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

["Zheng, Lv" <lv.zheng@intel.com>]

Hi, Dan

> From: Dan Williams [mailto:dan.j.williams@intel.com]
> Subject: Re: [PATCH v5] ACPICA: Tables: Add mechanism to allow to balance late stage acpi_get_table()
> independently
> 
> On Wed, Jun 7, 2017 at 2:14 PM, Rafael J. Wysocki <rafael@kernel.org> wrote:
> > On Wed, Jun 7, 2017 at 8:41 AM, Dan Williams <dan.j.williams@intel.com> wrote:
> >> On Tue, Jun 6, 2017 at 9:54 PM, Lv Zheng <lv.zheng@intel.com> 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.

I don't think they behave differently.

"kref" needn't consider unbalanced "get/put".
Because when the drivers(users) are deploying "kref",
they are responsible for ensuring balanced "get/put".
"kref" needn't take too much care about "overflow/underflow"
as if all users ensure balanced "get/put",
"overflow/underflow" is not possible.
Occurrence of "overflow/underflow" means bugs.
And can be further captured as "panic".

If "kref" considers to "warn_once" overflow/underflow users,
the logic in this commit can also be introduced to kref.
However it's useless as all users have ensured balanced "get/put".
Putting useless check than panic on hot path could be a waste.

> The difference is not necessarily bad, but hopefully it can be
> contained within the acpi core.

The old warning logic for table desc is just derived from utdelete.c.
Which reduces communication cost when the mechanism is upstreamed.

ACPICA table "validation_count" is deployed on top of old design.
Where "table unmap" is forbidden for late stage.
Thus there is no users ensuring balanced "get/put".
Under this circumstances, when we start to deploy balanced "get/put",
we need to consider all users as a whole.
You cannot say current unbalanced "get/put" users have bugs.
They are there just because of historical reasons.

Fortunately after applying this patch,
drivers should be able to have a better environment to use the new APIs.

Cheers,
Lv