On Thu, Jul 24, 2014 at 04:32:09PM +0200, Paolo Bonzini wrote:
> Changing the ACPI table size causes migration to break, and the memory
> hotplug work opened our eyes on how horribly we were breaking things in
> 2.0 already.
>
> The ACPI table size is rounded to the next 4k, which one would think
> gives some headroom. In practice this is not the case, because the user
> can control the ACPI table size (each CPU adds 97 bytes to the SSDT and
> 8 to the MADT) and so some "-smp" values will break the 4k boundary and
> fail to migrate. Similarly, PCI bridges add ~1870 bytes to the SSDT.
>
> To fix this, hard-code 64k as the maximum ACPI table size, which
> (despite being an order of magnitude smaller than 640k) should be enough
> for everyone.
>
> To fix migration from QEMU 2.0, compute the payload size of QEMU 2.0
> and always use that one. The previous patch shrunk the ACPI tables
> enough that the QEMU 2.0 size should always be enough.
>
> Migration from QEMU 1.7 should work for guests that have a number of CPUs
> other than 12, 13, 14, 54, 55, 56, 97, 98, 139, 140. It was already
> broken from QEMU 1.7 to QEMU 2.0 in the same way, though.
>
> Even with this patch, QEMU 1.7 and 2.0 have two different ideas of
> "-M pc-i440fx-2.0" when there are PCI bridges. Igor sent a patch to
> adopt the QEMU 1.7 definition. I think distributions should apply
> it if they move directly from QEMU 1.7 to 2.1+ without ever packaging
> version 2.0.
>
> Signed-off-by: Paolo Bonzini <pbonz...@redhat.com>
> ---
> replace magic constants with #defines [Igor]
> remove stray line from comment [Laszlo]
>
> hw/i386/acpi-build.c | 71
> +++++++++++++++++++++++++++++++++++++++++++++++++---
> hw/i386/pc_piix.c | 19 ++++++++++++++
> hw/i386/pc_q35.c | 5 ++++
> include/hw/i386/pc.h | 1 +
> 4 files changed, 92 insertions(+), 4 deletions(-)
>
> diff --git a/hw/i386/acpi-build.c b/hw/i386/acpi-build.c
> index ebc5f03..26d8dfa 100644
> --- a/hw/i386/acpi-build.c
> +++ b/hw/i386/acpi-build.c
> @@ -25,7 +25,9 @@
> #include <glib.h>
> #include "qemu-common.h"
> #include "qemu/bitmap.h"
> +#include "qemu/osdep.h"
> #include "qemu/range.h"
> +#include "qemu/error-report.h"
> #include "hw/pci/pci.h"
> #include "qom/cpu.h"
> #include "hw/i386/pc.h"
> @@ -52,6 +54,16 @@
> #include "qapi/qmp/qint.h"
> #include "qom/qom-qobject.h"
>
> +/* These are used to size the ACPI tables for -M pc-i440fx-1.7 and
> + * -M pc-i440fx-2.0.
Let's just say 2.0 and earlier?
> Even if the actual amount of AML generated grows
> + * a little bit, there should be plenty of free space since the DSDT
> + * shrunk by ~1.5k between QEMU 2.0 and QEMU 2.1.
> + */
> +#define ACPI_BUILD_CPU_AML_SIZE 97
> +#define ACPI_BUILD_BRIDGE_AML_SIZE 1875
Let's put _LEGACY_ somewhere here?
> +
> +#define ACPI_BUILD_TABLE_SIZE 0x10000
> +
> typedef struct AcpiCpuInfo {
> DECLARE_BITMAP(found_cpus, ACPI_CPU_HOTPLUG_ID_LIMIT);
> } AcpiCpuInfo;
> @@ -87,6 +99,8 @@ typedef struct AcpiBuildPciBusHotplugState {
> struct AcpiBuildPciBusHotplugState *parent;
> } AcpiBuildPciBusHotplugState;
>
> +unsigned bsel_alloc;
> +
Patch will be better contained if instead of using a global
bsel_alloc, we actually go and count the devices that
have ACPI_PCIHP_PROP_BSEL.
You can just scan all devices, or all pci devices, it
should not matter.
This way, this code will be local to the legacy path.
> static void acpi_get_dsdt(AcpiMiscInfo *info)
> {
> uint16_t *applesmc_sta;
> @@ -759,8 +773,8 @@ static void *acpi_set_bsel(PCIBus *bus, void *opaque)
> static void acpi_set_pci_info(void)
> {
> PCIBus *bus = find_i440fx(); /* TODO: Q35 support */
> - unsigned bsel_alloc = 0;
>
> + assert(bsel_alloc == 0);
> if (bus) {
> /* Scan all PCI buses. Set property to enable acpi based hotplug. */
> pci_for_each_bus_depth_first(bus, acpi_set_bsel, NULL, &bsel_alloc);
> @@ -1440,13 +1454,14 @@ static
> void acpi_build(PcGuestInfo *guest_info, AcpiBuildTables *tables)
> {
> GArray *table_offsets;
> - unsigned facs, dsdt, rsdt;
> + unsigned facs, ssdt, dsdt, rsdt;
> AcpiCpuInfo cpu;
> AcpiPmInfo pm;
> AcpiMiscInfo misc;
> AcpiMcfgInfo mcfg;
> PcPciInfo pci;
> uint8_t *u;
> + size_t aml_len = 0;
>
> acpi_get_cpu_info(&cpu);
> acpi_get_pm_info(&pm);
> @@ -1474,13 +1489,20 @@ void acpi_build(PcGuestInfo *guest_info,
> AcpiBuildTables *tables)
> dsdt = tables->table_data->len;
> build_dsdt(tables->table_data, tables->linker, &misc);
>
> + /* Count the size of the DSDT and SSDT, we will need it for legacy
> + * sizing of ACPI tables.
> + */
> + aml_len += tables->table_data->len - dsdt;
> +
> /* ACPI tables pointed to by RSDT */
> acpi_add_table(table_offsets, tables->table_data);
> build_fadt(tables->table_data, tables->linker, &pm, facs, dsdt);
>
> + ssdt = tables->table_data->len;
> acpi_add_table(table_offsets, tables->table_data);
> build_ssdt(tables->table_data, tables->linker, &cpu, &pm, &misc, &pci,
> guest_info);
> + aml_len += tables->table_data->len - ssdt;
>
> acpi_add_table(table_offsets, tables->table_data);
> build_madt(tables->table_data, tables->linker, &cpu, guest_info);
> @@ -1513,12 +1535,53 @@ void acpi_build(PcGuestInfo *guest_info,
> AcpiBuildTables *tables)
> /* RSDP is in FSEG memory, so allocate it separately */
> build_rsdp(tables->rsdp, tables->linker, rsdt);
>
> - /* We'll expose it all to Guest so align size to reduce
> + /* We'll expose it all to Guest so we want to reduce
> * chance of size changes.
> * RSDP is small so it's easy to keep it immutable, no need to
> * bother with alignment.
> + *
> + * We used to align the tables to 4k, but of course this would
> + * too simple to be enough. 4k turned out to be too small an
> + * alignment very soon, and in fact it is almost impossible to
> + * keep the table size stable for all (max_cpus, max_memory_slots)
> + * combinations. So the table size is always 64k for pc-i440fx-2.1
> + * and we give an error if the table grows beyond that limit.
> + *
> + * We still have the problem of migrating from "-M pc-i440fx-2.0". For
> + * that, we exploit the fact that QEMU 2.1 generates _smaller_ tables
> + * than 2.0 and we can always pad the smaller tables with zeros. We can
> + * then use the exact size of the 2.0 tables.
> + *
> + * All this is for PIIX4, since QEMU 2.0 didn't support Q35 migration.
> */
> - acpi_align_size(tables->table_data, 0x1000);
> + if (guest_info->legacy_acpi_table_size) {
> + /* Subtracting aml_len gives the size of fixed tables. Then add the
> + * size of the PIIX4 DSDT/SSDT in QEMU 2.0.
> + */
> + int legacy_aml_len =
> + guest_info->legacy_acpi_table_size +
> + ACPI_BUILD_CPU_AML_SIZE * max_cpus +
> + ACPI_BUILD_BRIDGE_AML_SIZE * (MAX(bsel_alloc, 1) - 1);
> + int legacy_table_size =
> + ROUND_UP(tables->table_data->len - aml_len + legacy_aml_len,
> 0x1000);
> + if (tables->table_data->len > legacy_table_size) {
> + /* -M pc-i440fx-2.0 doesn't support memory hotplug, so this
> should
> + * never happen.
> + */
> + error_report("This configuration is not supported with -M
> pc-i440fx-2.0.");
> + error_report("Please report this to qemu-devel@nongnu.org.");
Downstreams would need to patch this line out? Let's just drop the
second line.
> + exit(1);
so what happens here is 2.x -> 2.0 migration becomes broken, but
2.0 -> 2.x could still work.
I think I would prefer a warning instead.
> + }
> + g_array_set_size(tables->table_data, legacy_table_size);
> + } else {
> + if (tables->table_data->len > ACPI_BUILD_TABLE_SIZE) {
> + /* As of QEMU 2.1, this fires with 160 VCPUs and 255 memory
> slots. */
> + error_report("ACPI tables are larger than 64k. Please remove");
> + error_report("CPUs, NUMA nodes, memory slots or PCI bridges.");
> + exit(1);
> + }
> + g_array_set_size(tables->table_data, ACPI_BUILD_TABLE_SIZE);
Let's split the idea to use 64K always out, to a separate patch, ok?
> + }
>
> acpi_align_size(tables->linker, 0x1000);
>
> diff --git a/hw/i386/pc_piix.c b/hw/i386/pc_piix.c
> index 7081c08..c8a8090 100644
> --- a/hw/i386/pc_piix.c
> +++ b/hw/i386/pc_piix.c
> @@ -61,6 +61,7 @@ static const int ide_irq[MAX_IDE_BUS] = { 14, 15 };
>
> static bool has_pci_info;
> static bool has_acpi_build = true;
> +static int legacy_acpi_table_size;
> static bool smbios_defaults = true;
> static bool smbios_legacy_mode;
> /* Make sure that guest addresses aligned at 1Gbyte boundaries get mapped to
> @@ -163,6 +164,7 @@ static void pc_init1(MachineState *machine,
> guest_info = pc_guest_info_init(below_4g_mem_size, above_4g_mem_size);
>
> guest_info->has_acpi_build = has_acpi_build;
> + guest_info->legacy_acpi_table_size = legacy_acpi_table_size;
>
> guest_info->has_pci_info = has_pci_info;
> guest_info->isapc_ram_fw = !pci_enabled;
> @@ -297,6 +299,23 @@ static void pc_init_pci(MachineState *machine)
>
> static void pc_compat_2_0(MachineState *machine)
> {
> + /* This value depends on the actual DSDT and SSDT compiled into
> + * the source QEMU; unfortunately it depends on the binary and
> + * not on the machine type, so we cannot make pc-i440fx-1.7 work on
> + * both QEMU 1.7 and QEMU 2.0.
> + *
> + * Large variations cause migration to fail for more than one
> + * consecutive value of the "-smp" maxcpus option.
> + *
> + * For small variations of the kind caused by different iasl versions,
> + * the 4k rounding usually leaves slack. However, there could be still
> + * one or two values that break. For QEMU 1.7 and QEMU 2.0 the
> + * slack is only ~10 bytes before one "-smp maxcpus" value breaks!
> + *
> + * 6652 is valid for QEMU 2.0, the right value for pc-i440fx-1.7 on
> + * QEMU 1.7 it is 6414. For RHEL/CentOS 7.0 it is 6358.
> + */
> + legacy_acpi_table_size = 6652;
> smbios_legacy_mode = true;
> has_reserved_memory = false;
> }
> diff --git a/hw/i386/pc_q35.c b/hw/i386/pc_q35.c
> index f551961..c39ee98 100644
> --- a/hw/i386/pc_q35.c
> +++ b/hw/i386/pc_q35.c
> @@ -155,6 +155,11 @@ static void pc_q35_init(MachineState *machine)
> guest_info->has_acpi_build = has_acpi_build;
> guest_info->has_reserved_memory = has_reserved_memory;
>
> + /* Migration was not supported in 2.0 for Q35, so do not bother
> + * with this hack (see hw/i386/acpi-build.c).
> + */
> + guest_info->legacy_acpi_table_size = 0;
> +
> if (smbios_defaults) {
> MachineClass *mc = MACHINE_GET_CLASS(machine);
> /* These values are guest ABI, do not change */
> diff --git a/include/hw/i386/pc.h b/include/hw/i386/pc.h
> index 1c0c382..f4b9b2b 100644
> --- a/include/hw/i386/pc.h
> +++ b/include/hw/i386/pc.h
> @@ -94,6 +94,7 @@ struct PcGuestInfo {
> uint64_t *node_mem;
> uint64_t *node_cpu;
> FWCfgState *fw_cfg;
> + int legacy_acpi_table_size;
> bool has_acpi_build;
> bool has_reserved_memory;
> };
> --
> 1.8.3.1
