[PATCH 10/35] drm/amdkfd: register overlap system memory range
Felix Kuehling
Felix.Kuehling at amd.com
Thu Jan 7 03:01:02 UTC 2021
From: Philip Yang <Philip.Yang at amd.com>
No overlap range interval [start, last] exist in svms object interval
tree. If process registers new range which has overlap with old range,
the old range split into 2 ranges depending on the overlap happens at
head or tail part of old range.
Signed-off-by: Philip Yang <Philip.Yang at amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 297 ++++++++++++++++++++++++++-
1 file changed, 294 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 02918faa70d5..ad007261f54c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -293,6 +293,278 @@ static void svm_range_debug_dump(struct svm_range_list *svms)
}
}
+static bool
+svm_range_is_same_attrs(struct svm_range *old, struct svm_range *new)
+{
+ return (old->prefetch_loc == new->prefetch_loc &&
+ old->flags == new->flags &&
+ old->granularity == new->granularity);
+}
+
+static int
+svm_range_split_pages(struct svm_range *new, struct svm_range *old,
+ uint64_t start, uint64_t last)
+{
+ unsigned long old_start;
+ dma_addr_t *pages_addr;
+ uint64_t d;
+
+ old_start = old->it_node.start;
+ new->pages_addr = kvmalloc_array(new->npages,
+ sizeof(*new->pages_addr),
+ GFP_KERNEL | __GFP_ZERO);
+ if (!new->pages_addr)
+ return -ENOMEM;
+
+ d = new->it_node.start - old_start;
+ memcpy(new->pages_addr, old->pages_addr + d,
+ new->npages * sizeof(*new->pages_addr));
+
+ old->npages = last - start + 1;
+ old->it_node.start = start;
+ old->it_node.last = last;
+
+ pages_addr = kvmalloc_array(old->npages, sizeof(*pages_addr),
+ GFP_KERNEL);
+ if (!pages_addr) {
+ kvfree(new->pages_addr);
+ return -ENOMEM;
+ }
+
+ d = start - old_start;
+ memcpy(pages_addr, old->pages_addr + d,
+ old->npages * sizeof(*pages_addr));
+
+ kvfree(old->pages_addr);
+ old->pages_addr = pages_addr;
+
+ return 0;
+}
+
+/**
+ * svm_range_split_adjust - split range and adjust
+ *
+ * @new: new range
+ * @old: the old range
+ * @start: the old range adjust to start address in pages
+ * @last: the old range adjust to last address in pages
+ *
+ * Copy system memory pages, pages_addr or vram mm_nodes in old range to new
+ * range from new_start up to size new->npages, the remaining old range is from
+ * start to last
+ *
+ * Return:
+ * 0 - OK, -ENOMEM - out of memory
+ */
+static int
+svm_range_split_adjust(struct svm_range *new, struct svm_range *old,
+ uint64_t start, uint64_t last)
+{
+ int r = -EINVAL;
+
+ pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n",
+ new->svms, new->it_node.start, old->it_node.start,
+ old->it_node.last, start, last);
+
+ if (new->it_node.start < old->it_node.start ||
+ new->it_node.last > old->it_node.last) {
+ WARN_ONCE(1, "invalid new range start or last\n");
+ return -EINVAL;
+ }
+
+ if (old->pages_addr)
+ r = svm_range_split_pages(new, old, start, last);
+ else
+ WARN_ONCE(1, "split adjust invalid pages_addr and nodes\n");
+ if (r)
+ return r;
+
+ new->flags = old->flags;
+ new->preferred_loc = old->preferred_loc;
+ new->prefetch_loc = old->prefetch_loc;
+ new->actual_loc = old->actual_loc;
+ new->granularity = old->granularity;
+ bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
+ bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
+
+ return 0;
+}
+
+/**
+ * svm_range_split - split a range in 2 ranges
+ *
+ * @prange: the svm range to split
+ * @start: the remaining range start address in pages
+ * @last: the remaining range last address in pages
+ * @new: the result new range generated
+ *
+ * Two cases only:
+ * case 1: if start == prange->it_node.start
+ * prange ==> prange[start, last]
+ * new range [last + 1, prange->it_node.last]
+ *
+ * case 2: if last == prange->it_node.last
+ * prange ==> prange[start, last]
+ * new range [prange->it_node.start, start - 1]
+ *
+ * Context: Caller hold svms->rw_sem as write mode
+ *
+ * Return:
+ * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last
+ */
+static int
+svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last,
+ struct svm_range **new)
+{
+ uint64_t old_start = prange->it_node.start;
+ uint64_t old_last = prange->it_node.last;
+ struct svm_range_list *svms;
+ int r = 0;
+
+ pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms,
+ old_start, old_last, start, last);
+
+ if (old_start != start && old_last != last)
+ return -EINVAL;
+ if (start < old_start || last > old_last)
+ return -EINVAL;
+
+ svms = prange->svms;
+ if (old_start == start) {
+ *new = svm_range_new(svms, last + 1, old_last);
+ if (!*new)
+ return -ENOMEM;
+ r = svm_range_split_adjust(*new, prange, start, last);
+ } else {
+ *new = svm_range_new(svms, old_start, start - 1);
+ if (!*new)
+ return -ENOMEM;
+ r = svm_range_split_adjust(*new, prange, start, last);
+ }
+
+ return r;
+}
+
+static int
+svm_range_split_two(struct svm_range *prange, struct svm_range *new,
+ uint64_t start, uint64_t last,
+ struct list_head *insert_list,
+ struct list_head *update_list)
+{
+ struct svm_range *tail, *tail2;
+ int r;
+
+ r = svm_range_split(prange, prange->it_node.start, start - 1, &tail);
+ if (r)
+ return r;
+ r = svm_range_split(tail, start, last, &tail2);
+ if (r)
+ return r;
+ list_add(&tail2->list, insert_list);
+ list_add(&tail->list, insert_list);
+
+ if (!svm_range_is_same_attrs(prange, new))
+ list_add(&tail->update_list, update_list);
+
+ return 0;
+}
+
+static int
+svm_range_split_tail(struct svm_range *prange, struct svm_range *new,
+ uint64_t start, struct list_head *insert_list,
+ struct list_head *update_list)
+{
+ struct svm_range *tail;
+ int r;
+
+ r = svm_range_split(prange, prange->it_node.start, start - 1, &tail);
+ if (r)
+ return r;
+ list_add(&tail->list, insert_list);
+ if (!svm_range_is_same_attrs(prange, new))
+ list_add(&tail->update_list, update_list);
+
+ return 0;
+}
+
+static int
+svm_range_split_head(struct svm_range *prange, struct svm_range *new,
+ uint64_t last, struct list_head *insert_list,
+ struct list_head *update_list)
+{
+ struct svm_range *head;
+ int r;
+
+ r = svm_range_split(prange, last + 1, prange->it_node.last, &head);
+ if (r)
+ return r;
+ list_add(&head->list, insert_list);
+ if (!svm_range_is_same_attrs(prange, new))
+ list_add(&head->update_list, update_list);
+
+ return 0;
+}
+
+static int
+svm_range_split_add_front(struct svm_range *prange, struct svm_range *new,
+ uint64_t start, uint64_t last,
+ struct list_head *insert_list,
+ struct list_head *update_list)
+{
+ struct svm_range *front, *tail;
+ int r = 0;
+
+ front = svm_range_new(prange->svms, start, prange->it_node.start - 1);
+ if (!front)
+ return -ENOMEM;
+
+ list_add(&front->list, insert_list);
+ list_add(&front->update_list, update_list);
+
+ if (prange->it_node.last > last) {
+ pr_debug("split old in 2\n");
+ r = svm_range_split(prange, prange->it_node.start, last, &tail);
+ if (r)
+ return r;
+ list_add(&tail->list, insert_list);
+ }
+ if (!svm_range_is_same_attrs(prange, new))
+ list_add(&prange->update_list, update_list);
+
+ return 0;
+}
+
+struct svm_range *svm_range_clone(struct svm_range *old)
+{
+ struct svm_range *new;
+
+ new = svm_range_new(old->svms, old->it_node.start, old->it_node.last);
+ if (!new)
+ return NULL;
+
+ if (old->pages_addr) {
+ new->pages_addr = kvmalloc_array(new->npages,
+ sizeof(*new->pages_addr),
+ GFP_KERNEL);
+ if (!new->pages_addr) {
+ kfree(new);
+ return NULL;
+ }
+ memcpy(new->pages_addr, old->pages_addr,
+ old->npages * sizeof(*old->pages_addr));
+ }
+
+ new->flags = old->flags;
+ new->preferred_loc = old->preferred_loc;
+ new->prefetch_loc = old->prefetch_loc;
+ new->actual_loc = old->actual_loc;
+ new->granularity = old->granularity;
+ bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE);
+ bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE);
+
+ return new;
+}
+
/**
* svm_range_handle_overlap - split overlap ranges
* @svms: svm range list header
@@ -334,15 +606,27 @@ svm_range_handle_overlap(struct svm_range_list *svms, struct svm_range *new,
node = interval_tree_iter_first(&svms->objects, start, last);
while (node) {
struct interval_tree_node *next;
+ struct svm_range *old;
pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start,
node->last);
- prange = container_of(node, struct svm_range, it_node);
+ old = container_of(node, struct svm_range, it_node);
next = interval_tree_iter_next(node, start, last);
+ prange = svm_range_clone(old);
+ if (!prange) {
+ r = -ENOMEM;
+ goto out;
+ }
+
+ list_add(&old->remove_list, remove_list);
+ list_add(&prange->list, insert_list);
+
if (node->start < start && node->last > last) {
pr_debug("split in 2 ranges\n");
+ r = svm_range_split_two(prange, new, start, last,
+ insert_list, update_list);
start = last + 1;
} else if (node->start < start) {
@@ -352,11 +636,15 @@ svm_range_handle_overlap(struct svm_range_list *svms, struct svm_range *new,
*/
uint64_t old_last = node->last;
+ pr_debug("change old range last\n");
+ r = svm_range_split_tail(prange, new, start,
+ insert_list, update_list);
start = old_last + 1;
} else if (node->start == start && node->last > last) {
pr_debug("change old range start\n");
-
+ r = svm_range_split_head(prange, new, last,
+ insert_list, update_list);
start = last + 1;
} else if (node->start == start) {
@@ -364,12 +652,15 @@ svm_range_handle_overlap(struct svm_range_list *svms, struct svm_range *new,
pr_debug("found exactly same range\n");
else
pr_debug("next loop to add remaining range\n");
+ if (!svm_range_is_same_attrs(prange, new))
+ list_add(&prange->update_list, update_list);
start = node->last + 1;
} else { /* node->start > start */
pr_debug("add new range at front\n");
-
+ r = svm_range_split_add_front(prange, new, start, last,
+ insert_list, update_list);
start = node->last + 1;
}
--
2.29.2
More information about the amd-gfx
mailing list