Merge branch 'ipv6-expired-routes'
Kui-Feng Lee says: ==================== Remove expired routes with a separated list of routes. FIB6 GC walks trees of fib6_tables to remove expired routes. Walking a tree can be expensive if the number of routes in a table is big, even if most of them are permanent. Checking routes in a separated list of routes having expiration will avoid this potential issue. Background ========== The size of a Linux IPv6 routing table can become a big problem if not managed appropriately. Now, Linux has a garbage collector to remove expired routes periodically. However, this may lead to a situation in which the routing path is blocked for a long period due to an excessive number of routes. For example, years ago, there is a commit c7bb4b89 ("ipv6: tcp: drop silly ICMPv6 packet too big messages"). The root cause is that malicious ICMPv6 packets were sent back for every small packet sent to them. These packets add routes with an expiration time that prompts the GC to periodically check all routes in the tables, including permanent ones. Why Route Expires ================= Users can add IPv6 routes with an expiration time manually. However, the Neighbor Discovery protocol may also generate routes that can expire. For example, Router Advertisement (RA) messages may create a default route with an expiration time. [RFC 4861] For IPv4, it is not possible to set an expiration time for a route, and there is no RA, so there is no need to worry about such issues. Create Routes with Expires ========================== You can create routes with expires with the command. For example, ip -6 route add 2001:b000:591::3 via fe80::5054:ff:fe12:3457 \ dev enp0s3 expires 30 The route that has been generated will be deleted automatically in 30 seconds. GC of FIB6 ========== The function called fib6_run_gc() is responsible for performing garbage collection (GC) for the Linux IPv6 stack. It checks for the expiration of every route by traversing the trees of routing tables. The time taken to traverse a routing table increases with its size. Holding the routing table lock during traversal is particularly undesirable. Therefore, it is preferable to keep the lock for the shortest possible duration. Solution ======== The cause of the issue is keeping the routing table locked during the traversal of large trees. To solve this problem, we can create a separate list of routes that have expiration. This will prevent GC from checking permanent routes. Result ====== We conducted a test to measure the execution times of fib6_gc_timer_cb() and observed that it enhances the GC of FIB6. During the test, we added permanent routes with the following numbers: 1000, 3000, 6000, and 9000. Additionally, we added a route with an expiration time. Here are the average execution times for the kernel without the patch. - 120020 ns with 1000 permanent routes - 308920 ns with 3000 ... - 581470 ns with 6000 ... - 855310 ns with 9000 ... The kernel with the patch consistently takes around 14000 ns to execute, regardless of the number of permanent routes that are installed. Major changes from v7: - Fix warings raised by the patchwork. Major changes from v6: - Remove unnecessary check of tb6 in fib6_clean_expires_locked(). - Use ib6_clean_expires_locked() instead in fib6_purge_rt(). Major changes from v5: - Change the order of adding new routes to the GC list and starting GC timer. - Remove time measurements from the test case. - Stop forcing GC flush. Major changes from v4: - Detect existence of 'strace' in the test case. Major changes from v3: - Fix the type of arg according to feedback. - Add 1k temporary routes and 5K permanent routes in the test case. Measure time spending on GC with strace. Major changes from v2: - Remove unnecessary and incorrect sysctl restoring in the test case. Major changes from v1: - Moved gc_link to avoid creating a hole in fib6_info. - Moved fib6_set_expires*() and fib6_clean_expires*() to the header file and inlined. And removed duplicated lines. - Added a test case. ==================== Signed-off-by:David S. Miller <davem@davemloft.net>
Showing
Please register or sign in to comment