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

evil hack to work around non-fixed mmap requests

Browse Source
This commit is contained in:
cgd 1999-04-27 06:37:12 +00:00
parent c36a5f5190
commit 7e3ad9e963
1 changed files with 67 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

72
sys/compat/osf1/osf1_misc.c
View File

@ -1,4 +1,4 @@
/* $NetBSD: osf1_misc.c,v 1.24 1999/04/26 18:34:05 cgd Exp $ */
/* $NetBSD: osf1_misc.c,v 1.25 1999/04/27 06:37:12 cgd Exp $ */
/*
* Copyright (c) 1999 Christopher G. Demetriou. All rights reserved.
@ -75,6 +75,8 @@
#include <sys/exec.h>
#include <sys/vnode.h>
#include <sys/socketvar.h>
#include <vm/vm.h> /* XXX UVM headers are Cool */
#include <uvm/uvm.h> /* XXX see mmap emulation */
#include <compat/osf1/osf1.h>
#include <compat/osf1/osf1_syscall.h>
@ -305,10 +307,70 @@ osf1_sys_mmap(p, v, retval)
if (leftovers != 0)
return (EINVAL);
#if 0
/* XXX EVIL!!! */
SCARG(&a, flags) |= MAP_FIXED;
#endif
/*
* XXX The following code is evil.
*
* The OSF/1 mmap() function attempts to map non-fixed entries
* near the address that the user specified. Therefore, for
* non-fixed entires we try to find space in the address space
* starting at that address. If the user specified zero, we
* start looking at at least NBPG, so that programs can't
* accidentally live through deferencing NULL.
*
* The need for this kludgery is increased by the fact that
* the loader data segment is mapped at
* (end of user address space) - 1G, MAXDSIZ is 1G, and
* the VM system tries allocate non-fixed mappings _AFTER_
* (start of data) + MAXDSIZ. With the loader, of course,
* that means that it'll start trying at
* (end of user address space), and will never succeed!
*
* Notes:
*
* * Though we find space here, if something else (e.g. a second
* thread) were mucking with the address space the mapping
* we found might be used by another mmap(), and this call
* would clobber that region.
*
* * In general, tricks like this only work for MAP_ANON mappings,
* because of sharing/cache issues. That's not a problem on
* the Alpha, and though it's not good style to abuse that fact,
* there's little choice.
*
* * In order for this to be done right, the VM system should
* really try to use the requested 'addr' passed in to mmap()
* as a hint, even if non-fixed. If it's passed as zero,
* _maybe_ then try (start of data) + MAXDSIZ, or maybe
* provide a better way to avoid the data region altogether.
*/
if ((SCARG(&a, flags) & MAP_FIXED) == 0) {
vaddr_t addr = round_page(SCARG(&a, addr));
vsize_t size = round_page(SCARG(&a, len));
int fixed = 0;
vm_map_lock(&p->p_vmspace->vm_map);
/* if non-NULL address given, start looking there */
if (addr != 0 && uvm_map_findspace(&p->p_vmspace->vm_map,
addr, size, &addr, NULL, 0, 0) != NULL) {
fixed = 1;
goto done;
}
/* didn't find anything. take it again from the top. */
if (uvm_map_findspace(&p->p_vmspace->vm_map, NBPG, size, &addr,
NULL, 0, 0) != NULL) {
fixed = 1;
goto done;
}
done:
vm_map_unlock(&p->p_vmspace->vm_map);
if (fixed) {
SCARG(&a, flags) |= MAP_FIXED;
SCARG(&a, addr) = (void *)addr;
}
}
return sys_mmap(p, &a, retval);
}