Hi, On 10/4/22 10:24, Andy Shevchenko wrote: > On Mon, Oct 03, 2022 at 04:42:14PM +0200, Hans de Goede wrote: >> ACPI-2.0 says that the EC OpRegion handler must be available immediately >> (like the standard default OpRegion handlers): >> >> Quoting from the ACPI spec version 6.3: "6.5.4 _REG (Region) ... > > (Side note: As of today the revision 6.5 of the ACPI specification is > available) > >> 2. OSPM must make Embedded Controller operation regions, accessed via >> the Embedded Controllers described in ECDT, available before executing >> any control method. These operation regions may become inaccessible >> after OSPM runs _REG(EmbeddedControl, 0)." >> >> So acpi_bus_init() calls acpi_ec_ecdt_probe(), which calls >> acpi_install_address_space_handler() to install the EC's OpRegion >> handler, early on. >> >> This not only installs the OpRegion handler, but also calls the EC's >> _REG method. The _REG method call is a problem because it may rely on >> initialization done by the _INI methods of one of the PCI / _SB root devs, >> see for example: https://bugzilla.kernel.org/show_bug.cgi?id=214899 . >> >> Generally speaking _REG methods are executed when the ACPI-device they >> are part of has a driver bound to it. Where as _INI methods must be >> executed at table load time (according to the spec). The problem here >> is that the early acpi_install_address_space_handler() call causes >> the _REG handler to run too early. >> >> To allow fixing this the ACPICA code now allows to split the OpRegion >> handler installation and the executing of _REG into 2 separate steps. >> >> This commit uses this ACPICA functionality to fix the EC probe ordering >> by delaying the executing of _REG for ECDT described ECs till the matching >> EC device in the DSDT gets parsed and acpi_ec_add() for it gets called. >> This moves the calling of _REG for the EC on devices with an ECDT to >> the same point in time where it is called on devices without an ECDT table. > > ... > >> + if (call_reg && !test_bit(EC_FLAGS_EC_REG_CALLED, &ec->flags)) { >> + acpi_execute_reg_methods(ec->handle, ACPI_ADR_SPACE_EC); > >> + set_bit(EC_FLAGS_EC_REG_CALLED, &ec->flags); > > Just to be sure: do you need atomic operation here? Does it prevent of any > kind of races? Because if it had been the case, it would have been done via > test_and_set_bit() rather than testing and setting separated. AFAIK we do not need atomic operations here (I did not verify this). Also note that existing code a couple of lines above already uses the same pattern. Regards, Hans
