Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 1 | /* |
| 2 | * |
| 3 | * This file is provided under a dual BSD/GPLv2 license. When using or |
| 4 | * redistributing this file, you may do so under either license. |
| 5 | * |
| 6 | * GPL LICENSE SUMMARY |
| 7 | * |
| 8 | * Copyright(c) 2015 Intel Corporation. |
| 9 | * |
| 10 | * This program is free software; you can redistribute it and/or modify |
| 11 | * it under the terms of version 2 of the GNU General Public License as |
| 12 | * published by the Free Software Foundation. |
| 13 | * |
| 14 | * This program is distributed in the hope that it will be useful, but |
| 15 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
| 16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 17 | * General Public License for more details. |
| 18 | * |
| 19 | * BSD LICENSE |
| 20 | * |
| 21 | * Copyright(c) 2015 Intel Corporation. |
| 22 | * |
| 23 | * Redistribution and use in source and binary forms, with or without |
| 24 | * modification, are permitted provided that the following conditions |
| 25 | * are met: |
| 26 | * |
| 27 | * - Redistributions of source code must retain the above copyright |
| 28 | * notice, this list of conditions and the following disclaimer. |
| 29 | * - Redistributions in binary form must reproduce the above copyright |
| 30 | * notice, this list of conditions and the following disclaimer in |
| 31 | * the documentation and/or other materials provided with the |
| 32 | * distribution. |
| 33 | * - Neither the name of Intel Corporation nor the names of its |
| 34 | * contributors may be used to endorse or promote products derived |
| 35 | * from this software without specific prior written permission. |
| 36 | * |
| 37 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 38 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 39 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
| 40 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT |
| 41 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, |
| 42 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT |
| 43 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, |
| 44 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY |
| 45 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
| 46 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 47 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 48 | * |
| 49 | */ |
| 50 | #include <asm/page.h> |
| 51 | |
| 52 | #include "user_exp_rcv.h" |
| 53 | #include "trace.h" |
| 54 | |
Mitko Haralanov | b8abe34 | 2016-02-05 11:57:51 -0500 | [diff] [blame] | 55 | struct tid_group { |
| 56 | struct list_head list; |
| 57 | unsigned base; |
| 58 | u8 size; |
| 59 | u8 used; |
| 60 | u8 map; |
| 61 | }; |
| 62 | |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 63 | struct mmu_rb_node { |
| 64 | struct rb_node rbnode; |
| 65 | unsigned long virt; |
| 66 | unsigned long phys; |
| 67 | unsigned long len; |
| 68 | struct tid_group *grp; |
| 69 | u32 rcventry; |
| 70 | dma_addr_t dma_addr; |
| 71 | bool freed; |
| 72 | unsigned npages; |
| 73 | struct page *pages[0]; |
| 74 | }; |
| 75 | |
| 76 | enum mmu_call_types { |
| 77 | MMU_INVALIDATE_PAGE = 0, |
| 78 | MMU_INVALIDATE_RANGE = 1 |
| 79 | }; |
| 80 | |
| 81 | static const char * const mmu_types[] = { |
| 82 | "PAGE", |
| 83 | "RANGE" |
| 84 | }; |
| 85 | |
Mitko Haralanov | f88e0c8 | 2016-02-05 11:57:52 -0500 | [diff] [blame] | 86 | struct tid_pageset { |
| 87 | u16 idx; |
| 88 | u16 count; |
| 89 | }; |
| 90 | |
Mitko Haralanov | b8abe34 | 2016-02-05 11:57:51 -0500 | [diff] [blame] | 91 | #define EXP_TID_SET_EMPTY(set) (set.count == 0 && list_empty(&set.list)) |
| 92 | |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 93 | #define num_user_pages(vaddr, len) \ |
| 94 | (1 + (((((unsigned long)(vaddr) + \ |
| 95 | (unsigned long)(len) - 1) & PAGE_MASK) - \ |
| 96 | ((unsigned long)vaddr & PAGE_MASK)) >> PAGE_SHIFT)) |
| 97 | |
Mitko Haralanov | f88e0c8 | 2016-02-05 11:57:52 -0500 | [diff] [blame] | 98 | static void unlock_exp_tids(struct hfi1_ctxtdata *, struct exp_tid_set *, |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 99 | struct rb_root *); |
Mitko Haralanov | 7e7a436e | 2016-02-05 11:57:57 -0500 | [diff] [blame] | 100 | static u32 find_phys_blocks(struct page **, unsigned, struct tid_pageset *); |
Mitko Haralanov | f88e0c8 | 2016-02-05 11:57:52 -0500 | [diff] [blame] | 101 | static int set_rcvarray_entry(struct file *, unsigned long, u32, |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 102 | struct tid_group *, struct page **, unsigned); |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 103 | static inline int mmu_addr_cmp(struct mmu_rb_node *, unsigned long, |
| 104 | unsigned long); |
| 105 | static struct mmu_rb_node *mmu_rb_search_by_addr(struct rb_root *, |
Mitko Haralanov | b5eb3b2 | 2016-02-05 11:57:55 -0500 | [diff] [blame] | 106 | unsigned long); |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 107 | static inline struct mmu_rb_node *mmu_rb_search_by_entry(struct rb_root *, |
| 108 | u32); |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 109 | static int mmu_rb_insert_by_addr(struct rb_root *, struct mmu_rb_node *); |
| 110 | static int mmu_rb_insert_by_entry(struct rb_root *, struct mmu_rb_node *); |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 111 | static void mmu_notifier_mem_invalidate(struct mmu_notifier *, |
| 112 | unsigned long, unsigned long, |
| 113 | enum mmu_call_types); |
| 114 | static inline void mmu_notifier_page(struct mmu_notifier *, struct mm_struct *, |
| 115 | unsigned long); |
| 116 | static inline void mmu_notifier_range_start(struct mmu_notifier *, |
| 117 | struct mm_struct *, |
| 118 | unsigned long, unsigned long); |
Mitko Haralanov | f88e0c8 | 2016-02-05 11:57:52 -0500 | [diff] [blame] | 119 | static int program_rcvarray(struct file *, unsigned long, struct tid_group *, |
| 120 | struct tid_pageset *, unsigned, u16, struct page **, |
Mitko Haralanov | 7e7a436e | 2016-02-05 11:57:57 -0500 | [diff] [blame] | 121 | u32 *, unsigned *, unsigned *); |
Mitko Haralanov | 455d7f1 | 2016-02-05 11:57:56 -0500 | [diff] [blame] | 122 | static int unprogram_rcvarray(struct file *, u32, struct tid_group **); |
| 123 | static void clear_tid_node(struct hfi1_filedata *, u16, struct mmu_rb_node *); |
Mitko Haralanov | f88e0c8 | 2016-02-05 11:57:52 -0500 | [diff] [blame] | 124 | |
| 125 | static inline u32 rcventry2tidinfo(u32 rcventry) |
| 126 | { |
| 127 | u32 pair = rcventry & ~0x1; |
| 128 | |
| 129 | return EXP_TID_SET(IDX, pair >> 1) | |
| 130 | EXP_TID_SET(CTRL, 1 << (rcventry - pair)); |
| 131 | } |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 132 | |
Mitko Haralanov | b8abe34 | 2016-02-05 11:57:51 -0500 | [diff] [blame] | 133 | static inline void exp_tid_group_init(struct exp_tid_set *set) |
| 134 | { |
| 135 | INIT_LIST_HEAD(&set->list); |
| 136 | set->count = 0; |
| 137 | } |
| 138 | |
| 139 | static inline void tid_group_remove(struct tid_group *grp, |
| 140 | struct exp_tid_set *set) |
| 141 | { |
| 142 | list_del_init(&grp->list); |
| 143 | set->count--; |
| 144 | } |
| 145 | |
| 146 | static inline void tid_group_add_tail(struct tid_group *grp, |
| 147 | struct exp_tid_set *set) |
| 148 | { |
| 149 | list_add_tail(&grp->list, &set->list); |
| 150 | set->count++; |
| 151 | } |
| 152 | |
| 153 | static inline struct tid_group *tid_group_pop(struct exp_tid_set *set) |
| 154 | { |
| 155 | struct tid_group *grp = |
| 156 | list_first_entry(&set->list, struct tid_group, list); |
| 157 | list_del_init(&grp->list); |
| 158 | set->count--; |
| 159 | return grp; |
| 160 | } |
| 161 | |
| 162 | static inline void tid_group_move(struct tid_group *group, |
| 163 | struct exp_tid_set *s1, |
| 164 | struct exp_tid_set *s2) |
| 165 | { |
| 166 | tid_group_remove(group, s1); |
| 167 | tid_group_add_tail(group, s2); |
| 168 | } |
| 169 | |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 170 | static struct mmu_notifier_ops mn_opts = { |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 171 | .invalidate_page = mmu_notifier_page, |
| 172 | .invalidate_range_start = mmu_notifier_range_start, |
| 173 | }; |
| 174 | |
| 175 | /* |
| 176 | * Initialize context and file private data needed for Expected |
| 177 | * receive caching. This needs to be done after the context has |
| 178 | * been configured with the eager/expected RcvEntry counts. |
| 179 | */ |
| 180 | int hfi1_user_exp_rcv_init(struct file *fp) |
| 181 | { |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 182 | struct hfi1_filedata *fd = fp->private_data; |
| 183 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 184 | struct hfi1_devdata *dd = uctxt->dd; |
| 185 | unsigned tidbase; |
| 186 | int i, ret = 0; |
| 187 | |
| 188 | INIT_HLIST_NODE(&fd->mn.hlist); |
| 189 | spin_lock_init(&fd->rb_lock); |
| 190 | spin_lock_init(&fd->tid_lock); |
| 191 | spin_lock_init(&fd->invalid_lock); |
| 192 | fd->mn.ops = &mn_opts; |
| 193 | fd->tid_rb_root = RB_ROOT; |
| 194 | |
| 195 | if (!uctxt->subctxt_cnt || !fd->subctxt) { |
| 196 | exp_tid_group_init(&uctxt->tid_group_list); |
| 197 | exp_tid_group_init(&uctxt->tid_used_list); |
| 198 | exp_tid_group_init(&uctxt->tid_full_list); |
| 199 | |
| 200 | tidbase = uctxt->expected_base; |
| 201 | for (i = 0; i < uctxt->expected_count / |
| 202 | dd->rcv_entries.group_size; i++) { |
| 203 | struct tid_group *grp; |
| 204 | |
| 205 | grp = kzalloc(sizeof(*grp), GFP_KERNEL); |
| 206 | if (!grp) { |
| 207 | /* |
| 208 | * If we fail here, the groups already |
| 209 | * allocated will be freed by the close |
| 210 | * call. |
| 211 | */ |
| 212 | ret = -ENOMEM; |
| 213 | goto done; |
| 214 | } |
| 215 | grp->size = dd->rcv_entries.group_size; |
| 216 | grp->base = tidbase; |
| 217 | tid_group_add_tail(grp, &uctxt->tid_group_list); |
| 218 | tidbase += dd->rcv_entries.group_size; |
| 219 | } |
| 220 | } |
| 221 | |
| 222 | if (!HFI1_CAP_IS_USET(TID_UNMAP)) { |
| 223 | fd->invalid_tid_idx = 0; |
| 224 | fd->invalid_tids = kzalloc(uctxt->expected_count * |
| 225 | sizeof(u32), GFP_KERNEL); |
| 226 | if (!fd->invalid_tids) { |
| 227 | ret = -ENOMEM; |
| 228 | goto done; |
| 229 | } else { |
| 230 | /* |
| 231 | * Register MMU notifier callbacks. If the registration |
| 232 | * fails, continue but turn off the TID caching for |
| 233 | * all user contexts. |
| 234 | */ |
| 235 | ret = mmu_notifier_register(&fd->mn, current->mm); |
| 236 | if (ret) { |
| 237 | dd_dev_info(dd, |
| 238 | "Failed MMU notifier registration %d\n", |
| 239 | ret); |
| 240 | HFI1_CAP_USET(TID_UNMAP); |
| 241 | ret = 0; |
| 242 | } |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | if (HFI1_CAP_IS_USET(TID_UNMAP)) |
| 247 | fd->mmu_rb_insert = mmu_rb_insert_by_entry; |
| 248 | else |
| 249 | fd->mmu_rb_insert = mmu_rb_insert_by_addr; |
| 250 | |
| 251 | /* |
| 252 | * PSM does not have a good way to separate, count, and |
| 253 | * effectively enforce a limit on RcvArray entries used by |
| 254 | * subctxts (when context sharing is used) when TID caching |
| 255 | * is enabled. To help with that, we calculate a per-process |
| 256 | * RcvArray entry share and enforce that. |
| 257 | * If TID caching is not in use, PSM deals with usage on its |
| 258 | * own. In that case, we allow any subctxt to take all of the |
| 259 | * entries. |
| 260 | * |
| 261 | * Make sure that we set the tid counts only after successful |
| 262 | * init. |
| 263 | */ |
Mitko Haralanov | 455d7f1 | 2016-02-05 11:57:56 -0500 | [diff] [blame] | 264 | spin_lock(&fd->tid_lock); |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 265 | if (uctxt->subctxt_cnt && !HFI1_CAP_IS_USET(TID_UNMAP)) { |
| 266 | u16 remainder; |
| 267 | |
| 268 | fd->tid_limit = uctxt->expected_count / uctxt->subctxt_cnt; |
| 269 | remainder = uctxt->expected_count % uctxt->subctxt_cnt; |
| 270 | if (remainder && fd->subctxt < remainder) |
| 271 | fd->tid_limit++; |
| 272 | } else { |
| 273 | fd->tid_limit = uctxt->expected_count; |
| 274 | } |
Mitko Haralanov | 455d7f1 | 2016-02-05 11:57:56 -0500 | [diff] [blame] | 275 | spin_unlock(&fd->tid_lock); |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 276 | done: |
| 277 | return ret; |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 278 | } |
| 279 | |
| 280 | int hfi1_user_exp_rcv_free(struct hfi1_filedata *fd) |
| 281 | { |
Mitko Haralanov | 3abb33a | 2016-02-05 11:57:54 -0500 | [diff] [blame] | 282 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 283 | struct tid_group *grp, *gptr; |
| 284 | |
| 285 | /* |
| 286 | * The notifier would have been removed when the process'es mm |
| 287 | * was freed. |
| 288 | */ |
| 289 | if (current->mm && !HFI1_CAP_IS_USET(TID_UNMAP)) |
| 290 | mmu_notifier_unregister(&fd->mn, current->mm); |
| 291 | |
| 292 | kfree(fd->invalid_tids); |
| 293 | |
| 294 | if (!uctxt->cnt) { |
| 295 | if (!EXP_TID_SET_EMPTY(uctxt->tid_full_list)) |
| 296 | unlock_exp_tids(uctxt, &uctxt->tid_full_list, |
| 297 | &fd->tid_rb_root); |
| 298 | if (!EXP_TID_SET_EMPTY(uctxt->tid_used_list)) |
| 299 | unlock_exp_tids(uctxt, &uctxt->tid_used_list, |
| 300 | &fd->tid_rb_root); |
| 301 | list_for_each_entry_safe(grp, gptr, &uctxt->tid_group_list.list, |
| 302 | list) { |
| 303 | list_del_init(&grp->list); |
| 304 | kfree(grp); |
| 305 | } |
| 306 | spin_lock(&fd->rb_lock); |
| 307 | if (!RB_EMPTY_ROOT(&fd->tid_rb_root)) { |
| 308 | struct rb_node *node; |
| 309 | struct mmu_rb_node *rbnode; |
| 310 | |
| 311 | while ((node = rb_first(&fd->tid_rb_root))) { |
| 312 | rbnode = rb_entry(node, struct mmu_rb_node, |
| 313 | rbnode); |
| 314 | rb_erase(&rbnode->rbnode, &fd->tid_rb_root); |
| 315 | kfree(rbnode); |
| 316 | } |
| 317 | } |
| 318 | spin_unlock(&fd->rb_lock); |
| 319 | hfi1_clear_tids(uctxt); |
| 320 | } |
| 321 | return 0; |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 322 | } |
| 323 | |
Mitko Haralanov | b8abe34 | 2016-02-05 11:57:51 -0500 | [diff] [blame] | 324 | /* |
| 325 | * Write an "empty" RcvArray entry. |
| 326 | * This function exists so the TID registaration code can use it |
| 327 | * to write to unused/unneeded entries and still take advantage |
| 328 | * of the WC performance improvements. The HFI will ignore this |
| 329 | * write to the RcvArray entry. |
| 330 | */ |
| 331 | static inline void rcv_array_wc_fill(struct hfi1_devdata *dd, u32 index) |
| 332 | { |
| 333 | /* |
| 334 | * Doing the WC fill writes only makes sense if the device is |
| 335 | * present and the RcvArray has been mapped as WC memory. |
| 336 | */ |
| 337 | if ((dd->flags & HFI1_PRESENT) && dd->rcvarray_wc) |
| 338 | writeq(0, dd->rcvarray_wc + (index * 8)); |
| 339 | } |
| 340 | |
Mitko Haralanov | 7e7a436e | 2016-02-05 11:57:57 -0500 | [diff] [blame] | 341 | /* |
| 342 | * RcvArray entry allocation for Expected Receives is done by the |
| 343 | * following algorithm: |
| 344 | * |
| 345 | * The context keeps 3 lists of groups of RcvArray entries: |
| 346 | * 1. List of empty groups - tid_group_list |
| 347 | * This list is created during user context creation and |
| 348 | * contains elements which describe sets (of 8) of empty |
| 349 | * RcvArray entries. |
| 350 | * 2. List of partially used groups - tid_used_list |
| 351 | * This list contains sets of RcvArray entries which are |
| 352 | * not completely used up. Another mapping request could |
| 353 | * use some of all of the remaining entries. |
| 354 | * 3. List of full groups - tid_full_list |
| 355 | * This is the list where sets that are completely used |
| 356 | * up go. |
| 357 | * |
| 358 | * An attempt to optimize the usage of RcvArray entries is |
| 359 | * made by finding all sets of physically contiguous pages in a |
| 360 | * user's buffer. |
| 361 | * These physically contiguous sets are further split into |
| 362 | * sizes supported by the receive engine of the HFI. The |
| 363 | * resulting sets of pages are stored in struct tid_pageset, |
| 364 | * which describes the sets as: |
| 365 | * * .count - number of pages in this set |
| 366 | * * .idx - starting index into struct page ** array |
| 367 | * of this set |
| 368 | * |
| 369 | * From this point on, the algorithm deals with the page sets |
| 370 | * described above. The number of pagesets is divided by the |
| 371 | * RcvArray group size to produce the number of full groups |
| 372 | * needed. |
| 373 | * |
| 374 | * Groups from the 3 lists are manipulated using the following |
| 375 | * rules: |
| 376 | * 1. For each set of 8 pagesets, a complete group from |
| 377 | * tid_group_list is taken, programmed, and moved to |
| 378 | * the tid_full_list list. |
| 379 | * 2. For all remaining pagesets: |
| 380 | * 2.1 If the tid_used_list is empty and the tid_group_list |
| 381 | * is empty, stop processing pageset and return only |
| 382 | * what has been programmed up to this point. |
| 383 | * 2.2 If the tid_used_list is empty and the tid_group_list |
| 384 | * is not empty, move a group from tid_group_list to |
| 385 | * tid_used_list. |
| 386 | * 2.3 For each group is tid_used_group, program as much as |
| 387 | * can fit into the group. If the group becomes fully |
| 388 | * used, move it to tid_full_list. |
| 389 | */ |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 390 | int hfi1_user_exp_rcv_setup(struct file *fp, struct hfi1_tid_info *tinfo) |
| 391 | { |
Mitko Haralanov | 7e7a436e | 2016-02-05 11:57:57 -0500 | [diff] [blame] | 392 | int ret = 0, need_group = 0, pinned; |
| 393 | struct hfi1_filedata *fd = fp->private_data; |
| 394 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 395 | struct hfi1_devdata *dd = uctxt->dd; |
| 396 | unsigned npages, ngroups, pageidx = 0, pageset_count, npagesets, |
| 397 | tididx = 0, mapped, mapped_pages = 0; |
| 398 | unsigned long vaddr = tinfo->vaddr; |
| 399 | struct page **pages = NULL; |
| 400 | u32 *tidlist = NULL; |
| 401 | struct tid_pageset *pagesets = NULL; |
| 402 | |
| 403 | /* Get the number of pages the user buffer spans */ |
| 404 | npages = num_user_pages(vaddr, tinfo->length); |
| 405 | if (!npages) |
| 406 | return -EINVAL; |
| 407 | |
| 408 | if (npages > uctxt->expected_count) { |
| 409 | dd_dev_err(dd, "Expected buffer too big\n"); |
| 410 | return -EINVAL; |
| 411 | } |
| 412 | |
| 413 | /* Verify that access is OK for the user buffer */ |
| 414 | if (!access_ok(VERIFY_WRITE, (void __user *)vaddr, |
| 415 | npages * PAGE_SIZE)) { |
| 416 | dd_dev_err(dd, "Fail vaddr %p, %u pages, !access_ok\n", |
| 417 | (void *)vaddr, npages); |
| 418 | return -EFAULT; |
| 419 | } |
| 420 | |
| 421 | pagesets = kcalloc(uctxt->expected_count, sizeof(*pagesets), |
| 422 | GFP_KERNEL); |
| 423 | if (!pagesets) |
| 424 | return -ENOMEM; |
| 425 | |
| 426 | /* Allocate the array of struct page pointers needed for pinning */ |
| 427 | pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL); |
| 428 | if (!pages) { |
| 429 | ret = -ENOMEM; |
| 430 | goto bail; |
| 431 | } |
| 432 | |
| 433 | /* |
| 434 | * Pin all the pages of the user buffer. If we can't pin all the |
| 435 | * pages, accept the amount pinned so far and program only that. |
| 436 | * User space knows how to deal with partially programmed buffers. |
| 437 | */ |
| 438 | pinned = hfi1_acquire_user_pages(vaddr, npages, true, pages); |
| 439 | if (pinned <= 0) { |
| 440 | ret = pinned; |
| 441 | goto bail; |
| 442 | } |
| 443 | |
| 444 | /* Find sets of physically contiguous pages */ |
| 445 | npagesets = find_phys_blocks(pages, pinned, pagesets); |
| 446 | |
| 447 | /* |
| 448 | * We don't need to access this under a lock since tid_used is per |
| 449 | * process and the same process cannot be in hfi1_user_exp_rcv_clear() |
| 450 | * and hfi1_user_exp_rcv_setup() at the same time. |
| 451 | */ |
| 452 | spin_lock(&fd->tid_lock); |
| 453 | if (fd->tid_used + npagesets > fd->tid_limit) |
| 454 | pageset_count = fd->tid_limit - fd->tid_used; |
| 455 | else |
| 456 | pageset_count = npagesets; |
| 457 | spin_unlock(&fd->tid_lock); |
| 458 | |
| 459 | if (!pageset_count) |
| 460 | goto bail; |
| 461 | |
| 462 | ngroups = pageset_count / dd->rcv_entries.group_size; |
| 463 | tidlist = kcalloc(pageset_count, sizeof(*tidlist), GFP_KERNEL); |
| 464 | if (!tidlist) { |
| 465 | ret = -ENOMEM; |
| 466 | goto nomem; |
| 467 | } |
| 468 | |
| 469 | tididx = 0; |
| 470 | |
| 471 | /* |
| 472 | * From this point on, we are going to be using shared (between master |
| 473 | * and subcontexts) context resources. We need to take the lock. |
| 474 | */ |
| 475 | mutex_lock(&uctxt->exp_lock); |
| 476 | /* |
| 477 | * The first step is to program the RcvArray entries which are complete |
| 478 | * groups. |
| 479 | */ |
| 480 | while (ngroups && uctxt->tid_group_list.count) { |
| 481 | struct tid_group *grp = |
| 482 | tid_group_pop(&uctxt->tid_group_list); |
| 483 | |
| 484 | ret = program_rcvarray(fp, vaddr, grp, pagesets, |
| 485 | pageidx, dd->rcv_entries.group_size, |
| 486 | pages, tidlist, &tididx, &mapped); |
| 487 | /* |
| 488 | * If there was a failure to program the RcvArray |
| 489 | * entries for the entire group, reset the grp fields |
| 490 | * and add the grp back to the free group list. |
| 491 | */ |
| 492 | if (ret <= 0) { |
| 493 | tid_group_add_tail(grp, &uctxt->tid_group_list); |
| 494 | hfi1_cdbg(TID, |
| 495 | "Failed to program RcvArray group %d", ret); |
| 496 | goto unlock; |
| 497 | } |
| 498 | |
| 499 | tid_group_add_tail(grp, &uctxt->tid_full_list); |
| 500 | ngroups--; |
| 501 | pageidx += ret; |
| 502 | mapped_pages += mapped; |
| 503 | } |
| 504 | |
| 505 | while (pageidx < pageset_count) { |
| 506 | struct tid_group *grp, *ptr; |
| 507 | /* |
| 508 | * If we don't have any partially used tid groups, check |
| 509 | * if we have empty groups. If so, take one from there and |
| 510 | * put in the partially used list. |
| 511 | */ |
| 512 | if (!uctxt->tid_used_list.count || need_group) { |
| 513 | if (!uctxt->tid_group_list.count) |
| 514 | goto unlock; |
| 515 | |
| 516 | grp = tid_group_pop(&uctxt->tid_group_list); |
| 517 | tid_group_add_tail(grp, &uctxt->tid_used_list); |
| 518 | need_group = 0; |
| 519 | } |
| 520 | /* |
| 521 | * There is an optimization opportunity here - instead of |
| 522 | * fitting as many page sets as we can, check for a group |
| 523 | * later on in the list that could fit all of them. |
| 524 | */ |
| 525 | list_for_each_entry_safe(grp, ptr, &uctxt->tid_used_list.list, |
| 526 | list) { |
| 527 | unsigned use = min_t(unsigned, pageset_count - pageidx, |
| 528 | grp->size - grp->used); |
| 529 | |
| 530 | ret = program_rcvarray(fp, vaddr, grp, pagesets, |
| 531 | pageidx, use, pages, tidlist, |
| 532 | &tididx, &mapped); |
| 533 | if (ret < 0) { |
| 534 | hfi1_cdbg(TID, |
| 535 | "Failed to program RcvArray entries %d", |
| 536 | ret); |
| 537 | ret = -EFAULT; |
| 538 | goto unlock; |
| 539 | } else if (ret > 0) { |
| 540 | if (grp->used == grp->size) |
| 541 | tid_group_move(grp, |
| 542 | &uctxt->tid_used_list, |
| 543 | &uctxt->tid_full_list); |
| 544 | pageidx += ret; |
| 545 | mapped_pages += mapped; |
| 546 | need_group = 0; |
| 547 | /* Check if we are done so we break out early */ |
| 548 | if (pageidx >= pageset_count) |
| 549 | break; |
| 550 | } else if (WARN_ON(ret == 0)) { |
| 551 | /* |
| 552 | * If ret is 0, we did not program any entries |
| 553 | * into this group, which can only happen if |
| 554 | * we've screwed up the accounting somewhere. |
| 555 | * Warn and try to continue. |
| 556 | */ |
| 557 | need_group = 1; |
| 558 | } |
| 559 | } |
| 560 | } |
| 561 | unlock: |
| 562 | mutex_unlock(&uctxt->exp_lock); |
| 563 | nomem: |
| 564 | hfi1_cdbg(TID, "total mapped: tidpairs:%u pages:%u (%d)", tididx, |
| 565 | mapped_pages, ret); |
| 566 | if (tididx) { |
| 567 | spin_lock(&fd->tid_lock); |
| 568 | fd->tid_used += tididx; |
| 569 | spin_unlock(&fd->tid_lock); |
| 570 | tinfo->tidcnt = tididx; |
| 571 | tinfo->length = mapped_pages * PAGE_SIZE; |
| 572 | |
| 573 | if (copy_to_user((void __user *)(unsigned long)tinfo->tidlist, |
| 574 | tidlist, sizeof(tidlist[0]) * tididx)) { |
| 575 | /* |
| 576 | * On failure to copy to the user level, we need to undo |
| 577 | * everything done so far so we don't leak resources. |
| 578 | */ |
| 579 | tinfo->tidlist = (unsigned long)&tidlist; |
| 580 | hfi1_user_exp_rcv_clear(fp, tinfo); |
| 581 | tinfo->tidlist = 0; |
| 582 | ret = -EFAULT; |
| 583 | goto bail; |
| 584 | } |
| 585 | } |
| 586 | |
| 587 | /* |
| 588 | * If not everything was mapped (due to insufficient RcvArray entries, |
| 589 | * for example), unpin all unmapped pages so we can pin them nex time. |
| 590 | */ |
| 591 | if (mapped_pages != pinned) |
| 592 | hfi1_release_user_pages(&pages[mapped_pages], |
| 593 | pinned - mapped_pages, |
| 594 | false); |
| 595 | bail: |
| 596 | kfree(pagesets); |
| 597 | kfree(pages); |
| 598 | kfree(tidlist); |
| 599 | return ret > 0 ? 0 : ret; |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 600 | } |
| 601 | |
| 602 | int hfi1_user_exp_rcv_clear(struct file *fp, struct hfi1_tid_info *tinfo) |
| 603 | { |
Mitko Haralanov | 455d7f1 | 2016-02-05 11:57:56 -0500 | [diff] [blame] | 604 | int ret = 0; |
| 605 | struct hfi1_filedata *fd = fp->private_data; |
| 606 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 607 | u32 *tidinfo; |
| 608 | unsigned tididx; |
| 609 | |
| 610 | tidinfo = kcalloc(tinfo->tidcnt, sizeof(*tidinfo), GFP_KERNEL); |
| 611 | if (!tidinfo) |
| 612 | return -ENOMEM; |
| 613 | |
| 614 | if (copy_from_user(tidinfo, (void __user *)(unsigned long) |
| 615 | tinfo->tidlist, sizeof(tidinfo[0]) * |
| 616 | tinfo->tidcnt)) { |
| 617 | ret = -EFAULT; |
| 618 | goto done; |
| 619 | } |
| 620 | |
| 621 | mutex_lock(&uctxt->exp_lock); |
| 622 | for (tididx = 0; tididx < tinfo->tidcnt; tididx++) { |
| 623 | ret = unprogram_rcvarray(fp, tidinfo[tididx], NULL); |
| 624 | if (ret) { |
| 625 | hfi1_cdbg(TID, "Failed to unprogram rcv array %d", |
| 626 | ret); |
| 627 | break; |
| 628 | } |
| 629 | } |
| 630 | spin_lock(&fd->tid_lock); |
| 631 | fd->tid_used -= tididx; |
| 632 | spin_unlock(&fd->tid_lock); |
| 633 | tinfo->tidcnt = tididx; |
| 634 | mutex_unlock(&uctxt->exp_lock); |
| 635 | done: |
| 636 | kfree(tidinfo); |
| 637 | return ret; |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 638 | } |
| 639 | |
| 640 | int hfi1_user_exp_rcv_invalid(struct file *fp, struct hfi1_tid_info *tinfo) |
| 641 | { |
Mitko Haralanov | 455d7f1 | 2016-02-05 11:57:56 -0500 | [diff] [blame] | 642 | struct hfi1_filedata *fd = fp->private_data; |
| 643 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 644 | unsigned long *ev = uctxt->dd->events + |
| 645 | (((uctxt->ctxt - uctxt->dd->first_user_ctxt) * |
| 646 | HFI1_MAX_SHARED_CTXTS) + fd->subctxt); |
| 647 | u32 *array; |
| 648 | int ret = 0; |
| 649 | |
| 650 | if (!fd->invalid_tids) |
| 651 | return -EINVAL; |
| 652 | |
| 653 | /* |
| 654 | * copy_to_user() can sleep, which will leave the invalid_lock |
| 655 | * locked and cause the MMU notifier to be blocked on the lock |
| 656 | * for a long time. |
| 657 | * Copy the data to a local buffer so we can release the lock. |
| 658 | */ |
| 659 | array = kcalloc(uctxt->expected_count, sizeof(*array), GFP_KERNEL); |
| 660 | if (!array) |
| 661 | return -EFAULT; |
| 662 | |
| 663 | spin_lock(&fd->invalid_lock); |
| 664 | if (fd->invalid_tid_idx) { |
| 665 | memcpy(array, fd->invalid_tids, sizeof(*array) * |
| 666 | fd->invalid_tid_idx); |
| 667 | memset(fd->invalid_tids, 0, sizeof(*fd->invalid_tids) * |
| 668 | fd->invalid_tid_idx); |
| 669 | tinfo->tidcnt = fd->invalid_tid_idx; |
| 670 | fd->invalid_tid_idx = 0; |
| 671 | /* |
| 672 | * Reset the user flag while still holding the lock. |
| 673 | * Otherwise, PSM can miss events. |
| 674 | */ |
| 675 | clear_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev); |
| 676 | } else { |
| 677 | tinfo->tidcnt = 0; |
| 678 | } |
| 679 | spin_unlock(&fd->invalid_lock); |
| 680 | |
| 681 | if (tinfo->tidcnt) { |
| 682 | if (copy_to_user((void __user *)tinfo->tidlist, |
| 683 | array, sizeof(*array) * tinfo->tidcnt)) |
| 684 | ret = -EFAULT; |
| 685 | } |
| 686 | kfree(array); |
| 687 | |
| 688 | return ret; |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 689 | } |
| 690 | |
Mitko Haralanov | f88e0c8 | 2016-02-05 11:57:52 -0500 | [diff] [blame] | 691 | static u32 find_phys_blocks(struct page **pages, unsigned npages, |
| 692 | struct tid_pageset *list) |
| 693 | { |
| 694 | unsigned pagecount, pageidx, setcount = 0, i; |
| 695 | unsigned long pfn, this_pfn; |
| 696 | |
| 697 | if (!npages) |
| 698 | return 0; |
| 699 | |
| 700 | /* |
| 701 | * Look for sets of physically contiguous pages in the user buffer. |
| 702 | * This will allow us to optimize Expected RcvArray entry usage by |
| 703 | * using the bigger supported sizes. |
| 704 | */ |
| 705 | pfn = page_to_pfn(pages[0]); |
| 706 | for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) { |
| 707 | this_pfn = i < npages ? page_to_pfn(pages[i]) : 0; |
| 708 | |
| 709 | /* |
| 710 | * If the pfn's are not sequential, pages are not physically |
| 711 | * contiguous. |
| 712 | */ |
| 713 | if (this_pfn != ++pfn) { |
| 714 | /* |
| 715 | * At this point we have to loop over the set of |
| 716 | * physically contiguous pages and break them down it |
| 717 | * sizes supported by the HW. |
| 718 | * There are two main constraints: |
| 719 | * 1. The max buffer size is MAX_EXPECTED_BUFFER. |
| 720 | * If the total set size is bigger than that |
| 721 | * program only a MAX_EXPECTED_BUFFER chunk. |
| 722 | * 2. The buffer size has to be a power of two. If |
| 723 | * it is not, round down to the closes power of |
| 724 | * 2 and program that size. |
| 725 | */ |
| 726 | while (pagecount) { |
| 727 | int maxpages = pagecount; |
| 728 | u32 bufsize = pagecount * PAGE_SIZE; |
| 729 | |
| 730 | if (bufsize > MAX_EXPECTED_BUFFER) |
| 731 | maxpages = |
| 732 | MAX_EXPECTED_BUFFER >> |
| 733 | PAGE_SHIFT; |
| 734 | else if (!is_power_of_2(bufsize)) |
| 735 | maxpages = |
| 736 | rounddown_pow_of_two(bufsize) >> |
| 737 | PAGE_SHIFT; |
| 738 | |
| 739 | list[setcount].idx = pageidx; |
| 740 | list[setcount].count = maxpages; |
| 741 | pagecount -= maxpages; |
| 742 | pageidx += maxpages; |
| 743 | setcount++; |
| 744 | } |
| 745 | pageidx = i; |
| 746 | pagecount = 1; |
| 747 | pfn = this_pfn; |
| 748 | } else { |
| 749 | pagecount++; |
| 750 | } |
| 751 | } |
| 752 | return setcount; |
| 753 | } |
| 754 | |
| 755 | /** |
| 756 | * program_rcvarray() - program an RcvArray group with receive buffers |
| 757 | * @fp: file pointer |
| 758 | * @vaddr: starting user virtual address |
| 759 | * @grp: RcvArray group |
| 760 | * @sets: array of struct tid_pageset holding information on physically |
| 761 | * contiguous chunks from the user buffer |
| 762 | * @start: starting index into sets array |
| 763 | * @count: number of struct tid_pageset's to program |
| 764 | * @pages: an array of struct page * for the user buffer |
| 765 | * @tidlist: the array of u32 elements when the information about the |
| 766 | * programmed RcvArray entries is to be encoded. |
| 767 | * @tididx: starting offset into tidlist |
| 768 | * @pmapped: (output parameter) number of pages programmed into the RcvArray |
| 769 | * entries. |
| 770 | * |
| 771 | * This function will program up to 'count' number of RcvArray entries from the |
| 772 | * group 'grp'. To make best use of write-combining writes, the function will |
| 773 | * perform writes to the unused RcvArray entries which will be ignored by the |
| 774 | * HW. Each RcvArray entry will be programmed with a physically contiguous |
| 775 | * buffer chunk from the user's virtual buffer. |
| 776 | * |
| 777 | * Return: |
| 778 | * -EINVAL if the requested count is larger than the size of the group, |
| 779 | * -ENOMEM or -EFAULT on error from set_rcvarray_entry(), or |
| 780 | * number of RcvArray entries programmed. |
| 781 | */ |
| 782 | static int program_rcvarray(struct file *fp, unsigned long vaddr, |
| 783 | struct tid_group *grp, |
| 784 | struct tid_pageset *sets, |
| 785 | unsigned start, u16 count, struct page **pages, |
| 786 | u32 *tidlist, unsigned *tididx, unsigned *pmapped) |
| 787 | { |
| 788 | struct hfi1_filedata *fd = fp->private_data; |
| 789 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 790 | struct hfi1_devdata *dd = uctxt->dd; |
| 791 | u16 idx; |
| 792 | u32 tidinfo = 0, rcventry, useidx = 0; |
| 793 | int mapped = 0; |
| 794 | |
| 795 | /* Count should never be larger than the group size */ |
| 796 | if (count > grp->size) |
| 797 | return -EINVAL; |
| 798 | |
| 799 | /* Find the first unused entry in the group */ |
| 800 | for (idx = 0; idx < grp->size; idx++) { |
| 801 | if (!(grp->map & (1 << idx))) { |
| 802 | useidx = idx; |
| 803 | break; |
| 804 | } |
| 805 | rcv_array_wc_fill(dd, grp->base + idx); |
| 806 | } |
| 807 | |
| 808 | idx = 0; |
| 809 | while (idx < count) { |
| 810 | u16 npages, pageidx, setidx = start + idx; |
| 811 | int ret = 0; |
| 812 | |
| 813 | /* |
| 814 | * If this entry in the group is used, move to the next one. |
| 815 | * If we go past the end of the group, exit the loop. |
| 816 | */ |
| 817 | if (useidx >= grp->size) { |
| 818 | break; |
| 819 | } else if (grp->map & (1 << useidx)) { |
| 820 | rcv_array_wc_fill(dd, grp->base + useidx); |
| 821 | useidx++; |
| 822 | continue; |
| 823 | } |
| 824 | |
| 825 | rcventry = grp->base + useidx; |
| 826 | npages = sets[setidx].count; |
| 827 | pageidx = sets[setidx].idx; |
| 828 | |
| 829 | ret = set_rcvarray_entry(fp, vaddr + (pageidx * PAGE_SIZE), |
| 830 | rcventry, grp, pages + pageidx, |
| 831 | npages); |
| 832 | if (ret) |
| 833 | return ret; |
| 834 | mapped += npages; |
| 835 | |
| 836 | tidinfo = rcventry2tidinfo(rcventry - uctxt->expected_base) | |
| 837 | EXP_TID_SET(LEN, npages); |
| 838 | tidlist[(*tididx)++] = tidinfo; |
| 839 | grp->used++; |
| 840 | grp->map |= 1 << useidx++; |
| 841 | idx++; |
| 842 | } |
| 843 | |
| 844 | /* Fill the rest of the group with "blank" writes */ |
| 845 | for (; useidx < grp->size; useidx++) |
| 846 | rcv_array_wc_fill(dd, grp->base + useidx); |
| 847 | *pmapped = mapped; |
| 848 | return idx; |
| 849 | } |
| 850 | |
| 851 | static int set_rcvarray_entry(struct file *fp, unsigned long vaddr, |
| 852 | u32 rcventry, struct tid_group *grp, |
| 853 | struct page **pages, unsigned npages) |
| 854 | { |
| 855 | int ret; |
| 856 | struct hfi1_filedata *fd = fp->private_data; |
| 857 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 858 | struct mmu_rb_node *node; |
| 859 | struct hfi1_devdata *dd = uctxt->dd; |
| 860 | struct rb_root *root = &fd->tid_rb_root; |
| 861 | dma_addr_t phys; |
| 862 | |
| 863 | /* |
| 864 | * Allocate the node first so we can handle a potential |
| 865 | * failure before we've programmed anything. |
| 866 | */ |
| 867 | node = kzalloc(sizeof(*node) + (sizeof(struct page *) * npages), |
| 868 | GFP_KERNEL); |
| 869 | if (!node) |
| 870 | return -ENOMEM; |
| 871 | |
| 872 | phys = pci_map_single(dd->pcidev, |
| 873 | __va(page_to_phys(pages[0])), |
| 874 | npages * PAGE_SIZE, PCI_DMA_FROMDEVICE); |
| 875 | if (dma_mapping_error(&dd->pcidev->dev, phys)) { |
| 876 | dd_dev_err(dd, "Failed to DMA map Exp Rcv pages 0x%llx\n", |
| 877 | phys); |
| 878 | kfree(node); |
| 879 | return -EFAULT; |
| 880 | } |
| 881 | |
| 882 | node->virt = vaddr; |
| 883 | node->phys = page_to_phys(pages[0]); |
| 884 | node->len = npages * PAGE_SIZE; |
| 885 | node->npages = npages; |
| 886 | node->rcventry = rcventry; |
| 887 | node->dma_addr = phys; |
| 888 | node->grp = grp; |
| 889 | node->freed = false; |
| 890 | memcpy(node->pages, pages, sizeof(struct page *) * npages); |
| 891 | |
| 892 | spin_lock(&fd->rb_lock); |
| 893 | ret = fd->mmu_rb_insert(root, node); |
| 894 | spin_unlock(&fd->rb_lock); |
| 895 | |
| 896 | if (ret) { |
| 897 | hfi1_cdbg(TID, "Failed to insert RB node %u 0x%lx, 0x%lx %d", |
| 898 | node->rcventry, node->virt, node->phys, ret); |
| 899 | pci_unmap_single(dd->pcidev, phys, npages * PAGE_SIZE, |
| 900 | PCI_DMA_FROMDEVICE); |
| 901 | kfree(node); |
| 902 | return -EFAULT; |
| 903 | } |
| 904 | hfi1_put_tid(dd, rcventry, PT_EXPECTED, phys, ilog2(npages) + 1); |
Mitko Haralanov | 0b091fb | 2016-02-05 11:57:58 -0500 | [diff] [blame] | 905 | trace_hfi1_exp_tid_reg(uctxt->ctxt, fd->subctxt, rcventry, |
| 906 | npages, node->virt, node->phys, phys); |
Mitko Haralanov | f88e0c8 | 2016-02-05 11:57:52 -0500 | [diff] [blame] | 907 | return 0; |
| 908 | } |
| 909 | |
| 910 | static int unprogram_rcvarray(struct file *fp, u32 tidinfo, |
| 911 | struct tid_group **grp) |
| 912 | { |
| 913 | struct hfi1_filedata *fd = fp->private_data; |
| 914 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 915 | struct hfi1_devdata *dd = uctxt->dd; |
| 916 | struct mmu_rb_node *node; |
| 917 | u8 tidctrl = EXP_TID_GET(tidinfo, CTRL); |
| 918 | u32 tidbase = uctxt->expected_base, |
| 919 | tididx = EXP_TID_GET(tidinfo, IDX) << 1, rcventry; |
| 920 | |
| 921 | if (tididx >= uctxt->expected_count) { |
| 922 | dd_dev_err(dd, "Invalid RcvArray entry (%u) index for ctxt %u\n", |
| 923 | tididx, uctxt->ctxt); |
| 924 | return -EINVAL; |
| 925 | } |
| 926 | |
| 927 | if (tidctrl == 0x3) |
| 928 | return -EINVAL; |
| 929 | |
| 930 | rcventry = tidbase + tididx + (tidctrl - 1); |
| 931 | |
| 932 | spin_lock(&fd->rb_lock); |
| 933 | node = mmu_rb_search_by_entry(&fd->tid_rb_root, rcventry); |
| 934 | if (!node) { |
| 935 | spin_unlock(&fd->rb_lock); |
| 936 | return -EBADF; |
| 937 | } |
| 938 | rb_erase(&node->rbnode, &fd->tid_rb_root); |
| 939 | spin_unlock(&fd->rb_lock); |
| 940 | if (grp) |
| 941 | *grp = node->grp; |
| 942 | clear_tid_node(fd, fd->subctxt, node); |
| 943 | return 0; |
| 944 | } |
| 945 | |
| 946 | static void clear_tid_node(struct hfi1_filedata *fd, u16 subctxt, |
| 947 | struct mmu_rb_node *node) |
| 948 | { |
| 949 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 950 | struct hfi1_devdata *dd = uctxt->dd; |
| 951 | |
Mitko Haralanov | 0b091fb | 2016-02-05 11:57:58 -0500 | [diff] [blame] | 952 | trace_hfi1_exp_tid_unreg(uctxt->ctxt, fd->subctxt, node->rcventry, |
| 953 | node->npages, node->virt, node->phys, |
| 954 | node->dma_addr); |
| 955 | |
Mitko Haralanov | f88e0c8 | 2016-02-05 11:57:52 -0500 | [diff] [blame] | 956 | hfi1_put_tid(dd, node->rcventry, PT_INVALID, 0, 0); |
| 957 | /* |
| 958 | * Make sure device has seen the write before we unpin the |
| 959 | * pages. |
| 960 | */ |
| 961 | flush_wc(); |
| 962 | |
| 963 | pci_unmap_single(dd->pcidev, node->dma_addr, node->len, |
| 964 | PCI_DMA_FROMDEVICE); |
| 965 | hfi1_release_user_pages(node->pages, node->npages, true); |
| 966 | |
| 967 | node->grp->used--; |
| 968 | node->grp->map &= ~(1 << (node->rcventry - node->grp->base)); |
| 969 | |
| 970 | if (node->grp->used == node->grp->size - 1) |
| 971 | tid_group_move(node->grp, &uctxt->tid_full_list, |
| 972 | &uctxt->tid_used_list); |
| 973 | else if (!node->grp->used) |
| 974 | tid_group_move(node->grp, &uctxt->tid_used_list, |
| 975 | &uctxt->tid_group_list); |
| 976 | kfree(node); |
| 977 | } |
| 978 | |
| 979 | static void unlock_exp_tids(struct hfi1_ctxtdata *uctxt, |
| 980 | struct exp_tid_set *set, struct rb_root *root) |
| 981 | { |
| 982 | struct tid_group *grp, *ptr; |
| 983 | struct hfi1_filedata *fd = container_of(root, struct hfi1_filedata, |
| 984 | tid_rb_root); |
| 985 | int i; |
| 986 | |
| 987 | list_for_each_entry_safe(grp, ptr, &set->list, list) { |
| 988 | list_del_init(&grp->list); |
| 989 | |
| 990 | spin_lock(&fd->rb_lock); |
| 991 | for (i = 0; i < grp->size; i++) { |
| 992 | if (grp->map & (1 << i)) { |
| 993 | u16 rcventry = grp->base + i; |
| 994 | struct mmu_rb_node *node; |
| 995 | |
| 996 | node = mmu_rb_search_by_entry(root, rcventry); |
| 997 | if (!node) |
| 998 | continue; |
| 999 | rb_erase(&node->rbnode, root); |
| 1000 | clear_tid_node(fd, -1, node); |
| 1001 | } |
| 1002 | } |
| 1003 | spin_unlock(&fd->rb_lock); |
| 1004 | } |
| 1005 | } |
| 1006 | |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 1007 | static inline void mmu_notifier_page(struct mmu_notifier *mn, |
| 1008 | struct mm_struct *mm, unsigned long addr) |
| 1009 | { |
| 1010 | mmu_notifier_mem_invalidate(mn, addr, addr + PAGE_SIZE, |
| 1011 | MMU_INVALIDATE_PAGE); |
| 1012 | } |
| 1013 | |
| 1014 | static inline void mmu_notifier_range_start(struct mmu_notifier *mn, |
| 1015 | struct mm_struct *mm, |
| 1016 | unsigned long start, |
| 1017 | unsigned long end) |
| 1018 | { |
| 1019 | mmu_notifier_mem_invalidate(mn, start, end, MMU_INVALIDATE_RANGE); |
| 1020 | } |
| 1021 | |
| 1022 | static void mmu_notifier_mem_invalidate(struct mmu_notifier *mn, |
| 1023 | unsigned long start, unsigned long end, |
| 1024 | enum mmu_call_types type) |
| 1025 | { |
Mitko Haralanov | b5eb3b2 | 2016-02-05 11:57:55 -0500 | [diff] [blame] | 1026 | struct hfi1_filedata *fd = container_of(mn, struct hfi1_filedata, mn); |
| 1027 | struct hfi1_ctxtdata *uctxt = fd->uctxt; |
| 1028 | struct rb_root *root = &fd->tid_rb_root; |
| 1029 | struct mmu_rb_node *node; |
| 1030 | unsigned long addr = start; |
| 1031 | |
Mitko Haralanov | 0b091fb | 2016-02-05 11:57:58 -0500 | [diff] [blame] | 1032 | trace_hfi1_mmu_invalidate(uctxt->ctxt, fd->subctxt, mmu_types[type], |
| 1033 | start, end); |
| 1034 | |
Mitko Haralanov | b5eb3b2 | 2016-02-05 11:57:55 -0500 | [diff] [blame] | 1035 | spin_lock(&fd->rb_lock); |
| 1036 | while (addr < end) { |
| 1037 | node = mmu_rb_search_by_addr(root, addr); |
| 1038 | |
| 1039 | if (!node) { |
| 1040 | /* |
| 1041 | * Didn't find a node at this address. However, the |
| 1042 | * range could be bigger than what we have registered |
| 1043 | * so we have to keep looking. |
| 1044 | */ |
| 1045 | addr += PAGE_SIZE; |
| 1046 | continue; |
| 1047 | } |
| 1048 | |
| 1049 | /* |
| 1050 | * The next address to be looked up is computed based |
| 1051 | * on the node's starting address. This is due to the |
| 1052 | * fact that the range where we start might be in the |
| 1053 | * middle of the node's buffer so simply incrementing |
| 1054 | * the address by the node's size would result is a |
| 1055 | * bad address. |
| 1056 | */ |
| 1057 | addr = node->virt + (node->npages * PAGE_SIZE); |
| 1058 | if (node->freed) |
| 1059 | continue; |
| 1060 | |
Mitko Haralanov | 0b091fb | 2016-02-05 11:57:58 -0500 | [diff] [blame] | 1061 | trace_hfi1_exp_tid_inval(uctxt->ctxt, fd->subctxt, node->virt, |
| 1062 | node->rcventry, node->npages, |
| 1063 | node->dma_addr); |
Mitko Haralanov | b5eb3b2 | 2016-02-05 11:57:55 -0500 | [diff] [blame] | 1064 | node->freed = true; |
| 1065 | |
| 1066 | spin_lock(&fd->invalid_lock); |
| 1067 | if (fd->invalid_tid_idx < uctxt->expected_count) { |
| 1068 | fd->invalid_tids[fd->invalid_tid_idx] = |
| 1069 | rcventry2tidinfo(node->rcventry - |
| 1070 | uctxt->expected_base); |
| 1071 | fd->invalid_tids[fd->invalid_tid_idx] |= |
| 1072 | EXP_TID_SET(LEN, node->npages); |
| 1073 | if (!fd->invalid_tid_idx) { |
| 1074 | unsigned long *ev; |
| 1075 | |
| 1076 | /* |
| 1077 | * hfi1_set_uevent_bits() sets a user event flag |
| 1078 | * for all processes. Because calling into the |
| 1079 | * driver to process TID cache invalidations is |
| 1080 | * expensive and TID cache invalidations are |
| 1081 | * handled on a per-process basis, we can |
| 1082 | * optimize this to set the flag only for the |
| 1083 | * process in question. |
| 1084 | */ |
| 1085 | ev = uctxt->dd->events + |
| 1086 | (((uctxt->ctxt - |
| 1087 | uctxt->dd->first_user_ctxt) * |
| 1088 | HFI1_MAX_SHARED_CTXTS) + fd->subctxt); |
| 1089 | set_bit(_HFI1_EVENT_TID_MMU_NOTIFY_BIT, ev); |
| 1090 | } |
| 1091 | fd->invalid_tid_idx++; |
| 1092 | } |
| 1093 | spin_unlock(&fd->invalid_lock); |
| 1094 | } |
| 1095 | spin_unlock(&fd->rb_lock); |
Mitko Haralanov | f727a0c | 2016-02-05 11:57:46 -0500 | [diff] [blame] | 1096 | } |
| 1097 | |
| 1098 | static inline int mmu_addr_cmp(struct mmu_rb_node *node, unsigned long addr, |
| 1099 | unsigned long len) |
| 1100 | { |
| 1101 | if ((addr + len) <= node->virt) |
| 1102 | return -1; |
| 1103 | else if (addr >= node->virt && addr < (node->virt + node->len)) |
| 1104 | return 0; |
| 1105 | else |
| 1106 | return 1; |
| 1107 | } |
| 1108 | |
| 1109 | static inline int mmu_entry_cmp(struct mmu_rb_node *node, u32 entry) |
| 1110 | { |
| 1111 | if (entry < node->rcventry) |
| 1112 | return -1; |
| 1113 | else if (entry > node->rcventry) |
| 1114 | return 1; |
| 1115 | else |
| 1116 | return 0; |
| 1117 | } |
| 1118 | |
| 1119 | static struct mmu_rb_node *mmu_rb_search_by_addr(struct rb_root *root, |
| 1120 | unsigned long addr) |
| 1121 | { |
| 1122 | struct rb_node *node = root->rb_node; |
| 1123 | |
| 1124 | while (node) { |
| 1125 | struct mmu_rb_node *mnode = |
| 1126 | container_of(node, struct mmu_rb_node, rbnode); |
| 1127 | /* |
| 1128 | * When searching, use at least one page length for size. The |
| 1129 | * MMU notifier will not give us anything less than that. We |
| 1130 | * also don't need anything more than a page because we are |
| 1131 | * guaranteed to have non-overlapping buffers in the tree. |
| 1132 | */ |
| 1133 | int result = mmu_addr_cmp(mnode, addr, PAGE_SIZE); |
| 1134 | |
| 1135 | if (result < 0) |
| 1136 | node = node->rb_left; |
| 1137 | else if (result > 0) |
| 1138 | node = node->rb_right; |
| 1139 | else |
| 1140 | return mnode; |
| 1141 | } |
| 1142 | return NULL; |
| 1143 | } |
| 1144 | |
| 1145 | static inline struct mmu_rb_node *mmu_rb_search_by_entry(struct rb_root *root, |
| 1146 | u32 index) |
| 1147 | { |
| 1148 | struct mmu_rb_node *rbnode; |
| 1149 | struct rb_node *node; |
| 1150 | |
| 1151 | if (root && !RB_EMPTY_ROOT(root)) |
| 1152 | for (node = rb_first(root); node; node = rb_next(node)) { |
| 1153 | rbnode = rb_entry(node, struct mmu_rb_node, rbnode); |
| 1154 | if (rbnode->rcventry == index) |
| 1155 | return rbnode; |
| 1156 | } |
| 1157 | return NULL; |
| 1158 | } |
| 1159 | |
| 1160 | static int mmu_rb_insert_by_entry(struct rb_root *root, |
| 1161 | struct mmu_rb_node *node) |
| 1162 | { |
| 1163 | struct rb_node **new = &root->rb_node, *parent = NULL; |
| 1164 | |
| 1165 | while (*new) { |
| 1166 | struct mmu_rb_node *this = |
| 1167 | container_of(*new, struct mmu_rb_node, rbnode); |
| 1168 | int result = mmu_entry_cmp(this, node->rcventry); |
| 1169 | |
| 1170 | parent = *new; |
| 1171 | if (result < 0) |
| 1172 | new = &((*new)->rb_left); |
| 1173 | else if (result > 0) |
| 1174 | new = &((*new)->rb_right); |
| 1175 | else |
| 1176 | return 1; |
| 1177 | } |
| 1178 | |
| 1179 | rb_link_node(&node->rbnode, parent, new); |
| 1180 | rb_insert_color(&node->rbnode, root); |
| 1181 | return 0; |
| 1182 | } |
| 1183 | |
| 1184 | static int mmu_rb_insert_by_addr(struct rb_root *root, struct mmu_rb_node *node) |
| 1185 | { |
| 1186 | struct rb_node **new = &root->rb_node, *parent = NULL; |
| 1187 | |
| 1188 | /* Figure out where to put new node */ |
| 1189 | while (*new) { |
| 1190 | struct mmu_rb_node *this = |
| 1191 | container_of(*new, struct mmu_rb_node, rbnode); |
| 1192 | int result = mmu_addr_cmp(this, node->virt, node->len); |
| 1193 | |
| 1194 | parent = *new; |
| 1195 | if (result < 0) |
| 1196 | new = &((*new)->rb_left); |
| 1197 | else if (result > 0) |
| 1198 | new = &((*new)->rb_right); |
| 1199 | else |
| 1200 | return 1; |
| 1201 | } |
| 1202 | |
| 1203 | /* Add new node and rebalance tree. */ |
| 1204 | rb_link_node(&node->rbnode, parent, new); |
| 1205 | rb_insert_color(&node->rbnode, root); |
| 1206 | |
| 1207 | return 0; |
| 1208 | } |