github-mirror/linux
1
0
Fork 0
mirror of https://github.com/torvalds/linux.git synced 2026-05-26 16:12:59 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity Graph

nouveau/vmm: rewrite pte tracker using a struct and bitfields.

Browse Source 
I want to increase the counters here and start tracking LPTs as well
as there are certain situations where userspace with mixed page sizes
can cause ref/unrefs to live longer so need better reference counting.

This should be entirely non-functional.

Reviewed-by: Mary Guillemard <mary@mary.zone>
Tested-by: Mary Guillemard <mary@mary.zone>
Tested-by: Mel Henning <mhenning@darkrefraction.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Link: https://patch.msgid.link/20260204030208.2313241-2-airlied@gmail.com
This commit is contained in:
Dave Airlie 2026-02-04 13:00:05 +10:00
parent 750817a7c4
commit c4d53e567d
2 changed files with 31 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
View File

@ -53,7 +53,7 @@ nvkm_vmm_pt_new(const struct nvkm_vmm_desc *desc, bool sparse,
}
}
if (!(pgt = kzalloc(sizeof(*pgt) + lpte, GFP_KERNEL)))
if (!(pgt = kzalloc(sizeof(*pgt) + (sizeof(pgt->pte[0]) * lpte), GFP_KERNEL)))
return NULL;
pgt->page = page ? page->shift : 0;
pgt->sparse = sparse;
@ -208,7 +208,7 @@ nvkm_vmm_unref_sptes(struct nvkm_vmm_iter *it, struct nvkm_vmm_pt *pgt,
*/
for (lpti = ptei >> sptb; ptes; spti = 0, lpti++) {
const u32 pten = min(sptn - spti, ptes);
pgt->pte[lpti] -= pten;
pgt->pte[lpti].s.sptes -= pten;
ptes -= pten;
}
@ -218,9 +218,9 @@ nvkm_vmm_unref_sptes(struct nvkm_vmm_iter *it, struct nvkm_vmm_pt *pgt,
for (ptei = pteb = ptei >> sptb; ptei < lpti; pteb = ptei) {
/* Skip over any LPTEs that still have valid SPTEs. */
if (pgt->pte[pteb] & NVKM_VMM_PTE_SPTES) {
if (pgt->pte[pteb].s.sptes) {
for (ptes = 1, ptei++; ptei < lpti; ptes++, ptei++) {
if (!(pgt->pte[ptei] & NVKM_VMM_PTE_SPTES))
if (!(pgt->pte[ptei].s.sptes))
break;
}
continue;
@ -232,14 +232,14 @@ nvkm_vmm_unref_sptes(struct nvkm_vmm_iter *it, struct nvkm_vmm_pt *pgt,
*
* Determine how many LPTEs need to transition state.
*/
pgt->pte[ptei] &= ~NVKM_VMM_PTE_VALID;
pgt->pte[ptei].s.spte_valid = false;
for (ptes = 1, ptei++; ptei < lpti; ptes++, ptei++) {
if (pgt->pte[ptei] & NVKM_VMM_PTE_SPTES)
if (pgt->pte[ptei].s.sptes)
break;
pgt->pte[ptei] &= ~NVKM_VMM_PTE_VALID;
pgt->pte[ptei].s.spte_valid = false;
}
if (pgt->pte[pteb] & NVKM_VMM_PTE_SPARSE) {
if (pgt->pte[pteb].s.sparse) {
TRA(it, "LPTE %05x: U -> S %d PTEs", pteb, ptes);
pair->func->sparse(vmm, pgt->pt[0], pteb, ptes);
} else
@ -307,7 +307,7 @@ nvkm_vmm_ref_sptes(struct nvkm_vmm_iter *it, struct nvkm_vmm_pt *pgt,
*/
for (lpti = ptei >> sptb; ptes; spti = 0, lpti++) {
const u32 pten = min(sptn - spti, ptes);
pgt->pte[lpti] += pten;
pgt->pte[lpti].s.sptes += pten;
ptes -= pten;
}
@ -317,9 +317,9 @@ nvkm_vmm_ref_sptes(struct nvkm_vmm_iter *it, struct nvkm_vmm_pt *pgt,
for (ptei = pteb = ptei >> sptb; ptei < lpti; pteb = ptei) {
/* Skip over any LPTEs that already have valid SPTEs. */
if (pgt->pte[pteb] & NVKM_VMM_PTE_VALID) {
if (pgt->pte[pteb].s.spte_valid) {
for (ptes = 1, ptei++; ptei < lpti; ptes++, ptei++) {
if (!(pgt->pte[ptei] & NVKM_VMM_PTE_VALID))
if (!pgt->pte[ptei].s.spte_valid)
break;
}
continue;
@ -331,14 +331,14 @@ nvkm_vmm_ref_sptes(struct nvkm_vmm_iter *it, struct nvkm_vmm_pt *pgt,
*
* Determine how many LPTEs need to transition state.
*/
pgt->pte[ptei] |= NVKM_VMM_PTE_VALID;
pgt->pte[ptei].s.spte_valid = true;
for (ptes = 1, ptei++; ptei < lpti; ptes++, ptei++) {
if (pgt->pte[ptei] & NVKM_VMM_PTE_VALID)
if (pgt->pte[ptei].s.spte_valid)
break;
pgt->pte[ptei] |= NVKM_VMM_PTE_VALID;
pgt->pte[ptei].s.spte_valid = true;
}
if (pgt->pte[pteb] & NVKM_VMM_PTE_SPARSE) {
if (pgt->pte[pteb].s.sparse) {
const u32 spti = pteb * sptn;
const u32 sptc = ptes * sptn;
/* The entire LPTE is marked as sparse, we need
@ -386,7 +386,8 @@ nvkm_vmm_sparse_ptes(const struct nvkm_vmm_desc *desc,
pgt->pde[ptei++] = NVKM_VMM_PDE_SPARSE;
} else
if (desc->type == LPT) {
memset(&pgt->pte[ptei], NVKM_VMM_PTE_SPARSE, ptes);
union nvkm_pte_tracker sparse = { .s.sparse = 1 };
memset(&pgt->pte[ptei].u, sparse.u, ptes);
}
}
@ -398,7 +399,7 @@ nvkm_vmm_sparse_unref_ptes(struct nvkm_vmm_iter *it, bool pfn, u32 ptei, u32 pte
memset(&pt->pde[ptei], 0x00, sizeof(pt->pde[0]) * ptes);
else
if (it->desc->type == LPT)
memset(&pt->pte[ptei], 0x00, sizeof(pt->pte[0]) * ptes);
memset(&pt->pte[ptei].u, 0x00, sizeof(pt->pte[0]) * ptes);
return nvkm_vmm_unref_ptes(it, pfn, ptei, ptes);
}
@ -445,9 +446,9 @@ nvkm_vmm_ref_hwpt(struct nvkm_vmm_iter *it, struct nvkm_vmm_pt *pgd, u32 pdei)
* the SPTEs on some GPUs.
*/
for (ptei = pteb = 0; ptei < pten; pteb = ptei) {
bool spte = pgt->pte[ptei] & NVKM_VMM_PTE_SPTES;
bool spte = !!pgt->pte[ptei].s.sptes;
for (ptes = 1, ptei++; ptei < pten; ptes++, ptei++) {
bool next = pgt->pte[ptei] & NVKM_VMM_PTE_SPTES;
bool next = !!pgt->pte[ptei].s.sptes;
if (spte != next)
break;
}
@ -461,7 +462,7 @@ nvkm_vmm_ref_hwpt(struct nvkm_vmm_iter *it, struct nvkm_vmm_pt *pgd, u32 pdei)
} else {
desc->func->unmap(vmm, pt, pteb, ptes);
while (ptes--)
pgt->pte[pteb++] |= NVKM_VMM_PTE_VALID;
pgt->pte[pteb++].s.spte_valid = true;
}
}
} else {

14
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h
View File

@ -4,6 +4,15 @@
#include <core/memory.h>
enum nvkm_memory_target;
union nvkm_pte_tracker {
u8 u;
struct {
u8 sparse:1;
u8 spte_valid:1;
u8 sptes:6;
} s;
};
struct nvkm_vmm_pt {
/* Some GPUs have a mapping level with a dual page tables to
* support large and small pages in the same address-range.
@ -44,10 +53,7 @@ struct nvkm_vmm_pt {
*
* This information is used to manage LPTE state transitions.
*/
#define NVKM_VMM_PTE_SPARSE 0x80
#define NVKM_VMM_PTE_VALID 0x40
#define NVKM_VMM_PTE_SPTES 0x3f
u8 pte[];
union nvkm_pte_tracker pte[];
};
typedef void (*nvkm_vmm_pxe_func)(struct nvkm_vmm *,