1 /* 2 * Copyright 2014 IBM Corp. 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public License 6 * as published by the Free Software Foundation; either version 7 * 2 of the License, or (at your option) any later version. 8 */ 9 10 #include <linux/module.h> 11 #include <linux/kernel.h> 12 #include <linux/bitmap.h> 13 #include <linux/sched.h> 14 #include <linux/pid.h> 15 #include <linux/fs.h> 16 #include <linux/mm.h> 17 #include <linux/debugfs.h> 18 #include <linux/slab.h> 19 #include <linux/idr.h> 20 #include <linux/sched/mm.h> 21 #include <linux/mmu_context.h> 22 #include <asm/cputable.h> 23 #include <asm/current.h> 24 #include <asm/copro.h> 25 26 #include "cxl.h" 27 28 /* 29 * Allocates space for a CXL context. 30 */ 31 struct cxl_context *cxl_context_alloc(void) 32 { 33 return kzalloc(sizeof(struct cxl_context), GFP_KERNEL); 34 } 35 36 /* 37 * Initialises a CXL context. 38 */ 39 int cxl_context_init(struct cxl_context *ctx, struct cxl_afu *afu, bool master) 40 { 41 int i; 42 43 ctx->afu = afu; 44 ctx->master = master; 45 ctx->pid = NULL; /* Set in start work ioctl */ 46 mutex_init(&ctx->mapping_lock); 47 ctx->mapping = NULL; 48 ctx->tidr = 0; 49 ctx->assign_tidr = false; 50 51 if (cxl_is_power8()) { 52 spin_lock_init(&ctx->sste_lock); 53 54 /* 55 * Allocate the segment table before we put it in the IDR so that we 56 * can always access it when dereferenced from IDR. For the same 57 * reason, the segment table is only destroyed after the context is 58 * removed from the IDR. Access to this in the IOCTL is protected by 59 * Linux filesytem symantics (can't IOCTL until open is complete). 60 */ 61 i = cxl_alloc_sst(ctx); 62 if (i) 63 return i; 64 } 65 66 INIT_WORK(&ctx->fault_work, cxl_handle_fault); 67 68 init_waitqueue_head(&ctx->wq); 69 spin_lock_init(&ctx->lock); 70 71 ctx->irq_bitmap = NULL; 72 ctx->pending_irq = false; 73 ctx->pending_fault = false; 74 ctx->pending_afu_err = false; 75 76 INIT_LIST_HEAD(&ctx->irq_names); 77 INIT_LIST_HEAD(&ctx->extra_irq_contexts); 78 79 /* 80 * When we have to destroy all contexts in cxl_context_detach_all() we 81 * end up with afu_release_irqs() called from inside a 82 * idr_for_each_entry(). Hence we need to make sure that anything 83 * dereferenced from this IDR is ok before we allocate the IDR here. 84 * This clears out the IRQ ranges to ensure this. 85 */ 86 for (i = 0; i < CXL_IRQ_RANGES; i++) 87 ctx->irqs.range[i] = 0; 88 89 mutex_init(&ctx->status_mutex); 90 91 ctx->status = OPENED; 92 93 /* 94 * Allocating IDR! We better make sure everything's setup that 95 * dereferences from it. 96 */ 97 mutex_lock(&afu->contexts_lock); 98 idr_preload(GFP_KERNEL); 99 i = idr_alloc(&ctx->afu->contexts_idr, ctx, ctx->afu->adapter->min_pe, 100 ctx->afu->num_procs, GFP_NOWAIT); 101 idr_preload_end(); 102 mutex_unlock(&afu->contexts_lock); 103 if (i < 0) 104 return i; 105 106 ctx->pe = i; 107 if (cpu_has_feature(CPU_FTR_HVMODE)) { 108 ctx->elem = &ctx->afu->native->spa[i]; 109 ctx->external_pe = ctx->pe; 110 } else { 111 ctx->external_pe = -1; /* assigned when attaching */ 112 } 113 ctx->pe_inserted = false; 114 115 /* 116 * take a ref on the afu so that it stays alive at-least till 117 * this context is reclaimed inside reclaim_ctx. 118 */ 119 cxl_afu_get(afu); 120 return 0; 121 } 122 123 void cxl_context_set_mapping(struct cxl_context *ctx, 124 struct address_space *mapping) 125 { 126 mutex_lock(&ctx->mapping_lock); 127 ctx->mapping = mapping; 128 mutex_unlock(&ctx->mapping_lock); 129 } 130 131 static vm_fault_t cxl_mmap_fault(struct vm_fault *vmf) 132 { 133 struct vm_area_struct *vma = vmf->vma; 134 struct cxl_context *ctx = vma->vm_file->private_data; 135 u64 area, offset; 136 vm_fault_t ret; 137 138 offset = vmf->pgoff << PAGE_SHIFT; 139 140 pr_devel("%s: pe: %i address: 0x%lx offset: 0x%llx\n", 141 __func__, ctx->pe, vmf->address, offset); 142 143 if (ctx->afu->current_mode == CXL_MODE_DEDICATED) { 144 area = ctx->afu->psn_phys; 145 if (offset >= ctx->afu->adapter->ps_size) 146 return VM_FAULT_SIGBUS; 147 } else { 148 area = ctx->psn_phys; 149 if (offset >= ctx->psn_size) 150 return VM_FAULT_SIGBUS; 151 } 152 153 mutex_lock(&ctx->status_mutex); 154 155 if (ctx->status != STARTED) { 156 mutex_unlock(&ctx->status_mutex); 157 pr_devel("%s: Context not started, failing problem state access\n", __func__); 158 if (ctx->mmio_err_ff) { 159 if (!ctx->ff_page) { 160 ctx->ff_page = alloc_page(GFP_USER); 161 if (!ctx->ff_page) 162 return VM_FAULT_OOM; 163 memset(page_address(ctx->ff_page), 0xff, PAGE_SIZE); 164 } 165 get_page(ctx->ff_page); 166 vmf->page = ctx->ff_page; 167 vma->vm_page_prot = pgprot_cached(vma->vm_page_prot); 168 return 0; 169 } 170 return VM_FAULT_SIGBUS; 171 } 172 173 ret = vmf_insert_pfn(vma, vmf->address, (area + offset) >> PAGE_SHIFT); 174 175 mutex_unlock(&ctx->status_mutex); 176 177 return ret; 178 } 179 180 static const struct vm_operations_struct cxl_mmap_vmops = { 181 .fault = cxl_mmap_fault, 182 }; 183 184 /* 185 * Map a per-context mmio space into the given vma. 186 */ 187 int cxl_context_iomap(struct cxl_context *ctx, struct vm_area_struct *vma) 188 { 189 u64 start = vma->vm_pgoff << PAGE_SHIFT; 190 u64 len = vma->vm_end - vma->vm_start; 191 192 if (ctx->afu->current_mode == CXL_MODE_DEDICATED) { 193 if (start + len > ctx->afu->adapter->ps_size) 194 return -EINVAL; 195 196 if (cxl_is_power9()) { 197 /* 198 * Make sure there is a valid problem state 199 * area space for this AFU. 200 */ 201 if (ctx->master && !ctx->afu->psa) { 202 pr_devel("AFU doesn't support mmio space\n"); 203 return -EINVAL; 204 } 205 206 /* Can't mmap until the AFU is enabled */ 207 if (!ctx->afu->enabled) 208 return -EBUSY; 209 } 210 } else { 211 if (start + len > ctx->psn_size) 212 return -EINVAL; 213 214 /* Make sure there is a valid per process space for this AFU */ 215 if ((ctx->master && !ctx->afu->psa) || (!ctx->afu->pp_psa)) { 216 pr_devel("AFU doesn't support mmio space\n"); 217 return -EINVAL; 218 } 219 220 /* Can't mmap until the AFU is enabled */ 221 if (!ctx->afu->enabled) 222 return -EBUSY; 223 } 224 225 pr_devel("%s: mmio physical: %llx pe: %i master:%i\n", __func__, 226 ctx->psn_phys, ctx->pe , ctx->master); 227 228 vma->vm_flags |= VM_IO | VM_PFNMAP; 229 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 230 vma->vm_ops = &cxl_mmap_vmops; 231 return 0; 232 } 233 234 /* 235 * Detach a context from the hardware. This disables interrupts and doesn't 236 * return until all outstanding interrupts for this context have completed. The 237 * hardware should no longer access *ctx after this has returned. 238 */ 239 int __detach_context(struct cxl_context *ctx) 240 { 241 enum cxl_context_status status; 242 243 mutex_lock(&ctx->status_mutex); 244 status = ctx->status; 245 ctx->status = CLOSED; 246 mutex_unlock(&ctx->status_mutex); 247 if (status != STARTED) 248 return -EBUSY; 249 250 /* Only warn if we detached while the link was OK. 251 * If detach fails when hw is down, we don't care. 252 */ 253 WARN_ON(cxl_ops->detach_process(ctx) && 254 cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)); 255 flush_work(&ctx->fault_work); /* Only needed for dedicated process */ 256 257 /* 258 * Wait until no further interrupts are presented by the PSL 259 * for this context. 260 */ 261 if (cxl_ops->irq_wait) 262 cxl_ops->irq_wait(ctx); 263 264 /* release the reference to the group leader and mm handling pid */ 265 put_pid(ctx->pid); 266 267 cxl_ctx_put(); 268 269 /* Decrease the attached context count on the adapter */ 270 cxl_adapter_context_put(ctx->afu->adapter); 271 272 /* Decrease the mm count on the context */ 273 cxl_context_mm_count_put(ctx); 274 if (ctx->mm) 275 mm_context_remove_copro(ctx->mm); 276 ctx->mm = NULL; 277 278 return 0; 279 } 280 281 /* 282 * Detach the given context from the AFU. This doesn't actually 283 * free the context but it should stop the context running in hardware 284 * (ie. prevent this context from generating any further interrupts 285 * so that it can be freed). 286 */ 287 void cxl_context_detach(struct cxl_context *ctx) 288 { 289 int rc; 290 291 rc = __detach_context(ctx); 292 if (rc) 293 return; 294 295 afu_release_irqs(ctx, ctx); 296 wake_up_all(&ctx->wq); 297 } 298 299 /* 300 * Detach all contexts on the given AFU. 301 */ 302 void cxl_context_detach_all(struct cxl_afu *afu) 303 { 304 struct cxl_context *ctx; 305 int tmp; 306 307 mutex_lock(&afu->contexts_lock); 308 idr_for_each_entry(&afu->contexts_idr, ctx, tmp) { 309 /* 310 * Anything done in here needs to be setup before the IDR is 311 * created and torn down after the IDR removed 312 */ 313 cxl_context_detach(ctx); 314 315 /* 316 * We are force detaching - remove any active PSA mappings so 317 * userspace cannot interfere with the card if it comes back. 318 * Easiest way to exercise this is to unbind and rebind the 319 * driver via sysfs while it is in use. 320 */ 321 mutex_lock(&ctx->mapping_lock); 322 if (ctx->mapping) 323 unmap_mapping_range(ctx->mapping, 0, 0, 1); 324 mutex_unlock(&ctx->mapping_lock); 325 } 326 mutex_unlock(&afu->contexts_lock); 327 } 328 329 static void reclaim_ctx(struct rcu_head *rcu) 330 { 331 struct cxl_context *ctx = container_of(rcu, struct cxl_context, rcu); 332 333 if (cxl_is_power8()) 334 free_page((u64)ctx->sstp); 335 if (ctx->ff_page) 336 __free_page(ctx->ff_page); 337 ctx->sstp = NULL; 338 339 kfree(ctx->irq_bitmap); 340 341 /* Drop ref to the afu device taken during cxl_context_init */ 342 cxl_afu_put(ctx->afu); 343 344 kfree(ctx); 345 } 346 347 void cxl_context_free(struct cxl_context *ctx) 348 { 349 if (ctx->kernelapi && ctx->mapping) 350 cxl_release_mapping(ctx); 351 mutex_lock(&ctx->afu->contexts_lock); 352 idr_remove(&ctx->afu->contexts_idr, ctx->pe); 353 mutex_unlock(&ctx->afu->contexts_lock); 354 call_rcu(&ctx->rcu, reclaim_ctx); 355 } 356 357 void cxl_context_mm_count_get(struct cxl_context *ctx) 358 { 359 if (ctx->mm) 360 atomic_inc(&ctx->mm->mm_count); 361 } 362 363 void cxl_context_mm_count_put(struct cxl_context *ctx) 364 { 365 if (ctx->mm) 366 mmdrop(ctx->mm); 367 } 368