Age | Commit message (Collapse) | Author | Files | Lines |
|
Changing API ndo_xdp_xmit to take a struct xdp_frame instead of struct
xdp_buff. This brings xdp_return_frame and ndp_xdp_xmit in sync.
This builds towards changing the API further to become a bulk API,
because xdp_buff is not a queue-able object while xdp_frame is.
V4: Adjust for commit 59655a5b6c83 ("tuntap: XDP_TX can use native XDP")
V7: Adjust for commit d9314c474d4f ("i40e: add support for XDP_REDIRECT")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
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 d9314c474d4f ("i40e: add support for XDP_REDIRECT")
V8: Adjust for commit bd658dda4237 ("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>
|
|
This patch shows how it is possible to have both the driver local page
cache, which uses elevated refcnt for "catching"/avoiding SKB
put_page returns the page through the page allocator. And at the
same time, have pages getting returned to the page_pool from
ndp_xdp_xmit DMA completion.
The performance improvement for XDP_REDIRECT in this patch is really
good. Especially considering that (currently) the xdp_return_frame
API and page_pool_put_page() does per frame operations of both
rhashtable ID-lookup and locked return into (page_pool) ptr_ring.
(It is the plan to remove these per frame operation in a followup
patchset).
The benchmark performed was RX on mlx5 and XDP_REDIRECT out ixgbe,
with xdp_redirect_map (using devmap) . And the target/maximum
capability of ixgbe is 13Mpps (on this HW setup).
Before this patch for mlx5, XDP redirected frames were returned via
the page allocator. The single flow performance was 6Mpps, and if I
started two flows the collective performance drop to 4Mpps, because we
hit the page allocator lock (further negative scaling occurs).
Two test scenarios need to be covered, for xdp_return_frame API, which
is DMA-TX completion running on same-CPU or cross-CPU free/return.
Results were same-CPU=10Mpps, and cross-CPU=12Mpps. This is very
close to our 13Mpps max target.
The reason max target isn't reached in cross-CPU test, is likely due
to RX-ring DMA unmap/map overhead (which doesn't occur in ixgbe to
ixgbe testing). It is also planned to remove this unnecessary DMA
unmap in a later patchset
V2: Adjustments requested by Tariq
- Changed page_pool_create return codes not return NULL, only
ERR_PTR, as this simplifies err handling in drivers.
- Save a branch in mlx5e_page_release
- Correct page_pool size calc for MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ
V5: Updated patch desc
V8: Adjust for b0cedc844c00 ("net/mlx5e: Remove rq_headroom field from params")
V9:
- Adjust for 121e89275471 ("net/mlx5e: Refactor RQ XDP_TX indication")
- Adjust for 73281b78a37a ("net/mlx5e: Derive Striding RQ size from MTU")
- Correct handling if page_pool_create fail for MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ
V10: Req from Tariq
- Change pool_size calc for MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
New allocator type MEM_TYPE_PAGE_POOL for page_pool usage.
The registered allocator page_pool pointer is not available directly
from xdp_rxq_info, but it could be (if needed). For now, the driver
should keep separate track of the page_pool pointer, which it should
use for RX-ring page allocation.
As suggested by Saeed, to maintain a symmetric API it is the drivers
responsibility to allocate/create and free/destroy the page_pool.
Thus, after the driver have called xdp_rxq_info_unreg(), it is drivers
responsibility to free the page_pool, but with a RCU free call. This
is done easily via the page_pool helper page_pool_destroy() (which
avoids touching any driver code during the RCU callback, which could
happen after the driver have been unloaded).
V8: address issues found by kbuild test robot
- Address sparse should be static warnings
- Allow xdp.o to be compiled without page_pool.o
V9: Remove inline from .c file, compiler knows best
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Need a fast page recycle mechanism for ndo_xdp_xmit API for returning
pages on DMA-TX completion time, which have good cross CPU
performance, given DMA-TX completion time can happen on a remote CPU.
Refurbish my page_pool code, that was presented[1] at MM-summit 2016.
Adapted page_pool code to not depend the page allocator and
integration into struct page. The DMA mapping feature is kept,
even-though it will not be activated/used in this patchset.
[1] http://people.netfilter.org/hawk/presentations/MM-summit2016/generic_page_pool_mm_summit2016.pdf
V2: Adjustments requested by Tariq
- Changed page_pool_create return codes, don't return NULL, only
ERR_PTR, as this simplifies err handling in drivers.
V4: many small improvements and cleanups
- Add DOC comment section, that can be used by kernel-doc
- Improve fallback mode, to work better with refcnt based recycling
e.g. remove a WARN as pointed out by Tariq
e.g. quicker fallback if ptr_ring is empty.
V5: Fixed SPDX license as pointed out by Alexei
V6: Adjustments requested by Eric Dumazet
- Adjust ____cacheline_aligned_in_smp usage/placement
- Move rcu_head in struct page_pool
- Free pages quicker on destroy, minimize resources delayed an RCU period
- Remove code for forward/backward compat ABI interface
V8: Issues found by kbuild test robot
- Address sparse should be static warnings
- Only compile+link when a driver use/select page_pool,
mlx5 selects CONFIG_PAGE_POOL, although its first used in two patches
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Use the IDA infrastructure for getting a cyclic increasing ID number,
that is used for keeping track of each registered allocator per
RX-queue xdp_rxq_info. Instead of using the IDR infrastructure, which
uses a radix tree, use a dynamic rhashtable, for creating ID to
pointer lookup table, because this is faster.
The problem that is being solved here is that, the xdp_rxq_info
pointer (stored in xdp_buff) cannot be used directly, as the
guaranteed lifetime is too short. The info is needed on a
(potentially) remote CPU during DMA-TX completion time . In an
xdp_frame the xdp_mem_info is stored, when it got converted from an
xdp_buff, which is sufficient for the simple page refcnt based recycle
schemes.
For more advanced allocators there is a need to store a pointer to the
registered allocator. Thus, there is a need to guard the lifetime or
validity of the allocator pointer, which is done through this
rhashtable ID map to pointer. The removal and validity of of the
allocator and helper struct xdp_mem_allocator is guarded by RCU. The
allocator will be created by the driver, and registered with
xdp_rxq_info_reg_mem_model().
It is up-to debate who is responsible for freeing the allocator
pointer or invoking the allocator destructor function. In any case,
this must happen via RCU freeing.
Use the IDA infrastructure for getting a cyclic increasing ID number,
that is used for keeping track of each registered allocator per
RX-queue xdp_rxq_info.
V4: Per req of Jason Wang
- Use xdp_rxq_info_reg_mem_model() in all drivers implementing
XDP_REDIRECT, even-though it's not strictly necessary when
allocator==NULL for type MEM_TYPE_PAGE_SHARED (given it's zero).
V6: Per req of Alex Duyck
- Introduce rhashtable_lookup() call in later patch
V8: Address sparse should be static warnings (from kbuild test robot)
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Now all the users of ndo_xdp_xmit have been converted to use xdp_return_frame.
This enable a different memory model, thus activating another code path
in the xdp_return_frame API.
V2: Fixed issues pointed out by Tariq.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Also convert driver i40e, which very recently got XDP_REDIRECT support
in commit d9314c474d4f ("i40e: add support for XDP_REDIRECT").
V7: This patch got added in V7 of this patchset.
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>
|
|
The virtio_net driver assumes XDP frames are always released based on
page refcnt (via put_page). Thus, is only queues the XDP data pointer
address and uses virt_to_head_page() to retrieve struct page.
Use the XDP return API to get away from such assumptions. Instead
queue an xdp_frame, which allow us to use the xdp_return_frame API,
when releasing the frame.
V8: Avoid endianness issues (found by kbuild test robot)
V9: Change __virtnet_xdp_xmit from bool to int return value (found by Dan Carpenter)
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The tuntap driver invented it's own driver specific way of queuing
XDP packets, by storing the xdp_buff information in the top of
the XDP frame data.
Convert it over to use the more generic xdp_frame structure. The
main problem with the in-driver method is that the xdp_rxq_info pointer
cannot be trused/used when dequeueing the frame.
V3: Remove check based on feedback from Jason
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This is needed to convert drivers tuntap and virtio_net.
This is a generalization of what is done inside cpumap, which will be
converted later.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This is done to prepare for the next patch, and it is also
nice to move this XDP related struct out of filter.h.
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Extend struct ixgbe_tx_buffer to store the xdp_mem_info.
Notice that this could be optimized further by putting this into
a union in the struct ixgbe_tx_buffer, but this patchset
works towards removing this again. Thus, this is not done.
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>
|
|
This implements basic XDP redirect support in mlx5 driver.
Notice that the ndo_xdp_xmit() is NOT implemented, because that API
need some changes that this patchset is working towards.
The main purpose of this patch is have different drivers doing
XDP_REDIRECT to show how different memory models behave in a cross
driver world.
Update(pre-RFCv2 Tariq): Need to DMA unmap page before xdp_do_redirect,
as the return API does not exist yet to to keep this mapped.
Update(pre-RFCv3 Saeed): Don't mix XDP_TX and XDP_REDIRECT flushing,
introduce xdpsq.db.redirect_flush boolian.
V9: Adjust for commit 121e89275471 ("net/mlx5e: Refactor RQ XDP_TX indication")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
1. Added red_drops stats. Inbound packets dropped by RED, buffer exhaustion
2. Included fcs_err, jabber_err, l2_err and frame_err errors under
rx_errors
3. Included fifo_err, dmac_drop, red_drops, fw_err_pko, fw_err_link and
fw_err_drop under rx_dropped
4. Included max_collision_fail, max_deferral_fail, total_collisions,
fw_err_pko, fw_err_link, fw_err_drop and fw_err_pki under tx_dropped
5. Counting dma mapping errors
6. Added some firmware stats description and removed for some
Signed-off-by: Intiyaz Basha <intiyaz.basha@cavium.com>
Acked-by: Derek Chickles <derek.chickles@cavium.com>
Acked-by: Satanand Burla <satananda.burla@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This tracepoint was replaced by inet_sock_set_state in 563e0bb and not
used anywhere in the kernel anymore. Remove it.
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Replace magic number "0x5 << MAX_READ_REQUEST_SHIFT" with the
appropriate constant as defined in PCI core.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch adds missing values for the max read request size.
E.g. network driver r8169 uses a value of 4K.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Acked-by: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Switch stmmac_mode_ops to generic Hardware Interface Helpers instead of
using hard-coded callbacks. This makes the code more readable and more
flexible.
No functional change.
Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Switch stmmac_hwtimestamp to generic Hardware Interface Helpers instead
of using hard-coded callbacks. This makes the code more readable and
more flexible.
No functional change.
Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Switch stmmac_ops to generic Hardware Interface Helpers instead of using
hard-coded callbacks. This makes the code more readable and more
flexible.
No functional change.
Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Switch stmmac_dma_ops to generic Hardware Interface Helpers instead of
using hard-coded callbacks. This makes the code more readable and more
flexible.
No functional change.
Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Switch stmmac_desc_ops to generic Hardware Interface Helpers instead of
using hard-coded callbacks. This makes the code more readable and more
flexible.
No functional change.
Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This is a reference program showing how mmap() can be used
on TCP flows to implement receive zero copy.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Some networks can make sure TCP payload can exactly fit 4KB pages,
with well chosen MSS/MTU and architectures.
Implement mmap() system call so that applications can avoid
copying data without complex splice() games.
Note that a successful mmap( X bytes) on TCP socket is consuming
bytes, as if recvmsg() has been done. (tp->copied += X)
Only PROT_READ mappings are accepted, as skb page frags
are fundamentally shared and read only.
If tcp_mmap() finds data that is not a full page, or a patch of
urgent data, -EINVAL is returned, no bytes are consumed.
Application must fallback to recvmsg() to read the problematic sequence.
mmap() wont block, regardless of socket being in blocking or
non-blocking mode. If not enough bytes are in receive queue,
mmap() would return -EAGAIN, or -EIO if socket is in a state
where no other bytes can be added into receive queue.
An application might use SO_RCVLOWAT, poll() and/or ioctl( FIONREAD)
to efficiently use mmap()
On the sender side, MSG_EOR might help to clearly separate unaligned
headers and 4K-aligned chunks if necessary.
Tested:
mlx4 (cx-3) 40Gbit NIC, with tcp_mmap program provided in following patch.
MTU set to 4168 (4096 TCP payload, 40 bytes IPv6 header, 32 bytes TCP header)
Without mmap() (tcp_mmap -s)
received 32768 MB (0 % mmap'ed) in 8.13342 s, 33.7961 Gbit,
cpu usage user:0.034 sys:3.778, 116.333 usec per MB, 63062 c-switches
received 32768 MB (0 % mmap'ed) in 8.14501 s, 33.748 Gbit,
cpu usage user:0.029 sys:3.997, 122.864 usec per MB, 61903 c-switches
received 32768 MB (0 % mmap'ed) in 8.11723 s, 33.8635 Gbit,
cpu usage user:0.048 sys:3.964, 122.437 usec per MB, 62983 c-switches
received 32768 MB (0 % mmap'ed) in 8.39189 s, 32.7552 Gbit,
cpu usage user:0.038 sys:4.181, 128.754 usec per MB, 55834 c-switches
With mmap() on receiver (tcp_mmap -s -z)
received 32768 MB (100 % mmap'ed) in 8.03083 s, 34.2278 Gbit,
cpu usage user:0.024 sys:1.466, 45.4712 usec per MB, 65479 c-switches
received 32768 MB (100 % mmap'ed) in 7.98805 s, 34.4111 Gbit,
cpu usage user:0.026 sys:1.401, 43.5486 usec per MB, 65447 c-switches
received 32768 MB (100 % mmap'ed) in 7.98377 s, 34.4296 Gbit,
cpu usage user:0.028 sys:1.452, 45.166 usec per MB, 65496 c-switches
received 32768 MB (99.9969 % mmap'ed) in 8.01838 s, 34.281 Gbit,
cpu usage user:0.02 sys:1.446, 44.7388 usec per MB, 65505 c-switches
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
SO_RCVLOWAT is properly handled in tcp_poll(), so that POLLIN is only
generated when enough bytes are available in receive queue, after
David change (commit c7004482e8dc "tcp: Respect SO_RCVLOWAT in tcp_poll().")
But TCP still calls sk->sk_data_ready() for each chunk added in receive
queue, meaning thread is awaken, and goes back to sleep shortly after.
Tested:
tcp_mmap test program, receiving 32768 MB of data with SO_RCVLOWAT set to 512KB
-> Should get ~2 wakeups (c-switches) per MB, regardless of how many
(tiny or big) packets were received.
High speed (mostly full size GRO packets)
received 32768 MB (100 % mmap'ed) in 8.03112 s, 34.2266 Gbit,
cpu usage user:0.037 sys:1.404, 43.9758 usec per MB, 65497 c-switches
received 32768 MB (99.9954 % mmap'ed) in 7.98453 s, 34.4263 Gbit,
cpu usage user:0.03 sys:1.422, 44.3115 usec per MB, 65485 c-switches
Low speed (sender is ratelimited and sends 1-MSS at a time, so GRO is not helping)
received 22474.5 MB (100 % mmap'ed) in 6015.35 s, 0.0313414 Gbit,
cpu usage user:0.05 sys:1.586, 72.7952 usec per MB, 44950 c-switches
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
We should not delay acks if there are not enough bytes
in receive queue to satisfy SO_RCVLOWAT.
Since [E]POLLIN event is not going to be generated, there is little
hope for a delayed ack to be useful.
In fact, delaying ACK prevents sender from completing
the transfer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Applications might use SO_RCVLOWAT on TCP socket hoping to receive
one [E]POLLIN event only when a given amount of bytes are ready in socket
receive queue.
Problem is that receive autotuning is not aware of this constraint,
meaning sk_rcvbuf might be too small to allow all bytes to be stored.
Add a new (struct proto_ops)->set_rcvlowat method so that a protocol
can override the default setsockopt(SO_RCVLOWAT) behavior.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|