Our internal state machine gets screwed up if selection events are not
monotonically increasing in time, and we can enter a self-copy loop from
the proxy window that exhausts all pipes.
Snippet of logs when this occurs:
00:00:46.238 [wlr] [xwayland/selection/incoming.c:487] XCB_XFIXES_SELECTION_NOTIFY (selection=277, owner=4194626)
00:00:46.238 [wlr] [xwayland/selection/incoming.c:487] XCB_XFIXES_SELECTION_NOTIFY (selection=277, owner=2097153)
00:00:46.238 [wlr] [xwayland/selection/outgoing.c:378] XCB_SELECTION_REQUEST (time=58979563 owner=2097153, requestor=2097153 selection=277, target=279, property=278)
00:00:46.238 [wlr] [xwayland/selection/outgoing.c:397] ignoring old request from timestamp 58979563; expecting > 58979563
00:00:46.238 [wlr] [xwayland/selection/outgoing.c:29] SendEvent destination=2097153 SelectionNotify(31) time=58979563 requestor=2097153 selection=277 target=279 property=0
00:00:46.238 [wlr] [xwayland/selection/incoming.c:453] XCB_SELECTION_NOTIFY (selection=277, property=0, target=279)
Note that 2097153 is `selection->window`, and 4194626 is Emacs.
The race occurs if the selection owner changes back to our proxy window
between when we get `XCB_XFIXES_SELECTION_NOTIFY` for Emacs and when we
call `xcb_convert_selection` in `incoming.c:source_send` -- the
ConvertSelection request can end up hitting our proxy window, but the
timestamp will be rejected.
Fixes#2192.
There seems to be no reason why we can't service multiple Wayland-to-X11
transfers concurrently, so long as they are to different windows (or
possibly, same windows but different target properties?)
This commit removes the queuing logic, but retains the request
de-duplication from #2428.
Previously, the clipboard and primary selections shared the same window.
This was racey, and could have led to pasting failures.
On xfixes selection owner change notification, the logic for requesting
the supported mimetypes of the new owner's selection looks like:
xcb_convert_selection(
xwm->xcb_conn,
selection->window,
selection->atom,
xwm->atoms[TARGETS],
xwm->atoms[WL_SELECTION],
selection->timestamp
);
This means ask the selection owner to write its TARGETS for the
`selection->atom` selection (one of PRIMARY, CLIPBOARD, DND_SELECTION)
to `selection->window`'s WL_SELECTION atom.
However, `selection->window` is shared for both PRIMARY and CLIPBOARD
selections, and WL_SELECTION is used as the target atom in both cases.
So, there's a race when both selections change at the same time.
The CLIPBOARD selection might support mimetypes {A, B, C}, and the
PRIMARY only {A, B}. If the ConvertSelection requests/responses "cross
on the wire", so to speak, wlroots can end up believing that the PRIMARY
selection also supports C.
A Wayland client may then ask for the PRIMARY selection in C format,
which will fail with "convert selection failed".
This commit fixes this by using a separate window for PRIMARY and
CLIPBOARD target requests, so that WL_SELECTION can be used as the
target atom in both cases.
This commit introduces logic for using a new X11 window for each
incoming transfer, rather than having a global window for each selection
source.
This eliminates a whole class of bugs involving multiple concurrent
incoming transfers.
For now, we retain the outgoing transfer queue, and the selection
source-specific windows to support it. Source-specific windows are no
longer used in the incoming path, and will be removed in a future PR.
Refs #1497.
Previously, Xwayland could restart, and we'd get events for transfers
pointing to the previous (now freed) xwm instance. This led to
use-after-free segfaults.
Closes#2565.
This will hopefully be fixed in the future by having separate windows
for each X11-to-Wayland transfer, but until then, let's avoid a
compositor crash.
Previously, `transfer->incr` was being cleared on the next selection.
However, if the next selection was *also* incremental, it's possible
that `xwm_handle_selection_property_notify` would route us to
`xwm_get_incr_chunk` instead of `xwm_selection_get_data`.
Apart from reducing duplication, this has the positive side-effect of
allowing all deallocs to use
`xwm_selection_transfer_destroy_property_reply`, as opposed to the
latter and a mix of ad-hoc `free`s.
Previously, if the Wayland client died before an incremental transfer
was complete, the logs would be spammed by "write error to target fd" as
wlroots entered some control flow wherein it'd continually try
scheduling further writes to the already-dead pipe.
This commit contains no behavioral changes, but introduces explicit
handling for draining the X11 selection in case of Wayland client death.
If `xwm_data_source_write` failed, it's failed permanently. In fact, a
failing `xwm_data_source_write` sets `transfer->property_reply` to
null as part of its error handling.
Instead of relying on an indirect check (whether
`transfer->property_reply` is still non-null), explicitly use the return
value from `xwm_data_source_write`.
The `fd` is marked `O_NONBLOCK`, so `write` will never spuriously return
`EINTR`. Therefore, `write` failing is permanent, and we can return 0 to
make the return value meaningful.
wlroots would log "Unhandled NET_WM_STATE property change" log
messages for atoms we know about. Simplify the code structure
and remove these extra messages.
Previously, wlr_xwm_selection_transfer.source_fd meant:
- the source of data in a Wayland -> X11 copy (good)
- the destination of data in a X11 -> Wayland copy (confusing)
This made reading through xwayland/selection/incoming.c difficult: in
many places, "source" actually means "destination".
Previously, any error would be masked by an internal isatty call:
24:31:48.174 [DEBUG] [wlr] [xwayland/selection/incoming.c:386] XCB_SELECTION_NOTIFY (selection=277, property=278, target=256)
24:31:48.174 [ERROR] [wlr] [xwayland/selection/incoming.c:30] write error to target fd: Inappropriate ioctl for device
In certain situations windows can have their input field set to false
but still expect to receive input focus by passively listening to key
presses via a parent window. The ICCCM specification outlines how focus
should be given to clients.
Further reading: https://tronche.com/gui/x/icccm/sec-4.html#s-4.1.7
Relates to #2604
Xwayland has its own special handling for signals like SIGSEGV/SIGABRT.
Instead of leaving the job to the OS, it tries to walk up the call stack
(badly, because a lot of information is missing), print the stack trace
to stdout, then exit(1). This is very annoying because it prevents
Xwayland crashes from being easily debugged.
Xwayland has a flag "-core" that aborts instead of exiting. This allows
the OS to generate a coredump. It's far from perfect but better than
nothing, I guess.
We already mostly did this, but there were a couple of branches
(`calloc` failures) where we'd bail without letting the other side know.
Refs swaywm/sway#4007. Likely not going to be a real improvement there
(if `calloc` fails you're already pretty screwed), but it does address a
theoretical possibility.
It seems that if we ever try to reply to a selection request after
another has been sent by the same requestor (we reply in FIFO order),
the requestor never reads from it, and we end up stalling forever on a
transfer that will never complete.
It appears that `XCB_SELECTION_REQUEST` has some sort of singleton
semantics, and new requests for the same selection are meant to replace
outstanding older ones. I couldn't find a reference for this, but
empirically this does seem to be the case.
Real (contrived) case where we don't currently do this, and things break:
* run fcitx
* run Slack
* wl-copy < <(base64 /opt/firefox/libxul.so) # or some other large file
* focus Slack (no need to paste)
fcitx will send in an `XCB_SELECTION_REQUEST`, and we'll start
processing it. Immediately after, Slack sends its own. fcitx hangs for a
long, long time. In the meantime, Slack retries and sends another
selection request. We now have two pending requests from Slack.
Eventually fcitx gives up (or it can be `pkill`'d), and we start
processing the first request Slack gave us (FIFO). Slack (Electron?)
isn't listening on the other end anymore, and this transfer never
completes. The X11 clipboard becomes unusable until Slack is killed.
After this patch, the clipboard is immediately usable again after fcitx
bails. Also added a bunch of debug-level logging that makes diagnosing
this sort of issue easier.
Refs swaywm/sway#4007.
When debugging Xwayland-related issues, a common first step in debugging
has been to ask the reporter to move their real Xwayland to
/usr/bin/Xwayland.bin, and create a shell script starting Xwayland with
extra arguments under the original /usr/bin/Xwayland location.
Introducing a `WLR_XWAYLAND` environment variable makes this less
invasive, by allowing the user to swap out Xwayland without resorting to
global system changes (or source patches).
Fixes#2425.
wlroots can only handle one outgoing transfer at a time, so it keeps a
list of pending selections. The head of the list is the currently-active
selection, and when that transfer completes and is destroyed, the next
one is started.
The trouble is when you have a transfer to some app that is misbehaving.
fcitx is one such application. With really large transfers, fcitx will
hang and never wake up again. So, you can end up with a transfer list
that looks like this:
| T1: started | T2: pending | T3: pending | T4: pending |
The file descriptor for transfer T1 is registered in libwayland's epoll
loop. The rest are waiting in wlroots' list.
As a user, you want your clipboard back, so you `pkill fcitx`. Now
Xwayland sends `XCB_DESTROY_NOTIFY` to let us know to give up. We clean
up T4 first.
Due to a bug in wlroots code, we register the (fd, transfer data
pointer) pair for T1 with libwayland *again*, despite it already being
registered. We do this 2 more times as we remove T3 and T2.
Finally, we remove T1 and `free` all the memory associated with it,
before `close`-ing its transfer file descriptor.
However, we still have 3 copies of T1's file descriptor left in the
epoll loop, since we erroneously added them as part of removing T2/3/4.
When we `close` the file descriptor as part of T1's teardown, we
actually cause the epoll loop to wake up the next time around, saying
"this file descriptor has activity!" (it was closed, so `read`-ing would
normally return 0 to let us know of EOF).
But instead of returning 0, it returns -1 with `EBADF`, because the file
descriptor has already been closed. And finally, as part of error-handling
this, we access the transfer pointer, which was `free`'d. And we crash.
This one was awful to track down, but calls to `wlr_log` with %m have
the errno masked by the `isatty` call in `log_stderr`. Switch them to
`wlr_log_errno` instead.
Cue quality "how can read(2) POSSIBLY be returning ENOTTY?" moments.
Certain clients require this property to be set for expected behavior.
Most notably, steam client CSD maximize button no longer worked
after unmaximizing once, unless the state was changed by another
method. The state is unset whenever another surface gains focus.
