diff options
Diffstat (limited to 'kernel')
| -rw-r--r-- | kernel/gsp_sema_service.c | 61 | ||||
| -rw-r--r-- | kernel/gsp_sema_service.h | 3 | ||||
| -rw-r--r-- | kernel/main.c | 2 |
3 files changed, 66 insertions, 0 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index ea96595..eddc73d 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -8985,6 +8985,60 @@ static void svc_qmd_accum(struct nv_vgpu_device *vgpu, struct svc_channel *ch, static u64 svc_pass_hash; /* r391f: this decode pass's push content hash */ +/* + * r481 RISK-REDUCTION diagnostic (default 0): dump the full decoded method + * stream of the first few large GR-class pushes (len >= 1024) so the launch + * push's acquire/release method groups (or their absence) are visible + * verbatim. The streams_events class evidence shows the st2 vadd's copy-in + * running ~115 ms BEFORE its stream's H2D soft copies; the ordering the + * driver relied on is only provable by reading the push's own methods. + * Bounded: gsp_sema_push_raw dumps, at most once per push VA. + */ +unsigned gsp_sema_push_raw; +static u64 svc_pushraw_last_va; +static u32 svc_pushraw_n; + +static void svc_push_raw_dump(const u8 *buf, u32 len, u64 push_va, + u32 h_object) +{ + u32 off = 0, gi = 0; + + if (svc_pushraw_n >= gsp_sema_push_raw || !gsp_sema_push_raw) + return; + if (push_va == svc_pushraw_last_va) + return; + svc_pushraw_last_va = push_va; + svc_pushraw_n++; + pr_info("sema-pushraw: BEGIN h=0x%x va=0x%llx len=%u n=%u\n", + h_object, push_va, len, svc_pushraw_n); + while (off + 4 <= len && gi < 128) { + u32 hdr, cnt, addr, v[8], k, n; + + memcpy(&hdr, buf + off, 4); + if ((hdr >> 29) != 1) { + pr_info("sema-pushraw: off=0x%x raw=%08x (not a method header)\n", + off, hdr); + break; + } + cnt = (hdr >> 16) & 0xfff; + addr = (hdr & 0x1fff) << 2; + n = cnt > 8 ? 8 : cnt; + for (k = 0; k < n && off + 4 + 4 * k + 4 <= len; k++) + memcpy(&v[k], buf + off + 4 + 4 * k, 4); + pr_info("sema-pushraw: [%u] off=0x%x addr=0x%x cnt=%u v0=%08x v1=%08x v2=%08x v3=%08x v4=%08x\n", + gi, off, addr, cnt, + n > 0 ? v[0] : 0, n > 1 ? v[1] : 0, + n > 2 ? v[2] : 0, n > 3 ? v[3] : 0, + n > 4 ? v[4] : 0); + if (!cnt) + break; + off += 4 + cnt * 4; + gi++; + } + pr_info("sema-pushraw: END h=0x%x va=0x%llx consumed=0x%x/%u\n", + h_object, push_va, off, len); +} + static void svc_decode_stream(struct nv_vgpu_device *vgpu, const u8 *buf, u32 len, u64 page_pa, bool allow_ops, struct svc_channel *ch, u32 entry_idx, @@ -9049,6 +9103,13 @@ static void svc_decode_stream(struct nv_vgpu_device *vgpu, const u8 *buf, allow_ops = false; } + /* + * r481: verbatim method-stream dump for large GR-class pushes + * (default 0 = off). See svc_push_raw_dump. + */ + if (gsp_sema_push_raw && ch && !ch->engine_type && len >= 1024) + svc_push_raw_dump(buf, len, push_va, ch->h_object); + memset(&st, 0, sizeof(st)); st.push_va = push_va; /* diff --git a/kernel/gsp_sema_service.h b/kernel/gsp_sema_service.h index 3765cd2..305a027 100644 --- a/kernel/gsp_sema_service.h +++ b/kernel/gsp_sema_service.h @@ -313,6 +313,9 @@ extern bool gsp_sema_wait_gate; /* r401: bounded acquire-wait hold (fail-open) in ms; must exceed the * emulated launch completion latency for copy-after-launch ordering. */ extern unsigned gsp_sema_wait_hold_ms; +/* r481: verbatim method-stream dump for the first N large GR pushes + * (default 0 off). Diagnostic only; no behaviour change. */ +extern unsigned gsp_sema_push_raw; /* * r384 (guest boot-corruption lottery): fill-destination guards (default diff --git a/kernel/main.c b/kernel/main.c index 0110372..3ee5c23 100644 --- a/kernel/main.c +++ b/kernel/main.c @@ -2832,6 +2832,8 @@ module_param_named(sema_wait_gate, gsp_sema_wait_gate, bool, 0644); MODULE_PARM_DESC(sema_wait_gate, "honor C56F host-semaphore acquires before running a ring entry's work (default 1 since r408; 0 restores pass-through)"); module_param_named(sema_wait_hold_ms, gsp_sema_wait_hold_ms, uint, 0644); MODULE_PARM_DESC(sema_wait_hold_ms, "bounded acquire-wait hold before fail-open, ms (r401, default 4000; must exceed the emulated launch latency)"); +module_param_named(sema_push_raw, gsp_sema_push_raw, uint, 0644); +MODULE_PARM_DESC(sema_push_raw, "r481 diagnostic: dump the decoded method stream of the first N large GR pushes (default 0 off)"); MODULE_PARM_DESC(sema_copy_dst_pin_guard, "hold device-dst copies without a pin-resolved translation (default 1)"); /* |