skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag), DMA_TO_DEVICE)
is repeated across dozens of drivers and really wants a shorthand.
Add a macro which will count args and handle all possible number
from 2 to 5. Semantics:

skb_frag_dma_map(dev, frag) ->
__skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag), DMA_TO_DEVICE)

skb_frag_dma_map(dev, frag, offset) ->
__skb_frag_dma_map(dev, frag, offset, skb_frag_size(frag) - offset,
		   DMA_TO_DEVICE)

skb_frag_dma_map(dev, frag, offset, size) ->
__skb_frag_dma_map(dev, frag, offset, size, DMA_TO_DEVICE)

skb_frag_dma_map(dev, frag, offset, size, dir) ->
__skb_frag_dma_map(dev, frag, offset, size, dir)

No object code size changes for the existing callers. Users passing
less arguments also won't have bigger size comparing to the full
equivalent call.

Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

There are cases when we need to explicitly unroll loops. For example,
cache operations, filling DMA descriptors on very high speeds etc.
Add compiler-specific attribute macros to give the compiler a hint
that we'd like to unroll a loop.
Example usage:

 #define UNROLL_BATCH 8

	unrolled_count(UNROLL_BATCH)
	for (u32 i = 0; i < UNROLL_BATCH; i++)
		op(priv, i);

Note that sometimes the compilers won't unroll loops if they think this
would have worse optimization and perf than without unrolling, and that
unroll attributes are available only starting GCC 8. For older compiler
versions, no hints/attributes will be applied.
For better unrolling/parallelization, don't have any variables that
interfere between iterations except for the iterator itself.

Co-developed-by: Jose E. Marchesi <jose.marchesi@oracle.com> # pragmas
Signed-off-by: Jose E. Marchesi <jose.marchesi@oracle.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

In lots of places, bpf_prog pointer is used only for tracing or other
stuff that doesn't modify the structure itself. Same for net_device.
Address at least some of them and add `const` attributes there. The
object code didn't change, but that may prevent unwanted data
modifications and also allow more helpers to have const arguments.

Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

Lots of read-only helpers for &xdp_buff and &xdp_frame, such as getting
the frame length, skb_shared_info etc., don't have their arguments
marked with `const` for no reason. Add the missing annotations to leave
less place for mistakes and more for optimization.

Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

One may need to register memory model separately from xdp_rxq_info. One
simple example may be XDP test run code, but in general, it might be
useful when memory model registering is managed by one layer and then
XDP RxQ info by a different one.
Allow such scenarios by adding a simple helper which "attaches" an
already registered memory model to the desired xdp_rxq_info. As this
is mostly needed for Page Pool, add a special function to do that for
a &page_pool pointer.

Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

To make the system page pool usable as a source for allocating XDP
frames, we need to register it with xdp_reg_mem_model(), so that page
return works correctly. This is done in preparation for using the system
page_pool to convert XDP_PASS XSk frames to skbs; for the same reason,
make the per-cpu variable non-static so we can access it from other
source files as well (but w/o exporting).

Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

Currently, page_pool_put_page_bulk() indeed takes an array of pointers
to the data, not pages, despite the name. As one side effect, when
you're freeing frags from &skb_shared_info, xdp_return_frame_bulk()
converts page pointers to virtual addresses and then
page_pool_put_page_bulk() converts them back.
Make page_pool_put_page_bulk() actually handle array of pages. Pass
frags directly and use virt_to_page() when freeing xdpf->data, so that
the PP core will then get the compound head and take care of the rest.

Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

The main reason for this change was to allow mixing pages from different
&page_pools within one &xdp_buff/&xdp_frame. Why not?
Adjust xdp_return_frame_bulk() and page_pool_put_page_bulk(), so that
they won't be tied to a particular pool. Let the latter create a
separate bulk of pages which's PP is different and process it after
the main loop.
This greatly optimizes xdp_return_frame_bulk(): no more hashtable
lookups. Also make xdp_flush_frame_bulk() inline, as it's just one if +
function call + one u32 read, not worth extending the call ladder.

Co-developed-by: Toke Høiland-Jørgensen <toke@redhat.com> # iterative
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

Initially, xdp_frame::mem.id was used to search for the corresponding
&page_pool to return the page correctly.
However, after that struct page now contains a direct pointer to its PP,
further keeping of this field makes no sense. xdp_return_frame_bulk()
still uses it to do a lookup, but this is rather a leftover.
Remove xdp_frame::mem and replace it with ::mem_type, as only memory
type still matters and we need to know it to be able to free the frame
correctly.
As a cute side effect, we can now make every scalar field in &xdp_frame
of 4 byte width, speeding up accesses to them.

Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

The code piece which would attach a frag to &xdp_buff is almost
identical across the drivers supporting XDP multi-buffer on Rx.
Make it a generic elegant "oneliner".
Also, I see lots of drivers calculating frags_truesize as
`xdp->frame_sz * nr_frags`. I can't say this is fully correct, since
frags might be backed by chunks of different sizes, especially with
stuff like the header split. Even page_pool_alloc() can give you two
different truesizes on two subsequent requests to allocate the same
buffer size. Add a field to &skb_shared_info (unionized as there's no
free slot currently on x86_64) to track the "true" truesize. It can
be used later when updating an skb.

Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

The code which builds an skb from an &xdp_buff keeps multiplying itself
around the drivers with almost no changes. Let's try to stop that by
adding a generic function.
Unlike __xdp_build_skb_from_frame(), always allocate an skbuff head
using napi_build_skb() and make use of the available xdp_rxq pointer to
assign the Rx queue index. In case of PP-backed buffer, mark the skb to
be recycled, as every PP user's been switched to recycle skbs.

Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

After the series "XSk buff on a diet" by Maciej, the greatest pow-2
which &xdp_buff_xsk can be divided got reduced from 16 to 8 on x86_64.
Also, sizeof(xdp_buff_xsk) now is 120 bytes, which, taking the previous
sentence into account, leads to that it leaves 8 bytes at the end of
cacheline, which means an array of buffs will have its elements
messed between the cachelines chaotically.
Use __aligned_largest for this struct. This alignment is usually 16
bytes, which makes it fill two full cachelines and align an array
nicely. ___cacheline_aligned may be excessive here, especially on
arches with 128-256 byte CLs, as well as 32-bit arches (76 -> 96
bytes on MIPS32R2), while not doing better than _largest.

Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

When you register an XSk pool as XDP Rxq info memory model, you then
need to manually attach it after the registration.
Let the user combine both actions into one by just passing a pointer
to the pool directly to xdp_rxq_info_reg_mem_model(), which will take
care of calling xsk_pool_set_rxq_info(). This looks similar to how a
&page_pool gets registered and reduce repeating driver code.

Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

Currently, xsk_buff_add_frag() only adds a frag to the pool linked list,
not doing anything with the &xdp_buff. The drivers do that manually and
the logic is the same.
Make it really add an skb frag, just like xdp_buff_add_frag() does that,
and freeing frags on error if needed. This allows to remove repeating
code from i40e and ice and not add the same code again and again.

Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

Same as with converting &xdp_buff to skb on Rx, the code which allocates
a new skb and copies the XSk frame there is identical across the
drivers, so make it generic. This includes copying all the frags if they
are present in the original buff.
System percpu Page Pools help here a lot: when available, allocate pages
from there instead of the MM layer. This greatly improves XDP_PASS
performance on XSk: instead of page_alloc() + page_free(), the net core
recycles the same pages, so the only overhead left is memcpy()s.
Note that the passed buff gets freed if the conversion is done w/o any
error, assuming you don't need this buffer after you convert it to an
skb.

Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

Currently, when you send an XSk frame with metadata, you need to do
the following:

* call external xsk_buff_raw_get_dma();
* call inline xsk_buff_get_metadata(), which calls external
  xsk_buff_raw_get_data() and then do some inline checks.

This effectively means that the following piece:

addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;

is done twice per frame, plus you have 2 external calls per frame, plus
this:

	meta = pool->addrs + addr - pool->tx_metadata_len;
	if (unlikely(!xsk_buff_valid_tx_metadata(meta)))

is always inlined, even if there's no meta or it's invalid.

Add xsk_buff_raw_get_ctx() (xp_raw_get_ctx() to be precise) to do that
in one go. It returns a small structure with 2 fields: DMA address,
filled unconditionally, and metadata pointer, valid only if it's
present. The address correction is performed only once and you also
have only 1 external call per XSk frame, which does all the calculations
and checks outside of your hotpath. You only need to check
`if (ctx.meta)` for the metadata presence.

Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

Expand libeth's Page Pool functionality by adding native XDP support.
This means picking the appropriate headroom and DMA direction.
Also, register all the created &page_pools as XDP memory models.
A driver then can call xdp_rxq_info_attach_page_pool() when registering
its RxQ info.

Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

"Couple" is a bit humbly... Add the following functionality to libeth:

* XDP shared queues managing
* XDP_TX bulk sending infra
* .ndo_xdp_xmit() infra
* adding buffers to &xdp_buff
* running XDP prog and managing its verdict
* completing XDP Tx buffers
* ^ repeat everything for XSk

Suggested-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com> # lots of stuff
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>