#include <core/gpuobj.h>
#include <subdev/fb.h>
+struct nvkm_mmu_ptp {
+ struct nvkm_mmu_pt *pt;
+ struct list_head head;
+ u8 shift;
+ u16 mask;
+ u16 free;
+};
+
+static void
+nvkm_mmu_ptp_put(struct nvkm_mmu *mmu, bool force, struct nvkm_mmu_pt *pt)
+{
+ const int slot = pt->base >> pt->ptp->shift;
+ struct nvkm_mmu_ptp *ptp = pt->ptp;
+
+ /* If there were no free slots in the parent allocation before,
+ * there will be now, so return PTP to the cache.
+ */
+ if (!ptp->free)
+ list_add(&ptp->head, &mmu->ptp.list);
+ ptp->free |= BIT(slot);
+
+ /* If there's no more sub-allocations, destroy PTP. */
+ if (ptp->free == ptp->mask) {
+ nvkm_mmu_ptc_put(mmu, force, &ptp->pt);
+ list_del(&ptp->head);
+ kfree(ptp);
+ }
+
+ kfree(pt);
+}
+
+struct nvkm_mmu_pt *
+nvkm_mmu_ptp_get(struct nvkm_mmu *mmu, u32 size, bool zero)
+{
+ struct nvkm_mmu_pt *pt;
+ struct nvkm_mmu_ptp *ptp;
+ int slot;
+
+ if (!(pt = kzalloc(sizeof(*pt), GFP_KERNEL)))
+ return NULL;
+
+ ptp = list_first_entry_or_null(&mmu->ptp.list, typeof(*ptp), head);
+ if (!ptp) {
+ /* Need to allocate a new parent to sub-allocate from. */
+ if (!(ptp = kmalloc(sizeof(*ptp), GFP_KERNEL))) {
+ kfree(pt);
+ return NULL;
+ }
+
+ ptp->pt = nvkm_mmu_ptc_get(mmu, 0x1000, 0x1000, false);
+ if (!ptp->pt) {
+ kfree(ptp);
+ kfree(pt);
+ return NULL;
+ }
+
+ ptp->shift = order_base_2(size);
+ slot = nvkm_memory_size(ptp->pt->memory) >> ptp->shift;
+ ptp->mask = (1 << slot) - 1;
+ ptp->free = ptp->mask;
+ list_add(&ptp->head, &mmu->ptp.list);
+ }
+ pt->ptp = ptp;
+ pt->sub = true;
+
+ /* Sub-allocate from parent object, removing PTP from cache
+ * if there's no more free slots left.
+ */
+ slot = __ffs(ptp->free);
+ ptp->free &= ~BIT(slot);
+ if (!ptp->free)
+ list_del(&ptp->head);
+
+ pt->memory = pt->ptp->pt->memory;
+ pt->base = slot << ptp->shift;
+ pt->addr = pt->ptp->pt->addr + pt->base;
+ return pt;
+}
+
+struct nvkm_mmu_ptc {
+ struct list_head head;
+ struct list_head item;
+ u32 size;
+ u32 refs;
+};
+
+static inline struct nvkm_mmu_ptc *
+nvkm_mmu_ptc_find(struct nvkm_mmu *mmu, u32 size)
+{
+ struct nvkm_mmu_ptc *ptc;
+
+ list_for_each_entry(ptc, &mmu->ptc.list, head) {
+ if (ptc->size == size)
+ return ptc;
+ }
+
+ ptc = kmalloc(sizeof(*ptc), GFP_KERNEL);
+ if (ptc) {
+ INIT_LIST_HEAD(&ptc->item);
+ ptc->size = size;
+ ptc->refs = 0;
+ list_add(&ptc->head, &mmu->ptc.list);
+ }
+
+ return ptc;
+}
+
+void
+nvkm_mmu_ptc_put(struct nvkm_mmu *mmu, bool force, struct nvkm_mmu_pt **ppt)
+{
+ struct nvkm_mmu_pt *pt = *ppt;
+ if (pt) {
+ /* Handle sub-allocated page tables. */
+ if (pt->sub) {
+ mutex_lock(&mmu->ptp.mutex);
+ nvkm_mmu_ptp_put(mmu, force, pt);
+ mutex_unlock(&mmu->ptp.mutex);
+ return;
+ }
+
+ /* Either cache or free the object. */
+ mutex_lock(&mmu->ptc.mutex);
+ if (pt->ptc->refs < 8 /* Heuristic. */ && !force) {
+ list_add_tail(&pt->head, &pt->ptc->item);
+ pt->ptc->refs++;
+ } else {
+ nvkm_memory_unref(&pt->memory);
+ kfree(pt);
+ }
+ mutex_unlock(&mmu->ptc.mutex);
+ }
+}
+
+struct nvkm_mmu_pt *
+nvkm_mmu_ptc_get(struct nvkm_mmu *mmu, u32 size, u32 align, bool zero)
+{
+ struct nvkm_mmu_ptc *ptc;
+ struct nvkm_mmu_pt *pt;
+ int ret;
+
+ /* Sub-allocated page table (ie. GP100 LPT). */
+ if (align < 0x1000) {
+ mutex_lock(&mmu->ptp.mutex);
+ pt = nvkm_mmu_ptp_get(mmu, align, zero);
+ mutex_unlock(&mmu->ptp.mutex);
+ return pt;
+ }
+
+ /* Lookup cache for this page table size. */
+ mutex_lock(&mmu->ptc.mutex);
+ ptc = nvkm_mmu_ptc_find(mmu, size);
+ if (!ptc) {
+ mutex_unlock(&mmu->ptc.mutex);
+ return NULL;
+ }
+
+ /* If there's a free PT in the cache, reuse it. */
+ pt = list_first_entry_or_null(&ptc->item, typeof(*pt), head);
+ if (pt) {
+ if (zero)
+ nvkm_fo64(pt->memory, 0, 0, size >> 3);
+ list_del(&pt->head);
+ ptc->refs--;
+ mutex_unlock(&mmu->ptc.mutex);
+ return pt;
+ }
+ mutex_unlock(&mmu->ptc.mutex);
+
+ /* No such luck, we need to allocate. */
+ if (!(pt = kmalloc(sizeof(*pt), GFP_KERNEL)))
+ return NULL;
+ pt->ptc = ptc;
+ pt->sub = false;
+
+ ret = nvkm_memory_new(mmu->subdev.device, NVKM_MEM_TARGET_INST,
+ size, align, zero, &pt->memory);
+ if (ret) {
+ kfree(pt);
+ return NULL;
+ }
+
+ pt->base = 0;
+ pt->addr = nvkm_memory_addr(pt->memory);
+ return pt;
+}
+
+void
+nvkm_mmu_ptc_dump(struct nvkm_mmu *mmu)
+{
+ struct nvkm_mmu_ptc *ptc;
+ list_for_each_entry(ptc, &mmu->ptc.list, head) {
+ struct nvkm_mmu_pt *pt, *tt;
+ list_for_each_entry_safe(pt, tt, &ptc->item, head) {
+ nvkm_memory_unref(&pt->memory);
+ list_del(&pt->head);
+ kfree(pt);
+ }
+ }
+}
+
+static void
+nvkm_mmu_ptc_fini(struct nvkm_mmu *mmu)
+{
+ struct nvkm_mmu_ptc *ptc, *ptct;
+
+ list_for_each_entry_safe(ptc, ptct, &mmu->ptc.list, head) {
+ WARN_ON(!list_empty(&ptc->item));
+ list_del(&ptc->head);
+ kfree(ptc);
+ }
+}
+
+static void
+nvkm_mmu_ptc_init(struct nvkm_mmu *mmu)
+{
+ mutex_init(&mmu->ptc.mutex);
+ INIT_LIST_HEAD(&mmu->ptc.list);
+ mutex_init(&mmu->ptp.mutex);
+ INIT_LIST_HEAD(&mmu->ptp.list);
+}
+
void
nvkm_vm_map_at(struct nvkm_vma *vma, u64 delta, struct nvkm_mem *node)
{
delta += (u64)len << vma->node->type;
}
r = r->next;
- };
+ }
mmu->func->flush(vm);
}
mmu->func->map_pgt(vpgd->obj, pde, vpgt->mem);
}
- nvkm_memory_del(&pgt);
+ mmu->func->flush(vm);
+
+ nvkm_memory_unref(&pgt);
}
}
vm->pgt[0].refcount[0] = 1;
vm->pgt[0].mem[0] = pgt;
nvkm_memory_boot(pgt, vm);
+ vm->bootstrapped = true;
}
return ret;
return -ENOMEM;
}
- ret = nvkm_mm_init(&vm->mm, mm_offset >> 12, mm_length >> 12,
+ ret = nvkm_mm_init(&vm->mm, 0, mm_offset >> 12, mm_length >> 12,
block >> 12);
if (ret) {
vfree(vm->pgt);
}
if (*ptr) {
+ if ((*ptr)->bootstrapped && pgd)
+ nvkm_memory_unref(&(*ptr)->pgt[0].mem[0]);
nvkm_vm_unlink(*ptr, pgd);
kref_put(&(*ptr)->refcount, nvkm_vm_del);
}
nvkm_mmu_dtor(struct nvkm_subdev *subdev)
{
struct nvkm_mmu *mmu = nvkm_mmu(subdev);
+ void *data = mmu;
+
if (mmu->func->dtor)
- return mmu->func->dtor(mmu);
- return mmu;
+ data = mmu->func->dtor(mmu);
+
+ nvkm_mmu_ptc_fini(mmu);
+ return data;
}
static const struct nvkm_subdev_func
mmu->limit = func->limit;
mmu->dma_bits = func->dma_bits;
mmu->lpg_shift = func->lpg_shift;
+ nvkm_mmu_ptc_init(mmu);
}
int