| Message ID | 20241119-qcom-scm-missing-barriers-and-all-sort-of-srap-v1-1-7056127007a7@linaro.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | firmware: qcom: scm: Fixes for concurrency | expand |
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>
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);
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(-)