diff options
Diffstat (limited to 'kernel/gsp_sema_service.c')
| -rw-r--r-- | kernel/gsp_sema_service.c | 36 |
1 files changed, 22 insertions, 14 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index 53e2140..0354831 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -13396,20 +13396,28 @@ bool gsp_sema_pool_hold_fresh = 1; bool gsp_sema_copy_dst_pin_guard = 1; /* * r336/r34x wait gate: honor the C56F host-semaphore acquire ops captured - * for a ring entry before running its work. Still DEFAULT OFF (opt-in - * experiment; the module param sema_wait_gate=1 enables it), but the - * semantics are now the class contract's: op 0 = equal, 2 = strict GEQ, - * 3 = circular GEQ, 4 = AND, 5 = NOR; the wait set is captured PER ENTRY - * (the old single-last-wait global leaked a later entry's acquire into an - * earlier entry - the r340 over-hold); the entry's own release counts only - * when its method order precedes the acquire; pending releases our - * machinery is merely late about are drained first; and every hold is - * bounded at SVC_WAIT_HOLD_MS (100 ms) then fails open loudly ('give-up') - * so a value that can only be installed behind an upstream opdef hold - * cannot park a channel. Diagnostics: 'sema-wait: pass/hold/give-up' rows - * carry op, payload, observed value and decision. - */ -bool gsp_sema_wait_gate; + * for a ring entry before running its work. DEFAULT ON since r408: the + * r408 streams2d forensics reproduced the class the gate was built for + * with the gate off - the c2 readback (C56F ACQ_CIRC_GEQ pay=4 on the + * scale channel's progress counter 0x20400ffd0) ran at 106617.279 with + * cur=2, i.e. BEFORE the c2-add launch was even submitted (106617.411), + * read the not-yet-produced buffer (zeros) and the client failed with + * 'c2 all zeros'. With the gate on the same readback holds (sema-wait: + * hold ... cur=2 waited_ms=353) until the counter reaches 4 (pass ... + * cur=4 at 106812.259) and the ordering is correct. The semantics are + * the class contract's: op 0 = equal, 2 = strict GEQ, 3 = circular GEQ, + * 4 = AND, 5 = NOR; the wait set is captured PER ENTRY (the old + * single-last-wait global leaked a later entry's acquire into an earlier + * entry - the r340 over-hold); the entry's own release counts only when + * its method order precedes the acquire; pending releases our machinery + * is merely late about are drained first; and every hold is bounded at + * gsp_sema_wait_hold_ms (default 4000 ms, r401) then fails open loudly + * ('give-up') so a value that can only be installed behind an upstream + * opdef hold cannot park a channel. Diagnostics: 'sema-wait: + * pass/hold/give-up' rows carry op, payload, observed value and decision. + * sema_wait_gate=0 restores the pre-r336 pass-through for an A/B. + */ +bool gsp_sema_wait_gate = 1; /* * m169b: the guard above must never deadlock the boot's first client. The |