2022-11-09 14:44:47

by Zheng Wang

[permalink] [raw]
Subject: [PATCH v9] misc: sgi-gru: fix use-after-free error in gru_set_context_option, gru_fault and gru_handle_user_call_os

In some bad situation, the gts may be freed gru_check_chiplet_assignment.
The call chain can be gru_unload_context->gru_free_gru_context->gts_drop
and kfree finally. However, the caller didn't know if the gts is freed
or not and use it afterwards. This will trigger a Use after Free bug.

Fix it by introducing a return value to see if it's in error path or not.
Free the gts in caller if gru_check_chiplet_assignment check failed.

Fixes: 55484c45dbec ("gru: allow users to specify gru chiplet 2")
Signed-off-by: Zheng Wang <[email protected]>
Acked-by: Dimitri Sivanich <[email protected]>
---
v9:
- rewrite changelog and add comment in the code to make it more clear

v8:
- remove tested-by tag suggested by Greg

v7:
- fix some spelling problems suggested by Greg, change kernel test robot from reported-by tag to tested-by tag

v6:
- remove unused var checked by kernel test robot

v5:
- fix logical issue and remove unnecessary variable suggested by Dimitri Sivanich

v4:
- use VM_FAULT_NOPAGE as failure code in gru_fault and -EINVAL in other functions suggested by Yejian

v3:
- add preempt_enable and use VM_FAULT_NOPAGE as failure code suggested by Yejian

v2:
- commit message changes suggested by Greg

v1: https://lore.kernel.org/lkml/CAJedcCzY72jqgF-pCPtx66vXXwdPn-KMagZnqrxcpWw1NxTLaA@mail.gmail.com/
---
drivers/misc/sgi-gru/grufault.c | 14 ++++++++++++--
drivers/misc/sgi-gru/grumain.c | 22 ++++++++++++++++++----
drivers/misc/sgi-gru/grutables.h | 2 +-
3 files changed, 31 insertions(+), 7 deletions(-)

diff --git a/drivers/misc/sgi-gru/grufault.c b/drivers/misc/sgi-gru/grufault.c
index d7ef61e602ed..bdd515d33225 100644
--- a/drivers/misc/sgi-gru/grufault.c
+++ b/drivers/misc/sgi-gru/grufault.c
@@ -656,7 +656,9 @@ int gru_handle_user_call_os(unsigned long cb)
if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE)
goto exit;

- gru_check_context_placement(gts);
+ ret = gru_check_context_placement(gts);
+ if (ret)
+ goto err;

/*
* CCH may contain stale data if ts_force_cch_reload is set.
@@ -677,6 +679,10 @@ int gru_handle_user_call_os(unsigned long cb)
exit:
gru_unlock_gts(gts);
return ret;
+err:
+ gru_unlock_gts(gts);
+ gru_unload_context(gts, 1);
+ return -EINVAL;
}

/*
@@ -874,7 +880,11 @@ int gru_set_context_option(unsigned long arg)
} else {
gts->ts_user_blade_id = req.val1;
gts->ts_user_chiplet_id = req.val0;
- gru_check_context_placement(gts);
+ if (gru_check_context_placement(gts)) {
+ gru_unlock_gts(gts);
+ gru_unload_context(gts, 1);
+ return -EINVAL;
+ }
}
break;
case sco_gseg_owner:
diff --git a/drivers/misc/sgi-gru/grumain.c b/drivers/misc/sgi-gru/grumain.c
index 6706ef3c5977..816def778078 100644
--- a/drivers/misc/sgi-gru/grumain.c
+++ b/drivers/misc/sgi-gru/grumain.c
@@ -716,9 +716,10 @@ static int gru_check_chiplet_assignment(struct gru_state *gru,
* chiplet. Misassignment can occur if the process migrates to a different
* blade or if the user changes the selected blade/chiplet.
*/
-void gru_check_context_placement(struct gru_thread_state *gts)
+int gru_check_context_placement(struct gru_thread_state *gts)
{
struct gru_state *gru;
+ int ret = 0;

/*
* If the current task is the context owner, verify that the
@@ -726,15 +727,23 @@ void gru_check_context_placement(struct gru_thread_state *gts)
* references. Pthread apps use non-owner references to the CBRs.
*/
gru = gts->ts_gru;
+ /*
+ * If gru or gts->ts_tgid_owner isn't initialized properly, return
+ * success is fine, for it's not a deadly error. The related variable
+ * can be reconfigure in other function.The caller is responsible
+ * for their inspection, and reinitialization if needed.
+ */
if (!gru || gts->ts_tgid_owner != current->tgid)
- return;
+ return ret;

if (!gru_check_chiplet_assignment(gru, gts)) {
STAT(check_context_unload);
- gru_unload_context(gts, 1);
+ ret = -EINVAL;
} else if (gru_retarget_intr(gts)) {
STAT(check_context_retarget_intr);
}
+
+ return ret;
}


@@ -934,7 +943,12 @@ vm_fault_t gru_fault(struct vm_fault *vmf)
mutex_lock(&gts->ts_ctxlock);
preempt_disable();

- gru_check_context_placement(gts);
+ if (gru_check_context_placement(gts)) {
+ preempt_enable();
+ mutex_unlock(&gts->ts_ctxlock);
+ gru_unload_context(gts, 1);
+ return VM_FAULT_NOPAGE;
+ }

if (!gts->ts_gru) {
STAT(load_user_context);
diff --git a/drivers/misc/sgi-gru/grutables.h b/drivers/misc/sgi-gru/grutables.h
index 8c52776db234..640daf1994df 100644
--- a/drivers/misc/sgi-gru/grutables.h
+++ b/drivers/misc/sgi-gru/grutables.h
@@ -632,7 +632,7 @@ extern int gru_user_flush_tlb(unsigned long arg);
extern int gru_user_unload_context(unsigned long arg);
extern int gru_get_exception_detail(unsigned long arg);
extern int gru_set_context_option(unsigned long address);
-extern void gru_check_context_placement(struct gru_thread_state *gts);
+extern int gru_check_context_placement(struct gru_thread_state *gts);
extern int gru_cpu_fault_map_id(void);
extern struct vm_area_struct *gru_find_vma(unsigned long vaddr);
extern void gru_flush_all_tlb(struct gru_state *gru);
--
2.25.1



2022-11-09 23:08:16

by Dimitri Sivanich

[permalink] [raw]
Subject: Re: [PATCH v9] misc: sgi-gru: fix use-after-free error in gru_set_context_option, gru_fault and gru_handle_user_call_os

On Wed, Nov 09, 2022 at 10:17:01PM +0800, Zheng Wang wrote:
> In some bad situation, the gts may be freed gru_check_chiplet_assignment.
> The call chain can be gru_unload_context->gru_free_gru_context->gts_drop
> and kfree finally. However, the caller didn't know if the gts is freed
> or not and use it afterwards. This will trigger a Use after Free bug.
>
> Fix it by introducing a return value to see if it's in error path or not.
> Free the gts in caller if gru_check_chiplet_assignment check failed.
>
> Fixes: 55484c45dbec ("gru: allow users to specify gru chiplet 2")
> Signed-off-by: Zheng Wang <[email protected]>
> Acked-by: Dimitri Sivanich <[email protected]>
> ---
> v9:
> - rewrite changelog and add comment in the code to make it more clear
>
> v8:
> - remove tested-by tag suggested by Greg
>
> v7:
> - fix some spelling problems suggested by Greg, change kernel test robot from reported-by tag to tested-by tag
>
> v6:
> - remove unused var checked by kernel test robot
>
> v5:
> - fix logical issue and remove unnecessary variable suggested by Dimitri Sivanich
>
> v4:
> - use VM_FAULT_NOPAGE as failure code in gru_fault and -EINVAL in other functions suggested by Yejian
>
> v3:
> - add preempt_enable and use VM_FAULT_NOPAGE as failure code suggested by Yejian
>
> v2:
> - commit message changes suggested by Greg
>
> v1: https://lore.kernel.org/lkml/CAJedcCzY72jqgF-pCPtx66vXXwdPn-KMagZnqrxcpWw1NxTLaA@mail.gmail.com/
> ---
> drivers/misc/sgi-gru/grufault.c | 14 ++++++++++++--
> drivers/misc/sgi-gru/grumain.c | 22 ++++++++++++++++++----
> drivers/misc/sgi-gru/grutables.h | 2 +-
> 3 files changed, 31 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/misc/sgi-gru/grufault.c b/drivers/misc/sgi-gru/grufault.c
> index d7ef61e602ed..bdd515d33225 100644
> --- a/drivers/misc/sgi-gru/grufault.c
> +++ b/drivers/misc/sgi-gru/grufault.c
> @@ -656,7 +656,9 @@ int gru_handle_user_call_os(unsigned long cb)
> if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE)
> goto exit;
>
> - gru_check_context_placement(gts);
> + ret = gru_check_context_placement(gts);

Below we do not want to skip over the rest of the logic in either return case.
You will have to do the gru_find_lock_gts again after unloading the gru context,
then check the gts value, then return EINVAL if not set (same as earlier in the
function).

> + if (ret)
> + goto err;
>
> /*
> * CCH may contain stale data if ts_force_cch_reload is set.
> @@ -677,6 +679,10 @@ int gru_handle_user_call_os(unsigned long cb)
> exit:
> gru_unlock_gts(gts);
> return ret;
> +err:
> + gru_unlock_gts(gts);
> + gru_unload_context(gts, 1);
> + return -EINVAL;
> }
>
> /*
> @@ -874,7 +880,11 @@ int gru_set_context_option(unsigned long arg)
> } else {
> gts->ts_user_blade_id = req.val1;
> gts->ts_user_chiplet_id = req.val0;
> - gru_check_context_placement(gts);
> + if (gru_check_context_placement(gts)) {
> + gru_unlock_gts(gts);
> + gru_unload_context(gts, 1);

Looking at this again, I think we should return 0, as we originally would
have done in this case anyway.

> + return -EINVAL;
> + }
> }
> break;
> case sco_gseg_owner:
> diff --git a/drivers/misc/sgi-gru/grumain.c b/drivers/misc/sgi-gru/grumain.c
> index 6706ef3c5977..816def778078 100644
> --- a/drivers/misc/sgi-gru/grumain.c
> +++ b/drivers/misc/sgi-gru/grumain.c
> @@ -716,9 +716,10 @@ static int gru_check_chiplet_assignment(struct gru_state *gru,
> * chiplet. Misassignment can occur if the process migrates to a different
> * blade or if the user changes the selected blade/chiplet.
> */
> -void gru_check_context_placement(struct gru_thread_state *gts)
> +int gru_check_context_placement(struct gru_thread_state *gts)
> {
> struct gru_state *gru;
> + int ret = 0;
>
> /*
> * If the current task is the context owner, verify that the
> @@ -726,15 +727,23 @@ void gru_check_context_placement(struct gru_thread_state *gts)
> * references. Pthread apps use non-owner references to the CBRs.
> */
> gru = gts->ts_gru;
> + /*
> + * If gru or gts->ts_tgid_owner isn't initialized properly, return
> + * success is fine, for it's not a deadly error. The related variable
> + * can be reconfigure in other function.The caller is responsible
> + * for their inspection, and reinitialization if needed.
> + */

How about this instead?
"If gru or gts->ts_tgid_owner isn't initialized properly, return
success to indicate that the caller does not need to unload the gru
context. The caller is responsible for their inspection and
reinitialization if needed."

> if (!gru || gts->ts_tgid_owner != current->tgid)
> - return;
> + return ret;
>
> if (!gru_check_chiplet_assignment(gru, gts)) {
> STAT(check_context_unload);
> - gru_unload_context(gts, 1);
> + ret = -EINVAL;
> } else if (gru_retarget_intr(gts)) {
> STAT(check_context_retarget_intr);
> }
> +
> + return ret;
> }
>
>
> @@ -934,7 +943,12 @@ vm_fault_t gru_fault(struct vm_fault *vmf)
> mutex_lock(&gts->ts_ctxlock);
> preempt_disable();
>
> - gru_check_context_placement(gts);
> + if (gru_check_context_placement(gts)) {
> + preempt_enable();
> + mutex_unlock(&gts->ts_ctxlock);
> + gru_unload_context(gts, 1);
> + return VM_FAULT_NOPAGE;
> + }
>
> if (!gts->ts_gru) {
> STAT(load_user_context);
> diff --git a/drivers/misc/sgi-gru/grutables.h b/drivers/misc/sgi-gru/grutables.h
> index 8c52776db234..640daf1994df 100644
> --- a/drivers/misc/sgi-gru/grutables.h
> +++ b/drivers/misc/sgi-gru/grutables.h
> @@ -632,7 +632,7 @@ extern int gru_user_flush_tlb(unsigned long arg);
> extern int gru_user_unload_context(unsigned long arg);
> extern int gru_get_exception_detail(unsigned long arg);
> extern int gru_set_context_option(unsigned long address);
> -extern void gru_check_context_placement(struct gru_thread_state *gts);
> +extern int gru_check_context_placement(struct gru_thread_state *gts);
> extern int gru_cpu_fault_map_id(void);
> extern struct vm_area_struct *gru_find_vma(unsigned long vaddr);
> extern void gru_flush_all_tlb(struct gru_state *gru);
> --
> 2.25.1

2022-11-10 04:09:51

by Zheng Hacker

[permalink] [raw]
Subject: Re: [PATCH v9] misc: sgi-gru: fix use-after-free error in gru_set_context_option, gru_fault and gru_handle_user_call_os

Dimitri Sivanich <[email protected]> 于2022年11月10日周四 06:33写道:
>

> > diff --git a/drivers/misc/sgi-gru/grufault.c b/drivers/misc/sgi-gru/grufault.c
> > index d7ef61e602ed..bdd515d33225 100644
> > --- a/drivers/misc/sgi-gru/grufault.c
> > +++ b/drivers/misc/sgi-gru/grufault.c
> > @@ -656,7 +656,9 @@ int gru_handle_user_call_os(unsigned long cb)
> > if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE)
> > goto exit;
> >
> > - gru_check_context_placement(gts);
> > + ret = gru_check_context_placement(gts);
>
> Below we do not want to skip over the rest of the logic in either return case.
> You will have to do the gru_find_lock_gts again after unloading the gru context,
> then check the gts value, then return EINVAL if not set (same as earlier in the
> function).

Hello my friend! Thanks for your detailed review. I think you are right.
I'll see what I can do and make a revision.

>
> > + if (ret)
> > + goto err;
> >
> > /*
> > * CCH may contain stale data if ts_force_cch_reload is set.
> > @@ -677,6 +679,10 @@ int gru_handle_user_call_os(unsigned long cb)
> > exit:
> > gru_unlock_gts(gts);
> > return ret;
> > +err:
> > + gru_unlock_gts(gts);
> > + gru_unload_context(gts, 1);
> > + return -EINVAL;
> > }
> >
> > /*
> > @@ -874,7 +880,11 @@ int gru_set_context_option(unsigned long arg)
> > } else {
> > gts->ts_user_blade_id = req.val1;
> > gts->ts_user_chiplet_id = req.val0;
> > - gru_check_context_placement(gts);
> > + if (gru_check_context_placement(gts)) {
> > + gru_unlock_gts(gts);
> > + gru_unload_context(gts, 1);
>
> Looking at this again, I think we should return 0, as we originally would
> have done in this case anyway.

I think here we have finished the handling of the gts when it's illegal.
So return 0 is fine for me.

>
> > + return -EINVAL;
> > + }
> > }
> > break;
> > case sco_gseg_owner:

> > * If the current task is the context owner, verify that the
> > @@ -726,15 +727,23 @@ void gru_check_context_placement(struct gru_thread_state *gts)
> > * references. Pthread apps use non-owner references to the CBRs.
> > */
> > gru = gts->ts_gru;
> > + /*
> > + * If gru or gts->ts_tgid_owner isn't initialized properly, return
> > + * success is fine, for it's not a deadly error. The related variable
> > + * can be reconfigure in other function.The caller is responsible
> > + * for their inspection, and reinitialization if needed.
> > + */
>
> How about this instead?
> "If gru or gts->ts_tgid_owner isn't initialized properly, return
> success to indicate that the caller does not need to unload the gru
> context. The caller is responsible for their inspection and
> reinitialization if needed."
>

Yes, sounds much more clear for the developer, thanks for your suggestion :)

Best regards,
Zheng Wang