-----"Stefan Metzmacher" <metze@xxxxxxxxx> wrote: -----
>To: "Bernard Metzler" <bmt@xxxxxxxxxxxxxx>
>From: "Stefan Metzmacher" <metze@xxxxxxxxx>
>Date: 05/07/2021 01:37AM
>Cc: linux-rdma@xxxxxxxxxxxxxxx, "Stefan Metzmacher" <metze@xxxxxxxxx>
>Subject: [EXTERNAL] [PATCH 00/31] rdma/siw: fix a lot of deadlocks
>and use after free bugs
>
>Hi Bernard,
>
>while testing with my smbdirect driver I hit a lot of
>bugs in the siw.ko driver. They all cause problems where
>the siw driver was not able to unload anymore and I had to
>reboot the machine.
>
Hi Stefan,
Much appreciated!
These are quite some patches, and I will need some time
to go through. Would bee nice if those would be broken
down into smaller bundles (introduce non-blocking connect,
_siw_cep_close() subroutine, fixing cep reference counting,
smp_mb() after STag invalidation, ..). Anyway, many
thanks for the effort, it will improve the driver!
First comments:
A non blocking connect does really makes sense as you
are pointing out. I hope it doesn't complicate the CM
code even further.
I think we agreed upon not using BUG() and BUG_ON(),
so please don't introduce it.
'I hit a lot of bugs' is not very helpful, but just
a statement.
Thanks very much!
Bernard.
>I implemented:
>- a non blocking connect
>- fixed a lot of bugs where siw_cep_put() was called too often.
>- fixed bugs where siw_cm_upcall() confused the core IWCM logic
>
>I have some more changes to follow, but I wanted to send them
>finally out after having them one and a half year sitting in some
>private branch...
>
>Stefan Metzmacher (31):
> rdma/siw: fix warning in siw_proc_send()
> rdma/siw: call smp_mb() after mem->stag_valid = 0 in
> siw_invalidate_stag() too
> rdma/siw: remove superfluous siw_cep_put() from siw_connect() error
> path
> rdma/siw: let siw_accept() deferr RDMA_MODE until EVENT_ESTABLISHED
> rdma/siw: make use of kernel_{bind,connect,listen}()
> rdma/siw: make siw_cm_upcall() a noop without valid 'id'
> rdma/siw: split out a __siw_cep_terminate_upcall() function
> rdma/siw: use __siw_cep_terminate_upcall() for indirect
> SIW_CM_WORK_CLOSE_LLP
> rdma/siw: use __siw_cep_terminate_upcall() for
>SIW_CM_WORK_PEER_CLOSE
> rdma/siw: use __siw_cep_terminate_upcall() for
>SIW_CM_WORK_MPATIMEOUT
> rdma/siw: introduce SIW_EPSTATE_ACCEPTING/REJECTING for
> rdma_accept/rdma_reject
> rdma/siw: add some debugging of state and sk_state to the teardown
> process
> rdma/siw: handle SIW_EPSTATE_CONNECTING in
> __siw_cep_terminate_upcall()
> rdma/siw: let siw_connect() set AWAIT_MPAREP before
> siw_send_mpareqrep()
> rdma/siw: create a temporary copy of private data
> rdma/siw: use error and out logic at the end of siw_connect()
> rdma/siw: start mpa timer before calling siw_send_mpareqrep()
> rdma/siw: call the blocking kernel_bindconnect() just before
> siw_send_mpareqrep()
> rdma/siw: split out a __siw_cep_close() function
> rdma/siw: implement non-blocking connect.
> rdma/siw: let siw_listen_address() call siw_cep_alloc() first
> rdma/siw: let siw_listen_address() call siw_cep_set_inuse() early
> rdma/siw: make use of __siw_cep_close() in siw_accept()
> rdma/siw: do the full disassociation of cep and qp in
> siw_qp_llp_close()
> rdma/siw: fix double siw_cep_put() in siw_cm_work_handler()
> rdma/siw: make use of __siw_cep_close() in siw_cm_work_handler()
> rdma/siw: fix the "close" logic in siw_qp_cm_drop()
> rdma/siw: make use of __siw_cep_close() in siw_qp_cm_drop()
> rdma/siw: make use of __siw_cep_close() in siw_reject()
> rdma/siw: make use of __siw_cep_close() in siw_listen_address()
> rdma/siw: make use of __siw_cep_close() in siw_drop_listeners()
>
> drivers/infiniband/sw/siw/siw_cm.c | 537
>+++++++++++++++-----------
> drivers/infiniband/sw/siw/siw_cm.h | 3 +
> drivers/infiniband/sw/siw/siw_mem.c | 2 +
> drivers/infiniband/sw/siw/siw_qp.c | 3 +
> drivers/infiniband/sw/siw/siw_qp_rx.c | 2 +-
> 5 files changed, 316 insertions(+), 231 deletions(-)
>
>--
>2.25.1
>
>
