Aren
/
NetBSD
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Pull up following revision(s) (requested by riastradh in ticket #1623): 

sys/uvm/uvm_map.c: revision 1.396

uvm(9): Fix mmap optimization for topdown case.

PR kern/51393
This commit is contained in:
martin 2023-04-01 16:03:48 +00:00
parent e7f7424d79
commit 6cb45a21d3
1 changed files with 32 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

35
sys/uvm/uvm_map.c
View File

@ -1,4 +1,4 @@
/* $NetBSD: uvm_map.c,v 1.362.2.3 2023/04/01 16:00:28 martin Exp $ */
/* $NetBSD: uvm_map.c,v 1.362.2.4 2023/04/01 16:03:48 martin Exp $ */
/*
* Copyright (c) 1997 Charles D. Cranor and Washington University.
@ -66,7 +66,7 @@
*/
#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: uvm_map.c,v 1.362.2.3 2023/04/01 16:00:28 martin Exp $");
__KERNEL_RCSID(0, "$NetBSD: uvm_map.c,v 1.362.2.4 2023/04/01 16:03:48 martin Exp $");
#include "opt_ddb.h"
#include "opt_pax.h"
@ -1940,7 +1940,36 @@ uvm_map_findspace(struct vm_map *map, vaddr_t hint, vsize_t length,
* it, there would be four cases).
*/
if ((flags & UVM_FLAG_FIXED) == 0 && hint == vm_map_min(map)) {
if ((flags & UVM_FLAG_FIXED) == 0 &&
hint == (topdown ? vm_map_max(map) : vm_map_min(map))) {
/*
* The uvm_map_findspace algorithm is monotonic -- for
* topdown VM it starts with a high hint and returns a
* lower free address; for !topdown VM it starts with a
* low hint and returns a higher free address. As an
* optimization, start with the first (highest for
* topdown, lowest for !topdown) free address.
*
* XXX This `optimization' probably doesn't actually do
* much in practice unless userland explicitly passes
* the VM map's minimum or maximum address, which
* varies from machine to machine (VM_MAX/MIN_ADDRESS,
* e.g. 0x7fbfdfeff000 on amd64 but 0xfffffffff000 on
* aarch64) and may vary according to other factors
* like sysctl vm.user_va0_disable. In particular, if
* the user specifies 0 as a hint to mmap, then mmap
* will choose a default address which is usually _not_
* VM_MAX/MIN_ADDRESS but something else instead like
* VM_MAX_ADDRESS - stack size - guard page overhead,
* in which case this branch is never hit.
*
* In fact, this branch appears to have been broken for
* two decades between when topdown was introduced in
* ~2003 and when it was adapted to handle the topdown
* case without violating the monotonicity assertion in
* 2022. Maybe Someone^TM should either ditch the
* optimization or find a better way to do it.
*/
entry = map->first_free;
} else {
if (uvm_map_lookup_entry(map, hint, &entry)) {