On Mon, Jun 08, 2020 at 05:44:48PM -0700, Kevin Mitchell wrote:
When booting with ACPI unavailable or disabled, get_smp_config() ends up calling MP_processor_info() for each CPU found in the MPS table. Previously, this resulted in boot_cpu_physical_apicid getting unconditionally overwritten by the apicid of whatever processor had the CPU_BOOTPROCESSOR flag. This occurred even if boot_cpu_physical_apicid had already been more reliably determined in register_lapic_address() by calling read_apic_id() from the actual boot processor.
Ordinariliy, this is not a problem because the boot processor really is the one with the CPU_BOOTPROCESSOR flag. However, kexec is an exception in which the kernel may be booted from any processor regardless of the MPS table contents. In this case, boot_cpu_physical_apicid may not indicate the actual boot processor.
This was particularly problematic when the second kernel was booted with NR_CPUS fewer than the number of physical processors. It's the job of generic_processor_info() to decide which CPUs to bring up in this case. That obviously must include the real boot processor which it takes care to save a slot for. It relies upon the contents of boot_cpu_physical_apicid to do this, which if incorrect, may result in the boot processor getting left out.
This condition can be discovered by smp_sanity_check() and rectified by adding the boot processor to the phys_cpu_present_map with the warning "weird, boot CPU (#%d) not listed by the BIOS". However, commit 3e730dad3b6da ("x86/apic: Unify interrupt mode setup for SMP-capable system") caused setup_local_APIC() to be called before this could happen resulting in a BUG_ON(!apic->apic_id_registered()):
[ 0.655452] ------------[ cut here ]------------ [ 0.660610] Kernel BUG at setup_local_APIC+0x74/0x280 [verbose debug info unavailable] [ 0.669466] invalid opcode: 0000 [#1] SMP [ 0.673948] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.19.109.Ar-16509018.eostrunkkernel419 #1 [ 0.683670] Hardware name: Quanta Quanta LY6 (1LY6UZZ0FBC), BIOS 1.0.6.0-e7d6a55 11/26/2015 [ 0.693007] RIP: 0010:setup_local_APIC+0x74/0x280 [ 0.698264] Code: 80 e4 fe bf f0 00 00 00 89 c6 48 8b 05 0f 1a 8e 00 ff 50 10 e8 12 53 fd ff 48 8b 05 00 1a 8e 00 ff 90 a0 00 00 00 85 c0 75 02 <0f> 0b 48 8b 05 ed 19 8e 00 41 be 00 02 00 00 ff 90 b0 00 00 00 48 [ 0.719251] RSP: 0000:ffffffff81a03e20 EFLAGS: 00010246 [ 0.725091] RAX: 0000000000000000 RBX: 0000000000000003 RCX: 0000000000000000 [ 0.733066] RDX: 0000000000000000 RSI: 000000000000000f RDI: 0000000000000020 [ 0.741041] RBP: ffffffff81a03e98 R08: 0000000000000002 R09: 0000000000000000 [ 0.749014] R10: ffffffff81a204e0 R11: ffffffff81b50ea7 R12: 0000000000000000 [ 0.756989] R13: ffffffff81aef920 R14: ffffffff81af60a0 R15: 0000000000000000 [ 0.764965] FS: 0000000000000000(0000) GS:ffff888036800000(0000) knlGS:0000000000000000 [ 0.774007] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 0.780427] CR2: ffff888035c01000 CR3: 0000000035a08000 CR4: 00000000000006b0 [ 0.788401] Call Trace: [ 0.791137] ? amd_iommu_prepare+0x15/0x2a [ 0.795717] apic_bsp_setup+0x55/0x75 [ 0.799808] apic_intr_mode_init+0x169/0x16e [ 0.804579] x86_late_time_init+0x10/0x17 [ 0.809062] start_kernel+0x37e/0x3fe [ 0.813154] x86_64_start_reservations+0x2a/0x2c [ 0.818316] x86_64_start_kernel+0x72/0x75 [ 0.822886] secondary_startup_64+0xa4/0xb0 [ 0.827564] ---[ end trace 237b64da0fd9b22e ]---
This change avoids these issues by only setting boot_cpu_physical_apicid from the MPS table if it is not already set, which can occur in the construct_default_ISA_mptable() path. Otherwise, boot_cpu_physical_apicid will already have been set in register_lapic_address() and should therefore remain untouched.
Looking through all the places where boot_cpu_physical_apicid is accessed, nearly all of them assume that boot_cpu_physical_apicid should match read_apic_id() on the booting processor. The only place that might intend to use the BSP apicid listed in the MPS table is amd_numa_init(), which explicitly requires boot_cpu_physical_apicid to be the lowest apicid of all processors. Ironically, due to the early exit short circuit in early_get_smp_config(), it instead gets boot_cpu_physical_apicid = read_apic_id() rather than the MPS table BSP. The behaviour of amd_numa_init() is therefore unaffected by this change.
Fixes: 3e730dad3b6da ("x86/apic: Unify interrupt mode setup for SMP-capable system") Signed-off-by: Kevin Mitchell kevmitch@arista.com Cc: stable@vger.kernel.org
arch/x86/kernel/mpparse.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/arch/x86/kernel/mpparse.c b/arch/x86/kernel/mpparse.c index afac7ccce72f..6f22f09bfe11 100644 --- a/arch/x86/kernel/mpparse.c +++ b/arch/x86/kernel/mpparse.c @@ -64,7 +64,8 @@ static void __init MP_processor_info(struct mpc_cpu *m) if (m->cpuflag & CPU_BOOTPROCESSOR) { bootup_cpu = " (Bootup-CPU)";
boot_cpu_physical_apicid = m->apicid;
if (boot_cpu_physical_apicid == -1U)boot_cpu_physical_apicid = m->apicid;pr_info("Processor #%d%s\n", m->apicid, bootup_cpu); -- 2.26.2
We've moved on to our next kernel upgrade to linux-5.10 and are still seeing this same issue with the upstream kernel. We will therefore be porting this patch forward, but still wondering if there is any interest in getting this into the mainline kernel so more people get (more) correct code? Both patches still apply to the mainline (linux-5.18-rc7 right now). Are there any alternative suggestions for avoiding this BUG_ON on kexec?