[Accton] Check TX_DISABLE support before access in onlp_sfpi_control_…

[eric271110]

…get/set

platform:
as7535_28xb
as7926-40xfb
as7946-30xb
as7946-74xkb
as9716_32d
as9736_64d
as9737-32db
as9817-64nb
as9817-64
as9926-24

merge from [Accton][as9726-32d] Check TX_DISABLE support before access in onlp_sfpi_control_get/set (#191)

1. Flat-memory CMIS modules (typically passive DACs / simple AOCs) only have lower memory + page 00h. The page select register is silently discarded, so a request for page 01h or 10h actually still reads page 00h vendor info, returning plausible-looking but meaningless bytes.

2. control_get previously did not check the TX Disable Supported advertising bit at all; it walked straight to page 10h and read byte 0x82. control_set checked the advertising bit but missed the flat-memory case.

Both paths now:

• Read lower-memory byte 0x02 bit 7 (Flat_mem) first. Flat-memory modules return ONLP_STATUS_E_UNSUPPORTED immediately, without touching the page select registers.
• For paged modules, check the TX Disable Supported advertising bit on page 01h byte 0x9B bit 1, mirroring the existing set path.
• control_get's restore block preserves E_UNSUPPORTED instead of overwriting it with E_INTERNAL.

Verified on a real switch via libonlp.onlp_sfp_control_get(): an unsupported / flat-memory module now returns -10 (E_UNSUPPORTED) instead of a stale lane-mask byte.

[vincentchiang-ec]

test

[vincentchiang-ec]

Nice port overall — flat_mem detection, advertising-bit gating, bank-select order and the syslog cleanup all landed cleanly on most platforms. Two systematic issues showed up during review, and I've left inline notes at each site.

1. Restore-write clobbers E_UNSUPPORTED / OK (5 platforms × set+get = 10 sites). The final writeb(page_select, ADMIN_INFO) in the QSFP-DD block sets rv = E_INTERNAL on failure, which overwrites the rv = E_UNSUPPORTED we just carefully set for flat-memory / non-advertising modules (and also flips a successful write to E_INTERNAL). Regresses the exact behaviour PR #191 was written to introduce. Reference fix pattern is in as9726-32d sfpi.c (restore: label + rv = (rv == E_UNSUPPORTED) ? rv : E_INTERNAL) — please mirror it. Affected: as9716_32d, as9736-64d, as9926-24d, as9926-24db, as7926-40xfb.

2. as9817-64 / as9817-64-nb early-return pattern (2 platforms × set+get = 4 sites). These files use return E_INTERNAL inline in each error branch instead of the restore: label. The E_UNSUPPORTED paths (return E_UNSUPPORTED after restore) are fine — but the final "success" restore does return E_INTERNAL on restore failure, so a successful set/get becomes E_INTERNAL if only the ADMIN_INFO write itself fails. Same fix idea: either preserve the pre-restore return value or drop the failure-→ E_INTERNAL on the restore-only branch.

3. Missing & 0xff mask on *value = tx_disable (as9817-64 / as9817-64-nb only). Everywhere else the port applies *value = (tx_dis & 0xff) to avoid the sign-extension from dev_readb's int return; these two platforms still assign the raw signed int, so a byte with bit7 set will surface as a negative value in value. Trivial mask fix.

Nothing else looks off — flat-mem check offsets, page 01h advertising bit, bank-select-before-page-10h all match the reference on every platform.

[vincentchiang-ec]

Restore write clobbers E_UNSUPPORTED (control_set path). When the TX_DISABLE_SUPPORT advertising bit is clear a few lines above, rv is set to E_UNSUPPORTED; this block then unconditionally overwrites it to E_INTERNAL if the ADMIN_INFO restore itself fails. Same thing happens on the successful write path (turns rv = OK into E_INTERNAL).

Suggested fix — mirror the as9726-32d pattern in this repo:

restore:
    if ((onlp_sfpi_dev_writeb(port, PORT_EEPROM_DEVADDR,
                              QSFP_DD_EEPROM_OFFSET_PAGE_SELECT,
                              QSFP_DD_PAGE_ADMIN_INFO)) < 0) {
        syslog(LOG_ERR, "Failed to restore Page Select to Admin Info on port(%d)!", port);
    }
    if (rv < 0) {
        rv = (rv == ONLP_STATUS_E_UNSUPPORTED) ? rv : ONLP_STATUS_E_INTERNAL;
    }

Fork missed this platform — same pattern as the four flagged in the other comments.

[vincentchiang-ec]

Same restore-clobber issue on the control_get path — see the control_set comment above for the reference fix.

[vincentchiang-ec]

Restore write clobbers E_UNSUPPORTED (control_set path). When the earlier else { rv = ONLP_STATUS_E_UNSUPPORTED; } fires (advertising bit clear), this block promotes it to E_INTERNAL if the ADMIN_INFO restore itself fails. Also flips a successful write path (rv = OK) to E_INTERNAL.

Suggested fix — mirror the as9726-32d reference:

restore:
    if ((onlp_sfpi_dev_writeb(port, PORT_EEPROM_DEVADDR,
                              QSFP_DD_EEPROM_OFFSET_PAGE_SELECT,
                              QSFP_DD_PAGE_ADMIN_INFO)) < 0) {
        syslog(LOG_ERR, "Failed to restore Page Select to Admin Info on port(%d)!", port);
    }
    if (rv < 0) {
        rv = (rv == ONLP_STATUS_E_UNSUPPORTED) ? rv : ONLP_STATUS_E_INTERNAL;
    }

[vincentchiang-ec]

Same restore-clobber issue on the control_get path — see the control_set comment above.

[vincentchiang-ec]

Restore write clobbers E_UNSUPPORTED (control_set). Please mirror the as9726-32d restore: label + rv = (rv == E_UNSUPPORTED) ? rv : E_INTERNAL pattern — same story as the as9716_32d comment.

[vincentchiang-ec]

Successful control_get reported as E_INTERNAL when the restore write fails — same pattern as line 332 on the set side. Same fix suggestion.

[vincentchiang-ec]

Successful control_set reported as E_INTERNAL when the restore write fails. Every earlier error branch already return E_INTERNAL after its own restore, so this final restore only runs on the OK path — returning E_INTERNAL here turns a real success into a caller-visible failure. Same fix idea as the as9817-64 comment: either drop the return E_INTERNAL here or adopt the restore: label pattern from as9726-32d.

[vincentchiang-ec]

Missing & 0xff mask on *value = tx_disable. Sign-extension bug — dev_readb returns int and a byte with bit 7 set surfaces as a negative value. Please apply the *value = (tx_disable & 0xff) mask used elsewhere in this PR.

[vincentchiang-ec]

Successful control_get reported as E_INTERNAL when the restore write fails — same pattern as line 314. Same fix suggestion.

[vincentchiang-ec]

Same missing & 0xff mask on *value = tx_disable (QSFP+ path).

[vincentchiang-ec]

Follow-up context on the review above — I want to be fair about how much of this bug PR #194 actually introduced, since re-reading the pre-PR code changed my picture:

as9726-32d was already safe before PR #191 landed. It has used the goto restore + rv = (rv == E_UNSUPPORTED) ? rv : E_INTERNAL pattern for a long time. That's why it works as the reference here — PR #191 only added the flat-memory early-exit on top of an already-safe restore block.

The other 11 platforms use a different structure (inline writeb(ADMIN_INFO) with rv = E_INTERNAL on failure). Splitting the finding against that:

Path	Introduced by PR #194?
set — advertising bit clear → E_UNSUPPORTED clobbered	No, pre-existing. The 5 flagged platforms already had this bug on master; PR #194 didn't fix it (and I don't think it was expected to).
set — new flat-memory E_UNSUPPORTED clobbered	Yes, because PR #194 adds this new path into the already-broken restore.
get — E_UNSUPPORTED clobbered	Yes. control_get didn't produce E_UNSUPPORTED before, so this is fully a PR #194 introduction.
as9817-64 / as9817-64-nb — OK → E_INTERNAL on restore fail	No, pre-existing early-return structure.

So the practical takeaway isn't "PR #194 introduced all these bugs" — it's more that porting the flat-memory check via the restore: label + preserve-E_UNSUPPORTED pattern (i.e. actually mirroring as9726-32d structurally, not just its offsets) would fix both the newly-introduced and the pre-existing cases in one shot. That's the real value of adopting the reference pattern here — the fix isn't scope-creep, it's the same edit that makes the port complete.

(as9817-64 / as9817-64-nb are separate — their early-return structure is fine for E_UNSUPPORTED; only the OK-path return E_INTERNAL on restore-failure needs adjusting.)