> From: linux-acpi-owner@xxxxxxxxxxxxxxx > [mailto:linux-acpi-owner@xxxxxxxxxxxxxxx] On Behalf Of Zheng, Lv > Sent: Friday, July 26, 2013 9:54 AM > To: Rafael J. Wysocki > Cc: Wysocki, Rafael J; Brown, Len; linux-kernel@xxxxxxxxxxxxxxx; > linux-acpi@xxxxxxxxxxxxxxx > Subject: RE: [PATCH 06/13] ACPI/IPMI: Add reference counting for ACPI > operation region handlers > > > From: Rafael J. Wysocki [mailto:rjw@xxxxxxx] > > Sent: Friday, July 26, 2013 5:29 AM > > > > On Tuesday, July 23, 2013 04:09:43 PM Lv Zheng wrote: > > > This patch adds reference couting for ACPI operation region handlers > > > to fix races caused by the ACPICA address space callback invocations. > > > > > > ACPICA address space callback invocation is not suitable for Linux > > > CONFIG_MODULE=y execution environment. > > > > Actually, can you please explain to me what *exactly* the problem is? > > OK. I'll add race explanations in the next revision. > > The problem is there is no "lock" held inside ACPICA for invoking operation > region handlers. > Thus races happens between the acpi_remove/install_address_space_handler > and the handler/setup callbacks. This seems not a good explanation of the intent of this patch. I think the intent is here in the patch description: 1. It acts as a barrier for operation region callbacks - no callback will happen after acpi_unregister_region(). 2. acpi_unregister_region() is safe to be called in moudle->exit() functions. Hmm, maybe I need to re-order the patch description for this patch. Thanks for commenting. Best regards -Lv > > This is correct per ACPI specification. > As if there is interpreter locks held for invoking operation region handlers, the > timeout implemented inside the operation region handlers will make all locking > facilities (Acquire or Sleep,...) timed out. > Please refer to ACPI specification "5.5.2 Control Method Execution": > Interpretation of a Control Method is not preemptive, but it can block. When a > control method does block, OSPM can initiate or continue the execution of a > different control method. A control method can only assume that access to > global objects is exclusive for any period the control method does not block. > > So it is pretty much likely that ACPI IO transfers are locked inside the operation > region callback implementations. > Using locking facility to protect the callback invocation will risk dead locks. > > Thanks > -Lv > > > Rafael ��.n��������+%������w��{.n�����{�����ܨ}���Ơz�j:+v�����w����ޙ��&�)ߡ�a����z�ޗ���ݢj��w�f
