When I was doing PTP_SYS_OFFSET_PRECISE ioctl in VM which has 128 vCPUs, I got error returned occasionally. Then I checked the routine of "getcrosststamp". I found in kvm_arch_ptp_get_crosststamp() of x86, pvclock vcpu time info was got from hv_clock arrary which has only 64 elements. Hence this ioctl is executed on vCPU > 64, a wild/dangling pointer will be got, which had casued the error. To confirm this finding, I wrote a simple PTP_SYS_OFFSET_PRECISE ioctl test and used "taskset -c n" to run the test, when it was executed on vCPUs >= 64 it returned error. This patchset exposes this_cpu_pvti() to get per cpu pvclock vcpu time info of vCPUs >= 64 insdead of getting them from hv_clock arrary.
Zelin Deng (2): x86/kvmclock: Move this_cpu_pvti into kvmclock.h ptp: Fix ptp_kvm_getcrosststamp issue for x86 ptp_kvm
arch/x86/include/asm/kvmclock.h | 14 ++++++++++++++ arch/x86/kernel/kvmclock.c | 13 ++----------- drivers/ptp/ptp_kvm_x86.c | 9 ++------- 3 files changed, 18 insertions(+), 18 deletions(-)
There're other modules might use hv_clock_per_cpu variable like ptp_kvm, so move it into kvmclock.h and export the symbol to make it visiable to other modules.
Signed-off-by: Zelin Deng zelin.deng@linux.alibaba.com Cc: stable@vger.kernel.org --- arch/x86/include/asm/kvmclock.h | 14 ++++++++++++++ arch/x86/kernel/kvmclock.c | 13 ++----------- 2 files changed, 16 insertions(+), 11 deletions(-)
diff --git a/arch/x86/include/asm/kvmclock.h b/arch/x86/include/asm/kvmclock.h index eceea92..fcd1ad6 100644 --- a/arch/x86/include/asm/kvmclock.h +++ b/arch/x86/include/asm/kvmclock.h @@ -2,6 +2,20 @@ #ifndef _ASM_X86_KVM_CLOCK_H #define _ASM_X86_KVM_CLOCK_H
+#include <linux/percpu.h> + extern struct clocksource kvm_clock;
+extern struct pvclock_vsyscall_time_info *hv_clock_per_cpu; + +static inline struct pvclock_vcpu_time_info *this_cpu_pvti(void) +{ + return &this_cpu_read(hv_clock_per_cpu)->pvti; +} + +static inline struct pvclock_vsyscall_time_info *this_cpu_hvclock(void) +{ + return this_cpu_read(hv_clock_per_cpu); +} + #endif /* _ASM_X86_KVM_CLOCK_H */ diff --git a/arch/x86/kernel/kvmclock.c b/arch/x86/kernel/kvmclock.c index ad273e5..6c6b1b3 100644 --- a/arch/x86/kernel/kvmclock.c +++ b/arch/x86/kernel/kvmclock.c @@ -49,18 +49,9 @@ static int __init parse_no_kvmclock_vsyscall(char *arg) static struct pvclock_vsyscall_time_info hv_clock_boot[HVC_BOOT_ARRAY_SIZE] __bss_decrypted __aligned(PAGE_SIZE); static struct pvclock_wall_clock wall_clock __bss_decrypted; -static DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu); static struct pvclock_vsyscall_time_info *hvclock_mem; - -static inline struct pvclock_vcpu_time_info *this_cpu_pvti(void) -{ - return &this_cpu_read(hv_clock_per_cpu)->pvti; -} - -static inline struct pvclock_vsyscall_time_info *this_cpu_hvclock(void) -{ - return this_cpu_read(hv_clock_per_cpu); -} +DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu); +EXPORT_SYMBOL_GPL(hv_clock_per_cpu);
/* * The wallclock is the time of day when we booted. Since then, some time may
If PTP_SYS_OFFSET_PRECISE ioctl is executed on vCPU >= 64, struct pvclock_vcpu_time_info *src which is got by "src = &hv_clock[cpu].pvti" could be wild/dangling pointer, as hv_clock arrary has only HVC_BOOT_ARRAY_SIZE (64) elements. Therefore let's change it to "this_cpu_pvti()"
Fixes: 95a3d4454bb1 ("Switch kvmclock data to a PER_CPU variable") Signed-off-by: Zelin Deng zelin.deng@linux.alibaba.com Cc: stable@vger.kernel.org --- drivers/ptp/ptp_kvm_x86.c | 9 ++------- 1 file changed, 2 insertions(+), 7 deletions(-)
diff --git a/drivers/ptp/ptp_kvm_x86.c b/drivers/ptp/ptp_kvm_x86.c index 3dd519d..d0096cd 100644 --- a/drivers/ptp/ptp_kvm_x86.c +++ b/drivers/ptp/ptp_kvm_x86.c @@ -15,8 +15,6 @@ #include <linux/ptp_clock_kernel.h> #include <linux/ptp_kvm.h>
-struct pvclock_vsyscall_time_info *hv_clock; - static phys_addr_t clock_pair_gpa; static struct kvm_clock_pairing clock_pair;
@@ -28,8 +26,7 @@ int kvm_arch_ptp_init(void) return -ENODEV;
clock_pair_gpa = slow_virt_to_phys(&clock_pair); - hv_clock = pvclock_get_pvti_cpu0_va(); - if (!hv_clock) + if (!pvclock_get_pvti_cpu0_va()) return -ENODEV;
ret = kvm_hypercall2(KVM_HC_CLOCK_PAIRING, clock_pair_gpa, @@ -64,10 +61,8 @@ int kvm_arch_ptp_get_crosststamp(u64 *cycle, struct timespec64 *tspec, struct pvclock_vcpu_time_info *src; unsigned int version; long ret; - int cpu;
- cpu = smp_processor_id(); - src = &hv_clock[cpu].pvti; + src = this_cpu_pvti();
do { /*
On 29/09/21 07:13, Zelin Deng wrote:
When I was doing PTP_SYS_OFFSET_PRECISE ioctl in VM which has 128 vCPUs, I got error returned occasionally. Then I checked the routine of "getcrosststamp". I found in kvm_arch_ptp_get_crosststamp() of x86, pvclock vcpu time info was got from hv_clock arrary which has only 64 elements. Hence this ioctl is executed on vCPU > 64, a wild/dangling pointer will be got, which had casued the error. To confirm this finding, I wrote a simple PTP_SYS_OFFSET_PRECISE ioctl test and used "taskset -c n" to run the test, when it was executed on vCPUs >= 64 it returned error. This patchset exposes this_cpu_pvti() to get per cpu pvclock vcpu time info of vCPUs >= 64 insdead of getting them from hv_clock arrary.
Zelin Deng (2): x86/kvmclock: Move this_cpu_pvti into kvmclock.h ptp: Fix ptp_kvm_getcrosststamp issue for x86 ptp_kvm
arch/x86/include/asm/kvmclock.h | 14 ++++++++++++++ arch/x86/kernel/kvmclock.c | 13 ++----------- drivers/ptp/ptp_kvm_x86.c | 9 ++------- 3 files changed, 18 insertions(+), 18 deletions(-)
Queued, thanks.
Paolo
linux-stable-mirror@lists.linaro.org
-
Paolo Bonzini -
Zelin Deng