Re: [PATCH v5] ACPICA: Tables: Add mechanism to allow to balance late stage acpi_get_table() independently

From: Dan Williams
Date: Wed Jun 07 2017 - 17:24:56 EST


On Wed, Jun 7, 2017 at 2:14 PM, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
> On Wed, Jun 7, 2017 at 8:41 AM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
>> On Tue, Jun 6, 2017 at 9:54 PM, Lv Zheng <lv.zheng@xxxxxxxxx> wrote:
>>> Considering this case:
>>> 1. A program opens a sysfs table file 65535 times, it can increase
>>> validation_count and first increment cause the table to be mapped:
>>> validation_count = 65535
>>> 2. AML execution causes "Load" to be executed on the same table, this time
>>> it cannot increase validation_count, so validation_count remains:
>>> validation_count = 65535
>>> 3. The program closes sysfs table file 65535 times, it can decrease
>>> validation_count and the last decrement cause the table to be unmapped:
>>> validation_count = 0
>>> 4. AML code still accessing the loaded table, kernel crash can be observed.
>>>
>>> This is because orginally ACPICA doesn't support unmapping tables during
>>> OS late stage. So the current code only allows unmapping tables during OS
>>> early stage, and for late stage, no acpi_put_table() clones should be
>>> invoked, especially cases that can trigger frequent invocations of
>>> acpi_get_table()/acpi_put_table() are forbidden:
>>> 1. sysfs table accesses
>>> 2. dynamic Load/Unload opcode executions
>>> 3. acpi_load_table()
>>> 4. etc.
>>> Such frequent acpi_put_table() balance changes have to be done altogether.
>>>
>>> This philosophy is not convenient for Linux driver writers. Since the API
>>> is just there, developers will start to use acpi_put_table() during late
>>> stage. So we need to consider a better mechanism to allow them to safely
>>> invoke acpi_put_table().
>>>
>>> This patch provides such a mechanism by adding a validation_count
>>> threashold. When it is reached, the validation_count can no longer be
>>> incremented/decremented to invalidate the table descriptor (means
>>> preventing table unmappings) so that acpi_put_table() balance changes can be
>>> done independently to each others.
>>>
>>> Note: code added in acpi_tb_put_table() is actually a no-op but changes the
>>> warning message into a warning once message. Lv Zheng.
>>>
>>
>> This still seems to be unnecessary gymnastics to keep the validation
>> count around and make it work for random drivers.
>
> Well, I'm not sure I agree here.
>
> If we can make it work at one point, it should not be too hard to
> maintain that status.
>

I agree with that, my concern was with driver writers needing to be
worried about when it is safe to call acpi_put_table(). This reference
count behaves differently than other reference counts like kobjects.
The difference is not necessarily bad, but hopefully it can be
contained within the acpi core.