Every probe of a misbehaving switch so far measured it from the
outside. This puts the witnesses inside.
Each idle() pass increments a loop counter and stamps a checkpoint
after every phase: link polling, SFP, RX, TX, STP and the command
dispatch. A phase spanning two or more system ticks counts as slow and
its worst time is kept, in 5 ms units. handle_rx counts frames, so the
dump separates a flooded CPU from a stalled loop.
At boot the free stack area is painted, preserving the interrupt state
around the painting, and the dump reports how much of the pattern
survives next to the live stack pointer. The TCP connection table is
printed whole: state, timers, retransmissions, ports and peer, next to
the httpd transfer state, so a permanently occupied slot names its
occupant. With more than one slot configured every slot is listed.
Everything is raw hex; two dumps a known time apart give the rates.
The implementation sits in cmd_parser next to the other command
handlers and needs no storage beyond its counters.
Building with SDCC 4.6.0 emits tons of warnings e.g.:
cmd_parser.c:31: warning 356: object in read-only code space should be const
cmd_parser.c:36: warning 357: pointer to object in read-only code space should be pointer to const
These of course become fatal when building with --Werror.
Comparing resultant binaries, it seems that a single string is shifted
(deduplication?) but the numbers emitted at the end of the build are
otherwise identical.
idle() went to sleep before it looked at the receive buffer, so a packet that
arrived just after the loop dozed off waited for the next tick before anything
touched it. Every TCP round-trip paid that toll, which is why the web UI felt
sluggish even though no phase of the loop was anywhere near busy.
Check the receive buffer first and only sleep when it is empty. The read takes
its result straight from SFR_DATA_U16 rather than going through reg_read_m()
and the sfr_data[] copy, which is what the rest of the tree already does.
Measured on SWTGW218AS with this change alone:
web UI, median of 15 requests for the main page 88 ms -> 76 ms
CPU port under a one-way UDP stream 159 -> 230 packets/s
The loop still sleeps normally when there is nothing to do, so this does not
turn the core into a busy-wait: loops/ticks stays at 0.99 on an idle switch,
and rises above 1.0 only while packets are actually queued.
Costs 12 bytes of code and no extra XDATA.
handle_rx() picked up exactly one packet per iteration of the main loop. Since
the loop sleeps until the next tick, that capped the CPU port at roughly one
packet per tick no matter how little work it had to do. A burst of broadcast,
ARP or IGMP traffic aimed at the CPU queued up behind that and pushed out the
STP and LACP timers.
Drain up to RX_BUDGET packets in one pass instead. The budget matters: without
one a sustained flood would spin here and starve everything else in the loop.
The two register reads in the loop take their result straight from
SFR_DATA_U16 rather than going through reg_read_m() and the sfr_data[] copy,
which is what the rest of the tree already does.
Measured on SWTGW218AS with a one-way UDP stream, counting the packets the CPU
actually processed, with this change alone:
before 159 packets/s
RX_BUDGET 4 245 packets/s, loop iteration ~16 ms
RX_BUDGET 8 298 packets/s, loop iteration ~26 ms
Throughput is already most of the way there at 4, so the larger budget mostly
buys a longer worst-case delay for the other phases. STP stayed responsive in
both cases.
Costs 71 bytes of code and no extra XDATA.
boot.c calls it from BANK3 and only rtlplayground.c had a prototype, so
the compiler saw an implicit declaration. The call itself was fine, the
function lives in the common area, but the warning was real. Move the
prototype into the shared header.
Reported by vDorst.
sfp.c takes handle_sfp, sfp_module_read, sfp_apply_quirks, sfp_get_info,
sfp_read_field, sfp_print_info, setup_sfp_gpio, the quirk table and the module
state that goes with them. Everything exported is __banked: handle_sfp() and
sfp_print_info() are reached from cmd_parser.c in BANK2, the rest from the
common area, and both are bank boundaries.
The quirk table needed checking rather than assuming. Its entries are __code
pointers, and a __code pointer carries no bank, so passing one to a function in
a different bank hands it an offset read through the wrong window. That is safe
here because strcmp() sits in the common segment at 0x51f and is reached with a
plain lcall, so the bank register still says 3 while it runs. Verified in the
built image: the table at 0x35580 holds 0x563c and 0x5644, which resolve to
"QSFPTEK" and "QT-SFP+-T" inside BANK3.
Common segment 13393 -> 8432 bytes, leaving 7492 free rather than 2531. BANK3
holds 5710, BANK1 is unchanged and BANK2 grows by 31 for the trampolines.
All 34 machines in machine.h build and link, and the same walk over the
generated assembly finds no plain lcall crossing a bank boundary.
The common segment is the scarce one: sixteen kilobytes shared by every bank,
and before this it had about 2.5 kB left. Most of what moves here runs once at
startup and does not have to sit in it. The two SerDes helpers are the
exception: the link-change handler and the SFP polling both reach sds_config()
while the switch is running, so they are not startup-only. None of it is
reached from an interrupt.
boot.c takes setup_i2c, init_smi, nic_setup, rtl8373_revision, rtl8224_enable,
early_boot_handle_button, set_hostname_default and the two SerDes helpers.
BANK3 is enabled with the linker flag; nothing else in the build needed
changing.
Everything moved is declared __banked, since every caller is outside BANK3:
main() and friends run from the common area and rtl837x_init.c calls
sds_config() and sds_config_mac() from BANK2. The calls the moved code makes
outward were already banked, so that side needed no work.
Common segment 13393 -> 10137 bytes, leaving 5787 free rather than 2531.
BANK3 holds 3732. All 34 machines in machine.h build and link. Walking
the generated assembly for plain lcalls that cross a bank boundary finds none
among 2808 calls; those build cleanly and only fail at run time, so it seemed
worth checking by machine rather than by eye.
The lookup tables live in BANK1, so the routine switches the code bank to
reach them with movc and never switches it back. That is the contract the
__naked declaration already implied: same bank or common segment only.
The name now says it.
crc16_bank1 is the routine unchanged. crc16 becomes a __banked wrapper in
BANK1, so a caller from another bank has its own bank put back by
__sdcc_banked_ret. The two existing call sites take the direct path and
pay nothing.
isr_ext2() ends with PCON |= 1, putting the core to sleep from inside an
interrupt handler.
The DW8051 leaves idle when any enabled interrupt is activated, so the core
does not stay down for long. It naps until the next enabled interrupt request
turns up, with the 200 Hz system tick putting a ceiling of 5 ms on that, and
then carries on.
The sleep buys nothing either way, because the instruction after it is the
return. While it lasts the handler holds the priority latch, so everything at
its level or below waits, and that includes the tick, the console and the NIC
receive interrupt.
The handler still clears its flag, which is what it has to do. The bit numbers
beside the two masks were wrong: 0xef clears bit 4 and 0xdf clears bit 5,
not 7 and 6.
isr_ext2() ends with PCON |= 1, putting the core to sleep from inside an
interrupt handler.
The DW8051 leaves idle when any enabled interrupt is activated, so the core
does not stay down for long. It naps until the next enabled interrupt request
turns up, with the 200 Hz system tick putting a ceiling of 5 ms on that, and
then carries on.
The sleep buys nothing either way, because the instruction after it is the
return. While it lasts the handler holds the priority latch, so everything at
its level or below waits, and that includes the tick, the console and the NIC
receive interrupt.
The handler still clears its flag, which is what it has to do. The two wrong
comments next to the masks go as well: 0xef clears bit 4 and 0xdf clears bit 5,
not 7 and 6, and the masks already say that without help.
Add a `mac <aa:bb:cc:dd:ee:ff>` console command (and a bare `mac` to show
the current address) that overrides the management MAC at runtime. The
address is validated (unicast, globally administered, non-zero OUI) and
the static L2 management entry is re-registered so the change applies
without a reboot. Because it is a normal command, a line in the startup
config persists it across boots: execute_config() runs it before the final
static L2 entry and the derived hostname are set up.
Also point the keepLink KP-9000-9XHML-X V3.1 machine definition at the
factory MAC area (0x1FC000, same location other machines use) so a valid
factory MAC is picked up at boot instead of falling back to a locally
administered address.
The MAC parsing/printing lives in parse_mac()/print_mac() helpers next to
parse_ip()/print_ip() so future commands can reuse them, and a memcmp()
helper joins the project's own memcpy()/memset() (no libc is linked).
isr_ext2() ends with PCON |= 1. On the DW8051 any enabled interrupt request
ends idle, so the handler sleeps until the next request arrives, the system
tick at the latest, and only then returns. Everything at the same priority
waits with it, and the instruction after the idle is RETI, so the sleep buys
nothing.
The handler still clears its flag, which is what it has to do. The other four
uses of PCON |= 1 are in task context and are left alone.
uip_arp_ipin() is an empty macro with its body under #if 0, both inherited
from the uIP import and untouched since. That leaves uip_arp_update()
reachable only from uip_arp_arpin(), so an entry is refreshed only by ARP
addressed to us and ages out after UIP_ARP_MAXAGE no matter how much traffic
passed over it.
When it goes, the next transmission to that host hits the miss path of
uip_arp_out(), which overwrites the packet in uip_buf with an ARP request and
returns. For TCP that costs a retransmission. For UDP nothing retransmits, and
handle_tx() sends the UDP connections through uip_arp_out() too, so the
datagram is lost.
The body could not be enabled as written: uip_len -= sizeof(struct uip_eth_hdr)
is a leftover from when UIP_LLH_LEN was 0. Here it is ETHER_HEADER_SIZE +
RTL_FRAME_DESC_SIZE and uip.c reads the IP header at uip_buf[UIP_LLH_LEN], so
uip_len has to keep the link layer in it. The rest of the function is sound.
The EX0..EX3 handlers each emitted a single debug character through
write_char(), which busy-waits on the serial TX-empty flag that is set by
the serial interrupt. If one of these handlers runs while a byte is being
transmitted, it waits for a flag that the equal-or-lower-priority serial
interrupt cannot deliver, hanging the handler with the system tick stopped.
The characters were only ever debug markers, so drop them; the handlers now
just acknowledge the interrupt. No interrupt service routine should call the
blocking console path.
uip_arp_timer() was defined but never called, so ARP entries were only
refreshed on new learning and only removed under table pressure. A cached
entry for a peer whose MAC later changes (a gateway failover, a replaced
device) is then used until it is evicted, and traffic to anything reached
through it is sent to the old MAC and lost, while on-link traffic is
unaffected.
uIP expects the timer roughly every ten seconds; drive it from the
once-per-second housekeeping in idle().
The CPU's own MAC exists in the L2 table only as a learned entry,
created whenever the CPU transmits, and it ages out like any other
dynamic entry. Unknown unicast is not flooded to the CPU port, so on an
idle switch (no web session, no management traffic) the entry expires
after the aging time and frames for the management IP no longer reach
the CPU: pings die until the next CPU transmission (an ARP reply)
re-teaches the entry, then die again. Real management sessions keep the
entry alive continuously, which is why this is only visible on idle
switches.
Install a static entry for the management MAC on the CPU port after the
startup config has been replayed, so it lands in the final management
VLAN, and move it when the management VLAN is changed at runtime.
Entry mechanics, established on hardware: a table write with read-method
0 hashes the key in TBL_DATA_IN and places the entry itself (a search
returns index 0 on a miss, there is no software-visible insertion slot).
The age field in the third data word must be non-zero for the entry to
be visible to lookup at all, and the static flag (bit 0 of byte 1 of
that word) stops the aging engine from decrementing it: with the flag a
full-age entry survives untouched where an identical unflagged entry
ages to invisible within the aging period. With no management VLAN
configured the entry lives in VLAN 1.
Three ways the answer could mislead the caller, all in the paths that
matter most when something is wrong.
The capture sat in write_char(), so anything printed through
print_string_no_syslog() escaped it. That is exactly the replies of the
syslog commands, which deliberately avoid the syslog path: "syslog on"
answered with the bare address and no sentence around it, "syslog off"
answered with nothing at all. Moving the capture one level down, into
write_char_no_syslog(), catches both paths and costs nothing, since
write_char() reaches it anyway.
On a parse error send_bad_request() rewrote the buffer from the start,
throwing away the very text that explains the error and replacing it
with the 400 HTML page. Keeping the captured text and restating only the
status keeps the parser's own message; "400 NO" is as long as "200 OK",
so the header does not have to be rebuilt.
An overlong listing was cut at the buffer end with nothing to show for
it, so a clipped answer read as a short one. It now ends in a marker,
and the reserved room is accounted for in the guard.
While here: a command that configures something prints nothing, and an
empty body reads the same as "nothing ran", so a successful silent
command now answers "OK".
No internal RAM and no xdata; 174 bytes of BANK1, which has 7356 left.
POST /cmd runs the command and answers with an empty 200, so a caller cannot
tell "done" from "nothing happened", and whatever the command prints only
reaches the serial console. Reading a register over the network needs a second
channel just to see the answer.
write_char() is the one funnel every console helper goes through, and it already
tees into the syslog buffer. A third sink, live only while the httpd runs a
command that arrived over the network, puts the same bytes into the response
behind the header. Serial and syslog keep getting their copy.
The capture stops at the end of the response buffer, so a long listing is
truncated rather than overrunning it.
BANK1 grows 39 bytes and the common area 52. One byte of xdata, BANK2 unchanged.
stp_enabled is a flag, so say bool. cmpBytes() returns a comparison
result, so say int8_t. The three busy waits this branch adds read the
status straight out of the SFR instead of copying four bytes to xdata
first.
tcpip_output() decides whether to splice in the 802.1Q tag, and that
decision also moves the frame: the header is shifted forward over its
padding, so a tagged frame starts at uip_buf with the q_frame layout,
while an untagged one keeps the padding and the nonq_frame layout.
nic_tx_packet() took that decision a second time, from management_vlan
alone. That agreed while the sender suppressed the management VLAN around
the transmission, but not since the tag is skipped per frame for a
CPU-tagged one: the descriptor then still sits behind the padding while
the transfer is set up for the shifted layout. The frame goes out four
bytes early and its length is read from the offset where tx_seq and
chksum_flags live, so a sixty byte BPDU is sent as around 1800 bytes of
whatever follows it. Looped back it is no longer a BPDU, the port hears
nothing, auto edge promotes it and the loop stays open.
Record the decision where it is taken and let the transfer follow it.
A variable rather than an argument because internal RAM is full once the
aggregation module shares the image: the overlay area ends at 0x7f, and
an argument or any temporary for the condition no longer fits.
The port was frozen at 514 in the call that opens the connection, so a
collector listening anywhere else was simply out of reach, and plenty do,
because 514 is privileged and containerised collectors usually get a high
port instead. Working around it meant putting a proxy in front, which then
rewrites the source address and hides which device actually sent the line.
Add `syslog port <n>`. The default stays 514, so nothing changes for an
existing configuration. Setting it tears the connection down and builds a
new one, the same dance the address already does, because uip_udp_new()
freezes the port into the connection.
The value is parsed here rather than through atoi_short(), which reports
"no digits" but not overflow: 65540 would quietly wrap to 4 and we would
open a connection to a port nobody asked for. Checking before each multiply
keeps it in 16-bit arithmetic. Zero then covers all three ways to be wrong -
no digits, past 65535, or an explicit "0", which is not a port either.
Whitelist the new command in config.js as well, next to `syslog ip`, or the
line is dropped without a word when the configuration is saved and the
setting quietly fails to survive a reboot. The dedup entry has to precede
the general `^syslog\b` one, otherwise it would deduplicate against
`syslog on`.
itoa_short() is new: itoa() only handles a byte and print_short() prints
hex, so there was no way to show a port as a decimal number. Kept separate
from itoa() rather than widening it, because every existing caller passes a
byte and would start paying for 16-bit divisions it does not need.
The scratch variables for the parser live in xdata. Locals there would take
internal RAM, and this image has none left: data ends at 0x77, the overlay
occupies 0x78-0x7f and the stack starts at 0x80. Measured: no module's DSEG
grew.
Both conflicts are the register keyword removal meeting a change of ours,
so each takes main's signature and keeps what the branch was saying:
nic_rx_packet() still returns a bool, and port_ingress_filter_get() stays
declared.
main has since put its own copy of the STP module in BANK2, so the pragma
this branch carried is now duplicated. The one with main's comment stays.
cmd_parser.c conflicted twice: main rewrote the pvid command around the
new port-separator and atoi helpers, next to the line where this branch
delegates "stp" to stp_parse(). Both belong, so the STP delegation keeps
main's pvid body.
The linker caught what the merge could not see: main changed atoi_byte()
to return the digit count and leave the value in atoi_results_u8, where
it used to return non-zero on failure and write through a pointer. Note
the sense is inverted, so the three calls in rtl837x_stp.c are adjusted
rather than just re-arranged.
Only httpd.c conflicted: main added the buffered configuration upload
next to the pointers this branch had moved into xdata to free internal
RAM. Both belong, so the new globals sit above declarations that keep
their storage class.
The I2C controller transfers up to sixteen bytes per transaction and
page_impl.c already used that for sfp_send_data(), while sfp_read_reg()
asked for one byte and every caller looped. Reading a module therefore
cost one address phase per byte: 87 transactions when a module is
inserted, 52 for the sfp command, 36 for the vendor block in status.json.
sfp_read_block() replaces sfp_read_reg() and the callers that already
wanted a run of registers ask for it once: the vendor fields as three
16 byte pages, the diagnostics as one transfer, rate and encoding
together. That drops the three paths above to 8, 6 and 3 transactions,
and sfp_send_data() loses its copy of the transfer.
The vendor loops now run over 16..63 rather than 20..59 so the page base
is a multiple of sixteen and the index into the buffer is a single AND.
The four extra bytes at each end are read and discarded. The diagnostics
read asks for 16 bytes rather than the 15 it uses, because 16 is a width
the shipped firmware already exercises and 15 is not.
The device address, the bus selection and the start bit go into the
control register in one write now that the memory address is written
first, so a transfer touches that register once instead of three times.
The register reads take their result from the SFRs directly rather than
through the sfr_data mirror. The result is a bool and the destination is
sfp_buf, so a caller that cares about a failed transfer looks at the
return value instead of a flag.
Every caller gives up on the first failed read rather than carrying a
flag to the end, which is why the module read moved out of handle_sfp
into a function of its own. A module whose read fails is left marked as
absent, so the next poll retries it instead of configuring the SerDes
from bytes that never arrived.
BANK1 -194 bytes, BANK2 +382, common segment +44, xdata +15 for the
buffer, and one byte more of internal RAM free than before the series.
Built for all 25 machine definitions on sdcc 4.5.0; the tightest common
segment is 98 bytes free on SWTG024AS_V2_0, against 54 before this
series.
sfp_read_reg() sat in rtlplayground.c, so it occupied the common 16 KB
window that every bank shares, even though nothing outside the SFP paths
calls it. That window is the tightest resource in the image:
SWTG024AS_V2_0 and SWTG024AS_A_2_0_1_5C_1SFP had 54 bytes left in it.
rtl837x_pins.c is already in BANK2 and already holds the I2C bus helpers
this function calls, so the transfer belongs there. The function moves
verbatim and becomes __banked; the prototype in rtl837x_common.h says so,
which is what keeps the callers in BANK1 and BANK2 honest.
No behaviour change. The common segment gains 200 bytes on every machine:
159 to 359 free on SWTGW218AS, 54 to 254 on the two variants above.
BANK1 +6 bytes, BANK2 +336. Built on sdcc 4.5.0.
Saving the configuration works in Chrome and fails in Firefox, and the
difference is only how each browser splits the request. The handler
erased the config sector as soon as the request started and then parsed
the multipart body one TCP segment at a time, which requires every
boundary and every part header to fall inside a single segment. Firefox
splits inside a part header, so the parser lost its place and the
erased sector was left holding a truncated body or nothing at all. A
single-burst scripted post lost the whole body the same way.
The configuration is limited to two kilobytes, so the whole request body
now accumulates in xdata and is parsed only after the closing boundary
has arrived. The parts are walked in one pass, the part carrying a
filename is written to a freshly erased sector, and the client receives
a 200 instead of the previous silent close. No segmentation can confuse
this, since the parser only ever sees a complete body.
Locating the closing boundary first also bounds the walk over the parts,
since none can lie beyond it, so the length of the buffer is no longer
the bound and the test for the two trailing dashes is unnecessary.
The walk matches at offsets inside a buffer that is not terminated, so
neither existing helper fits: strcmp() goes on to compare the byte after
the match and is_word_x() demands a separator there. Add strstart() and
strstart_x() for that case, and use strlen_x() for the boundary length.
The firmware upload path still streams, since a megabyte cannot be
buffered, and is untouched.
The bounded waits were silent: on timeout the code went straight back to
its caller, and handle_rx() then read a frame the DMA may never have
delivered, which is worse than waiting longer.
Each of the three transfers now says so on the console, the two RX ones
report failure to handle_rx(), and handle_rx() acknowledges the packet
and gives up on it rather than parsing whatever is in the buffer. The
guard variable moved to the top of its function, so the block that held
it and its indentation are gone.
The comment said the ASIC never completes a TX when the egress port is in a
non-forwarding MSTP state. The guard is worth keeping either way, since an
unbounded spin in the DMA wait takes the whole main loop down, but the
mechanism is more than the evidence supports and the DMA into the TX ring
has no business knowing the egress port at all.
d00f
Matt Merhar
René van Dorst
gnacho
bloqaudio
logicog