I noticed while reading the code that we have a documentation hole that may cause memory leaks for clients that are unaware, in relation to completion callbacks. The situation arises as follows: for all commands with a completion callback, I checked that the code has clean semantics: either nbd_aio_FOO() returns -1 and we never call the callback cleanup, or nbd_aio_FOO() returns a cookie and we call the callback cleanup when the command is retired. And since completion callbacks can only ever be passed to nbd_aio_FOO() functions, it is simple enough to document that when nbd_aio_FOO() fails (possible when calling them while the state machine is in the wrong state, or with invalid flags, or a command that the server is unwilling to accept - all things we can check client-side without traffic to the server), then the callback function was NOT registered, and the user must clean any resources then and there to avoid a leak. Only when nbd_aio_FOO() succeeds will the responsibility for cleanup be handled by the callback. But for nbd_block_status and nbd_pread_structured, we have a second callback. And _these_ callbacks have a problem: we can return -1 for two different reasons, either because the command was never attempted (nbd_aio_FOO failed) and so the callback will never be cleaned, or because the command got sent to the server and the server failed it (the callback WILL be cleaned). As this is ambiguous, it risks being a memory leak. So I think we have a bug in our code, and need to strengthen our promise of whether the cleanup callback will be called, regardless of success or failure, while still minimizing any incompat changes that might cause a double-free for existing clients. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org
On 9/5/20 7:47 AM, Eric Blake wrote:> I noticed while reading the code that we have a documentation hole that > may cause memory leaks for clients that are unaware, in relation to > completion callbacks. > > The situation arises as follows: for all commands with a completion > callback, I checked that the code has clean semantics: either > nbd_aio_FOO() returns -1 and we never call the callback cleanup, or > nbd_aio_FOO() returns a cookie and we call the callback cleanup when the > command is retired. And since completion callbacks can only ever be > passed to nbd_aio_FOO() functions, it is simple enough to document that > when nbd_aio_FOO() fails (possible when calling them while the state > machine is in the wrong state, or with invalid flags, or a command that > the server is unwilling to accept - all things we can check client-side > without traffic to the server), then the callback function was NOT > registered, and the user must clean any resources then and there to > avoid a leak. Only when nbd_aio_FOO() succeeds will the responsibility > for cleanup be handled by the callback. > > But for nbd_block_status and nbd_pread_structured, we have a second > callback. And _these_ callbacks have a problem: we can return -1 for > two different reasons, either because the command was never attempted > (nbd_aio_FOO failed) and so the callback will never be cleaned, or > because the command got sent to the server and the server failed it (the > callback WILL be cleaned). As this is ambiguous, it risks being a > memory leak. So I think we have a bug in our code, and need to > strengthen our promise of whether the cleanup callback will be called, > regardless of success or failure, while still minimizing any incompat > changes that might cause a double-free for existing clients.Thinking about it more, it is probably easiest to just declare that we guarantee the cleanup callback is called regardless of failure mode, for all closures, and that we merely had a memory leak in earlier libnbd releases. Yes, this means that we have a risk of older apps hitting a double free if they cleaned up after a cookie of -1 to compensate for our failure to cleanup; but it is unlikely, since the code for an app to call the cleanup manually is more verbose, and since good apps are unlikely to call functions that are going to fail client-side to trigger the problem in the first place. I'll post a patch, including coverage in tests/closure-lifetimes.c, to demonstrate what I mean. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org
On 9/5/20 8:56 AM, Eric Blake wrote:> Thinking about it more, it is probably easiest to just declare that we > guarantee the cleanup callback is called regardless of failure mode, for > all closures, and that we merely had a memory leak in earlier libnbd > releases. Yes, this means that we have a risk of older apps hitting a > double free if they cleaned up after a cookie of -1 to compensate for > our failure to cleanup; but it is unlikely, since the code for an app to > call the cleanup manually is more verbose, and since good apps are > unlikely to call functions that are going to fail client-side to trigger > the problem in the first place. I'll post a patch, including coverage > in tests/closure-lifetimes.c, to demonstrate what I mean.Proof that common clients have a memory leak (rather than a risk of a double free error), and thus that this is worth fixing by guaranteeing closure cleanup even on client-side failures: $ export count=10 $ nbdkit null 1m --run 'export uri; /usr/bin/time -v nbdsh -c " h.set_request_structured_replies(False) def f (c, o, e, err): pass import os h.connect_uri(os.environ[\"uri\"]) for _ in range(int(os.environ[\"count\"])): try: h.block_status(512, 0, f) except nbd.Error as ex: pass print(\"got here\") "' 2>&1 | grep '\(here\|resident\)' got here print("got here") Maximum resident set size (kbytes): 14416 Average resident set size (kbytes): 0 $ count=10000 $ nbdkit null 1m --run 'export uri; /usr/bin/time -v nbdsh -c " h.set_request_structured_replies(False) def f (c, o, e, err): pass import os h.connect_uri(os.environ[\"uri\"]) for _ in range(int(os.environ[\"count\"])): try: h.block_status(512, 0, f) except nbd.Error as ex: pass print(\"got here\") "' 2>&1 | grep '\(here\|resident\)' got here print("got here") Maximum resident set size (kbytes): 17796 Average resident set size (kbytes): 0 Ideally, increasing the loop iterations should have no effect on the max memory usage. Replacing the h.set_request_structured_replies(False) line with h.add_meta_context(nbd.CONTEXT_BASE_ALLOCATION) (so that the block status request succeeds instead) proves the point; after that change, I'm seeing 17296k with count=1 and 17284k with count=10000 (that is, the memory usage is higher and somewhat less deterministic because there is now server interactions, but definitely no longer something that scales up as count increases). And in proving that, I found several _other_ bugs, now fixed: Python.ml was mapping Bool incorrectly (so that h.set_request_structured_replies(False) was often setting things to true instead); which warranted testsuite coverage of functions previously uncalled under Python or Ocaml testsuites, and flushed out bugs in ocaml NBD.set_tls and NBD.set_handshake_flags. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org
Possibly Parallel Threads
- Re: libnbd completion callback question
- Re: libnbd completion callback question
- Re: [PATCH libnbd] generator: Swap parameters of nbd_add_close_callback.
- [libnbd] notify API changes (was: Re: [libnbd PATCH 5/6] api: Add new nbd_aio_FOO_notify functions)
- [nbdkit PATCH] nbd: Another libnbd version bump