On Thu, Jun 29, 2006 at 02:36:38PM -0400, Michael wrote:
> Instead it borks here:
> trap type 0x7: pc=0xf0060274 npc=0xf0060278 psr=48001c3<S,PS>
> kernel: alignment fault trap
> stopped in pid 41.1 (cp) at netbsd:tmpfs_vptofh+0x14:
> std %g2, [%o1 + 8]
> db> bt
> tmpfs_vptofh(0x2d, 0xf44d5be4, 0xf3a20478, 0xf3a2c3b8, 0x0, 0xf3a2ee70) 
> at netbsd:nqsrv_getlease+0x288

This is easy:

struct fid {
        unsigned short  fid_len;                /* length of data in bytes */
        unsigned short  fid_reserved;           /* force longword alignment */
        char            fid_data[_VFS_MAXFIDSZ];/* data (variable length) */
};

If I calculated correctly, required alignement is 2 byte.

Now for tmpfs:

struct tmpfs_fid {
        uint16_t                tf_len;
        uint16_t                tf_pad;
        uint32_t                tf_gen;
        ino_t                   tf_id;
};

Required alignement here is that of ino_t, so 8 byte.

Now the code filling the fid is:

static int
tmpfs_vptofh(struct vnode *vp, struct fid *fhp)
{
        struct tmpfs_fid *tfhp;
        struct tmpfs_node *node;
        
        tfhp = (struct tmpfs_fid *)fhp;
        node = VP_TO_TMPFS_NODE(vp);

        tfhp->tf_len = sizeof(struct tmpfs_fid);
        tfhp->tf_id = node->tn_id;
        tfhp->tf_gen = node->tn_gen;

        return 0;
}

And if you look at the assembler code, the 

        tfhp->tf_id = node->tn_id;

line is implemented as the faulting "std" instruction.

Not gcc's fault - we have to fix it.

Martin