aboutsummaryrefslogtreecommitdiffstats
path: root/driver/socket.c (follow)
Commit message (Collapse)AuthorAgeFilesLines
* driver: socket: defer WskInit until sockets are actually createdJason A. Donenfeld2021-09-241-72/+68
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | It turns out that MINIPORT_INITIALIZE can actually delay system load, and we currently have reports of people's systems hanging indefinitely. WskCaptureNPI is known to deadlock if called too early in boot (say, from DriverEntry of a PnP driver), but it was thought that MINIPORT_INITIALIZE was sufficiently late that it was okay. Perhaps that assumption is incorrect. In case it is, this patch moves WSK initialization to when sockets are created, which always happens in the context of a user thread, which naturally happens late in boot and can block. We know empirically that MINIPORT_INITIALIZE can block system boot, by adding a `KeDelayExecutionThread(KernelMode, FALSE, &(LARGE_INTEGER){ .QuadPart = -SEC_TO_SYS_TIME_UNITS(300) });` to the top and noting that boot takes 5 minutes longer. So the theory that the assumption is incorrect is at least plausible. All this commit does is move the call to WskInit() from InitializeEx in device.c to SocketInit() in socket.c. The diff looks more verbose than it is because making WskInit static and removing its forward declaration required shuffling some functions around in socket.c, but no code changed during that shuffle. Reported-by: Oliver Freyermuth <freyermuth@physik.uni-bonn.de> Reported-by: Joshua Sjoding <joshua.sjoding@scjalliance.com> Reported-by: John-Paul Andreini <jandreini@geonerco.com> Reported-by: Arlo Clauser <Arlo@starcubedesign.com> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: do not use zero UDPv4 checksumsJason A. Donenfeld2021-09-201-7/+0
| | | | | | | | | | Some routers using hardware NAT can't handle this. A packet goes through initially, but then once the flow is established and it's offloaded to the NAT hardware, the zero checksum causes most packets to be dropped. So, unfortunately, we have to remove this optimization. Reported-by: Christian Årebrand <myspysgaddan@hotmail.com> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: bypass TDI entirelyJason A. Donenfeld2021-09-131-0/+12
| | | | | | | Not only will this improve performance, but it will eliminate a big issue with IP_PKTINFO. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: use IP_OPTIONS for cmsg hackJason A. Donenfeld2021-09-101-14/+10
| | | | Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: remember to copy cmsghack when copying endpointJason A. Donenfeld2021-09-091-23/+16
| | | | | | | Otherwise, we can't reply to incoming endpoints. Reported-by: Peter Whisker <peter.whisker@gmail.com> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: work around tcpip cmsg stripping bugJason A. Donenfeld2021-09-071-6/+32
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | If the cmsg only contains an IP_PKTINFO or IP_PKTINFO_EX, then on some systems, the entire control message will be stripped out early before passing it to WFP. Presumably this is an optimization gone bad. The lack of a valid controlMessage member in FWPS_INCOMING_METADATA_VALUES0 when callouts are triggered results in big problems. Specifically, problems occur with drivers like NFSDK or McAfee or various other applications users install that install callouts that capture an outgoing packet, and then later reinject it (with, say, FwpsInjectTransportSendAsync0). McAfee does this for their DPI. NFSDK does this for their userspace parser library. Various things seem to use this technique. The problem is that when IN_PKTINFO is stripped from FWPS_INCOMING_METADATA_VALUES0, then it's not subsequently passed to FwpsInjectTransportSendAsync0, so it's as if the packet was sent without IP_PKTINFO in the first place. This causes routing loops, and users have a dysfunctional tunnel with high CPU usage, as packets route round and round. These issues go away when the callout and FwpsInjectTransportSendAsync0 dance is removed, indicating that IN_PKTINFO is working as intended in the rest of the networking stack. It turns out that the faulty optimization only triggers if IP_PKTINFO is the only control message used. So, we tag on a second control message that (hopefully) does nothing; IP_WFP_REDIRECT_RECORDS seems like a reasonable candidate. It happens to be Windows 8+, so we disable it on Windows 7, which doesn't need the hack anyway. Adding an extra control message and potentially adding additional overhead to the egress path is pretty awful, but currently, I'm not aware of a better workaround. Reported-by: Keshav Kejriwal <kesh.kejriwal@gmail.com> Reported-by: Kai Haberzettl <khaberz@gmail.com> Reported-by: Seyed Mohammad Hossein Amirkhalili <hosami@gmail.com> Reported-by: Francky Meyer <francky.meyer@hotmail.fr> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: allocate IRPs on stackJason A. Donenfeld2021-08-121-59/+45
| | | | Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: group v4 and v6 sockoptsJason A. Donenfeld2021-08-121-9/+6
| | | | Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: relax cmsg alignment, but assert macros matchJason A. Donenfeld2021-08-111-0/+19
| | | | Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: rework loop routing, endpoint resolution, and reduce ctxJason A. Donenfeld2021-08-091-186/+213
| | | | Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: allow loop routingJason A. Donenfeld2021-08-081-15/+1
| | | | | | | | You can still create loops, but ever since we started making copies on both RX and TX, and freeing TX buffers really early (after encryption), we no longer have a stack chaining issue on free. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: use implicit CMSG_SPACE for total length, for v6Jason A. Donenfeld2021-08-041-2/+2
| | | | | | | | | It expects to receive the total space, with padding, for the cmsg length argument, so pad out cmsg_len to the data padding multiple, which amounts to the same thing. This will fix IPv6 endpoints. Reported-by: Darren VanBuren <onekopaka@theoks.net> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: allocate large rows on heapJason A. Donenfeld2021-08-041-6/+14
| | | | | | | | The most ridiculous function gets even more ridiculous: we need a heap allocation. We should probably find a new strategy here, as this is growing unwieldy. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: memory: move NBL pools to global scopeJason A. Donenfeld2021-08-041-1/+1
| | | | | | | This is preparation for the next commit, which will attempt to allocate everything at once for the RX path. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: ignore interfaces with down oper statusJason A. Donenfeld2021-08-031-4/+7
| | | | | | This should allow better transition between wifi and wired. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: increment local endpoint update generationJason A. Donenfeld2021-08-031-1/+1
| | | | | | | Otherwise we wind up retrying and doing the routing generation comparison at the top needlessly. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* driver: socket: flip mask around on route lookupJason A. Donenfeld2021-08-031-2/+2
| | | | | | | Very silly error that caused non-default routes to be ignored. Reported-by: Peter Whisker <peter.whisker@gmail.com> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* Introduce WireGuardNTJason A. Donenfeld2021-08-021-0/+991
Co-authored-by: Simon Rozman <simon@rozman.si> Signed-off-by: Simon Rozman <simon@rozman.si> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Copyright © 1996 – 2023 Jason A. Donenfeld. All Rights Reverse Engineered.