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 int cpu; 23 int buf_entries; 24 u64 last_idx; 25 spinlock_t lock; 26 }; 27 static DEFINE_PER_CPU(struct dtl, cpu_dtl); 28 29 static u8 dtl_event_mask = DTL_LOG_ALL; 30 31 32 /* 33 * Size of per-cpu log buffers. Firmware requires that the buffer does 34 * not cross a 4k boundary. 35 */ 36 static int dtl_buf_entries = N_DISPATCH_LOG; 37 38 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 39 struct dtl_ring { 40 u64 write_index; 41 struct dtl_entry *write_ptr; 42 struct dtl_entry *buf; 43 struct dtl_entry *buf_end; 44 }; 45 46 static DEFINE_PER_CPU(struct dtl_ring, dtl_rings); 47 48 static atomic_t dtl_count; 49 50 /* 51 * The cpu accounting code controls the DTL ring buffer, and we get 52 * given entries as they are processed. 53 */ 54 static void consume_dtle(struct dtl_entry *dtle, u64 index) 55 { 56 struct dtl_ring *dtlr = this_cpu_ptr(&dtl_rings); 57 struct dtl_entry *wp = dtlr->write_ptr; 58 struct lppaca *vpa = local_paca->lppaca_ptr; 59 60 if (!wp) 61 return; 62 63 *wp = *dtle; 64 barrier(); 65 66 /* check for hypervisor ring buffer overflow, ignore this entry if so */ 67 if (index + N_DISPATCH_LOG < be64_to_cpu(vpa->dtl_idx)) 68 return; 69 70 ++wp; 71 if (wp == dtlr->buf_end) 72 wp = dtlr->buf; 73 dtlr->write_ptr = wp; 74 75 /* incrementing write_index makes the new entry visible */ 76 smp_wmb(); 77 ++dtlr->write_index; 78 } 79 80 static int dtl_start(struct dtl *dtl) 81 { 82 struct dtl_ring *dtlr = &per_cpu(dtl_rings, dtl->cpu); 83 84 dtlr->buf = dtl->buf; 85 dtlr->buf_end = dtl->buf + dtl->buf_entries; 86 dtlr->write_index = 0; 87 88 /* setting write_ptr enables logging into our buffer */ 89 smp_wmb(); 90 dtlr->write_ptr = dtl->buf; 91 92 /* enable event logging */ 93 lppaca_of(dtl->cpu).dtl_enable_mask |= dtl_event_mask; 94 95 dtl_consumer = consume_dtle; 96 atomic_inc(&dtl_count); 97 return 0; 98 } 99 100 static void dtl_stop(struct dtl *dtl) 101 { 102 struct dtl_ring *dtlr = &per_cpu(dtl_rings, dtl->cpu); 103 104 dtlr->write_ptr = NULL; 105 smp_wmb(); 106 107 dtlr->buf = NULL; 108 109 /* restore dtl_enable_mask */ 110 lppaca_of(dtl->cpu).dtl_enable_mask = DTL_LOG_PREEMPT; 111 112 if (atomic_dec_and_test(&dtl_count)) 113 dtl_consumer = NULL; 114 } 115 116 static u64 dtl_current_index(struct dtl *dtl) 117 { 118 return per_cpu(dtl_rings, dtl->cpu).write_index; 119 } 120 121 #else /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 122 123 static int dtl_start(struct dtl *dtl) 124 { 125 unsigned long addr; 126 int ret, hwcpu; 127 128 /* Register our dtl buffer with the hypervisor. The HV expects the 129 * buffer size to be passed in the second word of the buffer */ 130 ((u32 *)dtl->buf)[1] = cpu_to_be32(DISPATCH_LOG_BYTES); 131 132 hwcpu = get_hard_smp_processor_id(dtl->cpu); 133 addr = __pa(dtl->buf); 134 ret = register_dtl(hwcpu, addr); 135 if (ret) { 136 printk(KERN_WARNING "%s: DTL registration for cpu %d (hw %d) " 137 "failed with %d\n", __func__, dtl->cpu, hwcpu, ret); 138 return -EIO; 139 } 140 141 /* set our initial buffer indices */ 142 lppaca_of(dtl->cpu).dtl_idx = 0; 143 144 /* ensure that our updates to the lppaca fields have occurred before 145 * we actually enable the logging */ 146 smp_wmb(); 147 148 /* enable event logging */ 149 lppaca_of(dtl->cpu).dtl_enable_mask = dtl_event_mask; 150 151 return 0; 152 } 153 154 static void dtl_stop(struct dtl *dtl) 155 { 156 int hwcpu = get_hard_smp_processor_id(dtl->cpu); 157 158 lppaca_of(dtl->cpu).dtl_enable_mask = 0x0; 159 160 unregister_dtl(hwcpu); 161 } 162 163 static u64 dtl_current_index(struct dtl *dtl) 164 { 165 return be64_to_cpu(lppaca_of(dtl->cpu).dtl_idx); 166 } 167 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 168 169 static int dtl_enable(struct dtl *dtl) 170 { 171 long int n_entries; 172 long int rc; 173 struct dtl_entry *buf = NULL; 174 175 if (!dtl_cache) 176 return -ENOMEM; 177 178 /* only allow one reader */ 179 if (dtl->buf) 180 return -EBUSY; 181 182 /* ensure there are no other conflicting dtl users */ 183 if (!read_trylock(&dtl_access_lock)) 184 return -EBUSY; 185 186 n_entries = dtl_buf_entries; 187 buf = kmem_cache_alloc_node(dtl_cache, GFP_KERNEL, cpu_to_node(dtl->cpu)); 188 if (!buf) { 189 printk(KERN_WARNING "%s: buffer alloc failed for cpu %d\n", 190 __func__, dtl->cpu); 191 read_unlock(&dtl_access_lock); 192 return -ENOMEM; 193 } 194 195 spin_lock(&dtl->lock); 196 rc = -EBUSY; 197 if (!dtl->buf) { 198 /* store the original allocation size for use during read */ 199 dtl->buf_entries = n_entries; 200 dtl->buf = buf; 201 dtl->last_idx = 0; 202 rc = dtl_start(dtl); 203 if (rc) 204 dtl->buf = NULL; 205 } 206 spin_unlock(&dtl->lock); 207 208 if (rc) { 209 read_unlock(&dtl_access_lock); 210 kmem_cache_free(dtl_cache, buf); 211 } 212 213 return rc; 214 } 215 216 static void dtl_disable(struct dtl *dtl) 217 { 218 spin_lock(&dtl->lock); 219 dtl_stop(dtl); 220 kmem_cache_free(dtl_cache, dtl->buf); 221 dtl->buf = NULL; 222 dtl->buf_entries = 0; 223 spin_unlock(&dtl->lock); 224 read_unlock(&dtl_access_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 void dtl_setup_file(struct dtl *dtl) 323 { 324 char name[10]; 325 326 sprintf(name, "cpu-%d", dtl->cpu); 327 328 debugfs_create_file(name, 0400, dtl_dir, dtl, &dtl_fops); 329 } 330 331 static int dtl_init(void) 332 { 333 int i; 334 335 if (!firmware_has_feature(FW_FEATURE_SPLPAR)) 336 return -ENODEV; 337 338 /* set up common debugfs structure */ 339 340 dtl_dir = debugfs_create_dir("dtl", powerpc_debugfs_root); 341 342 debugfs_create_x8("dtl_event_mask", 0600, dtl_dir, &dtl_event_mask); 343 debugfs_create_u32("dtl_buf_entries", 0400, dtl_dir, &dtl_buf_entries); 344 345 /* set up the per-cpu log structures */ 346 for_each_possible_cpu(i) { 347 struct dtl *dtl = &per_cpu(cpu_dtl, i); 348 spin_lock_init(&dtl->lock); 349 dtl->cpu = i; 350 351 dtl_setup_file(dtl); 352 } 353 354 return 0; 355 } 356 machine_arch_initcall(pseries, dtl_init); 357