Hi Bjorn,
Thanks for your pointers! See my comments inline. Aslo, can you please have
a look at my previous patch set v4 and check how many of your comments are
already addressed there. We may want to back to it then.
https://lkml.org/lkml/2016/2/4/646
Especially patches [0-6] which handle MMCONFIG refactoring.
On 05.03.2016 05:14, Bjorn Helgaas wrote:
On Fri, Mar 04, 2016 at 02:05:56PM +0530, Jayachandran Chandrashekaran
Nair wrote:
On Fri, Mar 4, 2016 at 4:21 AM, Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
Hi Tomasz, Jayachandran, et al,
On Tue, Feb 16, 2016 at 02:53:31PM +0100, Tomasz Nowicki wrote:
From: Jayachandran C <jchandra@xxxxxxxxxxxx>
Move pci_mmcfg_list handling to a drivers/acpi/pci_mcfg.c. This is
to share the API and code with ARM64 later. The corresponding
declarations are moved from asm/pci_x86.h to linux/pci-acpi.h
As a part of this we introduce three functions that can be
implemented by the arch code: pci_mmconfig_map_resource() to map a
mcfg entry, pci_mmconfig_unmap_resource to do the corresponding
unmap and pci_mmconfig_enabled to see if the arch setup of
mcfg entries was successful. We also provide weak implementations
of these, which will be used from ARM64. On x86, we retain the
old logic by providing platform specific implementation.
This patch is purely rearranging code, it should not have any
impact on the logic of MCFG parsing or list handling.
I definitely want to figure out how to make this work well on ARM64.
I need to ponder this some more, so these are just some initial
thoughts.
My first impression is that (a) the x86 MCFG code is an unmitigated
disaster, and (b) we're trying a little too hard to make that mess
generic. I think we might be better served if we came up with some
cleaner, more generic code that we can use for ARM64 today, and
migrate x86 toward that over time.
My concern is that if we elevate the current x86 code to be
"arch-independent", we will be perpetuating some interfaces and
designs that shouldn't be allowed to escape arch/x86.
I think the major decision is whether to maintain the pci_mmcfg_list
for all architectures or not. My initial plan was not to do this because
of the mess (basically the ECAM region info should be attached to
the pci root and not maintained in a separate list that needs locking),
The patch I posted initially https://patchwork.ozlabs.org/patch/553464/
had a much simpler way of handling the MCFG table without using
the list.
I agree that ECAM info should be attached to the PCI host controller.
That should simplify locking and hot-add and hot-removal of host
controllers.
I think pci_mmcfg_list is an implementation detail that may not need
to be generic. I certainly don't think it needs to be part of the
interface.
In x86 case it is not feasible to remove using the pci_mmcfg_list.
The only use of it outside is in xen that can be fixed up.
Some of the code that moved to drivers/acpi/pci_mcfg.c is not really
ACPI-specific, and could potentially be used for non-ACPI bridges that
support ECAM. I'd like to see that sort of code moved to a new file
like drivers/pci/ecam.c.
There's a little bit of overlap here with the ECAM code in
pci-host-generic.c. I'm not sure whether or how to include that, but
it's a very good example of how simple this *should* be: probe the
host bridge, discover the ECAM region, request the region, ioremap it,
done.
I had a similar approach in my initial patchset, please see the patch
above. The resource for ECAM is mapped similar to the the way
pci-host-generic.c handled it. An additional step I could do was to
move the common code (ioremap and mapbus) into a common
file and share the code with pci-host-generic.c
diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c
new file mode 100644
index 0000000..ea84365
--- /dev/null
+++ b/drivers/acpi/pci_mcfg.c
...
+int __weak pci_mmconfig_map_resource(struct device *dev,
+ struct pci_mmcfg_region *mcfg)
+{
+ struct resource *tmp;
+ void __iomem *vaddr;
+
+ tmp = insert_resource_conflict(&iomem_resource, &mcfg->res);
+ if (tmp) {
+ dev_warn(dev, "MMCONFIG %pR conflicts with %s %pR\n",
+ &mcfg->res, tmp->name, tmp);
+ return -EBUSY;
+ }
I think this is a mistake in the x86 MCFG support that we should not
carry over to a generic implementation. We should not use the MCFG
table for resource reservation because MCFG is not defined by the ACPI
spec and an OS need not include support for it. The platform must
indicate in some other, more generic way, that ECAM space is reserved.
This probably means ECAM space should be declared in a PNP0C02 _CRS
method (see the PCI Firmware Spec r3.0, sec 4.1.2, note 2).
We might need some kind of x86-specific quirk that does this, or a
pcibios hook or something here; I just don't think it should be
generic.
+int __init pci_mmconfig_parse_table(void)
+{
+ return acpi_sfi_table_parse(ACPI_SIG_MCFG, pci_parse_mcfg);
+}
I don't like the fact that we parse the entire MCFG table at once
here. I think we should look for the information we need when we are
claiming a PCI host bridge, e.g., in acpi_pci_root_add(). This might
not be a great fit for the way ACPI table management works, but I
think it's better to do things on-demand rather than just-in-case.
There is an overhead of looking up this table, and the information
available there is very limited (i.e, segment, start_bus, end_bus
and address). My approach in the above patch is to save this info
into an array at boot time and avoid multiple lookups.
We need to look up MCFG info once per host bridge, so I don't think
there's any performance issue here. But we do use acpi_table_parse(),
which is __init, and *that* is a reason why we might need to parse the
entire MCFG at boot-time. But this is the least of our worries in any
case.
diff --git a/include/linux/pci-acpi.h b/include/linux/pci-acpi.h
index 89ab057..e9450ef 100644
--- a/include/linux/pci-acpi.h
+++ b/include/linux/pci-acpi.h
@@ -106,6 +106,39 @@ extern const u8 pci_acpi_dsm_uuid[];
#define RESET_DELAY_DSM 0x08
#define FUNCTION_DELAY_DSM 0x09
+/* common API to maintain list of MCFG regions */
+/* "PCI MMCONFIG %04x [bus %02x-%02x]" */
+#define PCI_MMCFG_RESOURCE_NAME_LEN (22 + 4 + 2 + 2)
+
+struct pci_mmcfg_region {
+ struct list_head list;
+ struct resource res;
+ u64 address;
+ char __iomem *virt;
+ u16 segment;
+ u8 start_bus;
+ u8 end_bus;
+ char name[PCI_MMCFG_RESOURCE_NAME_LEN];
+};
+
+extern int pci_mmconfig_insert(struct device *dev, u16 seg, u8 start,
u8 end,
+ phys_addr_t addr);
+extern int pci_mmconfig_delete(u16 seg, u8 start, u8 end);
+
+extern struct pci_mmcfg_region *pci_mmconfig_lookup(int segment, int
bus);
+extern struct pci_mmcfg_region *pci_mmconfig_add(int segment, int
start,
+ int end, u64
addr);
+extern int pci_mmconfig_map_resource(struct device *dev,
+ struct pci_mmcfg_region *mcfg);
+extern void pci_mmconfig_unmap_resource(struct pci_mmcfg_region
*mcfg);
+extern int pci_mmconfig_enabled(void);
+extern int __init pci_mmconfig_parse_table(void);
+
+extern struct list_head pci_mmcfg_list;
+
+#define PCI_MMCFG_BUS_OFFSET(bus) ((bus) << 20)
+#define PCI_MMCFG_OFFSET(bus, devfn) ((bus) << 20 | (devfn) << 12)
+
With the exception of pci_mmconfig_parse_table(), nothing here is
ACPI-specific. I'd like to see the PCI ECAM-related interfaces
(hopefully not these exact ones, but a more rational set) put in
something like include/linux/pci-ecam.h.
#else /* CONFIG_ACPI */
static inline void acpi_pci_add_bus(struct pci_bus *bus) { }
static inline void acpi_pci_remove_bus(struct pci_bus *bus) { }
I can update this patch to
- drop the pci_mmcfg_list handling from generic case
- move common ECAM code so that it can be shared with
pci-host-generic.c
if that is what you are looking for. The code will end up looking much
simpler.
I think we should ignore x86 mmconfig for now. It is absurdly
complicated and I'm not sure it's fixable. I *do* want to keep
drivers/acpi/pci_root.c for all ACPI host bridges, including x86,
ia64, and arm64.
So I think we should write generic MCFG and ECAM support from scratch
for arm64. Something like this:
- Add an acpi_mcfg_init(), maybe in drivers/acpi/pci_mcfg.c, to be
called from acpi_init() to copy MCFG info to something we can
access after __init. This would not reserve resources, but
probably does have to ioremap() the regions to support
raw_pci_read().
As said, ECAM and ACPI specific code was isolated in previous patch. There,
I tried to leave x86 complication in arch/x86/ and extract generic
functionalities to driver/pci/ecam.c as the library.
- Implement raw_pci_read(), which is "special" because ACPI needs it
for PCI config access from AML. It's supposed to be "always
accessible" and we don't have a struct pci_bus *, so this probably
has to use the MCFG copy and the ioremap done above. Maybe it
should go in the same file. This is completely independent of
the PCI core and PCI data structures.
We were looking for the answer which would justify RAW PCI config accessors
being for ARM64 world. Unfortunately, nobody was able to show real use case
for ARM64. Do you see the reason we need this? Our conclusion was to leave
it empty for ARM64 which in turn makes code simpler. I am not ASWG member
while that was under discussion so I will ask Lorenzo to elaborate more on
this.
- Implement arm64 pci_acpi_scan_root() that calls
acpi_pci_root_create() with an .init_info() function that calls
acpi_pci_root_get_mcfg_addr() to read _CBA, and if that fails,
looks up the bus range in the MCFG copy from above. It should
call request_mem_region(). For a region from _CBA, it should call
ioremap(). For regions from MCFG it can probably use the ioremap
done by acpi_mcfg_init().
Yes, Expanding .init_info() to check for _CBA is good point.
I know acpi_pci_root_add() calls acpi_pci_root_get_mcfg_addr()
before calling pci_acpi_scan_root(), but I think that's wrong
because (a) some arches, e.g., ia64, don't use ECAM and (b) _CBA
and MCFG should be handled in the same place.
I know calling request_mem_region() here will probably be an
ordering problem because the PNP0C02 driver hasn't reserved
resources yet. But the host bridge driver is using the region and
it should reserve it.
- If we store the ECAM mapped base address in the sysdata or struct
pci_host_bridge, the normal config accessors can use
pci_generic_config_read() with a new generic .map_bus() function.
pci_generic_config_{read|write}() is what we want to use, actually we do
now, but ECAM region and sysdata association will remove ECAM region lookup
step (see patch 09/15 of this series).
On x86, the normal config access path is:
pci_read(struct pci_bus *, ...)
raw_pci_read(seg, bus#, ...)
raw_pci_ext_ops->read(seg, bus#, ...)
pci_mmcfg_read(seg, bus#, ...)
pci_dev_base
pci_mmconfig_lookup(seg, bus#)
I think this is somewhat backwards because we start with a pci_bus
pointer, so we *could* have a nice simple bus-specific accessor,
but we throw that pointer away, so pci_mmcfg_read() has to start
over and look up the ECAM offset from scratch, which makes it all
unnecessarily complicated.
As you pointed out raw_pci_{read|write} make things complicated, so IMO we
should either say they are absolutely necessary (and then think how to
simplify it) or just use simple bus-specific accessor (patch 02/15) e.g. for
ARM64.
