Jump target resolution should be in jit.c not offload.c.
No functional changes.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

If the program is simple and has only one adjust head call
with constant parameters, we can check that the call will
always succeed at translation time.  We need to track the
location of the call and make sure parameters are always
the same.  We also have to check the parameters against
datapath constraints and ETH_HLEN.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Support bpf_xdp_adjust_head().  We need to check whether the
packet offset after adjustment is within datapath's limits.
We also check if the frame is at least ETH_HLEN long (similar
to the kernel implementation).
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Add skeleton of verifier checks and translation handler
for call instructions.  Make sure jump target resolution
will not treat them as jumps.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

This patch add the optimization frontend, but adding a new eBPF IR scan
pass "nfp_bpf_opt_ldst_gather".

The pass will traverse the IR to recognize the load/store pairs sequences
that come from lowering of memory copy builtins.

The gathered memory copy information will be kept in the meta info
structure of the first load instruction in the sequence and will be
consumed by the optimization backend added in the previous patches.

NOTE: a sequence with cross memory access doesn't qualify this
optimization, i.e. if one load in the sequence will load from place that
has been written by previous store. This is because when we turn the
sequence into single CPP operation, we are reading all contents at once
into NFP transfer registers, then write them out as a whole. This is not
identical with what the original load/store sequence is doing.

Detecting cross memory access for two random pointers will be difficult,
fortunately under XDP/eBPF's restrictied runtime environment, the copy
normally happen among map, packet data and stack, they do not overlap with
each other.

And for cases supported by NFP, cross memory access will only happen on
PTR_TO_PACKET. Fortunately for this, there is ID information that we could
do accurate memory alias check.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

When the gathered copy length is bigger than 32-bytes and within 128-bytes
(the maximum length a single CPP Pull/Push request can finish), the
strategy of read/write are changeed into:

  * Read.
      - use direct reference mode when length is within 32-bytes.
      - use indirect mode when length is bigger than 32-bytes.

  * Write.
      - length <= 8-bytes
        use write8 (direct_ref).
      - length <= 32-byte and 4-bytes aligned
        use write32 (direct_ref).
      - length <= 32-bytes but not 4-bytes aligned
        use write8 (indirect_ref).
      - length > 32-bytes and 4-bytes aligned
        use write32 (indirect_ref).
      - length > 32-bytes and not 4-bytes aligned and <= 40-bytes
        use write32 (direct_ref) to finish the first 32-bytes.
        use write8 (direct_ref) to finish all remaining hanging part.
      - length > 32-bytes and not 4-bytes aligned
        use write32 (indirect_ref) to finish those 4-byte aligned parts.
        use write8 (direct_ref) to finish all remaining hanging part.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

For NFP, we want to re-group a sequence of load/store pairs lowered from
memcpy/memmove into single memory bulk operation which then could be
accelerated using NFP CPP bus.

This patch extends the existing load/store auxiliary information by adding
two new fields:

	struct bpf_insn *paired_st;
	s16 ldst_gather_len;

Both fields are supposed to be carried by the the load instruction at the
head of the sequence. "paired_st" is the corresponding store instruction at
the head and "ldst_gather_len" is the gathered length.

If "ldst_gather_len" is negative, then the sequence is doing memory
load/store in descending order, otherwise it is in ascending order. We need
this information to detect overlapped memory access.

This patch then optimize memory bulk copy when the copy length is within
32-bytes.

The strategy of read/write used is:

  * Read.
    Use read32 (direct_ref), always.

  * Write.
    - length <= 8-bytes
      write8 (direct_ref).
    - length <= 32-bytes and is 4-byte aligned
      write32 (direct_ref).
    - length <= 32-bytes but is not 4-byte aligned
      write8 (indirect_ref).

NOTE: the optimization should not change program semantics. The destination
register of the last load instruction should contain the same value before
and after this optimization.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Add support for emitting commands with field overwrites.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

When immed is used with No-Dest, the emitter should use reg.dst instead of
reg.areg for the destination, using the latter will actually encode
register zero.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

If any of the shift insns in the ld/shift sequence is jump destination,
don't do combination.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

If the mask insn in the ld/mask pair is jump destination, then don't do
combination.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

eBPF insns are internally organized as dual-list inside NFP offload JIT.
Random access to an insn needs to be done by either forward or backward
traversal along the list.

One place we need to do such traversal is at nfp_fixup_branches where one
traversal is needed for each jump insn to find the destination. Such
traversals could be avoided if jump destinations are collected through a
single travesal in a pre-scan pass, and such information could also be
useful in other places where jump destination info are needed.

This patch adds such jump destination collection in nfp_prog_prepare.
Suggested-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

This patch adds support for backward jump on NFP.

  - restrictions on backward jump in various functions have been removed.
  - nfp_fixup_branches now supports backward jump.

There is one thing to note, currently an input eBPF JMP insn may generate
several NFP insns, for example,

  NFP imm move insn A \
  NFP compare insn  B  --> 3 NFP insn jited from eBPF JMP insn M
  NFP branch insn   C /
  ---
  NFP insn X           --> 1 NFP insn jited from eBPF insn N
  ---
  ...

therefore, we are doing sanity check to make sure the last jited insn from
an eBPF JMP is a NFP branch instruction.

Once backward jump is allowed, it is possible an eBPF JMP insn is at the
end of the program. This is however causing trouble for the sanity check.
Because the sanity check requires the end index of the NFP insns jited from
one eBPF insn while only the start index is recorded before this patch that
we can only get the end index by:

  start_index_of_the_next_eBPF_insn - 1

or for the above example:

  start_index_of_eBPF_insn_N (which is the index of NFP insn X) - 1

nfp_fixup_branches was using nfp_for_each_insn_walk2 to expose *next* insn
to each iteration during the traversal so the last index could be
calculated from which. Now, it needs some extra code to handle the last
insn. Meanwhile, the use of walk2 is actually unnecessary, we could simply
use generic single instruction walk to do this, the next insn could be
easily calculated using list_next_entry.

So, this patch migrates the jump fixup traversal method to
*list_for_each_entry*, this simplifies the code logic a little bit.

The other thing to note is a new state variable "last_bpf_off" is
introduced to track the index of the last jited NFP insn. This is necessary
because NFP is generating special purposes epilogue sequences, so the index
of the last jited NFP insn is *not* always nfp_prog->prog_len - 1.
Suggested-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Following steps are taken in the driver to offload an XDP program:

XDP_SETUP_PROG:
 * prepare:
   - allocate program state;
   - run verifier (bpf_analyzer());
   - run translation;
 * load:
   - stop old program if needed;
   - load program;
   - enable BPF if not enabled;
 * clean up:
   - free program image.

With new infrastructure the flow will look like this:

BPF_OFFLOAD_VERIFIER_PREP:
  - allocate program state;
BPF_OFFLOAD_TRANSLATE:
   - run translation;
XDP_SETUP_PROG:
   - stop old program if needed;
   - load program;
   - enable BPF if not enabled;
BPF_OFFLOAD_DESTROY:
   - free program image.

Take advantage of the new infrastructure.  Allocation of driver
metadata has to be moved from jit.c to offload.c since it's now
done at a different stage.  Since there is no separate driver
private data for verification step, move temporary nfp_meta
pointer into nfp_prog.  We will now use user space context
offsets.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

struct nfp_prog is currently only used internally by the translator.
This means there is a lot of parameter passing going on, between
the translator and different stages of offload.  Simplify things
by allocating nfp_prog in offload.c already.

We will now use kmalloc() to allocate the program area and only
DMA map it for the time of loading (instead of allocating DMA
coherent memory upfront).
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Most of offload/translation prepare logic will be moved to
offload.c.  To help git generate more reasonable diffs
move nfp_prog_prepare() and nfp_prog_free() functions
there as a first step.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The register renumbering was removed and will not be coming back
in its old, naive form, given that it would be fundamentally
incompatible with calling functions.  Remove the leftovers.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Only support BPF_PROG_TYPE_SCHED_CLS programs in direct
action mode.  This simplifies preparing the offload since
there will now be only one mode of operation for that type
of program.  We need to know the attachment mode type of
cls_bpf programs, because exit codes are interpreted
differently for legacy vs DA mode.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch supports BPF_NEG under both BPF_ALU64 and BPF_ALU. LLVM recently
starts to generate it.

NOTE: BPF_NEG takes single operand which is an register and serve as both
input and output.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The current ALU_OP_NEG is Op encoding 0x4 for NPF ALU instruction. It is
actually performing "~B" operation which is bitwise NOT.

The using naming ALU_OP_NEG is misleading as NEG is -B which is not the
same as ~B.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Loading 64bit constants require up to 4 load immediates, since
we can only load 16 bits at a time.  If the 32bit halves of
the 64bit constant are the same, however, we can save a cycle
by doing a register move instead of two loads of 16 bits.

Note that we don't optimize the normal ALU64 load because even
though it's a 64 bit load the upper half of the register is
a coming from sign extension so we can load it in one cycle
anyway.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

If stack pointer has a different value on different paths
but the alignment to words (4B) remains the same, we can
set a new LMEM access pointer to the calculated value and
access whichever word it's pointing to.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

To access beyond 64th byte of the stack we need to set a new
stack pointer register (LMEM is accessed indirectly through
those pointers).  Add a function for encoding local CSR access
instruction.  Use stack pointer number 3.

Note that stack pointer registers allow us to index into 32
bytes of LMEM (with shift operations i.e. when operands are
restricted).  This means if access is crossing 32 byte boundary
we must not use offsetting, we have to set the pointer to the
exact address and move it with post-increments.

We depend on the datapath placing the stack base address in
GPR A22 for our use.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As long as the verifier tells us the stack offset exactly we
can render the LMEM reads quite easily.  Simply make sure that
the offset is constant for a given instruction and add it to
the instruction's offset.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When we are performing unaligned stack accesses in the 32-64B window
we have to do a read-modify-write cycle.  E.g. for reading 8 bytes
from address 17:

0:  tmp    = stack[16]
1:  gprLo  = tmp >> 8
2:  tmp    = stack[20]
3:  gprLo |= tmp << 24
4:  tmp    = stack[20]
5:  gprHi  = tmp >> 8
6:  tmp    = stack[24]
7:  gprHi |= tmp << 24

The load on line 4 is unnecessary, because tmp already contains data
from stack[20].

For write we can optimize both loads and writebacks away.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add simple stack read support, similar to write in every aspect,
but data flowing the other way.  Note that unlike write which can
be done in smaller than word quantities, if registers are loaded
with less-than-word of stack contents - the values have to be
zero extended.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Stack is implemented by the LMEM register file.  Unaligned accesses
to LMEM are not allowed.  Accesses also have to be 4B wide.

To support stack we need to make sure offsets of pointers are known
at translation time (for now) and perform correct load/mask/shift
operations.

Since we can access first 64B of LMEM without much effort support
only stacks not bigger than 64B.  Following commits will extend
the possible sizes beyond that.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The need to emitting a few nops will become more common soon
as we add stack and map support.  Add a helper.  This allows
for code to be shorter but also may be handy for marking the
nops with a "reason" to ease applying optimizations.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add support for direct packet access in TC, note that because
writing the packet will cause the verifier to generate a csum
fixup prologue we won't be able to offload packet writes from
TC, just yet, only the reads will work.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch adds ability to write packet contents using pre-validated
packet pointers (direct packet access).
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In direct packet access bound checks are already done, we can
simply dereference the packet pointer.

Verifier/parser logic needs to record pointer type.  Note that
although verifier does protect us from CTX vs other pointer
changes we will also want to differentiate between PACKET vs
MAP_VALUE or STACK, so we can add the check already.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Move data load into a separate function and separate it from
packet length checks of legacy I/O.  This makes the code more
readable and easier to reuse.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Sizes of fields in struct xdp_md/xdp_buff and some in sk_buff depend
on target architecture.  Take that into account and use struct xdp_buff,
not struct xdp_md.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Implement byte swaps with rotations, shifts and byte loads.
Remember to clear upper parts of the 64 bit registers.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Register move operation is encoded as alu no op.  This means
that one has to specify number of unused/none parameters to
the emit_alu().  Add a helper.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Now that we have BPF assemebler support in LLVM 6 we can easily
test all compare instructions (LLVM 4 didn't generate most of them
from C).  Fix the compare to immediates and refactor the order
of compare to regs to make sure they both follow the same pattern.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We optimize comparisons to immediate 0 as if (reg.lo | reg.hi).
The early return statement was missing, however, which means we
would generate two comparisons - optimized one followed by a
normal 2x 32 bit compare.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ld_field instruction has the following format in NFP assembler:

  ld_field[dst, 1000, src, <<24]

reoder parameters to emit_ld_field_any() to make it closer to
the familiar assembler order.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ld_field instruction is a bit special because the encoding uses
two source registers and one of them becomes the output.  We do
need to pass the dst register to our encoding helpers though,
otherwise the "write both banks" flag will not be observed.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Device expects the instructions in little endian.  Make sure we
byte swap on big endian hosts.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>