diff options
Diffstat (limited to 'kernel')
| -rw-r--r-- | kernel/gsp_phys_booter.c | 201 | ||||
| -rw-r--r-- | kernel/gsp_phys_gpu.c | 6 |
2 files changed, 193 insertions, 14 deletions
diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c index b1f4b68..0bafced 100644 --- a/kernel/gsp_phys_booter.c +++ b/kernel/gsp_phys_booter.c @@ -683,6 +683,27 @@ bool gsp_phys_boot_requested(void) */ #define GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC 33u /* + * Mode 34 (2026-10-06, experiment): the context-share golden lane. A + * FERMI_CONTEXT_SHARE_A (0x9067, flags SYNC -> VEID 0, VAS = the ladder + * SHARED_MANAGEMENT VAS) is allocated on the ladder TSG before the replay. + * The golden replay's channel A (0xcaf00054) is parented to the ladder TSG, + * passes hVASpace=0 and hContextShare (mutually exclusive) and keeps the + * captured unwired shape; its 9-entry promote (first init) runs on A, + * followed by A's TURING_A (0xc597) and a TURING_COMPUTE_A (0xc5c0, + * 0xcaf00090) class object. V6 (0xcaf00034) is allocated afterwards with + * the same share and constructs no class object of its own (a later V6 + * construct would re-commit the shared VEID's subcontext header over A's + * state); its FWCLIENT promote is suppressed. Stage 4 BIND/SCHEDULE/token/ + * GR-sem and every forwarded launch run on V6, which restores exactly the + * VEID/subcontext A's class state was written into. + * + * NOTE on naming: the requested macro name GSP_PHYS_RMCALL_GR_PROMOTE_CTXSHARE + * is already taken by mode 18 (the mode-8-shape context-share mode), so this + * mode is defined as GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE to disambiguate + * on the wire (gsp_phys_rmcall_gr_promote=34) and in logs. + */ +#define GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE 34u +/* * Mode-28 golden execution channel shape bits (gsp_phys_rmcall_golden_exec_shape, * default 7). Every combination keeps the captured flags 0x00200320 and the * FBMEM instance/RAMFC/mthd slices; the bits only choose the client-visible @@ -796,7 +817,7 @@ bool gsp_phys_boot_requested(void) */ #define GSP_PHYS_RMCALL_FWCLIENT_MASK_FECS_SYSMEM (1u << 16) #define GSP_PHYS_RMCALL_GR_PROMOTE_MAX \ - (GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC) + (GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) /* * Work-submit token notifier slot. The default index is @@ -1076,6 +1097,15 @@ bool gsp_phys_boot_requested(void) * SET_OBJECT binds this class in A's execution context. */ #define GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING_GR_OBJ 0xcaf0008cu +/* + * Mode 34: the TURING_COMPUTE_A (0xc5c0) class object constructed on the + * context-share golden channel A (0xcaf00054) after its 9-entry promote. + * V6 executes in A's shared VEID and constructs no object of its own, so + * this handle (not HANDLE_GR_OBJ) is the ladder's rmcall_gr_object for + * mode 34. 0xcaf00090 is free between the sibling handles and the + * 0xcaf01000 registered-buffer base. + */ +#define GSP_PHYS_RMCALL_HANDLE_GOLDEN_CTXSHARE_GR_OBJ 0xcaf00090u /* * Golden-image replay sizes (kernel_graphics.c:2058-2060, 2273-2276): the @@ -7706,6 +7736,8 @@ static const char *gsp_phys_rmcall_gr_promote_name(void) return "mode 31: mode-25 D first-init + TSG sibling A 9-entry promote after V6 (explicit TSG commit) + V6 execution"; case GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC: return "mode 33: mode 31 with the TSG sibling A wired as the object-carrying execution channel (ring/sysmem USERD/TSG parent, TURING_COMPUTE_A on A, execution swap to A)"; + case GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE: + return "mode 34: golden channel A (0xcaf00054) TSG-parented through FERMI_CONTEXT_SHARE_A (0x9067, VEID 0, ladder VAS), captured unwired shape + 9-entry promote + TURING_A + TURING_COMPUTE_A; V6 (0xcaf00034) executes on the same share with no object of its own and its promote suppressed"; default: return "off (PROMOTE_CTX never sent)"; } @@ -7898,7 +7930,9 @@ static int gsp_phys_rmcall_alloc_channel(void __iomem *bar0) boot_stage.rmcall_gr_promote == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT || boot_stage.rmcall_gr_promote == - GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC)) { + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC || + boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE)) { vars = &gsp_phys_rmcall_channel_variant_skipctx; attempts = 1; } @@ -8096,6 +8130,20 @@ static void gsp_phys_rmcall_golden_3d_on_ladder(void __iomem *bar0) boot_stage.rmcall_gr_promote == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC_LAST) return; + /* + * Mode 34: the context-share golden channel A already carries the + * class state (TURING_A + TURING_COMPUTE_A) that V6's shared VEID + * restores. V6 must not construct: a V6 GR object construct would + * re-commit the same subcontext header A wrote and overwrite that + * state (the context-share design: construct on A only, after the + * promote; B must not construct). + */ + if (boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) { + pr_info("vgsp-boot: rm-call: mode 34: no GOLDEN TURING_A on V6; the shared golden channel A 0x%x carries the class state V6 executes against\n", + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CHANNEL); + return; + } ret = gsp_phys_rmcall_alloc_gr_object_class(bar0, GSP_PHYS_RMCALL_CLASS_GR_3D, GSP_PHYS_RMCALL_HANDLE_GOLDEN_GR3D_LADDER, @@ -8363,7 +8411,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0) boot_stage.rmcall_gr_promote == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT || boot_stage.rmcall_gr_promote == - GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC) && + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC || + boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) && gsp_phys_rmcall_engine_is_gr()) { if (gsp_phys_rmcall_fwclient_mask_value() & GSP_PHYS_RMCALL_GOLDEN9_MASK_LADDER_VAS) { @@ -8477,7 +8527,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0) boot_stage.rmcall_gr_promote == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT || boot_stage.rmcall_gr_promote == - GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC) && + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC || + boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) && gsp_phys_rmcall_engine_is_gr()) { ch.flags = cpu_to_le32(GSP_PHYS_RMCALL_CHANNEL_FLAG_GOLDEN9); ch.internal_flags = cpu_to_le32( @@ -8591,6 +8643,27 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0) ch.h_va_space = cpu_to_le32(0); pr_info("vgsp-boot: rm-call: mode 30 GOLDEN SIBLING A: TSG-parented unwired captured shape (TURING_A will construct here for the first GR init + commit)\n"); } + /* + * Mode 34: the golden channel itself becomes the share carrier + * A. Per the hContextShare contract the channel must be a + * TSG child, must pass hVASpace=0 and takes its VAS from the + * share (the ladder's SHARED_MANAGEMENT VAS). The submission + * wiring stays deliberately absent: gpFifoOffset 0, hUserd + * handle 0 and hObjectBuffer 0 with the captured FBMEM + * instance/ramfc/mthd slices. The share must already exist + * (stage 3 allocates it before this replay). + */ + if (boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) { + ch_parent = boot_stage.rmcall_tsg; + ch.h_va_space = cpu_to_le32(0); + ch.h_context_share = + cpu_to_le32(boot_stage.rmcall_ctxshare); + pr_info("vgsp-boot: rm-call: mode 34 GOLDEN CHANNEL A (context-share carrier): hParent=ladder-TSG 0x%x hVASpace=0 hContextShare=0x%x (VEID 0, VAS 0x%x) captured unwired shape (gpFifoOffset=0 hUserdMemory[0]=0 hObjectBuffer=0, FBMEM instance/ramfc/mthd)\n", + boot_stage.rmcall_tsg, + boot_stage.rmcall_ctxshare, + boot_stage.rmcall_vas); + } pr_info("vgsp-boot: rm-call: GOLDEN9 channel shape: flags=0x%x (PRIVILEGED|USERD_INDEX_VALUE=3|USERD_INDEX_PAGE_FIXED, no SKIP_SCRUBBER/SKIP_CTXBUFFER_ALLOC) hObjectBuffer=0 gpFifoEntries=0x%x instance/userd/ramfc/mthd=FBMEM memdescs block=0x%llx (capture-matched apertures)\n", GSP_PHYS_RMCALL_CHANNEL_FLAG_GOLDEN9, GSP_PHYS_RMCALL_GOLDEN_GPFIFO_ENTRIES, golden_fb); @@ -8685,7 +8758,9 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0) boot_stage.rmcall_gr_promote == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT || boot_stage.rmcall_gr_promote == - GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC) && + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC || + boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) && gsp_phys_rmcall_engine_is_gr()) { ret = gsp_phys_rmcall_golden9_promote(bar0); if (ret) @@ -8734,6 +8809,44 @@ static int gsp_phys_rmcall_golden_replay(void __iomem *bar0) return ret; } + /* + * Mode 34: the TURING_COMPUTE_A (0xc5c0) class object on the same + * golden channel A, after the promote and after TURING_A (the + * promote-then-construct ordering with the working precedent on an + * unwired golden channel, mode 25; the wired sibling's post-promote + * construct was refused 0x57 in m43). This object is the class state + * the shared VEID carries for V6's forwarded methods; V6 constructs + * nothing of its own, so this construct is the last writer of the + * shared subcontext header. + */ + if (boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) { + ret = gsp_phys_rmcall_alloc_gr_object_class(bar0, + GSP_PHYS_RMCALL_CLASS_COMPUTE, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CTXSHARE_GR_OBJ, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CHANNEL, + "mode 34 TURING_COMPUTE_A on golden channel A", + &rpc, &status); + gsp_phys_rmcall_golden_log("TURING_COMPUTE_A on A", + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CTXSHARE_GR_OBJ, + GSP_PHYS_RMCALL_CLASS_COMPUTE, + status, rpc, ret); + if (ret) { + gsp_phys_rmcall_golden_fail("TURING_COMPUTE_A on A", + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CTXSHARE_GR_OBJ, + GSP_PHYS_RMCALL_CLASS_COMPUTE, + status, rpc, ret); + return ret; + } + boot_stage.rmcall_gr_object = + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CTXSHARE_GR_OBJ; + pr_info("vgsp-boot: rm-call: mode 34: golden channel A 0x%x now carries TURING_A 0x%x + TURING_COMPUTE_A 0x%x; rmcall_gr_object=0x%x, V6 executes and constructs nothing\n", + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CHANNEL, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_GR_OBJ, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CTXSHARE_GR_OBJ, + boot_stage.rmcall_gr_object); + } + pr_info("vgsp-boot: rm-call: GOLDEN replay complete: 6 objects (VAS=0x%x PBphys=0x%x PBvirt=0x%x USERD=0x%x channel=0x%x GR=0x%x)\n", GSP_PHYS_RMCALL_HANDLE_GOLDEN_VAS, GSP_PHYS_RMCALL_HANDLE_GOLDEN_PB_PHYS, @@ -10410,27 +10523,39 @@ static int gsp_phys_rmcall_gr_ctx_promote(void __iomem *bar0) mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC_LAST || mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING || mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT || - mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC) && + mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC || + mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) && (!gsp_phys_rmcall_golden_enabled() || !gsp_phys_gpu_enabled())) { pr_err("vgsp-boot: rm-call: mode %u requires gsp_phys_rmcall_golden=1 and gsp_phys_gpu=1 (the golden replay did not run); no PROMOTE_CTX is sent and the ladder aborts with no retry\n", mode); return -EINVAL; } /* - * Mode 30/31/33: the sibling A's 9-entry promote carries the group + * Mode 30/31/33/34: the sibling A's 9-entry promote carries the group * context commit; nothing is sent on the V6 channel. A multi-entry * promote containing MAIN is refused for a SKIP-ctx channel * (graphics_context.c: bKMDVirtualContext -> NV_ERR_INVALID_ARGUMENT) * and the single-MAIN exemption needs the pCtxPhysMemDesc the FWCLIENT * shapes never establish, so a V6 promote would abort stage 3 on 0x1f. + * Mode 34's promote ran on the context-share golden channel A inside + * the replay (skip_fecs=false), before V6 existed. */ if (mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING || mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT || - mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC) { + mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC || + mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) { if (mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING) pr_info("vgsp-boot: rm-call: mode 30: no PROMOTE_CTX on V6; the sibling A carries the group context commit (config_group ppEngCtxDesc), V6 inherits via the TSG commit/channel setup\n"); else if (mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT) pr_info("vgsp-boot: rm-call: mode 31: no PROMOTE_CTX on V6; sibling A's 9-entry promote ran after V6's alloc and the TSG loop (fifo_gv100.c:1480-1494) already committed V6\n"); + else if (mode == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) + pr_info("vgsp-boot: rm-call: mode 34: no PROMOTE_CTX on V6; golden channel A 0x%x (hContextShare=0x%x, TSG 0x%x) ran the 9-entry promote and carries TURING_A 0x%x + TURING_COMPUTE_A 0x%x; V6 0x%x executes in the shared VEID\n", + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CHANNEL, + boot_stage.rmcall_ctxshare, + boot_stage.rmcall_tsg, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_GR_OBJ, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CTXSHARE_GR_OBJ, + boot_stage.rmcall_channel); else pr_info("vgsp-boot: rm-call: mode 33: no PROMOTE_CTX on V6; wired sibling A's 9-entry promote ran (TSG commit) and A carries both the class object and the execution (hChannel=0x%x, hGr=0x%x)\n", GSP_PHYS_RMCALL_HANDLE_GOLDEN_SIBLING, @@ -12105,8 +12230,9 @@ static int gsp_phys_rmcall_alloc_ctxshare(void __iomem *bar0) &cs, sizeof(cs), false, NULL, "FERMI_CONTEXT_SHARE_A", true, NULL, &status); - pr_info("vgsp-boot: rm-call: FERMI_CONTEXT_SHARE_A alloc: hClass=0x%x status=0x%x subctxId=0x%x (hParent=hTsg=0x%x hVASpace=0x%x flags=SYNC)%s\n", - GSP_PHYS_RMCALL_CLASS_CTXSHARE, status, + pr_info("vgsp-boot: rm-call: FERMI_CONTEXT_SHARE_A alloc: hClass=0x%x hObject=0x%x status=0x%x subctxId=0x%x (hParent=hTsg=0x%x hVASpace=0x%x flags=SYNC)%s\n", + GSP_PHYS_RMCALL_CLASS_CTXSHARE, + ret ? 0 : GSP_PHYS_RMCALL_HANDLE_CTXSHARE, status, le32_to_cpu(cs.subctx_id), boot_stage.rmcall_tsg, boot_stage.rmcall_vas, ret ? " (rejected)" : ""); if (ret) @@ -12250,6 +12376,31 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0) boot_stage.rmcall_tsg = GSP_PHYS_RMCALL_HANDLE_TSG; /* + * Mode 34: allocate FERMI_CONTEXT_SHARE_A before the golden replay, so + * the replay's golden channel A can attach to it (hContextShare) and + * the ladder's V6 channel can attach to the same share later. The share + * is parented to the ladder TSG and uses the ladder SHARED_MANAGEMENT + * VAS (its construct allocates the VEID 0 GraphicsSubcontext in the + * TSG's GraphicsContext, and every channel of the TSG must pass it or + * the TSG flips to legacy mode). A missing shared-management VAS or a + * rejected share aborts stage 3 with no retry. + */ + if (gsp_phys_rmcall_engine_is_gr() && + boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) { + if (!gsp_phys_rmcall_vas_shared_value()) { + pr_err("vgsp-boot: rm-call: mode 34 needs the ladder SHARED_MANAGEMENT VAS (gsp_phys_rmcall_vas_shared=1): the context share's VAS is the ladder VAS and rides its page tables; aborting with no retry\n"); + return -EINVAL; + } + ret = gsp_phys_rmcall_alloc_ctxshare(bar0); + if (ret) { + pr_err("vgsp-boot: rm-call: stage 3 aborted: FERMI_CONTEXT_SHARE_A alloc failed (%d); mode 34 has no share for the golden channel A and V6, and the ladder is not retried\n", + ret); + return ret; + } + } + + /* * Golden-image replay (gsp_phys_rmcall_golden=1 + gsp_phys_gpu=1): * replay the stock client's post-load golden channel allocations after * the ladder's VAS/TSG exist and before the ladder's own channel/GR @@ -12274,7 +12425,9 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0) boot_stage.rmcall_gr_promote == GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_COMMIT || boot_stage.rmcall_gr_promote == - GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC) && + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_SIBLING_EXEC || + boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) && gsp_phys_rmcall_engine_is_gr()) { pr_err("vgsp-boot: rm-call: stage 3 aborted: mode-%u golden replay failed (%d); the golden construct is atomic and is not retried\n", boot_stage.rmcall_gr_promote, ret); @@ -12342,6 +12495,24 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0) &boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_RAMFC]); gsp_phys_rmcall_ctx_diag_decode("post-V6-alloc pre-A-commit"); } + /* + * Mode 34 readback point: A's 9-entry promote ran inside the replay + * before V6's alloc, so this is the first point where V6's + * instance/RAMFC can be checked against the already-committed TSG + * context; the decode is the same decisive fields (instance WFI pointer + * pair + both ENG_CTX_VALID target words). + */ + if (gsp_phys_rmcall_engine_is_gr() && + boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) { + gsp_phys_rmcall_ctx_diag_dump( + "INSTANCE (post-A-commit post-V6-alloc)", + &boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_INSTANCE]); + gsp_phys_rmcall_ctx_diag_dump( + "RAMFC (post-A-commit post-V6-alloc)", + &boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_RAMFC]); + gsp_phys_rmcall_ctx_diag_decode("post-A-commit post-V6-alloc"); + } /* * Mode 28: the golden execution lane. The golden replay allocated the @@ -12386,6 +12557,14 @@ static int gsp_phys_rmcall_stage3(void __iomem *bar0) GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_EXEC_LAST) { pr_info("vgsp-boot: rm-call: mode 29: skipping the TURING_COMPUTE_A (0xc5c0) object alloc: every GR object construct on the client-wired golden channel deadlocked the server (m3a-m3g); the golden execution lane runs object-less (the SET_OBJECT 0xc5c0 binding is the live experiment)\n"); boot_stage.rmcall_gr_object = 0; + } else if (boot_stage.rmcall_gr_promote == + GSP_PHYS_RMCALL_GR_PROMOTE_GOLDEN_CTXSHARE) { + pr_info("vgsp-boot: rm-call: mode 34: V6 0x%x constructs no GR object; the shared golden channel A 0x%x already carries TURING_A 0x%x + TURING_COMPUTE_A 0x%x (rmcall_gr_object=0x%x) and a V6 construct would overwrite the shared subcontext header\n", + boot_stage.rmcall_channel, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CHANNEL, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_GR_OBJ, + GSP_PHYS_RMCALL_HANDLE_GOLDEN_CTXSHARE_GR_OBJ, + boot_stage.rmcall_gr_object); } else { ret = gsp_phys_rmcall_alloc_gr_object(bar0); if (ret) { diff --git a/kernel/gsp_phys_gpu.c b/kernel/gsp_phys_gpu.c index 528caeb..339aa06 100644 --- a/kernel/gsp_phys_gpu.c +++ b/kernel/gsp_phys_gpu.c @@ -551,8 +551,8 @@ static int gsp_phys_rmcall_gr_promote_set(const char *val, return ret; if (gsp_phys_rmcall_gr_promote_param < 0 || - gsp_phys_rmcall_gr_promote_param > 33) { - pr_warn("vgsp: gsp_phys_rmcall_gr_promote=%d is outside 0..33; clamping to 0 (no PROMOTE_CTX)\n", + gsp_phys_rmcall_gr_promote_param > 34) { + pr_warn("vgsp: gsp_phys_rmcall_gr_promote=%d is outside 0..34; clamping to 0 (no PROMOTE_CTX)\n", gsp_phys_rmcall_gr_promote_param); gsp_phys_rmcall_gr_promote_param = 0; } @@ -571,7 +571,7 @@ module_param_cb(gsp_phys_rmcall_gr_promote, &gsp_phys_rmcall_gr_promote_ops, &gsp_phys_rmcall_gr_promote_param, 0644); MODULE_PARM_DESC(gsp_phys_rmcall_gr_promote, - "GR context-promote bisection (default 0: never send PROMOTE_CTX; 1: single entry init+nonmapped=0 size=raw+subctx, the p6h1b shape; 2: two-phase PA-only nonmapped then VA-only with the same fields; 3: single entry size=raw; 4: two-phase with a zero-field VA-only step; 5: batched entryCount=2 PA-only init + zero-field VA-only; 6: no-entry hVirtMemory, NV01_MEMORY_VIRTUAL (class 0x70) on the VAS at the context window + PROMOTE_CTX entryCount=0; 7: no-entry bare virtAddress+size, entryCount=0; 8: single MAIN entry like 1 on the KMD-managed channel (V6 SKIP_CTXBUFFER_ALLOC), server creates the virtual context memdesc itself; 9: same V6 channel with the stock client's zero-field VA-only MAIN entry, bInitialize=0; 10: same V6 channel, GPU_INITIALIZE_CTX (0x2080012d) installs the physical context memdesc then PROMOTE_CTX entryCount=0 bare virtAddress+size; 11: mode 10 with a server FBMEM context buffer, VIDMEM physAttr 0x4; 12: V5 UVM/vGPU-faithful PA-only MAIN entry (bInitialize=1 bNonmapped=1, no VA) and continue to BIND/SCHEDULE/submit; 13: mode 12 plus a PATCH PA-only promote first, sized from GET_ENGINE_CONTEXT_PROPERTIES engineId 0x10; 14: KMD channel (V6) + PATCH PA-only then the single-entry MAIN init+VA (the flow that books the context VA and commits ENG_CTX_VALID); 15: mode 14 with a server FBMEM MAIN context buffer and VIDMEM physAttr 0x4; 16: mode 14 plus a PM ctxsw PA-only promote first; 17: no client PROMOTE_CTX, paired with gsp_phys_srv_ctxbuf=1 so the GSP-RM server-manages and commits the context; 18: mode 8 shape plus a client FERMI_CONTEXT_SHARE_A on the TSG (hContextShare) whose construct allocates the GraphicsSubcontext; 19: the stock guest's exact 9-entry promote set (MAIN 2 MB PA+VA, PATCH, global buffers, FECS event, private access maps) on V5; 20: mode 6's no-entry hVirtMemory promote on the KMD-managed V6 channel (SKIP_CTXBUFFER_ALLOC), untested - the combination the V5 refusal (0x1f) points at; 21: mode 7's no-entry bare (virtAddress, size) promote on the V6 KMD-managed channel; 22: GPU_INITIALIZE_CTX then mode 20's no-entry hVirtMemory promote; 23: the hVirtMemory no-entry promote backed by the server-owned FBMEM context object (GSP_RM_ALLOC class 0x40) on the server-managed shared V5 channel - a physical memdesc reaches the per-buffer commit without the ADDR_VIRTUAL vaList gate; 24: the reference FWCLIENT non-3D set {MAIN, PM, PATCH, FECS_EVENT, PRIV_ACCESS_MAP, UNRESTRICTED_PRIV_ACCESS_MAP} as one control on the V6 KMD-managed channel, combined PA+VA entries where the ladder has a VA (MAIN PDB-mapped; PATCH/FECS/UNRESTRICTED booking-only), PM and PRIV_ACCESS_MAP init-only (no VA source), unsourceable buffers omitted and logged); 25: the real golden-client construct on the golden replay handles (captured channel flags 0x00200320 + single atomic 9-entry PROMOTE_CTX on the golden channel), then the V6 FWCLIENT follow-on promote on the golden-initialized GR-instance state; requires gsp_phys_rmcall_golden=1 and gsp_phys_rmcall_engine=1; 26/27: golden9 promote on the V6 execution channel (dead ends); 28: the golden execution lane - the golden channel gets a client GPFIFO ring (VA_GPFIFO), a sysmem host-writable USERD and the ladder TSG as parent, keeps the 9-entry promote, and the ladder switches execution to it (V6 stays unscheduled); 29: mode 28 minus the TURING_COMPUTE_A (0xc5c0) object construct on the client-wired golden channel (every GR object construct there deadlocks RM), i.e. the construct-last golden execution lane; 30: the TSG-sibling split: a clean unwired golden sibling A (captured shape, TSG-parented) runs the 9-entry promote and its own TURING_A first-init while V6 stays the execution channel (no swap) with its FWCLIENT promote suppressed; 31: mode-25's device-parented golden first-init (9-entry promote + TURING_A) followed by a clean TSG-parented sibling A whose 9-entry promote runs after V6's own constructs, so the TSG-wide server commit sets V6's instance pointer + RAMFC while V6 stays the execution channel and its own promote is suppressed; 33: mode 31 with the TSG sibling A wired as the object-carrying execution channel (mode-29 golden_exec_shape wiring: ring at VA_GPFIFO, 8 GPFIFO entries, ladder sysmem USERD, TSG parent, host USERD control offset 0 via shape bit7), A's 9-entry promote (skip_fecs), a TURING_COMPUTE_A (0xc5c0) object constructed on A after the accepted promote, and boot_stage.rmcall_channel swapped to A for stage 4 and all forwarded launches; requires gsp_phys_rmcall_golden=1, gsp_phys_gpu=1 and gsp_phys_rmcall_golden_exec_shape=0xE7; values outside 0..33 clamp to 0 with a warning"); + "GR context-promote bisection (default 0: never send PROMOTE_CTX; 1: single entry init+nonmapped=0 size=raw+subctx, the p6h1b shape; 2: two-phase PA-only nonmapped then VA-only with the same fields; 3: single entry size=raw; 4: two-phase with a zero-field VA-only step; 5: batched entryCount=2 PA-only init + zero-field VA-only; 6: no-entry hVirtMemory, NV01_MEMORY_VIRTUAL (class 0x70) on the VAS at the context window + PROMOTE_CTX entryCount=0; 7: no-entry bare virtAddress+size, entryCount=0; 8: single MAIN entry like 1 on the KMD-managed channel (V6 SKIP_CTXBUFFER_ALLOC), server creates the virtual context memdesc itself; 9: same V6 channel with the stock client's zero-field VA-only MAIN entry, bInitialize=0; 10: same V6 channel, GPU_INITIALIZE_CTX (0x2080012d) installs the physical context memdesc then PROMOTE_CTX entryCount=0 bare virtAddress+size; 11: mode 10 with a server FBMEM context buffer, VIDMEM physAttr 0x4; 12: V5 UVM/vGPU-faithful PA-only MAIN entry (bInitialize=1 bNonmapped=1, no VA) and continue to BIND/SCHEDULE/submit; 13: mode 12 plus a PATCH PA-only promote first, sized from GET_ENGINE_CONTEXT_PROPERTIES engineId 0x10; 14: KMD channel (V6) + PATCH PA-only then the single-entry MAIN init+VA (the flow that books the context VA and commits ENG_CTX_VALID); 15: mode 14 with a server FBMEM MAIN context buffer and VIDMEM physAttr 0x4; 16: mode 14 plus a PM ctxsw PA-only promote first; 17: no client PROMOTE_CTX, paired with gsp_phys_srv_ctxbuf=1 so the GSP-RM server-manages and commits the context; 18: mode 8 shape plus a client FERMI_CONTEXT_SHARE_A on the TSG (hContextShare) whose construct allocates the GraphicsSubcontext; 19: the stock guest's exact 9-entry promote set (MAIN 2 MB PA+VA, PATCH, global buffers, FECS event, private access maps) on V5; 20: mode 6's no-entry hVirtMemory promote on the KMD-managed V6 channel (SKIP_CTXBUFFER_ALLOC), untested - the combination the V5 refusal (0x1f) points at; 21: mode 7's no-entry bare (virtAddress, size) promote on the V6 KMD-managed channel; 22: GPU_INITIALIZE_CTX then mode 20's no-entry hVirtMemory promote; 23: the hVirtMemory no-entry promote backed by the server-owned FBMEM context object (GSP_RM_ALLOC class 0x40) on the server-managed shared V5 channel - a physical memdesc reaches the per-buffer commit without the ADDR_VIRTUAL vaList gate; 24: the reference FWCLIENT non-3D set {MAIN, PM, PATCH, FECS_EVENT, PRIV_ACCESS_MAP, UNRESTRICTED_PRIV_ACCESS_MAP} as one control on the V6 KMD-managed channel, combined PA+VA entries where the ladder has a VA (MAIN PDB-mapped; PATCH/FECS/UNRESTRICTED booking-only), PM and PRIV_ACCESS_MAP init-only (no VA source), unsourceable buffers omitted and logged); 25: the real golden-client construct on the golden replay handles (captured channel flags 0x00200320 + single atomic 9-entry PROMOTE_CTX on the golden channel), then the V6 FWCLIENT follow-on promote on the golden-initialized GR-instance state; requires gsp_phys_rmcall_golden=1 and gsp_phys_rmcall_engine=1; 26/27: golden9 promote on the V6 execution channel (dead ends); 28: the golden execution lane - the golden channel gets a client GPFIFO ring (VA_GPFIFO), a sysmem host-writable USERD and the ladder TSG as parent, keeps the 9-entry promote, and the ladder switches execution to it (V6 stays unscheduled); 29: mode 28 minus the TURING_COMPUTE_A (0xc5c0) object construct on the client-wired golden channel (every GR object construct there deadlocks RM), i.e. the construct-last golden execution lane; 30: the TSG-sibling split: a clean unwired golden sibling A (captured shape, TSG-parented) runs the 9-entry promote and its own TURING_A first-init while V6 stays the execution channel (no swap) with its FWCLIENT promote suppressed; 31: mode-25's device-parented golden first-init (9-entry promote + TURING_A) followed by a clean TSG-parented sibling A whose 9-entry promote runs after V6's own constructs, so the TSG-wide server commit sets V6's instance pointer + RAMFC while V6 stays the execution channel and its own promote is suppressed; 33: mode 31 with the TSG sibling A wired as the object-carrying execution channel (mode-29 golden_exec_shape wiring: ring at VA_GPFIFO, 8 GPFIFO entries, ladder sysmem USERD, TSG parent, host USERD control offset 0 via shape bit7), A's 9-entry promote (skip_fecs), a TURING_COMPUTE_A (0xc5c0) object constructed on A after the accepted promote, and boot_stage.rmcall_channel swapped to A for stage 4 and all forwarded launches; requires gsp_phys_rmcall_golden=1, gsp_phys_gpu=1 and gsp_phys_rmcall_golden_exec_shape=0xE7; 34: the context-share golden lane - a FERMI_CONTEXT_SHARE_A (0x9067, TSG-parented, ladder SHARED_MANAGEMENT VAS, flags SYNC -> VEID 0) is allocated before the replay; the golden replay's channel A (0xcaf00054) joins the ladder TSG through hContextShare with hVASpace=0 while keeping the captured unwired shape, runs the 9-entry promote and carries TURING_A (0xc597) + TURING_COMPUTE_A (0xc5c0, handle 0xcaf00090); V6 (0xcaf00034) attaches to the same share, constructs nothing (a later construct would overwrite the shared VEID's subcontext header) and its promote is suppressed, so stage 4 BIND/SCHEDULE/token/GR-sem and every forwarded launch run on V6 inside A's shared VEID/subcontext; requires gsp_phys_rmcall_golden=1, gsp_phys_gpu=1 and gsp_phys_rmcall_vas_shared=1; values outside 0..34 clamp to 0 with a warning"); int gsp_phys_rmcall_gr_promote_value(void) { |