diff options
Diffstat (limited to 'kernel/gsp_phys_booter.c')
| -rw-r--r-- | kernel/gsp_phys_booter.c | 30 |
1 files changed, 22 insertions, 8 deletions
diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c index 2f3202e..534a443 100644 --- a/kernel/gsp_phys_booter.c +++ b/kernel/gsp_phys_booter.c @@ -1668,6 +1668,20 @@ enum gsp_phys_rmcall_buf { #define GSP_PHYS_RMCALL_TIMEOUT_MS 2000u #define GSP_PHYS_RMCALL_POLL_MS 5u /* + * SESSION 148 perf (termbin review, verified in-tree): every RPC reply wait + * polled with msleep(5). With several quanta per round trip and ~8 + * serialized PDE_UPDATE forwards per forwarded GR launch, the 5 ms quantum + * dominated per-launch latency (~95% of the ~189 ms host service time per + * launch). Wait with sub-millisecond quanta instead: the reply is observed + * within one quantum, cutting each round trip to ~1 ms. POLL_MS remains for + * the poll accounting prints. + */ +#define GSP_PHYS_RMCALL_POLL_US 150u +static inline void gsp_phys_rmcall_wait_quantum(void) +{ + usleep_range(GSP_PHYS_RMCALL_POLL_US, GSP_PHYS_RMCALL_POLL_US + 200u); +} +/* * GR launch release poll window. A CE copy completes in single-digit ms, but * a GR launch can force the engine's first context switch for the channel; * the FECS ctxsw path is measured in fractions of a second, so the window is @@ -15371,7 +15385,7 @@ static int gsp_phys_rmcall_gr_sem_push_locked(void __iomem *bar0) gsp_phys_status_drain(NULL); return -ETIMEDOUT; } - msleep(GSP_PHYS_RMCALL_POLL_MS); + gsp_phys_rmcall_wait_quantum(); } gpget = get_unaligned_le32((u8 *)userd + gsp_phys_rmcall_userd_ctl_off() + GSP_PHYS_RMCALL_USERD_GPGET_OFF); @@ -15471,7 +15485,7 @@ static void gsp_phys_rmcall_ctx_image_probe(void __iomem *bar0) offs[k], poll); break; } - msleep(GSP_PHYS_RMCALL_POLL_MS); + gsp_phys_rmcall_wait_quantum(); } if (poll != GSP_PHYS_RMCALL_SEM_PAYLOAD) break; @@ -15572,7 +15586,7 @@ static int gsp_phys_rmcall_ce_copy_va(void __iomem *bar0, u64 src_va, *sem_out = poll; return -ETIMEDOUT; } - msleep(GSP_PHYS_RMCALL_POLL_MS); + gsp_phys_rmcall_wait_quantum(); } } @@ -15755,7 +15769,7 @@ int gsp_phys_backend_ce_copy_submit(u64 src_va, u64 dst_va, u64 sem_va, gsp_phys_work_timed_out); return -ETIMEDOUT; } - msleep(GSP_PHYS_RMCALL_POLL_MS); + gsp_phys_rmcall_wait_quantum(); } /* @@ -16274,7 +16288,7 @@ static int gsp_phys_guest_ce_copy_common(u64 src_hpa, u32 src_off, u64 dst_hpa, ret = -ETIMEDOUT; goto out; } - msleep(GSP_PHYS_RMCALL_POLL_MS); + gsp_phys_rmcall_wait_quantum(); } /* @@ -21235,7 +21249,7 @@ gr_page_unmapped: ret = -ETIMEDOUT; goto out; } - msleep(GSP_PHYS_RMCALL_POLL_MS); + gsp_phys_rmcall_wait_quantum(); } gsp_phys_gr_ok++; @@ -21640,7 +21654,7 @@ gr_page_unmapped: } if (time_after_eq(jiffies, dl)) break; - msleep(GSP_PHYS_RMCALL_POLL_MS); + gsp_phys_rmcall_wait_quantum(); } pr_info("vgsp-gr: outcome retry %u: fresh ack %s (qmd_rel=0x%08x want=0x%08x ch_rel=0x%08x after %u ms)\n", nw_attempts, @@ -22693,7 +22707,7 @@ static int gsp_phys_rmcall_stage4(void __iomem *bar0) gsp_phys_status_drain(NULL); return -ETIMEDOUT; } - msleep(GSP_PHYS_RMCALL_POLL_MS); + gsp_phys_rmcall_wait_quantum(); } if (memcmp(gsp_phys_chunk_page_va(src, 0), |