[patch] tmpfs mmap sync bug fix
Gleb Kurtsou
gleb.kurtsou at gmail.com
Fri Sep 18 17:57:11 UTC 2009
[ pjd@ CCed to make sure it doesn't apply to zfs ]
Mmaped pages can get out of sync in tmpfs. The bug is 100% reproducible
by:
# fsx -S 125 -d /tmpfs/file
It breaks at operation 42.
Fix is inspired by zfs, it calls vm_page_cache_free(). Reading zfs
sources, it looks like it doesn't check v_object->cache, but never the
less bug never shows up on there. Probably it's because of zfs using
VOP_BMAP to do page mapping. tmpfs uses default
vop_getpages/vop_putpages which invokes vop_read/vop_write accordingly.
Removing v_object->cache == NULL checks breaks things again.
I'm not entirely sure if it's correct fix and would appreciate vm guru
having a look at it, because I do the same thing in my pefs filesystem
and It works as expected for me:
http://blogs.freebsdish.org/gleb/2009/09/16/pefs-benchmark/
Thank,
Gleb.
-------------- next part --------------
diff --git a/sys/fs/tmpfs/tmpfs_vnops.c b/sys/fs/tmpfs/tmpfs_vnops.c
index db8ceea..59d94d7 100644
--- a/sys/fs/tmpfs/tmpfs_vnops.c
+++ b/sys/fs/tmpfs/tmpfs_vnops.c
@@ -444,7 +444,8 @@ tmpfs_mappedread(vm_object_t vobj, vm_object_t tobj, size_t len, struct uio *uio
offset = addr & PAGE_MASK;
tlen = MIN(PAGE_SIZE - offset, len);
- if ((vobj == NULL) || (vobj->resident_page_count == 0))
+ if ((vobj == NULL) ||
+ (vobj->resident_page_count == 0 && vobj->cache == NULL))
goto nocache;
VM_OBJECT_LOCK(vobj);
@@ -555,7 +556,8 @@ tmpfs_mappedwrite(vm_object_t vobj, vm_object_t tobj, size_t len, struct uio *ui
offset = addr & PAGE_MASK;
tlen = MIN(PAGE_SIZE - offset, len);
- if ((vobj == NULL) || (vobj->resident_page_count == 0)) {
+ if ((vobj == NULL) ||
+ (vobj->resident_page_count == 0 && vobj->cache == NULL)) {
vpg = NULL;
goto nocache;
}
@@ -573,6 +575,8 @@ lookupvpg:
VM_OBJECT_UNLOCK(vobj);
error = uiomove_fromphys(&vpg, offset, tlen, uio);
} else {
+ if (__predict_false(vobj->cache != NULL))
+ vm_page_cache_free(vobj, idx, idx + 1);
VM_OBJECT_UNLOCK(vobj);
vpg = NULL;
}
More information about the freebsd-fs
mailing list