On 05.08.2021 12:06, Stefano Garzarella wrote: > Caution: This is an external email. Be cautious while opening links or attachments. > > > > On Thu, Aug 05, 2021 at 11:33:12AM +0300, Arseny Krasnov wrote: >> On 04.08.2021 15:57, Stefano Garzarella wrote: >>> Caution: This is an external email. Be cautious while opening links or attachments. >>> >>> >>> >>> Hi Arseny, >>> >>> On Mon, Jul 26, 2021 at 07:31:33PM +0300, Arseny Krasnov wrote: >>>> This patchset implements support of MSG_EOR bit for SEQPACKET >>>> AF_VSOCK sockets over virtio transport. >>>> Idea is to distinguish concepts of 'messages' and 'records'. >>>> Message is result of sending calls: 'write()', 'send()', 'sendmsg()' >>>> etc. It has fixed maximum length, and it bounds are visible using >>>> return from receive calls: 'read()', 'recv()', 'recvmsg()' etc. >>>> Current implementation based on message definition above. >>> Okay, so the implementation we merged is wrong right? >>> Should we disable the feature bit in stable kernels that contain it? Or >>> maybe we can backport the fixes... >> Hi, >> >> No, this is correct and it is message boundary based. Idea of this >> patchset is to add extra boundaries marker which i think could be >> useful when we want to send data in seqpacket mode which length >> is bigger than maximum message length(this is limited by transport). >> Of course we can fragment big piece of data too small messages, but >> this >> requires to carry fragmentation info in data protocol. So In this case >> when we want to maintain boundaries receiver calls recvmsg() until >> MSG_EOR found. >> But when receiver knows, that data is fit in maximum datagram length, >> it doesn't care about checking MSG_EOR just calling recv() or >> read()(e.g. >> message based mode). > I'm not sure we should maintain boundaries of multiple send(), from > POSIX standard [1]: Yes, but also from POSIX: such calls like send() and sendmsg() operates with "message" and if we check recvmsg() we will find the following thing: For message-based sockets, such as SOCK_DGRAM and SOCK_SEQPACKET, the entire message shall be read in a single operation. If a message is too long to fit in the supplied buffers, and MSG_PEEK is not set in the flags argument, the excess bytes shall be discarded. I understand this, that send() boundaries also must be maintained. I've checked SEQPACKET in AF_UNIX and AX_25 - both doesn't support MSG_EOR, so send() boundaries must be supported. > > SOCK_SEQPACKET > Provides sequenced, reliable, bidirectional, connection-mode > transmission paths for records. A record can be sent using one or > more output operations and received using one or more input > operations, but a single operation never transfers part of more than > one record. Record boundaries are visible to the receiver via the > MSG_EOR flag. > > From my understanding a record could be sent with multiple send() and > received, for example, with a single recvmsg(). > The only boundary should be the MSG_EOR flag set by the user on the last > send() of a record. You are right, if we talking about "record". > > From send() description [2]: > > MSG_EOR > Terminates a record (if supported by the protocol). > > From recvmsg() description [3]: > > MSG_EOR > End-of-record was received (if supported by the protocol). > > Thanks, > Stefano > > [1] > https://pubs.opengroup.org/onlinepubs/9699919799/functions/socket.html > [2] https://pubs.opengroup.org/onlinepubs/9699919799/functions/send.html > [3] > https://pubs.opengroup.org/onlinepubs/9699919799/functions/recvmsg.html P.S.: seems SEQPACKET is too exotic thing that everyone implements it in own manner, because i've tested SCTP seqpacket implementation, and found that: 1) It doesn't support MSG_EOR bit at send side, but uses MSG_EOR at receiver side to mark MESSAGE boundary. 2) According POSIX any extra bytes that didn't fit in user's buffer must be dropped, but SCTP doesn't drop it - you can read rest of datagram in next calls. > >
