diff options
| author | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-14 12:44:32 -0400 |
|---|---|---|
| committer | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-14 12:44:32 -0400 |
| commit | 074f493c8bf51e487f5cdff30adbcbc65ad7b79b (patch) | |
| tree | ebd9b78e51c1d3b17d22cfb15b05ab4e06285b8d /kernel/gsp_phys_booter.c | |
| parent | 80af379d26334970462f7f14713ab2ca39747280 (diff) | |
| download | open-iov-074f493c8bf51e487f5cdff30adbcbc65ad7b79b.tar.gz open-iov-074f493c8bf51e487f5cdff30adbcbc65ad7b79b.zip | |
vgsp: user-lane GPFIFO depth must match the host ring modulus (8)
The host publishes every forwarded launch into the one-page ladder ring
modulo GSP_PHYS_RMCALL_GPFIFO_ENTRIES (8). Commit 3b82047c declared the
mode-35 user-lane channel with 1024 entries for capture fidelity; the
engine's GET/PUT then wrap at 1024 while the host wraps at 8, so at the
host's first wrap (GP_PUT 7 -> 0) the engine desyncs and never fetches
another entry: launches 1..7 green, the 8th and all later ones silent
(qmd_rel=5, ch_rel=0), which is the 'launch lottery'.
Live proof (r138): the failure fires exactly at slot=7 (first wrap) and
every launch after it fails; vecadd_loop (100 launches) dies at the wrap.
M1's two-launch acceptance never wrapped, so the regression was latent.
Set the declared count back to the host ring constant; the wrap-only
fidelity difference against the captured real-client ring is documented at
the define.
Diffstat (limited to 'kernel/gsp_phys_booter.c')
| -rw-r--r-- | kernel/gsp_phys_booter.c | 17 |
1 files changed, 15 insertions, 2 deletions
diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c index 72cda10..879fca4 100644 --- a/kernel/gsp_phys_booter.c +++ b/kernel/gsp_phys_booter.c @@ -1253,8 +1253,21 @@ bool gsp_phys_boot_requested(void) #define GSP_PHYS_RMCALL_CHANNEL_FLAG_USER_LANE \ (u32)(GSP_PHYS_RMCALL_CHANNEL_FLAG_PRIVILEGED | (4u << 8) | \ (1u << 21)) -/* Captured user-lane GPFIFO depth (#825: 1024 entries). */ -#define GSP_PHYS_RMCALL_USER_LANE_GPFIFO_ENTRIES 1024u +/* + * User-lane GPFIFO depth. MUST equal the host ring's push modulus + * (GSP_PHYS_RMCALL_GPFIFO_ENTRIES, 8): the host publishes into the + * one-page ladder ring modulo 8, so if the channel declares a different + * count the engine's GET/PUT wrap desyncs at the host's first wrap + * (GP_PUT 7 -> 0 while the engine's GET keeps counting to 8) and every + * later launch on the boot is never fetched (r138: launches 1..7 green, + * the 8th and all after it silent with `qmd_rel=5 ch_rel=0`). The captured + * #825 value (1024) describes the real client's own much larger ring, not + * this emulated one; the campaign's 2026-10-06 "fidelity" change to 1024 + * was a wrap-only difference that broke exactly here (M1's two-launch + * acceptance never wrapped). Keep the two ends equal. + */ +#define GSP_PHYS_RMCALL_USER_LANE_GPFIFO_ENTRIES \ + GSP_PHYS_RMCALL_GPFIFO_ENTRIES #define GSP_PHYS_RMCALL_CHANNEL_INTERNALFLAGS_USER_LANE 0x1du #define GSP_PHYS_RMCALL_TSG_FAULT_BUF_SIZE 0x2000u #define GSP_PHYS_RMCALL_USER_LANE_PM_RAW_SIZE 0xd100u |