1/*
2 * Virtual Processor Dispatch Trace Log
3 *
4 * (C) Copyright IBM Corporation 2009
5 *
6 * Author: Jeremy Kerr <jk@ozlabs.org>
7 *
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2, or (at your option)
11 * any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 */
22
23#include <linux/slab.h>
24#include <linux/debugfs.h>
25#include <linux/spinlock.h>
26#include <asm/smp.h>
27#include <asm/uaccess.h>
28#include <asm/firmware.h>
29#include <asm/lppaca.h>
30#include <asm/debug.h>
31#include <asm/plpar_wrappers.h>
32#include <asm/machdep.h>
33
34struct dtl {
35	struct dtl_entry	*buf;
36	struct dentry		*file;
37	int			cpu;
38	int			buf_entries;
39	u64			last_idx;
40	spinlock_t		lock;
41};
42static DEFINE_PER_CPU(struct dtl, cpu_dtl);
43
44/*
45 * Dispatch trace log event mask:
46 * 0x7: 0x1: voluntary virtual processor waits
47 *      0x2: time-slice preempts
48 *      0x4: virtual partition memory page faults
49 */
50static u8 dtl_event_mask = 0x7;
51
52
53/*
54 * Size of per-cpu log buffers. Firmware requires that the buffer does
55 * not cross a 4k boundary.
56 */
57static int dtl_buf_entries = N_DISPATCH_LOG;
58
59#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
60struct dtl_ring {
61	u64	write_index;
62	struct dtl_entry *write_ptr;
63	struct dtl_entry *buf;
64	struct dtl_entry *buf_end;
65	u8	saved_dtl_mask;
66};
67
68static DEFINE_PER_CPU(struct dtl_ring, dtl_rings);
69
70static atomic_t dtl_count;
71
72/*
73 * The cpu accounting code controls the DTL ring buffer, and we get
74 * given entries as they are processed.
75 */
76static void consume_dtle(struct dtl_entry *dtle, u64 index)
77{
78	struct dtl_ring *dtlr = this_cpu_ptr(&dtl_rings);
79	struct dtl_entry *wp = dtlr->write_ptr;
80	struct lppaca *vpa = local_paca->lppaca_ptr;
81
82	if (!wp)
83		return;
84
85	*wp = *dtle;
86	barrier();
87
88	/* check for hypervisor ring buffer overflow, ignore this entry if so */
89	if (index + N_DISPATCH_LOG < be64_to_cpu(vpa->dtl_idx))
90		return;
91
92	++wp;
93	if (wp == dtlr->buf_end)
94		wp = dtlr->buf;
95	dtlr->write_ptr = wp;
96
97	/* incrementing write_index makes the new entry visible */
98	smp_wmb();
99	++dtlr->write_index;
100}
101
102static int dtl_start(struct dtl *dtl)
103{
104	struct dtl_ring *dtlr = &per_cpu(dtl_rings, dtl->cpu);
105
106	dtlr->buf = dtl->buf;
107	dtlr->buf_end = dtl->buf + dtl->buf_entries;
108	dtlr->write_index = 0;
109
110	/* setting write_ptr enables logging into our buffer */
111	smp_wmb();
112	dtlr->write_ptr = dtl->buf;
113
114	/* enable event logging */
115	dtlr->saved_dtl_mask = lppaca_of(dtl->cpu).dtl_enable_mask;
116	lppaca_of(dtl->cpu).dtl_enable_mask |= dtl_event_mask;
117
118	dtl_consumer = consume_dtle;
119	atomic_inc(&dtl_count);
120	return 0;
121}
122
123static void dtl_stop(struct dtl *dtl)
124{
125	struct dtl_ring *dtlr = &per_cpu(dtl_rings, dtl->cpu);
126
127	dtlr->write_ptr = NULL;
128	smp_wmb();
129
130	dtlr->buf = NULL;
131
132	/* restore dtl_enable_mask */
133	lppaca_of(dtl->cpu).dtl_enable_mask = dtlr->saved_dtl_mask;
134
135	if (atomic_dec_and_test(&dtl_count))
136		dtl_consumer = NULL;
137}
138
139static u64 dtl_current_index(struct dtl *dtl)
140{
141	return per_cpu(dtl_rings, dtl->cpu).write_index;
142}
143
144#else /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
145
146static int dtl_start(struct dtl *dtl)
147{
148	unsigned long addr;
149	int ret, hwcpu;
150
151	/* Register our dtl buffer with the hypervisor. The HV expects the
152	 * buffer size to be passed in the second word of the buffer */
153	((u32 *)dtl->buf)[1] = DISPATCH_LOG_BYTES;
154
155	hwcpu = get_hard_smp_processor_id(dtl->cpu);
156	addr = __pa(dtl->buf);
157	ret = register_dtl(hwcpu, addr);
158	if (ret) {
159		printk(KERN_WARNING "%s: DTL registration for cpu %d (hw %d) "
160		       "failed with %d\n", __func__, dtl->cpu, hwcpu, ret);
161		return -EIO;
162	}
163
164	/* set our initial buffer indices */
165	lppaca_of(dtl->cpu).dtl_idx = 0;
166
167	/* ensure that our updates to the lppaca fields have occurred before
168	 * we actually enable the logging */
169	smp_wmb();
170
171	/* enable event logging */
172	lppaca_of(dtl->cpu).dtl_enable_mask = dtl_event_mask;
173
174	return 0;
175}
176
177static void dtl_stop(struct dtl *dtl)
178{
179	int hwcpu = get_hard_smp_processor_id(dtl->cpu);
180
181	lppaca_of(dtl->cpu).dtl_enable_mask = 0x0;
182
183	unregister_dtl(hwcpu);
184}
185
186static u64 dtl_current_index(struct dtl *dtl)
187{
188	return lppaca_of(dtl->cpu).dtl_idx;
189}
190#endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
191
192static int dtl_enable(struct dtl *dtl)
193{
194	long int n_entries;
195	long int rc;
196	struct dtl_entry *buf = NULL;
197
198	if (!dtl_cache)
199		return -ENOMEM;
200
201	/* only allow one reader */
202	if (dtl->buf)
203		return -EBUSY;
204
205	n_entries = dtl_buf_entries;
206	buf = kmem_cache_alloc_node(dtl_cache, GFP_KERNEL, cpu_to_node(dtl->cpu));
207	if (!buf) {
208		printk(KERN_WARNING "%s: buffer alloc failed for cpu %d\n",
209				__func__, dtl->cpu);
210		return -ENOMEM;
211	}
212
213	spin_lock(&dtl->lock);
214	rc = -EBUSY;
215	if (!dtl->buf) {
216		/* store the original allocation size for use during read */
217		dtl->buf_entries = n_entries;
218		dtl->buf = buf;
219		dtl->last_idx = 0;
220		rc = dtl_start(dtl);
221		if (rc)
222			dtl->buf = NULL;
223	}
224	spin_unlock(&dtl->lock);
225
226	if (rc)
227		kmem_cache_free(dtl_cache, buf);
228	return rc;
229}
230
231static void dtl_disable(struct dtl *dtl)
232{
233	spin_lock(&dtl->lock);
234	dtl_stop(dtl);
235	kmem_cache_free(dtl_cache, dtl->buf);
236	dtl->buf = NULL;
237	dtl->buf_entries = 0;
238	spin_unlock(&dtl->lock);
239}
240
241/* file interface */
242
243static int dtl_file_open(struct inode *inode, struct file *filp)
244{
245	struct dtl *dtl = inode->i_private;
246	int rc;
247
248	rc = dtl_enable(dtl);
249	if (rc)
250		return rc;
251
252	filp->private_data = dtl;
253	return 0;
254}
255
256static int dtl_file_release(struct inode *inode, struct file *filp)
257{
258	struct dtl *dtl = inode->i_private;
259	dtl_disable(dtl);
260	return 0;
261}
262
263static ssize_t dtl_file_read(struct file *filp, char __user *buf, size_t len,
264		loff_t *pos)
265{
266	long int rc, n_read, n_req, read_size;
267	struct dtl *dtl;
268	u64 cur_idx, last_idx, i;
269
270	if ((len % sizeof(struct dtl_entry)) != 0)
271		return -EINVAL;
272
273	dtl = filp->private_data;
274
275	/* requested number of entries to read */
276	n_req = len / sizeof(struct dtl_entry);
277
278	/* actual number of entries read */
279	n_read = 0;
280
281	spin_lock(&dtl->lock);
282
283	cur_idx = dtl_current_index(dtl);
284	last_idx = dtl->last_idx;
285
286	if (last_idx + dtl->buf_entries <= cur_idx)
287		last_idx = cur_idx - dtl->buf_entries + 1;
288
289	if (last_idx + n_req > cur_idx)
290		n_req = cur_idx - last_idx;
291
292	if (n_req > 0)
293		dtl->last_idx = last_idx + n_req;
294
295	spin_unlock(&dtl->lock);
296
297	if (n_req <= 0)
298		return 0;
299
300	i = last_idx % dtl->buf_entries;
301
302	/* read the tail of the buffer if we've wrapped */
303	if (i + n_req > dtl->buf_entries) {
304		read_size = dtl->buf_entries - i;
305
306		rc = copy_to_user(buf, &dtl->buf[i],
307				read_size * sizeof(struct dtl_entry));
308		if (rc)
309			return -EFAULT;
310
311		i = 0;
312		n_req -= read_size;
313		n_read += read_size;
314		buf += read_size * sizeof(struct dtl_entry);
315	}
316
317	/* .. and now the head */
318	rc = copy_to_user(buf, &dtl->buf[i], n_req * sizeof(struct dtl_entry));
319	if (rc)
320		return -EFAULT;
321
322	n_read += n_req;
323
324	return n_read * sizeof(struct dtl_entry);
325}
326
327static const struct file_operations dtl_fops = {
328	.open		= dtl_file_open,
329	.release	= dtl_file_release,
330	.read		= dtl_file_read,
331	.llseek		= no_llseek,
332};
333
334static struct dentry *dtl_dir;
335
336static int dtl_setup_file(struct dtl *dtl)
337{
338	char name[10];
339
340	sprintf(name, "cpu-%d", dtl->cpu);
341
342	dtl->file = debugfs_create_file(name, 0400, dtl_dir, dtl, &dtl_fops);
343	if (!dtl->file)
344		return -ENOMEM;
345
346	return 0;
347}
348
349static int dtl_init(void)
350{
351	struct dentry *event_mask_file, *buf_entries_file;
352	int rc, i;
353
354	if (!firmware_has_feature(FW_FEATURE_SPLPAR))
355		return -ENODEV;
356
357	/* set up common debugfs structure */
358
359	rc = -ENOMEM;
360	dtl_dir = debugfs_create_dir("dtl", powerpc_debugfs_root);
361	if (!dtl_dir) {
362		printk(KERN_WARNING "%s: can't create dtl root dir\n",
363				__func__);
364		goto err;
365	}
366
367	event_mask_file = debugfs_create_x8("dtl_event_mask", 0600,
368				dtl_dir, &dtl_event_mask);
369	buf_entries_file = debugfs_create_u32("dtl_buf_entries", 0400,
370				dtl_dir, &dtl_buf_entries);
371
372	if (!event_mask_file || !buf_entries_file) {
373		printk(KERN_WARNING "%s: can't create dtl files\n", __func__);
374		goto err_remove_dir;
375	}
376
377	/* set up the per-cpu log structures */
378	for_each_possible_cpu(i) {
379		struct dtl *dtl = &per_cpu(cpu_dtl, i);
380		spin_lock_init(&dtl->lock);
381		dtl->cpu = i;
382
383		rc = dtl_setup_file(dtl);
384		if (rc)
385			goto err_remove_dir;
386	}
387
388	return 0;
389
390err_remove_dir:
391	debugfs_remove_recursive(dtl_dir);
392err:
393	return rc;
394}
395machine_arch_initcall(pseries, dtl_init);
396