summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--kernel/gsp_phys_booter.c38
-rw-r--r--kernel/gsp_sema_service.c84
-rw-r--r--tools/cuda-ab/victory_mail.cu2
3 files changed, 59 insertions, 65 deletions
diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index 9e45937..fdd20db 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -16254,7 +16254,7 @@ static void gsp_phys_gr_rewrite_data_vars(u32 *w, u32 n_words,
{
u32 rew = 0, k;
- if (!req->n_data_wins)
+ if (!req->n_data)
return;
/* Scan every 4-byte boundary: the demo's bank-4 base pointer sits at
* a 4-aligned (not 8-aligned) offset, which an 8-stride scan misses. */
@@ -16265,12 +16265,11 @@ static void gsp_phys_gr_rewrite_data_vars(u32 *w, u32 n_words,
if (v < 0x10000000ULL || v >= (1ULL << 47) || (v & 3))
continue;
- wb = v & ~0x1fffffULL;
- for (x = 0; x < req->n_data_wins; x++) {
- if (req->data_wins[x].va_base != wb)
+ wb = v & ~0xfffULL;
+ for (x = 0; x < req->n_data; x++) {
+ if (req->data[x].va != wb)
continue;
- v = req->data_wins[x].reloc_base |
- (v & 0x1fffffULL);
+ v = req->data[x].reloc_va | (v & 0xfffULL);
w[k] = (u32)v;
w[k + 1] = (u32)(v >> 32);
rew++;
@@ -16296,20 +16295,30 @@ static int gsp_phys_rmcall_gr_data_map(void __iomem *bar0, u64 va, u64 hpa,
u32 epoch)
{
struct gsp_phys_boot_chunk *pt;
- u32 k, ip;
+ u32 ip;
u64 pte;
(void)bar0;
(void)epoch;
- if (va < GSP_PHYS_RMCALL_VA_GR_DATA_BASE)
+ /*
+ * m110: the engine's SM walker only honours the commit-time tree, so
+ * data pages are relocated into the main small-PT window
+ * (GSP_PHYS_RMCALL_VA_BASE + 0x10000 ..), whose slots above 0x10f are
+ * fresh PTE lines - the proven window-0 pattern. Slot 0x1f0 (VA
+ * 0x1f0000) is reserved for the mail injection; 0x180..0x18f stay
+ * clear of the legacy guest slot.
+ */
+ if (va < GSP_PHYS_RMCALL_VA_BASE + 0x10000ULL ||
+ va >= GSP_PHYS_RMCALL_VA_BASE + 0x200000ULL)
return -ERANGE;
- k = (u32)((va - GSP_PHYS_RMCALL_VA_GR_DATA_BASE) >> 21);
- if (k >= GSP_PHYS_RMCALL_GR_DATA_WINS)
+ ip = (u32)((va >> 12) & 0x1ffu);
+ if (ip < 0x110u || (ip >= 0x180u && ip < 0x190u))
return -ERANGE;
- pt = &boot_stage.rmcall_gr_data_pt[k];
+ pt = &boot_stage.rmcall_chunks[GSP_PHYS_RMCALL_BUF_PT];
+ if (pt->pages)
+ pt = pt;
if (!pt->pages)
return -EIO;
- ip = (u32)((va >> 12) & 0x1ffu);
pte = gsp_phys_rmcall_pte(hpa);
if (gsp_phys_rmcall_gr_pte_kind_value())
pte |= (u64)gsp_phys_rmcall_gr_pte_kind_value() << 56;
@@ -16661,9 +16670,8 @@ int gsp_phys_backend_guest_gr_submit(const struct gsp_phys_gr_submit *req,
boot_stage.rmcall_gr_data_pt
[GSP_PHYS_RMCALL_GR_DATA_WINS - 1].pages);
{
- u64 mva = GSP_PHYS_RMCALL_VA_GR_DATA_BASE +
- (((u64)GSP_PHYS_RMCALL_GR_DATA_WINS - 1)
- << 21);
+ u64 mva = GSP_PHYS_RMCALL_VA_BASE +
+ 0x1f0000ULL;
int mret = -EINVAL;
if (boot_stage.rmcall_gr_mail.pages)
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 844d2c9..caaaae3 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -9555,7 +9555,13 @@ static bool svc_gr_page_inline_covers(u64 page)
* Resolution only happens on the service thread; the work item turns the
* guest PAs into HPAs and the backend plants the identity PTEs.
*/
-#define SVC_GR_DATA_RELOC_BASE 0x100000000ULL
+#define SVC_GR_DATA_RELOC_BASE 0x110000ULL
+/*
+ * Per candidate pointer, map at most this many pages from the pointer's
+ * page onward (the engine-visible main-PT window has ~220 free 4 KB slots;
+ * the demo buffers are small on the native path).
+ */
+#define SVC_GR_DATA_PAGES_PER_PTR 64u
static void svc_gr_data_fill(struct nv_vgpu_device *vgpu)
{
@@ -9605,62 +9611,42 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu)
wins[n_cand++] = wv;
}
}
+ /* m110: one contiguous run of pages per candidate pointer, relocated
+ * into the engine-walkable main-PT window (0x110000 onward, fresh
+ * PTE lines; the commit-time snapshot only preserves that window). */
for (w = 0; w < n_cand; w++) {
- int wi;
+ u64 base = wins[w] & ~0x1fffffULL;
+ u32 p;
- for (wi = -1; wi <= 16; wi++) {
- u64 wv = wins[w] + (u64)wi * 0x200000ULL;
- u64 rbase;
- u32 got = 0, p, x;
+ for (p = 0; p < SVC_GR_DATA_PAGES_PER_PTR; p++) {
+ u64 va = wins[w] + (u64)p * SEMA_PAGE_SIZE;
+ u64 pa = 0;
+ u32 ap = 0, x;
+ struct svc_gr_data *rec;
- if (wi < 0 && wins[w] < 0x200000ULL)
- continue;
- if (wv >= (1ULL << 47))
+ if (va >= (1ULL << 47))
break;
- for (x = 0; x < svc_fwd_req.gr_n_data_wins; x++)
- if (svc_fwd_req.gr_data_wins[x].va_base == wv)
+ for (x = 0; x < svc_fwd_req.gr_n_data; x++)
+ if (svc_fwd_req.gr_data[x].va == va)
break;
- if (x == svc_fwd_req.gr_n_data_wins) {
- if (svc_fwd_req.gr_n_data_wins >=
- SVC_GR_DATA_WINS_MAX)
- return;
- svc_fwd_req.gr_data_wins[x].va_base = wv;
- rbase = SVC_GR_DATA_RELOC_BASE +
- ((u64)svc_fwd_req.gr_n_data_wins <<
- 21);
- svc_fwd_req.gr_data_wins[x].reloc_base =
- rbase;
- svc_fwd_req.gr_n_data_wins++;
- } else {
- rbase =
- svc_fwd_req.gr_data_wins[x].reloc_base;
- }
- for (p = 0; p < 512; p++) {
- u64 va = wv + (u64)p * SEMA_PAGE_SIZE;
- u64 pa = 0;
- u32 ap = 0;
- struct svc_gr_data *rec;
-
- if (svc_resolve_va(vgpu, va, &pa, &ap))
- continue;
- if (svc_fwd_req.gr_n_data >= SVC_GR_DATA_MAX)
- return;
- rec = &svc_fwd_req.gr_data
- [svc_fwd_req.gr_n_data++];
- rec->va = va;
- rec->pa = pa & ~(u64)(SEMA_PAGE_SIZE - 1);
- rec->reloc_va = rbase + (va & 0x1fffffULL);
- rec->ap = (u8)svc_ap_of(ap);
- got++;
- }
- if (!got && wi > 0)
+ if (x < svc_fwd_req.gr_n_data)
+ continue; /* already mapped */
+ if (svc_resolve_va(vgpu, va, &pa, &ap))
break;
if (svc_fwd_req.gr_n_data >= SVC_GR_DATA_MAX)
return;
- }
- }
- pr_info("vgsp-sema: svc-gr data fill: %u candidate ptrs, %u windows, %u pages resolved\n",
- n_cand, svc_fwd_req.gr_n_data_wins, svc_fwd_req.gr_n_data);
+ rec = &svc_fwd_req.gr_data[svc_fwd_req.gr_n_data];
+ rec->va = va;
+ rec->pa = pa & ~(u64)(SEMA_PAGE_SIZE - 1);
+ rec->reloc_va = SVC_GR_DATA_RELOC_BASE +
+ ((u64)svc_fwd_req.gr_n_data << 12);
+ rec->ap = (u8)svc_ap_of(ap);
+ svc_fwd_req.gr_n_data++;
+ }
+ (void)base;
+ }
+ pr_info("vgsp-sema: svc-gr data fill: %u candidate ptrs, %u pages resolved (reloc 0x%llx+)\n",
+ n_cand, svc_fwd_req.gr_n_data, (u64)SVC_GR_DATA_RELOC_BASE);
}
static int svc_gr_req_fill(struct nv_vgpu_device *vgpu,
diff --git a/tools/cuda-ab/victory_mail.cu b/tools/cuda-ab/victory_mail.cu
index 296f7e6..e8136c3 100644
--- a/tools/cuda-ab/victory_mail.cu
+++ b/tools/cuda-ab/victory_mail.cu
@@ -115,7 +115,7 @@ int main(void)
{
cudaDeviceProp prop;
const int len = (int)strlen(MAIL);
- const int n = 1 << 20;
+ const int n = 1 << 15; /* sized for the native GR path: the engine-visible data window is 2 MB */
size_t bytes = (size_t)n * sizeof(float);
char *d_mail, h_mail[128];
float *h_a, *h_b, *h_c;
generated by cgit v1.2.3 (git 2.39.1) at 2026-10-12 10:35:24 +0000