On 5/1/23 03:57, Tianyu Lan wrote:
From: Ashish Kalra <ashish.kalra@xxxxxxx>
Add checks in interrupt exit code paths in case of returns
to user mode to check if currently executing the #HV handler
then don't follow the irqentry_exit_to_user_mode path as
that can potentially cause the #HV handler to be
preempted and rescheduled on another CPU. Rescheduled #HV
handler on another cpu will cause interrupts to be handled
on a different cpu than the injected one, causing
invalid EOIs and missed/lost guest interrupts and
corresponding hangs and/or per-cpu IRQs handled on
non-intended cpu.
Signed-off-by: Ashish Kalra <ashish.kalra@xxxxxxx>
This should be merged into one of the appropriate #HV patches and just add
Ashish with a Co-developed-by: tag where appropriate. This would be
appropriate as a separate only if discovered after the series was merged.
---
Change since RFC v3:
* Add check of hv_handling_events in the do_exc_hv()
to avoid nested entry.
---
arch/x86/include/asm/idtentry.h | 66 +++++++++++++++++++++++++++++++++
arch/x86/kernel/sev.c | 37 +++++++++++++++++-
2 files changed, 102 insertions(+), 1 deletion(-)
diff --git a/arch/x86/include/asm/idtentry.h b/arch/x86/include/asm/idtentry.h
index b0f3501b2767..415b7e14c227 100644
--- a/arch/x86/include/asm/idtentry.h
+++ b/arch/x86/include/asm/idtentry.h
@@ -13,6 +13,10 @@
#include <asm/irq_stack.h>
+#ifdef CONFIG_AMD_MEM_ENCRYPT
+noinstr void irqentry_exit_hv_cond(struct pt_regs *regs, irqentry_state_t state);
+#endif
+
/**
* DECLARE_IDTENTRY - Declare functions for simple IDT entry points
* No error code pushed by hardware
@@ -176,6 +180,7 @@ __visible noinstr void func(struct pt_regs *regs, unsigned long error_code)
#define DECLARE_IDTENTRY_IRQ(vector, func) \
DECLARE_IDTENTRY_ERRORCODE(vector, func)
+#ifndef CONFIG_AMD_MEM_ENCRYPT
/**
* DEFINE_IDTENTRY_IRQ - Emit code for device interrupt IDT entry points
* @func: Function name of the entry point
@@ -205,6 +210,26 @@ __visible noinstr void func(struct pt_regs *regs, \
} \
\
static noinline void __##func(struct pt_regs *regs, u32 vector)
+#else
+
+#define DEFINE_IDTENTRY_IRQ(func) \
+static void __##func(struct pt_regs *regs, u32 vector); \
+ \
+__visible noinstr void func(struct pt_regs *regs, \
+ unsigned long error_code) \
+{ \
+ irqentry_state_t state = irqentry_enter(regs); \
+ u32 vector = (u32)(u8)error_code; \
+ \
+ instrumentation_begin(); \
+ kvm_set_cpu_l1tf_flush_l1d(); \
+ run_irq_on_irqstack_cond(__##func, regs, vector); \
+ instrumentation_end(); \
+ irqentry_exit_hv_cond(regs, state); \
+} \
+ \
+static noinline void __##func(struct pt_regs *regs, u32 vector)
+#endif
This seems like a lot of code duplication. Can't the difference (which is
the call to irqentry_exit() vs irqentry_exit_hv_cond() be #ifdef'd in the
current defintion?
Ditto for the changes below, too?
Thanks,
Tom
/**
* DECLARE_IDTENTRY_SYSVEC - Declare functions for system vector entry points
@@ -221,6 +246,7 @@ static noinline void __##func(struct pt_regs *regs, u32 vector)
#define DECLARE_IDTENTRY_SYSVEC(vector, func) \
DECLARE_IDTENTRY(vector, func)
+#ifndef CONFIG_AMD_MEM_ENCRYPT
/**
* DEFINE_IDTENTRY_SYSVEC - Emit code for system vector IDT entry points
* @func: Function name of the entry point
@@ -245,6 +271,26 @@ __visible noinstr void func(struct pt_regs *regs) \
} \
\
static noinline void __##func(struct pt_regs *regs)
+#else
+
+#define DEFINE_IDTENTRY_SYSVEC(func) \
+static void __##func(struct pt_regs *regs); \
+ \
+__visible noinstr void func(struct pt_regs *regs) \
+{ \
+ irqentry_state_t state = irqentry_enter(regs); \
+ \
+ instrumentation_begin(); \
+ kvm_set_cpu_l1tf_flush_l1d(); \
+ run_sysvec_on_irqstack_cond(__##func, regs); \
+ instrumentation_end(); \
+ irqentry_exit_hv_cond(regs, state); \
+} \
+ \
+static noinline void __##func(struct pt_regs *regs)
+#endif
+
+#ifndef CONFIG_AMD_MEM_ENCRYPT
/**
* DEFINE_IDTENTRY_SYSVEC_SIMPLE - Emit code for simple system vector IDT
@@ -274,6 +320,26 @@ __visible noinstr void func(struct pt_regs *regs) \
} \
\
static __always_inline void __##func(struct pt_regs *regs)
+#else
+
+#define DEFINE_IDTENTRY_SYSVEC_SIMPLE(func) \
+static __always_inline void __##func(struct pt_regs *regs); \
+ \
+__visible noinstr void func(struct pt_regs *regs) \
+{ \
+ irqentry_state_t state = irqentry_enter(regs); \
+ \
+ instrumentation_begin(); \
+ __irq_enter_raw(); \
+ kvm_set_cpu_l1tf_flush_l1d(); \
+ __##func(regs); \
+ __irq_exit_raw(); \
+ instrumentation_end(); \
+ irqentry_exit_hv_cond(regs, state); \
+} \
+ \
+static __always_inline void __##func(struct pt_regs *regs)
+#endif
/**
* DECLARE_IDTENTRY_XENCB - Declare functions for XEN HV callback entry point
diff --git a/arch/x86/kernel/sev.c b/arch/x86/kernel/sev.c
index b6becf158598..69b55075ddfe 100644
--- a/arch/x86/kernel/sev.c
+++ b/arch/x86/kernel/sev.c
@@ -149,6 +149,10 @@ struct sev_hv_doorbell_page {
struct sev_snp_runtime_data {
struct sev_hv_doorbell_page hv_doorbell_page;
+ /*
+ * Indication that we are currently handling #HV events.
+ */
+ bool hv_handling_events;
};
static DEFINE_PER_CPU(struct sev_snp_runtime_data*, snp_runtime_data);
@@ -204,6 +208,12 @@ static void do_exc_hv(struct pt_regs *regs)
{
union hv_pending_events pending_events;
+ /* Avoid nested entry. */
+ if (this_cpu_read(snp_runtime_data)->hv_handling_events)
+ return;
+
+ this_cpu_read(snp_runtime_data)->hv_handling_events = true;
+
while (sev_hv_pending()) {
pending_events.events = xchg(
&sev_snp_current_doorbell_page()->pending_events.events,
@@ -218,7 +228,7 @@ static void do_exc_hv(struct pt_regs *regs)
#endif
if (!pending_events.vector)
- return;
+ goto out;
if (pending_events.vector < FIRST_EXTERNAL_VECTOR) {
/* Exception vectors */
@@ -238,6 +248,9 @@ static void do_exc_hv(struct pt_regs *regs)
common_interrupt(regs, pending_events.vector);
}
}
+
+out:
+ this_cpu_read(snp_runtime_data)->hv_handling_events = false;
}
static __always_inline bool on_vc_stack(struct pt_regs *regs)
@@ -2542,3 +2555,25 @@ static int __init snp_init_platform_device(void)
return 0;
}
device_initcall(snp_init_platform_device);
+
+noinstr void irqentry_exit_hv_cond(struct pt_regs *regs, irqentry_state_t state)
+{
+ /*
+ * Check whether this returns to user mode, if so and if
+ * we are currently executing the #HV handler then we don't
+ * want to follow the irqentry_exit_to_user_mode path as
+ * that can potentially cause the #HV handler to be
+ * preempted and rescheduled on another CPU. Rescheduled #HV
+ * handler on another cpu will cause interrupts to be handled
+ * on a different cpu than the injected one, causing
+ * invalid EOIs and missed/lost guest interrupts and
+ * corresponding hangs and/or per-cpu IRQs handled on
+ * non-intended cpu.
+ */
+ if (user_mode(regs) &&
+ this_cpu_read(snp_runtime_data)->hv_handling_events)
+ return;
+
+ /* follow normal interrupt return/exit path */
+ irqentry_exit(regs, state);
+}
