]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
ceph: take the inode lock before acquiring cap refs
authorJeff Layton <jlayton@kernel.org>
Wed, 13 Nov 2019 14:10:27 +0000 (09:10 -0500)
committerIlya Dryomov <idryomov@gmail.com>
Thu, 14 Nov 2019 17:44:51 +0000 (18:44 +0100)
Most of the time, we (or the vfs layer) takes the inode_lock and then
acquires caps, but ceph_read_iter does the opposite, and that can lead
to a deadlock.

When there are multiple clients treading over the same data, we can end
up in a situation where a reader takes caps and then tries to acquire
the inode_lock. Another task holds the inode_lock and issues a request
to the MDS which needs to revoke the caps, but that can't happen until
the inode_lock is unwedged.

Fix this by having ceph_read_iter take the inode_lock earlier, before
attempting to acquire caps.

Fixes: 321fe13c9398 ("ceph: add buffered/direct exclusionary locking for reads and writes")
Link: https://tracker.ceph.com/issues/36348
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
fs/ceph/file.c

index bd77adb64bfd9812d25e0f0b92c0fb55689a9d7c..06efeaff3b5739fcee76435471aa39d8326460b1 100644 (file)
@@ -1264,14 +1264,24 @@ again:
        dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
             inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
 
+       if (iocb->ki_flags & IOCB_DIRECT)
+               ceph_start_io_direct(inode);
+       else
+               ceph_start_io_read(inode);
+
        if (fi->fmode & CEPH_FILE_MODE_LAZY)
                want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
        else
                want = CEPH_CAP_FILE_CACHE;
        ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1,
                            &got, &pinned_page);
-       if (ret < 0)
+       if (ret < 0) {
+               if (iocb->ki_flags & IOCB_DIRECT)
+                       ceph_end_io_direct(inode);
+               else
+                       ceph_end_io_read(inode);
                return ret;
+       }
 
        if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
            (iocb->ki_flags & IOCB_DIRECT) ||
@@ -1283,16 +1293,12 @@ again:
 
                if (ci->i_inline_version == CEPH_INLINE_NONE) {
                        if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
-                               ceph_start_io_direct(inode);
                                ret = ceph_direct_read_write(iocb, to,
                                                             NULL, NULL);
-                               ceph_end_io_direct(inode);
                                if (ret >= 0 && ret < len)
                                        retry_op = CHECK_EOF;
                        } else {
-                               ceph_start_io_read(inode);
                                ret = ceph_sync_read(iocb, to, &retry_op);
-                               ceph_end_io_read(inode);
                        }
                } else {
                        retry_op = READ_INLINE;
@@ -1303,11 +1309,10 @@ again:
                     inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
                     ceph_cap_string(got));
                ceph_add_rw_context(fi, &rw_ctx);
-               ceph_start_io_read(inode);
                ret = generic_file_read_iter(iocb, to);
-               ceph_end_io_read(inode);
                ceph_del_rw_context(fi, &rw_ctx);
        }
+
        dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
             inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
        if (pinned_page) {
@@ -1315,6 +1320,12 @@ again:
                pinned_page = NULL;
        }
        ceph_put_cap_refs(ci, got);
+
+       if (iocb->ki_flags & IOCB_DIRECT)
+               ceph_end_io_direct(inode);
+       else
+               ceph_end_io_read(inode);
+
        if (retry_op > HAVE_RETRIED && ret >= 0) {
                int statret;
                struct page *page = NULL;