diff --git a/src/borg/chunker.pyx b/src/borg/chunker.pyx
index e0caafbd27..5564d5dbe2 100644
--- a/src/borg/chunker.pyx
+++ b/src/borg/chunker.pyx
@@ -60,7 +60,12 @@ def dread(offset, size, fd=None, fh=-1):
             # UNIX only and, in case of block sizes that are not a multiple of the
             # system's page size, better be used with a bug fixed linux kernel > 4.6.0,
             # see comment/workaround in _chunker.c and borgbackup issue #907.
-            os.posix_fadvise(fh, offset, len(data), os.POSIX_FADV_DONTNEED)
+            try:
+                os.posix_fadvise(fh, offset, len(data), os.POSIX_FADV_DONTNEED)
+            except OSError:
+                # e.g. ESPIPE when reading from a pipe / FIFO. this is just an optimization of the
+                # cache usage, not critical for correct function, so we ignore these errors.
+                pass
         return data
     else:
         return fd.read(size)
diff --git a/src/borg/testsuite/chunker_pytest.py b/src/borg/testsuite/chunker_pytest.py
index 5e48fadf75..70bf027656 100644
--- a/src/borg/testsuite/chunker_pytest.py
+++ b/src/borg/testsuite/chunker_pytest.py
@@ -160,6 +160,33 @@ def test_chunkify_short_reads(header_size):
     assert [len(c) for c in chunks] == expected_sizes
 
 
+def test_chunkify_pipe():
+    # reading from a pipe via the OS-level file handle: reads come up short and posix_fadvise()
+    # fails with ESPIPE (e.g. on Linux) - neither must make the chunker stop or fail.
+    import threading
+    import time
+    data = os.urandom(3 * 65536 + 123)
+    r, w = os.pipe()
+
+    def writer():
+        with os.fdopen(w, 'wb') as f:
+            for start in range(0, len(data), 65536):
+                f.write(data[start:start + 65536])
+                f.flush()
+                time.sleep(0.05)  # writer pause: reads on the other side come up short
+
+    t = threading.Thread(target=writer)
+    t.start()
+    try:
+        chunker = ChunkerFixed(131072, sparse=False)
+        chunks = cf(chunker.chunkify(None, r))
+    finally:
+        os.close(r)
+        t.join()
+    assert b''.join(chunks) == data
+    assert [len(c) for c in chunks] == [131072, 65536 + 123]
+
+
 def test_buzhash_chunksize_distribution():
     data = os.urandom(1048576)
     min_exp, max_exp, mask = 10, 16, 14  # chunk size target 16kiB, clip at 1kiB and 64kiB