This bug occurs when:
- a new request arrives, one thread(let's call it A) is pending in optee_supp_req() with req->busy is initial value false.
- tee-supplicant is killed, then optee_supp_release() is called, this function calls list_del(&req->link), and set supp->ctx to NULL. And it also wake up process A.
- process A continues, it firstly checks supp->ctx which is NULL, then checks req->busy which is false, at last run list_del(&req->link). This triggers double list_del() and results kernel panic.
So let's set req->busy to true if optee_supp_release() has already called list_del(&req->link).
Signed-off-by: Zhizhou Zhang zhizhouzhang@asrmicro.com --- drivers/tee/optee/supp.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/tee/optee/supp.c b/drivers/tee/optee/supp.c index df35fc01fd3e..c8ac6636520a 100644 --- a/drivers/tee/optee/supp.c +++ b/drivers/tee/optee/supp.c @@ -62,6 +62,7 @@ void optee_supp_release(struct optee_supp *supp)
/* Abort all queued requests */ list_for_each_entry_safe(req, req_tmp, &supp->reqs, link) { + req->busy = true; list_del(&req->link); req->ret = TEEC_ERROR_COMMUNICATION; complete(&req->c);
Hi Zhizhou,
On Mon, Nov 19, 2018 at 02:41:39PM +0800, Zhizhou Zhang wrote:
This bug occurs when:
a new request arrives, one thread(let's call it A) is pending in optee_supp_req() with req->busy is initial value false.
tee-supplicant is killed, then optee_supp_release() is called, this function calls list_del(&req->link), and set supp->ctx to NULL. And it also wake up process A.
process A continues, it firstly checks supp->ctx which is NULL, then checks req->busy which is false, at last run list_del(&req->link). This triggers double list_del() and results kernel panic.
So let's set req->busy to true if optee_supp_release() has already called list_del(&req->link).
Signed-off-by: Zhizhou Zhang zhizhouzhang@asrmicro.com
drivers/tee/optee/supp.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/tee/optee/supp.c b/drivers/tee/optee/supp.c index df35fc01fd3e..c8ac6636520a 100644 --- a/drivers/tee/optee/supp.c +++ b/drivers/tee/optee/supp.c @@ -62,6 +62,7 @@ void optee_supp_release(struct optee_supp *supp) /* Abort all queued requests */ list_for_each_entry_safe(req, req_tmp, &supp->reqs, link) {
req->busy = true;-- 2.17.1
This seems to work, but is a bit confusing. It turns out the busy flag will only used to tell if the request is in the queue (and may need to be dequeued) or not.
How about renaming the flag to "in_queue" and update the assignments and tests appropriately to only indicate if it's in the queue or not?
That should work as well and be more clear on what's going on, or am I missing something?
Thanks, Jens
-----Original Message----- From: Jens Wiklander [mailto:jens.wiklander@linaro.org] Sent: Monday, November 19, 2018 6:09 PM To: Zhang Zhizhou(张治洲) zhizhouzhang@asrmicro.com Cc: tee-dev@lists.linaro.org; linux-arm-kernel@lists.infradead.org Subject: Re: [PATCH] tee: optee: avoid possible double list_del()
Hi Zhizhou,
On Mon, Nov 19, 2018 at 02:41:39PM +0800, Zhizhou Zhang wrote:
This bug occurs when:
a new request arrives, one thread(let's call it A) is pending in optee_supp_req() with req->busy is initial value false.
tee-supplicant is killed, then optee_supp_release() is called, this function calls list_del(&req->link), and set supp->ctx to NULL. And it also wake up process A.
process A continues, it firstly checks supp->ctx which is NULL, then checks req->busy which is false, at last run list_del(&req->link). This triggers double list_del() and results kernel panic.
So let's set req->busy to true if optee_supp_release() has already called list_del(&req->link).
Signed-off-by: Zhizhou Zhang zhizhouzhang@asrmicro.com
drivers/tee/optee/supp.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/tee/optee/supp.c b/drivers/tee/optee/supp.c index df35fc01fd3e..c8ac6636520a 100644 --- a/drivers/tee/optee/supp.c +++ b/drivers/tee/optee/supp.c @@ -62,6 +62,7 @@ void optee_supp_release(struct optee_supp *supp)
/* Abort all queued requests */ list_for_each_entry_safe(req, req_tmp, &supp->reqs, link) {
req->busy = true;-- 2.17.1
This seems to work, but is a bit confusing. It turns out the busy flag will only used to tell if the request is in the queue (and may need to be dequeued) or not.
How about renaming the flag to "in_queue" and update the assignments and tests appropriately to only indicate if it's in the queue or not?
You're right. I will send you a V2 patch later. Thanks!
That should work as well and be more clear on what's going on, or am I missing something?
When I was making this patch, I also had the same concern. This patch make busy doesn't mean busy any more. But I'm afraid of changing too much, so I sent you a only one line changed patch. ;-)
Thanks, Jens
-
Jens Wiklander -
Zhang Zhizhou(张治洲) -
Zhizhou Zhang