This change actually defines the (initial) metadata layout
that should be used by AF_XDP userspace (xsk_tx_metadata).
The first field is flags which requests appropriate offloads,
followed by the offload-specific fields. The supported per-device
offloads are exported via netlink (new xsk-flags).
The offloads themselves are still implemented in a bit of a
framework-y fashion that's left from my initial kfunc attempt.
I'm introducing new xsk_tx_metadata_ops which drivers are
supposed to implement. The drivers are also supposed
to call xsk_tx_metadata_request/xsk_tx_metadata_complete in
the right places. Since xsk_tx_metadata_{request,_complete}
are static inline, we don't incur any extra overhead doing
indirect calls.
The benefit of this scheme is as follows:
- keeps all metadata layout parsing away from driver code
- makes it easy to grep and see which drivers implement what
- don't need any extra flags to maintain to keep track of what
offloads are implemented; if the callback is implemented - the offload
is supported (used by netlink reporting code)
Two offloads are defined right now:
1. XDP_TXMD_FLAGS_CHECKSUM: skb-style csum_start+csum_offset
2. XDP_TXMD_FLAGS_TIMESTAMP: writes TX timestamp back into metadata
area upon completion (tx_timestamp field)
XDP_TXMD_FLAGS_TIMESTAMP is also implemented for XDP_COPY mode: it writes
SW timestamp from the skb destructor (note I'm reusing hwtstamps to pass
metadata pointer).
The struct is forward-compatible and can be extended in the future
by appending more fields.
Reviewed-by: Song Yoong Siang <yoong.siang.song@intel.com>
Signed-off-by: Stanislav Fomichev <sdf@google.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20231127190319.1190813-3-sdf@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
For zerocopy mode, tx_desc->addr can point to an arbitrary offset
and carry some TX metadata in the headroom. For copy mode, there
is no way currently to populate skb metadata.
Introduce new tx_metadata_len umem config option that indicates how many
bytes to treat as metadata. Metadata bytes come prior to tx_desc address
(same as in RX case).
The size of the metadata has mostly the same constraints as XDP:
- less than 256 bytes
- 8-byte aligned (compared to 4-byte alignment on xdp, due to 8-byte
timestamp in the completion)
- non-zero
This data is not interpreted in any way right now.
Reviewed-by: Song Yoong Siang <yoong.siang.song@intel.com>
Signed-off-by: Stanislav Fomichev <sdf@google.com>
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20231127190319.1190813-2-sdf@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Modify xskq_cons_read_desc_batch() in a way that each processed
descriptor will be checked if it is an EOP one or not and act
accordingly to that.
Change the behavior of mentioned function to break the processing when
stumbling upon invalid descriptor instead of skipping it. Furthermore,
let us give only full packets down to ZC driver.
With these two assumptions ZC drivers will not have to take care of an
intermediate state of incomplete frames, which will simplify its
implementations a lot.
Last but not least, stop processing when count of frags would exceed
max supported segments on underlying device.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-15-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Descriptors with zero length are not supported by many NICs. To preserve
uniform behavior discard any zero length desc as invvalid desc.
Signed-off-by: Tirthendu Sarkar <tirthendu.sarkar@intel.com>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-10-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
For transmitting an AF_XDP packet, allocate skb while processing the
first desc and copy data to it. The 'XDP_PKT_CONTD' flag in 'options'
field of the desc indicates the EOP status of the packet. If the current
desc is not EOP, store the skb, release the current desc and go
on to read the next descs.
Allocate a page for each subsequent desc, copy data to it and add it as
a frag in the skb stored in xsk. On processing EOP, transmit the skb
with frags. Addresses contained in descs have been already queued in
consumer queue and skb destructor updated the completion count.
On transmit failure cancel the releases, clear the descs from the
completion queue and consume the skb for retrying packet transmission.
For any invalid descriptor (invalid length/address/options) in the middle
of a packet, all pending descriptors will be dropped by xsk core along
with the invalid one and the next descriptor is treated as the start of
a new packet.
Maximum supported frames for a packet is MAX_SKB_FRAGS + 1. If it is
exceeded, all descriptors accumulated so far are dropped.
Signed-off-by: Tirthendu Sarkar <tirthendu.sarkar@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-9-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
In Tx path, xsk core reserves space for each desc to be transmitted in
the completion queue and it's address contained in it is stored in the
skb destructor arg. After successful transmission the skb destructor
submits the addr marking completion.
To handle multiple descriptors per packet, now along with reserving
space for each descriptor, the corresponding address is also stored in
completion queue. The number of pending descriptors are stored in skb
destructor arg and is used by the skb destructor to update completions.
Introduce 'skb' in xdp_sock to store a partially built packet when
__xsk_generic_xmit() must return before it sees the EOP descriptor for
the current packet so that packet building can resume in next call of
__xsk_generic_xmit().
Helper functions are introduced to set and get the pending descriptors
in the skb destructor arg. Also, wrappers are introduced for storing
descriptor addresses, submitting and cancelling (for unsuccessful
transmissions) the number of completions.
Signed-off-by: Tirthendu Sarkar <tirthendu.sarkar@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-7-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Use the 'options' field in xdp_desc as a packet continuity marker. Since
'options' field was unused till now and was expected to be set to 0, the
'eop' descriptor will have it set to 0, while the non-eop descriptors
will have to set it to 1. This ensures legacy applications continue to
work without needing any change for single-buffer packets.
Add helper functions and extend xskq_prod_reserve_desc() to use the
'options' field.
Signed-off-by: Tirthendu Sarkar <tirthendu.sarkar@intel.com>
Link: https://lore.kernel.org/r/20230719132421.584801-2-maciej.fijalkowski@intel.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Remove redundant (base_addr >= pool->addrs_cnt) comparison from the
conditional.
In particular, addr is computed as:
addr = base_addr + offset
... where base_addr and offset are stored as 48-bit and 16-bit unsigned
integers, respectively. The above sum cannot overflow u64 since base_addr
has a maximum value of 0x0000ffffffffffff and offset has a maximum value
of 0xffff (implying a maximum sum of 0x000100000000fffe). Since overflow
is impossible, it follows that addr >= base_addr.
Now if (base_addr >= pool->addrs_cnt), then clearly:
addr >= base_addr
>= pool->addrs_cnt
Thus, (base_addr >= pool->addrs_cnt) implies (addr >= pool->addrs_cnt).
Subsequently, the former comparison is unnecessary in the conditional
since for any boolean expressions A and B, (A || B) && (A -> B) is
equivalent to B.
Signed-off-by: Kal Conley <kal.conley@dectris.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20230411130025.19704-1-kal.conley@dectris.com
Perform the chunk boundary check like the page boundary check in
xp_desc_crosses_non_contig_pg(). This simplifies the implementation and
reduces the number of branches.
Signed-off-by: Kal Conley <kal.conley@dectris.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20230410121841.643254-1-kal.conley@dectris.com
Make sure unaligned descriptors that straddle the end of the UMEM are
considered invalid. Currently, descriptor validation is broken for
zero-copy mode which only checks descriptors at page granularity.
For example, descriptors in zero-copy mode that overrun the end of the
UMEM but not a page boundary are (incorrectly) considered valid. The
UMEM boundary check needs to happen before the page boundary and
contiguity checks in xp_desc_crosses_non_contig_pg(). Do this check in
xp_unaligned_validate_desc() instead like xp_check_unaligned() already
does.
Fixes: 2b43470add ("xsk: Introduce AF_XDP buffer allocation API")
Signed-off-by: Kal Conley <kal.conley@dectris.com>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/r/20230405235920.7305-2-kal.conley@dectris.com
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
When we try to start AF_XDP on some machines with long running time, due
to the machine's memory fragmentation problem, there is no sufficient
contiguous physical memory that will cause the start failure.
If the size of the queue is 8 * 1024, then the size of the desc[] is
8 * 1024 * 8 = 16 * PAGE, but we also add struct xdp_ring size, so it is
16page+. This is necessary to apply for a 4-order memory. If there are a
lot of queues, it is difficult to these machine with long running time.
Here, that we actually waste 15 pages. 4-Order memory is 32 pages, but
we only use 17 pages.
This patch replaces __get_free_pages() by vmalloc() to allocate memory
to solve these problems.
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Reviewed-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit d678cbd2f8 ("xsk: Fix handling of invalid descriptors in XSK TX
batching API") fixed batch API usage against set of descriptors with
invalid ones but introduced a problem when AF_XDP SW rings are smaller
than HW ones. Mismatch of reported Tx'ed frames between HW generator and
user space app was observed. It turned out that backpressure mechanism
became a bottleneck when the amount of produced descriptors to CQ is
lower than what we grabbed from XSK Tx ring.
Say that 512 entries had been taken from XSK Tx ring but we had only 490
free entries in CQ. Then callsite (ZC driver) will produce only 490
entries onto HW Tx ring but 512 entries will be released from Tx ring
and this is what will be seen by the user space.
In order to fix this case, mix XSK Tx/CQ ring interractions by moving
around internal functions and changing call order:
* pull out xskq_prod_nb_free() from xskq_prod_reserve_addr_batch()
up to xsk_tx_peek_release_desc_batch();
** move xskq_cons_release_n() into xskq_cons_read_desc_batch()
After doing so, algorithm can be described as follows:
1. lookup Tx entries
2. use value from 1. to reserve space in CQ (*)
3. Read from Tx ring as much descriptors as value from 2
3a. release descriptors from XSK Tx ring (**)
4. Finally produce addresses to CQ
Fixes: d678cbd2f8 ("xsk: Fix handling of invalid descriptors in XSK TX batching API")
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20220830121705.8618-1-maciej.fijalkowski@intel.com
xdpxceiver run on a AF_XDP ZC enabled driver revealed a problem with XSK
Tx batching API. There is a test that checks how invalid Tx descriptors
are handled by AF_XDP. Each valid descriptor is followed by invalid one
on Tx side whereas the Rx side expects only to receive a set of valid
descriptors.
In current xsk_tx_peek_release_desc_batch() function, the amount of
available descriptors is hidden inside xskq_cons_peek_desc_batch(). This
can be problematic in cases where invalid descriptors are present due to
the fact that xskq_cons_peek_desc_batch() returns only a count of valid
descriptors. This means that it is impossible to properly update XSK
ring state when calling xskq_cons_release_n().
To address this issue, pull out the contents of
xskq_cons_peek_desc_batch() so that callers (currently only
xsk_tx_peek_release_desc_batch()) will always be able to update the
state of ring properly, as total count of entries is now available and
use this value as an argument in xskq_cons_release_n(). By
doing so, xskq_cons_peek_desc_batch() can be dropped altogether.
Fixes: 9349eb3a9d ("xsk: Introduce batched Tx descriptor interfaces")
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20220607142200.576735-1-maciej.fijalkowski@intel.com
Simplify the mentioned helper function by removing ternary operator. The
expression that is there outputs the boolean value by itself.
This helper might be used in the hot path so this simplification can
also be considered as micro optimization.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20220413153015.453864-15-maciej.fijalkowski@intel.com
The error codes returned by xdp_do_redirect() when redirecting a frame
to an AF_XDP socket has not been very useful. A driver could not
distinguish between different errors. Prior this change the following
codes where used:
Socket not bound or incorrect queue/netdev: EINVAL
XDP frame/AF_XDP buffer size mismatch: ENOSPC
Could not allocate buffer (copy mode): ENOSPC
AF_XDP Rx buffer full: ENOSPC
After this change:
Socket not bound or incorrect queue/netdev: EINVAL
XDP frame/AF_XDP buffer size mismatch: ENOSPC
Could not allocate buffer (copy mode): ENOMEM
AF_XDP Rx buffer full: ENOBUFS
An AF_XDP zero-copy driver can now potentially determine if the
failure was due to a full Rx buffer, and if so stop processing more
frames, yielding to the userland AF_XDP application.
Signed-off-by: Björn Töpel <bjorn@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Link: https://lore.kernel.org/bpf/20220413153015.453864-2-maciej.fijalkowski@intel.com
Move desc_array from the driver to the pool. The reason behind this is
that we can then reuse this array as a temporary storage for descriptors
in all zero-copy drivers that use the batched interface. This will make
it easier to add batching to more drivers.
i40e is the only driver that has a batched Tx zero-copy
implementation, so no need to touch any other driver.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Alexander Lobakin <alexandr.lobakin@intel.com>
Link: https://lore.kernel.org/bpf/20220125160446.78976-6-maciej.fijalkowski@intel.com
Remove two dead stubs, sk_msg_clear_meta() was never
used, use of xskq_cons_is_full() got replaced by
xsk_tx_writeable() in v5.10.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20220126185412.2776254-1-kuba@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Add a new driver interface xsk_buff_alloc_batch() offering batched
buffer allocations to improve performance. The new interface takes
three arguments: the buffer pool to allocated from, a pointer to an
array of struct xdp_buff pointers which will contain pointers to the
allocated xdp_buffs, and an unsigned integer specifying the max number
of buffers to allocate. The return value is the actual number of
buffers that the allocator managed to allocate and it will be in the
range 0 <= N <= max, where max is the third parameter to the function.
u32 xsk_buff_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
u32 max);
A second driver interface is also introduced that need to be used in
conjunction with xsk_buff_alloc_batch(). It is a helper that sets the
size of struct xdp_buff and is used by the NIC Rx irq routine when
receiving a packet. This helper sets the three struct members data,
data_meta, and data_end. The two first ones is in the xsk_buff_alloc()
case set in the allocation routine and data_end is set when a packet
is received in the receive irq function. This unfortunately leads to
worse performance since the xdp_buff is touched twice with a long time
period in between leading to an extra cache miss. Instead, we fill out
the xdp_buff with all 3 fields at one single point in time in the
driver, when the size of the packet is known. Hence this helper. Note
that the driver has to use this helper (or set all three fields
itself) when using xsk_buff_alloc_batch(). xsk_buff_alloc() works as
before and does not require this.
void xsk_buff_set_size(struct xdp_buff *xdp, u32 size);
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210922075613.12186-3-magnus.karlsson@gmail.com
Fix broken Tx ring validation for AF_XDP. The commit under the Fixes
tag, fixed an off-by-one error in the validation but introduced
another error. Descriptors are now let through even if they straddle a
chunk boundary which they are not allowed to do in aligned mode. Worse
is that they are let through even if they straddle the end of the umem
itself, tricking the kernel to read data outside the allowed umem
region which might or might not be mapped at all.
Fix this by reintroducing the old code, but subtract the length by one
to fix the off-by-one error that the original patch was
addressing. The test chunk != chunk_end makes sure packets do not
straddle chunk boundraries. Note that packets of zero length are
allowed in the interface, therefore the test if the length is
non-zero.
Fixes: ac31565c21 ("xsk: Fix for xp_aligned_validate_desc() when len == chunk_size")
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Acked-by: Björn Töpel <bjorn@kernel.org>
Link: https://lore.kernel.org/bpf/20210618075805.14412-1-magnus.karlsson@gmail.com
When desc->len is equal to chunk_size, it is legal. But when the
xp_aligned_validate_desc() got chunk_end from desc->addr + desc->len
pointing to the next chunk during the check, it caused the check to
fail.
This problem was first introduced in bbff2f321a ("xsk: new descriptor
addressing scheme"). Later in 2b43470add ("xsk: Introduce AF_XDP buffer
allocation API") this piece of code was moved into the new function called
xp_aligned_validate_desc(). This function was then moved into xsk_queue.h
via 26062b185e ("xsk: Explicitly inline functions and move definitions").
Fixes: bbff2f321a ("xsk: new descriptor addressing scheme")
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20210428094424.54435-1-xuanzhuo@linux.alibaba.com
Currently, the AF_XDP rings uses general smp_{r,w,}mb() barriers on
the kernel-side. On most modern architectures
load-acquire/store-release barriers perform better, and results in
simpler code for circular ring buffers.
This change updates the XDP socket rings to use
load-acquire/store-release barriers.
It is important to note that changing from the old smp_{r,w,}mb()
barriers, to load-acquire/store-release barriers does not break
compatibility. The old semantics work with the new one, and vice
versa.
As pointed out by "Documentation/memory-barriers.txt" in the "SMP
BARRIER PAIRING" section:
"General barriers pair with each other, though they also pair with
most other types of barriers, albeit without multicopy atomicity.
An acquire barrier pairs with a release barrier, but both may also
pair with other barriers, including of course general barriers."
How different barriers behaves and pairs is outlined in
"tools/memory-model/Documentation/cheatsheet.txt".
In order to make sure that compatibility is not broken, LKMM herd7
based litmus tests can be constructed and verified.
We generalize the XDP socket ring to a one entry ring, and create two
scenarios; One where the ring is full, where only the consumer can
proceed, followed by the producer. One where the ring is empty, where
only the producer can proceed, followed by the consumer. Each scenario
is then expanded to four different tests: general producer/general
consumer, general producer/acqrel consumer, acqrel producer/general
consumer, acqrel producer/acqrel consumer. In total eight tests.
The empty ring test:
C spsc-rb+empty
// Simple one entry ring:
// prod cons allowed action prod cons
// 0 0 => prod => 1 0
// 0 1 => cons => 0 0
// 1 0 => cons => 1 1
// 1 1 => prod => 0 1
{}
// We start at prod==0, cons==0, data==0, i.e. nothing has been
// written to the ring. From here only the producer can start, and
// should write 1. Afterwards, consumer can continue and read 1 to
// data. Can we enter state prod==1, cons==1, but consumer observed
// the incorrect value of 0?
P0(int *prod, int *cons, int *data)
{
... producer
}
P1(int *prod, int *cons, int *data)
{
... consumer
}
exists( 1:d=0 /\ prod=1 /\ cons=1 );
The full ring test:
C spsc-rb+full
// Simple one entry ring:
// prod cons allowed action prod cons
// 0 0 => prod => 1 0
// 0 1 => cons => 0 0
// 1 0 => cons => 1 1
// 1 1 => prod => 0 1
{ prod = 1; }
// We start at prod==1, cons==0, data==1, i.e. producer has
// written 0, so from here only the consumer can start, and should
// consume 0. Afterwards, producer can continue and write 1 to
// data. Can we enter state prod==0, cons==1, but consumer observed
// the write of 1?
P0(int *prod, int *cons, int *data)
{
... producer
}
P1(int *prod, int *cons, int *data)
{
... consumer
}
exists( 1:d=1 /\ prod=0 /\ cons=1 );
where P0 and P1 are:
P0(int *prod, int *cons, int *data)
{
int p;
p = READ_ONCE(*prod);
if (READ_ONCE(*cons) == p) {
WRITE_ONCE(*data, 1);
smp_wmb();
WRITE_ONCE(*prod, p ^ 1);
}
}
P0(int *prod, int *cons, int *data)
{
int p;
p = READ_ONCE(*prod);
if (READ_ONCE(*cons) == p) {
WRITE_ONCE(*data, 1);
smp_store_release(prod, p ^ 1);
}
}
P1(int *prod, int *cons, int *data)
{
int c;
int d = -1;
c = READ_ONCE(*cons);
if (READ_ONCE(*prod) != c) {
smp_rmb();
d = READ_ONCE(*data);
smp_mb();
WRITE_ONCE(*cons, c ^ 1);
}
}
P1(int *prod, int *cons, int *data)
{
int c;
int d = -1;
c = READ_ONCE(*cons);
if (smp_load_acquire(prod) != c) {
d = READ_ONCE(*data);
smp_store_release(cons, c ^ 1);
}
}
The full LKMM litmus tests are found at [1].
On x86-64 systems the l2fwd AF_XDP xdpsock sample performance
increases by 1%. This is mostly due to that the smp_mb() is removed,
which is a relatively expensive operation on these
platforms. Weakly-ordered platforms, such as ARM64 might benefit even
more.
[1] https://github.com/bjoto/litmus-xsk
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/bpf/20210305094113.413544-2-bjorn.topel@gmail.com
Rollback the reservation in the completion ring when we get a
NETDEV_TX_BUSY. When this error is received from the driver, we are
supposed to let the user application retry the transmit again. And in
order to do this, we need to roll back the failed send so it can be
retried. Unfortunately, we did not cancel the reservation we had made
in the completion ring. By not doing this, we actually make the
completion ring one entry smaller per NETDEV_TX_BUSY error we get, and
after enough of these errors the completion ring will be of size zero
and transmit will stop working.
Fix this by cancelling the reservation when we get a NETDEV_TX_BUSY
error.
Fixes: 642e450b6b ("xsk: Do not discard packet when NETDEV_TX_BUSY")
Reported-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
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/20201218134525.13119-3-magnus.karlsson@gmail.com
xdp_return_frame_bulk() needs to pass a xdp_buff
to __xdp_return().
strlcpy got converted to strscpy but here it makes no
functional difference, so just keep the right code.
Conflicts:
net/netfilter/nf_tables_api.c
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Modify the tx writeable condition from the queue is not full to the
number of present tx queues is less than the half of the total number
of queues. Because the tx queue not full is a very short time, this will
cause a large number of EPOLLOUT events, and cause a large number of
process wake up.
Fixes: 35fcde7f8d ("xsk: support for Tx")
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/508fef55188d4e1160747ead64c6dcda36735880.1606555939.git.xuanzhuo@linux.alibaba.com
Introduce batched descriptor interfaces in the xsk core code for the
Tx path to be used in the driver to write a code path with higher
performance. This interface will be used by the i40e driver in the
next patch. Though other drivers would likely benefit from this new
interface too.
Note that batching is only implemented for the common case when
there is only one socket bound to the same device and queue id. When
this is not the case, we fall back to the old non-batched version of
the function.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/1605525167-14450-5-git-send-email-magnus.karlsson@gmail.com
Introduce one cache line worth of padding between the consumer pointer
and the flags field as well as between the flags field and the start
of the descriptors in all the lockless rings. This so that the x86 HW
adjacency prefetcher will not prefetch the adjacent pointer/field when
only one pointer/field is going to be used. This improves throughput
performance for the l2fwd sample app with 1% on my machine with HW
prefetching turned on in the BIOS.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/1605525167-14450-4-git-send-email-magnus.karlsson@gmail.com
Introduce one cache line worth of padding between the producer and
consumer pointers in all the lockless rings. This so that the HW
adjacency prefetcher will not prefetch the consumer pointer when the
producer pointer is used and vice versa. This improves throughput
performance for the l2fwd sample app with 2% on my machine with HW
prefetching turned on.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/1602166338-21378-1-git-send-email-magnus.karlsson@gmail.com
After 'peeking' the ring, the consumer, not the producer, reads the data.
Fix this mistake in the comments.
Fixes: 15d8c9162c ("xsk: Add function naming comments and reorder functions")
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20200928082344.17110-1-ciara.loftus@intel.com
Create and free the buffer pool independently from the umem. Move
these operations that are performed on the buffer pool from the
umem create and destroy functions to new create and destroy
functions just for the buffer pool. This so that in later commits
we can instantiate multiple buffer pools per umem when sharing a
umem between HW queues and/or devices. We also erradicate the
back pointer from the umem to the buffer pool as this will not
work when we introduce the possibility to have multiple buffer
pools per umem.
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/1598603189-32145-4-git-send-email-magnus.karlsson@intel.com
It can be useful for the user to know the reason behind a dropped packet.
Introduce new counters which track drops on the receive path caused by:
1. rx ring being full
2. fill ring being empty
Also, on the tx path introduce a counter which tracks the number of times
we attempt pull from the tx ring when it is empty.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200708072835.4427-2-ciara.loftus@intel.com
In order to reduce the number of function calls, the struct
xsk_buff_pool definition is moved to xsk_buff_pool.h. The functions
xp_get_dma(), xp_dma_sync_for_cpu(), xp_dma_sync_for_device(),
xp_validate_desc() and various helper functions are explicitly
inlined.
Further, move xp_get_handle() and xp_release() to xsk.c, to allow for
the compiler to perform inlining.
rfc->v1: Make sure xp_validate_desc() is inlined for Tx perf. (Maxim)
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200520192103.355233-15-bjorn.topel@gmail.com
There are no users of MEM_TYPE_ZERO_COPY. Remove all corresponding
code, including the "handle" member of struct xdp_buff.
rfc->v1: Fixed spelling in commit message. (Björn)
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200520192103.355233-13-bjorn.topel@gmail.com
In order to simplify AF_XDP zero-copy enablement for NIC driver
developers, a new AF_XDP buffer allocation API is added. The
implementation is based on a single core (single producer/consumer)
buffer pool for the AF_XDP UMEM.
A buffer is allocated using the xsk_buff_alloc() function, and
returned using xsk_buff_free(). If a buffer is disassociated with the
pool, e.g. when a buffer is passed to an AF_XDP socket, a buffer is
said to be released. Currently, the release function is only used by
the AF_XDP internals and not visible to the driver.
Drivers using this API should register the XDP memory model with the
new MEM_TYPE_XSK_BUFF_POOL type.
The API is defined in net/xdp_sock_drv.h.
The buffer type is struct xdp_buff, and follows the lifetime of
regular xdp_buffs, i.e. the lifetime of an xdp_buff is restricted to
a NAPI context. In other words, the API is not replacing xdp_frames.
In addition to introducing the API and implementations, the AF_XDP
core is migrated to use the new APIs.
rfc->v1: Fixed build errors/warnings for m68k and riscv. (kbuild test
robot)
Added headroom/chunk size getter. (Maxim/Björn)
v1->v2: Swapped SoBs. (Maxim)
v2->v3: Initialize struct xdp_buff member frame_sz. (Björn)
Add API to query the DMA address of a frame. (Maxim)
Do DMA sync for CPU till the end of the frame to handle
possible growth (frame_sz). (Maxim)
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200520192103.355233-6-bjorn.topel@gmail.com
Move the XSK_NEXT_PG_CONTIG_{MASK,SHIFT}, and
XDP_UMEM_USES_NEED_WAKEUP defines from xdp_sock.h to the AF_XDP
internal xsk.h file. Also, start using the BIT{,_ULL} macro instead of
explicit shifts.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200520192103.355233-5-bjorn.topel@gmail.com
Change two variables names so that it is clearer what they
represent. The first one is xsk_list that in fact only contains the
list of AF_XDP sockets with a Tx component. Change this to xsk_tx_list
for improved clarity. The second variable is size in the ring
structure. One might think that this is the size of the ring, but it
is in fact the size of the umem, copied into the ring structure to
improve performance. Rename this variable umem_size to avoid any
confusion.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Link: https://lore.kernel.org/bpf/1588599232-24897-2-git-send-email-magnus.karlsson@intel.com
The current codebase makes use of the zero-length array language
extension to the C90 standard, but the preferred mechanism to declare
variable-length types such as these ones is a flexible array member[1][2],
introduced in C99:
struct foo {
int stuff;
struct boo array[];
};
By making use of the mechanism above, we will get a compiler warning
in case the flexible array does not occur last in the structure, which
will help us prevent some kind of undefined behavior bugs from being
inadvertently introduced[3] to the codebase from now on.
Also, notice that, dynamic memory allocations won't be affected by
this change:
"Flexible array members have incomplete type, and so the sizeof operator
may not be applied. As a quirk of the original implementation of
zero-length arrays, sizeof evaluates to zero."[1]
This issue was found with the help of Coccinelle.
[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html
[2] https://github.com/KSPP/linux/issues/21
[3] commit 7649773293 ("cxgb3/l2t: Fix undefined behaviour")
Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit 4b638f13ba ("xsk: Eliminate the RX batch size")
introduced a much more lazy way of updating the global consumer
pointers from the kernel side, by only doing so when running out of
entries in the fill or Tx rings (the rings consumed by the
kernel). This can result in a deadlock with the user application if
the kernel requires more than one entry to proceed and the application
cannot put these entries in the fill ring because the kernel has not
updated the global consumer pointer since the ring is not empty.
Fix this by publishing the local kernel side consumer pointer whenever
we have completed Rx or Tx processing in the kernel. This way, user
space will have an up-to-date view of the consumer pointers whenever it
gets to execute in the one core case (application and driver on the
same core), or after a certain number of packets have been processed
in the two core case (application and driver on different cores).
A side effect of this patch is that the one core case gets better
performance, but the two core case gets worse. The reason that the one
core case improves is that updating the global consumer pointer is
relatively cheap since the application by definition is not running
when the kernel is (they are on the same core) and it is beneficial
for the application, once it gets to run, to have pointers that are
as up to date as possible since it then can operate on more packets
and buffers. In the two core case, the most important performance
aspect is to minimize the number of accesses to the global pointers
since they are shared between two cores and bounces between the caches
of those cores. This patch results in more updates to global state,
which means lower performance in the two core case.
Fixes: 4b638f13ba ("xsk: Eliminate the RX batch size")
Reported-by: Ryan Goodfellow <rgoodfel@isi.edu>
Reported-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Acked-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Link: https://lore.kernel.org/bpf/1581348432-6747-1-git-send-email-magnus.karlsson@intel.com
Add comments on how the ring access functions are named and how they
are supposed to be used for producers and consumers. The functions are
also reordered so that the consumer functions are in the beginning and
the producer functions in the end, for easier reference. Put this in a
separate patch as the diff might look a little odd, but no
functionality has changed in this patch.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-12-git-send-email-magnus.karlsson@intel.com
There are two unnecessary READ_ONCE of descriptor data. These are not
needed since the data is written by the producer before it signals
that the data is available by incrementing the producer pointer. As the
access to this producer pointer is serialized and the consumer always
reads the descriptor after it has read and synchronized with the
producer counter, the write of the descriptor will have fully
completed and it does not matter if the consumer has any read tearing.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-11-git-send-email-magnus.karlsson@intel.com
Change the names of the validation functions to better reflect what
they are doing. The uppermost ones are reading entries from the rings
and only the bottom ones validate entries. So xskq_cons_read_ is a
better prefix name.
Also change the xskq_cons_read_ functions to return a bool
as the the descriptor or address is already returned by reference
in the parameters. Everyone is using the return value as a bool
anyway.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-9-git-send-email-magnus.karlsson@intel.com
Simplify and refactor consumer ring functions. The consumer first
"peeks" to find descriptors or addresses that are available to
read from the ring, then reads them and finally "releases" these
descriptors once it is done. The two local variables cons_tail
and cons_head are turned into one single variable called
cached_cons. cached_tail referred to the cached value of the
global consumer pointer and will be stored in cached_cons. For
cached_head, we just use cached_prod instead as it was not used
for a consumer queue before. It also better reflects what it
really is now: a cached copy of the producer pointer.
The names of the functions are also renamed in the same manner as
the producer functions. The new functions are called xskq_cons_
followed by what it does.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-8-git-send-email-magnus.karlsson@intel.com
At this point, there are no users of the functions xskq_nb_avail and
xskq_nb_free that take any other number of entries argument than 1, so
let us get rid of the second argument that takes the number of
entries.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-7-git-send-email-magnus.karlsson@intel.com
In the xsk consumer ring code there is a variable called RX_BATCH_SIZE
that dictates the minimum number of entries that we try to grab from
the fill and Tx rings. In fact, the code always try to grab the
maximum amount of entries from these rings. The only thing this
variable does is to throw an error if there is less than 16 (as it is
defined) entries on the ring. There is no reason to do this and it
will just lead to weird behavior from user space's point of view. So
eliminate this variable.
With this change, we will be able to simplify the xskq_nb_free and
xskq_nb_avail code in the next commit.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-6-git-send-email-magnus.karlsson@intel.com
Adopt the naming of the producer ring access functions to have a
similar naming convention as the functions in libbpf, but adapted to
the kernel. You first reserve a number of entries that you later
submit to the global state of the ring. This is much clearer, IMO,
than the one that was in the kernel part. Once renamed, we also
discover that two functions are actually the same, so remove one of
them. Some of the primitive ring submission operations are also the
same so break these out into __xskq_prod_submit that the upper level
ring access functions can use.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-5-git-send-email-magnus.karlsson@intel.com
Currently, the xsk ring code has two cached producer pointers:
prod_head and prod_tail. This patch consolidates these two into a
single one called cached_prod to make the code simpler and easier to
maintain. This will be in line with the user space part of the the
code found in libbpf, that only uses a single cached pointer.
The Rx path only uses the two top level functions
xskq_produce_batch_desc and xskq_produce_flush_desc and they both use
prod_head and never prod_tail. So just move them over to
cached_prod.
The Tx XDP_DRV path uses xskq_produce_addr_lazy and
xskq_produce_flush_addr_n and unnecessarily operates on both prod_tail
and prod_head, so move them over to just use cached_prod by skipping
the intermediate step of updating prod_tail.
The Tx path in XDP_SKB mode uses xskq_reserve_addr and
xskq_produce_addr. They currently use both cached pointers, but we can
operate on the global producer pointer in xskq_produce_addr since it
has to be updated anyway, thus eliminating the use of both cached
pointers. We can also remove the xskq_nb_free in xskq_produce_addr
since it is already called in xskq_reserve_addr. No need to do it
twice.
When there is only one cached producer pointer, we can also simplify
xskq_nb_free by removing one argument.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-4-git-send-email-magnus.karlsson@intel.com
In order to set the correct return flags for poll, the xsk code has to
check if the Rx queue is empty and if the Tx queue is full. This code
was unnecessarily large and complex as it used the functions that are
used to update the local state from the global state (xskq_nb_free and
xskq_nb_avail). Since we are not doing this nor updating any data
dependent on this state, we can simplify the functions. Another
benefit from this is that we can also simplify the xskq_nb_free and
xskq_nb_avail functions in a later commit.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-3-git-send-email-magnus.karlsson@intel.com
The lazy update threshold was introduced to keep the producer and
consumer some distance apart in the completion ring. This was
important in the beginning of the development of AF_XDP as the ring
format as that point in time was very sensitive to the producer and
consumer being on the same cache line. This is not the case
anymore as the current ring format does not degrade in any noticeable
way when this happens. Moreover, this threshold makes it impossible
to run the system with rings that have less than 128 entries.
So let us remove this threshold and just get one entry from the ring
as in all other functions. This will enable us to remove this function
in a later commit. Note that xskq_produce_addr_lazy followed by
xskq_produce_flush_addr_n are still not the same function as
xskq_produce_addr() as it operates on another cached pointer.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/1576759171-28550-2-git-send-email-magnus.karlsson@intel.com
Currently, addresses are chunk size aligned. This means, we are very
restricted in terms of where we can place chunk within the umem. For
example, if we have a chunk size of 2k, then our chunks can only be placed
at 0,2k,4k,6k,8k... and so on (ie. every 2k starting from 0).
This patch introduces the ability to use unaligned chunks. With these
changes, we are no longer bound to having to place chunks at a 2k (or
whatever your chunk size is) interval. Since we are no longer dealing with
aligned chunks, they can now cross page boundaries. Checks for page
contiguity have been added in order to keep track of which pages are
followed by a physically contiguous page.
Signed-off-by: Kevin Laatz <kevin.laatz@intel.com>
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
This commit adds support for a new flag called need_wakeup in the
AF_XDP Tx and fill rings. When this flag is set, it means that the
application has to explicitly wake up the kernel Rx (for the bit in
the fill ring) or kernel Tx (for bit in the Tx ring) processing by
issuing a syscall. Poll() can wake up both depending on the flags
submitted and sendto() will wake up tx processing only.
The main reason for introducing this new flag is to be able to
efficiently support the case when application and driver is executing
on the same core. Previously, the driver was just busy-spinning on the
fill ring if it ran out of buffers in the HW and there were none on
the fill ring. This approach works when the application is running on
another core as it can replenish the fill ring while the driver is
busy-spinning. Though, this is a lousy approach if both of them are
running on the same core as the probability of the fill ring getting
more entries when the driver is busy-spinning is zero. With this new
feature the driver now sets the need_wakeup flag and returns to the
application. The application can then replenish the fill queue and
then explicitly wake up the Rx processing in the kernel using the
syscall poll(). For Tx, the flag is only set to one if the driver has
no outstanding Tx completion interrupts. If it has some, the flag is
zero as it will be woken up by a completion interrupt anyway.
As a nice side effect, this new flag also improves the performance of
the case where application and driver are running on two different
cores as it reduces the number of syscalls to the kernel. The kernel
tells user space if it needs to be woken up by a syscall, and this
eliminates many of the syscalls.
This flag needs some simple driver support. If the driver does not
support this, the Rx flag is always zero and the Tx flag is always
one. This makes any application relying on this feature default to the
old behaviour of not requiring any syscalls in the Rx path and always
having to call sendto() in the Tx path.
For backwards compatibility reasons, this feature has to be explicitly
turned on using a new bind flag (XDP_USE_NEED_WAKEUP). I recommend
that you always turn it on as it so far always have had a positive
performance impact.
The name and inspiration of the flag has been taken from io_uring by
Jens Axboe. Details about this feature in io_uring can be found in
http://kernel.dk/io_uring.pdf, section 8.3.
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>
