diff options
| author | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-09-24 10:48:08 -0400 |
|---|---|---|
| committer | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-09-24 10:48:08 -0400 |
| commit | 93431bdd896d9e6d9a25bebef0fb1af308dd64cd (patch) | |
| tree | 2c6b44e8b3fca4edd24a6a904e370a5c3e899d22 /kernel | |
| parent | 775c61c34fdccb32179d0cb6239db62b0b42e867 (diff) | |
| download | open-iov-93431bdd896d9e6d9a25bebef0fb1af308dd64cd.tar.gz open-iov-93431bdd896d9e6d9a25bebef0fb1af308dd64cd.zip | |
kernel: enlarge the pin table and its probe window; tools: bridge client mappings into the service
The pin table became the primary VA->PA store once push-buffer regions and
bridged client mappings started adding thousands of pages per boot; at
8192 slots with an 8-probe window, inserts were silently dropped and
lookups missed at half load. 32768 slots and a 32-probe window keep the
mapped set intact.
tools/guest-diag/{guest_pmap_all.py,bridge_sems.sh} translate the CUDA
client's /dev/nvidia* mappings through the guest's own pagemap and inject
each present page into the service as a sysmem pin. With the client's
semaphore pages pinned the service applies their releases into the page
the client polls (verified: applied 67 -> 109, misses stop growing for
those VAs), where before every write went to the vidmem shadow.
Diffstat (limited to 'kernel')
| -rw-r--r-- | kernel/gsp_sema_service.c | 13 |
1 files changed, 10 insertions, 3 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index 34b9ffd..168e017 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -157,7 +157,14 @@ static u32 svc_map_hash[SEMA_MAP_MAX * 2]; * (debug injection) and that must not be evicted by the learned cache, which * is content-addressed and gets thrashed by the lookup stream. */ -#define SVC_PIN_SLOTS 8192 +#define SVC_PIN_SLOTS 32768 +/* + * Linear-probe window for the pin table. The table is now the primary + * VA->PA store (push-buffer regions and bridged client mappings add + * thousands of pages at a time), so a short window would silently drop + * inserts and mis-resolve lookups once it is half full. + */ +#define SVC_PIN_PROBE 32 struct svc_pin { u64 va_page; @@ -965,7 +972,7 @@ static int svc_pin_lookup(u64 va, u64 *pa, int *ap) u32 h = (u32)((page >> SEMA_PAGE_SHIFT) * 2654435761u) & mask; u32 probe; - for (probe = 0; probe < 8; probe++) { + for (probe = 0; probe < SVC_PIN_PROBE; probe++) { struct svc_pin *p = &svc.pin[(h + probe) & mask]; if (!p->va_page) @@ -986,7 +993,7 @@ static void svc_pin_store(u64 va, u64 pa, int ap) u32 h = (u32)((page >> SEMA_PAGE_SHIFT) * 2654435761u) & mask; u32 probe; - for (probe = 0; probe < 8; probe++) { + for (probe = 0; probe < SVC_PIN_PROBE; probe++) { struct svc_pin *p = &svc.pin[(h + probe) & mask]; if (!p->va_page || p->va_page == page) { |