Date: Sat, 1 Oct 2016 14:54:39 +0300 From: Konstantin Belousov <kostikbel@gmail.com> To: Anton Yuzhaninov <citrin@citrin.ru> Cc: "freebsd-fs@freebsd.org" <freebsd-fs@freebsd.org> Subject: Re: UFS: unaligned read from GELI with 8k sectorsize Message-ID: <20161001115439.GY38409@kib.kiev.ua> In-Reply-To: <20161001114536.GX38409@kib.kiev.ua> References: <f84b069b-aeee-ff3c-d4f9-e2fe3caaddb1@citrin.ru> <20161001114536.GX38409@kib.kiev.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sat, Oct 01, 2016 at 02:45:36PM +0300, Konstantin Belousov wrote:
> On Sat, Oct 01, 2016 at 12:42:09AM -0400, Anton Yuzhaninov wrote:
> > Hi all.
> >
> > I'm trying to install FreeBSD with an encrypted root.
> >
> > Main difference from commonly used configuration - I want to use geli
> > with 8k sectorsize.
> >
> > I've created a geli provider, made UFS with 64k block and 8k fragment,
> > extracted files to this FS.
> >
> > While booting from installed system kernel panics after entering geli
> > passphrase:
> >
> > g_vfs_done():ada1p6.eli:[READ(offset=21938548736, length=8192)]error = 22
> > vnode_page_generic_gatpager_done: I/O read error 5
> > inid died (signal 6, exit 0)
> > panic: Going nowhere without my init!
> >
> > errno 22 is EINVAL and it probably returned because 21938548736 is not
> > multiple of 8192 (geli sectorsize).
> >
> > Why UFS tries to read with offset, which is not multiple of frag size?
> > And why this error happens only while mounting geli as root? If I boot
> > FreeBSD from a USB stick I can attach this geli, mount UFS and read all
> > files without problem.
> >
> > FreeBSD 11 / amd64.
>
> FreeBSD vnode pager assumes that it can read at page granularity.
> Since x86 page size is 4k, sometimes page-in has to occur not on the
> fragment boundary. In other words, fragment size > 4k are effectively
> not supported.
>
> Boot needs to execute files from the root mount, which results in the
> mmap(2) attempts on your file system. While mount and reads/writes do
> not involve the pager, which does not trigger further bugs in the
> buffer cache code. It should break if you try to execute badly aligned
> ELF binary from your stick, or just mmap() a file from the stick with
> non-8k aligned offset.
It might be not too hard to make this case working, although the speed
of the pagein will be detrimental. Try this, please.
diff --git a/sys/vm/vnode_pager.c b/sys/vm/vnode_pager.c
index a80a9c2..01e2ea2 100644
--- a/sys/vm/vnode_pager.c
+++ b/sys/vm/vnode_pager.c
@@ -796,7 +796,7 @@ vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count,
* getting pages via VOP_READ.
*/
error = VOP_BMAP(vp, foff / bsize, &bo, &bp->b_blkno, &after, &before);
- if (error == EOPNOTSUPP) {
+ if (error == EOPNOTSUPP || (error == 0 && bo->bo_bsize > PAGE_SIZE)) {
relpbuf(bp, freecnt);
VM_OBJECT_WLOCK(object);
for (i = 0; i < count; i++) {
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20161001115439.GY38409>