share_pool: Alloc sp memory on a specified memory node

ascend inclusion
category: feature
bugzilla: https://gitee.com/openeuler/kernel/issues/I4EUVI
CVE: NA

-------------------------------------------

Adding a function for getting node id, which can be used to alloc
share pool memory on a specified memory node.
Signed-off-by: NPeng Wu <wupeng58@huawei.com>
Reviewed-by: NDing Tianhong <dingtianhong@huawei.com>
Reviewed-by: Nchenweilong <chenweilong@huawei.com>
Reviewed-by: NTang Yizhou <tangyizhou@huawei.com>
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>
Reviewed-by: NWeilong Chen <chenweilong@huawei.com>
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>

include/linux/share_pool.h

@@ -10,6 +10,10 @@
 #define SP_HUGEPAGE		(1 << 0)
 #define SP_HUGEPAGE_ONLY	(1 << 1)
 #define SP_DVPP			(1 << 2)
+#define DEVICE_ID_MASK		0x3ff
+#define DEVICE_ID_SHIFT		32
+#define SP_FLAG_MASK		(SP_HUGEPAGE | SP_HUGEPAGE_ONLY | SP_DVPP | \
+				(_AC(DEVICE_ID_MASK, UL) << DEVICE_ID_SHIFT))
 
 #define SPG_ID_NONE	-1	/* not associated with sp_group, only for specified thread */
 #define SPG_ID_DEFAULT	0	/* use the spg id of current thread */
@@ -22,7 +26,7 @@
 #define SPG_ID_DVPP_PASS_THROUGH_MAX	899999
 #define SPG_ID_DVPP_PASS_THROUGH	900000
 
-#define MAX_DEVID 1	/* the max num of Da-vinci devices */
+#define MAX_DEVID 2	/* the max num of Da-vinci devices */
 
 /* to align the pointer to the (next) PMD boundary */
 #define PMD_ALIGN(addr)		ALIGN(addr, PMD_SIZE)
@@ -54,9 +58,9 @@ extern bool vmap_allow_huge;
  * |-------------------- 8T -------------------|---|------ 8T ------------|
  * |		Device 0	   |  Device 1 |...|                      |
  * |----------------------------------------------------------------------|
- * |- 16G -|- 16G -|- 16G -|- 16G -|     |     |   |                      |
+ * |------------- 16G -------------|    16G    |   |                      |
  * | DVPP GROUP0   | DVPP GROUP1   | ... | ... |...|  sp normal memory    |
- * | svm   |  sp   |  svm  |  sp   |     |     |   |                      |
+ * |     sp        |    sp         |     |     |   |                      |
  * |----------------------------------------------------------------------|
  *
  * The host SVM feature reserves 8T virtual memory by mmap, and due to the
@@ -181,6 +185,7 @@ extern void sp_proc_stat_drop(struct sp_proc_stat *stat);
 extern void spa_overview_show(struct seq_file *seq);
 extern void spg_overview_show(struct seq_file *seq);
 extern void proc_sharepool_init(void);
+extern int sp_node_id(struct vm_area_struct *vma);
 
 static inline struct task_struct *sp_get_task(struct mm_struct *mm)
 {
@@ -485,6 +490,10 @@ static inline void sp_free_pages(struct page *page, struct vm_struct *area)
 {
 }
 
+static inline int sp_node_id(struct vm_area_struct *vma)
+{
+	return numa_node_id();
+}
 #endif
 
 #endif /* LINUX_SHARE_POOL_H */

mm/share_pool.c

@@ -241,6 +241,7 @@ struct sp_area {
 	struct mm_struct *mm;		/* owner of k2u(task) */
 	unsigned long kva;		/* shared kva */
 	pid_t applier;			/* the original applier process */
+	int node_id;			/* memory node */
 };
 static DEFINE_SPINLOCK(sp_area_lock);
 static struct rb_root sp_area_root = RB_ROOT;
@@ -863,11 +864,13 @@ static struct sp_area *sp_alloc_area(unsigned long size, unsigned long flags,
 	unsigned long vend = MMAP_SHARE_POOL_16G_START;
 	unsigned long addr;
 	unsigned long size_align = PMD_ALIGN(size); /* va aligned to 2M */
+	int node_id = (flags >> DEVICE_ID_SHIFT) & DEVICE_ID_MASK;
 
 	if ((flags & SP_DVPP)) {
 		if (sp_area_customized == false) {
-			vstart = MMAP_SHARE_POOL_16G_START;
-			vend = MMAP_SHARE_POOL_16G_START + MMAP_SHARE_POOL_16G_SIZE;
+			vstart = MMAP_SHARE_POOL_16G_START +
+				node_id * MMAP_SHARE_POOL_16G_SIZE;
+			vend = vstart + MMAP_SHARE_POOL_16G_SIZE;
 		} else {
 			if (!spg) {
 				pr_err_ratelimited("share pool: don't allow k2u(task) in host svm multiprocess scene\n");
@@ -878,7 +881,7 @@ static struct sp_area *sp_alloc_area(unsigned long size, unsigned long flags,
 		}
 	}
 
-	spa = kmalloc(sizeof(struct sp_area), GFP_KERNEL);
+	spa = __kmalloc_node(sizeof(struct sp_area), GFP_KERNEL, node_id);
 	if (unlikely(!spa)) {
 		pr_err_ratelimited("share pool: alloc spa failed due to lack of memory\n");
 		return ERR_PTR(-ENOMEM);
@@ -973,6 +976,7 @@ static struct sp_area *sp_alloc_area(unsigned long size, unsigned long flags,
 	spa->mm = NULL;
 	spa->kva = 0;   /* NULL pointer */
 	spa->applier = applier;
+	spa->node_id = node_id;
 
 	if (spa_inc_usage(type, size, (flags & SP_DVPP))) {
 		err = ERR_PTR(-EINVAL);
@@ -1379,7 +1383,7 @@ void *sp_alloc(unsigned long size, unsigned long sp_flags, int spg_id)
 		return ERR_PTR(-EINVAL);
 	}
 
-	if (sp_flags & ~(SP_HUGEPAGE_ONLY | SP_HUGEPAGE | SP_DVPP)) {
+	if (sp_flags & (~SP_FLAG_MASK)) {
 		pr_err_ratelimited("share pool: allocation failed, invalid flag %lx\n", sp_flags);
 		return ERR_PTR(-EINVAL);
 	}
@@ -2606,7 +2610,8 @@ EXPORT_SYMBOL_GPL(sp_config_dvpp_range);
 static bool is_sp_normal_addr(unsigned long addr)
 {
 	return addr >= MMAP_SHARE_POOL_START &&
-		addr < MMAP_SHARE_POOL_16G_START + MMAP_SHARE_POOL_16G_SIZE;
+		addr < MMAP_SHARE_POOL_16G_START +
+			MAX_DEVID * MMAP_SHARE_POOL_16G_SIZE;
 }
 
 /**
@@ -2634,6 +2639,26 @@ bool is_sharepool_addr(unsigned long addr)
 }
 EXPORT_SYMBOL_GPL(is_sharepool_addr);
 
+int sp_node_id(struct vm_area_struct *vma)
+{
+	struct sp_area *spa;
+	int node_id = numa_node_id();
+
+	if (!enable_ascend_share_pool)
+		return node_id;
+
+	if (vma) {
+		spa = __find_sp_area(vma->vm_start);
+		if (spa) {
+			node_id = spa->node_id;
+			__sp_area_drop(spa);
+		}
+	}
+
+	return node_id;
+}
+EXPORT_SYMBOL_GPL(sp_node_id);
+
 static int __init mdc_default_group(char *s)
 {
 	enable_mdc_default_group = 1;
@@ -2999,6 +3024,16 @@ vm_fault_t sharepool_no_page(struct mm_struct *mm,
 	unsigned long haddr = address & huge_page_mask(h);
 	bool new_page = false;
 	int err;
+	int node_id;
+	struct sp_area *spa;
+
+	spa = __find_sp_area(vma->vm_start);
+	if (!spa) {
+		pr_err("share pool: vma is invalid, not from sp mmap\n");
+		return ret;
+	}
+	node_id = spa->node_id;
+	__sp_area_drop(spa);
 
 retry:
 	page = find_lock_page(mapping, idx);
@@ -3010,7 +3045,7 @@ vm_fault_t sharepool_no_page(struct mm_struct *mm,
 		page = alloc_huge_page(vma, haddr, 0);
 		if (IS_ERR(page)) {
 			page = alloc_huge_page_node(hstate_file(vma->vm_file),
-						    numa_mem_id());
+						    node_id);
 			if (!page)
 				page = ERR_PTR(-ENOMEM);
 		}