diff options
| -rw-r--r-- | kernel/gsp_sema_service.c | 73 | ||||
| -rw-r--r-- | kernel/gsp_sema_service.h | 1 | ||||
| -rw-r--r-- | kernel/main.c | 2 |
3 files changed, 75 insertions, 1 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index 8623f07..f6e85ec 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -1820,6 +1820,9 @@ static int svc_gmmu_walk(struct nv_vgpu_device *vgpu, u64 pdb, u32 entries, * index error. */ idx = (va >> 21) & 0xff; + if (trace) + pr_info("sema-walk: PD0 table=0x%llx idx=%u va=0x%llx tap=%d\n", + table, idx, va, tap); if (svc_read_ap(vgpu, table + 16 * idx, tap == SVC_APERTURE_FB ? SVC_APERTURE_FB : SVC_APERTURE_SYS, @@ -1829,6 +1832,9 @@ static int svc_gmmu_walk(struct nv_vgpu_device *vgpu, u64 pdb, u32 entries, : SVC_APERTURE_SYS, &small, sizeof(small))) return -ENOENT; + if (trace) + pr_info("sema-walk: PD0 big=0x%016llx small=0x%016llx\n", + big, small); /* Small half: 4 KB page table under it. */ sap = svc_pde_aperture(((small >> 1) & 0x3)); @@ -1840,6 +1846,12 @@ static int svc_gmmu_walk(struct nv_vgpu_device *vgpu, u64 pdb, u32 entries, *addr += va & 0xfff; return 0; } + if (trace) + pr_info("sema-walk: small miss sap=%d t=0x%llx\n", + sap, t); + } else if (trace) { + pr_info("sema-walk: small skip sap=%d small&1=%llu\n", + sap, small & 1); } /* Big half: a 2 MB page here, or a large (64 KB) page table. */ @@ -1872,6 +1884,9 @@ static int svc_gmmu_walk(struct nv_vgpu_device *vgpu, u64 pdb, u32 entries, return 0; } } + if (trace) + pr_info("sema-walk: no leaf va=0x%llx big=0x%llx small=0x%llx\n", + va, big, small); return -ENOENT; } @@ -5404,6 +5419,10 @@ static void svc_copy_snap_capture(struct nv_vgpu_device *vgpu, u64 dst, total >= 16 ? *(const u32 *)(buf + 12) : 0); } +static int svc_fwd_operand_hpa(struct nv_vgpu_device *vgpu, u8 kind, + u64 pa, u32 off, u32 bytes, u64 *hpa, + u32 *out_off); + static int svc_copy(struct nv_vgpu_device *vgpu, u64 dst, u64 src, size_t total, u32 dst_phys, u32 dst_aperture, u32 src_phys, u32 src_aperture, u64 *done_out) @@ -5636,6 +5655,25 @@ static int svc_copy(struct nv_vgpu_device *vgpu, u64 dst, u64 src, size_t total, dst_orig, dst, dst_aperture, dst_prov, dst_shadow, src_orig, src, src_aperture, src_prov, src_shadow, total); + /* + * r348 diagnostic: for a high-VA SYS source name the exact + * host PA this read will touch (gpa_to_hpa) - the direct + * comparison against the svc-fwd-hpa writeback line. + */ + if (src_orig >= SVC_VAS_FLOOR && + src_aperture == CE_APERTURE_SYS) { + u64 shpa = 0; + u32 soff = 0; + + if (!svc_fwd_operand_hpa(vgpu, SVC_APERTURE_SYS, + src, 0, SEMA_PAGE_SIZE, + &shpa, &soff) && !soff) + pr_info("sema-bigcopy: src-hpa va=0x%llx gpa=0x%llx -> hpa=0x%llx\n", + src_orig, src, shpa); + else + pr_info("sema-bigcopy: src-hpa va=0x%llx gpa=0x%llx -> UNMAPPED\n", + src_orig, src); + } pr_info("sema-bigcopy: src16=%16phN dst16_before=%16phN\n", sb, db); /* @@ -12032,6 +12070,14 @@ unsigned gsp_phys_gr_hold_max_ms = 3000; unsigned gsp_phys_gr_fill_hold; /* default 0: log-only, never hold */ /* + * r346: trace GMMU walks for unresolved high-VA operands (the arg-miss + * diagnostic re-walks each published vaspace for a failing base). When + * nonzero, those walks print the per-level entries (L0..L2, PD0 halves) + * so a failing VA shows exactly where the walk dies. + */ +unsigned gsp_sema_walk_trace; + +/* * F1-b: GR release-word authorship gate (default 1). While a forwarded GR * launch is in flight, the RAM-content sweep must not author the very QMD * completion release the backend poll is waiting for: a sweep write could @@ -15559,6 +15605,16 @@ static void svc_fwd_gr_work(void) data[gs.n_data].hpa = dhpa; data[gs.n_data].reloc_va = dg->reloc_va; /* + * r348 diagnostic: the last 32 high-VA data pages (the output + * operands live there) with the exact (gpa, ap) -> hpa the + * writeback will use - comparable against the D2H read's + * resolution line (sema-bigcopy) for the same VA. + */ + if (dg->va >= SVC_VAS_FLOOR && + i + 32 >= svc_fwd_req.gr_n_data) + pr_info("svc-fwd-hpa: va=0x%llx gpa=0x%llx ap=%u hpa=0x%llx bi=%u\n", + dg->va, dg->pa, dg->ap, dhpa, i); + /* * GRVAUDIT (diagnostic only): audit the m180 landing * authority against the PA this submit will put on the wire * for every tracked data page - the data[]/n_data set the @@ -16886,6 +16942,21 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu, rec->ap = (u8)svc_ap_of(ap); svc_fwd_req.gr_n_data++; /* + * r349: record the launch's own resolution as the + * ledger's landing authority for this VA. The read + * path's m171 read-after-write rule follows the + * ledger when it disagrees with the live translation, + * and the engine writeback below writes exactly this + * (pa, ap). Without the store the ledger kept a stale + * page for engine-written outputs (victory_reduce: + * the launch wrote FB 0x7c00000 while the ledger + * said SYS 0x7c00000; the D2H read followed the + * ledger and returned zeros while the correct sum + * sat on the written page). + */ + if (va >= SVC_VAS_FLOOR) + svc_ledger_store(va, rec->pa, rec->ap); + /* * m142: bounded per-page attribution log. A wrong data * page is invisible in the aggregate counters; these * lines let a post-mortem compare the chosen PA against @@ -16946,7 +17017,7 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu, wrc = svc_gmmu_walk(vgpu, svc.vas[vi].pdb, svc.vas[vi].entries, base, svc.vas[vi].ap, &wa, &wap, - false); + gsp_sema_walk_trace != 0); o += snprintf(vb + o, sizeof(vb) - o, " vas[%d] pdb=0x%llx ent=%u ap=%d rc=%d pa=0x%llx", vi, svc.vas[vi].pdb, diff --git a/kernel/gsp_sema_service.h b/kernel/gsp_sema_service.h index 6b411c9..fe5cae6 100644 --- a/kernel/gsp_sema_service.h +++ b/kernel/gsp_sema_service.h @@ -284,6 +284,7 @@ extern unsigned gsp_phys_gr_hold_max_ms; * hold and its absolute bound apply. */ extern unsigned gsp_phys_gr_fill_hold; +extern unsigned gsp_sema_walk_trace; /* * m164 CLASS-2: hold window (ms) for a GR launch push whose decode saw QMD diff --git a/kernel/main.c b/kernel/main.c index 47a1aba..bbaf845 100644 --- a/kernel/main.c +++ b/kernel/main.c @@ -2788,6 +2788,8 @@ MODULE_PARM_DESC(gsp_phys_gr_hold_max_ms, "Absolute GR fill hold bound in ms, ke */ module_param_named(gsp_phys_gr_fill_hold, gsp_phys_gr_fill_hold, uint, 0644); MODULE_PARM_DESC(gsp_phys_gr_fill_hold, "GR fill hold on unresolved operand pages (default 0: log-only, never hold)"); +module_param_named(gsp_sema_walk_trace, gsp_sema_walk_trace, uint, 0644); +MODULE_PARM_DESC(gsp_sema_walk_trace, "Trace GMMU walks for unresolved high-VA operands (default 0: off)"); /* * F1-b: while a forwarded GR launch is in flight, defer (do not apply) a |