19.03.2021 20:41, Dmitry Osipenko пишет:

19.03.2021 20:29, Dmitry Osipenko пишет:

quoted

+void cma_sysfs_alloc_pages_count(struct cma *cma, size_t count)
+{
+	atomic64_add(count, &cma->nr_pages_succeeded);
+}
+
+void cma_sysfs_fail_pages_count(struct cma *cma, size_t count)
+{
+	atomic64_add(count, &cma->nr_pages_failed);
+}

The atomic looks good, but aren't CMA allocations already protected by
the CMA core? Do we really need to worry about racing here?

Although, please scratch that. I see now that these functions are called
outside of the lock.