On Tue, Feb 02, 2021 at 09:13:09AM +0200, Leon Romanovsky wrote: > From: Yishai Hadas <yishaih@xxxxxxxxxx> > > Cleanup the synchronize_srcu() from the ODP flow as it was found to be a > very heavy time consumer as part of dereg_mr. > > For example de-registration of 10000 ODP MRs each with size of 2M > hugepage took 19.6 sec comparing de-registration of same number of non > ODP MRs that took 172 ms. > > The new locking scheme uses the wait_event() mechanism which follows the > use count of the MR instead of using synchronize_srcu(). > > By that change, the time required for the above test took 95 ms which is > even better than the non ODP flow. > > Once fully dropped the srcu usage, had to come with a lock to protect > the XA access. > > As part of using the above mechanism we could also clean the > num_deferred_work stuff and follow the use count instead. > > Signed-off-by: Yishai Hadas <yishaih@xxxxxxxxxx> > Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxx> > --- > Changelog: > v2: > * Add checks of xa_erase() result as an outcome of memory error injection tests. > * Found extra place where we can change open-coded logic to use mlx5r_store_odp_mkey(). > v1: https://lore.kernel.org/linux-rdma/20210128064812.1921519-1-leon@xxxxxxxxxx > * Deleted not-relevant comment implicit_get_child_mr(), I have no idea > why wrong version of this patch was sent as v0. > * Deleted two new break lines added by me to make code more uniformly > in before "return ..." (sometimes it has new line, sometimes doesn't). > v0: https://lore.kernel.org/linux-rdma/20210127143051.1873866-1-leon@xxxxxxxxxx > --- > drivers/infiniband/hw/mlx5/devx.c | 13 +- > drivers/infiniband/hw/mlx5/main.c | 5 - > drivers/infiniband/hw/mlx5/mlx5_ib.h | 31 ++- > drivers/infiniband/hw/mlx5/mr.c | 26 +-- > drivers/infiniband/hw/mlx5/odp.c | 224 +++++++------------ > drivers/net/ethernet/mellanox/mlx5/core/mr.c | 1 + > include/linux/mlx5/driver.h | 2 + > 7 files changed, 127 insertions(+), 175 deletions(-) Applied to for-next, thanks Jason
