[BUGFIX PATCH] crypto: ccree: fix resume race condition on init

[BUGFIX PATCH] crypto: ccree: fix resume race condition on init

[Gilad Ben-Yossef]

We were enabling autosuspend, which is using data set by the
hash module, prior to the hash module being inited, casuing
a crash on resume as part of the startup sequence if the race
was lost.

This was never a real problem because the PM infra was using low
res timers so we were always winning the race, until commit 8234f6734c5d
("PM-runtime: Switch autosuspend over to using hrtimers") changed that :-)

Fix this by seperating the PM setup and enablement and doing the
latter only at the end of the init sequence.

Signed-off-by: Gilad Ben-Yossef <gilad@benyossef.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: stable@kernel.org # v4.20
---
Herbert, could you please take this for 5.0-rc6 ? thanks.

 drivers/crypto/ccree/cc_driver.c |  7 ++++---
 drivers/crypto/ccree/cc_pm.c     | 13 ++++++-------
 drivers/crypto/ccree/cc_pm.h     |  3 +++
 3 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/drivers/crypto/ccree/cc_driver.c b/drivers/crypto/ccree/cc_driver.c
index 210cc86605e9..3bcc6c76e090 100644
--- a/drivers/crypto/ccree/cc_driver.c
+++ b/drivers/crypto/ccree/cc_driver.c
@@ -380,7 +380,7 @@ static int init_cc_resources(struct platform_device *plat_dev)
 	rc = cc_ivgen_init(new_drvdata);
 	if (rc) {
 		dev_err(dev, "cc_ivgen_init failed\n");
-		goto post_power_mgr_err;
+		goto post_buf_mgr_err;
 	}
 
 	/* Allocate crypto algs */
@@ -403,6 +403,9 @@ static int init_cc_resources(struct platform_device *plat_dev)
 		goto post_hash_err;
 	}
 
+	/* All set, we can allow autosuspend */
+	cc_pm_go(new_drvdata);
+
 	/* If we got here and FIPS mode is enabled
 	 * it means all FIPS test passed, so let TEE
 	 * know we're good.
@@ -417,8 +420,6 @@ static int init_cc_resources(struct platform_device *plat_dev)
 	cc_cipher_free(new_drvdata);
 post_ivgen_err:
 	cc_ivgen_fini(new_drvdata);
-post_power_mgr_err:
-	cc_pm_fini(new_drvdata);
 post_buf_mgr_err:
 	 cc_buffer_mgr_fini(new_drvdata);
 post_req_mgr_err:
diff --git a/drivers/crypto/ccree/cc_pm.c b/drivers/crypto/ccree/cc_pm.c
index d990f472e89f..6ff7e75ad90e 100644
--- a/drivers/crypto/ccree/cc_pm.c
+++ b/drivers/crypto/ccree/cc_pm.c
@@ -100,20 +100,19 @@ int cc_pm_put_suspend(struct device *dev)
 
 int cc_pm_init(struct cc_drvdata *drvdata)
 {
-	int rc = 0;
 	struct device *dev = drvdata_to_dev(drvdata);
 
 	/* must be before the enabling to avoid resdundent suspending */
 	pm_runtime_set_autosuspend_delay(dev, CC_SUSPEND_TIMEOUT);
 	pm_runtime_use_autosuspend(dev);
 	/* activate the PM module */
-	rc = pm_runtime_set_active(dev);
-	if (rc)
-		return rc;
-	/* enable the PM module*/
-	pm_runtime_enable(dev);
+	return pm_runtime_set_active(dev);
+}
 
-	return rc;
+/* enable the PM module*/
+void cc_pm_go(struct cc_drvdata *drvdata)
+{
+	pm_runtime_enable(drvdata_to_dev(drvdata));
 }
 
 void cc_pm_fini(struct cc_drvdata *drvdata)
diff --git a/drivers/crypto/ccree/cc_pm.h b/drivers/crypto/ccree/cc_pm.h
index 020a5403c58b..f62624357020 100644
--- a/drivers/crypto/ccree/cc_pm.h
+++ b/drivers/crypto/ccree/cc_pm.h
@@ -16,6 +16,7 @@
 extern const struct dev_pm_ops ccree_pm;
 
 int cc_pm_init(struct cc_drvdata *drvdata);
+void cc_pm_go(struct cc_drvdata *drvdata);
 void cc_pm_fini(struct cc_drvdata *drvdata);
 int cc_pm_suspend(struct device *dev);
 int cc_pm_resume(struct device *dev);
@@ -29,6 +30,8 @@ static inline int cc_pm_init(struct cc_drvdata *drvdata)
 	return 0;
 }
 
+static void cc_pm_go(struct cc_drvdata *drvdata) {}
+
 static inline void cc_pm_fini(struct cc_drvdata *drvdata) {}
 
 static inline int cc_pm_suspend(struct device *dev)
-- 
2.20.1

Re: [BUGFIX PATCH] crypto: ccree: fix resume race condition on init

[Herbert Xu]

On Thu, Feb 07, 2019 at 03:36:11PM +0200, Gilad Ben-Yossef wrote:
> We were enabling autosuspend, which is using data set by the
> hash module, prior to the hash module being inited, casuing
> a crash on resume as part of the startup sequence if the race
> was lost.
> 
> This was never a real problem because the PM infra was using low
> res timers so we were always winning the race, until commit 8234f6734c5d
> ("PM-runtime: Switch autosuspend over to using hrtimers") changed that :-)
> 
> Fix this by seperating the PM setup and enablement and doing the
> latter only at the end of the init sequence.
> 
> Signed-off-by: Gilad Ben-Yossef <gilad@benyossef.com>
> Cc: Vincent Guittot <vincent.guittot@linaro.org>
> Cc: stable@kernel.org # v4.20
> ---
> Herbert, could you please take this for 5.0-rc6 ? thanks.
> 
>  drivers/crypto/ccree/cc_driver.c |  7 ++++---
>  drivers/crypto/ccree/cc_pm.c     | 13 ++++++-------
>  drivers/crypto/ccree/cc_pm.h     |  3 +++
>  3 files changed, 13 insertions(+), 10 deletions(-)

Patch applied.  Thanks.
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt