1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Virtual Processor Dispatch Trace Log 4 * 5 * (C) Copyright IBM Corporation 2009 6 * 7 * Author: Jeremy Kerr <jk@ozlabs.org> 8 */ 9 10 #include <linux/slab.h> 11 #include <linux/spinlock.h> 12 #include <asm/smp.h> 13 #include <linux/uaccess.h> 14 #include <asm/firmware.h> 15 #include <asm/lppaca.h> 16 #include <asm/debugfs.h> 17 #include <asm/plpar_wrappers.h> 18 #include <asm/machdep.h> 19 20 struct dtl { 21 struct dtl_entry *buf; 22 struct dentry *file; 23 int cpu; 24 int buf_entries; 25 u64 last_idx; 26 spinlock_t lock; 27 }; 28 static DEFINE_PER_CPU(struct dtl, cpu_dtl); 29 30 /* 31 * Dispatch trace log event mask: 32 * 0x7: 0x1: voluntary virtual processor waits 33 * 0x2: time-slice preempts 34 * 0x4: virtual partition memory page faults 35 */ 36 static u8 dtl_event_mask = 0x7; 37 38 39 /* 40 * Size of per-cpu log buffers. Firmware requires that the buffer does 41 * not cross a 4k boundary. 42 */ 43 static int dtl_buf_entries = N_DISPATCH_LOG; 44 45 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 46 struct dtl_ring { 47 u64 write_index; 48 struct dtl_entry *write_ptr; 49 struct dtl_entry *buf; 50 struct dtl_entry *buf_end; 51 u8 saved_dtl_mask; 52 }; 53 54 static DEFINE_PER_CPU(struct dtl_ring, dtl_rings); 55 56 static atomic_t dtl_count; 57 58 /* 59 * The cpu accounting code controls the DTL ring buffer, and we get 60 * given entries as they are processed. 61 */ 62 static void consume_dtle(struct dtl_entry *dtle, u64 index) 63 { 64 struct dtl_ring *dtlr = this_cpu_ptr(&dtl_rings); 65 struct dtl_entry *wp = dtlr->write_ptr; 66 struct lppaca *vpa = local_paca->lppaca_ptr; 67 68 if (!wp) 69 return; 70 71 *wp = *dtle; 72 barrier(); 73 74 /* check for hypervisor ring buffer overflow, ignore this entry if so */ 75 if (index + N_DISPATCH_LOG < be64_to_cpu(vpa->dtl_idx)) 76 return; 77 78 ++wp; 79 if (wp == dtlr->buf_end) 80 wp = dtlr->buf; 81 dtlr->write_ptr = wp; 82 83 /* incrementing write_index makes the new entry visible */ 84 smp_wmb(); 85 ++dtlr->write_index; 86 } 87 88 static int dtl_start(struct dtl *dtl) 89 { 90 struct dtl_ring *dtlr = &per_cpu(dtl_rings, dtl->cpu); 91 92 dtlr->buf = dtl->buf; 93 dtlr->buf_end = dtl->buf + dtl->buf_entries; 94 dtlr->write_index = 0; 95 96 /* setting write_ptr enables logging into our buffer */ 97 smp_wmb(); 98 dtlr->write_ptr = dtl->buf; 99 100 /* enable event logging */ 101 dtlr->saved_dtl_mask = lppaca_of(dtl->cpu).dtl_enable_mask; 102 lppaca_of(dtl->cpu).dtl_enable_mask |= dtl_event_mask; 103 104 dtl_consumer = consume_dtle; 105 atomic_inc(&dtl_count); 106 return 0; 107 } 108 109 static void dtl_stop(struct dtl *dtl) 110 { 111 struct dtl_ring *dtlr = &per_cpu(dtl_rings, dtl->cpu); 112 113 dtlr->write_ptr = NULL; 114 smp_wmb(); 115 116 dtlr->buf = NULL; 117 118 /* restore dtl_enable_mask */ 119 lppaca_of(dtl->cpu).dtl_enable_mask = dtlr->saved_dtl_mask; 120 121 if (atomic_dec_and_test(&dtl_count)) 122 dtl_consumer = NULL; 123 } 124 125 static u64 dtl_current_index(struct dtl *dtl) 126 { 127 return per_cpu(dtl_rings, dtl->cpu).write_index; 128 } 129 130 #else /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 131 132 static int dtl_start(struct dtl *dtl) 133 { 134 unsigned long addr; 135 int ret, hwcpu; 136 137 /* Register our dtl buffer with the hypervisor. The HV expects the 138 * buffer size to be passed in the second word of the buffer */ 139 ((u32 *)dtl->buf)[1] = cpu_to_be32(DISPATCH_LOG_BYTES); 140 141 hwcpu = get_hard_smp_processor_id(dtl->cpu); 142 addr = __pa(dtl->buf); 143 ret = register_dtl(hwcpu, addr); 144 if (ret) { 145 printk(KERN_WARNING "%s: DTL registration for cpu %d (hw %d) " 146 "failed with %d\n", __func__, dtl->cpu, hwcpu, ret); 147 return -EIO; 148 } 149 150 /* set our initial buffer indices */ 151 lppaca_of(dtl->cpu).dtl_idx = 0; 152 153 /* ensure that our updates to the lppaca fields have occurred before 154 * we actually enable the logging */ 155 smp_wmb(); 156 157 /* enable event logging */ 158 lppaca_of(dtl->cpu).dtl_enable_mask = dtl_event_mask; 159 160 return 0; 161 } 162 163 static void dtl_stop(struct dtl *dtl) 164 { 165 int hwcpu = get_hard_smp_processor_id(dtl->cpu); 166 167 lppaca_of(dtl->cpu).dtl_enable_mask = 0x0; 168 169 unregister_dtl(hwcpu); 170 } 171 172 static u64 dtl_current_index(struct dtl *dtl) 173 { 174 return be64_to_cpu(lppaca_of(dtl->cpu).dtl_idx); 175 } 176 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 177 178 static int dtl_enable(struct dtl *dtl) 179 { 180 long int n_entries; 181 long int rc; 182 struct dtl_entry *buf = NULL; 183 184 if (!dtl_cache) 185 return -ENOMEM; 186 187 /* only allow one reader */ 188 if (dtl->buf) 189 return -EBUSY; 190 191 n_entries = dtl_buf_entries; 192 buf = kmem_cache_alloc_node(dtl_cache, GFP_KERNEL, cpu_to_node(dtl->cpu)); 193 if (!buf) { 194 printk(KERN_WARNING "%s: buffer alloc failed for cpu %d\n", 195 __func__, dtl->cpu); 196 return -ENOMEM; 197 } 198 199 spin_lock(&dtl->lock); 200 rc = -EBUSY; 201 if (!dtl->buf) { 202 /* store the original allocation size for use during read */ 203 dtl->buf_entries = n_entries; 204 dtl->buf = buf; 205 dtl->last_idx = 0; 206 rc = dtl_start(dtl); 207 if (rc) 208 dtl->buf = NULL; 209 } 210 spin_unlock(&dtl->lock); 211 212 if (rc) 213 kmem_cache_free(dtl_cache, buf); 214 return rc; 215 } 216 217 static void dtl_disable(struct dtl *dtl) 218 { 219 spin_lock(&dtl->lock); 220 dtl_stop(dtl); 221 kmem_cache_free(dtl_cache, dtl->buf); 222 dtl->buf = NULL; 223 dtl->buf_entries = 0; 224 spin_unlock(&dtl->lock); 225 } 226 227 /* file interface */ 228 229 static int dtl_file_open(struct inode *inode, struct file *filp) 230 { 231 struct dtl *dtl = inode->i_private; 232 int rc; 233 234 rc = dtl_enable(dtl); 235 if (rc) 236 return rc; 237 238 filp->private_data = dtl; 239 return 0; 240 } 241 242 static int dtl_file_release(struct inode *inode, struct file *filp) 243 { 244 struct dtl *dtl = inode->i_private; 245 dtl_disable(dtl); 246 return 0; 247 } 248 249 static ssize_t dtl_file_read(struct file *filp, char __user *buf, size_t len, 250 loff_t *pos) 251 { 252 long int rc, n_read, n_req, read_size; 253 struct dtl *dtl; 254 u64 cur_idx, last_idx, i; 255 256 if ((len % sizeof(struct dtl_entry)) != 0) 257 return -EINVAL; 258 259 dtl = filp->private_data; 260 261 /* requested number of entries to read */ 262 n_req = len / sizeof(struct dtl_entry); 263 264 /* actual number of entries read */ 265 n_read = 0; 266 267 spin_lock(&dtl->lock); 268 269 cur_idx = dtl_current_index(dtl); 270 last_idx = dtl->last_idx; 271 272 if (last_idx + dtl->buf_entries <= cur_idx) 273 last_idx = cur_idx - dtl->buf_entries + 1; 274 275 if (last_idx + n_req > cur_idx) 276 n_req = cur_idx - last_idx; 277 278 if (n_req > 0) 279 dtl->last_idx = last_idx + n_req; 280 281 spin_unlock(&dtl->lock); 282 283 if (n_req <= 0) 284 return 0; 285 286 i = last_idx % dtl->buf_entries; 287 288 /* read the tail of the buffer if we've wrapped */ 289 if (i + n_req > dtl->buf_entries) { 290 read_size = dtl->buf_entries - i; 291 292 rc = copy_to_user(buf, &dtl->buf[i], 293 read_size * sizeof(struct dtl_entry)); 294 if (rc) 295 return -EFAULT; 296 297 i = 0; 298 n_req -= read_size; 299 n_read += read_size; 300 buf += read_size * sizeof(struct dtl_entry); 301 } 302 303 /* .. and now the head */ 304 rc = copy_to_user(buf, &dtl->buf[i], n_req * sizeof(struct dtl_entry)); 305 if (rc) 306 return -EFAULT; 307 308 n_read += n_req; 309 310 return n_read * sizeof(struct dtl_entry); 311 } 312 313 static const struct file_operations dtl_fops = { 314 .open = dtl_file_open, 315 .release = dtl_file_release, 316 .read = dtl_file_read, 317 .llseek = no_llseek, 318 }; 319 320 static struct dentry *dtl_dir; 321 322 static int dtl_setup_file(struct dtl *dtl) 323 { 324 char name[10]; 325 326 sprintf(name, "cpu-%d", dtl->cpu); 327 328 dtl->file = debugfs_create_file(name, 0400, dtl_dir, dtl, &dtl_fops); 329 if (!dtl->file) 330 return -ENOMEM; 331 332 return 0; 333 } 334 335 static int dtl_init(void) 336 { 337 struct dentry *event_mask_file, *buf_entries_file; 338 int rc, i; 339 340 if (!firmware_has_feature(FW_FEATURE_SPLPAR)) 341 return -ENODEV; 342 343 /* set up common debugfs structure */ 344 345 rc = -ENOMEM; 346 dtl_dir = debugfs_create_dir("dtl", powerpc_debugfs_root); 347 if (!dtl_dir) { 348 printk(KERN_WARNING "%s: can't create dtl root dir\n", 349 __func__); 350 goto err; 351 } 352 353 event_mask_file = debugfs_create_x8("dtl_event_mask", 0600, 354 dtl_dir, &dtl_event_mask); 355 buf_entries_file = debugfs_create_u32("dtl_buf_entries", 0400, 356 dtl_dir, &dtl_buf_entries); 357 358 if (!event_mask_file || !buf_entries_file) { 359 printk(KERN_WARNING "%s: can't create dtl files\n", __func__); 360 goto err_remove_dir; 361 } 362 363 /* set up the per-cpu log structures */ 364 for_each_possible_cpu(i) { 365 struct dtl *dtl = &per_cpu(cpu_dtl, i); 366 spin_lock_init(&dtl->lock); 367 dtl->cpu = i; 368 369 rc = dtl_setup_file(dtl); 370 if (rc) 371 goto err_remove_dir; 372 } 373 374 return 0; 375 376 err_remove_dir: 377 debugfs_remove_recursive(dtl_dir); 378 err: 379 return rc; 380 } 381 machine_arch_initcall(pseries, dtl_init); 382