NVMe: Call put_nvmeq() before calling nvme_submit_sync_cmd()
We can't have preemption disabled when we call schedule(). Accept the
possibility that we'll get preempted, and it'll cost us some cacheline
bounces.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
diff --git a/drivers/block/nvme.c b/drivers/block/nvme.c
index 4bfed59..1c3cd6c 100644
--- a/drivers/block/nvme.c
+++ b/drivers/block/nvme.c
@@ -842,8 +842,13 @@
nvme_setup_prps(&c.common, sg, length);
nvmeq = get_nvmeq(ns);
- status = nvme_submit_sync_cmd(nvmeq, &c, &result);
+ /* Since nvme_submit_sync_cmd sleeps, we can't keep preemption
+ * disabled. We may be preempted at any point, and be rescheduled
+ * to a different CPU. That will cause cacheline bouncing, but no
+ * additional races since q_lock already protects against other CPUs.
+ */
put_nvmeq(nvmeq);
+ status = nvme_submit_sync_cmd(nvmeq, &c, &result);
nvme_unmap_user_pages(dev, io.opcode & 1, io.addr, length, sg, nents);
put_user(result, &uio->result);