x86/irq: Drop redundant irq_{enter,exit}() calls from posted MSI handler
RFC PATCH 1/1 x86/irq: Drop redundant irq_{enter,exit}() calls from posted MSI handler
Drop the unnecessary irq_{enter,exit}() calls from the posted MSI handler, as the wrapper provided by DEFINE_IDTENTRY_SYSVEC() does all of the necessary bookkeeping (the x86 specific entry/exit code provides a superset of functionality).
Note, in the series[1] that added DEFINE_IDTENTRY_SYSVEC(), the follow-up patches that actually use DEFINE_IDTENTRY_SYSVEC() drop all {entering,exiting}_irq() pairs (including entering_ack_irq()), where {entering,exiting}_irq() were little more than wrappers to irq_{enter,exit}().
TODO: Explain what, if anything, can go wrong?
---
RFC purely because I have no idea what can go wrong, and I'd like to document any potential issues in the changelog. I'm confident in the change itself, or I've failed miserably at reading the code. :-)
Found by inspection when looking at a bizarre code fetch from RIP=0 #PF that traces back to the irq_enter() (no reproducer for the #PF at this time). I don't know if the redundant bookkeeping contributed in any way, but the calls to irq_enter() and irq_exit() really stand out, because AFAICT, these are the *only* calls to those APIs in x86.
arch/x86/kernel/irq.c | 2 -- 1 file changed, 2 deletions(-)
arch/x86/kernel/irq.c +0 -2
@@ -430,7 +430,6 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_posted_msi_notification)/* Mark the handler active for intel_ack_posted_msi_irq() */ __this_cpu_write (posted_msi_handler_active, true);inc_irq_stat (POSTED_MSI_NOTIFICATION ); -irq_enter ();/* * Loop only MAX_POSTED_MSI_COALESCING_LOOP - 1 times here to take
@@ -456,7 +455,6 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_posted_msi_notification)handle_pending_pir (pid->pir, regs);apic_eoi (); -irq_exit ();__this_cpu_write (posted_msi_handler_active, false);set_irq_regs (old_regs); }
base-commit: dc59e4fea9d83f03bad6bddf3fa2e52491777482
Signature · 1 line
2.55.0.rc0.799.gd6f94ed593-goog
Quoted 52 lines · Sean Christopherson
Drop the unnecessary irq_{enter,exit}() calls from the posted MSI handler, as the wrapper provided by
DEFINE_IDTENTRY_SYSVEC()does all of the necessary bookkeeping (the x86 specific entry/exit code provides a superset of functionality).Note, in the series[1] that added
DEFINE_IDTENTRY_SYSVEC(), the follow-up patches that actually useDEFINE_IDTENTRY_SYSVEC()drop all {entering,exiting}_irq() pairs (includingentering_ack_irq()), where {entering,exiting}_irq() were little more than wrappers to irq_{enter,exit}().TODO: Explain what, if anything, can go wrong?
Fixes 1b03d82ba15e ("x86/irq: Install posted MSI notification handler")Signed-off-by Sean Christopherson <[email protected]>---
RFC purely because I have no idea what can go wrong, and I'd like to document any potential issues in the changelog. I'm confident in the change itself, or I've failed miserably at reading the code. :-)
Found by inspection when looking at a bizarre code fetch from RIP=0 #PF that traces back to the
irq_enter()(no reproducer for the #PF at this time). I don't know if the redundant bookkeeping contributed in any way, but the calls toirq_enter()andirq_exit()really stand out, because AFAICT, these are the *only* calls to those APIs in x86.arch/x86/kernel/irq.c | 2 -- 1 file changed, 2 deletions(-)arch/x86/kernel/irq.c +0
-2@@ -430,7 +430,6 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_posted_msi_notification)/* Mark the handler active for intel_ack_posted_msi_irq() */ __this_cpu_write (posted_msi_handler_active, true);inc_irq_stat (POSTED_MSI_NOTIFICATION ); -irq_enter ();/* * Loop only MAX_POSTED_MSI_COALESCING_LOOP - 1 times here to take @@ -456,7 +455,6 @@ DEFINE_IDTENTRY_SYSVEC(sysvec_posted_msi_notification)handle_pending_pir (pid->pir, regs);apic_eoi (); -irq_exit ();__this_cpu_write (posted_msi_handler_active, false);set_irq_regs (old_regs); }
Quoted 3 lines
base-commit:
dc59e4fea9d83f03bad6bddf3fa2e52491777482Signature · 1 line
2.55.0.rc0.799.gd6f94ed593-goog
Ping. I don't care terribly about landing the fix, but I would love input on what, if anything, could go wrong.