Ceph » CephFS

The RC for this issue is fixed by:

diff --git a/src/client/Client.cc b/src/client/Client.cc
index 9ccb39501a4..80c84a920d8 100644
--- a/src/client/Client.cc
+++ b/src/client/Client.cc
@@ -11824,12 +11824,11 @@ int Client::WriteEncMgr::read_modify_write(Context *_iofinish)
   ldout(cct, 10) << "start_block=" << start_block << " start_block_ofs=" << start_block_ofs
                 << " ofs_in_start_block=" << ofs_in_start_block << " end_block=" << end_block
                 << " end_block_ofs=" << end_block_ofs << " ofs_in_end_block=" << ofs_in_end_block
-                << dendl;
+                << " offset=" << offset
+                << " effective_size=" <<  in->effective_size() << " endoff=" << endoff << dendl;
   need_read_start = ofs_in_start_block > 0;
   need_read_end = (endoff < in->effective_size() && ofs_in_end_block < FSCRYPT_BLOCK_SIZE - 1 && start_block != end_block);
-  read_start_size = (need_read_start && need_read_end && start_block == end_block ?
-                     FSCRYPT_BLOCK_SIZE : ofs_in_start_block);
-  
+  read_start_size = FSCRYPT_BLOCK_SIZE; 
   bool need_read = need_read_start | need_read_end;
   ldout(cct, 10) << "need_read_start=" << need_read_start << "need_read_end=" << need_read_end
                  << " read_start_size=" << read_start_size << " need_read=" << need_read << dendl;

read_start_size was always ofs_in_start_block as it depends on need_read_end and have conflicting bool statements about start/end block. This caused data after the write portion to not be read. read_async would read up to the offset of where the write was occuring, write the pending data and leave the rest of fscrypt block empty. In fscrypt, the whole block needs to be read, overwrite starting at offset to len then recommit the whole fscrypt block.

There's also another case that is failing rmw:

reproducer:

import os

file = 'datafile31'
fd = os.open(file, os.O_RDONLY)
os.close(fd)

fd = os.open(file, os.O_RDWR|os.O_CREAT)

#reproducing sys calls after ffsb bench has started
fill = '\0'
size = 3192
offset = 60653568

s = ''
for i in range(0,size):
  s += fill

os.lseek(fd, offset, 0)
os.write(fd, str.encode(s))

os.close(fd)

The error we're getting

Inode.cc: In function 'int Inode::put_cap_ref(int)' thread 7f8aabfff6c0 time 2024-04-10T19:30:45.442916+0000
Inode.cc: 212: FAILED ceph_assert(cap_refs[c] > 0)

There's probably a missing buffered read that isn't happening that should.

As seen in note 2, the case of modification happening at start of block and contained within a single block is now addressed by:

diff --git a/src/client/Client.cc b/src/client/Client.cc
index 9ccb39501a4..3b45ed8453f 100644
--- a/src/client/Client.cc
+++ b/src/client/Client.cc
@@ -11825,11 +11825,10 @@ int Client::WriteEncMgr::read_modify_write(Context *_iofinish)
                 << " ofs_in_start_block=" << ofs_in_start_block << " end_block=" << end_block
                 << " end_block_ofs=" << end_block_ofs << " ofs_in_end_block=" << ofs_in_end_block
                 << dendl;
-  need_read_start = ofs_in_start_block > 0;
+  need_read_start = (ofs_in_start_block >= 0 && (endoff - offset) <= FSCRYPT_BLOCK_SIZE);