runtime: mark pages we return to kernel as NOHUGEPAGE

We return memory to the kernel with madvise(..., DONTNEED).
Also mark returned memory with NOHUGEPAGE to keep the kernel from
merging this memory into a huge page, effectively reallocating it.

Only known to be a problem on linux/{386,amd64,amd64p32} at the moment.
It may come up on other os/arch combinations in the future.

Fixes #8832

Change-Id: Ifffc6627a0296926e3f189a8a9b6e4bdb54c79eb
Reviewed-on: https://go-review.googlesource.com/5660Reviewed-by: Dmitry Vyukov <dvyukov@google.com>

src/runtime/arch1_386.go

@@ -12,4 +12,5 @@ const (
 	_PhysPageSize     = goos_nacl*65536 + (1-goos_nacl)*4096 // 4k normally; 64k on NaCl
 	_PCQuantum        = 1
 	_Int64Align       = 4
+	hugePageSize      = 1 << 21
 )

src/runtime/arch1_amd64.go

@@ -12,4 +12,5 @@ const (
 	_PhysPageSize     = 4096
 	_PCQuantum        = 1
 	_Int64Align       = 8
+	hugePageSize      = 1 << 21
 )

src/runtime/arch1_amd64p32.go

@@ -12,4 +12,5 @@ const (
 	_PhysPageSize     = 65536*goos_nacl + 4096*(1-goos_nacl)
 	_PCQuantum        = 1
 	_Int64Align       = 8
+	hugePageSize      = 1 << 21
 )

src/runtime/arch1_arm.go

@@ -12,4 +12,5 @@ const (
 	_PhysPageSize     = 65536*goos_nacl + 4096*(1-goos_nacl)
 	_PCQuantum        = 4
 	_Int64Align       = 4
+	hugePageSize      = 0
 )

src/runtime/arch1_ppc64.go

@@ -12,4 +12,5 @@ const (
 	_PhysPageSize     = 65536
 	_PCQuantum        = 4
 	_Int64Align       = 8
+	hugePageSize      = 0
 )

src/runtime/arch1_ppc64le.go

@@ -12,4 +12,5 @@ const (
 	_PhysPageSize     = 65536
 	_PCQuantum        = 4
 	_Int64Align       = 8
+	hugePageSize      = 0
 )

src/runtime/defs_linux_386.go

@@ -17,7 +17,9 @@ const (
 	_MAP_PRIVATE = 0x2
 	_MAP_FIXED   = 0x10
 
-	_MADV_DONTNEED = 0x4
+	_MADV_DONTNEED   = 0x4
+	_MADV_HUGEPAGE   = 0xe
+	_MADV_NOHUGEPAGE = 0xf
 
 	_SA_RESTART  = 0x10000000
 	_SA_ONSTACK  = 0x8000000

src/runtime/defs_linux_amd64.go

@@ -17,7 +17,9 @@ const (
 	_MAP_PRIVATE = 0x2
 	_MAP_FIXED   = 0x10
 
-	_MADV_DONTNEED = 0x4
+	_MADV_DONTNEED   = 0x4
+	_MADV_HUGEPAGE   = 0xe
+	_MADV_NOHUGEPAGE = 0xf
 
 	_SA_RESTART  = 0x10000000
 	_SA_ONSTACK  = 0x8000000

src/runtime/defs_linux_arm.go

@@ -6,14 +6,19 @@ const (
 	_ENOMEM = 0xc
 	_EAGAIN = 0xb
 
-	_PROT_NONE      = 0
-	_PROT_READ      = 0x1
-	_PROT_WRITE     = 0x2
-	_PROT_EXEC      = 0x4
-	_MAP_ANON       = 0x20
-	_MAP_PRIVATE    = 0x2
-	_MAP_FIXED      = 0x10
-	_MADV_DONTNEED  = 0x4
+	_PROT_NONE  = 0
+	_PROT_READ  = 0x1
+	_PROT_WRITE = 0x2
+	_PROT_EXEC  = 0x4
+
+	_MAP_ANON    = 0x20
+	_MAP_PRIVATE = 0x2
+	_MAP_FIXED   = 0x10
+
+	_MADV_DONTNEED   = 0x4
+	_MADV_HUGEPAGE   = 0xe
+	_MADV_NOHUGEPAGE = 0xf
+
 	_SA_RESTART     = 0x10000000
 	_SA_ONSTACK     = 0x8000000
 	_SA_RESTORER    = 0 // unused on ARM

src/runtime/defs_linux_ppc64.go

@@ -17,7 +17,9 @@ const (
 	_MAP_PRIVATE = 0x2
 	_MAP_FIXED   = 0x10
 
-	_MADV_DONTNEED = 0x4
+	_MADV_DONTNEED   = 0x4
+	_MADV_HUGEPAGE   = 0xe
+	_MADV_NOHUGEPAGE = 0xf
 
 	_SA_RESTART = 0x10000000
 	_SA_ONSTACK = 0x8000000

src/runtime/defs_linux_ppc64le.go

@@ -17,7 +17,9 @@ const (
 	_MAP_PRIVATE = 0x2
 	_MAP_FIXED   = 0x10
 
-	_MADV_DONTNEED = 0x4
+	_MADV_DONTNEED   = 0x4
+	_MADV_HUGEPAGE   = 0xe
+	_MADV_NOHUGEPAGE = 0xf
 
 	_SA_RESTART = 0x10000000
 	_SA_ONSTACK = 0x8000000

src/runtime/mem_linux.go

@@ -67,10 +67,29 @@ func sysAlloc(n uintptr, stat *uint64) unsafe.Pointer {
 }
 
 func sysUnused(v unsafe.Pointer, n uintptr) {
+	if hugePageSize != 0 && (uintptr(v)%hugePageSize != 0 || n%hugePageSize != 0) {
+		// See issue 8832
+		// Linux kernel bug: https://bugzilla.kernel.org/show_bug.cgi?id=93111
+		// Mark the region as NOHUGEPAGE so the kernel's khugepaged
+		// doesn't undo our DONTNEED request.  khugepaged likes to migrate
+		// regions which are only partially mapped to huge pages, including
+		// regions with some DONTNEED marks.  That needlessly allocates physical
+		// memory for our DONTNEED regions.
+		madvise(v, n, _MADV_NOHUGEPAGE)
+	}
 	madvise(v, n, _MADV_DONTNEED)
 }
 
 func sysUsed(v unsafe.Pointer, n uintptr) {
+	if hugePageSize != 0 {
+		// Undo the NOHUGEPAGE marks from sysUnused.  There is no alignment check
+		// around this call as spans may have been merged in the interim.
+		// Note that this might enable huge pages for regions which were
+		// previously disabled.  Unfortunately there is no easy way to detect
+		// what the previous state was, and in any case we probably want huge
+		// pages to back our heap if the kernel can arrange that.
+		madvise(v, n, _MADV_HUGEPAGE)
+	}
 }
 
 func sysFree(v unsafe.Pointer, n uintptr, stat *uint64) {