On Mon, Dec 03, 2018 at 03:20:40PM -0700, Jason Gunthorpe wrote: > On Mon, Dec 03, 2018 at 05:07:53PM -0500, Doug Ledford wrote: > > On Mon, 2018-12-03 at 20:58 +0000, Jason Gunthorpe wrote: > > > Using the new UVERBS_METHOD_INVOKE_WRITE interface all of the existing > > > write() commands can be immediately converted to execute over ioctl. > > > > > > If -DIOCTL_MODE=ioctl is set then only the IOCTL interface is supported, > > > and write will not be called. Otherwise the 'both' mode will auto > > > detect the kernel support and use it. > > > > > > Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxxxx> > > > libibverbs/cmd_fallback.c | 44 ++++++++++++++++++++++++++++++++++++++- > > > libibverbs/device.c | 27 ++++++++++++++++++++++++ > > > libibverbs/ibverbs.h | 1 + > > > 3 files changed, 71 insertions(+), 1 deletion(-) > > > > > > This needs the matching kernel patch: > > > > > > https://patchwork.kernel.org/patch/10706127/ > > > > > > a kernel headers update and the series at: > > > > > > https://github.com/linux-rdma/rdma-core/pull/434 > > > > > > To apply. > > > > And long term, this is what we need to resolve the security issue/no > > fork problem... > > Yes > > > We leave the write() interface enabled in the kernel for backward > > compatibility for the foreseeable future (this is the whole "no > > regressions" requirement from Linus...we were required to regress in > > order to resolve the security issue, but other than that, we should > > require users to update rdma-core to keep working systems working, so > > the existing write() interface needs to stay online). > > Yes > > > We enable this ioctl-can-wrap-writes feature in rdma-core and in the > > kernel > > Yes, but this patch just provides the code, the default compile option > is left as write() only. > > All the provider teams should test with ioctl() only mode and confirm > there aren't any crazy things missed. I was imagining we'd give > another month or two for that, the moven the rdma-core default to both > (ie prefer ioctl if available) > > At some point testing of the write() side will diminish, but now that > everything is the same flow it shouldn't be too hard to keep working. I'm in favour to set ioctl as default right now, so those "providers teams" will go an test this path right now and not in some unknown future. Thanks
Attachment:
signature.asc
Description: PGP signature
