This ensures that if the client or agent connects to the client-agent
"tunnel" while the other side is halfway through sending a multi part
message, the rest of the message gets discarded, and the connecting
party starts getting data at the beginning of the next message.
The agent message filter keeps track of messages as they are being send
reset the relevant filter to its initial state when one of the 2 ends
of the agent<->client "tunnel" disconnects.
read_from_vdi_port calls dispatch_vdi_port data, which will disconnect
the guest agent if it sends invalid data. It would then try to read more
data from the disconnected guest agent resulting in a NULL ptr dereference,
this patch fixes this.
write_to_vdi_port() was checking for reds->agent_state.connected to determine
wether it could write queued data. But agent_state.connected reflects if
*both* ends are connected. If the client has disconnected, but the guest agent
is still connected and some data is still pending (like a final clipboard
release from the client), then this data should be written to the guest agent.
We were calling reds_reset_vdp on client disconnect, which is not a good
idea. reds_reset_vdp does 3 things:
1) It resets the state related to reading chunks from the spicevmc virtio
port. If the client disconnects while the guest agent is in the middle
of sending a chunk, this will lead to an inconsistent state, and lots
of printing of "dispatch_vdi_port_data: invalid port" messages caused
by this inconsistent state sometimes followed by a segfault.
This can be triggered by copy and pasting something large (say
a screenshot) from the guest to the spice-gtk client, as the spice-gtk
client currently has a bug causing it to crash when receiving a multi
chunk vdagent messages. Without this patch (and with the spice-gtk bug
present) I can consistently reproduce this.
2) It clears any buffered writes from the client to the guest still pending
because the virtio port cannot consume data fast enough. Since the agent
itself is still running fine, throwing away writes for it because the
client has disconnected makes no sense. Esp, since on clean exit the
client may very well send a clipboard release message directly
before closing the connection, and this may get lost this way.
3) It sets client_agent_started to false, this is the only thing which
actually makes sense to do on client disconnect.
Note that since we no longer reset the vdp state on client disconnect, we
must now reset it on agent disconnect even if we don't have a client. So
the reds_reset_vdp call in reds_agent_remove() gets moved to the top,
above both the agent_state.connected and reds->peer checks which will
both fail in the no client case.
dispatch_vdi_port_data, was calling vdi_read_buf_release when no client
is connected to free the passed in buf. The only difference between
vdi_read_buf_release and directly adding the buffer back to the ring
with ring_add, is that vdi_read_buf_release calls read_from_vdi_port
after adding the buffer back. But dispatch_vdi_port_data only gets called
from read_from_vdi_port itself, thus this would lead to recursing into
read_from_vdi_port. read_from_vdi_port is protected against recursion and
will immediately return if called recursively. Thus calling
vdi_read_buf_release from dispatch_vdi_port_data is pointless, instead
simply putting the buffer back in the ring suffices.
Also bump SPICE_SERVER_VERSION to 0x000801 as 0.8.1 will be the
first version with the new API for this, and we need to be able to
detect the presence of this API in qemu.
The current assert(reds->agent_state.connected) tiggers if when
the agent disconnected there was still data waiting to be sent (for
instance if there is a bug in the client handling clipboard and it
is killed while a large clipboard transfer is in progress). So first
call to reds_agent_remove happens from spice_server_char_device_remove_interface,
and then it is called again (triggering the assert) from free_item_data
from read_from_vdi_port because of the channel destruction.
Other option would be to not call it from one of the paths - but that
is suboptimal:
* if there is no data in the pipe, the second call never happens.
* the second call has to be there anyway, because it may fail during
parsing data from the agent.
This patch fixes a segfault on this assert when a client starts passing
from guest agent to client a large clipboard and dies in the middle. There
is still another assert happening occasionally at marshaller which I don't
have a fix for (but it doesn't seem to be related).
Current master is calling red_channel_destroy() on incoming error, but
reds Channels still references it, which causes a double free() later
on (see valgrind report below).
Instead, on error condition, do like the rest of the channels and call
reds_disconnect(), which remove the references and call shutdown(),
which then call red_channel_destroy() and finally free the channel
with red_channel_destroy().
Note: the previous code intention was certainly to be able to keep the
rest of the channels connected when input channel has errors. This is
not addressed by this patch.
red_channel_shutdown:
==29792== Invalid read of size 8
==29792== at 0x4C6F063: red_channel_shutdown (red_channel.c:460)
==29792== by 0x4C51EFA: inputs_shutdown (inputs_channel.c:463)
==29792== by 0x4C48445: reds_shatdown_channels (reds.c:539)
==29792== by 0x4C4868A: reds_disconnect (reds.c:603)
==29792== by 0x4C519E9: main_channel_on_error (main_channel.c:765)
==29792== by 0x4C6E80A: red_channel_peer_on_incoming_error (red_channel.c:215)
==29792== by 0x4C6E22D: red_peer_handle_incoming (red_channel.c:87)
==29792== by 0x4C6E551: red_channel_receive (red_channel.c:154)
==29792== by 0x4C6F329: red_channel_event (red_channel.c:531)
==29792== by 0x41CB8C: main_loop_wait (vl.c:1365)
==29792== by 0x437CDE: kvm_main_loop (qemu-kvm.c:1589)
==29792== by 0x41FE9A: main (vl.c:1411)
==29792== Address 0x30b0f6d0 is 0 bytes inside a block of size 28,648 free'd
==29792== at 0x4A05372: free (vg_replace_malloc.c:366)
==29792== by 0x4C6F032: red_channel_destroy (red_channel.c:454)
==29792== by 0x4C6E80A: red_channel_peer_on_incoming_error (red_channel.c:215)
==29792== by 0x4C6E22D: red_peer_handle_incoming (red_channel.c:87)
==29792== by 0x4C6E551: red_channel_receive (red_channel.c:154)
==29792== by 0x4C6F329: red_channel_event (red_channel.c:531)
==29792== by 0x41CB8C: main_loop_wait (vl.c:1365)
==29792== by 0x437CDE: kvm_main_loop (qemu-kvm.c:1589)
==29792== by 0x41FE9A: main (vl.c:1411)
https://bugs.freedesktop.org/show_bug.cgi?id=34971
This reverts commit 5062433d8a.
red_channel_receive() can call red_channel_destroy() which frees
channel.
The condition bellow is then checked, which can access a freed
channel:
if (event & SPICE_WATCH_EVENT_WRITE || channel->send_data.blocked)
Reverting this commit solves the issue without any apparent
bugs/drawbacks, which kind of clears out the weird TODO.
handle_dev_input: cursor connect
==11826== Invalid read of size 4
==11826== at 0x4C6F83C: red_channel_event (red_channel.c:535)
==11826== by 0x41CB8C: main_loop_wait (vl.c:1365)
==11826== by 0x437CDE: kvm_main_loop (qemu-kvm.c:1589)
==11826== by 0x41FE9A: main (vl.c:1411)
==11826== Address 0x31fb00f0 is 96 bytes inside a block of size 28,648 free'd
==11826== at 0x4A05372: free (vg_replace_malloc.c:366)
==11826== by 0x4C6F536: red_channel_destroy (red_channel.c:453)
==11826== by 0x4C52B5D: inputs_channel_on_incoming_error (inputs_channel.c:449)
==11826== by 0x4C6ED0E: red_channel_peer_on_incoming_error (red_channel.c:215)
==11826== by 0x4C6E731: red_peer_handle_incoming (red_channel.c:87)
==11826== by 0x4C6EA55: red_channel_receive (red_channel.c:154)
==11826== by 0x4C6F82D: red_channel_event (red_channel.c:530)
==11826== by 0x41CB8C: main_loop_wait (vl.c:1365)
==11826== by 0x437CDE: kvm_main_loop (qemu-kvm.c:1589)
==11826== by 0x41FE9A: main (vl.c:1411)
==11826==
https://bugs.freedesktop.org/show_bug.cgi?id=34971
This allows later to have the callback table under RedChannel when
the callbacks actually get used by RedChannelClient. Since the cb's
are identical for different clients of the same channel it makes sense
to store the callback pointers in one place per channel. The rest of
the incoming and outgoing struct just gets moved to RedChannelClient.
Changes in display channel for a code size win.
A note about this and the previous cursor change: it will appear that we are
now (with these changes) releasing resources too early. This is not so - send
always has the option of blocking, which means after send you can not release
resources anyway, that's what the release_item callback is for. So both the
code before and now are doing the same accounting.
Use in main_channel. This is just for backward portability later
when multiple clients are introduced - needs to be considered (which
sockets do we want to export from libspiceserver?)
Introduce SpiceMarshaller param to all send's that do add_buf
Next patch will use marshaller in all functions that currently don't by
replacing red_channel_add_buf with marshaller add_ref. Note - currently
tunnel is broken due to wrong size in messages.
move all the ASSERT/PANIC/PANIC_ON/red_error/red_printf* macros
to a common file to be used with ring.h that is going to be used externally
(by spice-gtk).
Handling done in red_channel instead of per channel, using call backs
for the channel specific part.
Intended to reduce furthur reliance of channels on RedChannel struct.
The commit makes the code harder to understand because of the artificial
get_serial stuff, should later be fixed by having a joint migration
header with the serial (since all channels pass it).
For ussage in the send_item callback. It's only valid during this
time anyway (should make it return NULL in other occasions?)
No more direct usage of RedChannel.send_data.marshaller by channels.
