Mark instructions that use pointers to areas in the stack outside of the
current stack frame, and process them accordingly in mem_op_stack().
This way, we also support BPF-to-BPF calls where the caller passes a
pointer to data in its own stack frame to the callee (typically, when
the caller passes an address to one of its local variables located in
the stack, as an argument).

Thanks to Jakub and Jiong for figuring out how to deal with this case,
I just had to turn their email discussion into this patch.
Suggested-by: NJiong Wang <jiong.wang@netronome.com>
Suggested-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

When pre-processing the instructions, it is trivial to detect what
subprograms are using R6, R7, R8 or R9 as destination registers. If a
subprogram uses none of those, then we do not need to jump to the
subroutines dedicated to saving and restoring callee-saved registers in
its prologue and epilogue.

This patch introduces detection of callee-saved registers in subprograms
and prevents the JIT from adding calls to those subroutines whenever we
can: we save some instructions in the translated program, and some time
at runtime on BPF-to-BPF calls and returns.

If no subprogram needs to save those registers, we can avoid appending
the subroutines at the end of the program.
Signed-off-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Relocation for targets of BPF-to-BPF calls are required at the end of
translation. Update the nfp_fixup_branches() function in that regard.

When checking that the last instruction of each bloc is a branch, we
must account for the length of the instructions required to pop the
return address from the stack.
Signed-off-by: NQuentin Monnet <quentin.monnet@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 is the main patch for the logics of BPF-to-BPF calls in the nfp
driver.

The functions called on BPF_JUMP | BPF_CALL and BPF_JUMP | BPF_EXIT were
used to call helpers and exit from the program, respectively; make them
usable for calling into, or returning from, a BPF subprogram as well.

For all calls, push the return address as well as the callee-saved
registers (R6 to R9) to the stack, and pop them upon returning from the
calls. In order to limit the overhead in terms of instruction number,
this is done through dedicated subroutines. Jumping to the callee
actually consists in jumping to the subroutine, that "returns" to the
callee: this will require some fixup for passing the address in a later
patch. Similarly, returning consists in jumping to the subroutine, which
pops registers and then return directly to the caller (but no fixup is
needed here).

Return to the caller is performed with the RTN instruction newly added
to the JIT.

For the few steps where we need to know what subprogram an instruction
belongs to, the struct nfp_insn_meta is extended with a new subprog_idx
field.

Note that checks on the available stack size, to take into account the
additional requirements associated to BPF-to-BPF calls (storing R6-R9
and return addresses), are added in a later patch.
Signed-off-by: NQuentin Monnet <quentin.monnet@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>

Similarly to "exit" or "helper call" instructions, BPF-to-BPF calls will
require additional processing before translation starts, in order to
record and mark jump destinations.

We also mark the instructions where each subprogram begins. This will be
used in a following commit to determine where to add prologues for
subprograms.
Signed-off-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

The checks related to eBPF helper calls are performed each time the nfp
driver meets a BPF_JUMP | BPF_CALL instruction. However, these checks
are not relevant for BPF-to-BPF call (same instruction code, different
value in source register), so just skip the checks for such calls.

While at it, rename the function that runs those checks to make it clear
they apply to _helper_ calls only.
Signed-off-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

In order to support BPF-to-BPF calls in offloaded programs, the nfp
driver must collect information about the distinct subprograms: namely,
the number of subprograms composing the complete program and the stack
depth of those subprograms. The latter in particular is non-trivial to
collect, so we copy those elements from the kernel verifier via the
newly added post-verification hook. The struct nfp_prog is extended to
store this information. Stack depths are stored in an array of dedicated
structs.

Subprogram start indexes are not collected. Instead, meta instructions
associated to the start of a subprogram will be marked with a flag in a
later patch.
Signed-off-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

In preparation for support for BPF to BPF calls in offloaded programs,
rename the "stack_depth" field of the struct nfp_prog as
"stack_frame_depth". This is to make it clear that the field refers to
the maximum size of the current stack frame (as opposed to the maximum
size of the whole stack memory).
Signed-off-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

In current ABI the size of the messages carrying map elements was
statically defined to at most 16 words of key and 16 words of value
(NFP word is 4 bytes).  We should not make this assumption and use
the max key and value sizes from the BPF capability instead.

To make sure old kernels don't get surprised with larger (or smaller)
messages bump the FW ABI version to 3 when key/value size is different
than 16 words.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Up until now we only had per-vNIC BPF ABI version capabilities,
which are slightly awkward to use because bulk of the resources
and configuration does not relate to any particular vNIC.  Add
a new capability for global ABI version and check the per-vNIC
version are equal to it.  Assume the ABI version 2 if no explicit
version capability is present.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Add support for adjust_tail.  There are no FW changes needed but add
a FW capability just in case there would be any issue with previously
released FW, or we will have to change the ABI in the future.

The helper is trivial and shouldn't be used too often so just inline
the body of the function.  We add the delta to locally maintained
packet length register and check for overflow, since add of negative
value must overflow if result is positive.  Note that if delta of 0
would be allowed in the kernel this trick stops working and we need
one more instruction to compare lengths before and after the change.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Record perf maps by map ID, not raw kernel pointer.  This helps
with debug messages, because printing pointers to logs is frowned
upon, and makes debug easier for the users, as map ID is something
they should be more familiar with.  Note that perf maps are offload
neutral, therefore IDs won't be orphaned.

While at it use a rate limited print helper for the error message.
Reported-by: NKees Cook <keescook@chromium.org>
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NDirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Control queue is fairly low latency, and requires SKB allocations,
which means we can't even reach 0.5Msps with perf events.  Allow
perf events to be delivered to data queues.  This allows us to not
only use multiple queues, but also receive and deliver to user space
more than 5Msps per queue (Xeon E5-2630 v4 2.20GHz, no retpolines).
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NDirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

In preparation for SKB-less perf event handling make
nfp_bpf_event_output() take buffer address and length,
not SKB as parameters.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NDirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Create a higher-level entity to represent a device/ASIC to allow
programs and maps to be shared between device ports.  The extra
work is required to make sure we don't destroy BPF objects as
soon as the netdev for which they were loaded gets destroyed,
as other ports may still be using them.  When netdev goes away
all of its BPF objects will be moved to other netdevs of the
device, and only destroyed when last netdev is unregistered.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Acked-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

NFP doesn't have integer divide instruction, this patch use reciprocal
algorithm (the basic one, reciprocal_div) to emulate it.

For each u32 divide, we would need 11 instructions to finish the operation.

  7 (for multiplication) + 4 (various ALUs) = 11

Given NFP only supports multiplication no bigger than u32, we'd require
divisor and dividend no bigger than that as well.

Also eBPF doesn't support signed divide and has enforced this on C language
level by failing compilation. However LLVM assembler hasn't enforced this,
so it is possible for negative constant to leak in as a BPF_K operand
through assembly code, we reject such cases as well.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: NSong Liu <songliubraving@fb.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

NFP supports u16 and u32 multiplication. Multiplication is done 8-bits per
step, therefore we need 2 steps for u16 and 4 steps for u32.

We also need one start instruction to initialize the sequence and one or
two instructions to fetch the result depending on either you need the high
halve of u32 multiplication.

For ALU64, if either operand is beyond u32's value range, we reject it. One
thing to note, if the source operand is BPF_K, then we need to check "imm"
field directly, and we'd reject it if it is negative.  Because for ALU64,
"imm" (with s32 type) is expected to be sign extended to s64 which NFP mul
doesn't support. For ALU32, it is fine for "imm" be negative though,
because the result is 32-bits and here is no difference on the low halve
of result for signed/unsigned mul, so we will get correct result.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: NSong Liu <songliubraving@fb.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

NFP verifier hook is coping range information of the shift amount for
indirect shift operation so optimized shift sequences could be generated.

We want to use range info to do more things. For example, to decide whether
multiplication and divide are supported on the given range.

This patch simply let NFP verifier hook to copy range info for all operands
of all ALU operands.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: NSong Liu <songliubraving@fb.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

The two fields are a copy of umin and umax info of bpf_insn->src_reg
generated by verifier.

Rename to make their meaning clear.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: NSong Liu <songliubraving@fb.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

For indirect shifts, shift amount is not specified as constant, NFP needs
to get the shift amount through the low 5 bits of source A operand in
PREV_ALU, therefore extra instructions are needed compared with shifts by
constants.

Because NFP is 32-bit, so we are using register pair for 64-bit shifts and
therefore would need different instruction sequences depending on whether
shift amount is less than 32 or not.

NFP branch-on-bit-test instruction emitter is added by this patch and is
used for efficient runtime check on shift amount. We'd think the shift
amount is less than 32 if bit 5 is clear and greater or equal than 32
otherwise. Shift amount is greater than or equal to 64 will result in
undefined behavior.

This patch also use range info to avoid generating unnecessary runtime code
if we are certain shift amount is less than 32 or not.
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>

BPF has access to all internal FW datapath structures.  Including
the structure containing RX queue selection.  With little coordination
with the datapath we can let the offloaded BPF select the RX queue.
We just need a way to tell the datapath that queue selection has already
been done and it shouldn't overwrite it.  Define a bit to tell datapath
BPF already selected a queue (QSEL_SET), if the selected queue is not
enabled (>= number of enabled queues) datapath will perform normal RSS.

BPF queue selection on the NIC can be used to replace standard
datapath RSS with fully programmable BPF/XDP RSS.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Add support for the perf_event_output family of helpers.

The implementation on the NFP will not match the host code exactly.
The state of the host map and rings is unknown to the device, hence
device can't return errors when rings are not installed.  The device
simply packs the data into a firmware notification message and sends
it over to the host, returning success to the program.

There is no notion of a host CPU on the device when packets are being
processed.  Device will only offload programs which set BPF_F_CURRENT_CPU.
Still, if map index doesn't match CPU no error will be returned (see
above).

Dropped/lost firmware notification messages will not cause "lost
events" event on the perf ring, they are only visible via device
error counters.

Firmware notification messages may also get reordered in respect
to the packets which caused their generation.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

For asynchronous events originating from the device, like perf event
output, we need to be able to make sure that objects being referred
to by the FW message are valid on the host.  FW events can get queued
and reordered.  Even if we had a FW message "barrier" we should still
protect ourselves from bogus FW output.

Add a reverse-mapping hash table and record in it all raw map pointers
FW may refer to.  Only record neutral maps, i.e. perf event arrays.
These are currently the only objects FW can refer to.  Use RCU protection
on the read side, update side is under RTNL.

Since program vs map destruction order is slightly painful for offload
simply take an extra reference on all the recorded maps to make sure
they don't disappear.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Comparison instruction requires a subtraction.  If the constant
is negative we are more likely to fit it into a NFP instruction
directly if we change the sign and use addition.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

NFP has a prng register, which we can read to obtain a u32 worth
of pseudo random data.  Generate code for it.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>

Allow atomic add to be used even when the value is not guaranteed
to fit into a 16 bit immediate.  This requires the value to be pulled
as data, and therefore use of a transfer register and a context swap.

Track the information about possible lengths of the value, if it's
guaranteed to be larger than 16bits don't generate the code for the
optimized case at all.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>

Implement atomic add operation for 32 and 64 bit values.  Depend
on the verifier to ensure alignment.  Values have to be kept in
big endian and swapped upon read/write.  For now only support
atomic add of a constant.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>

Support calling map_delete_elem() FW helper from the datapath
programs.  For JIT checks and code are basically equivalent
to map lookups.  Similarly to other map helper key must be on
the stack.  Different pointer types are left for future extension.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>

Support calling map_update_elem() from the datapath programs
by calling into FW-provided helper.  Value pointer is passed
in LM pointer #2.  Keeping track of old state for arg3 is not
necessary, since LM pointer #2 will be always loaded in this
case, the trivial optimization for value at the bottom of the
stack can't be done here.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>

Our implementation has restriction on stack pointers for function
calls.  Move the common checks into a helper for reuse.  The state
has to be encapsulated into a structure to support parameters
other than BPF_REG_2.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>

This patch is the front end of this optimisation, it detects and marks
those packet reads that could be cached. Then the optimisation "backend"
will be activated automatically.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>

This patch assumes there is a packet data cache, and would try to read
packet data from the cache instead of from memory.

This patch only implements the optimisation "backend", it doesn't build
the packet data cache, so this optimisation is not enabled.

This patch has only enabled aligned packet data read, i.e. when the read
offset to the start of cache is REG_WIDTH aligned.
Signed-off-by: NJiong Wang <jiong.wang@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>

Commit 84ce5b98 ("scripts: kernel-doc: improve nested logic to
handle multiple identifiers") improved the handling of nested structure
definitions in scripts/kernel-doc, and changed the expected format of
documentation.  This causes new warnings to appear on W=1 builds.

Only comment changes.
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>

Use the recently added extack support for eBPF offload in the driver.
Signed-off-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

If an eBPF instruction is unknown to the driver JIT compiler, we can
reject the program at verification time.
Signed-off-by: NQuentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NJiong Wang <jiong.wang@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Plug in to the stack's map offload callbacks for BPF map offload.
Get next call needs some special handling on the FW side, since
we can't send a NULL pointer to the FW there is a get first entry
FW command.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Verify our current constraints on the location of the key are
met and generate the code for calling map lookup on the datapath.

New relocation types have to be added - for helpers and return
addresses.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Parse helper function and supported map FW TLV capabilities.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

Implement calls for FW map communication.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>

For map support we will need to send and receive control messages.
Add basic support for sending a message to FW, and waiting for a
reply.

Control messages are tagged with a 16 bit ID.  Add a simple ID
allocator and make sure we don't allow too many messages in flight,
to avoid request <> reply mismatches.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NQuentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>