Alessandro Rubini authored 8 years ago and Adam Wujek committed 8 years ago

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This commit removes the requirement for ARP frames to be broadcast
in the no-vlan case.  The simplified rule-set of a vlan setup already
accepts ARP unicast, and no harm is done.

Although we do not make ARP  requests so we won't receive unicast
replies, hosts do send unicast requests, to refresh an ARP entry
that is going to expire.

The next commit enacts this feature in software, as a side effect
of supporting pdelay PTP.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This commit adds the Kconfig mechanism and the pfilter rules.

The latency test socket transmists frames of a specific ether type;
the receiver must receive all such frames.  To minimize the load on the
packet filter, this adds a single rule: we already accept all PTPv2
frames, so we now "OR" to this the new ethernet type.   If the feature
is configured off, the rule uses the same PTPv2 ethertye, thus
turning in practice into a NOP.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

One rule only!

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This shortens the rule-set by one rule, making space for pdelay.

IP is acceptable whether unicast or broadcast. Thus, remove the IP_UNI
intermediate calculation and use only IP_OK.

Most of the lines in this commit are a renaming of FRAME_OUR_MAC to
FRAME_MAC_OK, with no real effect, because we'll coalesce all valid
macs in a single bit (I planned to use it in this commit, but I did
not, still I don't want to split the patch).

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This shortens the rule-set by one rule, making space for pdelay.

There is no need to use a temporary register once FOR_CPU is done.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This shortens the rule-set by two rules, making space for pdelay.

By accepting all ports from 0 to 511, we get all of our services,
including snmp that will be tested soon.  This relaxed check will give
us possibly some false positives (then discarded by the CPU) while
still protecting from most of an hostile or just probing  port sweep.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

These numbers are only available at build time; there is no run-time
command to change them as yet.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

WARNING: wr_switch_defconfig doesn't build, this is fixed in next commit

Unfortunately this commit makes the internal communication between
minic and lan asymmetric: on receive the tag is discarded (and the
longer header falls into the payload) while on transmit it must be
provided by lan.c.

The reason is that on receive we can trim 4 bytes from the payload,
but on send we can't add 4 bytes without a memmove.  The functions
receive pointers to two different header structures, so hopefully
user errors will be signalled by the compiler.

Still, users interact with net.c, which hides vlan completely from
them.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This changes the pfilter rule-set, to make it ready for vlan addition.
Moreover, it removes the choice between three rule-sets and uses one
that works for everyone (we'll add the vlan option later on).

Moreover, we prepare for the option of peer-delay PTP and UDP-based PTP.

As a side effect, one Kconfig option is removed.

We know for sure that what goes to the "streamer" protocol
is re-checked, so the streamer can get some "everything else", exactly
like the 7S wr-nic (that sends to the host everything unused).

This is the current rule set:

     - everything tagged is dropped
     - CPU: arp broadcast, not unicast (i.e. requests only).
     - CPU: PTP ethtype. Any mac address.
     - CPU: ICMP unicast, not broadcast.
     - CPU: UDP (uni/multicast), ports  bootpc, ptp-event, ptp-general.
     - Etherbone (class 7): UDP (uni/multicast) and port 0xebd0
     - Other fabric (class 6): anything not Etherbone (inverted class 7 bit)

This counts up to 28 rules. We have 4 more rules available and we may
add options, either at build time or run time to add udp ports or
other ethertypes.  Or not...

2 caveats:

  - frames for the CPU will have class bits 0x41 instead of 0x01, because
  everything not etherbone will receive bit 6. This is not a problem, as
  the CPU is not checking the class bits.

  - the wr-nic gateware, when using this new sw code base, must be
  changed to use classes 6 and 7 like everybody else, not classes 7
  and 5.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The description of class management was not completely clear or correct,
so it's now rewritten with some references to gateware files.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Actually, the "new" information is very little.  But now the
explanation of how it works fits in 80 columns (I personally need
this) and is laid out a little better.  Also, it describes the new
situation, activated by the next commits.

No code change at this point.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

If the Kconfig option is not set, there's no effect on code size.
If set, .bss increases by 256 bytes (the buffer) and .text by
another 200 bytes.

This adds timestamping support, so we know the duration of the
interrupt (because we found it's an important thing when problems
occur).

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Accessing data structures in the host to get target fields is broken:
we may have different data size or alignment, or whatever. The hack
of using "-m32" to be able to access lm32 fields from x86-64 is not
clean (we imply the i386 and target have the same size/alignment).
Even worse, not everybody has an x86-64 cross-compiler for i386.

Thus, we use OFFSET_OF in the target, with the lm32 cross compiler,
to build a binary table that is then accessed by the host.

In order to link the data structure in the host we use the assembler
instead of objcopy. With objcopy we'd need to spell out the bfd name
of the host, and we'd get horrible names like
   struct dump_info _binary_dump_info_host_bin_start[];

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Use an environment variable to check for spec byte order, and
fix the buffer before using it.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

It's still suboptimal, we need to autodetect with a in-binary helper,
but by now I'm analyzing old logs so I'd better not change wrpc
itself.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This fixes the problem with data sizes when building on a 64-bit host.
Now the dump of ppi makes sense, besides endianness.

Please note how this requires a cross-compile environment for i386
when you build on x86-64. Later on this requirement is lifted.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The tool is meant to print internal data structures of the wrpc,
by running either on a dump file or a live system.

The code is an edited copy of wr-switch-sw/userspace/tools/wrs_dump_shmem.c,
so the data-dumping mechanism is from there.

bugs at this point:
   - all endianness stuff is bad (in wrs it was local, here we cross-dump)
   - data sizes are wrong (see next commit)
   - we miss checks for consistency
   - we can't open /dev/mem (for spec for example) or a elf file
   - docs are missing too

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This just adds "void" in prototypes. I did the pfilter on a different
branch, and lost these changes of commit be2bd234

 wyhile moving pfilter
generation to tools/

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This changes pfilter-builder.c so it really build in user space, but removing
register setting and creating output files instead.

Running this program creates 3 files, that correspond to the previous
three #ifdef conditions.  The files are:

      rules-plain.bin    this is used for the no-etherbone configuration
      rules-ebone.bin    etherbone setup
      rules-e+nic.bin    etherbon plus 7solution's wr-nic packet filter

The files begin with a magix 0x11223344 word, that allows the soft-core
to fix any endianness difference (so no hairy mishaps are expected when
switching to a different soft-core). Then the 40-bit command words are
saved as 64-bit vaules, LSB-first. The output file is thus an odd number
of words and no 64-bit alignment is required.

The first three instructions of the packet filter are used to compare
the destination mac address of the frame.  We now use a fake mac address,
and the LM32 code will change it while programming the binary.

Please look at this commit while ignoring white-space, as it changes
indentation while turning #ifdef/#else into if()/else .

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

We are going to have a user-space tool to build the packet filter
images. The first step is copying dev/ep_pfilter.c, so we can
"git diff" the changes.  At this point the file is not
even built.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Because the golden bitstream for SPEC has syscon at offset 0x0.