On Fri, Jan 17, 2020 at 05:41:31PM +0100, Simon Goldschmidt wrote:
> + Jordy, who just found a bug here...
>
> Am 08.11.2019 um 17:24 schrieb Tom Rini:
> > New analysis by the tool has shown that we have some cases where we
> > weren't handling the error exit condition correctly. When we ran into
> > the ENOMEM case we wouldn't exit the function and thus incorrect things
> > could happen. Rework the unwinding such that we don't need a helper
> > function now and free what we may have allocated.
> >
> > Fixes: 18030d04d25d ("GPT: fix memory leaks identified by Coverity")
> > Reported-by: Coverity (CID: 275475, 275476)
> > Cc: Alison Chaiken <[email protected]>
> > Signed-off-by: Tom Rini <[email protected]>
> > ---
> > cmd/gpt.c | 35 ++++++++---------------------------
> > 1 file changed, 8 insertions(+), 27 deletions(-)
> >
> > diff --git a/cmd/gpt.c b/cmd/gpt.c
> > index 0c4349f4b249..2da8df60dca3 100644
> > --- a/cmd/gpt.c
> > +++ b/cmd/gpt.c
> > @@ -633,21 +633,6 @@ static int do_disk_guid(struct blk_desc *dev_desc,
> > char * const namestr)
> > }
> > #ifdef CONFIG_CMD_GPT_RENAME
> > -/*
> > - * There are 3 malloc() calls in set_gpt_info() and there is no info about
> > which
> > - * failed.
> > - */
> > -static void set_gpt_cleanup(char **str_disk_guid,
> > - disk_partition_t **partitions)
> > -{
> > -#ifdef CONFIG_RANDOM_UUID
> > - if (str_disk_guid)
> > - free(str_disk_guid);
> > -#endif
> > - if (partitions)
> > - free(partitions);
> > -}
> > -
> > static int do_rename_gpt_parts(struct blk_desc *dev_desc, char *subcomm,
> > char *name1, char *name2)
> > {
> > @@ -699,11 +684,7 @@ static int do_rename_gpt_parts(struct blk_desc
> > *dev_desc, char *subcomm,
> > &new_partitions, &part_count);
> > if (ret < 0) {
> > del_gpt_info();
> > - free(partitions_list);
> > - if (ret == -ENOMEM)
> > - set_gpt_cleanup(&str_disk_guid, &new_partitions);
> > - else
> > - goto out;
> > + goto out;
> > }
> > if (!strcmp(subcomm, "swap")) {
> > @@ -768,11 +749,7 @@ static int do_rename_gpt_parts(struct blk_desc
> > *dev_desc, char *subcomm,
> > */
> > if (ret < 0) {
> > del_gpt_info();
> > - free(partitions_list);
> > - if (ret == -ENOMEM)
> > - set_gpt_cleanup(&str_disk_guid, &new_partitions);
> > - else
> > - goto out;
> > + goto out;
> > }
> > debug("Writing new partition table\n");
> > @@ -797,8 +774,12 @@ static int do_rename_gpt_parts(struct blk_desc
> > *dev_desc, char *subcomm,
> > print_gpt_info();
> > del_gpt_info();
> > out:
> > - free(new_partitions);
> > - free(str_disk_guid);
> > +#ifdef CONFIG_RANDOM_UUID
> > + if (str_disk_guid)
>
> Looks good overall, but could it be required to initialize str_disk_guid and
> new_partitions to 0 to make this test here work? Because set_gpt_info does
> not always seem to set these pointers in the error case, so they could be
> left uninitialized?OK, so, looking at everything with the patch applied again. We do initialize new_partitions to NULL, so that's set. That leaves checking str_disk_guid. We will set (to a real value or NULL) str_disk_guid in set_gpt_info(). That function will NOT do anything if str_part is NULL. In our case, if str_part would be NULL we'll have errored out of do_rename_gpt_parts() before that call, so we're safe. If str_part can't be duplicated because we ran out of memory then yes, we can't be sure str_disk_guid will have been set. I'll send v2 shortly. Thanks! -- Tom
signature.asc
Description: PGP signature
