Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Nadav Amit <namit@vmware.com>
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     Andy Lutomirski <luto@kernel.org>, x86@kernel.org,
        linux-kernel@vger.kernel.org,
        Dave Hansen <dave.hansen@linux.intel.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>, Nadav Amit <namit@vmware.com>
Subject: [RFC 5/7] percpu: Assume preemption is disabled on per_cpu_ptr()
Date:   Thu, 18 Jul 2019 10:41:08 -0700
Message-Id: <20190718174110.4635-6-namit@vmware.com>
In-Reply-To: <20190718174110.4635-1-namit@vmware.com>
References: <20190718174110.4635-1-namit@vmware.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

When per_cpu_ptr() is used, the caller should have preemption disabled,
as otherwise the pointer is meaningless. If the user wants an "unstable"
pointer he should call raw_cpu_ptr().

Add an assertion to check that indeed preemption is disabled, and
distinguish between the two cases to allow further, per-arch
optimizations.

Signed-off-by: Nadav Amit <namit@vmware.com>
---
 include/asm-generic/percpu.h | 12 ++++++++++++
 include/linux/percpu-defs.h  | 33 ++++++++++++++++++++++++++++++++-
 2 files changed, 44 insertions(+), 1 deletion(-)

diff --git a/include/asm-generic/percpu.h b/include/asm-generic/percpu.h
index c2de013b2cf4..7853605f4210 100644
--- a/include/asm-generic/percpu.h
+++ b/include/asm-generic/percpu.h
@@ -36,6 +36,14 @@ extern unsigned long __per_cpu_offset[NR_CPUS];
 #define my_cpu_offset __my_cpu_offset
 #endif
 
+/*
+ * Determine the offset of the current active processor when preemption is
+ * disabled. Can be overriden by arch code.
+ */
+#ifndef __raw_my_cpu_offset
+#define __raw_my_cpu_offset __my_cpu_offset
+#endif
+
 /*
  * Arch may define arch_raw_cpu_ptr() to provide more efficient address
  * translations for raw_cpu_ptr().
@@ -44,6 +52,10 @@ extern unsigned long __per_cpu_offset[NR_CPUS];
 #define arch_raw_cpu_ptr(ptr) SHIFT_PERCPU_PTR(ptr, __my_cpu_offset)
 #endif
 
+#ifndef arch_raw_cpu_ptr_preemptable
+#define arch_raw_cpu_ptr_preemptable(ptr) SHIFT_PERCPU_PTR(ptr, __raw_my_cpu_offset)
+#endif
+
 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
 extern void setup_per_cpu_areas(void);
 #endif
diff --git a/include/linux/percpu-defs.h b/include/linux/percpu-defs.h
index a6fabd865211..13afca8a37e7 100644
--- a/include/linux/percpu-defs.h
+++ b/include/linux/percpu-defs.h
@@ -237,20 +237,51 @@ do {									\
 	SHIFT_PERCPU_PTR((ptr), per_cpu_offset((cpu)));			\
 })
 
+#ifndef arch_raw_cpu_ptr_preemption_disabled
+#define arch_raw_cpu_ptr_preemption_disabled(ptr)			\
+	arch_raw_cpu_ptr(ptr)
+#endif
+
+#define raw_cpu_ptr_preemption_disabled(ptr)				\
+({									\
+	__verify_pcpu_ptr(ptr);						\
+	arch_raw_cpu_ptr_preemption_disabled(ptr);			\
+})
+
+/*
+ * If preemption is enabled, we need to read the pointer atomically on
+ * raw_cpu_ptr(). However if it is disabled, we can use the
+ * raw_cpu_ptr_nopreempt(), which is potentially more efficient. Similarly, we
+ * can use the preemption-disabled version if the kernel is non-preemptable or
+ * if voluntary preemption is used.
+ */
+#ifdef CONFIG_PREEMPT
+
 #define raw_cpu_ptr(ptr)						\
 ({									\
 	__verify_pcpu_ptr(ptr);						\
 	arch_raw_cpu_ptr(ptr);						\
 })
 
+#else
+
+#define raw_cpu_ptr(ptr)	raw_cpu_ptr_preemption_disabled(ptr)
+
+#endif
+
 #ifdef CONFIG_DEBUG_PREEMPT
+/*
+ * Unlike other this_cpu_* operations, this_cpu_ptr() requires that preemption
+ * will be disabled. In contrast, raw_cpu_ptr() does not require that.
+ */
 #define this_cpu_ptr(ptr)						\
 ({									\
+	__this_cpu_preempt_check("ptr");				\
 	__verify_pcpu_ptr(ptr);						\
 	SHIFT_PERCPU_PTR(ptr, my_cpu_offset);				\
 })
 #else
-#define this_cpu_ptr(ptr) raw_cpu_ptr(ptr)
+#define this_cpu_ptr(ptr) raw_cpu_ptr_preemption_disabled(ptr)
 #endif
 
 #else	/* CONFIG_SMP */
-- 
2.17.1