On 1/7/19 2:06 PM, Eric Dumazet wrote: > > > On 01/07/2019 12:58 PM, David Ahern wrote: >> On 1/7/19 1:57 PM, Eric Dumazet wrote: >>> >>> >>> On 01/07/2019 12:41 PM, David Ahern wrote: >>>> From: David Ahern <dsah...@gmail.com> >>>> >>>> ip route uses ll_name_to_index to convert the user given device name to an >>>> index. At the moment ll_name_to_index uses if_nametoindex which is ioctl >>>> based and does not cache the result. When using a batch file this means >>>> the same device lookups can be done repeatedly adding unnecessary overhead >>>> (socket + ioctl call for each device lookup). >>>> >>>> Add a new function, ll_link_get, to send a netlink based RTM_GETLINK. If >>>> successful, cache the result in idx_head and name_head so future lookups >>>> can re-use the entry. >>>> >>>> With this change the time to install routes via a batch file is reduced >>>> from 30.7 seconds to 17.6 seconds (720,022 routes with 2 ecmp nexthops >>>> where the nexthop device is given). >>>> >>> >>> What time increase if we have 10,000 devices, and install one route ? >>> >>> Caching 10,000 devices would be quite a waste. >>> >> >> It only lookups up the device if ll_name_to_index is invoked for the >> device. >> >> This is NOT a link dump and cache; it only adds lookups to the cache. >> > > I see, maybe split this in two patches to avoid the confusion. > > Why is the RTNETLINK approach better than ioctl() ?
I could very well cache the if_nametoindex result but that is only name <--> idx. To avoid making assumptions on how the ll_map code is used in the future, RTM_GETLINK is used to leverage ll_remember_index and have the cache entries always be consistently filled. > > If this is the case, maybe if_nametoindex() should be fixed, instead of > carrying this locally in iproute2. > if_nametoindex has to do socket() - ioctl() - close() for every device lookup. It's not a bug and caching the socket descriptor makes assumptions that libc should not be making.
