On Tue, Nov 19, 2024 at 7:37 PM Krzysztof Kozlowski krzysztof.kozlowski@linaro.org wrote:

Commit 2e4955167ec5 ("firmware: qcom: scm: Fix __scm and waitq completion variable initialization") introduced a write barrier in probe function to store global '__scm' variable. It also claimed that it added a read barrier, because as we all known barriers are paired (see memory-barriers.txt: "Note that write barriers should normally be paired with read or address-dependency barriers"), however it did not really add it.

The offending commit used READ_ONCE() to access '__scm' global which is not a barrier.

The barrier is needed so the store to '__scm' will be properly visible. This is most likely not fatal in current driver design, because missing read barrier would mean qcom_scm_is_available() callers will access old value, NULL. Driver does not support unbinding and does not correctly handle probe failures, thus there is no risk of stale or old pointer in '__scm' variable.

However for code correctness, readability and to be sure that we did not mess up something in this tricky topic of SMP barriers, add a read barrier for accessing '__scm'. Change also comment from useless/obvious what does barrier do, to what is expected: which other parts of the code are involved here.

Fixes: 2e4955167ec5 ("firmware: qcom: scm: Fix __scm and waitq completion variable initialization") Cc: stable@vger.kernel.org Signed-off-by: Krzysztof Kozlowski krzysztof.kozlowski@linaro.org

---

drivers/firmware/qcom/qcom_scm.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/firmware/qcom/qcom_scm.c b/drivers/firmware/qcom/qcom_scm.c index 72bf87ddcd969834609cda2aa915b67505e93943..246d672e8f7f0e2a326a03a5af40cd434a665e67 100644 --- a/drivers/firmware/qcom/qcom_scm.c +++ b/drivers/firmware/qcom/qcom_scm.c @@ -1867,7 +1867,8 @@ static int qcom_scm_qseecom_init(struct qcom_scm *scm) */ bool qcom_scm_is_available(void) {

•   return !!READ_ONCE(__scm);
  

•   /* Paired with smp_store_release() in qcom_scm_probe */
  

•   return !!smp_load_acquire(&__scm);
  

} EXPORT_SYMBOL_GPL(qcom_scm_is_available);

@@ -2024,7 +2025,7 @@ static int qcom_scm_probe(struct platform_device *pdev) if (ret) return ret;

•   /* Let all above stores be available after this */
  

•   /* Paired with smp_load_acquire() in qcom_scm_is_available(). */
    smp_store_release(&__scm, scm);
  
    irq = platform_get_irq_optional(pdev, 0);
  
  

-- 2.43.0

I'm not an expert on barriers and SMP but the explanation sounds correct to me.

Reviewed-by: Bartosz Golaszewski bartosz.golaszewski@linaro.org

On 19/11/2024 19:33, Krzysztof Kozlowski wrote:

Commit 2e4955167ec5 ("firmware: qcom: scm: Fix __scm and waitq completion variable initialization") introduced a write barrier in probe function to store global '__scm' variable. We all known barriers are paired (see memory-barriers.txt: "Note that write barriers should normally be paired with read or address-dependency barriers"), therefore accessing it from concurrent contexts requires read barrier. Previous commit added such barrier in qcom_scm_is_available(), so let's use that directly.

Lack of this read barrier can result in fetching stale '__scm' variable value, NULL, and dereferencing it.

Fixes: ca61d6836e6f ("firmware: qcom: scm: fix a NULL-pointer dereference") Fixes: 449d0d84bcd8 ("firmware: qcom: scm: smc: switch to using the SCM allocator") Cc: stable@vger.kernel.org Signed-off-by: Krzysztof Kozlowski krzysztof.kozlowski@linaro.org

---

drivers/firmware/qcom/qcom_scm.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/firmware/qcom/qcom_scm.c b/drivers/firmware/qcom/qcom_scm.c index 246d672e8f7f0e2a326a03a5af40cd434a665e67..5d91b8e22844608f35432f1ba9c08d477d4ff762 100644 --- a/drivers/firmware/qcom/qcom_scm.c +++ b/drivers/firmware/qcom/qcom_scm.c @@ -217,7 +217,10 @@ static DEFINE_SPINLOCK(scm_query_lock); struct qcom_tzmem_pool *qcom_scm_get_tzmem_pool(void) {

• return __scm ? __scm->mempool : NULL;

• if (!qcom_scm_is_available())

• return NULL;
  

• return __scm->mempool;

I mentioned in commit msg that previous commit adds barrier in qcom_scm_get_tzmem_pool(), so to be clear: This depends on previous commit, because that barrier in qcom_scm_is_available() solves the control dependency here, assuming the minimal guarantee #1 ("On any given CPU, dependent memory accesses will be issued in order, with respect to itself.")

If this is inlined by compiler it will be:

scm = READ_ONCE(__scm); barrier() if (scm) return scm->mempool; else return NULL;

Which should be even safer than standard guarantee above (according to my understanding).

Best regards, Krzysztof