Intel drivers translate actions returned from XDP programs to their own
return codes that have the following mapping:
XDP_REDIRECT -> IXGBE_XDP_{REDIR,CONSUMED}
XDP_TX -> IXGBE_XDP_{TX,CONSUMED}
XDP_DROP -> IXGBE_XDP_CONSUMED
XDP_ABORTED -> IXGBE_XDP_CONSUMED
XDP_PASS -> IXGBE_XDP_PASS
Commit c7dd09fd46 ("ixgbe, xsk: Terminate Rx side of NAPI when XSK Rx
queue gets full") introduced new translation
XDP_REDIRECT -> IXGBE_XDP_EXIT
which is set when XSK RQ gets full and to indicate that driver should
stop further Rx processing. This happens for unsuccessful
xdp_do_redirect() so it is valuable to call trace_xdp_exception() for
this case. In order to avoid IXGBE_XDP_EXIT -> IXGBE_XDP_CONSUMED
overwrite, XDP_DROP case was moved above which in turn made the
'fallthrough' that is in XDP_ABORTED useless as it became the last label
in the switch statement.
Simply drop this leftover.
Fixes: c7dd09fd46 ("ixgbe, xsk: Terminate Rx side of NAPI when XSK Rx queue gets full")
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20220421132126.471515-2-maciej.fijalkowski@intel.com
Currently, when debugging AF_XDP workloads, one can correlate the -ENXIO
return code as the case that XSK is not in the bound state. Returning
same code from ndo_xsk_wakeup can be misleading and simply makes it
harder to follow what is going on.
Change ENXIOs in ixgbe's ndo_xsk_wakeup() implementation to EINVALs, so
that when probing it is clear that something is wrong on the driver
side, not the xsk_{recv,send}msg.
There is a -ENETDOWN that can happen from both kernel/driver sides
though, but I don't have a correct replacement for this on one of the
sides, so let's keep it that way.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20220413153015.453864-11-maciej.fijalkowski@intel.com
When XSK pool uses need_wakeup feature, correlate -ENOBUFS that was
returned from xdp_do_redirect() with a XSK Rx queue being full. In such
case, terminate the Rx processing that is being done on the current HW
Rx ring and let the user space consume descriptors from XSK Rx queue so
that there is room that driver can use later on.
Introduce new internal return code IXGBE_XDP_EXIT that will indicate case
described above.
Note that it does not affect Tx processing that is bound to the same
NAPI context, nor the other Rx rings.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20220413153015.453864-8-maciej.fijalkowski@intel.com
ixgbe_run_xdp_zc() suggests to compiler that XDP_REDIRECT is the most
probable action returned from BPF program that AF_XDP has in its
pipeline. Let's also bring this suggestion up to the callsite of
ixgbe_run_xdp_zc() so that compiler will be able to generate more
optimized code which in turn will make branch predictor happy.
Suggested-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20220413153015.453864-5-maciej.fijalkowski@intel.com
Commit c685c69fba ("ixgbe: don't do any AF_XDP zero-copy transmit if
netif is not OK") addressed the ring transient state when
MEM_TYPE_XSK_BUFF_POOL was being configured which in turn caused the
interface to through down/up. Maurice reported that when carrier is not
ok and xsk_pool is present on ring pair, ksoftirqd will consume 100% CPU
cycles due to the constant NAPI rescheduling as ixgbe_poll() states that
there is still some work to be done.
To fix this, do not set work_done to false for a !netif_carrier_ok().
Fixes: c685c69fba ("ixgbe: don't do any AF_XDP zero-copy transmit if netif is not OK")
Reported-by: Maurice Baijens <maurice.baijens@ellips.com>
Tested-by: Maurice Baijens <maurice.baijens@ellips.com>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For now, if the XDP prog returns XDP_PASS on XSK, the metadata
will be lost as it doesn't get copied to the skb.
Copy it along with the frame headers. Account its size on skb
allocation, and when copying just treat it as a part of the frame
and do a pull after to "move" it to the "reserved" zone.
net_prefetch() xdp->data_meta and align the copy size to speed-up
memcpy() a little and better match ixgbe_construct_skb().
Fixes: d0bcacd0a1 ("ixgbe: add AF_XDP zero-copy Rx support")
Suggested-by: Jesper Dangaard Brouer <brouer@redhat.com>
Suggested-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexander Lobakin <alexandr.lobakin@intel.com>
Reviewed-by: Michal Swiatkowski <michal.swiatkowski@linux.intel.com>
Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
{__,}napi_alloc_skb() allocates and reserves additional NET_SKB_PAD
+ NET_IP_ALIGN for any skb.
OTOH, ixgbe_construct_skb_zc() currently allocates and reserves
additional `xdp->data - xdp->data_hard_start`, which is
XDP_PACKET_HEADROOM for XSK frames.
There's no need for that at all as the frame is post-XDP and will
go only to the networking stack core.
Pass the size of the actual data only to __napi_alloc_skb() and
don't reserve anything. This will give enough headroom for stack
processing.
Fixes: d0bcacd0a1 ("ixgbe: add AF_XDP zero-copy Rx support")
Signed-off-by: Alexander Lobakin <alexandr.lobakin@intel.com>
Reviewed-by: Michal Swiatkowski <michal.swiatkowski@linux.intel.com>
Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
To not dereference bi->xdp each time in ixgbe_construct_skb_zc(),
pass bi->xdp as an argument instead of bi. We can also call
xsk_buff_free() outside of the function as well as assign bi->xdp
to NULL, which seems to make it closer to its name.
Suggested-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexander Lobakin <alexandr.lobakin@intel.com>
Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
In non trivial scenarios, the action id alone is not sufficient to
identify the program causing the warning. Before the previous patch,
the generated stack-trace pointed out at least the involved device
driver.
Let's additionally include the program name and id, and the relevant
device name.
If the user needs additional infos, he can fetch them via a kernel
probe, leveraging the arguments added here.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/bpf/ddb96bb975cbfddb1546cf5da60e77d5100b533c.1638189075.git.pabeni@redhat.com
Originally, ixgbe driver doesn't allow the mounting of xdpdrv if the
server is equipped with more than 64 cpus online. So it turns out that
the loading of xdpdrv causes the "NOMEM" failure.
Actually, we can adjust the algorithm and then make it work through
mapping the current cpu to some xdp ring with the protect of @tx_lock.
Here are some numbers before/after applying this patch with xdp-example
loaded on the eth0X:
As client (tx path):
Before After
TCP_STREAM send-64 734.14 714.20
TCP_STREAM send-128 1401.91 1395.05
TCP_STREAM send-512 5311.67 5292.84
TCP_STREAM send-1k 9277.40 9356.22 (not stable)
TCP_RR send-1 22559.75 21844.22
TCP_RR send-128 23169.54 22725.13
TCP_RR send-512 21670.91 21412.56
As server (rx path):
Before After
TCP_STREAM send-64 1416.49 1383.12
TCP_STREAM send-128 3141.49 3055.50
TCP_STREAM send-512 9488.73 9487.44
TCP_STREAM send-1k 9491.17 9356.22 (not stable)
TCP_RR send-1 23617.74 23601.60
...
Notice: the TCP_RR mode is unstable as the official document explains.
I tested many times with different parameters combined through netperf.
Though the result is not that accurate, I cannot see much influence on
this patch. The static key is places on the hot path, but it actually
shouldn't cause a huge regression theoretically.
Co-developed-by: Shujin Li <lishujin@kuaishou.com>
Signed-off-by: Shujin Li <lishujin@kuaishou.com>
Signed-off-by: Jason Xing <xingwanli@kuaishou.com>
Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In ixgbe_xsk_pool_enable(), if ixgbe_xsk_wakeup() fails,
We should restore the previous state and clean up the
resources. Add the missing clear af_xdp_zc_qps and unmap dma
to fix this bug.
Fixes: d49e286d35 ("ixgbe: add tracking of AF_XDP zero-copy state for each queue pair")
Fixes: 4a9b32f30f ("ixgbe: fix potential RX buffer starvation for AF_XDP")
Signed-off-by: Wang Hai <wanghai38@huawei.com>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://lore.kernel.org/r/20210817203736.3529939-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The Intel drivers all have rcu_read_lock()/rcu_read_unlock() pairs around
XDP program invocations. However, the actual lifetime of the objects
referred by the XDP program invocation is longer, all the way through to
the call to xdp_do_flush(), making the scope of the rcu_read_lock() too
small. This turns out to be harmless because it all happens in a single
NAPI poll cycle (and thus under local_bh_disable()), but it makes the
rcu_read_lock() misleading.
Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> # i40e
Cc: Jesse Brandeburg <jesse.brandeburg@intel.com>
Cc: Tony Nguyen <anthony.l.nguyen@intel.com>
Cc: intel-wired-lan@lists.osuosl.org
Link: https://lore.kernel.org/bpf/20210624160609.292325-12-toke@redhat.com
Add missing exception tracing to XDP when a number of different
errors can occur. The support was only partial. Several errors
where not logged which would confuse the user quite a lot not
knowing where and why the packets disappeared.
Fixes: 33fdc82f08 ("ixgbe: add support for XDP_TX action")
Fixes: d0bcacd0a1 ("ixgbe: add AF_XDP zero-copy Rx support")
Reported-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Tested-by: Vishakha Jambekar <vishakha.jambekar@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Optimize ixgbe_run_xdp_zc() for the XDP program verdict being
XDP_REDIRECT in the xsk zero-copy path. This path is only used when
having AF_XDP zero-copy on and in that case most packets will be
directed to user space. This provides a little under 100k extra
packets in throughput on my server when running l2fwd in xdpsock.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Tested-by: Vishakha Jambekar <vishakha.jambekar@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Test for dma_need_sync earlier to increase
performance. xsk_buff_dma_sync_for_cpu() takes an xdp_buff as
parameter and from that the xsk_buff_pool reference is dug out. Perf
shows that this dereference causes a lot of cache misses. But as the
buffer pool is now sent down to the driver at zero-copy initialization
time, we might as well use this pointer directly, instead of going via
the xsk_buff and we can do so already in xsk_buff_dma_sync_for_cpu()
instead of in xp_dma_sync_for_cpu. This gets rid of these cache
misses.
Throughput increases with 3% for the xdpsock l2fwd sample application
on my machine.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/1598603189-32145-11-git-send-email-magnus.karlsson@intel.com
Rename the AF_XDP zero-copy driver interface functions to better
reflect what they do after the replacement of umems with buffer
pools in the previous commit. Mostly it is about replacing the
umem name from the function names with xsk_buff and also have
them take the a buffer pool pointer instead of a umem. The
various ring functions have also been renamed in the process so
that they have the same naming convention as the internal
functions in xsk_queue.h. This so that it will be clearer what
they do and also for consistency.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/1598603189-32145-3-git-send-email-magnus.karlsson@intel.com
Replace the explicit umem reference passed to the driver in AF_XDP
zero-copy mode with the buffer pool instead. This in preparation for
extending the functionality of the zero-copy mode so that umems can be
shared between queues on the same netdev and also between netdevs. In
this commit, only an umem reference has been added to the buffer pool
struct. But later commits will add other entities to it. These are
going to be entities that are different between different queue ids
and netdevs even though the umem is shared between them.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/1598603189-32145-2-git-send-email-magnus.karlsson@intel.com
Convert all the remaining 'fall through" code comments to the newer
'fallthrough;' keyword.
Suggested-by: Joe Perches <joe@perches.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
In order to use standard 'xdp' prefix, rename convert_to_xdp_frame
utility routine in xdp_convert_buff_to_frame and replace all the
occurrences
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Link: https://lore.kernel.org/bpf/6344f739be0d1a08ab2b9607584c4d5478c8c083.1590698295.git.lorenzo@kernel.org
Move the AF_XDP zero-copy driver interface to its own include file
called xdp_sock_drv.h. This, hopefully, will make it more clear for
NIC driver implementors to know what functions to use for zero-copy
support.
v4->v5: Fix -Wmissing-prototypes by include header file. (Jakub)
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200520192103.355233-4-bjorn.topel@gmail.com
Intel drivers implement native AF_XDP zerocopy in separate C-files,
that have its own invocation of bpf_prog_run_xdp(). The setup of
xdp_buff is also handled in separately from normal code path.
This patch update XDP frame_sz for AF_XDP zerocopy drivers i40e, ice
and ixgbe, as the code changes needed are very similar. Introduce a
helper function xsk_umem_xdp_frame_sz() for calculating frame size.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Cc: intel-wired-lan@lists.osuosl.org
Cc: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/158945347511.97035.8536753731329475655.stgit@firesoul
Daniel Borkmann says:
====================
pull-request: bpf-next 2019-12-27
The following pull-request contains BPF updates for your *net-next* tree.
We've added 127 non-merge commits during the last 17 day(s) which contain
a total of 110 files changed, 6901 insertions(+), 2721 deletions(-).
There are three merge conflicts. Conflicts and resolution looks as follows:
1) Merge conflict in net/bpf/test_run.c:
There was a tree-wide cleanup c593642c8b ("treewide: Use sizeof_field() macro")
which gets in the way with b590cb5f80 ("bpf: Switch to offsetofend in
BPF_PROG_TEST_RUN"):
<<<<<<< HEAD
if (!range_is_zero(__skb, offsetof(struct __sk_buff, priority) +
sizeof_field(struct __sk_buff, priority),
=======
if (!range_is_zero(__skb, offsetofend(struct __sk_buff, priority),
>>>>>>> 7c8dce4b16
There are a few occasions that look similar to this. Always take the chunk with
offsetofend(). Note that there is one where the fields differ in here:
<<<<<<< HEAD
if (!range_is_zero(__skb, offsetof(struct __sk_buff, tstamp) +
sizeof_field(struct __sk_buff, tstamp),
=======
if (!range_is_zero(__skb, offsetofend(struct __sk_buff, gso_segs),
>>>>>>> 7c8dce4b16
Just take the one with offsetofend() /and/ gso_segs. Latter is correct due to
850a88cc40 ("bpf: Expose __sk_buff wire_len/gso_segs to BPF_PROG_TEST_RUN").
2) Merge conflict in arch/riscv/net/bpf_jit_comp.c:
(I'm keeping Bjorn in Cc here for a double-check in case I got it wrong.)
<<<<<<< HEAD
if (is_13b_check(off, insn))
return -1;
emit(rv_blt(tcc, RV_REG_ZERO, off >> 1), ctx);
=======
emit_branch(BPF_JSLT, RV_REG_T1, RV_REG_ZERO, off, ctx);
>>>>>>> 7c8dce4b16
Result should look like:
emit_branch(BPF_JSLT, tcc, RV_REG_ZERO, off, ctx);
3) Merge conflict in arch/riscv/include/asm/pgtable.h:
<<<<<<< HEAD
=======
#define VMALLOC_SIZE (KERN_VIRT_SIZE >> 1)
#define VMALLOC_END (PAGE_OFFSET - 1)
#define VMALLOC_START (PAGE_OFFSET - VMALLOC_SIZE)
#define BPF_JIT_REGION_SIZE (SZ_128M)
#define BPF_JIT_REGION_START (PAGE_OFFSET - BPF_JIT_REGION_SIZE)
#define BPF_JIT_REGION_END (VMALLOC_END)
/*
* Roughly size the vmemmap space to be large enough to fit enough
* struct pages to map half the virtual address space. Then
* position vmemmap directly below the VMALLOC region.
*/
#define VMEMMAP_SHIFT \
(CONFIG_VA_BITS - PAGE_SHIFT - 1 + STRUCT_PAGE_MAX_SHIFT)
#define VMEMMAP_SIZE BIT(VMEMMAP_SHIFT)
#define VMEMMAP_END (VMALLOC_START - 1)
#define VMEMMAP_START (VMALLOC_START - VMEMMAP_SIZE)
#define vmemmap ((struct page *)VMEMMAP_START)
>>>>>>> 7c8dce4b16
Only take the BPF_* defines from there and move them higher up in the
same file. Remove the rest from the chunk. The VMALLOC_* etc defines
got moved via 01f52e16b8 ("riscv: define vmemmap before pfn_to_page
calls"). Result:
[...]
#define __S101 PAGE_READ_EXEC
#define __S110 PAGE_SHARED_EXEC
#define __S111 PAGE_SHARED_EXEC
#define VMALLOC_SIZE (KERN_VIRT_SIZE >> 1)
#define VMALLOC_END (PAGE_OFFSET - 1)
#define VMALLOC_START (PAGE_OFFSET - VMALLOC_SIZE)
#define BPF_JIT_REGION_SIZE (SZ_128M)
#define BPF_JIT_REGION_START (PAGE_OFFSET - BPF_JIT_REGION_SIZE)
#define BPF_JIT_REGION_END (VMALLOC_END)
/*
* Roughly size the vmemmap space to be large enough to fit enough
* struct pages to map half the virtual address space. Then
* position vmemmap directly below the VMALLOC region.
*/
#define VMEMMAP_SHIFT \
(CONFIG_VA_BITS - PAGE_SHIFT - 1 + STRUCT_PAGE_MAX_SHIFT)
#define VMEMMAP_SIZE BIT(VMEMMAP_SHIFT)
#define VMEMMAP_END (VMALLOC_START - 1)
#define VMEMMAP_START (VMALLOC_START - VMEMMAP_SIZE)
[...]
Let me know if there are any other issues.
Anyway, the main changes are:
1) Extend bpftool to produce a struct (aka "skeleton") tailored and specific
to a provided BPF object file. This provides an alternative, simplified API
compared to standard libbpf interaction. Also, add libbpf extern variable
resolution for .kconfig section to import Kconfig data, from Andrii Nakryiko.
2) Add BPF dispatcher for XDP which is a mechanism to avoid indirect calls by
generating a branch funnel as discussed back in bpfconf'19 at LSF/MM. Also,
add various BPF riscv JIT improvements, from Björn Töpel.
3) Extend bpftool to allow matching BPF programs and maps by name,
from Paul Chaignon.
4) Support for replacing cgroup BPF programs attached with BPF_F_ALLOW_MULTI
flag for allowing updates without service interruption, from Andrey Ignatov.
5) Cleanup and simplification of ring access functions for AF_XDP with a
bonus of 0-5% performance improvement, from Magnus Karlsson.
6) Enable BPF JITs for x86-64 and arm64 by default. Also, final version of
audit support for BPF, from Daniel Borkmann and latter with Jiri Olsa.
7) Move and extend test_select_reuseport into BPF program tests under
BPF selftests, from Jakub Sitnicki.
8) Various BPF sample improvements for xdpsock for customizing parameters
to set up and benchmark AF_XDP, from Jay Jayatheerthan.
9) Improve libbpf to provide a ulimit hint on permission denied errors.
Also change XDP sample programs to attach in driver mode by default,
from Toke Høiland-Jørgensen.
10) Extend BPF test infrastructure to allow changing skb mark from tc BPF
programs, from Nikita V. Shirokov.
11) Optimize prologue code sequence in BPF arm32 JIT, from Russell King.
12) Fix xdp_redirect_cpu BPF sample to manually attach to tracepoints after
libbpf conversion, from Jesper Dangaard Brouer.
13) Minor misc improvements from various others.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Change the name of xsk_umem_discard_addr to xsk_umem_release_addr to
better reflect the new naming of the AF_XDP queue manipulation
functions. As this functions is used by drivers implementing support
for AF_XDP zero-copy, it requires a name change to these drivers. The
function xsk_umem_release_addr_rq has also changed name in the same
fashion.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-10-git-send-email-magnus.karlsson@intel.com
Use synchronize_rcu to wait until the XSK wakeup function finishes
before destroying the resources it uses:
1. ixgbe_down already calls synchronize_rcu after setting __IXGBE_DOWN.
2. After switching the XDP program, call synchronize_rcu to let
ixgbe_xsk_wakeup exit before the XDP program is freed.
3. Changing the number of channels brings the interface down.
4. Disabling UMEM sets __IXGBE_TX_DISABLED before closing hardware
resources and resetting xsk_umem. Check that bit in ixgbe_xsk_wakeup to
avoid using the XDP ring when it's already destroyed. synchronize_rcu is
called from ixgbe_txrx_ring_disable.
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20191217162023.16011-5-maximmi@mellanox.com
The need_wakeup flag for Tx might not be set for AF_XDP sockets that
are only used to send packets. This happens if there is at least one
outstanding packet that has not been completed by the hardware and we
get that corresponding completion (which will not generate an
interrupt since interrupts are disabled in the napi poll loop) between
the time we stopped processing the Tx completions and interrupts are
enabled again. In this case, the need_wakeup flag will have been
cleared at the end of the Tx completion processing as we believe we
will get an interrupt from the outstanding completion at a later point
in time. But if this completion interrupt occurs before interrupts
are enable, we lose it and should at that point really have set the
need_wakeup flag since there are no more outstanding completions that
can generate an interrupt to continue the processing. When this
happens, user space will see a Tx queue need_wakeup of 0 and skip
issuing a syscall, which means will never get into the Tx processing
again and we have a deadlock.
This patch introduces a quick fix for this issue by just setting the
need_wakeup flag for Tx to 1 all the time. I am working on a proper
fix for this that will toggle the flag appropriately, but it is more
challenging than I anticipated and I am afraid that this patch will
not be completed before the merge window closes, therefore this easier
fix for now. This fix has a negative performance impact in the range
of 0% to 4%. Towards the higher end of the scale if you have driver
and application on the same core and issue a lot of packets, and
towards no negative impact if you use two cores, lower transmission
speeds and/or a workload that also receives packets.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Daniel Borkmann says:
====================
pull-request: bpf-next 2019-09-16
The following pull-request contains BPF updates for your *net-next* tree.
The main changes are:
1) Now that initial BPF backend for gcc has been merged upstream, enable
BPF kselftest suite for bpf-gcc. Also fix a BE issue with access to
bpf_sysctl.file_pos, from Ilya.
2) Follow-up fix for link-vmlinux.sh to remove bash-specific extensions
related to recent work on exposing BTF info through sysfs, from Andrii.
3) AF_XDP zero copy fixes for i40e and ixgbe driver which caused umem
headroom to be added twice, from Ciara.
4) Refactoring work to convert sock opt tests into test_progs framework
in BPF kselftests, from Stanislav.
5) Fix a general protection fault in dev_map_hash_update_elem(), from Toke.
6) Cleanup to use BPF_PROG_RUN() macro in KCM, from Sami.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 7cbbf9f1fa ("ixgbe: fix xdp handle calculations") reintroduced
the addition of the umem headroom to the xdp handle in the ixgbe_zca_free,
ixgbe_alloc_buffer_slow_zc and ixgbe_alloc_buffer_zc functions. However,
the headroom is already added to the handle in the function
ixgbe_run_xdp_zc. This commit removes the latter addition and fixes the
case where the headroom is non-zero.
Fixes: 7cbbf9f1fa ("ixgbe: fix xdp handle calculations")
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tx code doesn't clear the descriptors' status after cleaning.
So, if the budget is larger than number of used elems in a ring, some
descriptors will be accounted twice and xsk_umem_complete_tx will move
prod_tail far beyond the prod_head breaking the completion queue ring.
Fix that by limiting the number of descriptors to clean by the number
of used descriptors in the Tx ring.
'ixgbe_clean_xdp_tx_irq()' function refactored to look more like
'ixgbe_xsk_clean_tx_ring()' since we're allowed to directly use
'next_to_clean' and 'next_to_use' indexes.
CC: stable@vger.kernel.org
Fixes: 8221c5eba8 ("ixgbe: add AF_XDP zero-copy Tx support")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Tested-by: William Tu <u9012063@gmail.com>
Tested-by: Eelco Chaudron <echaudro@redhat.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Currently, we don't add headroom to the handle in ixgbe_zca_free,
ixgbe_alloc_buffer_slow_zc and ixgbe_alloc_buffer_zc. The addition of the
headroom to the handle was removed in
commit d8c3061e5e ("ixgbe: modify driver for handling offsets"), which
will break things when headroom isvnon-zero. This patch fixes this and uses
xsk_umem_adjust_offset to add it appropritely based on the mode being run.
Fixes: d8c3061e5e ("ixgbe: modify driver for handling offsets")
Reported-by: Bjorn Topel <bjorn.topel@intel.com>
Signed-off-by: Kevin Laatz <kevin.laatz@intel.com>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
With the addition of the unaligned chunks option, we need to make sure we
handle the offsets accordingly based on the mode we are currently running
in. This patch modifies the driver to appropriately mask the address for
each case.
Signed-off-by: Kevin Laatz <kevin.laatz@intel.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Currently, the dma, addr and handle are modified when we reuse Rx buffers
in zero-copy mode. However, this is not required as the inputs to the
function are copies, not the original values themselves. As we use the
copies within the function, we can use the original 'obi' values
directly without having to mask and add the headroom.
Signed-off-by: Kevin Laatz <kevin.laatz@intel.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
This patch adds support for the need_wakeup feature of AF_XDP. If the
application has told the kernel that it might sleep using the new bind
flag XDP_USE_NEED_WAKEUP, the driver will then set this flag if it has
no more buffers on the NIC Rx ring and yield to the application. For
Tx, it will set the flag if it has no outstanding Tx completion
interrupts and return to the application.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
This commit replaces ndo_xsk_async_xmit with ndo_xsk_wakeup. This new
ndo provides the same functionality as before but with the addition of
a new flags field that is used to specifiy if Rx, Tx or both should be
woken up. The previous ndo only woke up Tx, as implied by the
name. The i40e and ixgbe drivers (which are all the supported ones)
are updated with this new interface.
This new ndo will be used by the new need_wakeup functionality of XDP
sockets that need to be able to wake up both Rx and Tx driver
processing.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Some drivers want to access the data transmitted in order to implement
acceleration features of the NICs. It is also useful in AF_XDP TX flow.
Change the xsk_umem_consume_tx API to return the whole xdp_desc, that
contains the data pointer, length and DMA address, instead of only the
latter two. Adapt the implementation of i40e and ixgbe to this change.
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: Björn Töpel <bjorn.topel@intel.com>
Cc: Magnus Karlsson <magnus.karlsson@intel.com>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
The total_packets count at ixgbe_clean_xdp_tx_irq is
always zero when testing with xdpsock -t -N. Set the gso_segs
to 1 to make the tx packet count correct.
Signed-off-by: William Tu <u9012063@gmail.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The tx bytecount is done twice. When running
'./xdpsock -t -N -i eth3' and 'ip -s link show dev eth3'
The avg packet size is 120 instead of 60. So remove the
extra one.
Signed-off-by: William Tu <u9012063@gmail.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
As current implementation of netdev already contains and provides
umems for us, we no longer have the need to contain these
structures in ixgbe_adapter.
Refactor the code to operate on netdev-provided umems.
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Here, we add a bitmap to the ixgbe_adapter that tracks if a
certain queue pair has been "zero-copy enabled" via the ndo_bpf.
The bitmap is used in ixgbe_xsk_umem, and enables zero-copy if
and only if XDP is enabled, the corresponding qid in the bitmap
is set, and the umem is non-NULL;
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Three conflicts, one of which, for marvell10g.c is non-trivial and
requires some follow-up from Heiner or someone else.
The issue is that Heiner converted the marvell10g driver over to
use the generic c45 code as much as possible.
However, in 'net' a bug fix appeared which makes sure that a new
local mask (MDIO_AN_10GBT_CTRL_ADV_NBT_MASK) with value 0x01e0
is cleared.
Signed-off-by: David S. Miller <davem@davemloft.net>
An issue has been found while testing zero-copy XDP that
causes a reset to be triggered. As it takes some time to
turn the carrier on after setting zc, and we already
start trying to transmit some packets, watchdog considers
this as an erroneous state and triggers a reset.
Don't do any work if netif carrier is not OK.
Fixes: 8221c5eba8 (ixgbe: add AF_XDP zero-copy Tx support)
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
When the RX rings are created they are also populated with buffers so
that packets can be received. Usually these are kernel buffers, but
for AF_XDP in zero-copy mode, these are user-space buffers and in this
case the application might not have sent down any buffers to the
driver at this point. And if no buffers are allocated at ring creation
time, no packets can be received and no interrupts will be generated so
the NAPI poll function that allocates buffers to the rings will never
get executed.
To rectify this, we kick the NAPI context of any queue with an
attached AF_XDP zero-copy socket in two places in the code. Once after
an XDP program has loaded and once after the umem is registered. This
take care of both cases: XDP program gets loaded first then AF_XDP
socket is created, and the reverse, AF_XDP socket is created first,
then XDP program is loaded.
Fixes: d0bcacd0a1 ("ixgbe: add AF_XDP zero-copy Rx support")
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Commit c9b47cc1fa ("xsk: fix bug when trying to use both copy and
zero-copy on one queue id") moved the umem query code to the AF_XDP
core, and therefore removed the need to query the netdevice for a
umem.
This patch removes XDP_QUERY_XSK_UMEM and all code that implement that
behavior, which is just dead code.
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
This patch adds zero-copy Tx support for AF_XDP sockets. It implements
the ndo_xsk_async_xmit netdev ndo and performs all the Tx logic from a
NAPI context. This means pulling egress packets from the Tx ring,
placing the frames on the NIC HW descriptor ring and completing sent
frames back to the application via the completion ring.
The regular XDP Tx ring is used for AF_XDP as well. This rationale for
this is as follows: XDP_REDIRECT guarantees mutual exclusion between
different NAPI contexts based on CPU id. In other words, a netdev can
XDP_REDIRECT to another netdev with a different NAPI context, since
the operation is bound to a specific core and each core has its own
hardware ring.
As the AF_XDP Tx action is running in the same NAPI context and using
the same ring, it will also be protected from XDP_REDIRECT actions
with the exact same mechanism.
As with AF_XDP Rx, all AF_XDP Tx specific functions are added to
ixgbe_xsk.c.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Tested-by: William Tu <u9012063@gmail.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch adds zero-copy Rx support for AF_XDP sockets. Instead of
allocating buffers of type MEM_TYPE_PAGE_SHARED, the Rx frames are
allocated as MEM_TYPE_ZERO_COPY when AF_XDP is enabled for a certain
queue.
All AF_XDP specific functions are added to a new file, ixgbe_xsk.c.
Note that when AF_XDP zero-copy is enabled, the XDP action XDP_PASS
will allocate a new buffer and copy the zero-copy frame prior passing
it to the kernel stack.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Tested-by: William Tu <u9012063@gmail.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
