Re: [Qemu-devel] [PATCH] migration: Fix handling fd protocol

[Yury Kotov]

18.04.2019, 20:01, "Dr. David Alan Gilbert" <address@hidden>:
> * Yury Kotov (address@hidden) wrote:
>>  18.04.2019, 19:03, "Dr. David Alan Gilbert" <address@hidden>:
>>  > * Yury Kotov (address@hidden) wrote:
>>  >>  18.04.2019, 17:20, "Dr. David Alan Gilbert" <address@hidden>:
>>  >>  > * Yury Kotov (address@hidden) wrote:
>>  >>  >>  15.04.2019, 14:30, "Dr. David Alan Gilbert" <address@hidden>:
>>  >>  >>  > * Daniel P. Berrangé (address@hidden) wrote:
>>  >>  >>  >>  On Mon, Apr 15, 2019 at 12:15:12PM +0100, Dr. David Alan 
>> Gilbert wrote:
>>  >>  >>  >>  > * Daniel P. Berrangé (address@hidden) wrote:
>>  >>  >>  >>  > > On Mon, Apr 15, 2019 at 01:33:21PM +0300, Yury Kotov wrote:
>>  >>  >>  >>  > > > 15.04.2019, 13:25, "Daniel P. Berrangé" <address@hidden>:
>>  >>  >>  >>  > > > > On Mon, Apr 15, 2019 at 01:17:06PM +0300, Yury Kotov 
>> wrote:
>>  >>  >>  >>  > > > >>  15.04.2019, 13:11, "Daniel P. Berrangé" 
>> <address@hidden>:
>>  >>  >>  >>  > > > >>  > On Mon, Apr 15, 2019 at 12:50:08PM +0300, Yury 
>> Kotov wrote:
>>  >>  >>  >>  > > > >>  >>  Hi,
>>  >>  >>  >>  > > > >>  >>
>>  >>  >>  >>  > > > >>  >>  Just to clarify. I see two possible solutions:
>>  >>  >>  >>  > > > >>  >>
>>  >>  >>  >>  > > > >>  >>  1) Since the migration code doesn't receive fd, 
>> it isn't responsible for
>>  >>  >>  >>  > > > >>  >>  closing it. So, it may be better to use 
>> migrate_fd_param for both
>>  >>  >>  >>  > > > >>  >>  incoming/outgoing and add dupping for 
>> migrate_fd_param. Thus, clients must
>>  >>  >>  >>  > > > >>  >>  close the fd themselves. But existing clients 
>> will have a leak.
>>  >>  >>  >>  > > > >>  >
>>  >>  >>  >>  > > > >>  > We can't break existing clients in this way as they 
>> are correctly
>>  >>  >>  >>  > > > >>  > using the monitor with its current semantics.
>>  >>  >>  >>  > > > >>  >
>>  >>  >>  >>  > > > >>  >>  2) If we don't duplicate fd, then at least we 
>> should remove fd from
>>  >>  >>  >>  > > > >>  >>  the corresponding list. Therefore, the solution 
>> is to fix qemu_close to find
>>  >>  >>  >>  > > > >>  >>  the list and remove fd from it. But qemu_close is 
>> currently consistent with
>>  >>  >>  >>  > > > >>  >>  qemu_open (which opens/dups fd), so adding 
>> additional logic might not be
>>  >>  >>  >>  > > > >>  >>  a very good idea.
>>  >>  >>  >>  > > > >>  >
>>  >>  >>  >>  > > > >>  > qemu_close is not appropriate place to deal with 
>> something speciifc
>>  >>  >>  >>  > > > >>  > to the montor.
>>  >>  >>  >>  > > > >>  >
>>  >>  >>  >>  > > > >>  >>  I don't see any other solution, but I might miss 
>> something.
>>  >>  >>  >>  > > > >>  >>  What do you think?
>>  >>  >>  >>  > > > >>  >
>>  >>  >>  >>  > > > >>  > All callers of monitor_get_fd() will close() the FD 
>> they get back.
>>  >>  >>  >>  > > > >>  > Thus monitor_get_fd() should remove it from the 
>> list when it returns
>>  >>  >>  >>  > > > >>  > it, and we should add API docs to monitor_get_fd() 
>> to explain this.
>>  >>  >>  >>  > > > >>  >
>>  >>  >>  >>  > > > >>  Ok, it sounds reasonable. But monitor_get_fd is only 
>> about outgoing migration.
>>  >>  >>  >>  > > > >>  But what about the incoming migration? It doesn't use 
>> monitor_get_fd but just
>>  >>  >>  >>  > > > >>  converts input string to int and use it as fd.
>>  >>  >>  >>  > > > >
>>  >>  >>  >>  > > > > The incoming migration expects the FD to be passed into 
>> QEMU by the mgmt
>>  >>  >>  >>  > > > > app when it is exec'ing the QEMU binary. It doesn't 
>> interact with the
>>  >>  >>  >>  > > > > monitor at all AFAIR.
>>  >>  >>  >>  > > > >
>>  >>  >>  >>  > > >
>>  >>  >>  >>  > > > Oh, sorry. This use case is not obvious. We used add-fd 
>> to pass fd for
>>  >>  >>  >>  > > > migrate-incoming and such way has described problems.
>>  >>  >>  >>  > >
>>  >>  >>  >>  > > That's a bug in your usage of QEMU IMHO, as the incoming 
>> code is not
>>  >>  >>  >>  > > designed to use add-fd.
>>  >>  >>  >>  >
>>  >>  >>  >>  > Hmm, that's true - although:
>>  >>  >>  >>  > a) It's very non-obvious
>>  >>  >>  >>  > b) Unfortunate, since it would go well with -incoming defer
>>  >>  >>  >>
>>  >>  >>  >>  Yeah I think this is a screw up on QMEU's part when introducing 
>> 'defer'.
>>  >>  >>  >>
>>  >>  >>  >>  We should have mandated use of 'add-fd' when using 'defer', 
>> since FD
>>  >>  >>  >>  inheritance-over-execve() should only be used for command line 
>> args,
>>  >>  >>  >>  not monitor commands.
>>  >>  >>  >>
>>  >>  >>  >>  Not sure how to best fix this is QEMU though without breaking 
>> back
>>  >>  >>  >>  compat for apps using 'defer' already.
>>  >>  >>  >
>>  >>  >>  > We could add mon-fd: transports that has the same behaviour as 
>> now for
>>  >>  >>  > outgoing, and for incoming uses the add-fd stash.
>>  >>  >>  >
>>  >>  >>
>>  >>  >>  Oh, I'm sorry again. I think my suggestion about monitor_fd_param 
>> wasn't
>>  >>  >>  relevant to this issue. If migrate-incoming + "fd:" + add-fd is an 
>> invalid use
>>  >>  >>  case, should we disallow this?
>>  >>  >>  I may add a check to fd_start_incoming_migration if fd is in mon 
>> fds list.
>>  >>  >>  But I'm afraid there are users like me who are already using this 
>> wrong use case.
>>  >>  >>  Because currently nothing in QEMU's docs disallow this.
>>  >>  >>
>>  >>  >>  So which solution is better in your opinion?
>>  >>  >>  1) Disallow fd's from mon fds list in fd_start_incoming_migration
>>  >>  >
>>  >>  > I'm surprised anything could be doing that - how would a user know 
>> what
>>  >>  > the correct fd index was?
>>  >>  >
>>  >>
>>  >>  Hmm, add-fd returns correct fd value. Maybe I din't catch you 
>> question...
>>  >
>>  > I don't understand, where does it return it?
>>  >
>>
>>  From misc.json:
>>  # Example:
>>  #
>>  # -> { "execute": "add-fd", "arguments": { "fdset-id": 1 } }
>>  # <- { "return": { "fdset-id": 1, "fd": 3 } }
>>  #
>>
>>  "fd": 3 is a valid fd for migrate-incoming(uri = "fd:3")
>
> Ah OK.
>
>>  >>  >>  2) Allow these fds, but dup them or close them correctly
>>  >>  >
>>  >>  > I think I'd leave the current (confusing) fd: as it is, maybe put a 
>> note
>>  >>  > in the manual.
>>  >>  >
>>  >>
>>  >>  So, using fd from fdset will be an undefined behavior, right?
>>  >
>>  > For incoming, yes.
>>  >
>>  >>  >>  And how to migrate-incoming defer through fd correctly?
>>  >>  >>  1) Add "mon-fd:" protocol to work with fds passed by 
>> "add-fd/remove-fd" commands
>>  >>  >>  as suggested by Dave
>>  >>  >
>>  >>  > That's my preference; it's explicitly named and consistent, and it
>>  >>  > doesn't touch the existing fd code.
>>  >>  >
>>  >>
>>  >>  Ok, but please tell me what you think of my suggestion (2) about using 
>> fd added
>>  >>  by the "getfd" command for incoming migration. It doesn't requires 
>> introducing
>>  >>  new protocol and will be consistent with outgoing migration through fd.
>>  >
>>  > I worry how qemu knows whether the command means it comes from the getfd
>>  > command or is actually a normal fd like now?
>>  > Can you give an example.
>>  >
>>
>>  getfd manages naming fds list.
>>  # -> { "execute": "getfd", "arguments": { "fdname": "fd1" } }
>>  So, for migrate (not incoming) is now valid migrate(uri="fd:fd1")
>>
>>  I want the same for migrate-incoming. If fdname is parseable int, then it is
>>  an old format. Otherwise - it is a name of fd added by addfd.
>>
>>  There is a function "monitor_fd_param" which do exactly what I mean:
>>  int monitor_fd_param(Monitor *mon, const char *fdname, Error **errp) {
>>      ... local vars ...
>>      if (!qemu_isdigit(fdname[0]) && mon) {
>>          fd = monitor_get_fd(mon, fdname, &local_err);
>>      } else {
>>          fd = qemu_parse_fd(fdname);
>>      }
>>      ... report err to errp ...
>>  }
>
> OK, if we're already using monitor_fd_param everywhere then I think
> we're already down the rat-hole of guessing whether we're an add-fd or
> fd by whether it's an integer, and I agree with you that we should
> just fix incoming to use that.
>
> Now, that means I guess we need to modify monitor_fd_param to tell us
> which type of fd it got, so we know whether to close it later?
>
> Dave
> P.S. I'm out from tomorrow for a weekish.
>

I think the right way is to check whether fd is added by add-fd and if so then
return error. Because IIUC the only valid usage of add-fd is to use
qemu_open("/dev/fdset/<fdset_id>") which finds suitable fd from fdset.
Such behavior is incompatible with fd:<fd_num> at all, as such syntax
doesn't imply the using of particular fd. But if so, why add-fd returns
the value of added fd?..

But if I'm right it's enought to:
1) Modify monitor_fd_param to check where fd comes from and disallow using
   fd of "add-fd",
2) As we discussed earlier, modify monitor_get_fd to remove named fd from its
   list before return,
3) Use monitor_fd_param in migrate_incoming for "fd:" proto.

I'm not insist. May be it's ok to use fd from fdset directly and so qemu_close
should be modifyed.

Just to clarify what I mean:
fdset is a struct:
struct MonFdset {
    int64_t id;
    QLIST_HEAD(, MonFdsetFd) fds;
    QLIST_HEAD(, MonFdsetFd) dup_fds;
    QLIST_ENTRY(MonFdset) next;
};

* add-fd appends new fd to "->fds" list.
* qemu_open("/dev/fdset/X", int perms) is looking for suitable (by perms) fd
  from fdset with id X, dup it and append "->dup_fds" list.
* qemu_close(int fd) tryes to find the fd in all "->dup_fds" lists
  of all fdsets and remove it. And closes fd anyway.

If not to disallow using fds added by add-fd then there are three
possible solutions:
1) dup fd in monitor_fd_param it the fd is from some fdset,
2) remove the fd from "->fds" list in qemu_close
3) don't close it in qemu_close, so client is responsible to close it by
   remove-fd.

Regards,
Yury

>>  >>  >
>>  >>  >>  2) My suggestion about monitor_fd_param and make "fd:" for
>>  >>  >>  migrate/migrate-incoming consistent. So user will be able to use
>>  >>  >>  getfd + migrate-incoming
>>  >>  >>  3) Both of them or something else
>>  >>  >>
>>  >>
>>