A lot of the performance gain comes from this patch.
While analysing performance overhead it was found that the largest CPU
stalls were caused when touching the struct page area. It is first read with
a READ_ONCE from build_skb_around via page_is_pfmemalloc(), and when freed
written by page_frag_free() call.
Measurements show that the prefetchw (W) variant operation is needed to
achieve the performance gain. We believe this optimization it two fold,
first the W-variant saves one step in the cache-coherency protocol, and
second it helps us to avoid the non-temporal prefetch HW optimizations and
bring this into all cache-levels. It might be worth investigating if
prefetch into L2 will have the same benefit.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Acked-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
As cpumap now batch consume xdp_frame's from the ptr_ring, it knows how many
SKBs it need to allocate. Thus, lets bulk allocate these SKBs via
kmem_cache_alloc_bulk() API, and use the previously introduced function
build_skb_around().
Notice that the flag __GFP_ZERO asks the slab/slub allocator to clear the
memory for us. This does clear a larger area than needed, but my micro
benchmarks on Intel CPUs show that this is slightly faster due to being a
cacheline aligned area is cleared for the SKBs. (For SLUB allocator, there
is a future optimization potential, because SKBs will with high probability
originate from same page. If we can find/identify continuous memory areas
then the Intel CPU memset rep stos will have a real performance gain.)
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Move ptr_ring dequeue outside loop, that allocate SKBs and calls network
stack, as these operations that can take some time. The ptr_ring is a
communication channel between CPUs, where we want to reduce/limit any
cacheline bouncing.
Do a concentrated bulk dequeue via ptr_ring_consume_batched, to shorten the
period and times the remote cacheline in ptr_ring is read
Batch size 8 is both to (1) limit BH-disable period, and (2) consume one
cacheline on 64-bit archs. After reducing the BH-disable section further
then we can consider changing this, while still thinking about L1 cacheline
size being active.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
We want to avoid leaking pointer info from xdp_frame (that is placed in
top of frame) like commit 6dfb970d3d ("xdp: avoid leaking info stored in
frame data on page reuse"), and followup commit 97e19cce05 ("bpf:
reserve xdp_frame size in xdp headroom") that reserve this headroom.
These changes also affected how cpumap constructed SKBs, as xdpf->headroom
size changed, the skb data starting point were in-effect shifted with 32
bytes (sizeof xdp_frame). This was still okay, as the cpumap frame_size
calculation also included xdpf->headroom which were reduced by same amount.
A bug was introduced in commit 77ea5f4cbe ("bpf/cpumap: make sure
frame_size for build_skb is aligned if headroom isn't"), where the
xdpf->headroom became part of the SKB_DATA_ALIGN rounding up. This
round-up to find the frame_size is in principle still correct as it does
not exceed the 2048 bytes frame_size (which is max for ixgbe and i40e),
but the 32 bytes offset of pkt_data_start puts this over the 2048 bytes
limit. This cause skb_shared_info to spill into next frame. It is a little
hard to trigger, as the SKB need to use above 15 skb_shinfo->frags[] as
far as I calculate. This does happen in practise for TCP streams when
skb_try_coalesce() kicks in.
KASAN can be used to detect these wrong memory accesses, I've seen:
BUG: KASAN: use-after-free in skb_try_coalesce+0x3cb/0x760
BUG: KASAN: wild-memory-access in skb_release_data+0xe2/0x250
Driver veth also construct a SKB from xdp_frame in this way, but is not
affected, as it doesn't reserve/deduct the room (used by xdp_frame) from
the SKB headroom. Instead is clears the pointers via xdp_scrub_frame(),
and allows SKB to use this area.
The fix in this patch is to do like veth and instead allow SKB to (re)use
the area occupied by xdp_frame, by clearing via xdp_scrub_frame(). (This
does kill the idea of the SKB being able to access (mem) info from this
area, but I guess it was a bad idea anyhow, and it was already killed by
the veth changes.)
Fixes: 77ea5f4cbe ("bpf/cpumap: make sure frame_size for build_skb is aligned if headroom isn't")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The frame_size passed to build_skb must be aligned, else it is
possible that the embedded struct skb_shared_info gets unaligned.
For correctness make sure that xdpf->headroom in included in the
alignment. No upstream drivers can hit this, as all XDP drivers provide
an aligned headroom. This was discovered when playing with implementing
XDP support for mvneta, which have a 2 bytes DSA header, and this
Marvell ARM64 platform didn't like doing atomic operations on an
unaligned skb_shinfo(skb)->dataref addresses.
Fixes: 1c601d829a ("bpf: cpumap xdp_buff to skb conversion and allocation")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Commits 109980b894 ("bpf: don't select potentially stale ri->map
from buggy xdp progs") and 7c30013133 ("bpf: fix ri->map_owner
pointer on bpf_prog_realloc") tried to mitigate that buggy programs
using bpf_redirect_map() helper call do not leave stale maps behind.
Idea was to add a map_owner cookie into the per CPU struct redirect_info
which was set to prog->aux by the prog making the helper call as a
proof that the map is not stale since the prog is implicitly holding
a reference to it. This owner cookie could later on get compared with
the program calling into BPF whether they match and therefore the
redirect could proceed with processing the map safely.
In (obvious) hindsight, this approach breaks down when tail calls are
involved since the original caller's prog->aux pointer does not have
to match the one from one of the progs out of the tail call chain,
and therefore the xdp buffer will be dropped instead of redirected.
A way around that would be to fix the issue differently (which also
allows to remove related work in fast path at the same time): once
the life-time of a redirect map has come to its end we use it's map
free callback where we need to wait on synchronize_rcu() for current
outstanding xdp buffers and remove such a map pointer from the
redirect info if found to be present. At that time no program is
using this map anymore so we simply invalidate the map pointers to
NULL iff they previously pointed to that instance while making sure
that the redirect path only reads out the map once.
Fixes: 97f91a7cf0 ("bpf: add bpf_redirect_map helper routine")
Fixes: 109980b894 ("bpf: don't select potentially stale ri->map from buggy xdp progs")
Reported-by: Sebastiano Miano <sebastiano.miano@polito.it>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Daniel Borkmann says:
====================
pull-request: bpf-next 2018-08-13
The following pull-request contains BPF updates for your *net-next* tree.
The main changes are:
1) Add driver XDP support for veth. This can be used in conjunction with
redirect of another XDP program e.g. sitting on NIC so the xdp_frame
can be forwarded to the peer veth directly without modification,
from Toshiaki.
2) Add a new BPF map type REUSEPORT_SOCKARRAY and prog type SK_REUSEPORT
in order to provide more control and visibility on where a SO_REUSEPORT
sk should be located, and the latter enables to directly select a sk
from the bpf map. This also enables map-in-map for application migration
use cases, from Martin.
3) Add a new BPF helper bpf_skb_ancestor_cgroup_id() that returns the id
of cgroup v2 that is the ancestor of the cgroup associated with the
skb at the ancestor_level, from Andrey.
4) Implement BPF fs map pretty-print support based on BTF data for regular
hash table and LRU map, from Yonghong.
5) Decouple the ability to attach BTF for a map from the key and value
pretty-printer in BPF fs, and enable further support of BTF for maps for
percpu and LPM trie, from Daniel.
6) Implement a better BPF sample of using XDP's CPU redirect feature for
load balancing SKB processing to remote CPU. The sample implements the
same XDP load balancing as Suricata does which is symmetric hash based
on IP and L4 protocol, from Jesper.
7) Revert adding NULL pointer check with WARN_ON_ONCE() in __xdp_return()'s
critical path as it is ensured that the allocator is present, from Björn.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit a26ca7c982 ("bpf: btf: Add pretty print support to
the basic arraymap") and 699c86d6ec ("bpf: btf: add pretty
print for hash/lru_hash maps") enabled support for BTF and
dumping via BPF fs for array and hash/lru map. However, both
can be decoupled from each other such that regular BPF maps
can be supported for attaching BTF key/value information,
while not all maps necessarily need to dump via map_seq_show_elem()
callback.
The basic sanity check which is a prerequisite for all maps
is that key/value size has to match in any case, and some maps
can have extra checks via map_check_btf() callback, e.g.
probing certain types or indicating no support in general. With
that we can also enable retrieving BTF info for per-cpu map
types and lpm.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
When removing a cpumap entry, a number of syncronization steps happen.
Eventually the teardown code __cpu_map_entry_free is invoked from/via
call_rcu.
The teardown code __cpu_map_entry_free() flushes remaining xdp_frames,
by invoking bq_flush_to_queue, which calls xdp_return_frame_rx_napi().
The issues is that the teardown code is not running in the RX NAPI
code path. Thus, it is not allowed to invoke the NAPI variant of
xdp_return_frame.
This bug was found and triggered by using the --stress-mode option to
the samples/bpf program xdp_redirect_cpu. It is hard to trigger,
because the ptr_ring have to be full and cpumap bulk queue max
contains 8 packets, and a remote CPU is racing to empty the ptr_ring
queue.
Fixes: 389ab7f01a ("xdp: introduce xdp_return_frame_rx_napi")
Tested-by: Jean-Tsung Hsiao <jhsiao@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
When sending an xdp_frame through xdp_do_redirect call, then error
cases can happen where the xdp_frame needs to be dropped, and
returning an -errno code isn't sufficient/possible any-longer
(e.g. for cpumap case). This is already fully supported, by simply
calling xdp_return_frame.
This patch is an optimization, which provides xdp_return_frame_rx_napi,
which is a faster variant for these error cases. It take advantage of
the protection provided by XDP RX running under NAPI protection.
This change is mostly relevant for drivers using the page_pool
allocator as it can take advantage of this. (Tested with mlx5).
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Changing API xdp_return_frame() to take struct xdp_frame as argument,
seems like a natural choice. But there are some subtle performance
details here that needs extra care, which is a deliberate choice.
When de-referencing xdp_frame on a remote CPU during DMA-TX
completion, result in the cache-line is change to "Shared"
state. Later when the page is reused for RX, then this xdp_frame
cache-line is written, which change the state to "Modified".
This situation already happens (naturally) for, virtio_net, tun and
cpumap as the xdp_frame pointer is the queued object. In tun and
cpumap, the ptr_ring is used for efficiently transferring cache-lines
(with pointers) between CPUs. Thus, the only option is to
de-referencing xdp_frame.
It is only the ixgbe driver that had an optimization, in which it can
avoid doing the de-reference of xdp_frame. The driver already have
TX-ring queue, which (in case of remote DMA-TX completion) have to be
transferred between CPUs anyhow. In this data area, we stored a
struct xdp_mem_info and a data pointer, which allowed us to avoid
de-referencing xdp_frame.
To compensate for this, a prefetchw is used for telling the cache
coherency protocol about our access pattern. My benchmarks show that
this prefetchw is enough to compensate the ixgbe driver.
V7: Adjust for commit d9314c474d ("i40e: add support for XDP_REDIRECT")
V8: Adjust for commit bd658dda42 ("net/mlx5e: Separate dma base address
and offset in dma_sync call")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The generic xdp_frame format, was inspired by the cpumap own internal
xdp_pkt format. It is now time to convert it over to the generic
xdp_frame format. The cpumap needs one extra field dev_rx.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce an xdp_return_frame API, and convert over cpumap as
the first user, given it have queued XDP frame structure to leverage.
V3: Cleanup and remove C99 style comments, pointed out by Alex Duyck.
V6: Remove comment that id will be added later (Req by Alex Duyck)
V8: Rename enum mem_type to xdp_mem_type (found by kbuild test robot)
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There're several implications after commit 0bf7800f17 ("ptr_ring:
try vmalloc() when kmalloc() fails") with the using of vmalloc() since
can't allow GFP_ATOMIC but mandate GFP_KERNEL. This will lead a WARN
since cpumap try to call with GFP_ATOMIC. Fortunately, entry
allocation of cpumap can only be done through syscall path which means
GFP_ATOMIC is not necessary, so fixing this by replacing GFP_ATOMIC
with GFP_KERNEL.
Reported-by: syzbot+1a240cdb1f4cc88819df@syzkaller.appspotmail.com
Fixes: 0bf7800f17 ("ptr_ring: try vmalloc() when kmalloc() fails")
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: akpm@linux-foundation.org
Cc: dhowells@redhat.com
Cc: hannes@cmpxchg.org
Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Fixes the following sparse warnings:
kernel/bpf/cpumap.c:146:6: warning:
symbol '__cpu_map_queue_destructor' was not declared. Should it be static?
kernel/bpf/cpumap.c:225:16: warning:
symbol 'cpu_map_build_skb' was not declared. Should it be static?
kernel/bpf/cpumap.c:340:26: warning:
symbol '__cpu_map_entry_alloc' was not declared. Should it be static?
kernel/bpf/cpumap.c:398:6: warning:
symbol '__cpu_map_entry_free' was not declared. Should it be static?
kernel/bpf/cpumap.c:441:6: warning:
symbol '__cpu_map_entry_replace' was not declared. Should it be static?
kernel/bpf/cpumap.c:454:5: warning:
symbol 'cpu_map_delete_elem' was not declared. Should it be static?
kernel/bpf/cpumap.c:467:5: warning:
symbol 'cpu_map_update_elem' was not declared. Should it be static?
kernel/bpf/cpumap.c:505:6: warning:
symbol 'cpu_map_free' was not declared. Should it be static?
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
All map types reimplement the field-by-field copy of union bpf_attr
members into struct bpf_map. Add a helper to perform this operation.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Discovered that the compiler laid-out asm code in suboptimal way
when studying perf report during benchmarking of cpumap. Help
the compiler by the marking unlikely code paths.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
As pointed out by Michael, commit 1c601d829a ("bpf: cpumap xdp_buff
to skb conversion and allocation") contains a classical example of the
potential lost wake-up problem.
We need to recheck the condition __ptr_ring_empty() after changing
current->state to TASK_INTERRUPTIBLE, this avoids a race between
wake_up_process() and schedule(). After this, a race with
wake_up_process() will simply change the state to TASK_RUNNING, and
the schedule() call not really put us to sleep.
Fixes: 1c601d829a ("bpf: cpumap xdp_buff to skb conversion and allocation")
Reported-by: "Michael S. Tsirkin" <mst@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds two tracepoint to the cpumap. One for the enqueue side
trace_xdp_cpumap_enqueue() and one for the kthread dequeue side
trace_xdp_cpumap_kthread().
To mitigate the tracepoint overhead, these are invoked during the
enqueue/dequeue bulking phases, thus amortizing the cost.
The obvious use-cases are for debugging and monitoring. The
non-intuitive use-case is using these as a feedback loop to know the
system load. One can imagine auto-scaling by reducing, adding or
activating more worker CPUs on demand.
V4: tracepoint remove time_limit info, instead add sched info
V8: intro struct bpf_cpu_map_entry members cpu+map_id in this patch
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch makes cpumap functional, by adding SKB allocation and
invoking the network stack on the dequeuing CPU.
For constructing the SKB on the remote CPU, the xdp_buff in converted
into a struct xdp_pkt, and it mapped into the top headroom of the
packet, to avoid allocating separate mem. For now, struct xdp_pkt is
just a cpumap internal data structure, with info carried between
enqueue to dequeue.
If a driver doesn't have enough headroom it is simply dropped, with
return code -EOVERFLOW. This will be picked up the xdp tracepoint
infrastructure, to allow users to catch this.
V2: take into account xdp->data_meta
V4:
- Drop busypoll tricks, keeping it more simple.
- Skip RPS and Generic-XDP-recursive-reinjection, suggested by Alexei
V5: correct RCU read protection around __netif_receive_skb_core.
V6: Setting TASK_RUNNING vs TASK_INTERRUPTIBLE based on talk with Rik van Riel
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch connects cpumap to the xdp_do_redirect_map infrastructure.
Still no SKB allocation are done yet. The XDP frames are transferred
to the other CPU, but they are simply refcnt decremented on the remote
CPU. This served as a good benchmark for measuring the overhead of
remote refcnt decrement. If driver page recycle cache is not
efficient then this, exposes a bottleneck in the page allocator.
A shout-out to MST's ptr_ring, which is the secret behind is being so
efficient to transfer memory pointers between CPUs, without constantly
bouncing cache-lines between CPUs.
V3: Handle !CONFIG_BPF_SYSCALL pointed out by kbuild test robot.
V4: Make Generic-XDP aware of cpumap type, but don't allow redirect yet,
as implementation require a separate upstream discussion.
V5:
- Fix a maybe-uninitialized pointed out by kbuild test robot.
- Restrict bpf-prog side access to cpumap, open when use-cases appear
- Implement cpu_map_enqueue() as a more simple void pointer enqueue
V6:
- Allow cpumap type for usage in helper bpf_redirect_map,
general bpf-prog side restriction moved to earlier patch.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The 'cpumap' is primarily used as a backend map for XDP BPF helper
call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
This patch implement the main part of the map. It is not connected to
the XDP redirect system yet, and no SKB allocation are done yet.
The main concern in this patch is to ensure the datapath can run
without any locking. This adds complexity to the setup and tear-down
procedure, which assumptions are extra carefully documented in the
code comments.
V2:
- make sure array isn't larger than NR_CPUS
- make sure CPUs added is a valid possible CPU
V3: fix nitpicks from Jakub Kicinski <kubakici@wp.pl>
V5:
- Restrict map allocation to root / CAP_SYS_ADMIN
- WARN_ON_ONCE if queue is not empty on tear-down
- Return -EPERM on memlock limit instead of -ENOMEM
- Error code in __cpu_map_entry_alloc() also handle ptr_ring_cleanup()
- Moved cpu_map_enqueue() to next patch
V6: all notice by Daniel Borkmann
- Fix err return code in cpu_map_alloc() introduced in V5
- Move cpu_possible() check after max_entries boundary check
- Forbid usage initially in check_map_func_compatibility()
V7:
- Fix alloc error path spotted by Daniel Borkmann
- Did stress test adding+removing CPUs from the map concurrently
- Fixed refcnt issue on cpu_map_entry, kthread started too soon
- Make sure packets are flushed during tear-down, involved use of
rcu_barrier() and kthread_run only exit after queue is empty
- Fix alloc error path in __cpu_map_entry_alloc() for ptr_ring
V8:
- Nitpicking comments and gramma by Edward Cree
- Fix missing semi-colon introduced in V7 due to rebasing
- Move struct bpf_cpu_map_entry members cpu+map_id to tracepoint patch
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
