1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Copyright 2014 IBM Corp. 4 */ 5 6 #include <linux/spinlock.h> 7 #include <linux/kernel.h> 8 #include <linux/module.h> 9 #include <linux/device.h> 10 #include <linux/mutex.h> 11 #include <linux/init.h> 12 #include <linux/list.h> 13 #include <linux/mm.h> 14 #include <linux/of.h> 15 #include <linux/slab.h> 16 #include <linux/idr.h> 17 #include <linux/pci.h> 18 #include <linux/sched/task.h> 19 20 #include <asm/cputable.h> 21 #include <asm/mmu.h> 22 #include <misc/cxl-base.h> 23 24 #include "cxl.h" 25 #include "trace.h" 26 27 static DEFINE_SPINLOCK(adapter_idr_lock); 28 static DEFINE_IDR(cxl_adapter_idr); 29 30 uint cxl_verbose; 31 module_param_named(verbose, cxl_verbose, uint, 0600); 32 MODULE_PARM_DESC(verbose, "Enable verbose dmesg output"); 33 34 const struct cxl_backend_ops *cxl_ops; 35 36 int cxl_afu_slbia(struct cxl_afu *afu) 37 { 38 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 39 40 pr_devel("cxl_afu_slbia issuing SLBIA command\n"); 41 cxl_p2n_write(afu, CXL_SLBIA_An, CXL_TLB_SLB_IQ_ALL); 42 while (cxl_p2n_read(afu, CXL_SLBIA_An) & CXL_TLB_SLB_P) { 43 if (time_after_eq(jiffies, timeout)) { 44 dev_warn(&afu->dev, "WARNING: CXL AFU SLBIA timed out!\n"); 45 return -EBUSY; 46 } 47 /* If the adapter has gone down, we can assume that we 48 * will PERST it and that will invalidate everything. 49 */ 50 if (!cxl_ops->link_ok(afu->adapter, afu)) 51 return -EIO; 52 cpu_relax(); 53 } 54 return 0; 55 } 56 57 static inline void _cxl_slbia(struct cxl_context *ctx, struct mm_struct *mm) 58 { 59 unsigned long flags; 60 61 if (ctx->mm != mm) 62 return; 63 64 pr_devel("%s matched mm - card: %i afu: %i pe: %i\n", __func__, 65 ctx->afu->adapter->adapter_num, ctx->afu->slice, ctx->pe); 66 67 spin_lock_irqsave(&ctx->sste_lock, flags); 68 trace_cxl_slbia(ctx); 69 memset(ctx->sstp, 0, ctx->sst_size); 70 spin_unlock_irqrestore(&ctx->sste_lock, flags); 71 mb(); 72 cxl_afu_slbia(ctx->afu); 73 } 74 75 static inline void cxl_slbia_core(struct mm_struct *mm) 76 { 77 struct cxl *adapter; 78 struct cxl_afu *afu; 79 struct cxl_context *ctx; 80 int card, slice, id; 81 82 pr_devel("%s called\n", __func__); 83 84 spin_lock(&adapter_idr_lock); 85 idr_for_each_entry(&cxl_adapter_idr, adapter, card) { 86 /* XXX: Make this lookup faster with link from mm to ctx */ 87 spin_lock(&adapter->afu_list_lock); 88 for (slice = 0; slice < adapter->slices; slice++) { 89 afu = adapter->afu[slice]; 90 if (!afu || !afu->enabled) 91 continue; 92 rcu_read_lock(); 93 idr_for_each_entry(&afu->contexts_idr, ctx, id) 94 _cxl_slbia(ctx, mm); 95 rcu_read_unlock(); 96 } 97 spin_unlock(&adapter->afu_list_lock); 98 } 99 spin_unlock(&adapter_idr_lock); 100 } 101 102 static struct cxl_calls cxl_calls = { 103 .cxl_slbia = cxl_slbia_core, 104 .owner = THIS_MODULE, 105 }; 106 107 int cxl_alloc_sst(struct cxl_context *ctx) 108 { 109 unsigned long vsid; 110 u64 ea_mask, size, sstp0, sstp1; 111 112 sstp0 = 0; 113 sstp1 = 0; 114 115 ctx->sst_size = PAGE_SIZE; 116 ctx->sst_lru = 0; 117 ctx->sstp = (struct cxl_sste *)get_zeroed_page(GFP_KERNEL); 118 if (!ctx->sstp) { 119 pr_err("cxl_alloc_sst: Unable to allocate segment table\n"); 120 return -ENOMEM; 121 } 122 pr_devel("SSTP allocated at 0x%p\n", ctx->sstp); 123 124 vsid = get_kernel_vsid((u64)ctx->sstp, mmu_kernel_ssize) << 12; 125 126 sstp0 |= (u64)mmu_kernel_ssize << CXL_SSTP0_An_B_SHIFT; 127 sstp0 |= (SLB_VSID_KERNEL | mmu_psize_defs[mmu_linear_psize].sllp) << 50; 128 129 size = (((u64)ctx->sst_size >> 8) - 1) << CXL_SSTP0_An_SegTableSize_SHIFT; 130 if (unlikely(size & ~CXL_SSTP0_An_SegTableSize_MASK)) { 131 WARN(1, "Impossible segment table size\n"); 132 return -EINVAL; 133 } 134 sstp0 |= size; 135 136 if (mmu_kernel_ssize == MMU_SEGSIZE_256M) 137 ea_mask = 0xfffff00ULL; 138 else 139 ea_mask = 0xffffffff00ULL; 140 141 sstp0 |= vsid >> (50-14); /* Top 14 bits of VSID */ 142 sstp1 |= (vsid << (64-(50-14))) & ~ea_mask; 143 sstp1 |= (u64)ctx->sstp & ea_mask; 144 sstp1 |= CXL_SSTP1_An_V; 145 146 pr_devel("Looked up %#llx: slbfee. %#llx (ssize: %x, vsid: %#lx), copied to SSTP0: %#llx, SSTP1: %#llx\n", 147 (u64)ctx->sstp, (u64)ctx->sstp & ESID_MASK, mmu_kernel_ssize, vsid, sstp0, sstp1); 148 149 /* Store calculated sstp hardware points for use later */ 150 ctx->sstp0 = sstp0; 151 ctx->sstp1 = sstp1; 152 153 return 0; 154 } 155 156 /* print buffer content as integers when debugging */ 157 void cxl_dump_debug_buffer(void *buf, size_t buf_len) 158 { 159 #ifdef DEBUG 160 int i, *ptr; 161 162 /* 163 * We want to regroup up to 4 integers per line, which means they 164 * need to be in the same pr_devel() statement 165 */ 166 ptr = (int *) buf; 167 for (i = 0; i * 4 < buf_len; i += 4) { 168 if ((i + 3) * 4 < buf_len) 169 pr_devel("%.8x %.8x %.8x %.8x\n", ptr[i], ptr[i + 1], 170 ptr[i + 2], ptr[i + 3]); 171 else if ((i + 2) * 4 < buf_len) 172 pr_devel("%.8x %.8x %.8x\n", ptr[i], ptr[i + 1], 173 ptr[i + 2]); 174 else if ((i + 1) * 4 < buf_len) 175 pr_devel("%.8x %.8x\n", ptr[i], ptr[i + 1]); 176 else 177 pr_devel("%.8x\n", ptr[i]); 178 } 179 #endif /* DEBUG */ 180 } 181 182 /* Find a CXL adapter by it's number and increase it's refcount */ 183 struct cxl *get_cxl_adapter(int num) 184 { 185 struct cxl *adapter; 186 187 spin_lock(&adapter_idr_lock); 188 if ((adapter = idr_find(&cxl_adapter_idr, num))) 189 get_device(&adapter->dev); 190 spin_unlock(&adapter_idr_lock); 191 192 return adapter; 193 } 194 195 static int cxl_alloc_adapter_nr(struct cxl *adapter) 196 { 197 int i; 198 199 idr_preload(GFP_KERNEL); 200 spin_lock(&adapter_idr_lock); 201 i = idr_alloc(&cxl_adapter_idr, adapter, 0, 0, GFP_NOWAIT); 202 spin_unlock(&adapter_idr_lock); 203 idr_preload_end(); 204 if (i < 0) 205 return i; 206 207 adapter->adapter_num = i; 208 209 return 0; 210 } 211 212 void cxl_remove_adapter_nr(struct cxl *adapter) 213 { 214 idr_remove(&cxl_adapter_idr, adapter->adapter_num); 215 } 216 217 struct cxl *cxl_alloc_adapter(void) 218 { 219 struct cxl *adapter; 220 221 if (!(adapter = kzalloc(sizeof(struct cxl), GFP_KERNEL))) 222 return NULL; 223 224 spin_lock_init(&adapter->afu_list_lock); 225 226 if (cxl_alloc_adapter_nr(adapter)) 227 goto err1; 228 229 if (dev_set_name(&adapter->dev, "card%i", adapter->adapter_num)) 230 goto err2; 231 232 /* start with context lock taken */ 233 atomic_set(&adapter->contexts_num, -1); 234 235 return adapter; 236 err2: 237 cxl_remove_adapter_nr(adapter); 238 err1: 239 kfree(adapter); 240 return NULL; 241 } 242 243 struct cxl_afu *cxl_alloc_afu(struct cxl *adapter, int slice) 244 { 245 struct cxl_afu *afu; 246 247 if (!(afu = kzalloc(sizeof(struct cxl_afu), GFP_KERNEL))) 248 return NULL; 249 250 afu->adapter = adapter; 251 afu->dev.parent = &adapter->dev; 252 afu->dev.release = cxl_ops->release_afu; 253 afu->slice = slice; 254 idr_init(&afu->contexts_idr); 255 mutex_init(&afu->contexts_lock); 256 spin_lock_init(&afu->afu_cntl_lock); 257 atomic_set(&afu->configured_state, -1); 258 afu->prefault_mode = CXL_PREFAULT_NONE; 259 afu->irqs_max = afu->adapter->user_irqs; 260 261 return afu; 262 } 263 264 int cxl_afu_select_best_mode(struct cxl_afu *afu) 265 { 266 if (afu->modes_supported & CXL_MODE_DIRECTED) 267 return cxl_ops->afu_activate_mode(afu, CXL_MODE_DIRECTED); 268 269 if (afu->modes_supported & CXL_MODE_DEDICATED) 270 return cxl_ops->afu_activate_mode(afu, CXL_MODE_DEDICATED); 271 272 dev_warn(&afu->dev, "No supported programming modes available\n"); 273 /* We don't fail this so the user can inspect sysfs */ 274 return 0; 275 } 276 277 int cxl_adapter_context_get(struct cxl *adapter) 278 { 279 int rc; 280 281 rc = atomic_inc_unless_negative(&adapter->contexts_num); 282 return rc ? 0 : -EBUSY; 283 } 284 285 void cxl_adapter_context_put(struct cxl *adapter) 286 { 287 atomic_dec_if_positive(&adapter->contexts_num); 288 } 289 290 int cxl_adapter_context_lock(struct cxl *adapter) 291 { 292 int rc; 293 /* no active contexts -> contexts_num == 0 */ 294 rc = atomic_cmpxchg(&adapter->contexts_num, 0, -1); 295 return rc ? -EBUSY : 0; 296 } 297 298 void cxl_adapter_context_unlock(struct cxl *adapter) 299 { 300 int val = atomic_cmpxchg(&adapter->contexts_num, -1, 0); 301 302 /* 303 * contexts lock taken -> contexts_num == -1 304 * If not true then show a warning and force reset the lock. 305 * This will happen when context_unlock was requested without 306 * doing a context_lock. 307 */ 308 if (val != -1) { 309 atomic_set(&adapter->contexts_num, 0); 310 WARN(1, "Adapter context unlocked with %d active contexts", 311 val); 312 } 313 } 314 315 static int __init init_cxl(void) 316 { 317 int rc = 0; 318 319 if (!tlbie_capable) 320 return -EINVAL; 321 322 if ((rc = cxl_file_init())) 323 return rc; 324 325 cxl_debugfs_init(); 326 327 /* 328 * we don't register the callback on P9. slb callack is only 329 * used for the PSL8 MMU and CX4. 330 */ 331 if (cxl_is_power8()) { 332 rc = register_cxl_calls(&cxl_calls); 333 if (rc) 334 goto err; 335 } 336 337 if (cpu_has_feature(CPU_FTR_HVMODE)) { 338 cxl_ops = &cxl_native_ops; 339 rc = pci_register_driver(&cxl_pci_driver); 340 } 341 #ifdef CONFIG_PPC_PSERIES 342 else { 343 cxl_ops = &cxl_guest_ops; 344 rc = platform_driver_register(&cxl_of_driver); 345 } 346 #endif 347 if (rc) 348 goto err1; 349 350 return 0; 351 err1: 352 if (cxl_is_power8()) 353 unregister_cxl_calls(&cxl_calls); 354 err: 355 cxl_debugfs_exit(); 356 cxl_file_exit(); 357 358 return rc; 359 } 360 361 static void exit_cxl(void) 362 { 363 if (cpu_has_feature(CPU_FTR_HVMODE)) 364 pci_unregister_driver(&cxl_pci_driver); 365 #ifdef CONFIG_PPC_PSERIES 366 else 367 platform_driver_unregister(&cxl_of_driver); 368 #endif 369 370 cxl_debugfs_exit(); 371 cxl_file_exit(); 372 if (cxl_is_power8()) 373 unregister_cxl_calls(&cxl_calls); 374 idr_destroy(&cxl_adapter_idr); 375 } 376 377 module_init(init_cxl); 378 module_exit(exit_cxl); 379 380 MODULE_DESCRIPTION("IBM Coherent Accelerator"); 381 MODULE_AUTHOR("Ian Munsie <imunsie@au1.ibm.com>"); 382 MODULE_LICENSE("GPL"); 383