Hi, Jakub,
On 22.12.2017 00:01, Jakub Kicinski wrote:
> We don't need the RTNL lock for all operations on offload state.
> We only need to hold it around ndo calls. The device offload
> initialization doesn't require it. The soon-to-come querying
> of the offload info will only need it partially. We will also
> be able to remove the waitqueue in following patches.
>
> Use struct rw_semaphore because map offload will require sleeping
> with the semaphore held for read.
>
> Suggested-by: Kirill Tkhai <ktk...@virtuozzo.com>
> Signed-off-by: Jakub Kicinski <jakub.kicin...@netronome.com>
> Reviewed-by: Quentin Monnet <quentin.mon...@netronome.com>
> ---
> v2:
> - use dev_get_by_index_rcu() instead of implicit lock dependencies;
> - use DECLARE_RWSEM() instead of init_rwsem() (Kirill).
> ---
> kernel/bpf/offload.c | 33 ++++++++++++++++++++++++++-------
> 1 file changed, 26 insertions(+), 7 deletions(-)
>
> diff --git a/kernel/bpf/offload.c b/kernel/bpf/offload.c
> index 8455b89d1bbf..f049073a37e6 100644
> --- a/kernel/bpf/offload.c
> +++ b/kernel/bpf/offload.c
> @@ -20,8 +20,12 @@
> #include <linux/netdevice.h>
> #include <linux/printk.h>
> #include <linux/rtnetlink.h>
> +#include <linux/rwsem.h>
>
> -/* protected by RTNL */
> +/* Protects bpf_prog_offload_devs and offload members of all progs.
> + * RTNL lock cannot be taken when holding this lock.
> + */
> +static DECLARE_RWSEM(bpf_devs_lock);
> static LIST_HEAD(bpf_prog_offload_devs);
>
> int bpf_prog_offload_init(struct bpf_prog *prog, union bpf_attr *attr)
> @@ -43,19 +47,30 @@ int bpf_prog_offload_init(struct bpf_prog *prog, union
> bpf_attr *attr)
> offload->prog = prog;
> init_waitqueue_head(&offload->verifier_done);
>
> - rtnl_lock();
> - offload->netdev = __dev_get_by_index(net, attr->prog_ifindex);
> + rcu_read_lock();
> + offload->netdev = dev_get_by_index_rcu(net, attr->prog_ifindex);
> if (!offload->netdev) {
> - rtnl_unlock();
> - kfree(offload);
> - return -EINVAL;
> + rcu_read_unlock();
> + goto err_free;
> }
> + dev_hold(offload->netdev);
> + rcu_read_unlock();
Small remark about this. There is already dev_get_by_index() in net/core/dev.c
with the functionality above. I haven't found in next patches the reason
we have to use directly rcu_read_lock() here.
So, it seems we may replace the above block with simple dev_get_by_index().
Everything else looks good for me.
>
> + down_write(&bpf_devs_lock);
> + if (offload->netdev->reg_state != NETREG_REGISTERED)
> + goto err_unlock;
> prog->aux->offload = offload;
> list_add_tail(&offload->offloads, &bpf_prog_offload_devs);
> - rtnl_unlock();
> + dev_put(offload->netdev);
> + up_write(&bpf_devs_lock);
>
> return 0;
> +err_unlock:
> + up_write(&bpf_devs_lock);
> + dev_put(offload->netdev);
> +err_free:
> + kfree(offload);
> + return -EINVAL;
> }
>
> static int __bpf_offload_ndo(struct bpf_prog *prog, enum bpf_netdev_command
> cmd,
> @@ -126,7 +141,9 @@ void bpf_prog_offload_destroy(struct bpf_prog *prog)
> wake_up(&offload->verifier_done);
>
> rtnl_lock();
> + down_write(&bpf_devs_lock);
> __bpf_prog_offload_destroy(prog);
> + up_write(&bpf_devs_lock);
> rtnl_unlock();
>
> kfree(offload);
> @@ -181,11 +198,13 @@ static int bpf_offload_notification(struct
> notifier_block *notifier,
> if (netdev->reg_state != NETREG_UNREGISTERING)
> break;
>
> + down_write(&bpf_devs_lock);
> list_for_each_entry_safe(offload, tmp, &bpf_prog_offload_devs,
> offloads) {
> if (offload->netdev == netdev)
> __bpf_prog_offload_destroy(offload->prog);
> }
> + up_write(&bpf_devs_lock);
> break;
> default:
> break;
Kirill
