On Sun, Oct 27, 2019 at 10:04:47PM +0200, Michal Kalderon wrote: > Several leaks and issues were found when running iWARP with kmemleak. > some apply to RoCE as well. > > This series fixes some memory leaks and some wrong methods of > synchronization which were used to wait for iWARP CM related events. > > Changes from V3 > - call xa_init for the qpids xarray. > - add another patch that calls xa_init_flags for srqs xarray. > > Changes from V2 > - Add a new separate patch that fixes the xarray api that was used > for the qps xarray, there was no need to use the _irq version of > the api. > > - Move xa_erase of qp_id to be right before the qp resources are > released. This fixes a race where the qp-id can be reassigned > before removed from the xarray. > > - Modify places that call kref_get_unless_zero to kref_get since we > already hold a valid pointer. > > - Comment about the usage of the same completion structure for two > different completions. > > - Add Fixes tag > > Changes from v1 > - When removing the qp from the xarray xa_erase should be used and > not xa_erase_irq as this can't be called from irq context. > > - Add xa_lock around loading a qp from the xarray and increase the > refcnt only under the xa_lock and only if not zero. This is to make > qedr more robust and not rely on the core/iwcm implementation to > assure correctness. > > - Complete the iwarp_cm_comp event only if the bit was turned on and > the destroy qp flow will attempt to look at the completion > > > Michal Kalderon (4): > RDMA/qedr: Fix srqs xarray initialization > RDMA/qedr: Fix qpids xarray api used > RDMA/qedr: Fix synchronization methods and memory leaks in qedr > RDMA/qedr: Fix memory leak in user qp and mr Applied to for-next, thanks Jason
