diff options
Diffstat (limited to 'kernel/gsp_sema_service.c')
| -rw-r--r-- | kernel/gsp_sema_service.c | 32 |
1 files changed, 28 insertions, 4 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index de92b8e..e0f1897 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -12815,7 +12815,17 @@ static void svc_gl_flush(struct nv_vgpu_device *vgpu, * references is covered by the candidate's own mapped pages. The * 448-word cap leaves headroom for the backend's 512-word method * block (2 bind words + the window + the 8-word release tail). - * Draw-synth/hold captures are never marked. + * + * eng24 (CONT.485): op_src=draw-synth is now included for the + * direct-GL (beef-family) lane when the raw window geometry holds. + * The draw-synth op_src only means the section's PROGRAM/PARAM/REL + * operands were synthesized (the beef push carries none); the WINDOW + * itself is the client's real method stream (BEGIN/END + RT setup + + * draw state), which is exactly what the verbatim 3D replay lane + * wants. Excluding 'd' sent these captures down the synthesized + * compute-QMD lane, where the launch waits for a QMD release the + * engine never writes (qmd_rel=0xc0ffee01 want=0x600df00d) instead + * of taking the X lane's proven ch_rel branch. * * SESSION 153-CONTINUED-10: op_src=hold is included. The GLX * server's per-frame CONTENT pushes carry the draw window but no @@ -12827,7 +12837,17 @@ static void svc_gl_flush(struct nv_vgpu_device *vgpu, * raw 3D method window the r929 replay lane wants; the synthesized * compute fields are what stream_3d bypasses. */ - if ((op_src[0] == 's' || op_src[0] == 'h') && svc_gr_cand.valid && + /* + * eng24 (CONT.485): the direct-GL lane's draw-synth captures ('d') + * join the replay lanes when the whole push survived in raw[] and + * the window starts at a real method (same geometry gates as 's'/'h'; + * the 'd' operands are synthetic but the window bytes are real). Gate + * to the beef family so no other client's capture changes lane. + */ + if ((op_src[0] == 's' || op_src[0] == 'h' || + (op_src[0] == 'd' && + (ch->h_object & 0xfffff000u) == 0xbeef0000u)) && + svc_gr_cand.valid && len == svc_gr_cand.push_n * 4 && svc_gl.start_off + 8u <= at_off && (at_off - svc_gl.start_off) / 4u <= 448u) { @@ -12837,7 +12857,9 @@ static void svc_gl_flush(struct nv_vgpu_device *vgpu, pr_info("sema-gl: 3D stream marked h=0x%x push=0x%llx window=%u words @off=0x%x (verbatim replay from the C597 bind)\n", ch->h_object, push_va, svc_gr_cand.stream_n, svc_gl.start_off); - } else if ((op_src[0] == 's' || op_src[0] == 'h') && + } else if ((op_src[0] == 's' || op_src[0] == 'h' || + (op_src[0] == 'd' && + (ch->h_object & 0xfffff000u) == 0xbeef0000u)) && svc_gr_cand.valid) { /* * SESSION 153-CONTINUED-45: a REAL-operand section whose @@ -12875,7 +12897,9 @@ static void svc_gl_flush(struct nv_vgpu_device *vgpu, * This is read-only and names the exact values the ENGINE will see; * an index size of 1 means the buffer is read as 16-bit indices. */ - if ((op_src[0] == 's' || op_src[0] == 'h') && + if ((op_src[0] == 's' || op_src[0] == 'h' || + (op_src[0] == 'd' && + (ch->h_object & 0xfffff000u) == 0xbeef0000u)) && svc_gl_draw_args_log < 64 && svc_gl.start_off < at_off && at_off <= len) { u32 o = svc_gl.start_off; |