On Wednesday, November 28, 2012 01:31:39 PM Toshi Kani wrote: > On Wed, 2012-11-28 at 19:28 +0100, Rafael J. Wysocki wrote: > > On Wednesday, November 28, 2012 09:54:43 AM Toshi Kani wrote: > > > > > > > > By using acpi_install_notify_handler(), each driver needs to > > > > > > > > walk > > > > > > > > through the entire ACPI namespace to find its associated ACPI > > > > > > > > devices > > > > > > > > and call it to register one by one. I think this is more work > > > > > > > > for > > > > > > > > non-ACPI drivers than defining acpi_driver. > > > > > > > > > > > > > > I'm not really sure what you mean. The drivers in question > > > > > > > already know > > > > > > > what the relevant ACPI device nodes are (because they need them > > > > > > > anyway > > > > > > > for other purposes), so they don't need to look for them > > > > > > > specifically and > > > > > > > acpi_install_notify_handler() doesn't do any namespace walking. > > > > > > > So what > > > > > > > you said above simply doesn't make sense from this viewpoint. > > > > > > > > > > > > Yes, if drivers already know the relevant ACPI devices, then > > > > > > walking the > > > > > > ACPI namespace is not necessary. I was referring the case like > > > > > > processor_driver.c, acpi_memhotplug.c, and container.c in my > > > > > > statement. > > > > > > > > > > BTW, when an ACPI device is marked as non-present, which is the case > > > > > before hot-add, we do not create an acpi_device object and therefore > > > > > do > > > > > not bind it with a driver. This is why these drivers walk the ACPI > > > > > namespace and install their notify handlers regardless of device > > > > > status. > > > > > > > > So maybe we should create struct acpi_device objects in that case too? > > > > > > I think it has some challenge as well. We bind an ACPI driver with > > > device_register(), which calls device_add()-> kobject_add(). So, all > > > non-present ACPI device objects will show up in sysfs, unless we can > > > change the core. This will change user interface. There can be quite > > > many non-present devices in ACPI namespace depending on FW > > > implementation. > > > > If additional devices appear in sysfs, that's not a problem. If there > > were fewer of them, that would be a real one. :-) > > I see. I guess this means that once we expose all non-present devices > in sysfs, we cannot go back to the current way. So, we need to be very > careful. Anyway, this model requires separate handling for static ACPI > [1] and dynamic ACPI [2], which may make the state model complicated. > > 1. Static ACPI - No creation / deletion of acpi_device at hot-plug. > 2. Dynamic ACPI - Create acpi_device at hot-add, delete at hot-remove.
Sure. The complication here is that we'll need to handle the removal of a bunch of struct acpi_device objects when a whole table goes away. However, first, we don't seem to handle table unloading now. At least acpi_unload_parent_table() is not called from anywhere as far as I can say. Second, we'll need to handle the removal of struct acpi_device objects _anyway_ on table unload, this way or another. > [1] ACPI namespace is static and contains the maximum possible config. > [2] ACPI namespace is dynamic. SSDT is loaded at hot-add, and unloaded > at hot-remove. Thanks, Rafael -- I speak only for myself. Rafael J. Wysocki, Intel Open Source Technology Center. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
