On Mon, May 18, 2020 at 11:08 AM James Morse <james.morse@xxxxxxx> wrote: > > Hi guys, > > On 12/05/2020 19:05, Dan Williams wrote: > > On Tue, May 12, 2020 at 9:28 AM Rafael J. Wysocki > > <rafael.j.wysocki@xxxxxxxxx> wrote: > >> Dan, > >> > >> Has this been addressed in the v2? > > > > No, this looks like a case I was concerned about, i.e. the GHES code > > is not being completely careful to avoid calling potentially sleeping > > functions with interrupts disabled. There is the nice comment that > > indicates that the fixmap should be used when ghes_notify_lock_irq() > > is held, but there seems to be no infrastructure to use / divert to > > the fixmap in the ghes_proc() path. > > ghes_map()/ghes_unmap() use the fixmap for reading the firmware provided records, > but this came through apei_read(), which claims to be IRQ and NMI safe... > > > > That needs to be reworked first. > > It seems the implementation was getting lucky before to hit the cached > > acpi_ioremap in this path under rcu_read_lock(), but it appears it > > should have always been using the fixmap. Ying, James, is my read > > correct? > > The path through this thing is pretty tortuous: The static HEST contains the address of > the pointer that firmware updates to point to CPER records when they are generated. This > pointer might be static (records are always in the same place), it might not. > > The address in the tables is static. ghes.c maps it in ghes_new(): > | rc = apei_map_generic_address(&generic->error_status_address); > > which happens before the ghes_add_timer()/request_irq()/ghes_nmi_add() stuff, so we should > always use the existing mapping. > > __ghes_peek_estatus() reads the pointer with apei_read(), which should use the mapping > from ghes_new(), then uses ghes_copy_tofrom_phys() which uses the fixmap to read the CPER > records. > > > Does apei_map_generic_address() no longer keep the GAR/address mapped? > (also possible I've totally mis-understood how ACPIs caching of mappings works!) Upon further investigation the problem appears to be that System-Memory OperationRegions are dynamically mapped at runtime for ASL code. This results in every unmap event triggering eviction from the cache and incurring synchronize_rcu_expedited(). The APEI code avoids this path by taking an extra reference at the beginning of time such that the rcu-walk through the cache at NMI time is guaranteed to both succeed, and not trigger an unmap event. So now I'm looking at whether System-Memory OperationRegions can be generically pre-mapped in a similar fashion.
