From 5edd19af18a36a4e22c570b1b969179e0ca1fe4c Mon Sep 17 00:00:00 2001 From: Cliff Wickman Date: Tue, 20 Jul 2010 18:09:05 -0500 Subject: x86, UV: Make kdump avoid stack dumps UV NMI callback's should not write stack dumps when a kdump is to be written. When invoking the crash kernel to write a dump, kdump_nmi_shootdown_cpus() uses NMI's to get all the cpu's to save their register context and halt. But the NMI interrupt handler runs a callback list. This patch sets a flag to prevent any of those callbacks from interfering with the halt of the cpu. For UV, which currently has the only callback to which this is relevant, the uv_handle_nmi() callback should not do dumping of stacks. The 'in_crash_kexec' flag is defined as an extern in kdebug.h firstly because x2apic_uv_x.c includes it. Secondly because some future callback might need the flag to know that it should not enter the debugger. (Such a scenario was in fact present in the 2.6.32 kernel, SuSE distribution, where a call to kdb needed to be avoided.) Signed-off-by: Cliff Wickman LKML-Reference: Signed-off-by: H. Peter Anvin --- arch/x86/include/asm/kdebug.h | 1 + arch/x86/kernel/apic/x2apic_uv_x.c | 4 ++++ arch/x86/kernel/crash.c | 3 +++ 3 files changed, 8 insertions(+) (limited to 'arch') diff --git a/arch/x86/include/asm/kdebug.h b/arch/x86/include/asm/kdebug.h index fa7c0b9..7a2910b 100644 --- a/arch/x86/include/asm/kdebug.h +++ b/arch/x86/include/asm/kdebug.h @@ -33,5 +33,6 @@ extern void __show_regs(struct pt_regs *regs, int all); extern void show_regs(struct pt_regs *regs); extern unsigned long oops_begin(void); extern void oops_end(unsigned long, struct pt_regs *, int signr); +extern int in_crash_kexec; #endif /* _ASM_X86_KDEBUG_H */ diff --git a/arch/x86/kernel/apic/x2apic_uv_x.c b/arch/x86/kernel/apic/x2apic_uv_x.c index e46f98f..7b598b8 100644 --- a/arch/x86/kernel/apic/x2apic_uv_x.c +++ b/arch/x86/kernel/apic/x2apic_uv_x.c @@ -604,6 +604,10 @@ int uv_handle_nmi(struct notifier_block *self, unsigned long reason, void *data) { if (reason != DIE_NMI_IPI) return NOTIFY_OK; + + if (in_crash_kexec) + /* do nothing if entering the crash kernel */ + return NOTIFY_OK; /* * Use a lock so only one cpu prints at a time * to prevent intermixed output. diff --git a/arch/x86/kernel/crash.c b/arch/x86/kernel/crash.c index ebd4c51..764c7c2 100644 --- a/arch/x86/kernel/crash.c +++ b/arch/x86/kernel/crash.c @@ -28,6 +28,8 @@ #include #include +int in_crash_kexec; + #if defined(CONFIG_SMP) && defined(CONFIG_X86_LOCAL_APIC) static void kdump_nmi_callback(int cpu, struct die_args *args) @@ -61,6 +63,7 @@ static void kdump_nmi_callback(int cpu, struct die_args *args) static void kdump_nmi_shootdown_cpus(void) { + in_crash_kexec = 1; nmi_shootdown_cpus(kdump_nmi_callback); disable_local_APIC(); -- cgit v1.1 From c4026cfd8febcd63dd278894108839f30e525a0e Mon Sep 17 00:00:00 2001 From: Cliff Wickman Date: Fri, 30 Jul 2010 14:10:55 -0500 Subject: x86, UV: Initialize BAU hub map Fix uninitialized uvhub_mask: - An unitialized bit map variable was causing initialization of non-existant hubs (this one causes boot panics). - And the bit map was too small for large machines. This patch makes it dynamic in size. - Fix the case where socket 0 has no enabled cpu's. Don't assume every hub has a socket 0. - uv_init_per_cpu() should be __init. Signed-off-by: Cliff Wickman Cc: # for .35.x LKML-Reference: Signed-off-by: Ingo Molnar --- arch/x86/kernel/tlb_uv.c | 26 ++++++++++++++------------ 1 file changed, 14 insertions(+), 12 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/tlb_uv.c b/arch/x86/kernel/tlb_uv.c index 59efb53..312ef02 100644 --- a/arch/x86/kernel/tlb_uv.c +++ b/arch/x86/kernel/tlb_uv.c @@ -1484,15 +1484,16 @@ calculate_destination_timeout(void) /* * initialize the bau_control structure for each cpu */ -static void uv_init_per_cpu(int nuvhubs) +static void __init uv_init_per_cpu(int nuvhubs) { int i; int cpu; int pnode; int uvhub; + int have_hmaster; short socket = 0; unsigned short socket_mask; - unsigned int uvhub_mask; + unsigned char *uvhub_mask; struct bau_control *bcp; struct uvhub_desc *bdp; struct socket_desc *sdp; @@ -1516,28 +1517,29 @@ static void uv_init_per_cpu(int nuvhubs) uvhub_descs = (struct uvhub_desc *) kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL); memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc)); + uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL); for_each_present_cpu(cpu) { bcp = &per_cpu(bau_control, cpu); memset(bcp, 0, sizeof(struct bau_control)); pnode = uv_cpu_hub_info(cpu)->pnode; uvhub = uv_cpu_hub_info(cpu)->numa_blade_id; - uvhub_mask |= (1 << uvhub); + *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8)); bdp = &uvhub_descs[uvhub]; bdp->num_cpus++; bdp->uvhub = uvhub; bdp->pnode = pnode; /* kludge: 'assuming' one node per socket, and assuming that disabling a socket just leaves a gap in node numbers */ - socket = (cpu_to_node(cpu) & 1);; + socket = (cpu_to_node(cpu) & 1); bdp->socket_mask |= (1 << socket); sdp = &bdp->socket[socket]; sdp->cpu_number[sdp->num_cpus] = cpu; sdp->num_cpus++; } - uvhub = 0; - while (uvhub_mask) { - if (!(uvhub_mask & 1)) - goto nexthub; + for (uvhub = 0; uvhub < nuvhubs; uvhub++) { + if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8)))) + continue; + have_hmaster = 0; bdp = &uvhub_descs[uvhub]; socket_mask = bdp->socket_mask; socket = 0; @@ -1551,8 +1553,10 @@ static void uv_init_per_cpu(int nuvhubs) bcp->cpu = cpu; if (i == 0) { smaster = bcp; - if (socket == 0) + if (!have_hmaster) { + have_hmaster++; hmaster = bcp; + } } bcp->cpus_in_uvhub = bdp->num_cpus; bcp->cpus_in_socket = sdp->num_cpus; @@ -1566,11 +1570,9 @@ nextsocket: socket++; socket_mask = (socket_mask >> 1); } -nexthub: - uvhub++; - uvhub_mask = (uvhub_mask >> 1); } kfree(uvhub_descs); + kfree(uvhub_mask); for_each_present_cpu(cpu) { bcp = &per_cpu(bau_control, cpu); bcp->baudisabled = 0; -- cgit v1.1 From 1d6225e8cc5598f2bc5c992f9c88b1137763e8e1 Mon Sep 17 00:00:00 2001 From: Cliff Wickman Date: Mon, 9 Aug 2010 16:11:22 -0500 Subject: x86, UV: Make kdump avoid stack dumps - fix !CONFIG_KEXEC breakage This replaces Version 1 of this patch, which broke the build when CONFIG_KEXEC and CONFIG_CRASH_DUMP were configured off. In that case the storage for the 'in_crash_kexec' flag was never built. This version defines that flag as 0 if CONFIG_KEXEC is not set. The patch is tested with all combinations of those two options. Signed-off-by: Cliff Wickman Cc: Andrew Morton LKML-Reference: Signed-off-by: H. Peter Anvin --- arch/x86/include/asm/kdebug.h | 5 +++++ 1 file changed, 5 insertions(+) (limited to 'arch') diff --git a/arch/x86/include/asm/kdebug.h b/arch/x86/include/asm/kdebug.h index 7a2910b..5bdfca8 100644 --- a/arch/x86/include/asm/kdebug.h +++ b/arch/x86/include/asm/kdebug.h @@ -33,6 +33,11 @@ extern void __show_regs(struct pt_regs *regs, int all); extern void show_regs(struct pt_regs *regs); extern unsigned long oops_begin(void); extern void oops_end(unsigned long, struct pt_regs *, int signr); +#ifdef CONFIG_KEXEC extern int in_crash_kexec; +#else +/* no crash dump is ever in progress if no crash kernel can be kexec'd */ +#define in_crash_kexec 0 +#endif #endif /* _ASM_X86_KDEBUG_H */ -- cgit v1.1