1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2016-20 Intel Corporation. */
3
4 #include <linux/file.h>
5 #include <linux/freezer.h>
6 #include <linux/highmem.h>
7 #include <linux/kthread.h>
8 #include <linux/miscdevice.h>
9 #include <linux/node.h>
10 #include <linux/pagemap.h>
11 #include <linux/ratelimit.h>
12 #include <linux/sched/mm.h>
13 #include <linux/sched/signal.h>
14 #include <linux/slab.h>
15 #include <linux/sysfs.h>
16 #include <linux/vmalloc.h>
17 #include <asm/sgx.h>
18 #include "driver.h"
19 #include "encl.h"
20 #include "encls.h"
21
22 struct sgx_epc_section sgx_epc_sections[SGX_MAX_EPC_SECTIONS];
23 static int sgx_nr_epc_sections;
24 static struct task_struct *ksgxd_tsk;
25 static DECLARE_WAIT_QUEUE_HEAD(ksgxd_waitq);
26 static DEFINE_XARRAY(sgx_epc_address_space);
27
28 /*
29 * These variables are part of the state of the reclaimer, and must be accessed
30 * with sgx_reclaimer_lock acquired.
31 */
32 static LIST_HEAD(sgx_active_page_list);
33 static DEFINE_SPINLOCK(sgx_reclaimer_lock);
34
35 static atomic_long_t sgx_nr_free_pages = ATOMIC_LONG_INIT(0);
36
37 /* Nodes with one or more EPC sections. */
38 static nodemask_t sgx_numa_mask;
39
40 /*
41 * Array with one list_head for each possible NUMA node. Each
42 * list contains all the sgx_epc_section's which are on that
43 * node.
44 */
45 static struct sgx_numa_node *sgx_numa_nodes;
46
47 static LIST_HEAD(sgx_dirty_page_list);
48
49 /*
50 * Reset post-kexec EPC pages to the uninitialized state. The pages are removed
51 * from the input list, and made available for the page allocator. SECS pages
52 * prepending their children in the input list are left intact.
53 *
54 * Return 0 when sanitization was successful or kthread was stopped, and the
55 * number of unsanitized pages otherwise.
56 */
__sgx_sanitize_pages(struct list_head * dirty_page_list)57 static unsigned long __sgx_sanitize_pages(struct list_head *dirty_page_list)
58 {
59 unsigned long left_dirty = 0;
60 struct sgx_epc_page *page;
61 LIST_HEAD(dirty);
62 int ret;
63
64 /* dirty_page_list is thread-local, no need for a lock: */
65 while (!list_empty(dirty_page_list)) {
66 if (kthread_should_stop())
67 return 0;
68
69 page = list_first_entry(dirty_page_list, struct sgx_epc_page, list);
70
71 /*
72 * Checking page->poison without holding the node->lock
73 * is racy, but losing the race (i.e. poison is set just
74 * after the check) just means __eremove() will be uselessly
75 * called for a page that sgx_free_epc_page() will put onto
76 * the node->sgx_poison_page_list later.
77 */
78 if (page->poison) {
79 struct sgx_epc_section *section = &sgx_epc_sections[page->section];
80 struct sgx_numa_node *node = section->node;
81
82 spin_lock(&node->lock);
83 list_move(&page->list, &node->sgx_poison_page_list);
84 spin_unlock(&node->lock);
85
86 continue;
87 }
88
89 ret = __eremove(sgx_get_epc_virt_addr(page));
90 if (!ret) {
91 /*
92 * page is now sanitized. Make it available via the SGX
93 * page allocator:
94 */
95 list_del(&page->list);
96 sgx_free_epc_page(page);
97 } else {
98 /* The page is not yet clean - move to the dirty list. */
99 list_move_tail(&page->list, &dirty);
100 left_dirty++;
101 }
102
103 cond_resched();
104 }
105
106 list_splice(&dirty, dirty_page_list);
107 return left_dirty;
108 }
109
sgx_reclaimer_age(struct sgx_epc_page * epc_page)110 static bool sgx_reclaimer_age(struct sgx_epc_page *epc_page)
111 {
112 struct sgx_encl_page *page = epc_page->owner;
113 struct sgx_encl *encl = page->encl;
114 struct sgx_encl_mm *encl_mm;
115 bool ret = true;
116 int idx;
117
118 idx = srcu_read_lock(&encl->srcu);
119
120 list_for_each_entry_rcu(encl_mm, &encl->mm_list, list) {
121 if (!mmget_not_zero(encl_mm->mm))
122 continue;
123
124 mmap_read_lock(encl_mm->mm);
125 ret = !sgx_encl_test_and_clear_young(encl_mm->mm, page);
126 mmap_read_unlock(encl_mm->mm);
127
128 mmput_async(encl_mm->mm);
129
130 if (!ret)
131 break;
132 }
133
134 srcu_read_unlock(&encl->srcu, idx);
135
136 if (!ret)
137 return false;
138
139 return true;
140 }
141
sgx_reclaimer_block(struct sgx_epc_page * epc_page)142 static void sgx_reclaimer_block(struct sgx_epc_page *epc_page)
143 {
144 struct sgx_encl_page *page = epc_page->owner;
145 unsigned long addr = page->desc & PAGE_MASK;
146 struct sgx_encl *encl = page->encl;
147 int ret;
148
149 sgx_zap_enclave_ptes(encl, addr);
150
151 mutex_lock(&encl->lock);
152
153 ret = __eblock(sgx_get_epc_virt_addr(epc_page));
154 if (encls_failed(ret))
155 ENCLS_WARN(ret, "EBLOCK");
156
157 mutex_unlock(&encl->lock);
158 }
159
__sgx_encl_ewb(struct sgx_epc_page * epc_page,void * va_slot,struct sgx_backing * backing)160 static int __sgx_encl_ewb(struct sgx_epc_page *epc_page, void *va_slot,
161 struct sgx_backing *backing)
162 {
163 struct sgx_pageinfo pginfo;
164 int ret;
165
166 pginfo.addr = 0;
167 pginfo.secs = 0;
168
169 pginfo.contents = (unsigned long)kmap_local_page(backing->contents);
170 pginfo.metadata = (unsigned long)kmap_local_page(backing->pcmd) +
171 backing->pcmd_offset;
172
173 ret = __ewb(&pginfo, sgx_get_epc_virt_addr(epc_page), va_slot);
174 set_page_dirty(backing->pcmd);
175 set_page_dirty(backing->contents);
176
177 kunmap_local((void *)(unsigned long)(pginfo.metadata -
178 backing->pcmd_offset));
179 kunmap_local((void *)(unsigned long)pginfo.contents);
180
181 return ret;
182 }
183
sgx_ipi_cb(void * info)184 void sgx_ipi_cb(void *info)
185 {
186 }
187
188 /*
189 * Swap page to the regular memory transformed to the blocked state by using
190 * EBLOCK, which means that it can no longer be referenced (no new TLB entries).
191 *
192 * The first trial just tries to write the page assuming that some other thread
193 * has reset the count for threads inside the enclave by using ETRACK, and
194 * previous thread count has been zeroed out. The second trial calls ETRACK
195 * before EWB. If that fails we kick all the HW threads out, and then do EWB,
196 * which should be guaranteed the succeed.
197 */
sgx_encl_ewb(struct sgx_epc_page * epc_page,struct sgx_backing * backing)198 static void sgx_encl_ewb(struct sgx_epc_page *epc_page,
199 struct sgx_backing *backing)
200 {
201 struct sgx_encl_page *encl_page = epc_page->owner;
202 struct sgx_encl *encl = encl_page->encl;
203 struct sgx_va_page *va_page;
204 unsigned int va_offset;
205 void *va_slot;
206 int ret;
207
208 encl_page->desc &= ~SGX_ENCL_PAGE_BEING_RECLAIMED;
209
210 va_page = list_first_entry(&encl->va_pages, struct sgx_va_page,
211 list);
212 va_offset = sgx_alloc_va_slot(va_page);
213 va_slot = sgx_get_epc_virt_addr(va_page->epc_page) + va_offset;
214 if (sgx_va_page_full(va_page))
215 list_move_tail(&va_page->list, &encl->va_pages);
216
217 ret = __sgx_encl_ewb(epc_page, va_slot, backing);
218 if (ret == SGX_NOT_TRACKED) {
219 ret = __etrack(sgx_get_epc_virt_addr(encl->secs.epc_page));
220 if (ret) {
221 if (encls_failed(ret))
222 ENCLS_WARN(ret, "ETRACK");
223 }
224
225 ret = __sgx_encl_ewb(epc_page, va_slot, backing);
226 if (ret == SGX_NOT_TRACKED) {
227 /*
228 * Slow path, send IPIs to kick cpus out of the
229 * enclave. Note, it's imperative that the cpu
230 * mask is generated *after* ETRACK, else we'll
231 * miss cpus that entered the enclave between
232 * generating the mask and incrementing epoch.
233 */
234 on_each_cpu_mask(sgx_encl_cpumask(encl),
235 sgx_ipi_cb, NULL, 1);
236 ret = __sgx_encl_ewb(epc_page, va_slot, backing);
237 }
238 }
239
240 if (ret) {
241 if (encls_failed(ret))
242 ENCLS_WARN(ret, "EWB");
243
244 sgx_free_va_slot(va_page, va_offset);
245 } else {
246 encl_page->desc |= va_offset;
247 encl_page->va_page = va_page;
248 }
249 }
250
sgx_reclaimer_write(struct sgx_epc_page * epc_page,struct sgx_backing * backing)251 static void sgx_reclaimer_write(struct sgx_epc_page *epc_page,
252 struct sgx_backing *backing)
253 {
254 struct sgx_encl_page *encl_page = epc_page->owner;
255 struct sgx_encl *encl = encl_page->encl;
256 struct sgx_backing secs_backing;
257 int ret;
258
259 mutex_lock(&encl->lock);
260
261 sgx_encl_ewb(epc_page, backing);
262 encl_page->epc_page = NULL;
263 encl->secs_child_cnt--;
264 sgx_encl_put_backing(backing);
265
266 if (!encl->secs_child_cnt && test_bit(SGX_ENCL_INITIALIZED, &encl->flags)) {
267 ret = sgx_encl_alloc_backing(encl, PFN_DOWN(encl->size),
268 &secs_backing);
269 if (ret)
270 goto out;
271
272 sgx_encl_ewb(encl->secs.epc_page, &secs_backing);
273
274 sgx_encl_free_epc_page(encl->secs.epc_page);
275 encl->secs.epc_page = NULL;
276
277 sgx_encl_put_backing(&secs_backing);
278 }
279
280 out:
281 mutex_unlock(&encl->lock);
282 }
283
284 /*
285 * Take a fixed number of pages from the head of the active page pool and
286 * reclaim them to the enclave's private shmem files. Skip the pages, which have
287 * been accessed since the last scan. Move those pages to the tail of active
288 * page pool so that the pages get scanned in LRU like fashion.
289 *
290 * Batch process a chunk of pages (at the moment 16) in order to degrade amount
291 * of IPI's and ETRACK's potentially required. sgx_encl_ewb() does degrade a bit
292 * among the HW threads with three stage EWB pipeline (EWB, ETRACK + EWB and IPI
293 * + EWB) but not sufficiently. Reclaiming one page at a time would also be
294 * problematic as it would increase the lock contention too much, which would
295 * halt forward progress.
296 */
sgx_reclaim_pages(void)297 static void sgx_reclaim_pages(void)
298 {
299 struct sgx_epc_page *chunk[SGX_NR_TO_SCAN];
300 struct sgx_backing backing[SGX_NR_TO_SCAN];
301 struct sgx_encl_page *encl_page;
302 struct sgx_epc_page *epc_page;
303 pgoff_t page_index;
304 int cnt = 0;
305 int ret;
306 int i;
307
308 spin_lock(&sgx_reclaimer_lock);
309 for (i = 0; i < SGX_NR_TO_SCAN; i++) {
310 if (list_empty(&sgx_active_page_list))
311 break;
312
313 epc_page = list_first_entry(&sgx_active_page_list,
314 struct sgx_epc_page, list);
315 list_del_init(&epc_page->list);
316 encl_page = epc_page->owner;
317
318 if (kref_get_unless_zero(&encl_page->encl->refcount) != 0)
319 chunk[cnt++] = epc_page;
320 else
321 /* The owner is freeing the page. No need to add the
322 * page back to the list of reclaimable pages.
323 */
324 epc_page->flags &= ~SGX_EPC_PAGE_RECLAIMER_TRACKED;
325 }
326 spin_unlock(&sgx_reclaimer_lock);
327
328 for (i = 0; i < cnt; i++) {
329 epc_page = chunk[i];
330 encl_page = epc_page->owner;
331
332 if (!sgx_reclaimer_age(epc_page))
333 goto skip;
334
335 page_index = PFN_DOWN(encl_page->desc - encl_page->encl->base);
336
337 mutex_lock(&encl_page->encl->lock);
338 ret = sgx_encl_alloc_backing(encl_page->encl, page_index, &backing[i]);
339 if (ret) {
340 mutex_unlock(&encl_page->encl->lock);
341 goto skip;
342 }
343
344 encl_page->desc |= SGX_ENCL_PAGE_BEING_RECLAIMED;
345 mutex_unlock(&encl_page->encl->lock);
346 continue;
347
348 skip:
349 spin_lock(&sgx_reclaimer_lock);
350 list_add_tail(&epc_page->list, &sgx_active_page_list);
351 spin_unlock(&sgx_reclaimer_lock);
352
353 kref_put(&encl_page->encl->refcount, sgx_encl_release);
354
355 chunk[i] = NULL;
356 }
357
358 for (i = 0; i < cnt; i++) {
359 epc_page = chunk[i];
360 if (epc_page)
361 sgx_reclaimer_block(epc_page);
362 }
363
364 for (i = 0; i < cnt; i++) {
365 epc_page = chunk[i];
366 if (!epc_page)
367 continue;
368
369 encl_page = epc_page->owner;
370 sgx_reclaimer_write(epc_page, &backing[i]);
371
372 kref_put(&encl_page->encl->refcount, sgx_encl_release);
373 epc_page->flags &= ~SGX_EPC_PAGE_RECLAIMER_TRACKED;
374
375 sgx_free_epc_page(epc_page);
376 }
377 }
378
sgx_should_reclaim(unsigned long watermark)379 static bool sgx_should_reclaim(unsigned long watermark)
380 {
381 return atomic_long_read(&sgx_nr_free_pages) < watermark &&
382 !list_empty(&sgx_active_page_list);
383 }
384
385 /*
386 * sgx_reclaim_direct() should be called (without enclave's mutex held)
387 * in locations where SGX memory resources might be low and might be
388 * needed in order to make forward progress.
389 */
sgx_reclaim_direct(void)390 void sgx_reclaim_direct(void)
391 {
392 if (sgx_should_reclaim(SGX_NR_LOW_PAGES))
393 sgx_reclaim_pages();
394 }
395
ksgxd(void * p)396 static int ksgxd(void *p)
397 {
398 set_freezable();
399
400 /*
401 * Sanitize pages in order to recover from kexec(). The 2nd pass is
402 * required for SECS pages, whose child pages blocked EREMOVE.
403 */
404 __sgx_sanitize_pages(&sgx_dirty_page_list);
405 WARN_ON(__sgx_sanitize_pages(&sgx_dirty_page_list));
406
407 while (!kthread_should_stop()) {
408 if (try_to_freeze())
409 continue;
410
411 wait_event_freezable(ksgxd_waitq,
412 kthread_should_stop() ||
413 sgx_should_reclaim(SGX_NR_HIGH_PAGES));
414
415 if (sgx_should_reclaim(SGX_NR_HIGH_PAGES))
416 sgx_reclaim_pages();
417
418 cond_resched();
419 }
420
421 return 0;
422 }
423
sgx_page_reclaimer_init(void)424 static bool __init sgx_page_reclaimer_init(void)
425 {
426 struct task_struct *tsk;
427
428 tsk = kthread_run(ksgxd, NULL, "ksgxd");
429 if (IS_ERR(tsk))
430 return false;
431
432 ksgxd_tsk = tsk;
433
434 return true;
435 }
436
current_is_ksgxd(void)437 bool current_is_ksgxd(void)
438 {
439 return current == ksgxd_tsk;
440 }
441
__sgx_alloc_epc_page_from_node(int nid)442 static struct sgx_epc_page *__sgx_alloc_epc_page_from_node(int nid)
443 {
444 struct sgx_numa_node *node = &sgx_numa_nodes[nid];
445 struct sgx_epc_page *page = NULL;
446
447 spin_lock(&node->lock);
448
449 if (list_empty(&node->free_page_list)) {
450 spin_unlock(&node->lock);
451 return NULL;
452 }
453
454 page = list_first_entry(&node->free_page_list, struct sgx_epc_page, list);
455 list_del_init(&page->list);
456 page->flags = 0;
457
458 spin_unlock(&node->lock);
459 atomic_long_dec(&sgx_nr_free_pages);
460
461 return page;
462 }
463
464 /**
465 * __sgx_alloc_epc_page() - Allocate an EPC page
466 *
467 * Iterate through NUMA nodes and reserve ia free EPC page to the caller. Start
468 * from the NUMA node, where the caller is executing.
469 *
470 * Return:
471 * - an EPC page: A borrowed EPC pages were available.
472 * - NULL: Out of EPC pages.
473 */
__sgx_alloc_epc_page(void)474 struct sgx_epc_page *__sgx_alloc_epc_page(void)
475 {
476 struct sgx_epc_page *page;
477 int nid_of_current = numa_node_id();
478 int nid_start, nid;
479
480 /*
481 * Try local node first. If it doesn't have an EPC section,
482 * fall back to the non-local NUMA nodes.
483 */
484 if (node_isset(nid_of_current, sgx_numa_mask))
485 nid_start = nid_of_current;
486 else
487 nid_start = next_node_in(nid_of_current, sgx_numa_mask);
488
489 nid = nid_start;
490 do {
491 page = __sgx_alloc_epc_page_from_node(nid);
492 if (page)
493 return page;
494
495 nid = next_node_in(nid, sgx_numa_mask);
496 } while (nid != nid_start);
497
498 return ERR_PTR(-ENOMEM);
499 }
500
501 /**
502 * sgx_mark_page_reclaimable() - Mark a page as reclaimable
503 * @page: EPC page
504 *
505 * Mark a page as reclaimable and add it to the active page list. Pages
506 * are automatically removed from the active list when freed.
507 */
sgx_mark_page_reclaimable(struct sgx_epc_page * page)508 void sgx_mark_page_reclaimable(struct sgx_epc_page *page)
509 {
510 spin_lock(&sgx_reclaimer_lock);
511 page->flags |= SGX_EPC_PAGE_RECLAIMER_TRACKED;
512 list_add_tail(&page->list, &sgx_active_page_list);
513 spin_unlock(&sgx_reclaimer_lock);
514 }
515
516 /**
517 * sgx_unmark_page_reclaimable() - Remove a page from the reclaim list
518 * @page: EPC page
519 *
520 * Clear the reclaimable flag and remove the page from the active page list.
521 *
522 * Return:
523 * 0 on success,
524 * -EBUSY if the page is in the process of being reclaimed
525 */
sgx_unmark_page_reclaimable(struct sgx_epc_page * page)526 int sgx_unmark_page_reclaimable(struct sgx_epc_page *page)
527 {
528 spin_lock(&sgx_reclaimer_lock);
529 if (page->flags & SGX_EPC_PAGE_RECLAIMER_TRACKED) {
530 /* The page is being reclaimed. */
531 if (list_empty(&page->list)) {
532 spin_unlock(&sgx_reclaimer_lock);
533 return -EBUSY;
534 }
535
536 list_del(&page->list);
537 page->flags &= ~SGX_EPC_PAGE_RECLAIMER_TRACKED;
538 }
539 spin_unlock(&sgx_reclaimer_lock);
540
541 return 0;
542 }
543
544 /**
545 * sgx_alloc_epc_page() - Allocate an EPC page
546 * @owner: the owner of the EPC page
547 * @reclaim: reclaim pages if necessary
548 *
549 * Iterate through EPC sections and borrow a free EPC page to the caller. When a
550 * page is no longer needed it must be released with sgx_free_epc_page(). If
551 * @reclaim is set to true, directly reclaim pages when we are out of pages. No
552 * mm's can be locked when @reclaim is set to true.
553 *
554 * Finally, wake up ksgxd when the number of pages goes below the watermark
555 * before returning back to the caller.
556 *
557 * Return:
558 * an EPC page,
559 * -errno on error
560 */
sgx_alloc_epc_page(void * owner,bool reclaim)561 struct sgx_epc_page *sgx_alloc_epc_page(void *owner, bool reclaim)
562 {
563 struct sgx_epc_page *page;
564
565 for ( ; ; ) {
566 page = __sgx_alloc_epc_page();
567 if (!IS_ERR(page)) {
568 page->owner = owner;
569 break;
570 }
571
572 if (list_empty(&sgx_active_page_list))
573 return ERR_PTR(-ENOMEM);
574
575 if (!reclaim) {
576 page = ERR_PTR(-EBUSY);
577 break;
578 }
579
580 if (signal_pending(current)) {
581 page = ERR_PTR(-ERESTARTSYS);
582 break;
583 }
584
585 sgx_reclaim_pages();
586 cond_resched();
587 }
588
589 if (sgx_should_reclaim(SGX_NR_LOW_PAGES))
590 wake_up(&ksgxd_waitq);
591
592 return page;
593 }
594
595 /**
596 * sgx_free_epc_page() - Free an EPC page
597 * @page: an EPC page
598 *
599 * Put the EPC page back to the list of free pages. It's the caller's
600 * responsibility to make sure that the page is in uninitialized state. In other
601 * words, do EREMOVE, EWB or whatever operation is necessary before calling
602 * this function.
603 */
sgx_free_epc_page(struct sgx_epc_page * page)604 void sgx_free_epc_page(struct sgx_epc_page *page)
605 {
606 struct sgx_epc_section *section = &sgx_epc_sections[page->section];
607 struct sgx_numa_node *node = section->node;
608
609 spin_lock(&node->lock);
610
611 page->owner = NULL;
612 if (page->poison)
613 list_add(&page->list, &node->sgx_poison_page_list);
614 else
615 list_add_tail(&page->list, &node->free_page_list);
616 page->flags = SGX_EPC_PAGE_IS_FREE;
617
618 spin_unlock(&node->lock);
619 atomic_long_inc(&sgx_nr_free_pages);
620 }
621
sgx_setup_epc_section(u64 phys_addr,u64 size,unsigned long index,struct sgx_epc_section * section)622 static bool __init sgx_setup_epc_section(u64 phys_addr, u64 size,
623 unsigned long index,
624 struct sgx_epc_section *section)
625 {
626 unsigned long nr_pages = size >> PAGE_SHIFT;
627 unsigned long i;
628
629 section->virt_addr = memremap(phys_addr, size, MEMREMAP_WB);
630 if (!section->virt_addr)
631 return false;
632
633 section->pages = vmalloc_array(nr_pages, sizeof(struct sgx_epc_page));
634 if (!section->pages) {
635 memunmap(section->virt_addr);
636 return false;
637 }
638
639 section->phys_addr = phys_addr;
640 xa_store_range(&sgx_epc_address_space, section->phys_addr,
641 phys_addr + size - 1, section, GFP_KERNEL);
642
643 for (i = 0; i < nr_pages; i++) {
644 section->pages[i].section = index;
645 section->pages[i].flags = 0;
646 section->pages[i].owner = NULL;
647 section->pages[i].poison = 0;
648 list_add_tail(§ion->pages[i].list, &sgx_dirty_page_list);
649 }
650
651 return true;
652 }
653
arch_is_platform_page(u64 paddr)654 bool arch_is_platform_page(u64 paddr)
655 {
656 return !!xa_load(&sgx_epc_address_space, paddr);
657 }
658 EXPORT_SYMBOL_GPL(arch_is_platform_page);
659
sgx_paddr_to_page(u64 paddr)660 static struct sgx_epc_page *sgx_paddr_to_page(u64 paddr)
661 {
662 struct sgx_epc_section *section;
663
664 section = xa_load(&sgx_epc_address_space, paddr);
665 if (!section)
666 return NULL;
667
668 return §ion->pages[PFN_DOWN(paddr - section->phys_addr)];
669 }
670
671 /*
672 * Called in process context to handle a hardware reported
673 * error in an SGX EPC page.
674 * If the MF_ACTION_REQUIRED bit is set in flags, then the
675 * context is the task that consumed the poison data. Otherwise
676 * this is called from a kernel thread unrelated to the page.
677 */
arch_memory_failure(unsigned long pfn,int flags)678 int arch_memory_failure(unsigned long pfn, int flags)
679 {
680 struct sgx_epc_page *page = sgx_paddr_to_page(pfn << PAGE_SHIFT);
681 struct sgx_epc_section *section;
682 struct sgx_numa_node *node;
683
684 /*
685 * mm/memory-failure.c calls this routine for all errors
686 * where there isn't a "struct page" for the address. But that
687 * includes other address ranges besides SGX.
688 */
689 if (!page)
690 return -ENXIO;
691
692 /*
693 * If poison was consumed synchronously. Send a SIGBUS to
694 * the task. Hardware has already exited the SGX enclave and
695 * will not allow re-entry to an enclave that has a memory
696 * error. The signal may help the task understand why the
697 * enclave is broken.
698 */
699 if (flags & MF_ACTION_REQUIRED)
700 force_sig(SIGBUS);
701
702 section = &sgx_epc_sections[page->section];
703 node = section->node;
704
705 spin_lock(&node->lock);
706
707 /* Already poisoned? Nothing more to do */
708 if (page->poison)
709 goto out;
710
711 page->poison = 1;
712
713 /*
714 * If the page is on a free list, move it to the per-node
715 * poison page list.
716 */
717 if (page->flags & SGX_EPC_PAGE_IS_FREE) {
718 list_move(&page->list, &node->sgx_poison_page_list);
719 goto out;
720 }
721
722 /*
723 * TBD: Add additional plumbing to enable pre-emptive
724 * action for asynchronous poison notification. Until
725 * then just hope that the poison:
726 * a) is not accessed - sgx_free_epc_page() will deal with it
727 * when the user gives it back
728 * b) results in a recoverable machine check rather than
729 * a fatal one
730 */
731 out:
732 spin_unlock(&node->lock);
733 return 0;
734 }
735
736 /*
737 * A section metric is concatenated in a way that @low bits 12-31 define the
738 * bits 12-31 of the metric and @high bits 0-19 define the bits 32-51 of the
739 * metric.
740 */
sgx_calc_section_metric(u64 low,u64 high)741 static inline u64 __init sgx_calc_section_metric(u64 low, u64 high)
742 {
743 return (low & GENMASK_ULL(31, 12)) +
744 ((high & GENMASK_ULL(19, 0)) << 32);
745 }
746
747 #ifdef CONFIG_NUMA
sgx_total_bytes_show(struct device * dev,struct device_attribute * attr,char * buf)748 static ssize_t sgx_total_bytes_show(struct device *dev, struct device_attribute *attr, char *buf)
749 {
750 return sysfs_emit(buf, "%lu\n", sgx_numa_nodes[dev->id].size);
751 }
752 static DEVICE_ATTR_RO(sgx_total_bytes);
753
arch_node_attr_is_visible(struct kobject * kobj,struct attribute * attr,int idx)754 static umode_t arch_node_attr_is_visible(struct kobject *kobj,
755 struct attribute *attr, int idx)
756 {
757 /* Make all x86/ attributes invisible when SGX is not initialized: */
758 if (nodes_empty(sgx_numa_mask))
759 return 0;
760
761 return attr->mode;
762 }
763
764 static struct attribute *arch_node_dev_attrs[] = {
765 &dev_attr_sgx_total_bytes.attr,
766 NULL,
767 };
768
769 const struct attribute_group arch_node_dev_group = {
770 .name = "x86",
771 .attrs = arch_node_dev_attrs,
772 .is_visible = arch_node_attr_is_visible,
773 };
774
arch_update_sysfs_visibility(int nid)775 static void __init arch_update_sysfs_visibility(int nid)
776 {
777 struct node *node = node_devices[nid];
778 int ret;
779
780 ret = sysfs_update_group(&node->dev.kobj, &arch_node_dev_group);
781
782 if (ret)
783 pr_err("sysfs update failed (%d), files may be invisible", ret);
784 }
785 #else /* !CONFIG_NUMA */
arch_update_sysfs_visibility(int nid)786 static void __init arch_update_sysfs_visibility(int nid) {}
787 #endif
788
sgx_page_cache_init(void)789 static bool __init sgx_page_cache_init(void)
790 {
791 u32 eax, ebx, ecx, edx, type;
792 u64 pa, size;
793 int nid;
794 int i;
795
796 sgx_numa_nodes = kmalloc_array(num_possible_nodes(), sizeof(*sgx_numa_nodes), GFP_KERNEL);
797 if (!sgx_numa_nodes)
798 return false;
799
800 for (i = 0; i < ARRAY_SIZE(sgx_epc_sections); i++) {
801 cpuid_count(SGX_CPUID, i + SGX_CPUID_EPC, &eax, &ebx, &ecx, &edx);
802
803 type = eax & SGX_CPUID_EPC_MASK;
804 if (type == SGX_CPUID_EPC_INVALID)
805 break;
806
807 if (type != SGX_CPUID_EPC_SECTION) {
808 pr_err_once("Unknown EPC section type: %u\n", type);
809 break;
810 }
811
812 pa = sgx_calc_section_metric(eax, ebx);
813 size = sgx_calc_section_metric(ecx, edx);
814
815 pr_info("EPC section 0x%llx-0x%llx\n", pa, pa + size - 1);
816
817 if (!sgx_setup_epc_section(pa, size, i, &sgx_epc_sections[i])) {
818 pr_err("No free memory for an EPC section\n");
819 break;
820 }
821
822 nid = numa_map_to_online_node(phys_to_target_node(pa));
823 if (nid == NUMA_NO_NODE) {
824 /* The physical address is already printed above. */
825 pr_warn(FW_BUG "Unable to map EPC section to online node. Fallback to the NUMA node 0.\n");
826 nid = 0;
827 }
828
829 if (!node_isset(nid, sgx_numa_mask)) {
830 spin_lock_init(&sgx_numa_nodes[nid].lock);
831 INIT_LIST_HEAD(&sgx_numa_nodes[nid].free_page_list);
832 INIT_LIST_HEAD(&sgx_numa_nodes[nid].sgx_poison_page_list);
833 node_set(nid, sgx_numa_mask);
834 sgx_numa_nodes[nid].size = 0;
835
836 /* Make SGX-specific node sysfs files visible: */
837 arch_update_sysfs_visibility(nid);
838 }
839
840 sgx_epc_sections[i].node = &sgx_numa_nodes[nid];
841 sgx_numa_nodes[nid].size += size;
842
843 sgx_nr_epc_sections++;
844 }
845
846 if (!sgx_nr_epc_sections) {
847 pr_err("There are zero EPC sections.\n");
848 return false;
849 }
850
851 for_each_online_node(nid) {
852 if (!node_isset(nid, sgx_numa_mask) &&
853 node_state(nid, N_MEMORY) && node_state(nid, N_CPU))
854 pr_info("node%d has both CPUs and memory but doesn't have an EPC section\n",
855 nid);
856 }
857
858 return true;
859 }
860
861 /*
862 * Update the SGX_LEPUBKEYHASH MSRs to the values specified by caller.
863 * Bare-metal driver requires to update them to hash of enclave's signer
864 * before EINIT. KVM needs to update them to guest's virtual MSR values
865 * before doing EINIT from guest.
866 */
sgx_update_lepubkeyhash(u64 * lepubkeyhash)867 void sgx_update_lepubkeyhash(u64 *lepubkeyhash)
868 {
869 int i;
870
871 WARN_ON_ONCE(preemptible());
872
873 for (i = 0; i < 4; i++)
874 wrmsrl(MSR_IA32_SGXLEPUBKEYHASH0 + i, lepubkeyhash[i]);
875 }
876
877 const struct file_operations sgx_provision_fops = {
878 .owner = THIS_MODULE,
879 };
880
881 static struct miscdevice sgx_dev_provision = {
882 .minor = MISC_DYNAMIC_MINOR,
883 .name = "sgx_provision",
884 .nodename = "sgx_provision",
885 .fops = &sgx_provision_fops,
886 };
887
888 /**
889 * sgx_set_attribute() - Update allowed attributes given file descriptor
890 * @allowed_attributes: Pointer to allowed enclave attributes
891 * @attribute_fd: File descriptor for specific attribute
892 *
893 * Append enclave attribute indicated by file descriptor to allowed
894 * attributes. Currently only SGX_ATTR_PROVISIONKEY indicated by
895 * /dev/sgx_provision is supported.
896 *
897 * Return:
898 * -0: SGX_ATTR_PROVISIONKEY is appended to allowed_attributes
899 * -EINVAL: Invalid, or not supported file descriptor
900 */
sgx_set_attribute(unsigned long * allowed_attributes,unsigned int attribute_fd)901 int sgx_set_attribute(unsigned long *allowed_attributes,
902 unsigned int attribute_fd)
903 {
904 CLASS(fd, f)(attribute_fd);
905
906 if (fd_empty(f))
907 return -EINVAL;
908
909 if (fd_file(f)->f_op != &sgx_provision_fops)
910 return -EINVAL;
911
912 *allowed_attributes |= SGX_ATTR_PROVISIONKEY;
913 return 0;
914 }
915 EXPORT_SYMBOL_GPL(sgx_set_attribute);
916
sgx_init(void)917 static int __init sgx_init(void)
918 {
919 int ret;
920 int i;
921
922 if (!cpu_feature_enabled(X86_FEATURE_SGX))
923 return -ENODEV;
924
925 if (!sgx_page_cache_init())
926 return -ENOMEM;
927
928 if (!sgx_page_reclaimer_init()) {
929 ret = -ENOMEM;
930 goto err_page_cache;
931 }
932
933 ret = misc_register(&sgx_dev_provision);
934 if (ret)
935 goto err_kthread;
936
937 /*
938 * Always try to initialize the native *and* KVM drivers.
939 * The KVM driver is less picky than the native one and
940 * can function if the native one is not supported on the
941 * current system or fails to initialize.
942 *
943 * Error out only if both fail to initialize.
944 */
945 ret = sgx_drv_init();
946
947 if (sgx_vepc_init() && ret)
948 goto err_provision;
949
950 return 0;
951
952 err_provision:
953 misc_deregister(&sgx_dev_provision);
954
955 err_kthread:
956 kthread_stop(ksgxd_tsk);
957
958 err_page_cache:
959 for (i = 0; i < sgx_nr_epc_sections; i++) {
960 vfree(sgx_epc_sections[i].pages);
961 memunmap(sgx_epc_sections[i].virt_addr);
962 }
963
964 return ret;
965 }
966
967 device_initcall(sgx_init);
968