xref: /linux/arch/x86/virt/vmx/tdx/tdx.c (revision 2672031b20f6681514bef14ddcfe8c62c2757d11)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright(c) 2023 Intel Corporation.
4  *
5  * Intel Trusted Domain Extensions (TDX) support
6  */
7 
8 #define pr_fmt(fmt)	"virt/tdx: " fmt
9 
10 #include <linux/types.h>
11 #include <linux/cache.h>
12 #include <linux/init.h>
13 #include <linux/errno.h>
14 #include <linux/printk.h>
15 #include <linux/cpu.h>
16 #include <linux/spinlock.h>
17 #include <linux/percpu-defs.h>
18 #include <linux/mutex.h>
19 #include <linux/list.h>
20 #include <linux/memblock.h>
21 #include <linux/memory.h>
22 #include <linux/minmax.h>
23 #include <linux/sizes.h>
24 #include <linux/pfn.h>
25 #include <linux/align.h>
26 #include <linux/sort.h>
27 #include <linux/log2.h>
28 #include <linux/acpi.h>
29 #include <linux/suspend.h>
30 #include <linux/acpi.h>
31 #include <asm/page.h>
32 #include <asm/special_insns.h>
33 #include <asm/msr-index.h>
34 #include <asm/msr.h>
35 #include <asm/cpufeature.h>
36 #include <asm/tdx.h>
37 #include <asm/intel-family.h>
38 #include <asm/processor.h>
39 #include <asm/mce.h>
40 #include "tdx.h"
41 
42 static u32 tdx_global_keyid __ro_after_init;
43 static u32 tdx_guest_keyid_start __ro_after_init;
44 static u32 tdx_nr_guest_keyids __ro_after_init;
45 
46 static DEFINE_PER_CPU(bool, tdx_lp_initialized);
47 
48 static struct tdmr_info_list tdx_tdmr_list;
49 
50 static enum tdx_module_status_t tdx_module_status;
51 static DEFINE_MUTEX(tdx_module_lock);
52 
53 /* All TDX-usable memory regions.  Protected by mem_hotplug_lock. */
54 static LIST_HEAD(tdx_memlist);
55 
56 typedef void (*sc_err_func_t)(u64 fn, u64 err, struct tdx_module_args *args);
57 
58 static inline void seamcall_err(u64 fn, u64 err, struct tdx_module_args *args)
59 {
60 	pr_err("SEAMCALL (0x%016llx) failed: 0x%016llx\n", fn, err);
61 }
62 
63 static inline void seamcall_err_ret(u64 fn, u64 err,
64 				    struct tdx_module_args *args)
65 {
66 	seamcall_err(fn, err, args);
67 	pr_err("RCX 0x%016llx RDX 0x%016llx R08 0x%016llx\n",
68 			args->rcx, args->rdx, args->r8);
69 	pr_err("R09 0x%016llx R10 0x%016llx R11 0x%016llx\n",
70 			args->r9, args->r10, args->r11);
71 }
72 
73 static inline int sc_retry_prerr(sc_func_t func, sc_err_func_t err_func,
74 				 u64 fn, struct tdx_module_args *args)
75 {
76 	u64 sret = sc_retry(func, fn, args);
77 
78 	if (sret == TDX_SUCCESS)
79 		return 0;
80 
81 	if (sret == TDX_SEAMCALL_VMFAILINVALID)
82 		return -ENODEV;
83 
84 	if (sret == TDX_SEAMCALL_GP)
85 		return -EOPNOTSUPP;
86 
87 	if (sret == TDX_SEAMCALL_UD)
88 		return -EACCES;
89 
90 	err_func(fn, sret, args);
91 	return -EIO;
92 }
93 
94 #define seamcall_prerr(__fn, __args)						\
95 	sc_retry_prerr(__seamcall, seamcall_err, (__fn), (__args))
96 
97 #define seamcall_prerr_ret(__fn, __args)					\
98 	sc_retry_prerr(__seamcall_ret, seamcall_err_ret, (__fn), (__args))
99 
100 /*
101  * Do the module global initialization once and return its result.
102  * It can be done on any cpu.  It's always called with interrupts
103  * disabled.
104  */
105 static int try_init_module_global(void)
106 {
107 	struct tdx_module_args args = {};
108 	static DEFINE_RAW_SPINLOCK(sysinit_lock);
109 	static bool sysinit_done;
110 	static int sysinit_ret;
111 
112 	lockdep_assert_irqs_disabled();
113 
114 	raw_spin_lock(&sysinit_lock);
115 
116 	if (sysinit_done)
117 		goto out;
118 
119 	/* RCX is module attributes and all bits are reserved */
120 	args.rcx = 0;
121 	sysinit_ret = seamcall_prerr(TDH_SYS_INIT, &args);
122 
123 	/*
124 	 * The first SEAMCALL also detects the TDX module, thus
125 	 * it can fail due to the TDX module is not loaded.
126 	 * Dump message to let the user know.
127 	 */
128 	if (sysinit_ret == -ENODEV)
129 		pr_err("module not loaded\n");
130 
131 	sysinit_done = true;
132 out:
133 	raw_spin_unlock(&sysinit_lock);
134 	return sysinit_ret;
135 }
136 
137 /**
138  * tdx_cpu_enable - Enable TDX on local cpu
139  *
140  * Do one-time TDX module per-cpu initialization SEAMCALL (and TDX module
141  * global initialization SEAMCALL if not done) on local cpu to make this
142  * cpu be ready to run any other SEAMCALLs.
143  *
144  * Always call this function via IPI function calls.
145  *
146  * Return 0 on success, otherwise errors.
147  */
148 int tdx_cpu_enable(void)
149 {
150 	struct tdx_module_args args = {};
151 	int ret;
152 
153 	if (!boot_cpu_has(X86_FEATURE_TDX_HOST_PLATFORM))
154 		return -ENODEV;
155 
156 	lockdep_assert_irqs_disabled();
157 
158 	if (__this_cpu_read(tdx_lp_initialized))
159 		return 0;
160 
161 	/*
162 	 * The TDX module global initialization is the very first step
163 	 * to enable TDX.  Need to do it first (if hasn't been done)
164 	 * before the per-cpu initialization.
165 	 */
166 	ret = try_init_module_global();
167 	if (ret)
168 		return ret;
169 
170 	ret = seamcall_prerr(TDH_SYS_LP_INIT, &args);
171 	if (ret)
172 		return ret;
173 
174 	__this_cpu_write(tdx_lp_initialized, true);
175 
176 	return 0;
177 }
178 EXPORT_SYMBOL_GPL(tdx_cpu_enable);
179 
180 /*
181  * Add a memory region as a TDX memory block.  The caller must make sure
182  * all memory regions are added in address ascending order and don't
183  * overlap.
184  */
185 static int add_tdx_memblock(struct list_head *tmb_list, unsigned long start_pfn,
186 			    unsigned long end_pfn, int nid)
187 {
188 	struct tdx_memblock *tmb;
189 
190 	tmb = kmalloc(sizeof(*tmb), GFP_KERNEL);
191 	if (!tmb)
192 		return -ENOMEM;
193 
194 	INIT_LIST_HEAD(&tmb->list);
195 	tmb->start_pfn = start_pfn;
196 	tmb->end_pfn = end_pfn;
197 	tmb->nid = nid;
198 
199 	/* @tmb_list is protected by mem_hotplug_lock */
200 	list_add_tail(&tmb->list, tmb_list);
201 	return 0;
202 }
203 
204 static void free_tdx_memlist(struct list_head *tmb_list)
205 {
206 	/* @tmb_list is protected by mem_hotplug_lock */
207 	while (!list_empty(tmb_list)) {
208 		struct tdx_memblock *tmb = list_first_entry(tmb_list,
209 				struct tdx_memblock, list);
210 
211 		list_del(&tmb->list);
212 		kfree(tmb);
213 	}
214 }
215 
216 /*
217  * Ensure that all memblock memory regions are convertible to TDX
218  * memory.  Once this has been established, stash the memblock
219  * ranges off in a secondary structure because memblock is modified
220  * in memory hotplug while TDX memory regions are fixed.
221  */
222 static int build_tdx_memlist(struct list_head *tmb_list)
223 {
224 	unsigned long start_pfn, end_pfn;
225 	int i, nid, ret;
226 
227 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
228 		/*
229 		 * The first 1MB is not reported as TDX convertible memory.
230 		 * Although the first 1MB is always reserved and won't end up
231 		 * to the page allocator, it is still in memblock's memory
232 		 * regions.  Skip them manually to exclude them as TDX memory.
233 		 */
234 		start_pfn = max(start_pfn, PHYS_PFN(SZ_1M));
235 		if (start_pfn >= end_pfn)
236 			continue;
237 
238 		/*
239 		 * Add the memory regions as TDX memory.  The regions in
240 		 * memblock has already guaranteed they are in address
241 		 * ascending order and don't overlap.
242 		 */
243 		ret = add_tdx_memblock(tmb_list, start_pfn, end_pfn, nid);
244 		if (ret)
245 			goto err;
246 	}
247 
248 	return 0;
249 err:
250 	free_tdx_memlist(tmb_list);
251 	return ret;
252 }
253 
254 static int read_sys_metadata_field(u64 field_id, u64 *data)
255 {
256 	struct tdx_module_args args = {};
257 	int ret;
258 
259 	/*
260 	 * TDH.SYS.RD -- reads one global metadata field
261 	 *  - RDX (in): the field to read
262 	 *  - R8 (out): the field data
263 	 */
264 	args.rdx = field_id;
265 	ret = seamcall_prerr_ret(TDH_SYS_RD, &args);
266 	if (ret)
267 		return ret;
268 
269 	*data = args.r8;
270 
271 	return 0;
272 }
273 
274 static int read_sys_metadata_field16(u64 field_id,
275 				     int offset,
276 				     struct tdx_tdmr_sysinfo *ts)
277 {
278 	u16 *ts_member = ((void *)ts) + offset;
279 	u64 tmp;
280 	int ret;
281 
282 	if (WARN_ON_ONCE(MD_FIELD_ID_ELE_SIZE_CODE(field_id) !=
283 			MD_FIELD_ID_ELE_SIZE_16BIT))
284 		return -EINVAL;
285 
286 	ret = read_sys_metadata_field(field_id, &tmp);
287 	if (ret)
288 		return ret;
289 
290 	*ts_member = tmp;
291 
292 	return 0;
293 }
294 
295 struct field_mapping {
296 	u64 field_id;
297 	int offset;
298 };
299 
300 #define TD_SYSINFO_MAP(_field_id, _offset) \
301 	{ .field_id = MD_FIELD_ID_##_field_id,	   \
302 	  .offset   = offsetof(struct tdx_tdmr_sysinfo, _offset) }
303 
304 /* Map TD_SYSINFO fields into 'struct tdx_tdmr_sysinfo': */
305 static const struct field_mapping fields[] = {
306 	TD_SYSINFO_MAP(MAX_TDMRS,	      max_tdmrs),
307 	TD_SYSINFO_MAP(MAX_RESERVED_PER_TDMR, max_reserved_per_tdmr),
308 	TD_SYSINFO_MAP(PAMT_4K_ENTRY_SIZE,    pamt_entry_size[TDX_PS_4K]),
309 	TD_SYSINFO_MAP(PAMT_2M_ENTRY_SIZE,    pamt_entry_size[TDX_PS_2M]),
310 	TD_SYSINFO_MAP(PAMT_1G_ENTRY_SIZE,    pamt_entry_size[TDX_PS_1G]),
311 };
312 
313 static int get_tdx_tdmr_sysinfo(struct tdx_tdmr_sysinfo *tdmr_sysinfo)
314 {
315 	int ret;
316 	int i;
317 
318 	/* Populate 'tdmr_sysinfo' fields using the mapping structure above: */
319 	for (i = 0; i < ARRAY_SIZE(fields); i++) {
320 		ret = read_sys_metadata_field16(fields[i].field_id,
321 						fields[i].offset,
322 						tdmr_sysinfo);
323 		if (ret)
324 			return ret;
325 	}
326 
327 	return 0;
328 }
329 
330 /* Calculate the actual TDMR size */
331 static int tdmr_size_single(u16 max_reserved_per_tdmr)
332 {
333 	int tdmr_sz;
334 
335 	/*
336 	 * The actual size of TDMR depends on the maximum
337 	 * number of reserved areas.
338 	 */
339 	tdmr_sz = sizeof(struct tdmr_info);
340 	tdmr_sz += sizeof(struct tdmr_reserved_area) * max_reserved_per_tdmr;
341 
342 	return ALIGN(tdmr_sz, TDMR_INFO_ALIGNMENT);
343 }
344 
345 static int alloc_tdmr_list(struct tdmr_info_list *tdmr_list,
346 			   struct tdx_tdmr_sysinfo *tdmr_sysinfo)
347 {
348 	size_t tdmr_sz, tdmr_array_sz;
349 	void *tdmr_array;
350 
351 	tdmr_sz = tdmr_size_single(tdmr_sysinfo->max_reserved_per_tdmr);
352 	tdmr_array_sz = tdmr_sz * tdmr_sysinfo->max_tdmrs;
353 
354 	/*
355 	 * To keep things simple, allocate all TDMRs together.
356 	 * The buffer needs to be physically contiguous to make
357 	 * sure each TDMR is physically contiguous.
358 	 */
359 	tdmr_array = alloc_pages_exact(tdmr_array_sz,
360 			GFP_KERNEL | __GFP_ZERO);
361 	if (!tdmr_array)
362 		return -ENOMEM;
363 
364 	tdmr_list->tdmrs = tdmr_array;
365 
366 	/*
367 	 * Keep the size of TDMR to find the target TDMR
368 	 * at a given index in the TDMR list.
369 	 */
370 	tdmr_list->tdmr_sz = tdmr_sz;
371 	tdmr_list->max_tdmrs = tdmr_sysinfo->max_tdmrs;
372 	tdmr_list->nr_consumed_tdmrs = 0;
373 
374 	return 0;
375 }
376 
377 static void free_tdmr_list(struct tdmr_info_list *tdmr_list)
378 {
379 	free_pages_exact(tdmr_list->tdmrs,
380 			tdmr_list->max_tdmrs * tdmr_list->tdmr_sz);
381 }
382 
383 /* Get the TDMR from the list at the given index. */
384 static struct tdmr_info *tdmr_entry(struct tdmr_info_list *tdmr_list,
385 				    int idx)
386 {
387 	int tdmr_info_offset = tdmr_list->tdmr_sz * idx;
388 
389 	return (void *)tdmr_list->tdmrs + tdmr_info_offset;
390 }
391 
392 #define TDMR_ALIGNMENT		SZ_1G
393 #define TDMR_ALIGN_DOWN(_addr)	ALIGN_DOWN((_addr), TDMR_ALIGNMENT)
394 #define TDMR_ALIGN_UP(_addr)	ALIGN((_addr), TDMR_ALIGNMENT)
395 
396 static inline u64 tdmr_end(struct tdmr_info *tdmr)
397 {
398 	return tdmr->base + tdmr->size;
399 }
400 
401 /*
402  * Take the memory referenced in @tmb_list and populate the
403  * preallocated @tdmr_list, following all the special alignment
404  * and size rules for TDMR.
405  */
406 static int fill_out_tdmrs(struct list_head *tmb_list,
407 			  struct tdmr_info_list *tdmr_list)
408 {
409 	struct tdx_memblock *tmb;
410 	int tdmr_idx = 0;
411 
412 	/*
413 	 * Loop over TDX memory regions and fill out TDMRs to cover them.
414 	 * To keep it simple, always try to use one TDMR to cover one
415 	 * memory region.
416 	 *
417 	 * In practice TDX supports at least 64 TDMRs.  A 2-socket system
418 	 * typically only consumes less than 10 of those.  This code is
419 	 * dumb and simple and may use more TMDRs than is strictly
420 	 * required.
421 	 */
422 	list_for_each_entry(tmb, tmb_list, list) {
423 		struct tdmr_info *tdmr = tdmr_entry(tdmr_list, tdmr_idx);
424 		u64 start, end;
425 
426 		start = TDMR_ALIGN_DOWN(PFN_PHYS(tmb->start_pfn));
427 		end   = TDMR_ALIGN_UP(PFN_PHYS(tmb->end_pfn));
428 
429 		/*
430 		 * A valid size indicates the current TDMR has already
431 		 * been filled out to cover the previous memory region(s).
432 		 */
433 		if (tdmr->size) {
434 			/*
435 			 * Loop to the next if the current memory region
436 			 * has already been fully covered.
437 			 */
438 			if (end <= tdmr_end(tdmr))
439 				continue;
440 
441 			/* Otherwise, skip the already covered part. */
442 			if (start < tdmr_end(tdmr))
443 				start = tdmr_end(tdmr);
444 
445 			/*
446 			 * Create a new TDMR to cover the current memory
447 			 * region, or the remaining part of it.
448 			 */
449 			tdmr_idx++;
450 			if (tdmr_idx >= tdmr_list->max_tdmrs) {
451 				pr_warn("initialization failed: TDMRs exhausted.\n");
452 				return -ENOSPC;
453 			}
454 
455 			tdmr = tdmr_entry(tdmr_list, tdmr_idx);
456 		}
457 
458 		tdmr->base = start;
459 		tdmr->size = end - start;
460 	}
461 
462 	/* @tdmr_idx is always the index of the last valid TDMR. */
463 	tdmr_list->nr_consumed_tdmrs = tdmr_idx + 1;
464 
465 	/*
466 	 * Warn early that kernel is about to run out of TDMRs.
467 	 *
468 	 * This is an indication that TDMR allocation has to be
469 	 * reworked to be smarter to not run into an issue.
470 	 */
471 	if (tdmr_list->max_tdmrs - tdmr_list->nr_consumed_tdmrs < TDMR_NR_WARN)
472 		pr_warn("consumed TDMRs reaching limit: %d used out of %d\n",
473 				tdmr_list->nr_consumed_tdmrs,
474 				tdmr_list->max_tdmrs);
475 
476 	return 0;
477 }
478 
479 /*
480  * Calculate PAMT size given a TDMR and a page size.  The returned
481  * PAMT size is always aligned up to 4K page boundary.
482  */
483 static unsigned long tdmr_get_pamt_sz(struct tdmr_info *tdmr, int pgsz,
484 				      u16 pamt_entry_size)
485 {
486 	unsigned long pamt_sz, nr_pamt_entries;
487 
488 	switch (pgsz) {
489 	case TDX_PS_4K:
490 		nr_pamt_entries = tdmr->size >> PAGE_SHIFT;
491 		break;
492 	case TDX_PS_2M:
493 		nr_pamt_entries = tdmr->size >> PMD_SHIFT;
494 		break;
495 	case TDX_PS_1G:
496 		nr_pamt_entries = tdmr->size >> PUD_SHIFT;
497 		break;
498 	default:
499 		WARN_ON_ONCE(1);
500 		return 0;
501 	}
502 
503 	pamt_sz = nr_pamt_entries * pamt_entry_size;
504 	/* TDX requires PAMT size must be 4K aligned */
505 	pamt_sz = ALIGN(pamt_sz, PAGE_SIZE);
506 
507 	return pamt_sz;
508 }
509 
510 /*
511  * Locate a NUMA node which should hold the allocation of the @tdmr
512  * PAMT.  This node will have some memory covered by the TDMR.  The
513  * relative amount of memory covered is not considered.
514  */
515 static int tdmr_get_nid(struct tdmr_info *tdmr, struct list_head *tmb_list)
516 {
517 	struct tdx_memblock *tmb;
518 
519 	/*
520 	 * A TDMR must cover at least part of one TMB.  That TMB will end
521 	 * after the TDMR begins.  But, that TMB may have started before
522 	 * the TDMR.  Find the next 'tmb' that _ends_ after this TDMR
523 	 * begins.  Ignore 'tmb' start addresses.  They are irrelevant.
524 	 */
525 	list_for_each_entry(tmb, tmb_list, list) {
526 		if (tmb->end_pfn > PHYS_PFN(tdmr->base))
527 			return tmb->nid;
528 	}
529 
530 	/*
531 	 * Fall back to allocating the TDMR's metadata from node 0 when
532 	 * no TDX memory block can be found.  This should never happen
533 	 * since TDMRs originate from TDX memory blocks.
534 	 */
535 	pr_warn("TDMR [0x%llx, 0x%llx): unable to find local NUMA node for PAMT allocation, fallback to use node 0.\n",
536 			tdmr->base, tdmr_end(tdmr));
537 	return 0;
538 }
539 
540 /*
541  * Allocate PAMTs from the local NUMA node of some memory in @tmb_list
542  * within @tdmr, and set up PAMTs for @tdmr.
543  */
544 static int tdmr_set_up_pamt(struct tdmr_info *tdmr,
545 			    struct list_head *tmb_list,
546 			    u16 pamt_entry_size[])
547 {
548 	unsigned long pamt_base[TDX_PS_NR];
549 	unsigned long pamt_size[TDX_PS_NR];
550 	unsigned long tdmr_pamt_base;
551 	unsigned long tdmr_pamt_size;
552 	struct page *pamt;
553 	int pgsz, nid;
554 
555 	nid = tdmr_get_nid(tdmr, tmb_list);
556 
557 	/*
558 	 * Calculate the PAMT size for each TDX supported page size
559 	 * and the total PAMT size.
560 	 */
561 	tdmr_pamt_size = 0;
562 	for (pgsz = TDX_PS_4K; pgsz < TDX_PS_NR; pgsz++) {
563 		pamt_size[pgsz] = tdmr_get_pamt_sz(tdmr, pgsz,
564 					pamt_entry_size[pgsz]);
565 		tdmr_pamt_size += pamt_size[pgsz];
566 	}
567 
568 	/*
569 	 * Allocate one chunk of physically contiguous memory for all
570 	 * PAMTs.  This helps minimize the PAMT's use of reserved areas
571 	 * in overlapped TDMRs.
572 	 */
573 	pamt = alloc_contig_pages(tdmr_pamt_size >> PAGE_SHIFT, GFP_KERNEL,
574 			nid, &node_online_map);
575 	if (!pamt)
576 		return -ENOMEM;
577 
578 	/*
579 	 * Break the contiguous allocation back up into the
580 	 * individual PAMTs for each page size.
581 	 */
582 	tdmr_pamt_base = page_to_pfn(pamt) << PAGE_SHIFT;
583 	for (pgsz = TDX_PS_4K; pgsz < TDX_PS_NR; pgsz++) {
584 		pamt_base[pgsz] = tdmr_pamt_base;
585 		tdmr_pamt_base += pamt_size[pgsz];
586 	}
587 
588 	tdmr->pamt_4k_base = pamt_base[TDX_PS_4K];
589 	tdmr->pamt_4k_size = pamt_size[TDX_PS_4K];
590 	tdmr->pamt_2m_base = pamt_base[TDX_PS_2M];
591 	tdmr->pamt_2m_size = pamt_size[TDX_PS_2M];
592 	tdmr->pamt_1g_base = pamt_base[TDX_PS_1G];
593 	tdmr->pamt_1g_size = pamt_size[TDX_PS_1G];
594 
595 	return 0;
596 }
597 
598 static void tdmr_get_pamt(struct tdmr_info *tdmr, unsigned long *pamt_base,
599 			  unsigned long *pamt_size)
600 {
601 	unsigned long pamt_bs, pamt_sz;
602 
603 	/*
604 	 * The PAMT was allocated in one contiguous unit.  The 4K PAMT
605 	 * should always point to the beginning of that allocation.
606 	 */
607 	pamt_bs = tdmr->pamt_4k_base;
608 	pamt_sz = tdmr->pamt_4k_size + tdmr->pamt_2m_size + tdmr->pamt_1g_size;
609 
610 	WARN_ON_ONCE((pamt_bs & ~PAGE_MASK) || (pamt_sz & ~PAGE_MASK));
611 
612 	*pamt_base = pamt_bs;
613 	*pamt_size = pamt_sz;
614 }
615 
616 static void tdmr_do_pamt_func(struct tdmr_info *tdmr,
617 		void (*pamt_func)(unsigned long base, unsigned long size))
618 {
619 	unsigned long pamt_base, pamt_size;
620 
621 	tdmr_get_pamt(tdmr, &pamt_base, &pamt_size);
622 
623 	/* Do nothing if PAMT hasn't been allocated for this TDMR */
624 	if (!pamt_size)
625 		return;
626 
627 	if (WARN_ON_ONCE(!pamt_base))
628 		return;
629 
630 	pamt_func(pamt_base, pamt_size);
631 }
632 
633 static void free_pamt(unsigned long pamt_base, unsigned long pamt_size)
634 {
635 	free_contig_range(pamt_base >> PAGE_SHIFT, pamt_size >> PAGE_SHIFT);
636 }
637 
638 static void tdmr_free_pamt(struct tdmr_info *tdmr)
639 {
640 	tdmr_do_pamt_func(tdmr, free_pamt);
641 }
642 
643 static void tdmrs_free_pamt_all(struct tdmr_info_list *tdmr_list)
644 {
645 	int i;
646 
647 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++)
648 		tdmr_free_pamt(tdmr_entry(tdmr_list, i));
649 }
650 
651 /* Allocate and set up PAMTs for all TDMRs */
652 static int tdmrs_set_up_pamt_all(struct tdmr_info_list *tdmr_list,
653 				 struct list_head *tmb_list,
654 				 u16 pamt_entry_size[])
655 {
656 	int i, ret = 0;
657 
658 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++) {
659 		ret = tdmr_set_up_pamt(tdmr_entry(tdmr_list, i), tmb_list,
660 				pamt_entry_size);
661 		if (ret)
662 			goto err;
663 	}
664 
665 	return 0;
666 err:
667 	tdmrs_free_pamt_all(tdmr_list);
668 	return ret;
669 }
670 
671 /*
672  * Convert TDX private pages back to normal by using MOVDIR64B to
673  * clear these pages.  Note this function doesn't flush cache of
674  * these TDX private pages.  The caller should make sure of that.
675  */
676 static void reset_tdx_pages(unsigned long base, unsigned long size)
677 {
678 	const void *zero_page = (const void *)page_address(ZERO_PAGE(0));
679 	unsigned long phys, end;
680 
681 	end = base + size;
682 	for (phys = base; phys < end; phys += 64)
683 		movdir64b(__va(phys), zero_page);
684 
685 	/*
686 	 * MOVDIR64B uses WC protocol.  Use memory barrier to
687 	 * make sure any later user of these pages sees the
688 	 * updated data.
689 	 */
690 	mb();
691 }
692 
693 static void tdmr_reset_pamt(struct tdmr_info *tdmr)
694 {
695 	tdmr_do_pamt_func(tdmr, reset_tdx_pages);
696 }
697 
698 static void tdmrs_reset_pamt_all(struct tdmr_info_list *tdmr_list)
699 {
700 	int i;
701 
702 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++)
703 		tdmr_reset_pamt(tdmr_entry(tdmr_list, i));
704 }
705 
706 static unsigned long tdmrs_count_pamt_kb(struct tdmr_info_list *tdmr_list)
707 {
708 	unsigned long pamt_size = 0;
709 	int i;
710 
711 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++) {
712 		unsigned long base, size;
713 
714 		tdmr_get_pamt(tdmr_entry(tdmr_list, i), &base, &size);
715 		pamt_size += size;
716 	}
717 
718 	return pamt_size / 1024;
719 }
720 
721 static int tdmr_add_rsvd_area(struct tdmr_info *tdmr, int *p_idx, u64 addr,
722 			      u64 size, u16 max_reserved_per_tdmr)
723 {
724 	struct tdmr_reserved_area *rsvd_areas = tdmr->reserved_areas;
725 	int idx = *p_idx;
726 
727 	/* Reserved area must be 4K aligned in offset and size */
728 	if (WARN_ON(addr & ~PAGE_MASK || size & ~PAGE_MASK))
729 		return -EINVAL;
730 
731 	if (idx >= max_reserved_per_tdmr) {
732 		pr_warn("initialization failed: TDMR [0x%llx, 0x%llx): reserved areas exhausted.\n",
733 				tdmr->base, tdmr_end(tdmr));
734 		return -ENOSPC;
735 	}
736 
737 	/*
738 	 * Consume one reserved area per call.  Make no effort to
739 	 * optimize or reduce the number of reserved areas which are
740 	 * consumed by contiguous reserved areas, for instance.
741 	 */
742 	rsvd_areas[idx].offset = addr - tdmr->base;
743 	rsvd_areas[idx].size = size;
744 
745 	*p_idx = idx + 1;
746 
747 	return 0;
748 }
749 
750 /*
751  * Go through @tmb_list to find holes between memory areas.  If any of
752  * those holes fall within @tdmr, set up a TDMR reserved area to cover
753  * the hole.
754  */
755 static int tdmr_populate_rsvd_holes(struct list_head *tmb_list,
756 				    struct tdmr_info *tdmr,
757 				    int *rsvd_idx,
758 				    u16 max_reserved_per_tdmr)
759 {
760 	struct tdx_memblock *tmb;
761 	u64 prev_end;
762 	int ret;
763 
764 	/*
765 	 * Start looking for reserved blocks at the
766 	 * beginning of the TDMR.
767 	 */
768 	prev_end = tdmr->base;
769 	list_for_each_entry(tmb, tmb_list, list) {
770 		u64 start, end;
771 
772 		start = PFN_PHYS(tmb->start_pfn);
773 		end   = PFN_PHYS(tmb->end_pfn);
774 
775 		/* Break if this region is after the TDMR */
776 		if (start >= tdmr_end(tdmr))
777 			break;
778 
779 		/* Exclude regions before this TDMR */
780 		if (end < tdmr->base)
781 			continue;
782 
783 		/*
784 		 * Skip over memory areas that
785 		 * have already been dealt with.
786 		 */
787 		if (start <= prev_end) {
788 			prev_end = end;
789 			continue;
790 		}
791 
792 		/* Add the hole before this region */
793 		ret = tdmr_add_rsvd_area(tdmr, rsvd_idx, prev_end,
794 				start - prev_end,
795 				max_reserved_per_tdmr);
796 		if (ret)
797 			return ret;
798 
799 		prev_end = end;
800 	}
801 
802 	/* Add the hole after the last region if it exists. */
803 	if (prev_end < tdmr_end(tdmr)) {
804 		ret = tdmr_add_rsvd_area(tdmr, rsvd_idx, prev_end,
805 				tdmr_end(tdmr) - prev_end,
806 				max_reserved_per_tdmr);
807 		if (ret)
808 			return ret;
809 	}
810 
811 	return 0;
812 }
813 
814 /*
815  * Go through @tdmr_list to find all PAMTs.  If any of those PAMTs
816  * overlaps with @tdmr, set up a TDMR reserved area to cover the
817  * overlapping part.
818  */
819 static int tdmr_populate_rsvd_pamts(struct tdmr_info_list *tdmr_list,
820 				    struct tdmr_info *tdmr,
821 				    int *rsvd_idx,
822 				    u16 max_reserved_per_tdmr)
823 {
824 	int i, ret;
825 
826 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++) {
827 		struct tdmr_info *tmp = tdmr_entry(tdmr_list, i);
828 		unsigned long pamt_base, pamt_size, pamt_end;
829 
830 		tdmr_get_pamt(tmp, &pamt_base, &pamt_size);
831 		/* Each TDMR must already have PAMT allocated */
832 		WARN_ON_ONCE(!pamt_size || !pamt_base);
833 
834 		pamt_end = pamt_base + pamt_size;
835 		/* Skip PAMTs outside of the given TDMR */
836 		if ((pamt_end <= tdmr->base) ||
837 				(pamt_base >= tdmr_end(tdmr)))
838 			continue;
839 
840 		/* Only mark the part within the TDMR as reserved */
841 		if (pamt_base < tdmr->base)
842 			pamt_base = tdmr->base;
843 		if (pamt_end > tdmr_end(tdmr))
844 			pamt_end = tdmr_end(tdmr);
845 
846 		ret = tdmr_add_rsvd_area(tdmr, rsvd_idx, pamt_base,
847 				pamt_end - pamt_base,
848 				max_reserved_per_tdmr);
849 		if (ret)
850 			return ret;
851 	}
852 
853 	return 0;
854 }
855 
856 /* Compare function called by sort() for TDMR reserved areas */
857 static int rsvd_area_cmp_func(const void *a, const void *b)
858 {
859 	struct tdmr_reserved_area *r1 = (struct tdmr_reserved_area *)a;
860 	struct tdmr_reserved_area *r2 = (struct tdmr_reserved_area *)b;
861 
862 	if (r1->offset + r1->size <= r2->offset)
863 		return -1;
864 	if (r1->offset >= r2->offset + r2->size)
865 		return 1;
866 
867 	/* Reserved areas cannot overlap.  The caller must guarantee. */
868 	WARN_ON_ONCE(1);
869 	return -1;
870 }
871 
872 /*
873  * Populate reserved areas for the given @tdmr, including memory holes
874  * (via @tmb_list) and PAMTs (via @tdmr_list).
875  */
876 static int tdmr_populate_rsvd_areas(struct tdmr_info *tdmr,
877 				    struct list_head *tmb_list,
878 				    struct tdmr_info_list *tdmr_list,
879 				    u16 max_reserved_per_tdmr)
880 {
881 	int ret, rsvd_idx = 0;
882 
883 	ret = tdmr_populate_rsvd_holes(tmb_list, tdmr, &rsvd_idx,
884 			max_reserved_per_tdmr);
885 	if (ret)
886 		return ret;
887 
888 	ret = tdmr_populate_rsvd_pamts(tdmr_list, tdmr, &rsvd_idx,
889 			max_reserved_per_tdmr);
890 	if (ret)
891 		return ret;
892 
893 	/* TDX requires reserved areas listed in address ascending order */
894 	sort(tdmr->reserved_areas, rsvd_idx, sizeof(struct tdmr_reserved_area),
895 			rsvd_area_cmp_func, NULL);
896 
897 	return 0;
898 }
899 
900 /*
901  * Populate reserved areas for all TDMRs in @tdmr_list, including memory
902  * holes (via @tmb_list) and PAMTs.
903  */
904 static int tdmrs_populate_rsvd_areas_all(struct tdmr_info_list *tdmr_list,
905 					 struct list_head *tmb_list,
906 					 u16 max_reserved_per_tdmr)
907 {
908 	int i;
909 
910 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++) {
911 		int ret;
912 
913 		ret = tdmr_populate_rsvd_areas(tdmr_entry(tdmr_list, i),
914 				tmb_list, tdmr_list, max_reserved_per_tdmr);
915 		if (ret)
916 			return ret;
917 	}
918 
919 	return 0;
920 }
921 
922 /*
923  * Construct a list of TDMRs on the preallocated space in @tdmr_list
924  * to cover all TDX memory regions in @tmb_list based on the TDX module
925  * TDMR global information in @tdmr_sysinfo.
926  */
927 static int construct_tdmrs(struct list_head *tmb_list,
928 			   struct tdmr_info_list *tdmr_list,
929 			   struct tdx_tdmr_sysinfo *tdmr_sysinfo)
930 {
931 	int ret;
932 
933 	ret = fill_out_tdmrs(tmb_list, tdmr_list);
934 	if (ret)
935 		return ret;
936 
937 	ret = tdmrs_set_up_pamt_all(tdmr_list, tmb_list,
938 			tdmr_sysinfo->pamt_entry_size);
939 	if (ret)
940 		return ret;
941 
942 	ret = tdmrs_populate_rsvd_areas_all(tdmr_list, tmb_list,
943 			tdmr_sysinfo->max_reserved_per_tdmr);
944 	if (ret)
945 		tdmrs_free_pamt_all(tdmr_list);
946 
947 	/*
948 	 * The tdmr_info_list is read-only from here on out.
949 	 * Ensure that these writes are seen by other CPUs.
950 	 * Pairs with a smp_rmb() in is_pamt_page().
951 	 */
952 	smp_wmb();
953 
954 	return ret;
955 }
956 
957 static int config_tdx_module(struct tdmr_info_list *tdmr_list, u64 global_keyid)
958 {
959 	struct tdx_module_args args = {};
960 	u64 *tdmr_pa_array;
961 	size_t array_sz;
962 	int i, ret;
963 
964 	/*
965 	 * TDMRs are passed to the TDX module via an array of physical
966 	 * addresses of each TDMR.  The array itself also has certain
967 	 * alignment requirement.
968 	 */
969 	array_sz = tdmr_list->nr_consumed_tdmrs * sizeof(u64);
970 	array_sz = roundup_pow_of_two(array_sz);
971 	if (array_sz < TDMR_INFO_PA_ARRAY_ALIGNMENT)
972 		array_sz = TDMR_INFO_PA_ARRAY_ALIGNMENT;
973 
974 	tdmr_pa_array = kzalloc(array_sz, GFP_KERNEL);
975 	if (!tdmr_pa_array)
976 		return -ENOMEM;
977 
978 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++)
979 		tdmr_pa_array[i] = __pa(tdmr_entry(tdmr_list, i));
980 
981 	args.rcx = __pa(tdmr_pa_array);
982 	args.rdx = tdmr_list->nr_consumed_tdmrs;
983 	args.r8 = global_keyid;
984 	ret = seamcall_prerr(TDH_SYS_CONFIG, &args);
985 
986 	/* Free the array as it is not required anymore. */
987 	kfree(tdmr_pa_array);
988 
989 	return ret;
990 }
991 
992 static int do_global_key_config(void *unused)
993 {
994 	struct tdx_module_args args = {};
995 
996 	return seamcall_prerr(TDH_SYS_KEY_CONFIG, &args);
997 }
998 
999 /*
1000  * Attempt to configure the global KeyID on all physical packages.
1001  *
1002  * This requires running code on at least one CPU in each package.
1003  * TDMR initialization) will fail will fail if any package in the
1004  * system has no online CPUs.
1005  *
1006  * This code takes no affirmative steps to online CPUs.  Callers (aka.
1007  * KVM) can ensure success by ensuring sufficient CPUs are online and
1008  * can run SEAMCALLs.
1009  */
1010 static int config_global_keyid(void)
1011 {
1012 	cpumask_var_t packages;
1013 	int cpu, ret = -EINVAL;
1014 
1015 	if (!zalloc_cpumask_var(&packages, GFP_KERNEL))
1016 		return -ENOMEM;
1017 
1018 	/*
1019 	 * Hardware doesn't guarantee cache coherency across different
1020 	 * KeyIDs.  The kernel needs to flush PAMT's dirty cachelines
1021 	 * (associated with KeyID 0) before the TDX module can use the
1022 	 * global KeyID to access the PAMT.  Given PAMTs are potentially
1023 	 * large (~1/256th of system RAM), just use WBINVD.
1024 	 */
1025 	wbinvd_on_all_cpus();
1026 
1027 	for_each_online_cpu(cpu) {
1028 		/*
1029 		 * The key configuration only needs to be done once per
1030 		 * package and will return an error if configured more
1031 		 * than once.  Avoid doing it multiple times per package.
1032 		 */
1033 		if (cpumask_test_and_set_cpu(topology_physical_package_id(cpu),
1034 					packages))
1035 			continue;
1036 
1037 		/*
1038 		 * TDH.SYS.KEY.CONFIG cannot run concurrently on
1039 		 * different cpus.  Do it one by one.
1040 		 */
1041 		ret = smp_call_on_cpu(cpu, do_global_key_config, NULL, true);
1042 		if (ret)
1043 			break;
1044 	}
1045 
1046 	free_cpumask_var(packages);
1047 	return ret;
1048 }
1049 
1050 static int init_tdmr(struct tdmr_info *tdmr)
1051 {
1052 	u64 next;
1053 
1054 	/*
1055 	 * Initializing a TDMR can be time consuming.  To avoid long
1056 	 * SEAMCALLs, the TDX module may only initialize a part of the
1057 	 * TDMR in each call.
1058 	 */
1059 	do {
1060 		struct tdx_module_args args = {
1061 			.rcx = tdmr->base,
1062 		};
1063 		int ret;
1064 
1065 		ret = seamcall_prerr_ret(TDH_SYS_TDMR_INIT, &args);
1066 		if (ret)
1067 			return ret;
1068 		/*
1069 		 * RDX contains 'next-to-initialize' address if
1070 		 * TDH.SYS.TDMR.INIT did not fully complete and
1071 		 * should be retried.
1072 		 */
1073 		next = args.rdx;
1074 		cond_resched();
1075 		/* Keep making SEAMCALLs until the TDMR is done */
1076 	} while (next < tdmr->base + tdmr->size);
1077 
1078 	return 0;
1079 }
1080 
1081 static int init_tdmrs(struct tdmr_info_list *tdmr_list)
1082 {
1083 	int i;
1084 
1085 	/*
1086 	 * This operation is costly.  It can be parallelized,
1087 	 * but keep it simple for now.
1088 	 */
1089 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++) {
1090 		int ret;
1091 
1092 		ret = init_tdmr(tdmr_entry(tdmr_list, i));
1093 		if (ret)
1094 			return ret;
1095 	}
1096 
1097 	return 0;
1098 }
1099 
1100 static int init_tdx_module(void)
1101 {
1102 	struct tdx_tdmr_sysinfo tdmr_sysinfo;
1103 	int ret;
1104 
1105 	/*
1106 	 * To keep things simple, assume that all TDX-protected memory
1107 	 * will come from the page allocator.  Make sure all pages in the
1108 	 * page allocator are TDX-usable memory.
1109 	 *
1110 	 * Build the list of "TDX-usable" memory regions which cover all
1111 	 * pages in the page allocator to guarantee that.  Do it while
1112 	 * holding mem_hotplug_lock read-lock as the memory hotplug code
1113 	 * path reads the @tdx_memlist to reject any new memory.
1114 	 */
1115 	get_online_mems();
1116 
1117 	ret = build_tdx_memlist(&tdx_memlist);
1118 	if (ret)
1119 		goto out_put_tdxmem;
1120 
1121 	ret = get_tdx_tdmr_sysinfo(&tdmr_sysinfo);
1122 	if (ret)
1123 		goto err_free_tdxmem;
1124 
1125 	/* Allocate enough space for constructing TDMRs */
1126 	ret = alloc_tdmr_list(&tdx_tdmr_list, &tdmr_sysinfo);
1127 	if (ret)
1128 		goto err_free_tdxmem;
1129 
1130 	/* Cover all TDX-usable memory regions in TDMRs */
1131 	ret = construct_tdmrs(&tdx_memlist, &tdx_tdmr_list, &tdmr_sysinfo);
1132 	if (ret)
1133 		goto err_free_tdmrs;
1134 
1135 	/* Pass the TDMRs and the global KeyID to the TDX module */
1136 	ret = config_tdx_module(&tdx_tdmr_list, tdx_global_keyid);
1137 	if (ret)
1138 		goto err_free_pamts;
1139 
1140 	/* Config the key of global KeyID on all packages */
1141 	ret = config_global_keyid();
1142 	if (ret)
1143 		goto err_reset_pamts;
1144 
1145 	/* Initialize TDMRs to complete the TDX module initialization */
1146 	ret = init_tdmrs(&tdx_tdmr_list);
1147 	if (ret)
1148 		goto err_reset_pamts;
1149 
1150 	pr_info("%lu KB allocated for PAMT\n", tdmrs_count_pamt_kb(&tdx_tdmr_list));
1151 
1152 out_put_tdxmem:
1153 	/*
1154 	 * @tdx_memlist is written here and read at memory hotplug time.
1155 	 * Lock out memory hotplug code while building it.
1156 	 */
1157 	put_online_mems();
1158 	return ret;
1159 
1160 err_reset_pamts:
1161 	/*
1162 	 * Part of PAMTs may already have been initialized by the
1163 	 * TDX module.  Flush cache before returning PAMTs back
1164 	 * to the kernel.
1165 	 */
1166 	wbinvd_on_all_cpus();
1167 	/*
1168 	 * According to the TDX hardware spec, if the platform
1169 	 * doesn't have the "partial write machine check"
1170 	 * erratum, any kernel read/write will never cause #MC
1171 	 * in kernel space, thus it's OK to not convert PAMTs
1172 	 * back to normal.  But do the conversion anyway here
1173 	 * as suggested by the TDX spec.
1174 	 */
1175 	tdmrs_reset_pamt_all(&tdx_tdmr_list);
1176 err_free_pamts:
1177 	tdmrs_free_pamt_all(&tdx_tdmr_list);
1178 err_free_tdmrs:
1179 	free_tdmr_list(&tdx_tdmr_list);
1180 err_free_tdxmem:
1181 	free_tdx_memlist(&tdx_memlist);
1182 	goto out_put_tdxmem;
1183 }
1184 
1185 static int __tdx_enable(void)
1186 {
1187 	int ret;
1188 
1189 	ret = init_tdx_module();
1190 	if (ret) {
1191 		pr_err("module initialization failed (%d)\n", ret);
1192 		tdx_module_status = TDX_MODULE_ERROR;
1193 		return ret;
1194 	}
1195 
1196 	pr_info("module initialized\n");
1197 	tdx_module_status = TDX_MODULE_INITIALIZED;
1198 
1199 	return 0;
1200 }
1201 
1202 /**
1203  * tdx_enable - Enable TDX module to make it ready to run TDX guests
1204  *
1205  * This function assumes the caller has: 1) held read lock of CPU hotplug
1206  * lock to prevent any new cpu from becoming online; 2) done both VMXON
1207  * and tdx_cpu_enable() on all online cpus.
1208  *
1209  * This function requires there's at least one online cpu for each CPU
1210  * package to succeed.
1211  *
1212  * This function can be called in parallel by multiple callers.
1213  *
1214  * Return 0 if TDX is enabled successfully, otherwise error.
1215  */
1216 int tdx_enable(void)
1217 {
1218 	int ret;
1219 
1220 	if (!boot_cpu_has(X86_FEATURE_TDX_HOST_PLATFORM))
1221 		return -ENODEV;
1222 
1223 	lockdep_assert_cpus_held();
1224 
1225 	mutex_lock(&tdx_module_lock);
1226 
1227 	switch (tdx_module_status) {
1228 	case TDX_MODULE_UNINITIALIZED:
1229 		ret = __tdx_enable();
1230 		break;
1231 	case TDX_MODULE_INITIALIZED:
1232 		/* Already initialized, great, tell the caller. */
1233 		ret = 0;
1234 		break;
1235 	default:
1236 		/* Failed to initialize in the previous attempts */
1237 		ret = -EINVAL;
1238 		break;
1239 	}
1240 
1241 	mutex_unlock(&tdx_module_lock);
1242 
1243 	return ret;
1244 }
1245 EXPORT_SYMBOL_GPL(tdx_enable);
1246 
1247 static bool is_pamt_page(unsigned long phys)
1248 {
1249 	struct tdmr_info_list *tdmr_list = &tdx_tdmr_list;
1250 	int i;
1251 
1252 	/* Ensure that all remote 'tdmr_list' writes are visible: */
1253 	smp_rmb();
1254 
1255 	/*
1256 	 * The TDX module is no longer returning TDX_SYS_NOT_READY and
1257 	 * is initialized.  The 'tdmr_list' was initialized long ago
1258 	 * and is now read-only.
1259 	 */
1260 	for (i = 0; i < tdmr_list->nr_consumed_tdmrs; i++) {
1261 		unsigned long base, size;
1262 
1263 		tdmr_get_pamt(tdmr_entry(tdmr_list, i), &base, &size);
1264 
1265 		if (phys >= base && phys < (base + size))
1266 			return true;
1267 	}
1268 
1269 	return false;
1270 }
1271 
1272 /*
1273  * Return whether the memory page at the given physical address is TDX
1274  * private memory or not.
1275  *
1276  * This can be imprecise for two known reasons:
1277  * 1. PAMTs are private memory and exist before the TDX module is
1278  *    ready and TDH_PHYMEM_PAGE_RDMD works.  This is a relatively
1279  *    short window that occurs once per boot.
1280  * 2. TDH_PHYMEM_PAGE_RDMD reflects the TDX module's knowledge of the
1281  *    page.  However, the page can still cause #MC until it has been
1282  *    fully converted to shared using 64-byte writes like MOVDIR64B.
1283  *    Buggy hosts might still leave #MC-causing memory in place which
1284  *    this function can not detect.
1285  */
1286 static bool paddr_is_tdx_private(unsigned long phys)
1287 {
1288 	struct tdx_module_args args = {
1289 		.rcx = phys & PAGE_MASK,
1290 	};
1291 	u64 sret;
1292 
1293 	if (!boot_cpu_has(X86_FEATURE_TDX_HOST_PLATFORM))
1294 		return false;
1295 
1296 	/* Get page type from the TDX module */
1297 	sret = __seamcall_ret(TDH_PHYMEM_PAGE_RDMD, &args);
1298 
1299 	/*
1300 	 * The SEAMCALL will not return success unless there is a
1301 	 * working, "ready" TDX module.  Assume an absence of TDX
1302 	 * private pages until SEAMCALL is working.
1303 	 */
1304 	if (sret)
1305 		return false;
1306 
1307 	/*
1308 	 * SEAMCALL was successful -- read page type (via RCX):
1309 	 *
1310 	 *  - PT_NDA:	Page is not used by the TDX module
1311 	 *  - PT_RSVD:	Reserved for Non-TDX use
1312 	 *  - Others:	Page is used by the TDX module
1313 	 *
1314 	 * Note PAMT pages are marked as PT_RSVD but they are also TDX
1315 	 * private memory.
1316 	 */
1317 	switch (args.rcx) {
1318 	case PT_NDA:
1319 		return false;
1320 	case PT_RSVD:
1321 		return is_pamt_page(phys);
1322 	default:
1323 		return true;
1324 	}
1325 }
1326 
1327 /*
1328  * Some TDX-capable CPUs have an erratum.  A write to TDX private
1329  * memory poisons that memory, and a subsequent read of that memory
1330  * triggers #MC.
1331  *
1332  * Help distinguish erratum-triggered #MCs from a normal hardware one.
1333  * Just print additional message to show such #MC may be result of the
1334  * erratum.
1335  */
1336 const char *tdx_dump_mce_info(struct mce *m)
1337 {
1338 	if (!m || !mce_is_memory_error(m) || !mce_usable_address(m))
1339 		return NULL;
1340 
1341 	if (!paddr_is_tdx_private(m->addr))
1342 		return NULL;
1343 
1344 	return "TDX private memory error. Possible kernel bug.";
1345 }
1346 
1347 static __init int record_keyid_partitioning(u32 *tdx_keyid_start,
1348 					    u32 *nr_tdx_keyids)
1349 {
1350 	u32 _nr_mktme_keyids, _tdx_keyid_start, _nr_tdx_keyids;
1351 	int ret;
1352 
1353 	/*
1354 	 * IA32_MKTME_KEYID_PARTIONING:
1355 	 *   Bit [31:0]:	Number of MKTME KeyIDs.
1356 	 *   Bit [63:32]:	Number of TDX private KeyIDs.
1357 	 */
1358 	ret = rdmsr_safe(MSR_IA32_MKTME_KEYID_PARTITIONING, &_nr_mktme_keyids,
1359 			&_nr_tdx_keyids);
1360 	if (ret || !_nr_tdx_keyids)
1361 		return -EINVAL;
1362 
1363 	/* TDX KeyIDs start after the last MKTME KeyID. */
1364 	_tdx_keyid_start = _nr_mktme_keyids + 1;
1365 
1366 	*tdx_keyid_start = _tdx_keyid_start;
1367 	*nr_tdx_keyids = _nr_tdx_keyids;
1368 
1369 	return 0;
1370 }
1371 
1372 static bool is_tdx_memory(unsigned long start_pfn, unsigned long end_pfn)
1373 {
1374 	struct tdx_memblock *tmb;
1375 
1376 	/*
1377 	 * This check assumes that the start_pfn<->end_pfn range does not
1378 	 * cross multiple @tdx_memlist entries.  A single memory online
1379 	 * event across multiple memblocks (from which @tdx_memlist
1380 	 * entries are derived at the time of module initialization) is
1381 	 * not possible.  This is because memory offline/online is done
1382 	 * on granularity of 'struct memory_block', and the hotpluggable
1383 	 * memory region (one memblock) must be multiple of memory_block.
1384 	 */
1385 	list_for_each_entry(tmb, &tdx_memlist, list) {
1386 		if (start_pfn >= tmb->start_pfn && end_pfn <= tmb->end_pfn)
1387 			return true;
1388 	}
1389 	return false;
1390 }
1391 
1392 static int tdx_memory_notifier(struct notifier_block *nb, unsigned long action,
1393 			       void *v)
1394 {
1395 	struct memory_notify *mn = v;
1396 
1397 	if (action != MEM_GOING_ONLINE)
1398 		return NOTIFY_OK;
1399 
1400 	/*
1401 	 * Empty list means TDX isn't enabled.  Allow any memory
1402 	 * to go online.
1403 	 */
1404 	if (list_empty(&tdx_memlist))
1405 		return NOTIFY_OK;
1406 
1407 	/*
1408 	 * The TDX memory configuration is static and can not be
1409 	 * changed.  Reject onlining any memory which is outside of
1410 	 * the static configuration whether it supports TDX or not.
1411 	 */
1412 	if (is_tdx_memory(mn->start_pfn, mn->start_pfn + mn->nr_pages))
1413 		return NOTIFY_OK;
1414 
1415 	return NOTIFY_BAD;
1416 }
1417 
1418 static struct notifier_block tdx_memory_nb = {
1419 	.notifier_call = tdx_memory_notifier,
1420 };
1421 
1422 static void __init check_tdx_erratum(void)
1423 {
1424 	/*
1425 	 * These CPUs have an erratum.  A partial write from non-TD
1426 	 * software (e.g. via MOVNTI variants or UC/WC mapping) to TDX
1427 	 * private memory poisons that memory, and a subsequent read of
1428 	 * that memory triggers #MC.
1429 	 */
1430 	switch (boot_cpu_data.x86_model) {
1431 	case INTEL_FAM6_SAPPHIRERAPIDS_X:
1432 	case INTEL_FAM6_EMERALDRAPIDS_X:
1433 		setup_force_cpu_bug(X86_BUG_TDX_PW_MCE);
1434 	}
1435 }
1436 
1437 void __init tdx_init(void)
1438 {
1439 	u32 tdx_keyid_start, nr_tdx_keyids;
1440 	int err;
1441 
1442 	err = record_keyid_partitioning(&tdx_keyid_start, &nr_tdx_keyids);
1443 	if (err)
1444 		return;
1445 
1446 	pr_info("BIOS enabled: private KeyID range [%u, %u)\n",
1447 			tdx_keyid_start, tdx_keyid_start + nr_tdx_keyids);
1448 
1449 	/*
1450 	 * The TDX module itself requires one 'global KeyID' to protect
1451 	 * its metadata.  If there's only one TDX KeyID, there won't be
1452 	 * any left for TDX guests thus there's no point to enable TDX
1453 	 * at all.
1454 	 */
1455 	if (nr_tdx_keyids < 2) {
1456 		pr_err("initialization failed: too few private KeyIDs available.\n");
1457 		return;
1458 	}
1459 
1460 	/*
1461 	 * At this point, hibernation_available() indicates whether or
1462 	 * not hibernation support has been permanently disabled.
1463 	 */
1464 	if (hibernation_available()) {
1465 		pr_err("initialization failed: Hibernation support is enabled\n");
1466 		return;
1467 	}
1468 
1469 	err = register_memory_notifier(&tdx_memory_nb);
1470 	if (err) {
1471 		pr_err("initialization failed: register_memory_notifier() failed (%d)\n",
1472 				err);
1473 		return;
1474 	}
1475 
1476 #if defined(CONFIG_ACPI) && defined(CONFIG_SUSPEND)
1477 	pr_info("Disable ACPI S3. Turn off TDX in the BIOS to use ACPI S3.\n");
1478 	acpi_suspend_lowlevel = NULL;
1479 #endif
1480 
1481 	/*
1482 	 * Just use the first TDX KeyID as the 'global KeyID' and
1483 	 * leave the rest for TDX guests.
1484 	 */
1485 	tdx_global_keyid = tdx_keyid_start;
1486 	tdx_guest_keyid_start = tdx_keyid_start + 1;
1487 	tdx_nr_guest_keyids = nr_tdx_keyids - 1;
1488 
1489 	setup_force_cpu_cap(X86_FEATURE_TDX_HOST_PLATFORM);
1490 
1491 	check_tdx_erratum();
1492 }
1493