[RFC PATCH 37/37] arm64/mm: attempt speculative mm faults first

[Michel Lespinasse <michel@lespinasse.org>]

Attempt speculative mm fault handling first, and fall back to the
existing (non-speculative) code if that fails.

This follows the lines of the x86 speculative fault handling code,
but with some minor arch differences such as the way that the
VM_FAULT_BADACCESS case is handled.

Signed-off-by: Michel Lespinasse <michel@lespinasse.org>
---
 arch/arm64/mm/fault.c | 52 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index f37d4e3830b7..3757bfbb457a 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c
@@ -25,6 +25,7 @@
 #include <linux/perf_event.h>
 #include <linux/preempt.h>
 #include <linux/hugetlb.h>
+#include <linux/vm_event_item.h>
 
 #include <asm/acpi.h>
 #include <asm/bug.h>
@@ -530,6 +531,9 @@ static int __kprobes do_page_fault(unsigned long far, unsigned int esr,
 	unsigned long vm_flags = VM_ACCESS_FLAGS;
 	unsigned int mm_flags = FAULT_FLAG_DEFAULT;
 	unsigned long addr = untagged_addr(far);
+	struct vm_area_struct *vma;
+	struct vm_area_struct pvma;
+	unsigned long seq;
 
 	if (kprobe_page_fault(regs, esr))
 		return 0;
@@ -564,6 +568,53 @@ static int __kprobes do_page_fault(unsigned long far, unsigned int esr,
 
 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, addr);
 
+	/* Only try spf for multithreaded user space faults. */
+	if (!(mm_flags & FAULT_FLAG_USER) || atomic_read(&mm->mm_users) == 1)
+		goto no_spf;
+
+	count_vm_event(SPF_ATTEMPT);
+	seq = mmap_seq_read_start(mm);
+	if (seq & 1) {
+		count_vm_spf_event(SPF_ABORT_ODD);
+		goto spf_abort;
+	}
+	rcu_read_lock();
+	vma = find_vma(mm, addr);
+	if (!vma || vma->vm_start > addr) {
+		rcu_read_unlock();
+		count_vm_spf_event(SPF_ABORT_UNMAPPED);
+		goto spf_abort;
+	}
+	if (!vma_can_speculate(vma, mm_flags)) {
+		rcu_read_unlock();
+		count_vm_spf_event(SPF_ABORT_NO_SPECULATE);
+		goto spf_abort;
+	}
+	pvma = *vma;
+	rcu_read_unlock();
+	if (!mmap_seq_read_check(mm, seq, SPF_ABORT_VMA_COPY))
+		goto spf_abort;
+	vma = &pvma;
+	if (!(vma->vm_flags & vm_flags)) {
+		count_vm_spf_event(SPF_ABORT_ACCESS_ERROR);
+		goto spf_abort;
+	}
+	fault = do_handle_mm_fault(vma, addr & PAGE_MASK,
+			mm_flags | FAULT_FLAG_SPECULATIVE, seq, regs);
+
+	/* Quick path to respond to signals */
+	if (fault_signal_pending(fault, regs)) {
+		if (!user_mode(regs))
+			goto no_context;
+		return 0;
+	}
+	if (!(fault & VM_FAULT_RETRY))
+		goto done;
+
+spf_abort:
+	count_vm_event(SPF_ABORT);
+no_spf:
+
 	/*
 	 * As per x86, we may deadlock here. However, since the kernel only
 	 * validly references user space from well defined areas of the code,
@@ -604,6 +655,7 @@ static int __kprobes do_page_fault(unsigned long far, unsigned int esr,
 		}
 	}
 	mmap_read_unlock(mm);
+done:
 
 	/*
 	 * Handle the "normal" (no error) case first.
-- 
2.20.1