[LU-13118] change client instance to respect ASLR

• Type: Improvement
• Resolution: Unresolved
• Priority: Minor
• Fix Version/s: Upstream
• Assignee: James A Simmons
• Reporter: Andreas Dilger

Description

The client mount currently uses the superblock address as the unique configuration instance ("unsigned long cfg_instance" in the code) to distinguish multiple client mountpoints on the same node for the purpose of processing configuration records, which are treated as independent Lustre clients from the server point of view.

This cfg_instance can actually be any unique value as it is mostly just used as an identifier in messages and tunable parameters in procfs and sysfs. One place where it is actually checked is config_log_find(), but it only compares for identical instance numbers, and does not interpret the value itself.

To comply with ASLR requirements, it would be better to use some other identifier for cfg_instance. It does not need to be unique across all clients, only within a single client. One option would be a 64-bit random number, which has about a 1-in-4B chance of collision for multiple mounts on the same client, which is generally uncommon, but not impossible to hit. Another option would be to use the client mount UUID, but this is a much larger value and changes would be needed to many parts of the code. A compromise might be to use the 16-byte UUID folded over itself (i.e. first half XOR second half) to fit into an 8-byte value.

On the server, class_config_llog_handler() and config_log_find_or_add() use cfg_instance to pass an obd device reference, but this is only used to get the obd_name field. We may be able to use config_log_find() on the server to find the obd device back again? We might consider to just add the obd_name string into the cfg_instance itself? If we also added the client hostname into the cfg_instance this might be convenient to allow having tunable parameters that could be client specific (e.g. "llite.*_node27.max_pages_per_rpc"), but that might be overloading this too much.

On the test front, the cfg_instance is used to distinguish client vs. server OSC devices, mostly using "*-osc-[^M]*" or variants of this. There are a few tests that are using "*osc-[\-0\-9a\-f]*" but they should be made consistent.

Issue Links

• is related to: LU-13499 client UUID is truncated
• is related to: LU-12521 print_instance() incorrect if fsname contains a dash
• is related to: LU-11803 sanity test 255c fails with 'Ladvise test 12, bad lock count, returned 1, actual 0'
• is related to: LU-11809 conf-sanity test 28A hangs on file system mount
• is related to: LU-12511 Prepare lustre for adoption into the linux kernel
• is related to: LU-18076 add mechanisms to get client UUID from mountpoint