osd/PrimaryLogPG: fix the extent length error of the sync read

The extent.length of the sync read is changed to 0, when
the read is failed. This makes the retry of the read with
a offset of 0 to read the total object.
Let the extent.length as it is, if the return code is -EAGAIN,
which will make the read try again later with the correct
extent.length.

Fixes: https://tracker.ceph.com/issues/37680
Signed-off-by: Xiaofei Cui <cuixiaofei@sangfor.com.cn>
(cherry picked from commit fd5d2d6d6eb39151d6d96105e83bfd39749d79d5)

diff --git a/src/osd/PrimaryLogPG.cc b/src/osd/PrimaryLogPG.cc
index 5d0805cc05bcfbae32f7eebb2e3d449cc9e156c8..0a4465c997eecbe793f239db5f0641e859d716b0 100644 (file)
--- a/src/osd/PrimaryLogPG.cc
+++ b/src/osd/PrimaryLogPG.cc
@@ -5470,7 +5470,12 @@ int PrimaryLogPG::do_read(OpContext *ctx, OSDOp& osd_op) {
     }
     if (r >= 0)
       op.extent.length = r;
-    else {
+    else if (r == -EAGAIN) {
+      // EAGAIN should not change the length of extent or count the read op.
+      dout(10) << " read got " << r << " / " << op.extent.length
+              << " bytes from obj " << soid << ". try again." << dendl;
+      return -EAGAIN;
+    } else {
       result = r;
       op.extent.length = 0;
     }