xref: /OK3568_Linux_fs/kernel/mm/gup_benchmark.c (revision 4882a59341e53eb6f0b4789bf948001014eff981)
1*4882a593Smuzhiyun #include <linux/kernel.h>
2*4882a593Smuzhiyun #include <linux/mm.h>
3*4882a593Smuzhiyun #include <linux/slab.h>
4*4882a593Smuzhiyun #include <linux/uaccess.h>
5*4882a593Smuzhiyun #include <linux/ktime.h>
6*4882a593Smuzhiyun #include <linux/debugfs.h>
7*4882a593Smuzhiyun 
8*4882a593Smuzhiyun #define GUP_FAST_BENCHMARK	_IOWR('g', 1, struct gup_benchmark)
9*4882a593Smuzhiyun #define GUP_BENCHMARK		_IOWR('g', 2, struct gup_benchmark)
10*4882a593Smuzhiyun #define PIN_FAST_BENCHMARK	_IOWR('g', 3, struct gup_benchmark)
11*4882a593Smuzhiyun #define PIN_BENCHMARK		_IOWR('g', 4, struct gup_benchmark)
12*4882a593Smuzhiyun #define PIN_LONGTERM_BENCHMARK	_IOWR('g', 5, struct gup_benchmark)
13*4882a593Smuzhiyun 
14*4882a593Smuzhiyun struct gup_benchmark {
15*4882a593Smuzhiyun 	__u64 get_delta_usec;
16*4882a593Smuzhiyun 	__u64 put_delta_usec;
17*4882a593Smuzhiyun 	__u64 addr;
18*4882a593Smuzhiyun 	__u64 size;
19*4882a593Smuzhiyun 	__u32 nr_pages_per_call;
20*4882a593Smuzhiyun 	__u32 flags;
21*4882a593Smuzhiyun 	__u64 expansion[10];	/* For future use */
22*4882a593Smuzhiyun };
23*4882a593Smuzhiyun 
put_back_pages(unsigned int cmd,struct page ** pages,unsigned long nr_pages)24*4882a593Smuzhiyun static void put_back_pages(unsigned int cmd, struct page **pages,
25*4882a593Smuzhiyun 			   unsigned long nr_pages)
26*4882a593Smuzhiyun {
27*4882a593Smuzhiyun 	unsigned long i;
28*4882a593Smuzhiyun 
29*4882a593Smuzhiyun 	switch (cmd) {
30*4882a593Smuzhiyun 	case GUP_FAST_BENCHMARK:
31*4882a593Smuzhiyun 	case GUP_BENCHMARK:
32*4882a593Smuzhiyun 		for (i = 0; i < nr_pages; i++)
33*4882a593Smuzhiyun 			put_page(pages[i]);
34*4882a593Smuzhiyun 		break;
35*4882a593Smuzhiyun 
36*4882a593Smuzhiyun 	case PIN_FAST_BENCHMARK:
37*4882a593Smuzhiyun 	case PIN_BENCHMARK:
38*4882a593Smuzhiyun 	case PIN_LONGTERM_BENCHMARK:
39*4882a593Smuzhiyun 		unpin_user_pages(pages, nr_pages);
40*4882a593Smuzhiyun 		break;
41*4882a593Smuzhiyun 	}
42*4882a593Smuzhiyun }
43*4882a593Smuzhiyun 
verify_dma_pinned(unsigned int cmd,struct page ** pages,unsigned long nr_pages)44*4882a593Smuzhiyun static void verify_dma_pinned(unsigned int cmd, struct page **pages,
45*4882a593Smuzhiyun 			      unsigned long nr_pages)
46*4882a593Smuzhiyun {
47*4882a593Smuzhiyun 	unsigned long i;
48*4882a593Smuzhiyun 	struct page *page;
49*4882a593Smuzhiyun 
50*4882a593Smuzhiyun 	switch (cmd) {
51*4882a593Smuzhiyun 	case PIN_FAST_BENCHMARK:
52*4882a593Smuzhiyun 	case PIN_BENCHMARK:
53*4882a593Smuzhiyun 	case PIN_LONGTERM_BENCHMARK:
54*4882a593Smuzhiyun 		for (i = 0; i < nr_pages; i++) {
55*4882a593Smuzhiyun 			page = pages[i];
56*4882a593Smuzhiyun 			if (WARN(!page_maybe_dma_pinned(page),
57*4882a593Smuzhiyun 				 "pages[%lu] is NOT dma-pinned\n", i)) {
58*4882a593Smuzhiyun 
59*4882a593Smuzhiyun 				dump_page(page, "gup_benchmark failure");
60*4882a593Smuzhiyun 				break;
61*4882a593Smuzhiyun 			}
62*4882a593Smuzhiyun 		}
63*4882a593Smuzhiyun 		break;
64*4882a593Smuzhiyun 	}
65*4882a593Smuzhiyun }
66*4882a593Smuzhiyun 
__gup_benchmark_ioctl(unsigned int cmd,struct gup_benchmark * gup)67*4882a593Smuzhiyun static int __gup_benchmark_ioctl(unsigned int cmd,
68*4882a593Smuzhiyun 		struct gup_benchmark *gup)
69*4882a593Smuzhiyun {
70*4882a593Smuzhiyun 	ktime_t start_time, end_time;
71*4882a593Smuzhiyun 	unsigned long i, nr_pages, addr, next;
72*4882a593Smuzhiyun 	int nr;
73*4882a593Smuzhiyun 	struct page **pages;
74*4882a593Smuzhiyun 	int ret = 0;
75*4882a593Smuzhiyun 	bool needs_mmap_lock =
76*4882a593Smuzhiyun 		cmd != GUP_FAST_BENCHMARK && cmd != PIN_FAST_BENCHMARK;
77*4882a593Smuzhiyun 
78*4882a593Smuzhiyun 	if (gup->size > ULONG_MAX)
79*4882a593Smuzhiyun 		return -EINVAL;
80*4882a593Smuzhiyun 
81*4882a593Smuzhiyun 	nr_pages = gup->size / PAGE_SIZE;
82*4882a593Smuzhiyun 	pages = kvcalloc(nr_pages, sizeof(void *), GFP_KERNEL);
83*4882a593Smuzhiyun 	if (!pages)
84*4882a593Smuzhiyun 		return -ENOMEM;
85*4882a593Smuzhiyun 
86*4882a593Smuzhiyun 	if (needs_mmap_lock && mmap_read_lock_killable(current->mm)) {
87*4882a593Smuzhiyun 		ret = -EINTR;
88*4882a593Smuzhiyun 		goto free_pages;
89*4882a593Smuzhiyun 	}
90*4882a593Smuzhiyun 
91*4882a593Smuzhiyun 	i = 0;
92*4882a593Smuzhiyun 	nr = gup->nr_pages_per_call;
93*4882a593Smuzhiyun 	start_time = ktime_get();
94*4882a593Smuzhiyun 	for (addr = gup->addr; addr < gup->addr + gup->size; addr = next) {
95*4882a593Smuzhiyun 		if (nr != gup->nr_pages_per_call)
96*4882a593Smuzhiyun 			break;
97*4882a593Smuzhiyun 
98*4882a593Smuzhiyun 		next = addr + nr * PAGE_SIZE;
99*4882a593Smuzhiyun 		if (next > gup->addr + gup->size) {
100*4882a593Smuzhiyun 			next = gup->addr + gup->size;
101*4882a593Smuzhiyun 			nr = (next - addr) / PAGE_SIZE;
102*4882a593Smuzhiyun 		}
103*4882a593Smuzhiyun 
104*4882a593Smuzhiyun 		/* Filter out most gup flags: only allow a tiny subset here: */
105*4882a593Smuzhiyun 		gup->flags &= FOLL_WRITE;
106*4882a593Smuzhiyun 
107*4882a593Smuzhiyun 		switch (cmd) {
108*4882a593Smuzhiyun 		case GUP_FAST_BENCHMARK:
109*4882a593Smuzhiyun 			nr = get_user_pages_fast(addr, nr, gup->flags,
110*4882a593Smuzhiyun 						 pages + i);
111*4882a593Smuzhiyun 			break;
112*4882a593Smuzhiyun 		case GUP_BENCHMARK:
113*4882a593Smuzhiyun 			nr = get_user_pages(addr, nr, gup->flags, pages + i,
114*4882a593Smuzhiyun 					    NULL);
115*4882a593Smuzhiyun 			break;
116*4882a593Smuzhiyun 		case PIN_FAST_BENCHMARK:
117*4882a593Smuzhiyun 			nr = pin_user_pages_fast(addr, nr, gup->flags,
118*4882a593Smuzhiyun 						 pages + i);
119*4882a593Smuzhiyun 			break;
120*4882a593Smuzhiyun 		case PIN_BENCHMARK:
121*4882a593Smuzhiyun 			nr = pin_user_pages(addr, nr, gup->flags, pages + i,
122*4882a593Smuzhiyun 					    NULL);
123*4882a593Smuzhiyun 			break;
124*4882a593Smuzhiyun 		case PIN_LONGTERM_BENCHMARK:
125*4882a593Smuzhiyun 			nr = pin_user_pages(addr, nr,
126*4882a593Smuzhiyun 					    gup->flags | FOLL_LONGTERM,
127*4882a593Smuzhiyun 					    pages + i, NULL);
128*4882a593Smuzhiyun 			break;
129*4882a593Smuzhiyun 		default:
130*4882a593Smuzhiyun 			ret = -EINVAL;
131*4882a593Smuzhiyun 			goto unlock;
132*4882a593Smuzhiyun 		}
133*4882a593Smuzhiyun 
134*4882a593Smuzhiyun 		if (nr <= 0)
135*4882a593Smuzhiyun 			break;
136*4882a593Smuzhiyun 		i += nr;
137*4882a593Smuzhiyun 	}
138*4882a593Smuzhiyun 	end_time = ktime_get();
139*4882a593Smuzhiyun 
140*4882a593Smuzhiyun 	/* Shifting the meaning of nr_pages: now it is actual number pinned: */
141*4882a593Smuzhiyun 	nr_pages = i;
142*4882a593Smuzhiyun 
143*4882a593Smuzhiyun 	gup->get_delta_usec = ktime_us_delta(end_time, start_time);
144*4882a593Smuzhiyun 	gup->size = addr - gup->addr;
145*4882a593Smuzhiyun 
146*4882a593Smuzhiyun 	/*
147*4882a593Smuzhiyun 	 * Take an un-benchmark-timed moment to verify DMA pinned
148*4882a593Smuzhiyun 	 * state: print a warning if any non-dma-pinned pages are found:
149*4882a593Smuzhiyun 	 */
150*4882a593Smuzhiyun 	verify_dma_pinned(cmd, pages, nr_pages);
151*4882a593Smuzhiyun 
152*4882a593Smuzhiyun 	start_time = ktime_get();
153*4882a593Smuzhiyun 
154*4882a593Smuzhiyun 	put_back_pages(cmd, pages, nr_pages);
155*4882a593Smuzhiyun 
156*4882a593Smuzhiyun 	end_time = ktime_get();
157*4882a593Smuzhiyun 	gup->put_delta_usec = ktime_us_delta(end_time, start_time);
158*4882a593Smuzhiyun 
159*4882a593Smuzhiyun unlock:
160*4882a593Smuzhiyun 	if (needs_mmap_lock)
161*4882a593Smuzhiyun 		mmap_read_unlock(current->mm);
162*4882a593Smuzhiyun free_pages:
163*4882a593Smuzhiyun 	kvfree(pages);
164*4882a593Smuzhiyun 	return ret;
165*4882a593Smuzhiyun }
166*4882a593Smuzhiyun 
gup_benchmark_ioctl(struct file * filep,unsigned int cmd,unsigned long arg)167*4882a593Smuzhiyun static long gup_benchmark_ioctl(struct file *filep, unsigned int cmd,
168*4882a593Smuzhiyun 		unsigned long arg)
169*4882a593Smuzhiyun {
170*4882a593Smuzhiyun 	struct gup_benchmark gup;
171*4882a593Smuzhiyun 	int ret;
172*4882a593Smuzhiyun 
173*4882a593Smuzhiyun 	switch (cmd) {
174*4882a593Smuzhiyun 	case GUP_FAST_BENCHMARK:
175*4882a593Smuzhiyun 	case GUP_BENCHMARK:
176*4882a593Smuzhiyun 	case PIN_FAST_BENCHMARK:
177*4882a593Smuzhiyun 	case PIN_BENCHMARK:
178*4882a593Smuzhiyun 	case PIN_LONGTERM_BENCHMARK:
179*4882a593Smuzhiyun 		break;
180*4882a593Smuzhiyun 	default:
181*4882a593Smuzhiyun 		return -EINVAL;
182*4882a593Smuzhiyun 	}
183*4882a593Smuzhiyun 
184*4882a593Smuzhiyun 	if (copy_from_user(&gup, (void __user *)arg, sizeof(gup)))
185*4882a593Smuzhiyun 		return -EFAULT;
186*4882a593Smuzhiyun 
187*4882a593Smuzhiyun 	ret = __gup_benchmark_ioctl(cmd, &gup);
188*4882a593Smuzhiyun 	if (ret)
189*4882a593Smuzhiyun 		return ret;
190*4882a593Smuzhiyun 
191*4882a593Smuzhiyun 	if (copy_to_user((void __user *)arg, &gup, sizeof(gup)))
192*4882a593Smuzhiyun 		return -EFAULT;
193*4882a593Smuzhiyun 
194*4882a593Smuzhiyun 	return 0;
195*4882a593Smuzhiyun }
196*4882a593Smuzhiyun 
197*4882a593Smuzhiyun static const struct file_operations gup_benchmark_fops = {
198*4882a593Smuzhiyun 	.open = nonseekable_open,
199*4882a593Smuzhiyun 	.unlocked_ioctl = gup_benchmark_ioctl,
200*4882a593Smuzhiyun };
201*4882a593Smuzhiyun 
gup_benchmark_init(void)202*4882a593Smuzhiyun static int gup_benchmark_init(void)
203*4882a593Smuzhiyun {
204*4882a593Smuzhiyun 	debugfs_create_file_unsafe("gup_benchmark", 0600, NULL, NULL,
205*4882a593Smuzhiyun 				   &gup_benchmark_fops);
206*4882a593Smuzhiyun 
207*4882a593Smuzhiyun 	return 0;
208*4882a593Smuzhiyun }
209*4882a593Smuzhiyun 
210*4882a593Smuzhiyun late_initcall(gup_benchmark_init);
211