Conversation
alobakin
force-pushed
the
iavf-xdp
branch
2 times, most recently
from
March 1, 2023 16:57
41f3642 to
8862fe7
Compare
alobakin
force-pushed
the
iavf-xdp
branch
2 times, most recently
from
March 3, 2023 18:04
777041a to
6e1d7f5
Compare
alobakin
force-pushed
the
iavf-xdp
branch
4 times, most recently
from
March 10, 2023 09:39
fe96a55 to
e327e47
Compare
Not a secret there's a ton of code duplication between two and more Intel ethernet modules. Before introducing new changes, which would need to be copied over again, start decoupling the already existing duplicate functionality into a new module, which will be shared between several Intel Ethernet drivers. Add the lookup table which converts 8/10-bit hardware packet type into a parsed bitfield structure for easy checking packet format parameters, such as payload level, IP version, etc. This is currently used by i40e, ice and iavf and it's all the same in all three drivers. The only difference introduced in this implementation is that instead of defining a 256 (or 1024 in case of ice) element array, add unlikely() condition to limit the input to 154 (current maximum non-reserved packet type). There's no reason to waste 600 (or even 3600) bytes only to not hurt very unlikely exception packets. The hash computation function now takes payload level directly as a pkt_hash_type. There's a couple cases when non-IP ptypes are marked as L3 payload and in the previous versions their hash level would be 2, not 3. But skb_set_hash() only sees difference between L4 and non-L4, thus this won't change anything at all. The module is behind the hidden Kconfig symbol, which the drivers will select when needed. The exports are behind 'LIBIE' namespace to limit the scope of the functions. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Ever since build_skb() became stable, the old way with allocating an skb for storing the headers separately, which will be then copied manually, was slower, less flexible and thus obsolete. * it had higher pressure on MM since it actually allocates new pages, which then get split and refcount-biased (NAPI page cache); * it implies memcpy() of packet headers (40+ bytes per each frame); * the actual header length was calculated via eth_get_headlen(), which invokes Flow Dissector and thus wastes a bunch of CPU cycles; * XDP makes it even more weird since it requires headroom for long and also tailroom for some time (since mbuf landed). Take a look at the ice driver, which is built around work-arounds to make XDP work with it. Even on some quite low-end hardware (not a common case for 100G NICs) it was performing worse. The only advantage "legacy-rx" had is that it didn't require any reserved headroom and tailroom. But iavf didn't use this, as it always splits pages into two halves of 2k, while that save would only be useful when striding. And again, XDP effectively removes that sole pro. There's a train of features to land in IAVF soon: Page Pool, XDP, XSk, multi-buffer etc. Each new would require adding more and more Danse Macabre for absolutely no reason, besides making hotpath less and less effective. Remove the "feature" with all the related code. This includes at least one very hot branch (typically hit on each new frame), which was either always-true or always-false at least for a complete NAPI bulk of 64 frames, the whole private flags cruft and so on. Some stats: Function: add/remove: 0/2 grow/shrink: 0/7 up/down: 0/-774 (-774) RO Data: add/remove: 0/1 grow/shrink: 0/0 up/down: 0/-40 (-40) Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
The Rx hotpath code of IAVF is not well-optimized TBH. Before doing any further buffer model changes, shake it up a bit. Notably: 1. Cache more variables on the stack. DMA device, Rx page size, NTC -- these are the most common things used all throughout the hotpath, often in loops on each iteration. Instead of fetching (or even calculating, as with the page size) them from the ring all the time, cache them on the stack at the beginning of the NAPI polling callback. NTC will be written back at the end, the rest are used read-only, so no sync needed. 2. Don't move the recycled buffers around the ring. The idea of passing the page of the right-now-recycled-buffer to a different buffer, in this case, the first one that needs to be allocated, moreover, on each new frame, is fundamentally wrong. It involves a few o' fetches, branches and then writes (and one Rx buffer struct is at least 32 bytes) where they're completely unneeded, but gives no good -- the result is the same as if we'd recycle it inplace, at the same position where it was used. So drop this and let the main refilling function take care of all the buffers, which were processed and now need to be recycled/refilled. 3. Don't allocate with %GPF_ATOMIC on ifup. This involved introducing the @gfp parameter to a couple functions. Doesn't change anything for Rx -> softirq. 4. 1 budget unit == 1 descriptor, not skb. There could be underflow when receiving a lot of fragmented frames. If each of them would consist of 2 frags, it means that we'd process 64 descriptors at the point where we pass the 32th skb to the stack. But the driver would count that only as a half, which could make NAPI re-enable interrupts prematurely and create unnecessary CPU load. 5. Shortcut !size case. It's super rare, but possible -- for example, if the last buffer of the fragmented frame contained only FCS, which was then stripped by the HW. Instead of checking for size several times when processing, quickly reuse the buffer and jump to the skb fields part. 6. Refill the ring after finishing the polling loop. Previously, the loop wasn't starting a new iteration after the 64th desc, meaning that we were always leaving 16 buffers non-refilled until the next NAPI poll. It's better to refill them while they're still hot, so do that right after exiting the loop as well. For a full cycle of 64 descs, there will be 4 refills of 16 descs from now on. Function: add/remove: 4/2 grow/shrink: 0/5 up/down: 473/-647 (-174) + up to 2% performance. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
As an intermediate step, remove all page splitting/recyclig code. Just always allocate a new page and don't touch its refcount, so that it gets freed by the core stack later. The change allows to greatly simplify certain parts of the code: Function: add/remove: 2/3 grow/shrink: 0/5 up/down: 543/-963 (-420) &iavf_rx_buf can even now retire in favor of just storing an array of pages used for Rx. Their DMA addresses can be stored in page::dma_addr -- use Page Pool's function for that. No surprise perf loses up to 30% here, but that regression will go away once PP lands. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
The current scheme with trying to pick the smallest buffer possible for the current MTU in order to flip/split pages is not very optimal. For example, on default MTU of 1500 it gives only 192 bytes of headroom, while XDP may require up to 258. But this also involves unnecessary code complication, which sometimes is even hard to follow. As page split is no more, always allocate order-0 pages. This optimizes performance a bit and drops some bytes off the object code. Next, always pick the maximum buffer length available for this %PAGE_SIZE to set it up in the hardware. This means it now becomes a constant value, which also has its positive impact. On x64 this means (without XDP): 4096 page 64 head, 320 tail 3712 HW buffer size 3686 max MTU w/o frags Previously, the maximum MTU w/o splitting a frame into several buffers was 3046. Increased buffer size allows us to reach the maximum frame size w/ frags supported by HW: 16382 bytes (MTU 16356). Reflect it in the netdev config as well. Relying on max single buffer size when calculating MTU was not correct. Move around a couple of fields in &iavf_ring after ::rx_buf_len removal to reduce holes and improve cache locality. Instead of providing the Rx definitions, which can and will be reused in rest of the drivers, exclusively for IAVF, do that in the libie header. Non-PP drivers could still use at least some of them and lose a couple copied lines. Function: add/remove: 0/0 grow/shrink: 3/9 up/down: 18/-265 (-247) + even reclaims a half percent of performance, nice. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Add a new flag, %PP_FLAG_DMA_MAP_WEAK, whill will tell PP to map pages with %DMA_ATTR_WEAK_ORDERING. To keep the code simple and optimized, map the following PP flags to DMA map attr flags: %PP_FLAG_DMA_MAP => %DMA_ATTR_SKIP_CPU_SYNC %PP_FLAG_DMA_MAP_WEAK => %DMA_ATTR_WEAK_ORDERING The first pair is done to be able to just pass it directly to dma_map_page_attrs(). When a driver wants Page Pool to maintain DMA mappings, it always sets this flag. Page Pool always skips CPU syncs when mapping to do that separately later, so having those two 1:1 avoids introducing ifs and/or bit-ors and keeps the code more compact. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Each driver is responsible for syncing buffers written by HW for CPU before accessing them. Almost each PP-enabled driver uses the same pattern, which could be shorthanded into a static inline to make driver code a little bit more compact. Introduce a pair of such functions. The first one takes the actual size of the data written by HW and is the main one to be used on Rx. The second picks max_len from the PP params and is designed for more extreme cases when the size is unknown, but the buffer still needs to be synced. Also constify pointer arguments of page_pool_get_dma_dir() and page_pool_get_dma_addr() to give a bit more room for optimization, as both of them are read-only. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Now that the IAVF driver simply uses dev_alloc_page() + free_page() with no custom recycling logics and one whole page per frame, it can easily be switched to using Page Pool API instead. Introduce libie_rx_page_pool_create(), a wrapper for creating a PP with the default libie settings applicable to all Intel hardware, and replace the alloc/free calls with the corresponding PP functions, including the newly added sync-for-CPU helpers. Use skb_mark_for_recycle() to bring back the recycling and restore the initial performance. From the important object code changes, worth mentioning that __iavf_alloc_rx_pages() is now inlined due to the greatly reduced size. The resulting driver is on par with the pre-series code and 1-2% slower than the "optimized" version right before the recycling removal. But the number of locs and object code bytes slaughtered is much more important here after all, not speaking of that there's still a vast space for optimization and improvements. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Next stop, per-queue private stats. They have only subtle differences from driver to driver and can easily be resolved. Define common structures, inline helpers and Ethtool helpers to collect, update and export the statistics. Use u64_stats_t right from the start, as well as the corresponding helpers to ensure tear-free operations. For the NAPI parts of both Rx and Tx, also define small onstack containers to update them in polling loops and then sync the actual containers once a loop ends. The drivers will be switched to use this API later on a per-driver basis, along with conversion to PP. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Expand the libie generic per-queue stats with the generic Page Pool stats provided by the API itself, when CONFIG_PAGE_POOL is enable. When it's not, there'll be no such fields in the stats structure, so no space wasted. They are also a bit special in terms of how they are obtained. One &page_pool accumulates statistics until it's destroyed obviously, which happens on ifdown. So, in order to not lose any statistics, get the stats and store in the queue container before destroying a pool. This container survives ifups/downs, so it basically stores the statistics accumulated since the very first pool was allocated on this queue. When it's needed to export the stats, first get the numbers from this container and then add the "live" numbers -- the ones that the current active pool returns. The result values will always represent the actual device-lifetime* stats. There's a cast from &page_pool_stats to `u64 *` in a couple functions, but they are guarded with stats asserts to make sure it's safe to do. FWIW it saves a lot of object code. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
iavf is pretty much ready for using the generic libie stats, so drop all the custom code and just use generic definitions. The only thing is that it previously lacked the counter of Tx queue stops. It's present in the other drivers, so add it here as well. The rest is straightforward. There were two fields in the Tx stats struct, which didn't belong there. The first one has never been used, wipe it; and move the other to the queue structure. Plus move around a couple fields in &iavf_ring to account stats structs' alignment. Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Currently, the test relies on that only dropped ("xmitted") frames will
be recycled and if a frame became an skb, it will be freed later by the
stack and never come back to its page_pool.
So, it easily gets broken by trying to recycle skbs:
test_xdp_do_redirect:PASS:pkt_count_xdp 0 nsec
test_xdp_do_redirect:FAIL:pkt_count_zero unexpected pkt_count_zero:
actual 9936 != expected 2
test_xdp_do_redirect:PASS:pkt_count_tc 0 nsec
That huge mismatch happened because after the TC ingress hook zeroes the
magic, the page gets recycled when skb is freed, not returned to the MM
layer. "Live frames" mode initializes only new pages and keeps the
recycled ones as is by design, so they appear with zeroed magic on the
Rx path again.
Expand the possible magic values from two: 0 (was "xmitted"/dropped or
did hit the TC hook) and 0x42 (hit the input XDP prog) to three: the new
one will mark frames hit the TC hook, so that they will elide both
@pkt_count_zero and @pkt_count_xdp. They can then be recycled to their
page_pool or returned to the page allocator, this won't affect the
counters anyhow. Just make sure to mark them as "input" (0x42) when they
appear on the Rx path again.
Also make an enum from those magics, so that they will be always visible
and can be changed in just one place anytime. This also eases adding any
new marks later on.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
skb_mark_for_recycle() is guarded with CONFIG_PAGE_POOL, this creates unneeded complication when using it in the generic code. For now, it's only used in the drivers always selecting Page Pool, so this works. Move the guards so that preprocessor will cut out only the operation itself and the function will still be a noop on !PAGE_POOL systems, but available there as well. No functional changes. Reported-by: kernel test robot <lkp@intel.com> Link: https://lore.kernel.org/oe-kbuild-all/202303020342.Wi2PRFFH-lkp@intel.com Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
__xdp_build_skb_from_frame() state(d): /* Until page_pool get SKB return path, release DMA here */ Page Pool got skb pages recycling in April 2021, but missed this function. xdp_release_frame() is relevant only for Page Pool backed frames and it detaches the page from the corresponding page_pool in order to make it freeable via page_frag_free(). It can instead just mark the output skb as eligible for recycling if the frame is backed by a pp. No change for other memory model types (the same condition check as before). cpumap redirect and veth on Page Pool drivers now become zero-alloc (or almost). Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
__xdp_build_skb_from_frame() was the last user of
{__,}xdp_release_frame(), which detaches pages from the page_pool.
All the consumers now recycle Page Pool skbs and page, except mlx5,
stmmac and tsnep drivers, which use page_pool_release_page() directly
(might change one day). It's safe to assume this functionality is not
needed anymore and can be removed (in favor of recycling).
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
The current version of Intel 'ice' driver allows for using zero for the ring lenghth in 'configure queue' VIRTCHNL message. Such a value indicates the ring should not be configured. Implement the same handling in i40e driver. Instead of returning an 'invalid parameter' error for zero-sized rings, just skip that ring during queue pair configuration. That unified handling is needed for AF_XDP implementation for 'iavf' driver. In that use case we sometimes need to configure Tx ring only for a given queue pair. Signed-off-by: Michal Kubiak <michal.kubiak@intel.com>
The XDP and AF_XDP feature is initialized using .ndo functions. Those functions are always synchronous and may require some serious queues reconfiguration including changing the number of queues. Performing such a reconfiguration implies sending a bunch of VIRTCHNL messages to the PF in order to disable queues, re-enable and re-configure them, or update the RSS LUT. By definition, those VIRTCHNL messages are sent asynchronously, so the result of each VIRTCHNL operation can be received from the PF via admin queue after some time. Moreover, the previous implementation of some VIRTCHNL functions (e.g. 'iavf_disable_queues()' or 'iavf_enable_queues()' does not allow to call them selectively for specific queues only. In order to addres those problems and cover all scenarios of XDP and AF_XDP initialization, implement a polling mechanism with a timeout for blocking the execution of XDP .ndo functions until the result of VIRTCHNL operation on PF is known to the driver. Also, refactor the existing VIRTCHNL API by adding functions for selective queue enabling, disabling and configuration. Signed-off-by: Michal Kubiak <michal.kubiak@intel.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
XDP and XSk parts are untested, has some TODOs to finish + lacks latest improvements and fixes from bpf-next's ice.