powerpc already uses GENERIC_ENTRY for the irq entry/exit path. That
means the generic entry code handles all the arch-independent
thread_info flags (TIF_SIGPENDING, TIF_NOTIFY_RESUME, TIF_NOTIFY_SIGNAL,
TIF_NEED_RESCHED, TIF_NEED_RESCHED_LAZY, TIF_UPROBE, TIF_PATCH_PENDING,
TIF_MEMDIE, TIF_RSEQ, TIF_HRTIMER_REARM, TIF_USER_RETURN_NOTIFY) on
behalf of the arch. The only remaining blocker was that powerpc defined
its own bit numbers for those flags, which conflicted with the layout
required by the common header include/asm-generic/thread_info_tif.h.
This patch selects HAVE_GENERIC_TIF_BITS and switches powerpc's
thread_info.h over to include <asm-generic/thread_info_tif.h> for the
arch-independent bits, using HAVE_TIF_NEED_RESCHED_LAZY to opt in to
the lazy-preemption flag. The powerpc-private bits (TIF_RESTORE_TM,
TIF_SINGLESTEP, TIF_RESTOREALL, TIF_NOERROR, TIF_EMULATE_STACK_STORE,
TIF_ELF2ABI, TIF_POLLING_NRFLAG, TIF_32BIT, TIF_SYSCALL_RET) are
renumbered to start at bit 16, which is the reserved range for arch use.
Enabling HAVE_GENERIC_TIF_BITS unlocks the following features that were
previously unavailable on powerpc because their Kconfig dependencies
could not be satisfied:
1. RSEQ_SLICE_EXTENSION (CONFIG_RSEQ_SLICE_EXTENSION)
Allows userspace to register an RSEQ critical section and request
a limited time-slice extension when the task is about to be
preempted. The kernel grants the extension (via the RSEQ shared
data ABI) so the critical section can complete, avoiding the
thundering-herd problem where many threads spin on the same
lock/resource after a preemption. This feature requires a
dedicated TIF_RSEQ bit, which only exists when HAVE_GENERIC_TIF_BITS
is set. Depends on: RSEQ && HIGH_RES_TIMERS && GENERIC_ENTRY &&
HAVE_GENERIC_TIF_BITS.
2. HRTIMER_REARM_DEFERRED (CONFIG_HRTIMER_REARM_DEFERRED)
Defers re-arming of the hrtimer interrupt to the return-to-user
path, using the dedicated TIF_HRTIMER_REARM bit. This avoids
redundant hrtimer_interrupt() invocations within the same tick
period and reduces latency jitter for high-resolution timer users
(e.g., PREEMPT_RT workloads, audio, real-time tasks). Depends on:
GENERIC_ENTRY && HAVE_GENERIC_TIF_BITS && HIGH_RES_TIMERS &&
SCHED_HRTICK.
3. Dedicated TIF_RSEQ flag (fast-path RSEQ handling)
Without HAVE_GENERIC_TIF_BITS, TIF_RSEQ falls back to aliasing
TIF_NOTIFY_RESUME, forcing the generic entry slow-path to run on
every return-to-user for all RSEQ-registered threads. With the
dedicated bit, the fast path can skip RSEQ work unless the flag
is actually set, reducing overhead in the common case.
4. TIF_USER_RETURN_NOTIFY
The generic bit for the user-return-notifier subsystem
(kernel/user-return-notifier.c). Used by KVM on x86 to lazily
restore guest MSRs; having the generic bit available means future
powerpc KVM work can use the same infrastructure.
The arch-specific bits that do not exist in the generic header retain
their private numbering in the powerpc-reserved range (bits 16+), so
no existing powerpc functionality is affected.
Tested on powerpc (POWER11 pSeries CONFIG_RSEQ=y +
CONFIG_RSEQ_SLICE_EXTENSION=y + CONFIG_HRTIMER_REARM_DEFERRED=y):
* tools/testing/selftests/rseq/ -- all tests pass (basic_test,
param_test, basic_percpu_ops_test, syscall_errors_test,
slice_test via run_timeslice_test.sh)
* tools/testing/selftests/timers/hrtimer-rearm.sh -- deferred=1
events observed; no new dmesg warnings
* Boot and basic workload testing shows no regression
Signed-off-by: Mukesh Kumar Chaurasiya (IBM) <[email protected]>
---
This patch is applicabe over
https://lore.kernel.org/all/961c058f3f3e03045552186c083a68a0fde3f43f.1791975972.git.msucha...@suse.de/
Changelog:
V1 -> V2:
- Commit message updated
V1: https://lore.kernel.org/all/[email protected]/
arch/powerpc/Kconfig | 1 +
arch/powerpc/include/asm/thread_info.h | 36 +++++++++-----------------
2 files changed, 13 insertions(+), 24 deletions(-)
diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig
index 7f7910f38a67..e5527b56159a 100644
--- a/arch/powerpc/Kconfig
+++ b/arch/powerpc/Kconfig
@@ -255,6 +255,7 @@ config PPC
select HAVE_FUNCTION_GRAPH_TRACER
select HAVE_FUNCTION_TRACER if !COMPILE_TEST && (PPC64 ||
(PPC32 && CC_IS_GCC))
select HAVE_GCC_PLUGINS
+ select HAVE_GENERIC_TIF_BITS
select HAVE_HARDLOCKUP_DETECTOR_ARCH if PPC_BOOK3S_64 && SMP
select HAVE_HARDLOCKUP_DETECTOR_PERF if PERF_EVENTS &&
HAVE_PERF_EVENTS_NMI
select HAVE_HW_BREAKPOINT if PERF_EVENTS && (PPC_BOOK3S
|| PPC_8xx)
diff --git a/arch/powerpc/include/asm/thread_info.h
b/arch/powerpc/include/asm/thread_info.h
index 4b7ec1db3034..4df3f4b5bd56 100644
--- a/arch/powerpc/include/asm/thread_info.h
+++ b/arch/powerpc/include/asm/thread_info.h
@@ -93,44 +93,32 @@ void arch_setup_new_exec(void);
#endif /* __ASSEMBLER__ */
+#define HAVE_TIF_NEED_RESCHED_LAZY
+
+#include <asm-generic/thread_info_tif.h>
/*
* thread information flag bit numbers
*/
-#define TIF_SIGPENDING 1 /* signal pending */
-#define TIF_NEED_RESCHED 2 /* rescheduling necessary */
-#define TIF_NOTIFY_SIGNAL 3 /* signal notifications exist */
-#define TIF_RESTORE_TM 5 /* need to restore TM FP/VEC/VSX */
-#define TIF_PATCH_PENDING 6 /* pending live patching update */
-#define TIF_SINGLESTEP 8 /* singlestepping active */
-#define TIF_NEED_RESCHED_LAZY 9 /* Scheduler driven lazy preemption */
-#define TIF_RESTOREALL 11 /* Restore all regs (implies NOERROR) */
-#define TIF_NOERROR 12 /* Force successful syscall return */
-#define TIF_NOTIFY_RESUME 13 /* callback before returning to user */
-#define TIF_UPROBE 14 /* breakpointed or single-stepping */
-#define TIF_EMULATE_STACK_STORE 16 /* Is an instruction emulation
+#define TIF_RESTORE_TM 16 /* need to restore TM FP/VEC/VSX */
+#define TIF_SINGLESTEP 17 /* singlestepping active */
+#define TIF_RESTOREALL 18 /* Restore all regs (implies NOERROR) */
+#define TIF_NOERROR 19 /* Force successful syscall return */
+#define TIF_EMULATE_STACK_STORE 20 /* Is an instruction emulation
for stack store? */
-#define TIF_MEMDIE 17 /* is terminating due to OOM killer */
#if defined(CONFIG_PPC64)
-#define TIF_ELF2ABI 18 /* function descriptors must die! */
+#define TIF_ELF2ABI 21 /* function descriptors must die! */
#endif
-#define TIF_POLLING_NRFLAG 19 /* true if poll_idle() is polling
TIF_NEED_RESCHED */
-#define TIF_32BIT 20 /* 32 bit binary */
-#define TIF_SYSCALL_RET 21 /* syscall error value set */
+#define TIF_POLLING_NRFLAG 22 /* true if poll_idle() is polling
TIF_NEED_RESCHED */
+#define TIF_32BIT 23 /* 32 bit binary */
+#define TIF_SYSCALL_RET 24 /* syscall error value set */
/* as above, but as bit values */
-#define _TIF_SIGPENDING (1<<TIF_SIGPENDING)
-#define _TIF_NEED_RESCHED (1<<TIF_NEED_RESCHED)
-#define _TIF_NEED_RESCHED_LAZY (1<<TIF_NEED_RESCHED_LAZY)
-#define _TIF_NOTIFY_SIGNAL (1<<TIF_NOTIFY_SIGNAL)
#define _TIF_POLLING_NRFLAG (1<<TIF_POLLING_NRFLAG)
#define _TIF_32BIT (1<<TIF_32BIT)
#define _TIF_RESTORE_TM (1<<TIF_RESTORE_TM)
-#define _TIF_PATCH_PENDING (1<<TIF_PATCH_PENDING)
#define _TIF_SINGLESTEP (1<<TIF_SINGLESTEP)
#define _TIF_RESTOREALL (1<<TIF_RESTOREALL)
#define _TIF_NOERROR (1<<TIF_NOERROR)
-#define _TIF_NOTIFY_RESUME (1<<TIF_NOTIFY_RESUME)
-#define _TIF_UPROBE (1<<TIF_UPROBE)
#define _TIF_EMULATE_STACK_STORE (1<<TIF_EMULATE_STACK_STORE)
#define _TIF_USER_WORK_MASK (_TIF_SIGPENDING | _TIF_NEED_RESCHED | \
--
2.56.0
