runtime: add arenaBaseOffset on aix/ppc64

On AIX, addresses returned by mmap are between 0x0a00000000000000
and 0x0afffffffffffff. The previous solution to handle these large
addresses was to increase the arena size up to 60 bits addresses,
cf CL 138736.

However, with the new page allocator, the 60bit heap addresses are
causing huge memory allocations, especially by (s *pageAlloc).init. mmap
and munmap syscalls dealing with these allocations are reducing
performances of every Go programs.

In order to avoid these allocations, arenaBaseOffset is set to
0x0a00000000000000 and heap addresses are on 48bit, as others operating
systems.

Updates: #35451

Change-Id: Ice916b8578f76703428ec12a82024147a7592bc0
Reviewed-on: https://go-review.googlesource.com/c/go/+/206841
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>

1 files changed, 10 insertions, 12 deletions

diff --git a/src/runtime/malloc.go b/src/runtime/malloc.go
index 39c5fa2a25..47ed470504 100644
--- a/src/runtime/malloc.go
+++ b/src/runtime/malloc.go
@@ -192,10 +192,6 @@ const (
 	// exceed Go's 48 bit limit, it's extremely unlikely in
 	// practice.
 	//
-	// On aix/ppc64, the limits is increased to 1<<60 to accept addresses
-	// returned by mmap syscall. These are in range:
-	//  0x0a00000000000000 - 0x0afffffffffffff
-	//
 	// On 32-bit platforms, we accept the full 32-bit address
 	// space because doing so is cheap.
 	// mips32 only has access to the low 2GB of virtual memory, so
@@ -210,7 +206,7 @@ const (
 	// arenaBaseOffset to offset into the top 4 GiB.
 	//
 	// WebAssembly currently has a limit of 4GB linear memory.
-	heapAddrBits = (_64bit*(1-sys.GoarchWasm)*(1-sys.GoosAix)*(1-sys.GoosDarwin*sys.GoarchArm64))*48 + (1-_64bit+sys.GoarchWasm)*(32-(sys.GoarchMips+sys.GoarchMipsle)) + 60*sys.GoosAix + 33*sys.GoosDarwin*sys.GoarchArm64
+	heapAddrBits = (_64bit*(1-sys.GoarchWasm)*(1-sys.GoosDarwin*sys.GoarchArm64))*48 + (1-_64bit+sys.GoarchWasm)*(32-(sys.GoarchMips+sys.GoarchMipsle)) + 33*sys.GoosDarwin*sys.GoarchArm64
 
 	// maxAlloc is the maximum size of an allocation. On 64-bit,
 	// it's theoretically possible to allocate 1<<heapAddrBits bytes. On
@@ -229,7 +225,6 @@ const (
 	//       Platform  Addr bits  Arena size  L1 entries   L2 entries
 	// --------------  ---------  ----------  ----------  -----------
 	//       */64-bit         48        64MB           1    4M (32MB)
-	//     aix/64-bit         60       256MB        4096    4M (32MB)
 	// windows/64-bit         48         4MB          64    1M  (8MB)
 	//       */32-bit         32         4MB           1  1024  (4KB)
 	//     */mips(le)         31         4MB           1   512  (2KB)
@@ -251,7 +246,7 @@ const (
 	// logHeapArenaBytes is log_2 of heapArenaBytes. For clarity,
 	// prefer using heapArenaBytes where possible (we need the
 	// constant to compute some other constants).
-	logHeapArenaBytes = (6+20)*(_64bit*(1-sys.GoosWindows)*(1-sys.GoosAix)*(1-sys.GoarchWasm)) + (2+20)*(_64bit*sys.GoosWindows) + (2+20)*(1-_64bit) + (8+20)*sys.GoosAix + (2+20)*sys.GoarchWasm
+	logHeapArenaBytes = (6+20)*(_64bit*(1-sys.GoosWindows)*(1-sys.GoarchWasm)) + (2+20)*(_64bit*sys.GoosWindows) + (2+20)*(1-_64bit) + (2+20)*sys.GoarchWasm
 
 	// heapArenaBitmapBytes is the size of each heap arena's bitmap.
 	heapArenaBitmapBytes = heapArenaBytes / (sys.PtrSize * 8 / 2)
@@ -271,10 +266,7 @@ const (
 	// We use the L1 map on 64-bit Windows because the arena size
 	// is small, but the address space is still 48 bits, and
 	// there's a high cost to having a large L2.
-	//
-	// We use the L1 map on aix/ppc64 to keep the same L2 value
-	// as on Linux.
-	arenaL1Bits = 6*(_64bit*sys.GoosWindows) + 12*sys.GoosAix
+	arenaL1Bits = 6 * (_64bit * sys.GoosWindows)
 
 	// arenaL2Bits is the number of bits of the arena number
 	// covered by the second level arena index.
@@ -301,9 +293,15 @@ const (
 	// bits. This offset lets us handle "negative" addresses (or
 	// high addresses if viewed as unsigned).
 	//
+	// On aix/ppc64, this offset allows to keep the heapAddrBits to
+	// 48. Otherwize, it would be 60 in order to handle mmap addresses
+	// (in range 0x0a00000000000000 - 0x0afffffffffffff). But in this
+	// case, the memory reserved in (s *pageAlloc).init for chunks
+	// is causing important slowdowns.
+	//
 	// On other platforms, the user address space is contiguous
 	// and starts at 0, so no offset is necessary.
-	arenaBaseOffset = sys.GoarchAmd64 * (1 << 47)
+	arenaBaseOffset = sys.GoarchAmd64*(1<<47) + (^0x0a00000000000000+1)&uintptrMask*sys.GoosAix
 
 	// Max number of threads to run garbage collection.
 	// 2, 3, and 4 are all plausible maximums depending