Re: [PATCH v2] Revert "ACPI: Add memory semantics to acpi_os_map_memory()"

[Lorenzo Pieralisi]

On Wed, Sep 22, 2021 at 01:11:26PM +0200, Ard Biesheuvel wrote:
> On Tue, 21 Sept 2021 at 12:05, Lorenzo Pieralisi
> <lorenzo.pieralisi@xxxxxxx> wrote:
> >
> > On Mon, Sep 20, 2021 at 07:32:56PM +0200, Rafael J. Wysocki wrote:
> > > On Mon, Sep 20, 2021 at 7:03 PM Lorenzo Pieralisi
> > > <lorenzo.pieralisi@xxxxxxx> wrote:
> > > >
> > > > On Thu, Sep 16, 2021 at 05:08:27PM +0100, Lorenzo Pieralisi wrote:
> > > > > On Fri, Sep 10, 2021 at 07:28:49PM +0200, Ard Biesheuvel wrote:
> > > > > > On Fri, 10 Sept 2021 at 16:32, Jia He <justin.he@xxxxxxx> wrote:
> > > > > > >
> > > > > > > This reverts commit 437b38c51162f8b87beb28a833c4d5dc85fa864e.
> > > > > > >
> > > > > > > After this commit, a boot panic is alway hit on an Ampere EMAG server
> > > > > > > with call trace as follows:
> > > > > > >  Internal error: synchronous external abort: 96000410 [#1] SMP
> > > > > > >  Modules linked in:
> > > > > > >  CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.14.0+ #462
> > > > > > >  Hardware name: MiTAC RAPTOR EV-883832-X3-0001/RAPTOR, BIOS 0.14 02/22/2019
> > > > > > >  pstate: 60000005 (nZCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> > > > > > > [...snip...]
> > > > > > >  Call trace:
> > > > > > >   acpi_ex_system_memory_space_handler+0x26c/0x2c8
> > > > > > >   acpi_ev_address_space_dispatch+0x228/0x2c4
> > > > > > >   acpi_ex_access_region+0x114/0x268
> > > > > > >   acpi_ex_field_datum_io+0x128/0x1b8
> > > > > > >   acpi_ex_extract_from_field+0x14c/0x2ac
> > > > > > >   acpi_ex_read_data_from_field+0x190/0x1b8
> > > > > > >   acpi_ex_resolve_node_to_value+0x1ec/0x288
> > > > > > >   acpi_ex_resolve_to_value+0x250/0x274
> > > > > > >   acpi_ds_evaluate_name_path+0xac/0x124
> > > > > > >   acpi_ds_exec_end_op+0x90/0x410
> > > > > > >   acpi_ps_parse_loop+0x4ac/0x5d8
> > > > > > >   acpi_ps_parse_aml+0xe0/0x2c8
> > > > > > >   acpi_ps_execute_method+0x19c/0x1ac
> > > > > > >   acpi_ns_evaluate+0x1f8/0x26c
> > > > > > >   acpi_ns_init_one_device+0x104/0x140
> > > > > > >   acpi_ns_walk_namespace+0x158/0x1d0
> > > > > > >   acpi_ns_initialize_devices+0x194/0x218
> > > > > > >   acpi_initialize_objects+0x48/0x50
> > > > > > >   acpi_init+0xe0/0x498
> > > > > > >
> > > > > > > As mentioned by Lorenzo:
> > > > > > >   "We are forcing memory semantics mappings to PROT_NORMAL_NC, which
> > > > > > >   eMAG does not like at all and I'd need to understand why. It looks
> > > > > > >   like the issue happen in SystemMemory Opregion handler."
> > > > > > >
> > > > > > > Hence just revert it before everything is clear.
> > > > > > >
> > > > > >
> > > > > > Can we try to find the root cause first? -rc1 is not even out yet, and
> > > > > > reverting it now means we can not resubmit it until the next merge
> > > > > > window.
> > > > >
> > > > > I am waiting to debug this on an eMAG but I noticed something that
> > > > > I wanted to bring up.
> > > > >
> > > > > SystemMemory Operation region handler - ie
> > > > >
> > > > > acpi_ex_system_memory_space_handler()
> > > > >
> > > > > maps the Operation Region (that AFAICS is MMIO, it is _not_ memory)
> > > > > with acpi_os_map_memory() and I believe that's what is causing this
> > > > > bug.
> > > > >
> > > > > On the other hand, acpi_os_map_generic_address(), to handle spaceid
> > > > > ACPI_ADR_SPACE_SYSTEM_MEMORY, uses acpi_os_map_iomem() that is more
> > > > > in line with my expectations.
> > > >
> > > > Hi Rafael,
> > > >
> > > > I wanted to ask please if you have any insights on why
> > > >
> > > > (1) acpi_ex_system_memory_space_handler()
> > > > (2) acpi_os_map_generic_address()
> > > >
> > > > Use two different calls to map memory for the _same_ address space ID
> > > > (SystemMemory).
> > > >
> > > > (3) acpi_os_map_memory()
> > > > vs
> > > > (4) acpi_os_map_iomem()
> > >
> > > I don't really have a good answer here.
> > >
> > > On x86 this doesn't really matter and that's where
> > > acpi_ex_system_memory_space_handler() was first introduced.  It is not
> > > only used for IOMEM (there are SystemMemory operation regions in RAM),
> > > but since it may be in IOMEM, it should assume so.
> > >
> > > > I am struggling to understand why (1) uses (3) ("memory semantics") when
> > > > (2) uses (4) - it is actually unclear how the distinction between
> > > > the two mapping APIs is to be drawn and on what basis one should
> > > > choose which one to use.
> > > >
> > > > I am still waiting to grab some HW to debug this report but the issue
> > > > here is that we are mapping an OpRegion SystemMemory with (3) in the
> > > > memory space handler and given the patch we are reverting we end up
> > > > mapping the operation region with normal non-cacheable memory attributes
> > > > that probably the physical address range behind the OpRegion does not
> > > > support.
> > >
> > > If that is the case, there needs to be a mechanism to decide what kind
> > > of mapping to use for SystemMemory operation regions based on the type
> > > of physical memory the address range in question is located in.
> >
> > Thank you Rafael. The mechanism we are currently relying on is the EFI
> > memory map but if the Opregion address is not described there then we
> > are left with a default choice to make (theoretically I may also parse
> > all _CRS in the namespace to find whether a resource include the
> > Opregion and I may infer attributes from the _CRS resource entry).
> >
> 
> I'm not sure that would help, as I would expected the memory described
> by _CRS to be mostly mutually exclusive from memory used by OpRegions.
> 
> > Maybe we should update the ACPI specs to enforce it; with current
> > firmware the idea of using the OS expected *usage* of memory (ie
> > memory vs IO) described by the mapping function prototype can't work
> > as this revert shows (even though it would be better if I manage
> > to find what the precise issue is).
> >
> > We can't map something with specific attributes if we don't know
> > whether the physical address space backing the region supports it.
> >
> 
> We don't have a a safe default in either direction, so I agree this is
> a hole in the specs.
> 
> > I am left with little choice: I assume the best thing I could do
> > to fix the original bug is to use ioremap_* in acpi_data_show()
> > instead of acpi_os_map/unmap_memory() to map that memory with
> > specific attributes (for BERT error regions, they must be RAM
> > so, _hopefully_, we know it can be mapped with eg normal memory
> > mappings).
> >
> > Thoughts ?
> >
> 
> One thing I just realized is that the EFI memory map is not a complete
> solution to begin with, as it may not cover hot/coldplugged memory
> regions that are only described via ACPI.
> 
> Did you make any progress with the eMAG?

I manage to get the ACPI tables dump. The fault is triggered on
a SystemMemory OPregion access (FYI - should be a reset register),
probably (but on this only Ampere can help us) because the MMIO
range in question does not support the AXI attributes assigned
by the NormalNC mapping.

I believe mapping SystemMemory Opregions as NormalNC does not make
much sense anyway.

The UEFI specs seem to hint that the ACPI Op-region cacheability
attributes must be determined through the UEFI memory map, not
sure whether that means that the OpRegion itself _must_ be in
the EFI memory map.

I believe we need to go on with the revert and find a way to fix the
BERT error region mappings, to make them NormalNC so that we can do
unaligned accesses on them.

What to do specs side - to be debated, we have to do something because
it is impossible to handle it sensibly otherwise.

Lorenzo