On Thu, May 17, 2018 at 08:55:04PM -0400, Oleg Drokin wrote:
> Hello!
>
> I just updated to the tip of master for smatch and it seems uninitialized variables detection is way too overzealous.
> While it did find one valid usage, I cannot stop imagining that it basically assumes that if you
> pass a pointer to uninitialized variable somewhere, the variable would remain unitialized at least sometimes.
>
> E.g. this code:
> struct page *ll_get_dir_page(struct inode *dir, struct md_op_data *op_data,
> __u64 offset, struct ll_dir_chain *chain)
> {
> struct md_callback cb_op;
> struct page *page;
> int rc;
>
> cb_op.md_blocking_ast = ll_md_blocking_ast;
> rc = md_read_page(ll_i2mdexp(dir), op_data, &cb_op, offset, &page);
> if (rc != 0)
> return ERR_PTR(rc);
>
> return page;
> }
>
> Produces:
> lustre/llite/dir.c:156 ll_get_dir_page() error: uninitialized symbol 'page’.
I don't get this warning on my tree. I pushed some changes 11 days ago
which we possibly related to this so maybe that made a difference?
Can you run `smatch_data/db/smdb.py return_states md_read_page` from
the top level directory? It should mark the parameter as an
UNTRACKED_PARAM:
drivers/staging/lustre/lustre/lmv/lmv_obd.c | md_read_page | 1191 | s32min-s32max | UNTRACKED_PARAM | 4 | $ | |
You can just rebuild part of the DB to test this:
kchecker --info drivers/staging/lustre/lustre/lmv/lmv_obd.c > warns.txt
~/path/to/smatch_data/db/reload_partial.sh warns.txt
> I went through everythign down the stack and I cannot see how.
>
> The actual code could be seen here:
> https://git.hpdd.intel.com/?p=fs/lustre-release.git;a=blob;f=lustre/llite/dir.c;h=44b174726677b780f01c4bd252d599deb08ac083;hb=refs/heads/master
>
> Am I missing something? This is on by default even without —-spammy
>
> Also is —two-pass supposed to do anything worthwhile? So far the most visible effect is complaining
> how basically every function call return is “unused”.
No. --two-passes isn't useful. It was supposed to warn about unused
returns but I haven't been testing it so I didn't notice when it broke.
In theory, --two-passes is the future because it would make parsing
loops a lot better. But the problem is it makes everything twice as
slow...
regards,
dan carpenter
--
To unsubscribe from this list: send the line "unsubscribe smatch" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
