x86: Add a __copy_from_user_nmi

From: Andi Kleen <[email protected]>

Add a inlined __ variant of copy_from_user_nmi. The inlined variant allows the user to:

- batch the access_ok check for multiple accesses
- avoid having a pagefault_disable/enable on every access if the caller already ensures disabled page faults due to its context.
- get all the optimizations in copy_*_user for small constant sized transfers

It is just a define to __copy_from_user_inatomic.

---
 arch/x86/include/asm/uaccess.h | 9 +++++++++
 1 file changed, 9 insertions(+)

index a8df874..1d0766c 100644
@@ -745,5 +745,14 @@ copy_to_user(void __user *to, const void *from, unsigned long n)
 #undef __copy_from_user_overflow
 #undef __copy_to_user_overflow
 
+/*
+ * We rely on the nested NMI work to allow atomic faults from the NMI path; the
+ * nested NMI paths are careful to preserve CR2.
+ *
+ * Caller must use pagefault_enable/disable, or run in interrupt context,
+ * and also do a uaccess_ok() check
+ */
+#define __copy_from_user_nmi __copy_from_user_inatomic
+
 #endif /* _ASM_X86_UACCESS_H */
 

2.4.3

From: Andi Kleen <[email protected]>

Change the perf user stack walking to use the new __copy_from_user_nmi, and split each access into word sized transfer sizes. This allows to inline the complete access and optimize it all into a single load.

The main advantage is that this avoids the overhead of double page faults. When normal copy_from_user fails it reexecutes the copy to compute an accurate number of non copied bytes. This leads to executing the expensive page fault twice.

While walking stacks having a fault at some point is relatively common (typically when some part of the program isn't compiled with frame pointers), so this is a large overhead.

With the optimized copies we avoid this problem because they only do all accesses once. And of course they're much faster too when the access does not fault because they're just single instructions instead of complex function calls.

While profiling a kernel build with -g, the patch brings down the average time of the PMI handler from 966ns to 552ns (-43%)

---
 arch/x86/kernel/cpu/perf_event.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)

index 4562cf0..15bf8b3 100644
@@ -2255,7 +2255,12 @@ perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
 		frame.next_frame     = 0;
 		frame.return_address = 0;
 
-		bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
+		if (!access_ok(VERIFY_READ, fp, 8))
+			break;
+		bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
+		if (bytes != 0)
+			break;
+		bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
 		if (bytes != 0)
 			break;
 

@@ -2307,7 +2312,13 @@ perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
 		frame.next_frame	     = NULL;
 		frame.return_address = 0;
 
-		bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
+		if (!access_ok(VERIFY_READ, fp, 16))
+			break;
+
+		bytes = __copy_from_user_nmi(&frame.next_frame, fp, 8);
+		if (bytes != 0)
+			break;
+		bytes = __copy_from_user_nmi(&frame.return_address, fp+8, 8);
 		if (bytes != 0)
 			break;
 

2.4.3

That's true. I'll add the pagefault_disable/enable.

-Andi