Re: [Qemu-devel] [PATCH] Virt: ACPI: fix qemu assert due to re-assigned table data address

[Andrew Jones <drjones@redhat.com>]

On Tue, Dec 26, 2017 at 07:54:15PM +0800, Shannon Zhao wrote:
> 
> 
> On 2017/12/26 19:48, Andrew Jones wrote:
> > On Fri, Dec 22, 2017 at 02:52:47PM +0800, Shannon Zhao wrote:
> >> acpi_data_push uses g_array_set_size to resize the memory size. If there is no
> >> enough contiguous memory, the address will be changed. If we use the old value,
> >> it will assert.
> >> qemu-kvm: hw/acpi/bios-linker-loader.c:214: bios_linker_loader_add_checksum:
> >> Assertion `start_offset < file->blob->len' failed.`
> >>
> >> Signed-off-by: Shannon Zhao <zhaoshenglong@huawei.com>
> >> ---
> >>  hw/arm/virt-acpi-build.c | 18 +++++++++++-------
> >>  1 file changed, 11 insertions(+), 7 deletions(-)
> >>
> >> diff --git a/hw/arm/virt-acpi-build.c b/hw/arm/virt-acpi-build.c
> >> index 3d78ff6..5901142 100644
> >> --- a/hw/arm/virt-acpi-build.c
> >> +++ b/hw/arm/virt-acpi-build.c
> >> @@ -453,6 +453,7 @@ build_spcr(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> >>      AcpiSerialPortConsoleRedirection *spcr;
> >>      const MemMapEntry *uart_memmap = &vms->memmap[VIRT_UART];
> >>      int irq = vms->irqmap[VIRT_UART] + ARM_SPI_BASE;
> >> +    int spcr_start = table_data->len;
> >>  
> >>      spcr = acpi_data_push(table_data, sizeof(*spcr));
> >>  
> >> @@ -476,8 +477,8 @@ build_spcr(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> >>      spcr->pci_device_id = 0xffff;  /* PCI Device ID: not a PCI device */
> >>      spcr->pci_vendor_id = 0xffff;  /* PCI Vendor ID: not a PCI device */
> >>  
> >> -    build_header(linker, table_data, (void *)spcr, "SPCR", sizeof(*spcr), 2,
> >> -                 NULL, NULL);
> >> +    build_header(linker, table_data, (void *)(table_data->data + spcr_start),
> >> +                 "SPCR", table_data->len - spcr_start, 2, NULL, NULL);
> >>  }
> > 
> > We don't need to change build_spcr(), as acpi_data_push() is only called
> > once, so spcr == new table_data->data + old table_data->len and new
> > table_data->len - spcr == sizeof(*spcr) (the size used in the only
> > acpi_data_push() call)
> > 
> >>  
> >>  static void
> >> @@ -512,8 +513,8 @@ build_srat(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> >>          mem_base += numa_info[i].node_mem;
> >>      }
> >>  
> >> -    build_header(linker, table_data, (void *)srat, "SRAT",
> >> -                 table_data->len - srat_start, 3, NULL, NULL);
> >> +    build_header(linker, table_data, (void *)(table_data->data + srat_start),
> >> +                 "SRAT", table_data->len - srat_start, 3, NULL, NULL);
> > 
> > Yes, we need this fix, as there are many acpi_data_push() calls in this
> > function. I guess this was the table that triggered the assert.
> > 
> >>  }
> >>  
> >>  static void
> >> @@ -522,6 +523,7 @@ build_mcfg(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> >>      AcpiTableMcfg *mcfg;
> >>      const MemMapEntry *memmap = vms->memmap;
> >>      int len = sizeof(*mcfg) + sizeof(mcfg->allocation[0]);
> >> +    int mcfg_start = table_data->len;
> >>  
> >>      mcfg = acpi_data_push(table_data, len);
> >>      mcfg->allocation[0].address = cpu_to_le64(memmap[VIRT_PCIE_ECAM].base);
> >> @@ -532,7 +534,8 @@ build_mcfg(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> >>      mcfg->allocation[0].end_bus_number = (memmap[VIRT_PCIE_ECAM].size
> >>                                            / PCIE_MMCFG_SIZE_MIN) - 1;
> >>  
> >> -    build_header(linker, table_data, (void *)mcfg, "MCFG", len, 1, NULL, NULL);
> >> +    build_header(linker, table_data, (void *)(table_data->data + mcfg_start),
> >> +                 "MCFG", len, 1, NULL, NULL);
> >>  }
> > 
> > No need to change this one.
> > 
> >>  
> >>  /* GTDT */
> >> @@ -651,6 +654,7 @@ build_madt(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> >>  static void build_fadt(GArray *table_data, BIOSLinker *linker,
> >>                         VirtMachineState *vms, unsigned dsdt_tbl_offset)
> >>  {
> >> +    int fadt_start = table_data->len;
> >>      AcpiFadtDescriptorRev5_1 *fadt = acpi_data_push(table_data, sizeof(*fadt));
> >>      unsigned xdsdt_entry_offset = (char *)&fadt->x_dsdt - table_data->data;
> >>      uint16_t bootflags;
> >> @@ -681,8 +685,8 @@ static void build_fadt(GArray *table_data, BIOSLinker *linker,
> >>          ACPI_BUILD_TABLE_FILE, xdsdt_entry_offset, sizeof(fadt->x_dsdt),
> >>          ACPI_BUILD_TABLE_FILE, dsdt_tbl_offset);
> >>  
> >> -    build_header(linker, table_data,
> >> -                 (void *)fadt, "FACP", sizeof(*fadt), 5, NULL, NULL);
> >> +    build_header(linker, table_data, (void *)(table_data->data + fadt_start),
> >> +                 "FACP", table_data->len - fadt_start, 5, NULL, NULL);
> >>  }
> > 
> > No need to change this one either.
> > 
> >>  
> >>  /* DSDT */
> >> -- 
> >> 2.0.4
> >>
> >>
> > 
> > Please respin only changing the one that needs the fix.
> > 
> Hi Andrew,
> 
> Thanks for your comments. What you said is right that only the
> build_srat needs to be fixed but I thought we need to unify the style
> and avoid new issues if add some acpi_data_push in other functions in
> the future.

Our style is unified with the x86 style which doesn't bother with the
_start offsets when only one push is used. I agree it's a bit error
prone, though, so I won't insist on not changing the ones that don't
need it. However if you choose to keep them, then I think the commit
message should be updated to explain that only SRAT needs the change,
and the other changes are for future-proofing and pattern consistency.

Thanks,
drew