diff options
Diffstat (limited to 'kernel/gsp_phys_booter.c')
| -rw-r--r-- | kernel/gsp_phys_booter.c | 69 |
1 files changed, 57 insertions, 12 deletions
diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c index 2c0fdc0..badf14f 100644 --- a/kernel/gsp_phys_booter.c +++ b/kernel/gsp_phys_booter.c @@ -12765,6 +12765,12 @@ EXPORT_SYMBOL_GPL(gsp_phys_backend_guest_ce_copy); #define GSP_PHYS_RMCALL_GR_LOAD_INLINE_QMD_DATA 0x320u /* Header count: ADDRESS_A + ADDRESS_B + 64 QMD words. */ #define GSP_PHYS_RMCALL_GR_QMD_INLINE_WORDS 66u +/* Whole run: one method header + the inline-QMD data words. */ +#define GSP_PHYS_RMCALL_GR_QMD_WINDOW_WORDS 67u +/* Dwords the per-channel compute init emits (see build_gr_init). */ +#define GSP_PHYS_RMCALL_GR_INIT_WORDS 152u +/* PCAS(4) + REPORT_SEM(8). */ +#define GSP_PHYS_RMCALL_GR_TAIL_WORDS 12u #define GSP_PHYS_RMCALL_GR_QMD_REL_PAYLOAD GSP_PHYS_RMCALL_SEM_PAYLOAD #define GSP_PHYS_RMCALL_GR_SEM_PAYLOAD 0xc0ffee02u /* @@ -12874,21 +12880,56 @@ static u32 *gsp_phys_rmcall_build_gr_init(u32 *p) * channel long before the launch), so the replay must not either. * Returns the block size in dwords. */ -static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb, const u32 *qmd, - u64 qmd_va, u64 sem_va, - u32 sem_payload) +static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb, + const struct gsp_phys_gr_submit *req, + const u32 *qmd, u64 qmd_va, + u64 sem_va, u32 sem_payload) { u32 *p = pb; p = gsp_phys_rmcall_build_gr_init(p); - *p++ = ((u32)GSP_PHYS_RMCALL_METHOD_IMM << 29) | - (GSP_PHYS_RMCALL_GR_QMD_INLINE_WORDS << 16) | - ((u32)GSP_PHYS_RMCALL_GR_SUBCHANNEL << 13) | - (GSP_PHYS_RMCALL_GR_INLINE_QMD_ADDRESS_A >> 2); - *p++ = (u32)(qmd_va >> 40); - *p++ = (u32)((qmd_va >> 8) & 0xffffffffu); - memcpy(p, qmd, GSP_PHYS_GR_QMD_WORDS * 4); - p += GSP_PHYS_GR_QMD_WORDS; + /* + * Verbatim replay: copy the captured GRCE prelude and the guest's + * own inline-QMD run, keeping the method headers, and relocate only + * the ADDRESS_A/B data words and the 64 QMD words exactly like the + * rebuilt window. Only a contiguous 0x318 run is replayed; anything + * else falls back to the rebuilt window. + */ + if (req && req->raw && + req->raw_words >= + (u32)req->raw_qmd_hdr + GSP_PHYS_RMCALL_GR_QMD_WINDOW_WORDS && + req->raw_qmd_end == + (u32)req->raw_qmd_hdr + GSP_PHYS_RMCALL_GR_QMD_WINDOW_WORDS && + GSP_PHYS_RMCALL_GR_INIT_WORDS + + (u32)req->raw_qmd_hdr + GSP_PHYS_RMCALL_GR_QMD_WINDOW_WORDS + + GSP_PHYS_RMCALL_GR_TAIL_WORDS <= + GSP_PHYS_RMCALL_GR_METHOD_WORDS) { + u32 pre = (u32)req->raw_qmd_hdr + + GSP_PHYS_RMCALL_GR_QMD_WINDOW_WORDS; + + memcpy(p, req->raw, pre * 4); + p[req->raw_qmd_hdr + 1] = (u32)(qmd_va >> 40); + p[req->raw_qmd_hdr + 2] = + (u32)((qmd_va >> 8) & 0xffffffffu); + memcpy(p + req->raw_qmd_hdr + 3, qmd, + GSP_PHYS_GR_QMD_WORDS * 4); + p += pre; + pr_info("vgsp-gr: verbatim replay prelude+window words=%u qmd_hdr=%u\n", + pre, req->raw_qmd_hdr); + } else { + if (req && req->raw) + pr_info("vgsp-gr: raw prelude unusable n=%u hdr=%u end=%u; rebuilt window\n", + req->raw_words, req->raw_qmd_hdr, + req->raw_qmd_end); + *p++ = ((u32)GSP_PHYS_RMCALL_METHOD_IMM << 29) | + (GSP_PHYS_RMCALL_GR_QMD_INLINE_WORDS << 16) | + ((u32)GSP_PHYS_RMCALL_GR_SUBCHANNEL << 13) | + (GSP_PHYS_RMCALL_GR_INLINE_QMD_ADDRESS_A >> 2); + *p++ = (u32)(qmd_va >> 40); + *p++ = (u32)((qmd_va >> 8) & 0xffffffffu); + memcpy(p, qmd, GSP_PHYS_GR_QMD_WORDS * 4); + p += GSP_PHYS_GR_QMD_WORDS; + } p = gsp_phys_rmcall_push_gr(p, GSP_PHYS_RMCALL_GR_SEND_PCAS_A, (u32)(qmd_va >> 8)); p = gsp_phys_rmcall_push_gr(p, @@ -13554,7 +13595,7 @@ gr_page_unmapped: qmd_dst, GSP_PHYS_GR_QMD_WORDS); } - words = gsp_phys_rmcall_build_gr_launch(methods, qmd, + words = gsp_phys_rmcall_build_gr_launch(methods, req, qmd, qmd_dst, ch_sem_va, want_ch); memcpy(gsp_phys_chunk_page_va(pb, 0), methods, words * 4); @@ -16029,6 +16070,10 @@ static_assert(offsetof(struct gsp_phys_gr_inline, data) == 12 && static_assert((2u + 1u + GSP_PHYS_RMCALL_GR_QMD_INLINE_WORDS + 8u) <= GSP_PHYS_RMCALL_GR_METHOD_WORDS, "the GR launch method block (compute init + SET_OBJECT + inline-QMD run + PCAS + REPORT_SEMAPHORE) must fit the bound"); +static_assert(GSP_PHYS_RMCALL_GR_INIT_WORDS + GSP_PHYS_GR_RAW_WORDS + + GSP_PHYS_RMCALL_GR_TAIL_WORDS <= + GSP_PHYS_RMCALL_GR_METHOD_WORDS, + "the verbatim GR launch block (compute init + raw tail + PCAS + REPORT_SEMAPHORE) must fit the bound"); static_assert(GSP_PHYS_RMCALL_GR_METHOD_WORDS * 4u <= GSP_PHYS_RMCALL_COPY_SIZE, "the GR launch method block must fit one module pushbuffer page"); |