This patch adds support in the kernel for offloading in the NIC Tx and Rx
checksumming for encapsulated packets (such as VXLAN and IP GRE).
For Tx encapsulation offload, the driver will need to set the right bits
in netdev->hw_enc_features. The protocol driver will have to set the
skb->encapsulation bit and populate the inner headers, so the NIC driver will
use those inner headers to calculate the csum in hardware.
For Rx encapsulation offload, the driver will need to set again the
skb->encapsulation flag and the skb->ip_csum to CHECKSUM_UNNECESSARY.
In that case the protocol driver should push the decapsulated packet up
to the stack, again with CHECKSUM_UNNECESSARY. In ether case, the protocol
driver should set the skb->encapsulation flag back to zero. Finally the
protocol driver should have NETIF_F_RXCSUM flag set in its features.
Signed-off-by: Joseph Gasparakis <joseph.gasparakis@intel.com>
Signed-off-by: Peter P Waskiewicz Jr <peter.p.waskiewicz.jr@intel.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Orphaning frags for zero copy skbs needs to allocate data in atomic
context so is has a chance to fail. If it does we currently discard
the skb which is safe, but we don't report anything to the caller,
so it can not recover by e.g. disabling zero copy.
Add an API to free skb reporting such errors: this is used
by tun in case orphaning frags fails.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Even if skb is marked for zero copy, net core might still decide
to copy it later which is somewhat slower than a copy in user context:
besides copying the data we need to pin/unpin the pages.
Add a parameter reporting such cases through zero copy callback:
if this happens a lot, device can take this into account
and switch to copying in user context.
This patch updates all users but ignores the passed value for now:
it will be used by follow-up patches.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Network device drivers can communicate a Toeplitz hash in skb->rxhash,
but devices differ in their hashing capabilities. All compute a 5-tuple
hash for TCP over IPv4, but for other connection-oriented protocols,
they may compute only a 3-tuple. This breaks RPS load balancing, e.g.,
for TCP over IPv6 flows. Additionally, for GRE and other tunnels,
the kernel computes a 5-tuple hash over the inner packet if possible,
but devices do not.
This patch recomputes the rxhash in software in all cases where it
cannot be certain that a 5-tuple was computed. Device drivers can avoid
recomputation by setting the skb->l4_rxhash flag.
Recomputing adds cycles to each packet when RPS is enabled or the
packet arrives over a tunnel. A comparison of 200x TCP_STREAM between
two servers running unmodified netnext with rxhash computation
in hardware vs software (using ethtool -K eth0 rxhash [on|off]) shows
how much time is spent in __skb_get_rxhash in this worst case:
0.03% swapper [kernel.kallsyms] [k] __skb_get_rxhash
0.03% swapper [kernel.kallsyms] [k] __skb_get_rxhash
0.05% swapper [kernel.kallsyms] [k] __skb_get_rxhash
With 200x TCP_RR it increases to
0.10% netperf [kernel.kallsyms] [k] __skb_get_rxhash
0.10% netperf [kernel.kallsyms] [k] __skb_get_rxhash
0.10% netperf [kernel.kallsyms] [k] __skb_get_rxhash
I considered having the patch explicitly skips recomputation when it knows
that it will not improve the hash (TCP over IPv4), but that conditional
complicates code without saving many cycles in practice, because it has
to take place after flow dissector.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Over time, skb recycling infrastructure got litle interest and
many bugs. Generic rx path skb allocation is now using page
fragments for efficient GRO / TCP coalescing, and recyling
a tx skb for rx path is not worth the pain.
Last identified bug is that fat skbs can be recycled
and it can endup using high order pages after few iterations.
With help from Maxime Bizon, who pointed out that commit
87151b8689 (net: allow pskb_expand_head() to get maximum tailroom)
introduced this regression for recycled skbs.
Instead of fixing this bug, lets remove skb recycling.
Drivers wanting really hot skbs should use build_skb() anyway,
to allocate/populate sk_buff right before netif_receive_skb()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Maxime Bizon <mbizon@freebox.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
In order to avoid false drop_monitor indications, we should
call consume_skb() if skb_clone() was successful.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The skb->pfmemalloc flag gets set to true iff during the slab allocation
of data in __alloc_skb that the the PFMEMALLOC reserves were used. If
page splitting is used, it is possible that pages will be allocated from
the PFMEMALLOC reserve without propagating this information to the skb.
This patch propagates page->pfmemalloc from pages allocated for fragments
to the skb.
It works by reintroducing and expanding the skb_alloc_page() API to take
an skb. If the page was allocated from pfmemalloc reserves, it is
automatically copied. If the driver allocates the page before the skb, it
should call skb_propagate_pfmemalloc() after the skb is allocated to
ensure the flag is copied properly.
Failure to do so is not critical. The resulting driver may perform slower
if it is used for swap-over-NBD or swap-over-NFS but it should not result
in failure.
[davem@davemloft.net: API rename and consistency]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: David S. Miller <davem@davemloft.net>
Cc: Neil Brown <neilb@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Christie <michaelc@cs.wisc.edu>
Cc: Eric B Munson <emunson@mgebm.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Change the skb allocation API to indicate RX usage and use this to fall
back to the PFMEMALLOC reserve when needed. SKBs allocated from the
reserve are tagged in skb->pfmemalloc. If an SKB is allocated from the
reserve and the socket is later found to be unrelated to page reclaim, the
packet is dropped so that the memory remains available for page reclaim.
Network protocols are expected to recover from this packet loss.
[a.p.zijlstra@chello.nl: Ideas taken from various patches]
[davem@davemloft.net: Use static branches, coding style corrections]
[sebastian@breakpoint.cc: Avoid unnecessary cast, fix !CONFIG_NET build]
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: David S. Miller <davem@davemloft.net>
Cc: Neil Brown <neilb@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Christie <michaelc@cs.wisc.edu>
Cc: Eric B Munson <emunson@mgebm.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Christoph Lameter <cl@linux.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Many places do
if ((skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY))
skb_copy_ubufs(skb, gfp_mask);
to copy and invoke frag destructors if necessary.
Add an inline helper for this.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Orphaning skb in dev_hard_start_xmit() makes bonding behavior
unfriendly for applications sending big UDP bursts : Once packets
pass the bonding device and come to real device, they might hit a full
qdisc and be dropped. Without orphaning, the sender is automatically
throttled because sk->sk_wmemalloc reaches sk->sk_sndbuf (assuming
sk_sndbuf is not too big)
We could try to defer the orphaning adding another test in
dev_hard_start_xmit(), but all this seems of little gain,
now that BQL tends to make packets more likely to be parked
in Qdisc queues instead of NIC TX ring, in cases where performance
matters.
Reverts commits :
fc6055a5ba net: Introduce skb_orphan_try()
87fd308cfc net: skb_tx_hash() fix relative to skb_orphan_try()
and removes SKBTX_DRV_NEEDS_SK_REF flag
Reported-and-bisected-by: Jean-Michel Hautbois <jhautbois@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Oliver Hartkopp <socketcan@hartkopp.net>
Acked-by: Oliver Hartkopp <socketcan@hartkopp.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
At the beginning of __skb_cow, headroom gets set to a minimum of
NET_SKB_PAD. This causes unnecessary reallocations if the buffer was not
cloned and the headroom is just below NET_SKB_PAD, but still more than the
amount requested by the caller.
This was showing up frequently in my tests on VLAN tx, where
vlan_insert_tag calls skb_cow_head(skb, VLAN_HLEN).
Locally generated packets should have enough headroom, and for forward
paths, we already have NET_SKB_PAD bytes of headroom, so we don't need to
add any extra space here.
Signed-off-by: Felix Fietkau <nbd@openwrt.org>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move tcp_try_coalesce() protocol independent part to
skb_try_coalesce().
skb_try_coalesce() can be used in IPv4 defrag and IPv6 reassembly,
to build optimized skbs (less sk_buff, and possibly less 'headers')
skb_try_coalesce() is zero copy, unless the copy can fit in destination
header (its a rare case)
kfree_skb_partial() is also moved to net/core/skbuff.c and exported,
because IPv6 will need it in patch (ipv6: use skb coalescing in
reassembly).
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix two issues introduced in commit a1c7fff7e1
( net: netdev_alloc_skb() use build_skb() )
- Must be IRQ safe (non NAPI drivers can use it)
- Must not leak the frag if build_skb() fails to allocate sk_buff
This patch introduces netdev_alloc_frag() for drivers willing to
use build_skb() instead of __netdev_alloc_skb() variants.
Factorize code so that :
__dev_alloc_skb() is a wrapper around __netdev_alloc_skb(), and
dev_alloc_skb() a wrapper around netdev_alloc_skb()
Use __GFP_COLD flag.
Almost all network drivers now benefit from skb->head_frag
infrastructure.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/ethernet/intel/e1000e/param.c
drivers/net/wireless/iwlwifi/iwl-agn-rx.c
drivers/net/wireless/iwlwifi/iwl-trans-pcie-rx.c
drivers/net/wireless/iwlwifi/iwl-trans.h
Resolved the iwlwifi conflict with mainline using 3-way diff posted
by John Linville and Stephen Rothwell. In 'net' we added a bug
fix to make iwlwifi report a more accurate skb->truesize but this
conflicted with RX path changes that happened meanwhile in net-next.
In e1000e a conflict arose in the validation code for settings of
adapter->itr. 'net-next' had more sophisticated logic so that
logic was used.
Signed-off-by: David S. Miller <davem@davemloft.net>
With the recent changes for how we compute the skb truesize it occurs to me
we are probably going to have a lot of calls to skb_end_pointer -
skb->head. Instead of running all over the place doing that it would make
more sense to just make it a separate inline skb_end_offset(skb) that way
we can return the correct value without having gcc having to do all the
optimization to cancel out skb->head - skb->head.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds support for a skb_head_is_locked helper function. It is
meant to be used any time we are considering transferring the head from
skb->head to a paged frag. If the head is locked it means we cannot remove
the head from the skb so it must be copied or we must take the skb as a
whole.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
remove useless casts and rename variables for less confusion.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
skb->head is currently allocated from kmalloc(). This is convenient but
has the drawback the data cannot be converted to a page fragment if
needed.
We have three spots were it hurts :
1) GRO aggregation
When a linear skb must be appended to another skb, GRO uses the
frag_list fallback, very inefficient since we keep all struct sk_buff
around. So drivers enabling GRO but delivering linear skbs to network
stack aren't enabling full GRO power.
2) splice(socket -> pipe).
We must copy the linear part to a page fragment.
This kind of defeats splice() purpose (zero copy claim)
3) TCP coalescing.
Recently introduced, this permits to group several contiguous segments
into a single skb. This shortens queue lengths and save kernel memory,
and greatly reduce probabilities of TCP collapses. This coalescing
doesnt work on linear skbs (or we would need to copy data, this would be
too slow)
Given all these issues, the following patch introduces the possibility
of having skb->head be a fragment in itself. We use a new skb flag,
skb->head_frag to carry this information.
build_skb() is changed to accept a frag_size argument. Drivers willing
to provide a page fragment instead of kmalloc() data will set a non zero
value, set to the fragment size.
Then, on situations we need to convert the skb head to a frag in itself,
we can check if skb->head_frag is set and avoid the copies or various
fallbacks we have.
This means drivers currently using frags could be updated to avoid the
current skb->head allocation and reduce their memory footprint (aka skb
truesize). (thats 512 or 1024 bytes saved per skb). This also makes
bpf/netfilter faster since the 'first frag' will be part of skb linear
part, no need to copy data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Maciej Żenczykowski <maze@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Matt Carlson <mcarlson@broadcom.com>
Cc: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The skb struct ubuf_info callback gets passed struct ubuf_info
itself, not the arg value as the field name and the function signature
seem to imply. Rename the arg field to ctx to match usage,
add documentation and change the callback argument type
to make usage clear and to have compiler check correctness.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Marc Merlin reported many order-1 allocations failures in TX path on its
wireless setup, that dont make any sense with MTU=1500 network, and non
SG capable hardware.
After investigation, it turns out TCP uses sk_stream_alloc_skb() and
used as a convention skb_tailroom(skb) to know how many bytes of data
payload could be put in this skb (for non SG capable devices)
Note : these skb used kmalloc-4096 (MTU=1500 + MAX_HEADER +
sizeof(struct skb_shared_info) being above 2048)
Later, mac80211 layer need to add some bytes at the tail of skb
(IEEE80211_ENCRYPT_TAILROOM = 18 bytes) and since no more tailroom is
available has to call pskb_expand_head() and request order-1
allocations.
This patch changes sk_stream_alloc_skb() so that only
sk->sk_prot->max_header bytes of headroom are reserved, and use a new
skb field, avail_size to hold the data payload limit.
This way, order-0 allocations done by TCP stack can leave more than 2 KB
of tailroom and no more allocation is performed in mac80211 layer (or
any layer needing some tailroom)
avail_size is unioned with mark/dropcount, since mark will be set later
in IP stack for output packets. Therefore, skb size is unchanged.
Reported-by: Marc MERLIN <marc@merlins.org>
Tested-by: Marc MERLIN <marc@merlins.org>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
