpi_os_acquire_mutex
acpi_os_sleep
E(Interpreter)
X(Interpreter)
Object evaluation:
acpi_ns_evaluate
E(Interpreter)
acpi_ps_execute_method
X(Interpreter)
acpi_ns_load_table
acpi_ev_initialize_region
address_space.Setup
address_space.Handler
acpi_os_wait_semaphore
acpi_os_acquire_mutex
acpi_os_sleep
E(Interpreter)
X(Interpreter)
Where:
1. E(Interpreter) means acquire(MTX_INTERPRETER, MTX_NAMESPACE);
2. X(Interpreter) means release(MTX_NAMESPACE, MTX_INTERPRETER);
After this change, we can see:
1. All namespace nodes creations are locked by the namespace lock.
2. All namespace nodes referencing are locked with the same lock.
3. But we also can notice a defact that, all namespace nodes deletions
could be affected by this change. As a consequence,
acpi_ns_delete_namespace_subtree() may delete a static namespace node that
is still referenced by the interpreter (for example, the parser scopes).
Currently, we needn't worry about the last defact because in ACPICA, table
unloading is not fully functioning, its design strictly relies on the fact
that when the namespace deletion happens, either the AML table or the OSPMs
should have been notified and thus either the AML table or the OSPMs
shouldn't reference deletion-related namespace nodes during the namespace
deletion. And this change still works with the above restrictions applied.
While making this a-step-forward helps us to correct the wrong grammar to
pull many things back to the correct rail. And pulling things back to the
correct rail in return makes it possible for us to support fully
functioning table unloading after doing many cleanups.
While this patch is generated, all namespace locks are examined to ensure
that they can meet either of the following pattens:
1. L(Namespace)
U(Namespace)
2. E(Interpreter)
X(Interpreter)
3. E(Interpreter)
X(Interpreter)
L(Namespace)
U(Namespace)
E(Interpreter)
X(Interpreter)
We ensure this by adding X(Interpreter)/E(Interpreter) or removing
U(Namespace)/L(Namespace) for those currently are executed in the following
order:
E(Interpreter)
L(Namespace)
U(Namespace)
X(Interpreter)
And adding E(Interpreter)/X(Interpreter) for those currently are executed
in the following order:
X(Interpreter)
E(Interpreter)
Originally, the interpreter lock is held for the execution AML opcodes, the
namespace lock is held for the named object creation AML opcodes. Since
they are actually same in MS interpreter (can all be executed during the
table loading), we can combine the 2 locks and tune the locking code better
in this way. Lv Zheng.
Link: https://bugzilla.kernel.org/show_bug.cgi?id=153541 # [1]
Link: https://bugzilla.kernel.org/show_bug.cgi?id=121701 # [1]
Link: https://bugs.acpica.org/show_bug.cgi?id=1323
Link: https://github.com/acpica/acpica/commit/767ee533
Reported-and-tested-by: Mika Westerberg <mika.westerberg@linux.intel.com>
Reported-and-tested-by: Greg White <gwhite@kupulau.com>
Reported-and-tested-by: Dutch Guy <lucht_piloot@gmx.net>
Signed-off-by: Lv Zheng <lv.zheng@intel.com>
Signed-off-by: Bob Moore <robert.moore@intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
74f51b80a0