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/dtl.h> 16 #include <asm/lppaca.h> 17 #include <asm/debugfs.h> 18 #include <asm/plpar_wrappers.h> 19 #include <asm/machdep.h> 20 21 struct dtl { 22 struct dtl_entry *buf; 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 static u8 dtl_event_mask = DTL_LOG_ALL; 31 32 33 /* 34 * Size of per-cpu log buffers. Firmware requires that the buffer does 35 * not cross a 4k boundary. 36 */ 37 static int dtl_buf_entries = N_DISPATCH_LOG; 38 39 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 40 struct dtl_ring { 41 u64 write_index; 42 struct dtl_entry *write_ptr; 43 struct dtl_entry *buf; 44 struct dtl_entry *buf_end; 45 }; 46 47 static DEFINE_PER_CPU(struct dtl_ring, dtl_rings); 48 49 static atomic_t dtl_count; 50 51 /* 52 * The cpu accounting code controls the DTL ring buffer, and we get 53 * given entries as they are processed. 54 */ 55 static void consume_dtle(struct dtl_entry *dtle, u64 index) 56 { 57 struct dtl_ring *dtlr = this_cpu_ptr(&dtl_rings); 58 struct dtl_entry *wp = dtlr->write_ptr; 59 struct lppaca *vpa = local_paca->lppaca_ptr; 60 61 if (!wp) 62 return; 63 64 *wp = *dtle; 65 barrier(); 66 67 /* check for hypervisor ring buffer overflow, ignore this entry if so */ 68 if (index + N_DISPATCH_LOG < be64_to_cpu(vpa->dtl_idx)) 69 return; 70 71 ++wp; 72 if (wp == dtlr->buf_end) 73 wp = dtlr->buf; 74 dtlr->write_ptr = wp; 75 76 /* incrementing write_index makes the new entry visible */ 77 smp_wmb(); 78 ++dtlr->write_index; 79 } 80 81 static int dtl_start(struct dtl *dtl) 82 { 83 struct dtl_ring *dtlr = &per_cpu(dtl_rings, dtl->cpu); 84 85 dtlr->buf = dtl->buf; 86 dtlr->buf_end = dtl->buf + dtl->buf_entries; 87 dtlr->write_index = 0; 88 89 /* setting write_ptr enables logging into our buffer */ 90 smp_wmb(); 91 dtlr->write_ptr = dtl->buf; 92 93 /* enable event logging */ 94 lppaca_of(dtl->cpu).dtl_enable_mask |= dtl_event_mask; 95 96 dtl_consumer = consume_dtle; 97 atomic_inc(&dtl_count); 98 return 0; 99 } 100 101 static void dtl_stop(struct dtl *dtl) 102 { 103 struct dtl_ring *dtlr = &per_cpu(dtl_rings, dtl->cpu); 104 105 dtlr->write_ptr = NULL; 106 smp_wmb(); 107 108 dtlr->buf = NULL; 109 110 /* restore dtl_enable_mask */ 111 lppaca_of(dtl->cpu).dtl_enable_mask = DTL_LOG_PREEMPT; 112 113 if (atomic_dec_and_test(&dtl_count)) 114 dtl_consumer = NULL; 115 } 116 117 static u64 dtl_current_index(struct dtl *dtl) 118 { 119 return per_cpu(dtl_rings, dtl->cpu).write_index; 120 } 121 122 #else /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 123 124 static int dtl_start(struct dtl *dtl) 125 { 126 unsigned long addr; 127 int ret, hwcpu; 128 129 /* Register our dtl buffer with the hypervisor. The HV expects the 130 * buffer size to be passed in the second word of the buffer */ 131 ((u32 *)dtl->buf)[1] = cpu_to_be32(DISPATCH_LOG_BYTES); 132 133 hwcpu = get_hard_smp_processor_id(dtl->cpu); 134 addr = __pa(dtl->buf); 135 ret = register_dtl(hwcpu, addr); 136 if (ret) { 137 printk(KERN_WARNING "%s: DTL registration for cpu %d (hw %d) " 138 "failed with %d\n", __func__, dtl->cpu, hwcpu, ret); 139 return -EIO; 140 } 141 142 /* set our initial buffer indices */ 143 lppaca_of(dtl->cpu).dtl_idx = 0; 144 145 /* ensure that our updates to the lppaca fields have occurred before 146 * we actually enable the logging */ 147 smp_wmb(); 148 149 /* enable event logging */ 150 lppaca_of(dtl->cpu).dtl_enable_mask = dtl_event_mask; 151 152 return 0; 153 } 154 155 static void dtl_stop(struct dtl *dtl) 156 { 157 int hwcpu = get_hard_smp_processor_id(dtl->cpu); 158 159 lppaca_of(dtl->cpu).dtl_enable_mask = 0x0; 160 161 unregister_dtl(hwcpu); 162 } 163 164 static u64 dtl_current_index(struct dtl *dtl) 165 { 166 return be64_to_cpu(lppaca_of(dtl->cpu).dtl_idx); 167 } 168 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */ 169 170 static int dtl_enable(struct dtl *dtl) 171 { 172 long int n_entries; 173 long int rc; 174 struct dtl_entry *buf = NULL; 175 176 if (!dtl_cache) 177 return -ENOMEM; 178 179 /* only allow one reader */ 180 if (dtl->buf) 181 return -EBUSY; 182 183 /* ensure there are no other conflicting dtl users */ 184 if (!read_trylock(&dtl_access_lock)) 185 return -EBUSY; 186 187 n_entries = dtl_buf_entries; 188 buf = kmem_cache_alloc_node(dtl_cache, GFP_KERNEL, cpu_to_node(dtl->cpu)); 189 if (!buf) { 190 printk(KERN_WARNING "%s: buffer alloc failed for cpu %d\n", 191 __func__, dtl->cpu); 192 read_unlock(&dtl_access_lock); 193 return -ENOMEM; 194 } 195 196 spin_lock(&dtl->lock); 197 rc = -EBUSY; 198 if (!dtl->buf) { 199 /* store the original allocation size for use during read */ 200 dtl->buf_entries = n_entries; 201 dtl->buf = buf; 202 dtl->last_idx = 0; 203 rc = dtl_start(dtl); 204 if (rc) 205 dtl->buf = NULL; 206 } 207 spin_unlock(&dtl->lock); 208 209 if (rc) { 210 read_unlock(&dtl_access_lock); 211 kmem_cache_free(dtl_cache, buf); 212 } 213 214 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 read_unlock(&dtl_access_lock); 226 } 227 228 /* file interface */ 229 230 static int dtl_file_open(struct inode *inode, struct file *filp) 231 { 232 struct dtl *dtl = inode->i_private; 233 int rc; 234 235 rc = dtl_enable(dtl); 236 if (rc) 237 return rc; 238 239 filp->private_data = dtl; 240 return 0; 241 } 242 243 static int dtl_file_release(struct inode *inode, struct file *filp) 244 { 245 struct dtl *dtl = inode->i_private; 246 dtl_disable(dtl); 247 return 0; 248 } 249 250 static ssize_t dtl_file_read(struct file *filp, char __user *buf, size_t len, 251 loff_t *pos) 252 { 253 long int rc, n_read, n_req, read_size; 254 struct dtl *dtl; 255 u64 cur_idx, last_idx, i; 256 257 if ((len % sizeof(struct dtl_entry)) != 0) 258 return -EINVAL; 259 260 dtl = filp->private_data; 261 262 /* requested number of entries to read */ 263 n_req = len / sizeof(struct dtl_entry); 264 265 /* actual number of entries read */ 266 n_read = 0; 267 268 spin_lock(&dtl->lock); 269 270 cur_idx = dtl_current_index(dtl); 271 last_idx = dtl->last_idx; 272 273 if (last_idx + dtl->buf_entries <= cur_idx) 274 last_idx = cur_idx - dtl->buf_entries + 1; 275 276 if (last_idx + n_req > cur_idx) 277 n_req = cur_idx - last_idx; 278 279 if (n_req > 0) 280 dtl->last_idx = last_idx + n_req; 281 282 spin_unlock(&dtl->lock); 283 284 if (n_req <= 0) 285 return 0; 286 287 i = last_idx % dtl->buf_entries; 288 289 /* read the tail of the buffer if we've wrapped */ 290 if (i + n_req > dtl->buf_entries) { 291 read_size = dtl->buf_entries - i; 292 293 rc = copy_to_user(buf, &dtl->buf[i], 294 read_size * sizeof(struct dtl_entry)); 295 if (rc) 296 return -EFAULT; 297 298 i = 0; 299 n_req -= read_size; 300 n_read += read_size; 301 buf += read_size * sizeof(struct dtl_entry); 302 } 303 304 /* .. and now the head */ 305 rc = copy_to_user(buf, &dtl->buf[i], n_req * sizeof(struct dtl_entry)); 306 if (rc) 307 return -EFAULT; 308 309 n_read += n_req; 310 311 return n_read * sizeof(struct dtl_entry); 312 } 313 314 static const struct file_operations dtl_fops = { 315 .open = dtl_file_open, 316 .release = dtl_file_release, 317 .read = dtl_file_read, 318 .llseek = no_llseek, 319 }; 320 321 static struct dentry *dtl_dir; 322 323 static void dtl_setup_file(struct dtl *dtl) 324 { 325 char name[10]; 326 327 sprintf(name, "cpu-%d", dtl->cpu); 328 329 debugfs_create_file(name, 0400, dtl_dir, dtl, &dtl_fops); 330 } 331 332 static int dtl_init(void) 333 { 334 int i; 335 336 if (!firmware_has_feature(FW_FEATURE_SPLPAR)) 337 return -ENODEV; 338 339 /* set up common debugfs structure */ 340 341 dtl_dir = debugfs_create_dir("dtl", powerpc_debugfs_root); 342 343 debugfs_create_x8("dtl_event_mask", 0600, dtl_dir, &dtl_event_mask); 344 debugfs_create_u32("dtl_buf_entries", 0400, dtl_dir, &dtl_buf_entries); 345 346 /* set up the per-cpu log structures */ 347 for_each_possible_cpu(i) { 348 struct dtl *dtl = &per_cpu(cpu_dtl, i); 349 spin_lock_init(&dtl->lock); 350 dtl->cpu = i; 351 352 dtl_setup_file(dtl); 353 } 354 355 return 0; 356 } 357 machine_arch_initcall(pseries, dtl_init); 358