New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
ice: add XDP mbuf support #4424
Conversation
Rx path is going to be modified in a way that fragmented frame will be gathered within xdp_buff in the first place. This approach implies that underlying buffer has to provide tailroom for skb_shared_info. This is currently the case when ring uses build_skb but not when legacy-rx knob is turned on. This case configures 2k Rx buffers and has no way to provide either headroom or tailroom - FWIW it currently has XDP_PACKET_HEADROOM which is broken and in here it is removed. 2k Rx buffers were used so driver in this setting was able to support 9k MTU as it can chain up to 5 Rx buffers. With offset configuring HW writing 2k of a data was passing the half of the page which broke the assumption of our internal page recycling tricks. Now if above got fixed and legacy-rx path would be left as is, when referring to skb_shared_info via xdp_get_shared_info_from_buff(), packet's content would be corrupted again. Hence size of Rx buffer needs to be lowered and therefore supported MTU. This operation will allow us to keep the unified data path and with 8k MTU users (if any of legacy-rx) would still be good to go. However, tendency is to drop the support for this code path at some point. Add ICE_RXBUF_1664 as vsi::rx_buf_len and ICE_MAX_FRAME_LEGACY_RX (8320) as vsi::max_frame for legacy-rx. For bigger page sizes configure 3k Rx buffers, not 2k. Since headroom support is removed, disable data_meta support on legacy-rx. When preparing XDP buff, rely on ice_rx_ring::rx_offset setting when deciding whether to support data_meta or not. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
In preparation for XDP multi-buffer support, let's store xdp_buff on Rx ring struct. This will allow us to combine fragmented frames across separate NAPI cycles in the same way as currently skb fragments are handled. This means that skb pointer on Rx ring will become redundant and will be removed. For now it is kept and layout of Rx ring struct was not inspected, some member movement will be needed later on so that will be the time to take care of it. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
This will allow us to avoid carrying additional auxiliary array of page counts when dealing with XDP multi buffer support. Previously combining fragmented frame to skb was not affected in the same way as XDP would be as whole frame is needed to be in place before executing XDP prog. Therefore, when going through HW Rx descriptors one-by-one, calls to ice_put_rx_buf() need to be taken *after* running XDP prog on a potentially multi buffered frame, so some additional storage of page count is needed. By adding page count to rx buf, it will make it easier to walk through processed entries at the end of rx cleaning routine and decide whether or not buffers should be recycled. While at it, bump ice_rx_buf::pagecnt_bias from u16 up to u32. It was proven many times that calculations on variables smaller than standard register size are harmful. This was also the case during experiments with embedding page count to ice_rx_buf - when this was added as u16 it had a performance impact. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Plan is to move ice_put_rx_buf() to the end of ice_clean_rx_irq() so in order to keep the ability of walking through HW Rx descriptors, pull out next_to_clean handling out of ice_put_rx_buf(). Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
This might be in future used by ZC driver and might potentially yield a minor performance boost. While at it, constify arguments that ice_is_non_eop() takes, since they are pointers and this will help compiler while generating asm. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Currently calls to ice_put_rx_buf() are sprinkled through ice_clean_rx_irq() - first place is for explicit flow director's descriptor handling, second is after running XDP prog and the last one is after taking care of skb. 1st callsite was actually only for ntc bump purpose, as Rx buffer to be recycled is not even passed to a function. It is possible to walk through Rx buffers processed in particular NAPI cycle by caching ntc from beginning of the ice_clean_rx_irq(). To do so, let us store XDP verdict inside ice_rx_buf, so action we need to take on will be known. For XDP prog absence, just store ICE_XDP_PASS as a verdict. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
This should have been used in there from day 1, let us address that before introducing XDP multi-buffer support for Rx side. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Currently ice_finalize_xdp_rx() is called only when xdp_prog is present on VSI, which is a good thing. However, this optimization can be enhanced and check only if any of the XDP_TX/XDP_REDIRECT took place in current Rx processing. Non-zero value of @xdp_xmit indicates that xdp_prog is present on VSI. This way XDP_DROP-based workloads will not suffer from unnecessary calls to external function. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
SKB path calculates truesize on three different functions, which could be avoided as xdp_buff carries the already calculated truesize under xdp_buff::frame_sz. If ice_add_rx_frag() is adjusted to take the xdp_buff as an input just like functions responsible for creating sk_buff initially, codebase could be simplified by removing these redundant recalculations and rely on xdp_buff::frame_sz instead. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Ice driver needs to be a bit reworked on Rx data path in order to support multi-buffer XDP. For skb path, it currently works in a way that Rx ring carries pointer to skb so if driver didn't manage to combine fragmented frame at current NAPI instance, it can restore the state on next instance and keep looking for last fragment (so descriptor with EOP bit set). What needs to be achieved is that xdp_buff needs to be combined in such way (linear + frags part) in the first place. Then skb will be ready to go in case of XDP_PASS or BPF program being not present on interface. If BPF program is there, it would work on multi-buffer XDP. At this point xdp_buff resides directly on Rx ring, so given the fact that skb will be built straight from xdp_buff, there will be no further need to carry skb on Rx ring. Besides removing skb pointer from Rx ring, lots of members have been moved around within ice_rx_ring. First and foremost reason was to place rx_buf with xdp_buff on the same cacheline. This means that once we touch rx_buf (which is a preceding step before touching xdp_buff), xdp_buff will already be hot in cache. Second thing was that xdp_rxq is used rather rarely and it occupies a separate cacheline, so maybe it is better to have it at the end of ice_rx_ring. Other change that affects ice_rx_ring is the introduction of ice_rx_ring::first_desc. Its purpose is twofold - first is to propagate rx_buf->act to all the parts of current xdp_buff after running XDP program, so that ice_put_rx_buf() that got moved out of the main Rx processing loop will be able to tak an appriopriate action on each buffer. Second is for ice_construct_skb(). ice_construct_skb() has a copybreak mechanism which had an explicit impact on xdp_buff->skb conversion in the new approach when legacy Rx flag is toggled. It works in a way that linear part is 256 bytes long, if frame is bigger than that, remaining bytes are going as a frag to skb_shared_info. This means while memcpying frags from xdp_buff to newly allocated skb, care needs to be taken when picking the destination frag array entry. Upon the time ice_construct_skb() is called, when dealing with fragmented frame, current rx_buf points to the *last* fragment, but copybreak needs to be done against the first one. That's where ice_rx_ring::first_desc helps. When frame building spans across NAPI polls (DD bit is not set on current descriptor and xdp->data is not NULL) with current Rx buffer handling state there might be some problems. Since calls to ice_put_rx_buf() were pulled out of the main Rx processing loop and were scoped from cached_ntc to current ntc, remember that now mentioned function relies on rx_buf->act, which is set within ice_run_xdp(). ice_run_xdp() is called when EOP bit was found, so currently we could put Rx buffer with rx_buf->act being *uninitialized*. To address this, change scoping to rely on first_desc on both boundaries instead. This also implies that cleaned_count which is used as an input to ice_alloc_rx_buffers() and tells how many new buffers should be refilled has to be adjusted. If it stayed as is, what could happen is a case where ntc would go over ntu. Therefore, remove cleaned_count altogether and use against allocing routine newly introduced ICE_RX_DESC_UNUSED() macro which is an equivalent of ICE_DESC_UNUSED() dedicated for Rx side and based on struct ice_rx_ring::first_desc instead of next_to_clean. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Similarly as for Rx side in previous patch, logic on XDP Tx in ice driver needs to be adjusted for multi-buffer support. Specifically, the way how HW Tx descriptors are produced and cleaned. Currently, XDP_TX works on strict ring boundaries, meaning it sets RS bit (on producer side) / looks up DD bit (on consumer/cleaning side) every quarter of the ring. It means that if for example multi buffer frame would span across the ring quarter boundary (say that frame consists of 4 frames and we start from 62 descriptor where ring is sized to 256 entries), RS bit would be produced in the middle of multi buffer frame, which would be a broken behavior as it needs to be set on the last descriptor of the frame. To make it work, set RS bit at the last descriptor from the batch of frames that XDP_TX action was used on and make the first entry remember the index of last descriptor with RS bit set. This way, cleaning side can take the index of descriptor with RS bit, look up DD bit's presence and clean from first entry to last. In order to clean up the code base introduce the common ice_set_rs_bit() which will return index of descriptor that got RS bit produced on so that standard driver can store this within proper ice_tx_buf and ZC driver can simply ignore return value. Co-developed-by: Martyna Szapar-Mudlaw <martyna.szapar-mudlaw@linux.intel.com> Signed-off-by: Martyna Szapar-Mudlaw <martyna.szapar-mudlaw@linux.intel.com> Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Now that both ZC and standard XDP data paths stopped using Tx logic based on next_dd and next_rs fields, we can safely remove these fields and shrink Tx ring structure. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Let us store pointer to xdp_buff that came from xsk_buff_pool on tx_buf so that it will be possible to recycle it via xsk_buff_free() on Tx cleaning side. This way it is not necessary to do expensive copy to another xdp_buff backed by a newly allocated page. Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Upstream branch: c1a3daf |
c149f9d
to
4e9ef36
Compare
Upstream branch: 10d1b0e Pull request is NOT updated. Failed to apply https://patchwork.kernel.org/project/netdevbpf/list/?series=717481
conflict:
|
At least one diff in series https://patchwork.kernel.org/project/netdevbpf/list/?series=717481 irrelevant now. Closing PR. |
Pull request for series with
subject: ice: add XDP mbuf support
version: 1
url: https://patchwork.kernel.org/project/netdevbpf/list/?series=717481