Re: [Qemu-arm] [RFC PATCH 0/4] ARM virt: ACPI memory hotplug support

[Igor Mammedov]

On Thu, 28 Feb 2019 13:27:54 +0100
Laszlo Ersek <address@hidden> wrote:

> On 02/28/19 11:12, Auger Eric wrote:
> > Hi Laszlo,
> > 
> > On 2/27/19 9:14 PM, Laszlo Ersek wrote:  
> >> On 02/27/19 13:55, Shameerali Kolothum Thodi wrote:  
> >>> Hi Laszlo,
> >>>  
> >>>> -----Original Message-----
> >>>> From: Shameerali Kolothum Thodi
> >>>> Sent: 25 February 2019 09:54
> >>>> To: 'Laszlo Ersek' <address@hidden>; Auger Eric <address@hidden>;
> >>>> address@hidden; address@hidden;
> >>>> address@hidden; address@hidden; address@hidden
> >>>> Cc: xuwei (O) <address@hidden>; Linuxarm <address@hidden>; Ard
> >>>> Biesheuvel <address@hidden>; Leif Lindholm (Linaro address)
> >>>> <address@hidden>
> >>>> Subject: RE: [RFC PATCH 0/4] ARM virt: ACPI memory hotplug support  
> >>>
> >>> [...]
> >>>    
> >>>>>>> The root cause seems to be EDK2 ACPI table checksum failure
> >>>>>>> as NFIT table is getting updated on hot-add. This needs further
> >>>>>>> investigation.  
> >>>>>> + Ard, Leif, Laszlo if they have any idea of what is missing/wrong.  
> >>>>>
> >>>>> Huh, very interesting; I usually don't expect my sanity checks to fire
> >>>>> in practice. :)
> >>>>>
> >>>>> The message
> >>>>>
> >>>>>   ProcessCmdAddChecksum: invalid checksum range in "etc/acpi/tables"
> >>>>>
> >>>>> is logged by OVMF's and ArmVirtQemu's ACPI Platform DXE Driver when it
> >>>>> finds an invalid COMMAND_ADD_CHECKSUM command in QEMU's ACPI
> >>>>> linker/loader script.
> >>>>>
> >>>>> Please see the command definition in QEMU's
> >>>>> "hw/acpi/bios-linker-loader.c". In particular, please refer to the
> >>>>> function bios_linker_loader_add_checksum(), which builds the command
> >>>>> structure, and documents the fields.
> >>>>>
> >>>>> (You may also refer to QEMU_LOADER_ADD_CHECKSUM in file
> >>>>> "OvmfPkg/AcpiPlatformDxe/QemuLoader.h" in the edk2 source tree, for the
> >>>>> same information.)
> >>>>>
> >>>>> The error message is logged if:
> >>>>> - the offset at which the checksum should be stored falls outside of the
> >>>>> size of the fw_cfg blob, or
> >>>>> - the range over which the checksum should be calculated falls outside
> >>>>> (at least in part) of the fw_cfg blob.
> >>>>>
> >>>>> To me this suggests that QEMU generates an invalid
> >>>>> COMMAND_ADD_CHECKSUM
> >>>>> command for the firmware.
> >>>>>
> >>>>> ... I've tried to skim the patches briefly. I think there must be an
> >>>>> error in the DSDT building logic that is only active on reboot if an
> >>>>> nvdimm module was hot-added before the reboot.  
> >>>>
> >>>> Thanks for taking a look and the pointers. I will debug this further
> >>>> and get back.  
> >>>
> >>> The root cause of the issue seems to be UEFI not seeing the updated acpi
> >>> table blob size on reboot once a new NFIT table is added(nvdimm hot 
> >>> added).
> >>>
> >>> Please see the debug logs below,
> >>>
> >>> Initial Guest boot
> >>> ---------------------------
> >>>
> >>> Debug logs from Qemu:
> >>>
> >>> build_header: acpi sig DSDT len 0x5127
> >>> build_header: acpi sig FACP len 0x10c
> >>> build_header: acpi sig APIC len 0xa8
> >>> build_header: acpi sig GTDT len 0x60
> >>> build_header: acpi sig MCFG len 0x3c
> >>> build_header: acpi sig SPCR len 0x50
> >>> build_header: acpi sig SRAT len 0x92
> >>> build_header: acpi sig SSDT len 0x38f
> >>> build_header: acpi sig XSDT len 0x5c
> >>> virt_acpi_build: acpi table_blob len 0x5844
> >>>
> >>> Debug logs from UEFI:
> >>>
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x9 Start=0x0 
> >>> Length=0x5127 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x5130 
> >>> Start=0x5127 Length=0x10C Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x523C 
> >>> Start=0x5233 Length=0xA8 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x52E4 
> >>> Start=0x52DB Length=0x60 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x5344 
> >>> Start=0x533B Length=0x3C Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x5380 
> >>> Start=0x5377 Length=0x50 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x53D0 
> >>> Start=0x53C7 Length=0x92 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x5462 
> >>> Start=0x5459 Length=0x38F Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x57F1 
> >>> Start=0x57E8 Length=0x5C Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/rsdp" ResultOffset=0x8 Start=0x0 
> >>> Length=0x14 Blob->Size=0x24
> >>> ProcessCmdAddChecksum: File="etc/acpi/rsdp" ResultOffset=0x20 Start=0x0 
> >>> Length=0x24 Blob->Size=0x24
> >>> InstallQemuFwCfgTables: installed 8 tables
> >>>
> >>> Guest Reboot after ndimm hot added
> >>> ------------------------------------
> >>>
> >>> Debug logs from Qemu:
> >>>
> >>> build_header: acpi sig DSDT len 0x5127
> >>> build_header: acpi sig FACP len 0x10c
> >>> build_header: acpi sig APIC len 0xa8
> >>> build_header: acpi sig GTDT len 0x60
> >>> build_header: acpi sig MCFG len 0x3c
> >>> build_header: acpi sig SPCR len 0x50
> >>> build_header: acpi sig SRAT len 0x92
> >>> build_header: acpi sig SSDT len 0x38f
> >>> build_header: acpi sig NFIT len 0xe0  -->New
> >>> build_header: acpi sig XSDT len 0x64
> >>> virt_acpi_build: acpi table_blob len 0x592c -->blob len updated
> >>>
> >>> Debug logs from UEFI:
> >>>
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x9 Start=0x0 
> >>> Length=0x5127 Blob->Size=0x5844  -->Wrong blob size.
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x5130 
> >>> Start=0x5127 Length=0x10C Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x523C 
> >>> Start=0x5233 Length=0xA8 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x52E4 
> >>> Start=0x52DB Length=0x60 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x5344 
> >>> Start=0x533B Length=0x3C Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x5380 
> >>> Start=0x5377 Length=0x50 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x53D0 
> >>> Start=0x53C7 Length=0x92 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x5462 
> >>> Start=0x5459 Length=0x38F Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: File="etc/acpi/tables" ResultOffset=0x57F1 
> >>> Start=0x57E8 Length=0xE0 Blob->Size=0x5844
> >>> ProcessCmdAddChecksum: invalid checksum range in "etc/acpi/tables"
> >>> OnRootBridgesConnected: InstallAcpiTables: Protocol Error
> >>>
> >>>
> >>> To me it seems on ARM vit acpi path, the blob len is calculated based
> >>> on actual tables and is updated only in virt_acpi_setup() --> 
> >>> acpi_add_rom_blob()
> >>> path. I had a look at the x86 code and it looks like, there, the blob len 
> >>> gets updated
> >>> with an additional buffer to take care of table resizing[1].
> >>>
> >>> As a hack i added the same to ARM virt and it seems to resolve the issue.
> >>> I am not sure this is the best approach to fix this though.
> >>>
> >>> Please let me know your thoughts.
> >>>
> >>> Thanks,
> >>> Shameer
> >>>
> >>> diff --git a/hw/arm/virt-acpi-build.c b/hw/arm/virt-acpi-build.c
> >>> index 132414c..4291553 100644
> >>> --- a/hw/arm/virt-acpi-build.c
> >>> +++ b/hw/arm/virt-acpi-build.c
> >>> @@ -50,6 +50,8 @@
> >>>  #define ARM_SPI_BASE 32
> >>>  #define ACPI_POWER_BUTTON_DEVICE "PWRB"
> >>>
> >>> +#define ACPI_BUILD_TABLE_SIZE    0x20000
> >>> +
> >>>  static void acpi_dsdt_add_cpus(Aml *scope, int smp_cpus)
> >>>  {
> >>>      uint16_t i;
> >>> @@ -886,6 +888,10 @@ void virt_acpi_build(VirtMachineState *vms, 
> >>> AcpiBuildTables *tables)
> >>>          build_rsdp(tables->rsdp, tables->linker, &rsdp_data);
> >>>      }
> >>>
> >>> +    /* Make sure we have a buffer in case we need to resize the tables. 
> >>> */
> >>> +    g_array_set_size(tables_blob, ROUND_UP(acpi_data_len(tables_blob),
> >>> +                     ACPI_BUILD_TABLE_SIZE));
> >>> +
> >>>      /* Cleanup memory that's no longer used. */
> >>>      g_array_free(table_offsets, true);
> >>>  }
> >>>
> >>> [1] https://github.com/qemu/qemu/blob/master/hw/i386/acpi-build.c#L2792  
> >>
> >> Nice analysis, thanks.
> >>
> >> I think the line that you reference, i.e.
> >>
> >>   acpi_align_size(tables_blob, ACPI_BUILD_TABLE_SIZE);
> >>
> >> in acpi_build() [hw/i386/acpi-build.c] masks this issue for x86 only as
> >> a side effect. To my understanding, the alignment / padding exists there
> >> for migration compatibility. It doesn't exist for updating the size of
> >> the ACPI blobs in fw_cfg across reboots. The issue is masked because the
> >> alignment is large enough (un-changed) to contain the regenerated blobs
> >> as well.>
> >> Given that the "virt" machine type is versioned, I think migration
> >> compat is a valid concern there too. This in itself would justify a
> >> similar padding.  
> > I don't understand the migration compat issue. Please could you elaborate?  
> 
> git-blame explains it to some extent -- please see commit 07fb61760cde
> ("pc: hack for migration compatibility from QEMU 2.0", 2014-07-28).
> 
> I don't remember any details at this point that the commit does not
> state. (I see that I reviewed the patch back then, so perhaps the
> mailing list archive has some discussion.)
> 
> Interestingly, the commit message refers to "memory hotplug work" too.
> 
> ... Ahh, wait, I do remember the main issue now. Here's the thing. The
> ACPI payload that QEMU generates for the firmware is considered a part
> of the firmware itself. Therefore, it is not versioned -- because the
> firmware itself is not versioned. (In other words, if you migrate a VM
> from one host to another host, and that other host has different
> firmware that the VM will pick up after re-launch (from cold boot), then
> the firmware will change in the VM.)
> 
> By considering ACPI a part of the firmware, QEMU never versioned the
> ACPI payload, just like the actual firmware was never versioned. In
> other words, if you have machine type Foo on qemu release Bar, and
> machine type Foo on qemu release Baz, compat properties and such will
> ensure that the virtual hardware looks the same to the guest, but QEMU
> will *not* ensure that the ACPI payload generated at QEMU startup (more
> precisely, at "machine done") will be identical. Despite the fact that
> both QEMU instances use machine type Foo.
> 
> Now, combine this with the feature that fw_cfg has been backed by RAM
> Blocks, for a quite long time now (this wasn't always the case, but it
> has been for multiple years now). The end result is that the RAM
> block(s) holding the initial ACPI payload may differ between releases
> Bar and Baz, within the same machine type Foo. This means that migration
> between them will fail, due to RAMBlock size difference.
> 
> Hence the padding -- it tries to cancel out small variances in ACPI
> payload size.
> 
> >>
> >> I don't know if we want to specifically care about size-changing
> >> ACPI-regen across reboot. I believe measures for that specific use case
> >> don't exist in x86 machine types either.  
> > The NFIT redimensioning should exit on x86 too?  
> 
> That's not my point. My point was that the padding, which was originally
> supposed to mask variances in ACPI payload size across *QEMU releases*,
> for migration compat, ended up masking a variance of different origin:
> namely ACPI regeneration at reboot (with different contents). In other
> words, we never implemented any specific measures for this
> resize-on-reboot issue, instead we allowed the migration compat code
> (the padding) to take care of it as well.
> 
> In virt, there is no such ACPI padding code (for migration compat) --
> for whatever reason --, and so it *also* cannot take care of the
> resize-on-reboot problem.
What's describe above is a bit outdate, let me paint a rough current state
in chronological order (I haven't checked how it actually works now).
--
1. how it work(s|ed):
     * tables are generated during machine_done and registered with fw_cfg
     * then oops: we have device_add that could be called later but before
       guest runs and we don't have complete picture wrt CRS resources for
       PCI as that is programmed by firmware, so here goes a trick to
       regenerate tables when firmware accesses its fwcfg entry
       (see how build_state->patched is used), byt that time FW has intialized
       PCI to the extent necessary to build valid tables.
     * then we do hot add ACPI table are not updated until reset
       (acpi_build_reset) and all dance with patching repeats when FW
       gets tables form QEMU.
     * at those versions of QEMU tables size was static
     
2. Later on hotplugging bridges (I think) caused issues on reboot/migration
as it caused DSDT size increase, that made as invent padding race train-wreak.
2.2 then there were cases where tables changed size depending where a device
was cold or hot plugged and may be some other issues I don't recall.

3. With padding race being a lost game (there always were border line configs
where it broke migration), Michael introduced re-sizable blobs 42d859001d for
acpi table blobs, which could be resized on target to accommodate larger
tables blob during migration.

4. Even with resizable memory regions for the sake of backward migration
compatibility we needed to keep padding we used with legacy machine. It's
still broken in some cases but mostly works (we declared padding approach
broken by design and left it as is)
--

Thankfully arm/virt ACPI support materialized after all of the above
so we don't have legacy padding obligations at all. Hence suggest to
find were the bug is instead of trying to start yet another padding race.
 
> [...]
> 
> Thanks
> Laszlo