libibverbs: Set vmr->access in ibv_cmd_reg_mr_ex()
[ Upstream commit 45c66ed5603b7a11491214fb3dc4a1d0b1626277 ]
ibv_cmd_reg_mr_ex() did not record the MR's access flags in vmr->access,
unlike ibv_cmd_reg_mr(). ibv_dereg_mr() decides whether to re-enable fork
tracking from verbs_get_mr(mr)->access:
if (... && !(access & IBV_ACCESS_ON_DEMAND))
ibv_dofork_range(addr, length);
With vmr->access left zero, an on-demand (ODP) MR registered through
ibv_reg_mr_ex() / ibv_cmd_reg_mr_ex() reads access == 0 at dereg, so
ibv_dofork_range() runs even though registration skipped the matching
ibv_dontfork_range() for ODP (need_fork = !(ON_DEMAND || FD)). The
unbalanced dofork corrupts the fork-range accounting when ibv_fork_init()
is enabled.
Set vmr->access = mr_init_attr->access on the success path, mirroring
ibv_cmd_reg_mr(), so the dereg fork decision is correct. (The write-ABI
fallback path already goes through ibv_cmd_reg_mr(), which sets it.)
Fixes: ca61708a8838 ("verbs: Add ibv_cmd_reg_mr_ex() to be used by drivers")
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Signed-off-by: Nicolas Morey <nmorey@suse.com>
This is the userspace components for the Linux Kernel's drivers/infiniband subsystem. Specifically this contains the userspace libraries for the following device nodes:
The userspace component of the libibverbs RDMA kernel drivers are included under the providers/ directory. Support for the following Kernel RDMA drivers is included:
Additional service daemons are provided for:
This project uses a cmake based build system. Quick start:
$ bash build.sh
build/bin will contain the sample programs and build/lib will contain the shared libraries. The build is configured to run all the programs ‘in-place’ and cannot be installed.
$ apt-get install build-essential cmake gcc libudev-dev libnl-3-dev libnl-route-3-dev ninja-build pkg-config valgrind python3-dev cython3 python3-docutils pandoc
Supported releases:
$ dnf builddep redhat/rdma-core.spec
NOTE: Fedora Core uses the name ‘ninja-build’ for the ‘ninja’ command.
$ zypper install cmake gcc libnl3-devel libudev-devel ninja pkg-config valgrind-devel python3-devel python3-Cython python3-docutils pandoc
Install required packages:
$ yum install cmake gcc libnl3-devel libudev-devel make pkgconfig valgrind-devel
Developers on CentOS 7 or Amazon Linux 2 are suggested to install more modern tooling for the best experience.
CentOS 7:
$ yum install epel-release $ yum install cmake3 ninja-build pandoc
Amazon Linux 2:
$ amazon-linux-extras install epel $ yum install cmake3 ninja-build pandoc
NOTE: EPEL uses the name ‘ninja-build’ for the ‘ninja’ command, and ‘cmake3’ for the ‘cmake’ command.
To set up software RDMA on an existing interface with either of the available drivers, use the following commands, substituting <DRIVER> with the name of the driver of your choice (rdma_rxe or siw) and <TYPE> with the type corresponding to the driver (rxe or siw).
# modprobe <DRIVER> # rdma link add <NAME> type <TYPE> netdev <DEVICE>
Please note that you need version of iproute2 recent enough is required for the command above to work.
You can use either ibv_devices or rdma link to verify that the device was successfully added.
Bugs should be reported to the linux-rdma@vger.kernel.org mailing list In your bug report, please include:
Information about your system:
How to reproduce the bug.
If the bug is a crash, the exact output printed out when the crash occurred, including any kernel messages produced.
See Contributing to rdma-core.
Stable versions are released regularly with backported fixes (see Documentation/stable.md) The current minimum version still maintained is ‘v33.X’