vgsp: point inline-QMD launch at the guest QMD buffer + GR post-mortem logging

1 files changed, 72 insertions, 1 deletions

diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index b95bd81..7113663 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -13024,6 +13024,7 @@ int gsp_phys_backend_guest_gr_submit(const struct gsp_phys_gr_submit *req,
 	bool qmd_rel = false, ch_rel = false;
 	u64 qmd_rel_hpa = 0, ch_rel_hpa = 0;
 	u64 ch_sem_va = GSP_PHYS_RMCALL_VA_SEM + 4;
+	u64 qmd_dst = GSP_PHYS_RMCALL_VA_SRC;
 	u32 qmd_rel_off = 0, ch_rel_off = 0;
 	u32 poll = 0, w0 = 0, w1 = 0;
 	int ret = 0;
@@ -13351,6 +13352,38 @@ gr_page_unmapped:
 	}
 
 	/*
+	 * Inline-QMD destination: the CUDA driver points ADDRESS_A/B at its
+	 * own QMD buffer (turing_launch.c:1629) and the IQ2M copies the 64
+	 * inline words there before scheduling. Map the guest QMD page and
+	 * use it as the destination when the caller resolved it; a Tier-1
+	 * install keeps the guest VA (the guest sees its own QMD), a Tier-2
+	 * alias keeps the launch runnable, and an unresolved page keeps the
+	 * module scratch VA.
+	 */
+	if (req->qmd_hpa) {
+		u64 qmd_page = req->qmd_va & ~(u64)(GSP_PHYS_PAGE_SIZE - 1);
+
+		if (qmd_page &&
+		    !gsp_phys_rmcall_gr_map_va_locked(bar0, qmd_page,
+						      req->qmd_hpa, epoch)) {
+			qmd_dst = qmd_page;
+		} else {
+			u64 sh_va = 0;
+
+			if (n_shadow < GSP_PHYS_RMCALL_GR_SHADOW_PAGES &&
+			    !gsp_phys_rmcall_gr_map_shadow_locked(
+				    n_shadow, req->qmd_hpa, &sh_va)) {
+				n_shadow++;
+				qmd_dst = sh_va +
+					  (req->qmd_va &
+					   (GSP_PHYS_PAGE_SIZE - 1));
+			}
+		}
+		pr_info("vgsp-gr: inline-QMD dest va=0x%llx hpa=0x%llx -> 0x%llx (shadow=%u)\n",
+			req->qmd_va, req->qmd_hpa, qmd_dst, n_shadow);
+	}
+
+	/*
 	 * Channel release: author it into the guest semaphore page when the
 	 * caller resolved one, else into the module semaphore's second word.
 	 */
@@ -13380,9 +13413,12 @@ gr_page_unmapped:
 	wmb();
 
 	words = gsp_phys_rmcall_build_gr_launch(methods, qmd,
-						GSP_PHYS_RMCALL_VA_SRC,
+						qmd_dst,
 						ch_sem_va, want_ch);
 	memcpy(gsp_phys_chunk_page_va(pb, 0), methods, words * 4);
+	pr_info("vgsp-gr: qmd w0..3=%08x %08x %08x %08x w4..7=%08x %08x %08x %08x w23..25=%08x %08x %08x w32/33=%08x %08x w48/49=%08x %08x\n",
+		qmd[0], qmd[1], qmd[2], qmd[3], qmd[4], qmd[5], qmd[6], qmd[7],
+		qmd[23], qmd[24], qmd[25], qmd[32], qmd[33], qmd[48], qmd[49]);
 
 	ret = gsp_phys_rmcall_publish_locked(bar0, boot_stage.rmcall_va_pb,
 					     words, &sub);
@@ -13436,6 +13472,41 @@ gr_page_unmapped:
 					}
 				}
 			}
+			/*
+			 * Post-mortem: if the IQ2M landed the QMD in the guest
+			 * buffer, the inline window was processed; GP_GET shows
+			 * the front-end consumed the entry; the release word is
+			 * re-read one last time.
+			 */
+			{
+				u8 *ud = gsp_phys_chunk_page_va(
+					&boot_stage.rmcall_chunks
+						 [GSP_PHYS_RMCALL_BUF_USERD],
+					0);
+				u32 gpget_post = ud ?
+					get_unaligned_le32(
+						ud + GSP_PHYS_RMCALL_USERD_GPGET_OFF)
+					: 0;
+				u32 q0 = 0, q1 = 0, q23 = 0, q24 = 0, q25 = 0;
+				u32 relw = 0;
+
+				if (req->qmd_hpa) {
+					u8 *qp = (u8 *)__va(req->qmd_hpa);
+
+					q0 = get_unaligned_le32(qp);
+					q1 = get_unaligned_le32(qp + 4);
+					q23 = get_unaligned_le32(qp + 92);
+					q24 = get_unaligned_le32(qp + 96);
+					q25 = get_unaligned_le32(qp + 100);
+				}
+				if (qmd_rel_hpa)
+					relw = READ_ONCE(
+						*(u32 *)((u8 *)__va(qmd_rel_hpa) +
+							 qmd_rel_off));
+				pr_info("vgsp-gr: post-mortem gpget=%u->%u qmd_page w0=%08x w1=%08x w23..25=%08x %08x %08x relw=%08x\n",
+					sub.gpget, gpget_post, q0, q1, q23, q24,
+					q25, relw);
+			}
 			poll = w1;
 			ret = -ETIMEDOUT;
 			goto out;