'lustre_cfg_new()' can return ERR_PTR(-ENOMEM).
Handle these errors and propagate the error code to the callers.
Error handling has been rearranged in 'lustre_process_log()' with the
addition of a label in order to free some resources.
Signed-off-by: Christophe JAILLET <[email protected]>
---
drivers/staging/lustre/lustre/obdclass/obd_mount.c | 16 ++++++++++++++--
1 file changed, 14 insertions(+), 2 deletions(-)
diff --git a/drivers/staging/lustre/lustre/obdclass/obd_mount.c b/drivers/staging/lustre/lustre/obdclass/obd_mount.c
index 59fbc29aae94..5473615cd338 100644
--- a/drivers/staging/lustre/lustre/obdclass/obd_mount.c
+++ b/drivers/staging/lustre/lustre/obdclass/obd_mount.c
@@ -89,11 +89,14 @@ int lustre_process_log(struct super_block *sb, char *logname,
lustre_cfg_bufs_set(bufs, 2, cfg, sizeof(*cfg));
lustre_cfg_bufs_set(bufs, 3, &sb, sizeof(sb));
lcfg = lustre_cfg_new(LCFG_LOG_START, bufs);
+ if (IS_ERR(lcfg)) {
+ rc = PTR_ERR(lcfg);
+ goto out_free;
+ }
+
rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
lustre_cfg_free(lcfg);
- kfree(bufs);
-
if (rc == -EINVAL)
LCONSOLE_ERROR_MSG(0x15b, "%s: The configuration from log '%s' failed from the MGS (%d). Make sure this client and the MGS are running compatible versions of Lustre.\n",
mgc->obd_name, logname, rc);
@@ -104,6 +107,9 @@ int lustre_process_log(struct super_block *sb, char *logname,
rc);
/* class_obd_list(); */
+
+out_free:
+ kfree(bufs);
return rc;
}
EXPORT_SYMBOL(lustre_process_log);
@@ -127,6 +133,9 @@ int lustre_end_log(struct super_block *sb, char *logname,
if (cfg)
lustre_cfg_bufs_set(&bufs, 2, cfg, sizeof(*cfg));
lcfg = lustre_cfg_new(LCFG_LOG_END, &bufs);
+ if (IS_ERR(lcfg))
+ return PTR_ERR(lcfg);
+
rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
lustre_cfg_free(lcfg);
return rc;
@@ -159,6 +168,9 @@ static int do_lcfg(char *cfgname, lnet_nid_t nid, int cmd,
lustre_cfg_bufs_set_string(&bufs, 4, s4);
lcfg = lustre_cfg_new(cmd, &bufs);
+ if (IS_ERR(lcfg))
+ return PTR_ERR(lcfg);
+
lcfg->lcfg_nid = nid;
rc = class_process_config(lcfg);
lustre_cfg_free(lcfg);
--
2.9.3
Hello!
On Nov 6, 2016, at 12:11 PM, Christophe JAILLET wrote:
> 'lustre_cfg_new()' can return ERR_PTR(-ENOMEM).
> Handle these errors and propagate the error code to the callers.
>
> Error handling has been rearranged in 'lustre_process_log()' with the
> addition of a label in order to free some resources.
I wonder if we should just make it return NULL on allocation failure,
and then at least the other error handling that is there (i.e. in your other patch)
would become correct.
This would make handling in mgc_apply_recover_logs incorrect, but it's already
geared towards this sort of handling anyway, as it discards the passed error
and sets ENOMEM unconditionally (just need to revert 3092c34a in a way).
Thanks!
>
> Signed-off-by: Christophe JAILLET <[email protected]>
> ---
> drivers/staging/lustre/lustre/obdclass/obd_mount.c | 16 ++++++++++++++--
> 1 file changed, 14 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/staging/lustre/lustre/obdclass/obd_mount.c b/drivers/staging/lustre/lustre/obdclass/obd_mount.c
> index 59fbc29aae94..5473615cd338 100644
> --- a/drivers/staging/lustre/lustre/obdclass/obd_mount.c
> +++ b/drivers/staging/lustre/lustre/obdclass/obd_mount.c
> @@ -89,11 +89,14 @@ int lustre_process_log(struct super_block *sb, char *logname,
> lustre_cfg_bufs_set(bufs, 2, cfg, sizeof(*cfg));
> lustre_cfg_bufs_set(bufs, 3, &sb, sizeof(sb));
> lcfg = lustre_cfg_new(LCFG_LOG_START, bufs);
> + if (IS_ERR(lcfg)) {
> + rc = PTR_ERR(lcfg);
> + goto out_free;
> + }
> +
> rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
> lustre_cfg_free(lcfg);
>
> - kfree(bufs);
> -
> if (rc == -EINVAL)
> LCONSOLE_ERROR_MSG(0x15b, "%s: The configuration from log '%s' failed from the MGS (%d). Make sure this client and the MGS are running compatible versions of Lustre.\n",
> mgc->obd_name, logname, rc);
> @@ -104,6 +107,9 @@ int lustre_process_log(struct super_block *sb, char *logname,
> rc);
>
> /* class_obd_list(); */
> +
> +out_free:
> + kfree(bufs);
> return rc;
> }
> EXPORT_SYMBOL(lustre_process_log);
> @@ -127,6 +133,9 @@ int lustre_end_log(struct super_block *sb, char *logname,
> if (cfg)
> lustre_cfg_bufs_set(&bufs, 2, cfg, sizeof(*cfg));
> lcfg = lustre_cfg_new(LCFG_LOG_END, &bufs);
> + if (IS_ERR(lcfg))
> + return PTR_ERR(lcfg);
> +
> rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
> lustre_cfg_free(lcfg);
> return rc;
> @@ -159,6 +168,9 @@ static int do_lcfg(char *cfgname, lnet_nid_t nid, int cmd,
> lustre_cfg_bufs_set_string(&bufs, 4, s4);
>
> lcfg = lustre_cfg_new(cmd, &bufs);
> + if (IS_ERR(lcfg))
> + return PTR_ERR(lcfg);
> +
> lcfg->lcfg_nid = nid;
> rc = class_process_config(lcfg);
> lustre_cfg_free(lcfg);
> --
> 2.9.3
> On Nov 6, 2016, at 12:11 PM, Christophe JAILLET wrote:
>
> > 'lustre_cfg_new()' can return ERR_PTR(-ENOMEM).
> > Handle these errors and propagate the error code to the callers.
> >
> > Error handling has been rearranged in 'lustre_process_log()' with the
> > addition of a label in order to free some resources.
>
> I wonder if we should just make it return NULL on allocation failure,
> and then at least the other error handling that is there (i.e. in your other patch)
> would become correct.
> This would make handling in mgc_apply_recover_logs incorrect, but it's already
> geared towards this sort of handling anyway, as it discards the passed error
> and sets ENOMEM unconditionally (just need to revert 3092c34a in a way).
The header lustre_cfg.h is meant to be a UAPI header file. It is used for
our userland tools but with the current shape of lustre_cfg.h upstream our
tools will not build with it. So having kzalloc and kfree in this header
is incorrect. To do this right I need to update our user land tools as
well so we should hold off on these patches.
On Mon, Nov 07, 2016 at 04:10:16AM +0000, James Simmons wrote:
>
> > On Nov 6, 2016, at 12:11 PM, Christophe JAILLET wrote:
> >
> > > 'lustre_cfg_new()' can return ERR_PTR(-ENOMEM).
> > > Handle these errors and propagate the error code to the callers.
> > >
> > > Error handling has been rearranged in 'lustre_process_log()' with the
> > > addition of a label in order to free some resources.
> >
> > I wonder if we should just make it return NULL on allocation failure,
> > and then at least the other error handling that is there (i.e. in your other patch)
> > would become correct.
> > This would make handling in mgc_apply_recover_logs incorrect, but it's already
> > geared towards this sort of handling anyway, as it discards the passed error
> > and sets ENOMEM unconditionally (just need to revert 3092c34a in a way).
>
> The header lustre_cfg.h is meant to be a UAPI header file. It is used for
> our userland tools but with the current shape of lustre_cfg.h upstream our
> tools will not build with it. So having kzalloc and kfree in this header
> is incorrect. To do this right I need to update our user land tools as
> well so we should hold off on these patches.
Ok, but the code as-is today is incorrect, so that should get fixed
somehow, soon...
thanks,
greg k-h
On Nov 6, 2016, at 10:26, Drokin, Oleg <[email protected]> wrote:
>
> Hello!
>
> On Nov 6, 2016, at 12:11 PM, Christophe JAILLET wrote:
>
>> 'lustre_cfg_new()' can return ERR_PTR(-ENOMEM).
>> Handle these errors and propagate the error code to the callers.
>>
>> Error handling has been rearranged in 'lustre_process_log()' with the
>> addition of a label in order to free some resources.
>
> I wonder if we should just make it return NULL on allocation failure,
> and then at least the other error handling that is there (i.e. in your other patch)
> would become correct.
> This would make handling in mgc_apply_recover_logs incorrect, but it's already
> geared towards this sort of handling anyway, as it discards the passed error
> and sets ENOMEM unconditionally (just need to revert 3092c34a in a way).
I'd agree with Oleg that returning NULL is the preferable solution here.
There are also callers of lustre_cfg_new() in class_config_llog_handler(),
do_lcfg(), and lustre_end_log() that do not check error returns at all that
should be fixed at the same time.
Cheers, Andreas
>>
>> Signed-off-by: Christophe JAILLET <[email protected]>
>> ---
>> drivers/staging/lustre/lustre/obdclass/obd_mount.c | 16 ++++++++++++++--
>> 1 file changed, 14 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/staging/lustre/lustre/obdclass/obd_mount.c b/drivers/staging/lustre/lustre/obdclass/obd_mount.c
>> index 59fbc29aae94..5473615cd338 100644
>> --- a/drivers/staging/lustre/lustre/obdclass/obd_mount.c
>> +++ b/drivers/staging/lustre/lustre/obdclass/obd_mount.c
>> @@ -89,11 +89,14 @@ int lustre_process_log(struct super_block *sb, char *logname,
>> lustre_cfg_bufs_set(bufs, 2, cfg, sizeof(*cfg));
>> lustre_cfg_bufs_set(bufs, 3, &sb, sizeof(sb));
>> lcfg = lustre_cfg_new(LCFG_LOG_START, bufs);
>> + if (IS_ERR(lcfg)) {
>> + rc = PTR_ERR(lcfg);
>> + goto out_free;
>> + }
>> +
>> rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
>> lustre_cfg_free(lcfg);
>>
>> - kfree(bufs);
>> -
>> if (rc == -EINVAL)
>> LCONSOLE_ERROR_MSG(0x15b, "%s: The configuration from log '%s' failed from the MGS (%d). Make sure this client and the MGS are running compatible versions of Lustre.\n",
>> mgc->obd_name, logname, rc);
>> @@ -104,6 +107,9 @@ int lustre_process_log(struct super_block *sb, char *logname,
>> rc);
>>
>> /* class_obd_list(); */
>> +
>> +out_free:
>> + kfree(bufs);
>> return rc;
>> }
>> EXPORT_SYMBOL(lustre_process_log);
>> @@ -127,6 +133,9 @@ int lustre_end_log(struct super_block *sb, char *logname,
>> if (cfg)
>> lustre_cfg_bufs_set(&bufs, 2, cfg, sizeof(*cfg));
>> lcfg = lustre_cfg_new(LCFG_LOG_END, &bufs);
>> + if (IS_ERR(lcfg))
>> + return PTR_ERR(lcfg);
>> +
>> rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
>> lustre_cfg_free(lcfg);
>> return rc;
>> @@ -159,6 +168,9 @@ static int do_lcfg(char *cfgname, lnet_nid_t nid, int cmd,
>> lustre_cfg_bufs_set_string(&bufs, 4, s4);
>>
>> lcfg = lustre_cfg_new(cmd, &bufs);
>> + if (IS_ERR(lcfg))
>> + return PTR_ERR(lcfg);
>> +
>> lcfg->lcfg_nid = nid;
>> rc = class_process_config(lcfg);
>> lustre_cfg_free(lcfg);
>> --
>> 2.9.3
>
On Nov 7, 2016, at 4:33 PM, Dilger, Andreas wrote:
> On Nov 6, 2016, at 10:26, Drokin, Oleg <[email protected]> wrote:
>>
>> Hello!
>>
>> On Nov 6, 2016, at 12:11 PM, Christophe JAILLET wrote:
>>
>>> 'lustre_cfg_new()' can return ERR_PTR(-ENOMEM).
>>> Handle these errors and propagate the error code to the callers.
>>>
>>> Error handling has been rearranged in 'lustre_process_log()' with the
>>> addition of a label in order to free some resources.
>>
>> I wonder if we should just make it return NULL on allocation failure,
>> and then at least the other error handling that is there (i.e. in your other patch)
>> would become correct.
>> This would make handling in mgc_apply_recover_logs incorrect, but it's already
>> geared towards this sort of handling anyway, as it discards the passed error
>> and sets ENOMEM unconditionally (just need to revert 3092c34a in a way).
>
> I'd agree with Oleg that returning NULL is the preferable solution here.
>
> There are also callers of lustre_cfg_new() in class_config_llog_handler(),
> do_lcfg(), and lustre_end_log() that do not check error returns at all that
> should be fixed at the same time.
This patch was actually doing it.
>
> Cheers, Andreas
>
>>>
>>> Signed-off-by: Christophe JAILLET <[email protected]>
>>> ---
>>> drivers/staging/lustre/lustre/obdclass/obd_mount.c | 16 ++++++++++++++--
>>> 1 file changed, 14 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/staging/lustre/lustre/obdclass/obd_mount.c b/drivers/staging/lustre/lustre/obdclass/obd_mount.c
>>> index 59fbc29aae94..5473615cd338 100644
>>> --- a/drivers/staging/lustre/lustre/obdclass/obd_mount.c
>>> +++ b/drivers/staging/lustre/lustre/obdclass/obd_mount.c
>>> @@ -89,11 +89,14 @@ int lustre_process_log(struct super_block *sb, char *logname,
>>> lustre_cfg_bufs_set(bufs, 2, cfg, sizeof(*cfg));
>>> lustre_cfg_bufs_set(bufs, 3, &sb, sizeof(sb));
>>> lcfg = lustre_cfg_new(LCFG_LOG_START, bufs);
>>> + if (IS_ERR(lcfg)) {
>>> + rc = PTR_ERR(lcfg);
>>> + goto out_free;
>>> + }
>>> +
>>> rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
>>> lustre_cfg_free(lcfg);
>>>
>>> - kfree(bufs);
>>> -
>>> if (rc == -EINVAL)
>>> LCONSOLE_ERROR_MSG(0x15b, "%s: The configuration from log '%s' failed from the MGS (%d). Make sure this client and the MGS are running compatible versions of Lustre.\n",
>>> mgc->obd_name, logname, rc);
>>> @@ -104,6 +107,9 @@ int lustre_process_log(struct super_block *sb, char *logname,
>>> rc);
>>>
>>> /* class_obd_list(); */
>>> +
>>> +out_free:
>>> + kfree(bufs);
>>> return rc;
>>> }
>>> EXPORT_SYMBOL(lustre_process_log);
>>> @@ -127,6 +133,9 @@ int lustre_end_log(struct super_block *sb, char *logname,
>>> if (cfg)
>>> lustre_cfg_bufs_set(&bufs, 2, cfg, sizeof(*cfg));
>>> lcfg = lustre_cfg_new(LCFG_LOG_END, &bufs);
>>> + if (IS_ERR(lcfg))
>>> + return PTR_ERR(lcfg);
>>> +
>>> rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
>>> lustre_cfg_free(lcfg);
>>> return rc;
>>> @@ -159,6 +168,9 @@ static int do_lcfg(char *cfgname, lnet_nid_t nid, int cmd,
>>> lustre_cfg_bufs_set_string(&bufs, 4, s4);
>>>
>>> lcfg = lustre_cfg_new(cmd, &bufs);
>>> + if (IS_ERR(lcfg))
>>> + return PTR_ERR(lcfg);
>>> +
>>> lcfg->lcfg_nid = nid;
>>> rc = class_process_config(lcfg);
>>> lustre_cfg_free(lcfg);
>>> --
>>> 2.9.3
>>