xref: /freebsd/sys/amd64/sgx/sgx.c (revision 7899f917b1c0ea178f1d2be0cfb452086d079d23)
1 /*-
2  * Copyright (c) 2017 Ruslan Bukin <br@bsdpad.com>
3  * All rights reserved.
4  *
5  * This software was developed by BAE Systems, the University of Cambridge
6  * Computer Laboratory, and Memorial University under DARPA/AFRL contract
7  * FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent Computing
8  * (TC) research program.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * Design overview.
34  *
35  * The driver provides character device for mmap(2) and ioctl(2) system calls
36  * allowing user to manage isolated compartments ("enclaves") in user VA space.
37  *
38  * The driver duties is EPC pages management, enclave management, user data
39  * validation.
40  *
41  * This driver requires Intel SGX support from hardware.
42  *
43  * /dev/sgx:
44  *    .mmap:
45  *        sgx_mmap_single() allocates VM object with following pager
46  *        operations:
47  *              a) sgx_pg_ctor():
48  *                  VM object constructor does nothing
49  *              b) sgx_pg_dtor():
50  *                  VM object destructor destroys the SGX enclave associated
51  *                  with the object: it frees all the EPC pages allocated for
52  *                  enclave and removes the enclave.
53  *              c) sgx_pg_fault():
54  *                  VM object fault handler does nothing
55  *
56  *    .ioctl:
57  *        sgx_ioctl():
58  *               a) SGX_IOC_ENCLAVE_CREATE
59  *                   Adds Enclave SECS page: initial step of enclave creation.
60  *               b) SGX_IOC_ENCLAVE_ADD_PAGE
61  *                   Adds TCS, REG pages to the enclave.
62  *               c) SGX_IOC_ENCLAVE_INIT
63  *                   Finalizes enclave creation.
64  *
65  * Enclave lifecycle:
66  *          .-- ECREATE  -- Add SECS page
67  *   Kernel |   EADD     -- Add TCS, REG pages
68  *    space |   EEXTEND  -- Measure the page (take unique hash)
69  *    ENCLS |   EPA      -- Allocate version array page
70  *          '-- EINIT    -- Finalize enclave creation
71  *   User   .-- EENTER   -- Go to entry point of enclave
72  *    space |   EEXIT    -- Exit back to main application
73  *    ENCLU '-- ERESUME  -- Resume enclave execution (e.g. after exception)
74  *
75  * Enclave lifecycle from driver point of view:
76  *  1) User calls mmap() on /dev/sgx: we allocate a VM object
77  *  2) User calls ioctl SGX_IOC_ENCLAVE_CREATE: we look for the VM object
78  *     associated with user process created on step 1, create SECS physical
79  *     page and store it in enclave's VM object queue by special index
80  *     SGX_SECS_VM_OBJECT_INDEX.
81  *  3) User calls ioctl SGX_IOC_ENCLAVE_ADD_PAGE: we look for enclave created
82  *     on step 2, create TCS or REG physical page and map it to specified by
83  *     user address of enclave VM object.
84  *  4) User finalizes enclave creation with ioctl SGX_IOC_ENCLAVE_INIT call.
85  *  5) User can freely enter to and exit from enclave using ENCLU instructions
86  *     from userspace: the driver does nothing here.
87  *  6) User proceed munmap(2) system call (or the process with enclave dies):
88  *     we destroy the enclave associated with the object.
89  *
90  * EPC page types and their indexes in VM object queue:
91  *   - PT_SECS index is special and equals SGX_SECS_VM_OBJECT_INDEX (-1);
92  *   - PT_TCS and PT_REG indexes are specified by user in addr field of ioctl
93  *     request data and determined as follows:
94  *       pidx = OFF_TO_IDX(addp->addr - vmh->base);
95  *   - PT_VA index is special, created for PT_REG, PT_TCS and PT_SECS pages
96  *     and determined by formula:
97  *       va_page_idx = - SGX_VA_PAGES_OFFS - (page_idx / SGX_VA_PAGE_SLOTS);
98  *     PT_VA page can hold versions of up to 512 pages, and slot for each
99  *     page in PT_VA page is determined as follows:
100  *       va_slot_idx = page_idx % SGX_VA_PAGE_SLOTS;
101  *   - PT_TRIM is unused.
102  *
103  * Locking:
104  *    SGX ENCLS set of instructions have limitations on concurrency:
105  *    some instructions can't be executed same time on different CPUs.
106  *    We use sc->mtx_encls lock around them to prevent concurrent execution.
107  *    sc->mtx lock is used to manage list of created enclaves and the state of
108  *    SGX driver.
109  *
110  * Eviction of EPC pages:
111  *    Eviction support is not implemented in this driver, however the driver
112  *    manages VA (version array) pages: it allocates a VA slot for each EPC
113  *    page. This will be required for eviction support in future.
114  *    VA pages and slots are currently unused.
115  *
116  * Intel® 64 and IA-32 Architectures Software Developer's Manual
117  * https://software.intel.com/en-us/articles/intel-sdm
118  */
119 
120 #include <sys/param.h>
121 #include <sys/systm.h>
122 #include <sys/ioccom.h>
123 #include <sys/malloc.h>
124 #include <sys/kernel.h>
125 #include <sys/lock.h>
126 #include <sys/mutex.h>
127 #include <sys/rwlock.h>
128 #include <sys/conf.h>
129 #include <sys/module.h>
130 #include <sys/proc.h>
131 #include <sys/vmem.h>
132 #include <sys/vmmeter.h>
133 
134 #include <vm/vm.h>
135 #include <vm/vm_param.h>
136 #include <vm/vm_extern.h>
137 #include <vm/vm_kern.h>
138 #include <vm/vm_page.h>
139 #include <vm/vm_map.h>
140 #include <vm/vm_object.h>
141 #include <vm/vm_pager.h>
142 #include <vm/vm_phys.h>
143 #include <vm/vm_radix.h>
144 #include <vm/pmap.h>
145 
146 #include <machine/md_var.h>
147 #include <machine/specialreg.h>
148 #include <machine/cpufunc.h>
149 #include <machine/sgx.h>
150 #include <machine/sgxreg.h>
151 
152 #include <amd64/sgx/sgxvar.h>
153 
154 #define	SGX_DEBUG
155 #undef	SGX_DEBUG
156 
157 #ifdef	SGX_DEBUG
158 #define	dprintf(fmt, ...)	printf(fmt, ##__VA_ARGS__)
159 #else
160 #define	dprintf(fmt, ...)
161 #endif
162 
163 static struct cdev_pager_ops sgx_pg_ops;
164 struct sgx_softc sgx_sc;
165 
166 static int
167 sgx_get_epc_page(struct sgx_softc *sc, struct epc_page **epc)
168 {
169 	vmem_addr_t addr;
170 	int i;
171 
172 	if (vmem_alloc(sc->vmem_epc, PAGE_SIZE, M_FIRSTFIT | M_NOWAIT,
173 	    &addr) == 0) {
174 		i = (addr - sc->epc_base) / PAGE_SIZE;
175 		*epc = &sc->epc_pages[i];
176 		return (0);
177 	}
178 
179 	return (ENOMEM);
180 }
181 
182 static void
183 sgx_put_epc_page(struct sgx_softc *sc, struct epc_page *epc)
184 {
185 	vmem_addr_t addr;
186 
187 	if (epc == NULL)
188 		return;
189 
190 	addr = (epc->index * PAGE_SIZE) + sc->epc_base;
191 	vmem_free(sc->vmem_epc, addr, PAGE_SIZE);
192 }
193 
194 static int
195 sgx_va_slot_init_by_index(struct sgx_softc *sc, vm_object_t object,
196     uint64_t idx)
197 {
198 	struct epc_page *epc;
199 	vm_page_t page;
200 	vm_page_t p;
201 	int ret;
202 
203 	VM_OBJECT_ASSERT_WLOCKED(object);
204 
205 	p = vm_page_lookup(object, idx);
206 	if (p == NULL) {
207 		ret = sgx_get_epc_page(sc, &epc);
208 		if (ret) {
209 			dprintf("%s: No free EPC pages available.\n",
210 			    __func__);
211 			return (ret);
212 		}
213 
214 		mtx_lock(&sc->mtx_encls);
215 		sgx_epa((void *)epc->base);
216 		mtx_unlock(&sc->mtx_encls);
217 
218 		page = PHYS_TO_VM_PAGE(epc->phys);
219 
220 		page->valid = VM_PAGE_BITS_ALL;
221 		vm_page_insert(page, object, idx);
222 	}
223 
224 	return (0);
225 }
226 
227 static int
228 sgx_va_slot_init(struct sgx_softc *sc,
229     struct sgx_enclave *enclave,
230     uint64_t addr)
231 {
232 	vm_pindex_t pidx;
233 	uint64_t va_page_idx;
234 	uint64_t idx;
235 	vm_object_t object;
236 	int ret;
237 
238 	object = enclave->object;
239 
240 	VM_OBJECT_ASSERT_WLOCKED(object);
241 
242 	pidx = OFF_TO_IDX(addr);
243 
244 	va_page_idx = pidx / SGX_VA_PAGE_SLOTS;
245 	idx = - SGX_VA_PAGES_OFFS - va_page_idx;
246 
247 	ret = sgx_va_slot_init_by_index(sc, object, idx);
248 
249 	return (ret);
250 }
251 
252 static int
253 sgx_mem_find(struct sgx_softc *sc, uint64_t addr,
254     vm_map_entry_t *entry0, vm_object_t *object0)
255 {
256 	vm_map_t map;
257 	vm_map_entry_t entry;
258 	vm_object_t object;
259 
260 	map = &curproc->p_vmspace->vm_map;
261 
262 	vm_map_lock_read(map);
263 	if (!vm_map_lookup_entry(map, addr, &entry)) {
264 		vm_map_unlock_read(map);
265 		dprintf("%s: Can't find enclave.\n", __func__);
266 		return (EINVAL);
267 	}
268 
269 	object = entry->object.vm_object;
270 	if (object == NULL || object->handle == NULL) {
271 		vm_map_unlock_read(map);
272 		return (EINVAL);
273 	}
274 
275 	if (object->type != OBJT_MGTDEVICE ||
276 	    object->un_pager.devp.ops != &sgx_pg_ops) {
277 		vm_map_unlock_read(map);
278 		return (EINVAL);
279 	}
280 
281 	vm_object_reference(object);
282 
283 	*object0 = object;
284 	*entry0 = entry;
285 	vm_map_unlock_read(map);
286 
287 	return (0);
288 }
289 
290 static int
291 sgx_enclave_find(struct sgx_softc *sc, uint64_t addr,
292     struct sgx_enclave **encl)
293 {
294 	struct sgx_vm_handle *vmh;
295 	struct sgx_enclave *enclave;
296 	vm_map_entry_t entry;
297 	vm_object_t object;
298 	int ret;
299 
300 	ret = sgx_mem_find(sc, addr, &entry, &object);
301 	if (ret)
302 		return (ret);
303 
304 	vmh = object->handle;
305 	if (vmh == NULL) {
306 		vm_object_deallocate(object);
307 		return (EINVAL);
308 	}
309 
310 	enclave = vmh->enclave;
311 	if (enclave == NULL || enclave->object == NULL) {
312 		vm_object_deallocate(object);
313 		return (EINVAL);
314 	}
315 
316 	*encl = enclave;
317 
318 	return (0);
319 }
320 
321 static int
322 sgx_enclave_alloc(struct sgx_softc *sc, struct secs *secs,
323     struct sgx_enclave **enclave0)
324 {
325 	struct sgx_enclave *enclave;
326 
327 	enclave = malloc(sizeof(struct sgx_enclave),
328 	    M_SGX, M_WAITOK | M_ZERO);
329 
330 	enclave->base = secs->base;
331 	enclave->size = secs->size;
332 
333 	*enclave0 = enclave;
334 
335 	return (0);
336 }
337 
338 static void
339 sgx_epc_page_remove(struct sgx_softc *sc,
340     struct epc_page *epc)
341 {
342 
343 	mtx_lock(&sc->mtx_encls);
344 	sgx_eremove((void *)epc->base);
345 	mtx_unlock(&sc->mtx_encls);
346 }
347 
348 static void
349 sgx_page_remove(struct sgx_softc *sc, vm_page_t p)
350 {
351 	struct epc_page *epc;
352 	vm_paddr_t pa;
353 	uint64_t offs;
354 
355 	(void)vm_page_remove(p);
356 
357 	dprintf("%s: p->pidx %ld\n", __func__, p->pindex);
358 
359 	pa = VM_PAGE_TO_PHYS(p);
360 	epc = &sc->epc_pages[0];
361 	offs = (pa - epc->phys) / PAGE_SIZE;
362 	epc = &sc->epc_pages[offs];
363 
364 	sgx_epc_page_remove(sc, epc);
365 	sgx_put_epc_page(sc, epc);
366 }
367 
368 static void
369 sgx_enclave_remove(struct sgx_softc *sc,
370     struct sgx_enclave *enclave)
371 {
372 	vm_object_t object;
373 	vm_page_t p, p_secs, p_next;
374 
375 	mtx_lock(&sc->mtx);
376 	TAILQ_REMOVE(&sc->enclaves, enclave, next);
377 	mtx_unlock(&sc->mtx);
378 
379 	object = enclave->object;
380 
381 	VM_OBJECT_WLOCK(object);
382 
383 	/*
384 	 * First remove all the pages except SECS,
385 	 * then remove SECS page.
386 	 */
387 restart:
388 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
389 		if (p->pindex == SGX_SECS_VM_OBJECT_INDEX)
390 			continue;
391 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0)
392 			goto restart;
393 		sgx_page_remove(sc, p);
394 	}
395 	p_secs = vm_page_grab(object, SGX_SECS_VM_OBJECT_INDEX,
396 	    VM_ALLOC_NOCREAT);
397 	/* Now remove SECS page */
398 	if (p_secs != NULL)
399 		sgx_page_remove(sc, p_secs);
400 
401 	KASSERT(TAILQ_EMPTY(&object->memq) == 1, ("not empty"));
402 	KASSERT(object->resident_page_count == 0, ("count"));
403 
404 	VM_OBJECT_WUNLOCK(object);
405 }
406 
407 static int
408 sgx_measure_page(struct sgx_softc *sc, struct epc_page *secs,
409     struct epc_page *epc, uint16_t mrmask)
410 {
411 	int i, j;
412 	int ret;
413 
414 	mtx_lock(&sc->mtx_encls);
415 
416 	for (i = 0, j = 1; i < PAGE_SIZE; i += 0x100, j <<= 1) {
417 		if (!(j & mrmask))
418 			continue;
419 
420 		ret = sgx_eextend((void *)secs->base,
421 		    (void *)(epc->base + i));
422 		if (ret == SGX_EFAULT) {
423 			mtx_unlock(&sc->mtx_encls);
424 			return (ret);
425 		}
426 	}
427 
428 	mtx_unlock(&sc->mtx_encls);
429 
430 	return (0);
431 }
432 
433 static int
434 sgx_secs_validate(struct sgx_softc *sc, struct secs *secs)
435 {
436 	struct secs_attr *attr;
437 	int i;
438 
439 	if (secs->size == 0)
440 		return (EINVAL);
441 
442 	/* BASEADDR must be naturally aligned on an SECS.SIZE boundary. */
443 	if (secs->base & (secs->size - 1))
444 		return (EINVAL);
445 
446 	/* SECS.SIZE must be at least 2 pages. */
447 	if (secs->size < 2 * PAGE_SIZE)
448 		return (EINVAL);
449 
450 	if ((secs->size & (secs->size - 1)) != 0)
451 		return (EINVAL);
452 
453 	attr = &secs->attributes;
454 
455 	if (attr->reserved1 != 0 ||
456 	    attr->reserved2 != 0 ||
457 	    attr->reserved3 != 0)
458 		return (EINVAL);
459 
460 	for (i = 0; i < SECS_ATTR_RSV4_SIZE; i++)
461 		if (attr->reserved4[i])
462 			return (EINVAL);
463 
464 	/*
465 	 * Intel® Software Guard Extensions Programming Reference
466 	 * 6.7.2 Relevant Fields in Various Data Structures
467 	 * 6.7.2.1 SECS.ATTRIBUTES.XFRM
468 	 * XFRM[1:0] must be set to 0x3.
469 	 */
470 	if ((attr->xfrm & 0x3) != 0x3)
471 		return (EINVAL);
472 
473 	if (!attr->mode64bit)
474 		return (EINVAL);
475 
476 	if (secs->size > sc->enclave_size_max)
477 		return (EINVAL);
478 
479 	for (i = 0; i < SECS_RSV1_SIZE; i++)
480 		if (secs->reserved1[i])
481 			return (EINVAL);
482 
483 	for (i = 0; i < SECS_RSV2_SIZE; i++)
484 		if (secs->reserved2[i])
485 			return (EINVAL);
486 
487 	for (i = 0; i < SECS_RSV3_SIZE; i++)
488 		if (secs->reserved3[i])
489 			return (EINVAL);
490 
491 	for (i = 0; i < SECS_RSV4_SIZE; i++)
492 		if (secs->reserved4[i])
493 			return (EINVAL);
494 
495 	return (0);
496 }
497 
498 static int
499 sgx_tcs_validate(struct tcs *tcs)
500 {
501 	int i;
502 
503 	if ((tcs->flags) ||
504 	    (tcs->ossa & (PAGE_SIZE - 1)) ||
505 	    (tcs->ofsbasgx & (PAGE_SIZE - 1)) ||
506 	    (tcs->ogsbasgx & (PAGE_SIZE - 1)) ||
507 	    ((tcs->fslimit & 0xfff) != 0xfff) ||
508 	    ((tcs->gslimit & 0xfff) != 0xfff))
509 		return (EINVAL);
510 
511 	for (i = 0; i < nitems(tcs->reserved3); i++)
512 		if (tcs->reserved3[i])
513 			return (EINVAL);
514 
515 	return (0);
516 }
517 
518 static void
519 sgx_tcs_dump(struct sgx_softc *sc, struct tcs *t)
520 {
521 
522 	dprintf("t->flags %lx\n", t->flags);
523 	dprintf("t->ossa %lx\n", t->ossa);
524 	dprintf("t->cssa %x\n", t->cssa);
525 	dprintf("t->nssa %x\n", t->nssa);
526 	dprintf("t->oentry %lx\n", t->oentry);
527 	dprintf("t->ofsbasgx %lx\n", t->ofsbasgx);
528 	dprintf("t->ogsbasgx %lx\n", t->ogsbasgx);
529 	dprintf("t->fslimit %x\n", t->fslimit);
530 	dprintf("t->gslimit %x\n", t->gslimit);
531 }
532 
533 static int
534 sgx_pg_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
535     vm_ooffset_t foff, struct ucred *cred, u_short *color)
536 {
537 	struct sgx_vm_handle *vmh;
538 
539 	vmh = handle;
540 	if (vmh == NULL) {
541 		dprintf("%s: vmh not found.\n", __func__);
542 		return (0);
543 	}
544 
545 	dprintf("%s: vmh->base %lx foff 0x%lx size 0x%lx\n",
546 	    __func__, vmh->base, foff, size);
547 
548 	return (0);
549 }
550 
551 static void
552 sgx_pg_dtor(void *handle)
553 {
554 	struct sgx_vm_handle *vmh;
555 	struct sgx_softc *sc;
556 
557 	vmh = handle;
558 	if (vmh == NULL) {
559 		dprintf("%s: vmh not found.\n", __func__);
560 		return;
561 	}
562 
563 	sc = vmh->sc;
564 	if (sc == NULL) {
565 		dprintf("%s: sc is NULL\n", __func__);
566 		return;
567 	}
568 
569 	if (vmh->enclave == NULL) {
570 		dprintf("%s: Enclave not found.\n", __func__);
571 		return;
572 	}
573 
574 	sgx_enclave_remove(sc, vmh->enclave);
575 
576 	free(vmh->enclave, M_SGX);
577 	free(vmh, M_SGX);
578 }
579 
580 static int
581 sgx_pg_fault(vm_object_t object, vm_ooffset_t offset,
582     int prot, vm_page_t *mres)
583 {
584 
585 	/*
586 	 * The purpose of this trivial handler is to handle the race
587 	 * when user tries to access mmaped region before or during
588 	 * enclave creation ioctl calls.
589 	 */
590 
591 	dprintf("%s: offset 0x%lx\n", __func__, offset);
592 
593 	return (VM_PAGER_FAIL);
594 }
595 
596 static struct cdev_pager_ops sgx_pg_ops = {
597 	.cdev_pg_ctor = sgx_pg_ctor,
598 	.cdev_pg_dtor = sgx_pg_dtor,
599 	.cdev_pg_fault = sgx_pg_fault,
600 };
601 
602 static void
603 sgx_insert_epc_page_by_index(vm_page_t page, vm_object_t object,
604     vm_pindex_t pidx)
605 {
606 
607 	VM_OBJECT_ASSERT_WLOCKED(object);
608 
609 	page->valid = VM_PAGE_BITS_ALL;
610 	vm_page_insert(page, object, pidx);
611 }
612 
613 static void
614 sgx_insert_epc_page(struct sgx_enclave *enclave,
615     struct epc_page *epc, uint64_t addr)
616 {
617 	vm_pindex_t pidx;
618 	vm_page_t page;
619 
620 	VM_OBJECT_ASSERT_WLOCKED(enclave->object);
621 
622 	pidx = OFF_TO_IDX(addr);
623 	page = PHYS_TO_VM_PAGE(epc->phys);
624 
625 	sgx_insert_epc_page_by_index(page, enclave->object, pidx);
626 }
627 
628 static int
629 sgx_ioctl_create(struct sgx_softc *sc, struct sgx_enclave_create *param)
630 {
631 	struct sgx_vm_handle *vmh;
632 	vm_map_entry_t entry;
633 	vm_page_t p;
634 	struct page_info pginfo;
635 	struct secinfo secinfo;
636 	struct sgx_enclave *enclave;
637 	struct epc_page *epc;
638 	struct secs *secs;
639 	vm_object_t object;
640 	vm_page_t page;
641 	int ret;
642 
643 	epc = NULL;
644 	secs = NULL;
645 	enclave = NULL;
646 	object = NULL;
647 
648 	/* SGX Enclave Control Structure (SECS) */
649 	secs = malloc(PAGE_SIZE, M_SGX, M_WAITOK | M_ZERO);
650 	ret = copyin((void *)param->src, secs, sizeof(struct secs));
651 	if (ret) {
652 		dprintf("%s: Can't copy SECS.\n", __func__);
653 		goto error;
654 	}
655 
656 	ret = sgx_secs_validate(sc, secs);
657 	if (ret) {
658 		dprintf("%s: SECS validation failed.\n", __func__);
659 		goto error;
660 	}
661 
662 	ret = sgx_mem_find(sc, secs->base, &entry, &object);
663 	if (ret) {
664 		dprintf("%s: Can't find vm_map.\n", __func__);
665 		goto error;
666 	}
667 
668 	vmh = object->handle;
669 	if (!vmh) {
670 		dprintf("%s: Can't find vmh.\n", __func__);
671 		ret = ENXIO;
672 		goto error;
673 	}
674 
675 	dprintf("%s: entry start %lx offset %lx\n",
676 	    __func__, entry->start, entry->offset);
677 	vmh->base = (entry->start - entry->offset);
678 
679 	ret = sgx_enclave_alloc(sc, secs, &enclave);
680 	if (ret) {
681 		dprintf("%s: Can't alloc enclave.\n", __func__);
682 		goto error;
683 	}
684 	enclave->object = object;
685 	enclave->vmh = vmh;
686 
687 	memset(&secinfo, 0, sizeof(struct secinfo));
688 	memset(&pginfo, 0, sizeof(struct page_info));
689 	pginfo.linaddr = 0;
690 	pginfo.srcpge = (uint64_t)secs;
691 	pginfo.secinfo = &secinfo;
692 	pginfo.secs = 0;
693 
694 	ret = sgx_get_epc_page(sc, &epc);
695 	if (ret) {
696 		dprintf("%s: Failed to get free epc page.\n", __func__);
697 		goto error;
698 	}
699 	enclave->secs_epc_page = epc;
700 
701 	VM_OBJECT_WLOCK(object);
702 	p = vm_page_lookup(object, SGX_SECS_VM_OBJECT_INDEX);
703 	if (p) {
704 		VM_OBJECT_WUNLOCK(object);
705 		/* SECS page already added. */
706 		ret = ENXIO;
707 		goto error;
708 	}
709 
710 	ret = sgx_va_slot_init_by_index(sc, object,
711 	    - SGX_VA_PAGES_OFFS - SGX_SECS_VM_OBJECT_INDEX);
712 	if (ret) {
713 		VM_OBJECT_WUNLOCK(object);
714 		dprintf("%s: Can't init va slot.\n", __func__);
715 		goto error;
716 	}
717 
718 	mtx_lock(&sc->mtx);
719 	if ((sc->state & SGX_STATE_RUNNING) == 0) {
720 		mtx_unlock(&sc->mtx);
721 		/* Remove VA page that was just created for SECS page. */
722 		p = vm_page_grab(enclave->object,
723 		    - SGX_VA_PAGES_OFFS - SGX_SECS_VM_OBJECT_INDEX,
724 		    VM_ALLOC_NOCREAT);
725 		sgx_page_remove(sc, p);
726 		VM_OBJECT_WUNLOCK(object);
727 		goto error;
728 	}
729 	mtx_lock(&sc->mtx_encls);
730 	ret = sgx_ecreate(&pginfo, (void *)epc->base);
731 	mtx_unlock(&sc->mtx_encls);
732 	if (ret == SGX_EFAULT) {
733 		dprintf("%s: gp fault\n", __func__);
734 		mtx_unlock(&sc->mtx);
735 		/* Remove VA page that was just created for SECS page. */
736 		p = vm_page_grab(enclave->object,
737 		    - SGX_VA_PAGES_OFFS - SGX_SECS_VM_OBJECT_INDEX,
738 		    VM_ALLOC_NOCREAT);
739 		sgx_page_remove(sc, p);
740 		VM_OBJECT_WUNLOCK(object);
741 		goto error;
742 	}
743 
744 	TAILQ_INSERT_TAIL(&sc->enclaves, enclave, next);
745 	mtx_unlock(&sc->mtx);
746 
747 	vmh->enclave = enclave;
748 
749 	page = PHYS_TO_VM_PAGE(epc->phys);
750 	sgx_insert_epc_page_by_index(page, enclave->object,
751 	    SGX_SECS_VM_OBJECT_INDEX);
752 
753 	VM_OBJECT_WUNLOCK(object);
754 
755 	/* Release the reference. */
756 	vm_object_deallocate(object);
757 
758 	free(secs, M_SGX);
759 
760 	return (0);
761 
762 error:
763 	free(secs, M_SGX);
764 	sgx_put_epc_page(sc, epc);
765 	free(enclave, M_SGX);
766 	vm_object_deallocate(object);
767 
768 	return (ret);
769 }
770 
771 static int
772 sgx_ioctl_add_page(struct sgx_softc *sc,
773     struct sgx_enclave_add_page *addp)
774 {
775 	struct epc_page *secs_epc_page;
776 	struct sgx_enclave *enclave;
777 	struct sgx_vm_handle *vmh;
778 	struct epc_page *epc;
779 	struct page_info pginfo;
780 	struct secinfo secinfo;
781 	vm_object_t object;
782 	void *tmp_vaddr;
783 	uint64_t page_type;
784 	struct tcs *t;
785 	uint64_t addr;
786 	uint64_t pidx;
787 	vm_page_t p;
788 	int ret;
789 
790 	tmp_vaddr = NULL;
791 	epc = NULL;
792 	object = NULL;
793 
794 	/* Find and get reference to VM object. */
795 	ret = sgx_enclave_find(sc, addp->addr, &enclave);
796 	if (ret) {
797 		dprintf("%s: Failed to find enclave.\n", __func__);
798 		goto error;
799 	}
800 
801 	object = enclave->object;
802 	KASSERT(object != NULL, ("vm object is NULL\n"));
803 	vmh = object->handle;
804 
805 	ret = sgx_get_epc_page(sc, &epc);
806 	if (ret) {
807 		dprintf("%s: Failed to get free epc page.\n", __func__);
808 		goto error;
809 	}
810 
811 	memset(&secinfo, 0, sizeof(struct secinfo));
812 	ret = copyin((void *)addp->secinfo, &secinfo,
813 	    sizeof(struct secinfo));
814 	if (ret) {
815 		dprintf("%s: Failed to copy secinfo.\n", __func__);
816 		goto error;
817 	}
818 
819 	tmp_vaddr = malloc(PAGE_SIZE, M_SGX, M_WAITOK | M_ZERO);
820 	ret = copyin((void *)addp->src, tmp_vaddr, PAGE_SIZE);
821 	if (ret) {
822 		dprintf("%s: Failed to copy page.\n", __func__);
823 		goto error;
824 	}
825 
826 	page_type = (secinfo.flags & SECINFO_FLAGS_PT_M) >>
827 	    SECINFO_FLAGS_PT_S;
828 	if (page_type != SGX_PT_TCS && page_type != SGX_PT_REG) {
829 		dprintf("%s: page can't be added.\n", __func__);
830 		goto error;
831 	}
832 	if (page_type == SGX_PT_TCS) {
833 		t = (struct tcs *)tmp_vaddr;
834 		ret = sgx_tcs_validate(t);
835 		if (ret) {
836 			dprintf("%s: TCS page validation failed.\n",
837 			    __func__);
838 			goto error;
839 		}
840 		sgx_tcs_dump(sc, t);
841 	}
842 
843 	addr = (addp->addr - vmh->base);
844 	pidx = OFF_TO_IDX(addr);
845 
846 	VM_OBJECT_WLOCK(object);
847 	p = vm_page_lookup(object, pidx);
848 	if (p) {
849 		VM_OBJECT_WUNLOCK(object);
850 		/* Page already added. */
851 		ret = ENXIO;
852 		goto error;
853 	}
854 
855 	ret = sgx_va_slot_init(sc, enclave, addr);
856 	if (ret) {
857 		VM_OBJECT_WUNLOCK(object);
858 		dprintf("%s: Can't init va slot.\n", __func__);
859 		goto error;
860 	}
861 
862 	secs_epc_page = enclave->secs_epc_page;
863 	memset(&pginfo, 0, sizeof(struct page_info));
864 	pginfo.linaddr = (uint64_t)addp->addr;
865 	pginfo.srcpge = (uint64_t)tmp_vaddr;
866 	pginfo.secinfo = &secinfo;
867 	pginfo.secs = (uint64_t)secs_epc_page->base;
868 
869 	mtx_lock(&sc->mtx_encls);
870 	ret = sgx_eadd(&pginfo, (void *)epc->base);
871 	if (ret == SGX_EFAULT) {
872 		dprintf("%s: gp fault on eadd\n", __func__);
873 		mtx_unlock(&sc->mtx_encls);
874 		VM_OBJECT_WUNLOCK(object);
875 		goto error;
876 	}
877 	mtx_unlock(&sc->mtx_encls);
878 
879 	ret = sgx_measure_page(sc, enclave->secs_epc_page, epc, addp->mrmask);
880 	if (ret == SGX_EFAULT) {
881 		dprintf("%s: gp fault on eextend\n", __func__);
882 		sgx_epc_page_remove(sc, epc);
883 		VM_OBJECT_WUNLOCK(object);
884 		goto error;
885 	}
886 
887 	sgx_insert_epc_page(enclave, epc, addr);
888 
889 	VM_OBJECT_WUNLOCK(object);
890 
891 	/* Release the reference. */
892 	vm_object_deallocate(object);
893 
894 	free(tmp_vaddr, M_SGX);
895 
896 	return (0);
897 
898 error:
899 	free(tmp_vaddr, M_SGX);
900 	sgx_put_epc_page(sc, epc);
901 	vm_object_deallocate(object);
902 
903 	return (ret);
904 }
905 
906 static int
907 sgx_ioctl_init(struct sgx_softc *sc, struct sgx_enclave_init *initp)
908 {
909 	struct epc_page *secs_epc_page;
910 	struct sgx_enclave *enclave;
911 	struct thread *td;
912 	void *tmp_vaddr;
913 	void *einittoken;
914 	void *sigstruct;
915 	vm_object_t object;
916 	int retry;
917 	int ret;
918 
919 	td = curthread;
920 	tmp_vaddr = NULL;
921 	object = NULL;
922 
923 	dprintf("%s: addr %lx, sigstruct %lx, einittoken %lx\n",
924 	    __func__, initp->addr, initp->sigstruct, initp->einittoken);
925 
926 	/* Find and get reference to VM object. */
927 	ret = sgx_enclave_find(sc, initp->addr, &enclave);
928 	if (ret) {
929 		dprintf("%s: Failed to find enclave.\n", __func__);
930 		goto error;
931 	}
932 
933 	object = enclave->object;
934 
935 	tmp_vaddr = malloc(PAGE_SIZE, M_SGX, M_WAITOK | M_ZERO);
936 	sigstruct = tmp_vaddr;
937 	einittoken = (void *)((uint64_t)sigstruct + PAGE_SIZE / 2);
938 
939 	ret = copyin((void *)initp->sigstruct, sigstruct,
940 	    SGX_SIGSTRUCT_SIZE);
941 	if (ret) {
942 		dprintf("%s: Failed to copy SIGSTRUCT page.\n", __func__);
943 		goto error;
944 	}
945 
946 	ret = copyin((void *)initp->einittoken, einittoken,
947 	    SGX_EINITTOKEN_SIZE);
948 	if (ret) {
949 		dprintf("%s: Failed to copy EINITTOKEN page.\n", __func__);
950 		goto error;
951 	}
952 
953 	secs_epc_page = enclave->secs_epc_page;
954 	retry = 16;
955 	do {
956 		mtx_lock(&sc->mtx_encls);
957 		ret = sgx_einit(sigstruct, (void *)secs_epc_page->base,
958 		    einittoken);
959 		mtx_unlock(&sc->mtx_encls);
960 		dprintf("%s: sgx_einit returned %d\n", __func__, ret);
961 	} while (ret == SGX_UNMASKED_EVENT && retry--);
962 
963 	if (ret) {
964 		dprintf("%s: Failed init enclave: %d\n", __func__, ret);
965 		td->td_retval[0] = ret;
966 		ret = 0;
967 	}
968 
969 error:
970 	free(tmp_vaddr, M_SGX);
971 
972 	/* Release the reference. */
973 	vm_object_deallocate(object);
974 
975 	return (ret);
976 }
977 
978 static int
979 sgx_ioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags,
980     struct thread *td)
981 {
982 	struct sgx_enclave_add_page *addp;
983 	struct sgx_enclave_create *param;
984 	struct sgx_enclave_init *initp;
985 	struct sgx_softc *sc;
986 	int ret;
987 	int len;
988 
989 	sc = &sgx_sc;
990 
991 	len = IOCPARM_LEN(cmd);
992 
993 	dprintf("%s: cmd %lx, addr %lx, len %d\n",
994 	    __func__, cmd, (uint64_t)addr, len);
995 
996 	if (len > SGX_IOCTL_MAX_DATA_LEN)
997 		return (EINVAL);
998 
999 	switch (cmd) {
1000 	case SGX_IOC_ENCLAVE_CREATE:
1001 		param = (struct sgx_enclave_create *)addr;
1002 		ret = sgx_ioctl_create(sc, param);
1003 		break;
1004 	case SGX_IOC_ENCLAVE_ADD_PAGE:
1005 		addp = (struct sgx_enclave_add_page *)addr;
1006 		ret = sgx_ioctl_add_page(sc, addp);
1007 		break;
1008 	case SGX_IOC_ENCLAVE_INIT:
1009 		initp = (struct sgx_enclave_init *)addr;
1010 		ret = sgx_ioctl_init(sc, initp);
1011 		break;
1012 	default:
1013 		return (EINVAL);
1014 	}
1015 
1016 	return (ret);
1017 }
1018 
1019 static int
1020 sgx_mmap_single(struct cdev *cdev, vm_ooffset_t *offset,
1021     vm_size_t mapsize, struct vm_object **objp, int nprot)
1022 {
1023 	struct sgx_vm_handle *vmh;
1024 	struct sgx_softc *sc;
1025 
1026 	sc = &sgx_sc;
1027 
1028 	dprintf("%s: mapsize 0x%lx, offset %lx\n",
1029 	    __func__, mapsize, *offset);
1030 
1031 	vmh = malloc(sizeof(struct sgx_vm_handle),
1032 	    M_SGX, M_WAITOK | M_ZERO);
1033 	vmh->sc = sc;
1034 	vmh->size = mapsize;
1035 	vmh->mem = cdev_pager_allocate(vmh, OBJT_MGTDEVICE, &sgx_pg_ops,
1036 	    mapsize, nprot, *offset, NULL);
1037 	if (vmh->mem == NULL) {
1038 		free(vmh, M_SGX);
1039 		return (ENOMEM);
1040 	}
1041 
1042 	VM_OBJECT_WLOCK(vmh->mem);
1043 	vm_object_set_flag(vmh->mem, OBJ_PG_DTOR);
1044 	VM_OBJECT_WUNLOCK(vmh->mem);
1045 
1046 	*objp = vmh->mem;
1047 
1048 	return (0);
1049 }
1050 
1051 static struct cdevsw sgx_cdevsw = {
1052 	.d_version =		D_VERSION,
1053 	.d_ioctl =		sgx_ioctl,
1054 	.d_mmap_single =	sgx_mmap_single,
1055 	.d_name =		"Intel SGX",
1056 };
1057 
1058 static int
1059 sgx_get_epc_area(struct sgx_softc *sc)
1060 {
1061 	vm_offset_t epc_base_vaddr;
1062 	u_int cp[4];
1063 	int error;
1064 	int i;
1065 
1066 	cpuid_count(SGX_CPUID, 0x2, cp);
1067 
1068 	sc->epc_base = ((uint64_t)(cp[1] & 0xfffff) << 32) +
1069 	    (cp[0] & 0xfffff000);
1070 	sc->epc_size = ((uint64_t)(cp[3] & 0xfffff) << 32) +
1071 	    (cp[2] & 0xfffff000);
1072 	sc->npages = sc->epc_size / SGX_PAGE_SIZE;
1073 
1074 	if (sc->epc_size == 0 || sc->epc_base == 0) {
1075 		printf("%s: Incorrect EPC data: EPC base %lx, size %lu\n",
1076 		    __func__, sc->epc_base, sc->epc_size);
1077 		return (EINVAL);
1078 	}
1079 
1080 	if (cp[3] & 0xffff)
1081 		sc->enclave_size_max = (1 << ((cp[3] >> 8) & 0xff));
1082 	else
1083 		sc->enclave_size_max = SGX_ENCL_SIZE_MAX_DEF;
1084 
1085 	epc_base_vaddr = (vm_offset_t)pmap_mapdev_attr(sc->epc_base,
1086 	    sc->epc_size, VM_MEMATTR_DEFAULT);
1087 
1088 	sc->epc_pages = malloc(sizeof(struct epc_page) * sc->npages,
1089 	    M_DEVBUF, M_WAITOK | M_ZERO);
1090 
1091 	for (i = 0; i < sc->npages; i++) {
1092 		sc->epc_pages[i].base = epc_base_vaddr + SGX_PAGE_SIZE * i;
1093 		sc->epc_pages[i].phys = sc->epc_base + SGX_PAGE_SIZE * i;
1094 		sc->epc_pages[i].index = i;
1095 	}
1096 
1097 	sc->vmem_epc = vmem_create("SGX EPC", sc->epc_base, sc->epc_size,
1098 	    PAGE_SIZE, PAGE_SIZE, M_FIRSTFIT | M_WAITOK);
1099 	if (sc->vmem_epc == NULL) {
1100 		printf("%s: Can't create vmem arena.\n", __func__);
1101 		free(sc->epc_pages, M_SGX);
1102 		return (EINVAL);
1103 	}
1104 
1105 	error = vm_phys_fictitious_reg_range(sc->epc_base,
1106 	    sc->epc_base + sc->epc_size, VM_MEMATTR_DEFAULT);
1107 	if (error) {
1108 		printf("%s: Can't register fictitious space.\n", __func__);
1109 		free(sc->epc_pages, M_SGX);
1110 		return (EINVAL);
1111 	}
1112 
1113 	return (0);
1114 }
1115 
1116 static void
1117 sgx_put_epc_area(struct sgx_softc *sc)
1118 {
1119 
1120 	vm_phys_fictitious_unreg_range(sc->epc_base,
1121 	    sc->epc_base + sc->epc_size);
1122 
1123 	free(sc->epc_pages, M_SGX);
1124 }
1125 
1126 static int
1127 sgx_load(void)
1128 {
1129 	struct sgx_softc *sc;
1130 	int error;
1131 
1132 	sc = &sgx_sc;
1133 
1134 	if ((cpu_stdext_feature & CPUID_STDEXT_SGX) == 0)
1135 		return (ENXIO);
1136 
1137 	error = sgx_get_epc_area(sc);
1138 	if (error) {
1139 		printf("%s: Failed to get Processor Reserved Memory area.\n",
1140 		    __func__);
1141 		return (ENXIO);
1142 	}
1143 
1144 	mtx_init(&sc->mtx_encls, "SGX ENCLS", NULL, MTX_DEF);
1145 	mtx_init(&sc->mtx, "SGX driver", NULL, MTX_DEF);
1146 
1147 	TAILQ_INIT(&sc->enclaves);
1148 
1149 	sc->sgx_cdev = make_dev(&sgx_cdevsw, 0, UID_ROOT, GID_WHEEL,
1150 	    0600, "isgx");
1151 
1152 	sc->state |= SGX_STATE_RUNNING;
1153 
1154 	printf("SGX initialized: EPC base 0x%lx size %ld (%d pages)\n",
1155 	    sc->epc_base, sc->epc_size, sc->npages);
1156 
1157 	return (0);
1158 }
1159 
1160 static int
1161 sgx_unload(void)
1162 {
1163 	struct sgx_softc *sc;
1164 
1165 	sc = &sgx_sc;
1166 
1167 	if ((sc->state & SGX_STATE_RUNNING) == 0)
1168 		return (0);
1169 
1170 	mtx_lock(&sc->mtx);
1171 	if (!TAILQ_EMPTY(&sc->enclaves)) {
1172 		mtx_unlock(&sc->mtx);
1173 		return (EBUSY);
1174 	}
1175 	sc->state &= ~SGX_STATE_RUNNING;
1176 	mtx_unlock(&sc->mtx);
1177 
1178 	destroy_dev(sc->sgx_cdev);
1179 
1180 	vmem_destroy(sc->vmem_epc);
1181 	sgx_put_epc_area(sc);
1182 
1183 	mtx_destroy(&sc->mtx_encls);
1184 	mtx_destroy(&sc->mtx);
1185 
1186 	return (0);
1187 }
1188 
1189 static int
1190 sgx_handler(module_t mod, int what, void *arg)
1191 {
1192 	int error;
1193 
1194 	switch (what) {
1195 	case MOD_LOAD:
1196 		error = sgx_load();
1197 		break;
1198 	case MOD_UNLOAD:
1199 		error = sgx_unload();
1200 		break;
1201 	default:
1202 		error = 0;
1203 		break;
1204 	}
1205 
1206 	return (error);
1207 }
1208 
1209 static moduledata_t sgx_kmod = {
1210 	"sgx",
1211 	sgx_handler,
1212 	NULL
1213 };
1214 
1215 DECLARE_MODULE(sgx, sgx_kmod, SI_SUB_LAST, SI_ORDER_ANY);
1216 MODULE_VERSION(sgx, 1);
1217