1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copied from arch/arm64/kernel/cpufeature.c
5 * Copyright (C) 2015 ARM Ltd.
6 * Copyright (C) 2017 SiFive
9 #include <linux/acpi.h>
10 #include <linux/bitmap.h>
11 #include <linux/ctype.h>
12 #include <linux/log2.h>
13 #include <linux/memory.h>
14 #include <linux/module.h>
17 #include <asm/alternative.h>
18 #include <asm/cacheflush.h>
19 #include <asm/cpufeature.h>
20 #include <asm/hwcap.h>
21 #include <asm/hwprobe.h>
22 #include <asm/patch.h>
23 #include <asm/processor.h>
24 #include <asm/vector.h>
26 #include "copy-unaligned.h"
28 #define NUM_ALPHA_EXTS ('z' - 'a' + 1)
30 #define MISALIGNED_ACCESS_JIFFIES_LG2 1
31 #define MISALIGNED_BUFFER_SIZE 0x4000
32 #define MISALIGNED_COPY_SIZE ((MISALIGNED_BUFFER_SIZE / 2) - 0x80)
34 unsigned long elf_hwcap __read_mostly;
37 static DECLARE_BITMAP(riscv_isa, RISCV_ISA_EXT_MAX) __read_mostly;
39 /* Per-cpu ISA extensions. */
40 struct riscv_isainfo hart_isa[NR_CPUS];
42 /* Performance information */
43 DEFINE_PER_CPU(long, misaligned_access_speed);
46 * riscv_isa_extension_base() - Get base extension word
48 * @isa_bitmap: ISA bitmap to use
49 * Return: base extension word as unsigned long value
51 * NOTE: If isa_bitmap is NULL then Host ISA bitmap will be used.
53 unsigned long riscv_isa_extension_base(const unsigned long *isa_bitmap)
59 EXPORT_SYMBOL_GPL(riscv_isa_extension_base);
62 * __riscv_isa_extension_available() - Check whether given extension
65 * @isa_bitmap: ISA bitmap to use
66 * @bit: bit position of the desired extension
67 * Return: true or false
69 * NOTE: If isa_bitmap is NULL then Host ISA bitmap will be used.
71 bool __riscv_isa_extension_available(const unsigned long *isa_bitmap, int bit)
73 const unsigned long *bmap = (isa_bitmap) ? isa_bitmap : riscv_isa;
75 if (bit >= RISCV_ISA_EXT_MAX)
78 return test_bit(bit, bmap) ? true : false;
80 EXPORT_SYMBOL_GPL(__riscv_isa_extension_available);
82 static bool riscv_isa_extension_check(int id)
85 case RISCV_ISA_EXT_ZICBOM:
86 if (!riscv_cbom_block_size) {
87 pr_err("Zicbom detected in ISA string, disabling as no cbom-block-size found\n");
89 } else if (!is_power_of_2(riscv_cbom_block_size)) {
90 pr_err("Zicbom disabled as cbom-block-size present, but is not a power-of-2\n");
94 case RISCV_ISA_EXT_ZICBOZ:
95 if (!riscv_cboz_block_size) {
96 pr_err("Zicboz detected in ISA string, disabling as no cboz-block-size found\n");
98 } else if (!is_power_of_2(riscv_cboz_block_size)) {
99 pr_err("Zicboz disabled as cboz-block-size present, but is not a power-of-2\n");
108 #define __RISCV_ISA_EXT_DATA(_name, _id) { \
110 .property = #_name, \
115 * The canonical order of ISA extension names in the ISA string is defined in
116 * chapter 27 of the unprivileged specification.
118 * Ordinarily, for in-kernel data structures, this order is unimportant but
119 * isa_ext_arr defines the order of the ISA string in /proc/cpuinfo.
121 * The specification uses vague wording, such as should, when it comes to
122 * ordering, so for our purposes the following rules apply:
124 * 1. All multi-letter extensions must be separated from other extensions by an
127 * 2. Additional standard extensions (starting with 'Z') must be sorted after
128 * single-letter extensions and before any higher-privileged extensions.
130 * 3. The first letter following the 'Z' conventionally indicates the most
131 * closely related alphabetical extension category, IMAFDQLCBKJTPVH.
132 * If multiple 'Z' extensions are named, they must be ordered first by
133 * category, then alphabetically within a category.
135 * 3. Standard supervisor-level extensions (starting with 'S') must be listed
136 * after standard unprivileged extensions. If multiple supervisor-level
137 * extensions are listed, they must be ordered alphabetically.
139 * 4. Standard machine-level extensions (starting with 'Zxm') must be listed
140 * after any lower-privileged, standard extensions. If multiple
141 * machine-level extensions are listed, they must be ordered
144 * 5. Non-standard extensions (starting with 'X') must be listed after all
145 * standard extensions. If multiple non-standard extensions are listed, they
146 * must be ordered alphabetically.
148 * An example string following the order is:
149 * rv64imadc_zifoo_zigoo_zafoo_sbar_scar_zxmbaz_xqux_xrux
151 * New entries to this struct should follow the ordering rules described above.
153 const struct riscv_isa_ext_data riscv_isa_ext[] = {
154 __RISCV_ISA_EXT_DATA(i, RISCV_ISA_EXT_i),
155 __RISCV_ISA_EXT_DATA(m, RISCV_ISA_EXT_m),
156 __RISCV_ISA_EXT_DATA(a, RISCV_ISA_EXT_a),
157 __RISCV_ISA_EXT_DATA(f, RISCV_ISA_EXT_f),
158 __RISCV_ISA_EXT_DATA(d, RISCV_ISA_EXT_d),
159 __RISCV_ISA_EXT_DATA(q, RISCV_ISA_EXT_q),
160 __RISCV_ISA_EXT_DATA(c, RISCV_ISA_EXT_c),
161 __RISCV_ISA_EXT_DATA(b, RISCV_ISA_EXT_b),
162 __RISCV_ISA_EXT_DATA(k, RISCV_ISA_EXT_k),
163 __RISCV_ISA_EXT_DATA(j, RISCV_ISA_EXT_j),
164 __RISCV_ISA_EXT_DATA(p, RISCV_ISA_EXT_p),
165 __RISCV_ISA_EXT_DATA(v, RISCV_ISA_EXT_v),
166 __RISCV_ISA_EXT_DATA(h, RISCV_ISA_EXT_h),
167 __RISCV_ISA_EXT_DATA(zicbom, RISCV_ISA_EXT_ZICBOM),
168 __RISCV_ISA_EXT_DATA(zicboz, RISCV_ISA_EXT_ZICBOZ),
169 __RISCV_ISA_EXT_DATA(zicntr, RISCV_ISA_EXT_ZICNTR),
170 __RISCV_ISA_EXT_DATA(zicond, RISCV_ISA_EXT_ZICOND),
171 __RISCV_ISA_EXT_DATA(zicsr, RISCV_ISA_EXT_ZICSR),
172 __RISCV_ISA_EXT_DATA(zifencei, RISCV_ISA_EXT_ZIFENCEI),
173 __RISCV_ISA_EXT_DATA(zihintpause, RISCV_ISA_EXT_ZIHINTPAUSE),
174 __RISCV_ISA_EXT_DATA(zihpm, RISCV_ISA_EXT_ZIHPM),
175 __RISCV_ISA_EXT_DATA(zba, RISCV_ISA_EXT_ZBA),
176 __RISCV_ISA_EXT_DATA(zbb, RISCV_ISA_EXT_ZBB),
177 __RISCV_ISA_EXT_DATA(zbs, RISCV_ISA_EXT_ZBS),
178 __RISCV_ISA_EXT_DATA(smaia, RISCV_ISA_EXT_SMAIA),
179 __RISCV_ISA_EXT_DATA(smstateen, RISCV_ISA_EXT_SMSTATEEN),
180 __RISCV_ISA_EXT_DATA(ssaia, RISCV_ISA_EXT_SSAIA),
181 __RISCV_ISA_EXT_DATA(sscofpmf, RISCV_ISA_EXT_SSCOFPMF),
182 __RISCV_ISA_EXT_DATA(sstc, RISCV_ISA_EXT_SSTC),
183 __RISCV_ISA_EXT_DATA(svinval, RISCV_ISA_EXT_SVINVAL),
184 __RISCV_ISA_EXT_DATA(svnapot, RISCV_ISA_EXT_SVNAPOT),
185 __RISCV_ISA_EXT_DATA(svpbmt, RISCV_ISA_EXT_SVPBMT),
188 const size_t riscv_isa_ext_count = ARRAY_SIZE(riscv_isa_ext);
190 static void __init riscv_parse_isa_string(unsigned long *this_hwcap, struct riscv_isainfo *isainfo,
191 unsigned long *isa2hwcap, const char *isa)
194 * For all possible cpus, we have already validated in
195 * the boot process that they at least contain "rv" and
196 * whichever of "32"/"64" this kernel supports, and so this
197 * section can be skipped.
202 const char *ext = isa++;
203 const char *ext_end = isa;
204 bool ext_long = false, ext_err = false;
209 * Workaround for invalid single-letter 's' & 'u' (QEMU).
210 * No need to set the bit in riscv_isa as 's' & 'u' are
211 * not valid ISA extensions. It works unless the first
212 * multi-letter extension in the ISA string begins with
213 * "Su" and is not prefixed with an underscore.
215 if (ext[-1] != '_' && ext[1] == 'u') {
227 * Before attempting to parse the extension itself, we find its end.
228 * As multi-letter extensions must be split from other multi-letter
229 * extensions with an "_", the end of a multi-letter extension will
230 * either be the null character or the "_" at the start of the next
231 * multi-letter extension.
233 * Next, as the extensions version is currently ignored, we
234 * eliminate that portion. This is done by parsing backwards from
235 * the end of the extension, removing any numbers. This may be a
236 * major or minor number however, so the process is repeated if a
237 * minor number was found.
239 * ext_end is intended to represent the first character *after* the
240 * name portion of an extension, but will be decremented to the last
241 * character itself while eliminating the extensions version number.
242 * A simple re-increment solves this problem.
245 for (; *isa && *isa != '_'; ++isa)
246 if (unlikely(!isalnum(*isa)))
250 if (unlikely(ext_err))
253 if (!isdigit(ext_end[-1]))
256 while (isdigit(*--ext_end))
259 if (tolower(ext_end[0]) != 'p' || !isdigit(ext_end[-1])) {
264 while (isdigit(*--ext_end))
271 * Things are a little easier for single-letter extensions, as they
272 * are parsed forwards.
274 * After checking that our starting position is valid, we need to
275 * ensure that, when isa was incremented at the start of the loop,
276 * that it arrived at the start of the next extension.
278 * If we are already on a non-digit, there is nothing to do. Either
279 * we have a multi-letter extension's _, or the start of an
282 * Otherwise we have found the current extension's major version
283 * number. Parse past it, and a subsequent p/minor version number
284 * if present. The `p` extension must not appear immediately after
285 * a number, so there is no fear of missing it.
288 if (unlikely(!isalpha(*ext))) {
296 while (isdigit(*++isa))
299 if (tolower(*isa) != 'p')
302 if (!isdigit(*++isa)) {
307 while (isdigit(*++isa))
314 * The parser expects that at the start of an iteration isa points to the
315 * first character of the next extension. As we stop parsing an extension
316 * on meeting a non-alphanumeric character, an extra increment is needed
317 * where the succeeding extension is a multi-letter prefixed with an "_".
322 #define SET_ISA_EXT_MAP(name, bit) \
324 if ((ext_end - ext == strlen(name)) && \
325 !strncasecmp(ext, name, strlen(name)) && \
326 riscv_isa_extension_check(bit)) \
327 set_bit(bit, isainfo->isa); \
330 if (unlikely(ext_err))
333 int nr = tolower(*ext) - 'a';
335 if (riscv_isa_extension_check(nr)) {
336 *this_hwcap |= isa2hwcap[nr];
337 set_bit(nr, isainfo->isa);
340 for (int i = 0; i < riscv_isa_ext_count; i++)
341 SET_ISA_EXT_MAP(riscv_isa_ext[i].name,
342 riscv_isa_ext[i].id);
344 #undef SET_ISA_EXT_MAP
348 static void __init riscv_fill_hwcap_from_isa_string(unsigned long *isa2hwcap)
350 struct device_node *node;
353 struct acpi_table_header *rhct;
357 if (!acpi_disabled) {
358 status = acpi_get_table(ACPI_SIG_RHCT, 0, &rhct);
359 if (ACPI_FAILURE(status))
363 for_each_possible_cpu(cpu) {
364 struct riscv_isainfo *isainfo = &hart_isa[cpu];
365 unsigned long this_hwcap = 0;
368 node = of_cpu_device_node_get(cpu);
370 pr_warn("Unable to find cpu node\n");
374 rc = of_property_read_string(node, "riscv,isa", &isa);
377 pr_warn("Unable to find \"riscv,isa\" devicetree entry\n");
381 rc = acpi_get_riscv_isa(rhct, cpu, &isa);
383 pr_warn("Unable to get ISA for the hart - %d\n", cpu);
388 riscv_parse_isa_string(&this_hwcap, isainfo, isa2hwcap, isa);
391 * These ones were as they were part of the base ISA when the
392 * port & dt-bindings were upstreamed, and so can be set
393 * unconditionally where `i` is in riscv,isa on DT systems.
396 set_bit(RISCV_ISA_EXT_ZICSR, isainfo->isa);
397 set_bit(RISCV_ISA_EXT_ZIFENCEI, isainfo->isa);
398 set_bit(RISCV_ISA_EXT_ZICNTR, isainfo->isa);
399 set_bit(RISCV_ISA_EXT_ZIHPM, isainfo->isa);
403 * All "okay" hart should have same isa. Set HWCAP based on
404 * common capabilities of every "okay" hart, in case they don't
408 elf_hwcap &= this_hwcap;
410 elf_hwcap = this_hwcap;
412 if (bitmap_empty(riscv_isa, RISCV_ISA_EXT_MAX))
413 bitmap_copy(riscv_isa, isainfo->isa, RISCV_ISA_EXT_MAX);
415 bitmap_and(riscv_isa, riscv_isa, isainfo->isa, RISCV_ISA_EXT_MAX);
418 if (!acpi_disabled && rhct)
419 acpi_put_table((struct acpi_table_header *)rhct);
422 static int __init riscv_fill_hwcap_from_ext_list(unsigned long *isa2hwcap)
426 for_each_possible_cpu(cpu) {
427 unsigned long this_hwcap = 0;
428 struct device_node *cpu_node;
429 struct riscv_isainfo *isainfo = &hart_isa[cpu];
431 cpu_node = of_cpu_device_node_get(cpu);
433 pr_warn("Unable to find cpu node\n");
437 if (!of_property_present(cpu_node, "riscv,isa-extensions")) {
438 of_node_put(cpu_node);
442 for (int i = 0; i < riscv_isa_ext_count; i++) {
443 if (of_property_match_string(cpu_node, "riscv,isa-extensions",
444 riscv_isa_ext[i].property) < 0)
447 if (!riscv_isa_extension_check(riscv_isa_ext[i].id))
450 /* Only single letter extensions get set in hwcap */
451 if (strnlen(riscv_isa_ext[i].name, 2) == 1)
452 this_hwcap |= isa2hwcap[riscv_isa_ext[i].id];
454 set_bit(riscv_isa_ext[i].id, isainfo->isa);
457 of_node_put(cpu_node);
460 * All "okay" harts should have same isa. Set HWCAP based on
461 * common capabilities of every "okay" hart, in case they don't.
464 elf_hwcap &= this_hwcap;
466 elf_hwcap = this_hwcap;
468 if (bitmap_empty(riscv_isa, RISCV_ISA_EXT_MAX))
469 bitmap_copy(riscv_isa, isainfo->isa, RISCV_ISA_EXT_MAX);
471 bitmap_and(riscv_isa, riscv_isa, isainfo->isa, RISCV_ISA_EXT_MAX);
474 if (bitmap_empty(riscv_isa, RISCV_ISA_EXT_MAX))
480 #ifdef CONFIG_RISCV_ISA_FALLBACK
481 bool __initdata riscv_isa_fallback = true;
483 bool __initdata riscv_isa_fallback;
484 static int __init riscv_isa_fallback_setup(char *__unused)
486 riscv_isa_fallback = true;
489 early_param("riscv_isa_fallback", riscv_isa_fallback_setup);
492 void __init riscv_fill_hwcap(void)
494 char print_str[NUM_ALPHA_EXTS + 1];
495 unsigned long isa2hwcap[26] = {0};
498 isa2hwcap['i' - 'a'] = COMPAT_HWCAP_ISA_I;
499 isa2hwcap['m' - 'a'] = COMPAT_HWCAP_ISA_M;
500 isa2hwcap['a' - 'a'] = COMPAT_HWCAP_ISA_A;
501 isa2hwcap['f' - 'a'] = COMPAT_HWCAP_ISA_F;
502 isa2hwcap['d' - 'a'] = COMPAT_HWCAP_ISA_D;
503 isa2hwcap['c' - 'a'] = COMPAT_HWCAP_ISA_C;
504 isa2hwcap['v' - 'a'] = COMPAT_HWCAP_ISA_V;
506 if (!acpi_disabled) {
507 riscv_fill_hwcap_from_isa_string(isa2hwcap);
509 int ret = riscv_fill_hwcap_from_ext_list(isa2hwcap);
511 if (ret && riscv_isa_fallback) {
512 pr_info("Falling back to deprecated \"riscv,isa\"\n");
513 riscv_fill_hwcap_from_isa_string(isa2hwcap);
518 * We don't support systems with F but without D, so mask those out
521 if ((elf_hwcap & COMPAT_HWCAP_ISA_F) && !(elf_hwcap & COMPAT_HWCAP_ISA_D)) {
522 pr_info("This kernel does not support systems with F but not D\n");
523 elf_hwcap &= ~COMPAT_HWCAP_ISA_F;
526 if (elf_hwcap & COMPAT_HWCAP_ISA_V) {
527 riscv_v_setup_vsize();
529 * ISA string in device tree might have 'v' flag, but
530 * CONFIG_RISCV_ISA_V is disabled in kernel.
531 * Clear V flag in elf_hwcap if CONFIG_RISCV_ISA_V is disabled.
533 if (!IS_ENABLED(CONFIG_RISCV_ISA_V))
534 elf_hwcap &= ~COMPAT_HWCAP_ISA_V;
537 memset(print_str, 0, sizeof(print_str));
538 for (i = 0, j = 0; i < NUM_ALPHA_EXTS; i++)
539 if (riscv_isa[0] & BIT_MASK(i))
540 print_str[j++] = (char)('a' + i);
541 pr_info("riscv: base ISA extensions %s\n", print_str);
543 memset(print_str, 0, sizeof(print_str));
544 for (i = 0, j = 0; i < NUM_ALPHA_EXTS; i++)
545 if (elf_hwcap & BIT_MASK(i))
546 print_str[j++] = (char)('a' + i);
547 pr_info("riscv: ELF capabilities %s\n", print_str);
550 unsigned long riscv_get_elf_hwcap(void)
554 hwcap = (elf_hwcap & ((1UL << RISCV_ISA_EXT_BASE) - 1));
556 if (!riscv_v_vstate_ctrl_user_allowed())
557 hwcap &= ~COMPAT_HWCAP_ISA_V;
562 void check_unaligned_access(int cpu)
564 u64 start_cycles, end_cycles;
568 unsigned long start_jiffies, now;
572 long speed = RISCV_HWPROBE_MISALIGNED_SLOW;
574 page = alloc_pages(GFP_NOWAIT, get_order(MISALIGNED_BUFFER_SIZE));
576 pr_warn("Can't alloc pages to measure memcpy performance");
580 /* Make an unaligned destination buffer. */
581 dst = (void *)((unsigned long)page_address(page) | 0x1);
582 /* Unalign src as well, but differently (off by 1 + 2 = 3). */
583 src = dst + (MISALIGNED_BUFFER_SIZE / 2);
587 __riscv_copy_words_unaligned(dst, src, MISALIGNED_COPY_SIZE);
589 start_jiffies = jiffies;
590 while ((now = jiffies) == start_jiffies)
594 * For a fixed amount of time, repeatedly try the function, and take
595 * the best time in cycles as the measurement.
597 while (time_before(jiffies, now + (1 << MISALIGNED_ACCESS_JIFFIES_LG2))) {
598 start_cycles = get_cycles64();
599 /* Ensure the CSR read can't reorder WRT to the copy. */
601 __riscv_copy_words_unaligned(dst, src, MISALIGNED_COPY_SIZE);
602 /* Ensure the copy ends before the end time is snapped. */
604 end_cycles = get_cycles64();
605 if ((end_cycles - start_cycles) < word_cycles)
606 word_cycles = end_cycles - start_cycles;
610 __riscv_copy_bytes_unaligned(dst, src, MISALIGNED_COPY_SIZE);
611 start_jiffies = jiffies;
612 while ((now = jiffies) == start_jiffies)
615 while (time_before(jiffies, now + (1 << MISALIGNED_ACCESS_JIFFIES_LG2))) {
616 start_cycles = get_cycles64();
618 __riscv_copy_bytes_unaligned(dst, src, MISALIGNED_COPY_SIZE);
620 end_cycles = get_cycles64();
621 if ((end_cycles - start_cycles) < byte_cycles)
622 byte_cycles = end_cycles - start_cycles;
627 /* Don't divide by zero. */
628 if (!word_cycles || !byte_cycles) {
629 pr_warn("cpu%d: rdtime lacks granularity needed to measure unaligned access speed\n",
635 if (word_cycles < byte_cycles)
636 speed = RISCV_HWPROBE_MISALIGNED_FAST;
638 ratio = div_u64((byte_cycles * 100), word_cycles);
639 pr_info("cpu%d: Ratio of byte access time to unaligned word access is %d.%02d, unaligned accesses are %s\n",
643 (speed == RISCV_HWPROBE_MISALIGNED_FAST) ? "fast" : "slow");
645 per_cpu(misaligned_access_speed, cpu) = speed;
648 __free_pages(page, get_order(MISALIGNED_BUFFER_SIZE));
651 static int check_unaligned_access_boot_cpu(void)
653 check_unaligned_access(0);
657 arch_initcall(check_unaligned_access_boot_cpu);
659 void riscv_user_isa_enable(void)
661 if (riscv_cpu_has_extension_unlikely(smp_processor_id(), RISCV_ISA_EXT_ZICBOZ))
662 csr_set(CSR_SENVCFG, ENVCFG_CBZE);
665 #ifdef CONFIG_RISCV_ALTERNATIVE
667 * Alternative patch sites consider 48 bits when determining when to patch
668 * the old instruction sequence with the new. These bits are broken into a
669 * 16-bit vendor ID and a 32-bit patch ID. A non-zero vendor ID means the
670 * patch site is for an erratum, identified by the 32-bit patch ID. When
671 * the vendor ID is zero, the patch site is for a cpufeature. cpufeatures
672 * further break down patch ID into two 16-bit numbers. The lower 16 bits
673 * are the cpufeature ID and the upper 16 bits are used for a value specific
674 * to the cpufeature and patch site. If the upper 16 bits are zero, then it
675 * implies no specific value is specified. cpufeatures that want to control
676 * patching on a per-site basis will provide non-zero values and implement
677 * checks here. The checks return true when patching should be done, and
680 static bool riscv_cpufeature_patch_check(u16 id, u16 value)
686 case RISCV_ISA_EXT_ZICBOZ:
688 * Zicboz alternative applications provide the maximum
689 * supported block size order, or zero when it doesn't
690 * matter. If the current block size exceeds the maximum,
691 * then the alternative cannot be applied.
693 return riscv_cboz_block_size <= (1U << value);
699 void __init_or_module riscv_cpufeature_patch_func(struct alt_entry *begin,
700 struct alt_entry *end,
703 struct alt_entry *alt;
704 void *oldptr, *altptr;
707 if (stage == RISCV_ALTERNATIVES_EARLY_BOOT)
710 for (alt = begin; alt < end; alt++) {
711 if (alt->vendor_id != 0)
714 id = PATCH_ID_CPUFEATURE_ID(alt->patch_id);
716 if (id >= RISCV_ISA_EXT_MAX) {
717 WARN(1, "This extension id:%d is not in ISA extension list", id);
721 if (!__riscv_isa_extension_available(NULL, id))
724 value = PATCH_ID_CPUFEATURE_VALUE(alt->patch_id);
725 if (!riscv_cpufeature_patch_check(id, value))
728 oldptr = ALT_OLD_PTR(alt);
729 altptr = ALT_ALT_PTR(alt);
731 mutex_lock(&text_mutex);
732 patch_text_nosync(oldptr, altptr, alt->alt_len);
733 riscv_alternative_fix_offsets(oldptr, alt->alt_len, oldptr - altptr);
734 mutex_unlock(&text_mutex);