Re: [PATCH v5 6/7] acpi/ghes: add support for generic error injection via QAPI

From: Igor Mammedov
Date: Thu Aug 08 2024 - 04:11:24 EST

On Wed, 7 Aug 2024 15:25:47 +0100
Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote:

> On Tue, 6 Aug 2024 16:31:13 +0200
> Igor Mammedov <imammedo@xxxxxxxxxx> wrote:
>
> > On Fri, 2 Aug 2024 23:44:01 +0200
> > Mauro Carvalho Chehab <mchehab+huawei@xxxxxxxxxx> wrote:
> >
> > > Provide a generic interface for error injection via GHESv2.
> > >
> > > This patch is co-authored:
> > > - original ghes logic to inject a simple ARM record by Shiju Jose;
> > > - generic logic to handle block addresses by Jonathan Cameron;
> > > - generic GHESv2 error inject by Mauro Carvalho Chehab;
> > >
> > > Co-authored-by: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>
> > > Co-authored-by: Shiju Jose <shiju.jose@xxxxxxxxxx>
> > > Co-authored-by: Mauro Carvalho Chehab <mchehab+huawei@xxxxxxxxxx>
> > > Cc: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>
> > > Cc: Shiju Jose <shiju.jose@xxxxxxxxxx>
> > > Signed-off-by: Mauro Carvalho Chehab <mchehab+huawei@xxxxxxxxxx>
> > > ---
> > > hw/acpi/ghes.c | 159 ++++++++++++++++++++++++++++++++++++++---
> > > hw/acpi/ghes_cper.c | 2 +-
> > > include/hw/acpi/ghes.h | 3 +
> > > 3 files changed, 152 insertions(+), 12 deletions(-)
> > >
> > > diff --git a/hw/acpi/ghes.c b/hw/acpi/ghes.c
> > > index a745dcc7be5e..e125c9475773 100644
> > > --- a/hw/acpi/ghes.c
> > > +++ b/hw/acpi/ghes.c
> > > @@ -395,23 +395,22 @@ void acpi_ghes_add_fw_cfg(AcpiGhesState *ags, FWCfgState *s,
> > > ags->present = true;
> > > }
> > >
> > > +static uint64_t ghes_get_state_start_address(void)
> >
> > ghes_get_hardware_errors_address() might better reflect what address it will return
> >
> > > +{
> > > + AcpiGedState *acpi_ged_state =
> > > + ACPI_GED(object_resolve_path_type("", TYPE_ACPI_GED, NULL));
> > > + AcpiGhesState *ags = &acpi_ged_state->ghes_state;
> > > +
> > > + return le64_to_cpu(ags->ghes_addr_le);
> > > +}
> > > +
> > > int acpi_ghes_record_errors(uint8_t source_id, uint64_t physical_address)
> > > {
> > > uint64_t error_block_addr, read_ack_register_addr, read_ack_register = 0;
> > > - uint64_t start_addr;
> > > + uint64_t start_addr = ghes_get_state_start_address();
> > > bool ret = -1;
> > > - AcpiGedState *acpi_ged_state;
> > > - AcpiGhesState *ags;
> > > -
> > > assert(source_id < ACPI_HEST_SRC_ID_RESERVED);
> > >
> > > - acpi_ged_state = ACPI_GED(object_resolve_path_type("", TYPE_ACPI_GED,
> > > - NULL));
> > > - g_assert(acpi_ged_state);
> > > - ags = &acpi_ged_state->ghes_state;
> > > -
> > > - start_addr = le64_to_cpu(ags->ghes_addr_le);
> > > -
> > > if (physical_address) {
> > > start_addr += source_id * sizeof(uint64_t);
> >
> > above should be a separate patch
> >
> > >
> > > @@ -448,9 +447,147 @@ int acpi_ghes_record_errors(uint8_t source_id, uint64_t physical_address)
> > > return ret;
> > > }
> > >
> > > +/*
> > > + * Error register block data layout
> > > + *
> > > + * | +---------------------+ ges.ghes_addr_le
> > > + * | |error_block_address0 |
> > > + * | +---------------------+
> > > + * | |error_block_address1 |
> > > + * | +---------------------+ --+--
> > > + * | | ............. | GHES_ADDRESS_SIZE
> > > + * | +---------------------+ --+--
> > > + * | |error_block_addressN |
> > > + * | +---------------------+
> > > + * | | read_ack0 |
> > > + * | +---------------------+ --+--
> > > + * | | read_ack1 | GHES_ADDRESS_SIZE
> > > + * | +---------------------+ --+--
> > > + * | | ............. |
> > > + * | +---------------------+
> > > + * | | read_ackN |
> > > + * | +---------------------+ --+--
> > > + * | | CPER | |
> > > + * | | .... | GHES_MAX_RAW_DATA_LENGT
> > > + * | | CPER | |
> > > + * | +---------------------+ --+--
> > > + * | | .......... |
> > > + * | +---------------------+
> > > + * | | CPER |
> > > + * | | .... |
> > > + * | | CPER |
> > > + * | +---------------------+
> > > + */
> >
> > no need to duplicate docs/specs/acpi_hest_ghes.rst,
> > I'd just reffer to it and maybe add short comment as to why it's mentioned.
> >
> > > +/* Map from uint32_t notify to entry offset in GHES */
> > > +static const uint8_t error_source_to_index[] = { 0xff, 0xff, 0xff, 0xff,
> > > + 0xff, 0xff, 0xff, 1, 0};
> > > +
> > > +static bool ghes_get_addr(uint32_t notify, uint64_t *error_block_addr,
> > > + uint64_t *read_ack_addr)
> > > +{
> > > + uint64_t base;
> > > +
> > > + if (notify >= ACPI_GHES_NOTIFY_RESERVED) {
> > > + return false;
> > > + }
> > > +
> > > + /* Find and check the source id for this new CPER */
> > > + if (error_source_to_index[notify] == 0xff) {
> > > + return false;
> > > + }
> > > +
> > > + base = ghes_get_state_start_address();
> > > +
> > > + *read_ack_addr = base +
> > > + ACPI_GHES_ERROR_SOURCE_COUNT * sizeof(uint64_t) +
> > > + error_source_to_index[notify] * sizeof(uint64_t);
> > > +
> > > + /* Could also be read back from the error_block_address register */
> > > + *error_block_addr = base +
> > > + ACPI_GHES_ERROR_SOURCE_COUNT * sizeof(uint64_t) +
> > > + ACPI_GHES_ERROR_SOURCE_COUNT * sizeof(uint64_t) +
> > > + error_source_to_index[notify] * ACPI_GHES_MAX_RAW_DATA_LENGTH;
> > > +
> > > + return true;
> > > +}
> >
> > I don't like all this pointer math, which is basically a reverse engineered
> > QEMU actions on startup + guest provided etc/hardware_errors address.
> >
> > For once, it assumes error_source_to_index[] matches order in which HEST
> > error sources were described, which is fragile.
> >
> > 2nd: migration-wive it's disaster, since old/new HEST/hardware_errors tables
> > in RAM migrated from older version might not match above assumptions
> > of target QEMU.
> >
> > I see 2 ways to rectify it:
> > 1st: preferred/cleanest would be to tell QEMU (via fw_cfg) address of HEST table
> > in guest RAM, like we do with etc/hardware_errors, see
> > build_ghes_error_table()
> > ...
> > tell firmware to write hardware_errors GPA into
> > and then fetch from HEST table in RAM, the guest patched error/ack addresses
> > for given source_id
> >
> > code-wise: relatively simple once one wraps their own head over
> > how this whole APEI thing works in QEMU
> > workflow is described in docs/specs/acpi_hest_ghes.rst
> > look to me as sufficient to grasp it.
> > (but my view is very biased given my prior knowledge,
> > aka: docs/comments/examples wrt acpi patching are good enough)
> > (if it's not clear how to do it, ask me for pointers)
>
> Hi Igor, I think I follow what you mean but maybe this question will reveal
> otherwise. HEST is currently in ACPI_BUILD_TABLE_FILE.
> Would you suggest splitting it to it's own file, or using table_offsets
> to get the offset in ACPI_BUILD_TABLE_FILE GPA?
yep, offset taken right before HEST is to be created
doc comment for bios_linker_loader_write_pointer() explains how it works

we need something like:
bios_linker_loader_write_pointer(linker,
ACPI_HEST_TABLE_ADDR_FW_CFG_FILE, 0, sizeof(uint64_t),
ACPI_BUILD_TABLE_FILE, hest_offset_within_ACPI_BUILD_TABLE_FILE);

to register new file see:
a08a64627 ACPI: Record the Generic Error Status Block address
and to avoid copy past error maybe
136fc6aa2 ACPI: Avoid infinite recursion when dump-vmstat
for this needs to be limited to new machine types and keep
old ones without this new feature. (I'd use hw_compat_ machinery for that)

while at it we should rename
ACPI_GHES_DATA_ADDR_FW_CFG_FILE -> ACPI_GHES_ERRORS_ADDR_FW_CFG_FILE