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.
CONFIG_UPLOAD_BUF at 2560 capped a configuration upload at about 2.2K
while the config sector holds 4K. Size the buffer as CONFIG_LEN plus
room for the multipart framing so the whole sector is usable.
config_take() wrote the payload without checking it against the sector:
safe before only because the buffer could not hold an oversized one.
Reject a payload that does not fit CONFIG_LEN, terminator included,
instead of writing past the erased sector.
send_config() streams a configuration larger than the TCP output buffer
through the cont_addr/cont_len continuation, but set cont_addr to the
offset within the config instead of a flash address, so the tail was read
from code space. The file server sets cont_addr absolute; do the same by
adding CONFIG_START.
The bug was unreachable while CONFIG_UPLOAD_BUF capped uploads below
TCP_OUTBUF_SIZE, but a stored config near the full sector exposes it.
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 single-digit arm of the ingress parser guards with p - '1' > 9,
which no digit can satisfy: the largest, '9', gives 8. The digit that
needed rejecting is '0', which gives -1 and indexes one byte before
phys_to_log_port, so "ingress 0 t" reads out of bounds and applies the
ingress mode to whatever port number that byte happens to contain.
Ports are 1-based, so reject anything below '1'; values above '9' are
already excluded by the isnumber check before this.
The MAC table listing tests bit 0 of byte 2 of the third table data
word for the static flag, but the flag lives in bit 0 of byte 1: an
entry written with byte 1 bit 0 set survives the aging engine
indefinitely where an identical entry without it ages out, and reads
back with exactly that bit set through both the address and the
next-entry read methods. Byte 2 of that word reads zero for learned and
static entries alike, so every entry has always been listed as learned
and a static entry has never been visible as such in the table listing.
parse_isolate() accepted a two digit port and mapped it to logical port 9,
the CPU port, while port_isolate() and port_isolation_get() both refuse
anything above machine.max_port. Setting the isolation of the CPU port was
therefore declined without a word and reading it always answered no
members, whatever the hardware held.
Bound the port to the front panel, so the command says what it does. The
digit is checked before it indexes phys_to_log_port[], which a non numeric
argument used to read past.
handle_rx() dispatched to igmp_packet_handler() on the destination
address alone, so an ordinary IGMPv3 report off the wire reached the
handler and could write a table entry whether or not anyone had asked
for snooping. The STP branch right above it is gated on stpEnabled;
this brings the IGMP branch in line.
Snooping state lived only in the per-port registers, and the receive
path cannot afford to read one per packet, so the flag shadows it:
igmp_enable() sets it, igmp_setup() clears it, and igmp_setup() runs
from both the boot path and "igmp off".
While here, igmp off becomes an explicit subcommand instead of the
fall-through, and an unrecognised igmp subcommand prints the usage
line rather than silently turning snooping off.
Six bytes of BANK1 and one of xdata, no internal RAM.
The two flags added with the JSON fix sit in data, where internal RAM is
full enough that this branch stopped linking for some toolchains. __bit
puts them in the bit area instead and hands three bytes back to the
stack: SSEG goes from 131 to 134 on SWTGW218AS.
Patch by vDorst on the pull request.
lag show has always printed the groups as 1 to 4 while lag <n> took the
number literally, so typing what you saw configured the group beside it.
Both lag and lag hash count from one now, matching how ports are numbered
everywhere else, and reject anything outside 1 to 4. Subtracting '1' makes
0 wrap well past three, so one test covers both ends.
The port argument indexed machine.phys_to_log_port, which holds nine
entries, before it was checked, and a two digit argument reaches 109. It is
bounded before the table is touched rather than after.
port_lag_members_set() and port_lag_hash_set() complained about a group out
of range and then wrote the registers anyway, past the four the groups
occupy. They return instead.
lag hash also read cmd_words_b[1] without checking a word was there, and
now shares the error path parse_lag() already had.
Both the L2 page and the ports page in #335 need to page through /l2.json
and decode the same fields, and the second copy arrived carrying the two
bugs the first one had only just been fixed for. Rather than keep two
copies in step by hand, the transport and the decoding move to main.js,
which every page already loads, and each page keeps only what is its own.
walkL2(onDone) pages through the table once, parses idx and vlan out of
hex, maps the port to a physical number or to 'CPU', and calls
onDone(entries, ok). It stops on a wrapped index, an empty page or 4096
entries, all of which set ok. A page that comes back as anything other than
200, or with a body JSON.parse rejects, is asked for again at the same
index up to three times; only once those run out does the walk end with ok
clear, so a caller can tell a finished table from a partial one. l2.js
keeps the s and l to label mapping, since that needs the page's own
translations, redraws only when ok is set, and restarts the walk from its
callback either way.
Two things change while moving:
The next request goes out from the previous reply rather than from a
setInterval that fires whether or not the last one came back. The httpd
serves one connection at a time, so a timer that outruns the responses only
queues work it cannot use.
A walk that reaches 4096 entries hands over what it collected. Before it
threw the entries away and cleared its own interval, which left the page
unable to refresh again until it was reloaded.
The retry is not a new idea, it is the old behaviour written down. The
previous code ignored anything that was not a 200 and let the interval ask
for the same index again, so a blip never disturbed the table on screen.
Dropping that on the way to a chained walk would have made every timeout
redraw the page with a truncated table, which at one connection at a time
is not a rare event.
Driven with a scripted server in node, running the helper itself rather
than a copy of it: an empty table gives 0 entries in 1 request; three pages
ending in a repeated index give 61 entries in 3 requests, asking for 0, 30
and 60; an empty page ends the walk after 2; a 500 and a malformed body are
each retried at the same index and then complete normally, asking 0, 30, 30
and 31; three failures in a row end the walk with ok clear and the 30
entries already collected; 4096 entries in one page end it with ok set; the
CPU port decodes to 'CPU'; vlan and idx come back as numbers.
main.js grows by 1331 bytes and l2.js loses 1039, so 292 bytes of flash.
Worth stating where they land: main.js is loaded by every page, so pages
that never walk the table now carry the helper too. That is the cost of
having the decoding exist exactly once, which is the point of the move.
sfp_read_reg() waited for the transfer to finish and then read the output
register whatever the outcome, so an address nothing acknowledged came back
as an ordinary byte and no caller could tell it apart from data. The vendor
SDK looks at bit 1 of the control register for exactly this, and we did
not.
A failure now sets sfp_i2c_fail and the read returns 0xff, which is already
the value sfp_apply_quirks() reads as either a failed transfer or a voltage
the spec does not allow, so that test starts being true when it should be.
The insertion path and the sfp command clear the flag first and say so
afterwards, rather than presenting the bytes as though they came from the
module.
What this deliberately does not do is act on the failure. Skipping
sds_config() when the rate read failed is the obvious next step, but a
module that raises the bit spuriously would then never be configured at
all, which is worse than what happens today, and I have no way to judge how
often the bit is right. That decision belongs with someone holding the
board.
It also leaves the other half of the rewrite alone, reading and writing up
to sixteen bytes per transaction. doc/sfp.md describes only the single byte
path and does not name a length field, and guessing at a register I cannot
test is how the last attempt at this function went wrong.
40 bytes of the common segment, 51 of BANK2 and 1 of xdata, nothing in
BANK1 or internal RAM. Built for SWTGW218AS and KP_9000_6XHML_X2 on sdcc
4.5.0. Not tested on hardware: shorting the clock line, as in #342, should
now print the failure line instead of a plausible looking byte.
l2.js reads the last index of a reply to know where the next page starts.
With the firmware side of this branch an empty table answers [], and
s[s.length-1] then throws on undefined. It used to answer commas with
nothing between them and throw in JSON.parse instead, so this is the same
case reaching a different line rather than a new one. An empty reply now renders what
has been collected and starts the next pass from zero.
The 4096 entry cap compared the array against the number instead of its
length, so it never fired: an empty array and a 5000 element one both
compare false. Comparing the length restores what the check was for.
The MAC table listing wrote its separator once per iteration rather than
once per object. An entry the table engine reports as invalid produces no
object, so it contributed a bare comma, and two in a row give ",," which
JSON.parse rejects. The whole table then fails to load, not just the row
that was missing. The separator now goes before each object and the
closing bracket after the loop, which is the shape send_vlanlist already
uses further down the file.
The next index for an invalid entry was computed as h | low + 1, and the
addition binds tighter than the or. That agrees with (h | low) + 1 except
when the low byte reads 0xff and bit 8 of the index is already set, eight
of the 4096 combinations. There the result is the start of the current
block of 256 rather than the start of the next one, so the walk repeats a
block it has already covered. Reading the index once after the branch
rather than once in each arm removes the second copy of that expression
along with the bug.
The VLAN now comes first in each object. It is taken from the same
L2_DATA_OUT_B read that decides whether the entry is valid, which saves
reading that register a second time. The page addresses the fields by
name, so the order they arrive in does not matter to it.
A bound check on the output buffer goes in for consistency with
send_vlanlist. Thirty entries of at most 74 bytes plus the brackets fit in
the 2500 byte buffer with 179 to spare, so nothing changes today, but the
margin was nowhere stated and L2_MAX_TRANSFER is a tunable.
5 bytes of BANK1, nothing in BANK2, xdata or internal RAM. Built for SWTGW218AS
and KP_9000_6XHML_X2 on sdcc 4.5.0.
__sdcc_banked_call and __sdcc_banked_ret were assembled into GSFINAL, which
sits in the startup path: GSINIT ends exactly where GSFINAL begins, so the
processor falls into it rather than being sent there. It works today only
because this object comes after every C object on the link line, so the
LJMP to __sdcc_program_startup is laid down first and the helpers land
behind it. Reordering that line, or moving main() into another module,
would put the helper body at the fallthrough address instead, and the board
would not come up out of a build that reports nothing wrong.
SDCC's own crtbank.asm declares the area order and then puts both symbols
in HOME, so the file takes that name and that preamble as well.
GSFINAL now holds the three byte jump and nothing else. Of 401 symbols 17
change address, every one in the startup region, and between the reset
vector and 0x0094 not a byte differs, so no interrupt vector is disturbed.
Run on a SWTGW218AS: it came back after about 42 seconds reporting the new
build, with its stored configuration byte identical and every link at the
speed it had before.
Wrapping REG_SET restored the guard in front of the hash default, and
that exposed three things about the line it guards.
The test was against zero. The register does not read zero: it comes out
of reset holding source port number plus both MAC fields, both IP fields
and the L4 source port, which the header now names LAG_HASH_RESET.
Measured on an SWTGW218AS, where all four groups read 0x3f after a cold
boot and a value written before a power cycle is gone afterwards. With
the guard working and the test unreachable, the default would never be
installed, where before it was installed on every call. Testing against
the reset value restores the intent, and zero is still accepted in case
another device does reset that way.
The write went to the base address while the read that decides it used
the group offset, so a group other than zero was tested and group zero
was written. Both ends use the offset now.
The range check printed a complaint and carried on. It returns, which
matters more now that the hash write also uses the group number to build
an address.
Thirty bytes of BANK1.
The member mask of an aggregation group is decoded by hand in two places,
the lag command and the JSON behind the aggregation page, and every branch
that touches trunks adds another copy.
port_lag_members_get() sits next to port_lag_members_set() and both readers
call it. It answers from the hardware, so it covers a group configured with
lag and one a protocol brought up, without either having to say so.
It reads through reg_read() rather than reg_read_m(), so sfr_data is left
alone. Neither caller looked at it afterwards; both read the hash register
next.
The handler took one raw character of the request line and passed it to
send_counters(), which uses it to index machine.phys_to_log_port. That
array has nine entries and the character is whatever the client sent, so
the read ran up to 246 entries past the end and the result went on to
STAT_GET as a port number. is_word() accepts any request whose name is
followed by a question mark, so nothing constrained the byte to a digit.
Bounding it where it is read keeps the check beside the assumption it
protects and needs nothing from the machine description. sdcc leaves
plain char unsigned and the subtraction wraps in eight bits, so a byte
below '0' comes out above 200 and one upper test covers both ends:
exactly '0' to '8' now reach send_counters. The compiled test is
add a,#0xf7 followed by jnc, which I read back out of the assembly rather
than assuming.
Out of range answers 400 by the path the other malformed requests already
take, rather than an empty array. An empty array would have been worse
than useless here, since the statistics page calls BigInt on the first
element before it looks at the length. The page asks only for index zero
to the port count minus one, so nothing that answered before stops
answering, and a non-200 reply makes its handler do nothing at all.
11 bytes of BANK1, nothing in the common segment, BANK2, xdata or
internal RAM. Built for SWTGW218AS and KP_9000_6XHML_X2 on sdcc 4.5.0.
Mark create_build_dir as PHONY so directory creation is never skipped.
Also add html_data.h as order-only prerequisite to the .c pattern rule,
fixing another the race where httpd/httpd.c and httpd/page_impl.c are
compiled before the generated header exists.
Replace file-path prerequisite tools/output/fileadder with order-only
dependency on the tools PHONY target.
Also add tools as order-only dependency to the final .bin target which
invokes all of the tools.
igmp_setup() writes the per-port IGMP configuration twice, once with the
value spelled out and once with the same number assembled from the
constants: IGMP_MAX_GROUP | IGMP_PROTOCOL_ENABLE | IGMP_FLOOD is exactly
0x00ff7c15. The second loop carries the comment block explaining the bit
layout, so the literal one is the one to drop.
Today the cost is one redundant register write, because REG_SET is not a
single statement and the unbraced loop body only ever reaches index
machine.max_port + 1. Once the macro is wrapped it becomes one redundant
write per port on every boot and on every "igmp off", which is what makes
this worth removing rather than leaving.
Fifty nine bytes of BANK1 on SWTGW218AS, nothing anywhere else.
Both macros expand to a run of statements joined by backslashes with
nothing around them, so as the unbraced body of an if or a for only the
first assignment belongs to that body. The other three and the
reg_write() call sit after it and run once, unconditionally, with
whatever the loop counter ended on. Wrapping each macro into a single
statement is what every call site already assumes it to be.
This hands the compiler no new room around the SFR writes, which is worth
showing rather than asserting. Building the whole image before and after
and comparing the generated assembly module by module, with label
numbering, block scope suffixes and the version string normalised away,
three modules differ: rtl837x_igmp, rtl837x_port and rtl837x_leds. Ten
modules call these macros, so the other seven come out identical, and so
does everything else in the image.
Those three differ because they hold the five unbraced uses. Two of them
want a fix rather than only the brace, and that is left to the commits
that follow.
Six bytes of BANK1 on SWTGW218AS, nothing anywhere else.
* uip: parenthesise UIP_LLH_LEN
The macro expands to a bare sum, so wherever it is subtracted the second term
gets added instead. UIP_TCP_MSS - and with it UIP_RECEIVE_WINDOW - therefore
comes out 24 bytes above the buffer's real capacity. UIP_APPDATA_SIZE and
UIP_REASS_BUFSIZE are wrong the same way, though neither is reachable today.
The additions, uip_buf[UIP_LLH_LEN] and friends, were right by luck.
* uip: keep the advertised MSS below the buffer edge
Deriving the MSS straight from the buffer size makes the switch advertise
exactly the segment that fills uip_buf to its last byte, and a peer that takes
it literally corrupts every large upload: the firmware image arrives fully
acknowledged, with no retransmissions on the wire, yet the CRC over the
streamed body never matches and the flash write is abandoned.
Isolated by changing nothing but the segment size, same buffer and same file:
1490-byte segments fail four times out of four, 1460-byte segments succeed,
745-byte segments succeed. Linux halves its segments against a window this
small, so only macOS on a jumbo link ever produces a full-size segment - which
is why the failure hides so well.
Where exactly the full segment breaks the stream is not pinned down yet; until
it is, the advertised MSS stays a step below the edge.
* uip: size the buffer to the largest frame the CPU port accepts
UIP_TCP_MSS derives from UIP_CONF_BUFFER_SIZE, and the buffer was large enough
for frames the hardware will never deliver, so the switch advertised a segment
size no peer could usefully reach. A client on a jumbo-MTU link took it at its
word and the oversized replies went nowhere.
Size the buffer to the ingress limit instead. ICMP bypasses MSS and so probes
the hardware directly: on a SWTGW218AS a 1502-byte payload is answered and 1503
never arrives, which puts the largest frame the NIC hands us at 1556 bytes of
uip_buf. UIP_TCP_MSS then derives to 1490, the same edge measured over TCP.
Frames above the limit are dropped by the NIC rather than written to the
buffer - an 8 kB ping leaves the switch untouched - so nothing overruns it.
Frees 644 bytes of XDATA.
* uip: trim these comments, one of which had stopped being true
The note above UIP_CONF_BUFFER_SIZE claimed the MSS derives from it as 1490.
It does not: the commit that follows pins the MSS at 1460 on purpose, a step
below that ceiling, because a segment filling the buffer to its last byte
corrupts large uploads. Left as it was, the file argued with itself.
Both blocks are shorter now. What justifies the numbers stays, which is the
ICMP measurement behind 1556 and the four-out-of-four failure behind 1460.
What went is the storytelling around them, which belongs in this thread rather
than in a config header.
* uip: derive the MSS from the buffer again, minus explicit headroom
The review asked why the buffer size and the MSS are both set by hand when one
used to follow from the other. They answer different questions, but the gap
between them is a number in its own right, so it gets a name now:
UIP_CONF_BUFFER_EXTRA, and UIP_TCP_MSS goes back to being derived.
The headroom is where the measurement lives. A segment that fills uip_buf to
its last byte corrupts large uploads: with nothing but the segment size
changing, 1490 fails four times out of four and 1460 succeeds. With the buffer
sized to the frame the NIC accepts, an extra of 30 lands on 1460.
Deriving it the other way round does not work. Sizing the buffer from a 1460
byte MSS gives 1526, which is 30 bytes under the frame the NIC actually
delivers. A 1502 byte ICMP payload occupies 1556 bytes of uip_buf and is
answered today, and it would stop fitting.
The generated image is byte for byte the same as the one with 1460 written out,
so the expression lands on the value that was measured.
---------
Co-authored-by: d00f <tokyusho@chatik.pl>
René van Dorst
logicog
feelfree69
d00f
bloqaudio
d00f
Priit Laes