With client encryption, a buffered write to an encrypted file can fail on the client due to a failed lock check in osc_encrypt_pagecache_blocks(). osc_brw_prep_request() then returns -EINVAL, and the whole write RPC fails. A file small enough to be written by a single RPC ends up empty (0 bytes), and rewriting it usually succeeds.

osc_brw_prep_request() checks for a lock on the folio (and takes it if not already locked). But other threads can take the folio lock for a short time while the folio is under writeback, and release it before the check in osc_encrypt_pagecache_blocks(), e.g. a page fault through mmap, memory management related tasks such as reclaim, migration and compaction in async mode, invalidate_mapping_pages()/mapping_evict_folio() (e.g. drop_caches), and the readahead/filemap_read helpers.