3 * Common boot and setup code.
5 * Copyright (C) 2001 PPC64 Team, IBM Corp
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
13 #include <linux/export.h>
14 #include <linux/string.h>
15 #include <linux/sched.h>
16 #include <linux/init.h>
17 #include <linux/kernel.h>
18 #include <linux/reboot.h>
19 #include <linux/delay.h>
20 #include <linux/initrd.h>
21 #include <linux/seq_file.h>
22 #include <linux/ioport.h>
23 #include <linux/console.h>
24 #include <linux/utsname.h>
25 #include <linux/tty.h>
26 #include <linux/root_dev.h>
27 #include <linux/notifier.h>
28 #include <linux/cpu.h>
29 #include <linux/unistd.h>
30 #include <linux/serial.h>
31 #include <linux/serial_8250.h>
32 #include <linux/memblock.h>
33 #include <linux/pci.h>
34 #include <linux/lockdep.h>
35 #include <linux/memory.h>
36 #include <linux/nmi.h>
38 #include <asm/debugfs.h>
40 #include <asm/kdump.h>
42 #include <asm/processor.h>
43 #include <asm/pgtable.h>
46 #include <asm/machdep.h>
49 #include <asm/cputable.h>
50 #include <asm/dt_cpu_ftrs.h>
51 #include <asm/sections.h>
52 #include <asm/btext.h>
53 #include <asm/nvram.h>
54 #include <asm/setup.h>
56 #include <asm/iommu.h>
57 #include <asm/serial.h>
58 #include <asm/cache.h>
61 #include <asm/firmware.h>
64 #include <asm/kexec.h>
65 #include <asm/code-patching.h>
66 #include <asm/livepatch.h>
68 #include <asm/cputhreads.h>
69 #include <asm/hw_irq.h>
70 #include <asm/feature-fixups.h>
75 #define DBG(fmt...) udbg_printf(fmt)
80 int spinning_secondaries;
83 struct ppc64_caches ppc64_caches = {
93 EXPORT_SYMBOL_GPL(ppc64_caches);
95 #if defined(CONFIG_PPC_BOOK3E) && defined(CONFIG_SMP)
96 void __init setup_tlb_core_data(void)
100 BUILD_BUG_ON(offsetof(struct tlb_core_data, lock) != 0);
102 for_each_possible_cpu(cpu) {
103 int first = cpu_first_thread_sibling(cpu);
106 * If we boot via kdump on a non-primary thread,
107 * make sure we point at the thread that actually
110 if (cpu_first_thread_sibling(boot_cpuid) == first)
113 paca_ptrs[cpu]->tcd_ptr = &paca_ptrs[first]->tcd;
116 * If we have threads, we need either tlbsrx.
117 * or e6500 tablewalk mode, or else TLB handlers
118 * will be racy and could produce duplicate entries.
119 * Should we panic instead?
121 WARN_ONCE(smt_enabled_at_boot >= 2 &&
122 !mmu_has_feature(MMU_FTR_USE_TLBRSRV) &&
123 book3e_htw_mode != PPC_HTW_E6500,
124 "%s: unsupported MMU configuration\n", __func__);
131 static char *smt_enabled_cmdline;
133 /* Look for ibm,smt-enabled OF option */
134 void __init check_smt_enabled(void)
136 struct device_node *dn;
137 const char *smt_option;
139 /* Default to enabling all threads */
140 smt_enabled_at_boot = threads_per_core;
142 /* Allow the command line to overrule the OF option */
143 if (smt_enabled_cmdline) {
144 if (!strcmp(smt_enabled_cmdline, "on"))
145 smt_enabled_at_boot = threads_per_core;
146 else if (!strcmp(smt_enabled_cmdline, "off"))
147 smt_enabled_at_boot = 0;
152 rc = kstrtoint(smt_enabled_cmdline, 10, &smt);
154 smt_enabled_at_boot =
155 min(threads_per_core, smt);
158 dn = of_find_node_by_path("/options");
160 smt_option = of_get_property(dn, "ibm,smt-enabled",
164 if (!strcmp(smt_option, "on"))
165 smt_enabled_at_boot = threads_per_core;
166 else if (!strcmp(smt_option, "off"))
167 smt_enabled_at_boot = 0;
175 /* Look for smt-enabled= cmdline option */
176 static int __init early_smt_enabled(char *p)
178 smt_enabled_cmdline = p;
181 early_param("smt-enabled", early_smt_enabled);
183 #endif /* CONFIG_SMP */
185 /** Fix up paca fields required for the boot cpu */
186 static void __init fixup_boot_paca(void)
188 /* The boot cpu is started */
189 get_paca()->cpu_start = 1;
190 /* Allow percpu accesses to work until we setup percpu data */
191 get_paca()->data_offset = 0;
192 /* Mark interrupts disabled in PACA */
193 irq_soft_mask_set(IRQS_DISABLED);
196 static void __init configure_exceptions(void)
199 * Setup the trampolines from the lowmem exception vectors
200 * to the kdump kernel when not using a relocatable kernel.
202 setup_kdump_trampoline();
204 /* Under a PAPR hypervisor, we need hypercalls */
205 if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
206 /* Enable AIL if possible */
207 pseries_enable_reloc_on_exc();
210 * Tell the hypervisor that we want our exceptions to
211 * be taken in little endian mode.
213 * We don't call this for big endian as our calling convention
214 * makes us always enter in BE, and the call may fail under
215 * some circumstances with kdump.
217 #ifdef __LITTLE_ENDIAN__
218 pseries_little_endian_exceptions();
221 /* Set endian mode using OPAL */
222 if (firmware_has_feature(FW_FEATURE_OPAL))
223 opal_configure_cores();
225 /* AIL on native is done in cpu_ready_for_interrupts() */
229 static void cpu_ready_for_interrupts(void)
232 * Enable AIL if supported, and we are in hypervisor mode. This
233 * is called once for every processor.
235 * If we are not in hypervisor mode the job is done once for
236 * the whole partition in configure_exceptions().
238 if (cpu_has_feature(CPU_FTR_HVMODE) &&
239 cpu_has_feature(CPU_FTR_ARCH_207S)) {
240 unsigned long lpcr = mfspr(SPRN_LPCR);
241 mtspr(SPRN_LPCR, lpcr | LPCR_AIL_3);
245 * Set HFSCR:TM based on CPU features:
246 * In the special case of TM no suspend (P9N DD2.1), Linux is
247 * told TM is off via the dt-ftrs but told to (partially) use
248 * it via OPAL_REINIT_CPUS_TM_SUSPEND_DISABLED. So HFSCR[TM]
249 * will be off from dt-ftrs but we need to turn it on for the
252 if (cpu_has_feature(CPU_FTR_HVMODE)) {
253 if (cpu_has_feature(CPU_FTR_TM_COMP))
254 mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) | HFSCR_TM);
256 mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) & ~HFSCR_TM);
259 /* Set IR and DR in PACA MSR */
260 get_paca()->kernel_msr = MSR_KERNEL;
263 unsigned long spr_default_dscr = 0;
265 void __init record_spr_defaults(void)
267 if (early_cpu_has_feature(CPU_FTR_DSCR))
268 spr_default_dscr = mfspr(SPRN_DSCR);
272 * Early initialization entry point. This is called by head.S
273 * with MMU translation disabled. We rely on the "feature" of
274 * the CPU that ignores the top 2 bits of the address in real
275 * mode so we can access kernel globals normally provided we
276 * only toy with things in the RMO region. From here, we do
277 * some early parsing of the device-tree to setup out MEMBLOCK
278 * data structures, and allocate & initialize the hash table
279 * and segment tables so we can start running with translation
282 * It is this function which will call the probe() callback of
283 * the various platform types and copy the matching one to the
284 * global ppc_md structure. Your platform can eventually do
285 * some very early initializations from the probe() routine, but
286 * this is not recommended, be very careful as, for example, the
287 * device-tree is not accessible via normal means at this point.
290 void __init early_setup(unsigned long dt_ptr)
292 static __initdata struct paca_struct boot_paca;
294 /* -------- printk is _NOT_ safe to use here ! ------- */
296 /* Try new device tree based feature discovery ... */
297 if (!dt_cpu_ftrs_init(__va(dt_ptr)))
298 /* Otherwise use the old style CPU table */
299 identify_cpu(0, mfspr(SPRN_PVR));
301 /* Assume we're on cpu 0 for now. Don't write to the paca yet! */
302 initialise_paca(&boot_paca, 0);
303 setup_paca(&boot_paca);
306 /* -------- printk is now safe to use ------- */
308 /* Enable early debugging if any specified (see udbg.h) */
311 DBG(" -> early_setup(), dt_ptr: 0x%lx\n", dt_ptr);
314 * Do early initialization using the flattened device
315 * tree, such as retrieving the physical memory map or
316 * calculating/retrieving the hash table size.
318 early_init_devtree(__va(dt_ptr));
320 /* Now we know the logical id of our boot cpu, setup the paca. */
321 if (boot_cpuid != 0) {
322 /* Poison paca_ptrs[0] again if it's not the boot cpu */
323 memset(&paca_ptrs[0], 0x88, sizeof(paca_ptrs[0]));
325 setup_paca(paca_ptrs[boot_cpuid]);
329 * Configure exception handlers. This include setting up trampolines
330 * if needed, setting exception endian mode, etc...
332 configure_exceptions();
334 /* Apply all the dynamic patching */
335 apply_feature_fixups();
336 setup_feature_keys();
338 /* Initialize the hash table or TLB handling */
342 * After firmware and early platform setup code has set things up,
343 * we note the SPR values for configurable control/performance
344 * registers, and use those as initial defaults.
346 record_spr_defaults();
349 * At this point, we can let interrupts switch to virtual mode
350 * (the MMU has been setup), so adjust the MSR in the PACA to
351 * have IR and DR set and enable AIL if it exists
353 cpu_ready_for_interrupts();
356 * We enable ftrace here, but since we only support DYNAMIC_FTRACE, it
357 * will only actually get enabled on the boot cpu much later once
358 * ftrace itself has been initialized.
360 this_cpu_enable_ftrace();
362 DBG(" <- early_setup()\n");
364 #ifdef CONFIG_PPC_EARLY_DEBUG_BOOTX
366 * This needs to be done *last* (after the above DBG() even)
368 * Right after we return from this function, we turn on the MMU
369 * which means the real-mode access trick that btext does will
370 * no longer work, it needs to switch to using a real MMU
371 * mapping. This call will ensure that it does
374 #endif /* CONFIG_PPC_EARLY_DEBUG_BOOTX */
378 void early_setup_secondary(void)
380 /* Mark interrupts disabled in PACA */
381 irq_soft_mask_set(IRQS_DISABLED);
383 /* Initialize the hash table or TLB handling */
384 early_init_mmu_secondary();
387 * At this point, we can let interrupts switch to virtual mode
388 * (the MMU has been setup), so adjust the MSR in the PACA to
389 * have IR and DR set.
391 cpu_ready_for_interrupts();
394 #endif /* CONFIG_SMP */
396 void panic_smp_self_stop(void)
404 #if defined(CONFIG_SMP) || defined(CONFIG_KEXEC_CORE)
405 static bool use_spinloop(void)
407 if (IS_ENABLED(CONFIG_PPC_BOOK3S)) {
409 * See comments in head_64.S -- not all platforms insert
410 * secondaries at __secondary_hold and wait at the spin
413 if (firmware_has_feature(FW_FEATURE_OPAL))
419 * When book3e boots from kexec, the ePAPR spin table does
422 return of_property_read_bool(of_chosen, "linux,booted-from-kexec");
425 void smp_release_cpus(void)
433 DBG(" -> smp_release_cpus()\n");
435 /* All secondary cpus are spinning on a common spinloop, release them
436 * all now so they can start to spin on their individual paca
437 * spinloops. For non SMP kernels, the secondary cpus never get out
438 * of the common spinloop.
441 ptr = (unsigned long *)((unsigned long)&__secondary_hold_spinloop
443 *ptr = ppc_function_entry(generic_secondary_smp_init);
445 /* And wait a bit for them to catch up */
446 for (i = 0; i < 100000; i++) {
449 if (spinning_secondaries == 0)
453 DBG("spinning_secondaries = %d\n", spinning_secondaries);
455 DBG(" <- smp_release_cpus()\n");
457 #endif /* CONFIG_SMP || CONFIG_KEXEC_CORE */
460 * Initialize some remaining members of the ppc64_caches and systemcfg
462 * (at least until we get rid of them completely). This is mostly some
463 * cache informations about the CPU that will be used by cache flush
464 * routines and/or provided to userland
467 static void init_cache_info(struct ppc_cache_info *info, u32 size, u32 lsize,
472 info->line_size = lsize;
473 info->block_size = bsize;
474 info->log_block_size = __ilog2(bsize);
476 info->blocks_per_page = PAGE_SIZE / bsize;
478 info->blocks_per_page = 0;
481 info->assoc = 0xffff;
483 info->assoc = size / (sets * lsize);
486 static bool __init parse_cache_info(struct device_node *np,
488 struct ppc_cache_info *info)
490 static const char *ipropnames[] __initdata = {
493 "i-cache-block-size",
496 static const char *dpropnames[] __initdata = {
499 "d-cache-block-size",
502 const char **propnames = icache ? ipropnames : dpropnames;
503 const __be32 *sizep, *lsizep, *bsizep, *setsp;
504 u32 size, lsize, bsize, sets;
509 lsize = bsize = cur_cpu_spec->dcache_bsize;
510 sizep = of_get_property(np, propnames[0], NULL);
512 size = be32_to_cpu(*sizep);
513 setsp = of_get_property(np, propnames[1], NULL);
515 sets = be32_to_cpu(*setsp);
516 bsizep = of_get_property(np, propnames[2], NULL);
517 lsizep = of_get_property(np, propnames[3], NULL);
521 lsize = be32_to_cpu(*lsizep);
523 bsize = be32_to_cpu(*bsizep);
524 if (sizep == NULL || bsizep == NULL || lsizep == NULL)
528 * OF is weird .. it represents fully associative caches
529 * as "1 way" which doesn't make much sense and doesn't
530 * leave room for direct mapped. We'll assume that 0
531 * in OF means direct mapped for that reason.
538 init_cache_info(info, size, lsize, bsize, sets);
543 void __init initialize_cache_info(void)
545 struct device_node *cpu = NULL, *l2, *l3 = NULL;
548 DBG(" -> initialize_cache_info()\n");
551 * All shipping POWER8 machines have a firmware bug that
552 * puts incorrect information in the device-tree. This will
553 * be (hopefully) fixed for future chips but for now hard
554 * code the values if we are running on one of these
556 pvr = PVR_VER(mfspr(SPRN_PVR));
557 if (pvr == PVR_POWER8 || pvr == PVR_POWER8E ||
558 pvr == PVR_POWER8NVL) {
559 /* size lsize blk sets */
560 init_cache_info(&ppc64_caches.l1i, 0x8000, 128, 128, 32);
561 init_cache_info(&ppc64_caches.l1d, 0x10000, 128, 128, 64);
562 init_cache_info(&ppc64_caches.l2, 0x80000, 128, 0, 512);
563 init_cache_info(&ppc64_caches.l3, 0x800000, 128, 0, 8192);
565 cpu = of_find_node_by_type(NULL, "cpu");
568 * We're assuming *all* of the CPUs have the same
569 * d-cache and i-cache sizes... -Peter
572 if (!parse_cache_info(cpu, false, &ppc64_caches.l1d))
573 DBG("Argh, can't find dcache properties !\n");
575 if (!parse_cache_info(cpu, true, &ppc64_caches.l1i))
576 DBG("Argh, can't find icache properties !\n");
579 * Try to find the L2 and L3 if any. Assume they are
580 * unified and use the D-side properties.
582 l2 = of_find_next_cache_node(cpu);
585 parse_cache_info(l2, false, &ppc64_caches.l2);
586 l3 = of_find_next_cache_node(l2);
590 parse_cache_info(l3, false, &ppc64_caches.l3);
595 /* For use by binfmt_elf */
596 dcache_bsize = ppc64_caches.l1d.block_size;
597 icache_bsize = ppc64_caches.l1i.block_size;
599 cur_cpu_spec->dcache_bsize = dcache_bsize;
600 cur_cpu_spec->icache_bsize = icache_bsize;
602 DBG(" <- initialize_cache_info()\n");
606 * This returns the limit below which memory accesses to the linear
607 * mapping are guarnateed not to cause an architectural exception (e.g.,
608 * TLB or SLB miss fault).
610 * This is used to allocate PACAs and various interrupt stacks that
611 * that are accessed early in interrupt handlers that must not cause
612 * re-entrant interrupts.
614 __init u64 ppc64_bolted_size(void)
616 #ifdef CONFIG_PPC_BOOK3E
617 /* Freescale BookE bolts the entire linear mapping */
618 /* XXX: BookE ppc64_rma_limit setup seems to disagree? */
619 if (early_mmu_has_feature(MMU_FTR_TYPE_FSL_E))
620 return linear_map_top;
621 /* Other BookE, we assume the first GB is bolted */
624 /* BookS radix, does not take faults on linear mapping */
625 if (early_radix_enabled())
628 /* BookS hash, the first segment is bolted */
629 if (early_mmu_has_feature(MMU_FTR_1T_SEGMENT))
630 return 1UL << SID_SHIFT_1T;
631 return 1UL << SID_SHIFT;
635 static void *__init alloc_stack(unsigned long limit, int cpu)
639 pa = memblock_alloc_base_nid(THREAD_SIZE, THREAD_SIZE, limit,
640 early_cpu_to_node(cpu), MEMBLOCK_NONE);
642 pa = memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit);
644 panic("cannot allocate stacks");
650 void __init irqstack_early_init(void)
652 u64 limit = ppc64_bolted_size();
656 * Interrupt stacks must be in the first segment since we
657 * cannot afford to take SLB misses on them. They are not
658 * accessed in realmode.
660 for_each_possible_cpu(i) {
661 softirq_ctx[i] = alloc_stack(limit, i);
662 hardirq_ctx[i] = alloc_stack(limit, i);
666 #ifdef CONFIG_PPC_BOOK3E
667 void __init exc_lvl_early_init(void)
671 for_each_possible_cpu(i) {
674 sp = alloc_stack(ULONG_MAX, i);
676 paca_ptrs[i]->crit_kstack = sp + THREAD_SIZE;
678 sp = alloc_stack(ULONG_MAX, i);
680 paca_ptrs[i]->dbg_kstack = sp + THREAD_SIZE;
682 sp = alloc_stack(ULONG_MAX, i);
683 mcheckirq_ctx[i] = sp;
684 paca_ptrs[i]->mc_kstack = sp + THREAD_SIZE;
687 if (cpu_has_feature(CPU_FTR_DEBUG_LVL_EXC))
688 patch_exception(0x040, exc_debug_debug_book3e);
693 * Emergency stacks are used for a range of things, from asynchronous
694 * NMIs (system reset, machine check) to synchronous, process context.
695 * We set preempt_count to zero, even though that isn't necessarily correct. To
696 * get the right value we'd need to copy it from the previous thread_info, but
697 * doing that might fault causing more problems.
698 * TODO: what to do with accounting?
700 static void emerg_stack_init_thread_info(struct thread_info *ti, int cpu)
704 ti->preempt_count = 0;
707 klp_init_thread_info(ti);
711 * Stack space used when we detect a bad kernel stack pointer, and
712 * early in SMP boots before relocation is enabled. Exclusive emergency
713 * stack for machine checks.
715 void __init emergency_stack_init(void)
721 * Emergency stacks must be under 256MB, we cannot afford to take
722 * SLB misses on them. The ABI also requires them to be 128-byte
725 * Since we use these as temporary stacks during secondary CPU
726 * bringup, machine check, system reset, and HMI, we need to get
727 * at them in real mode. This means they must also be within the RMO
730 * The IRQ stacks allocated elsewhere in this file are zeroed and
731 * initialized in kernel/irq.c. These are initialized here in order
732 * to have emergency stacks available as early as possible.
734 limit = min(ppc64_bolted_size(), ppc64_rma_size);
736 for_each_possible_cpu(i) {
737 struct thread_info *ti;
739 ti = alloc_stack(limit, i);
740 memset(ti, 0, THREAD_SIZE);
741 emerg_stack_init_thread_info(ti, i);
742 paca_ptrs[i]->emergency_sp = (void *)ti + THREAD_SIZE;
744 #ifdef CONFIG_PPC_BOOK3S_64
745 /* emergency stack for NMI exception handling. */
746 ti = alloc_stack(limit, i);
747 memset(ti, 0, THREAD_SIZE);
748 emerg_stack_init_thread_info(ti, i);
749 paca_ptrs[i]->nmi_emergency_sp = (void *)ti + THREAD_SIZE;
751 /* emergency stack for machine check exception handling. */
752 ti = alloc_stack(limit, i);
753 memset(ti, 0, THREAD_SIZE);
754 emerg_stack_init_thread_info(ti, i);
755 paca_ptrs[i]->mc_emergency_sp = (void *)ti + THREAD_SIZE;
761 #define PCPU_DYN_SIZE ()
763 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
765 return memblock_alloc_try_nid(size, align, __pa(MAX_DMA_ADDRESS),
766 MEMBLOCK_ALLOC_ACCESSIBLE,
767 early_cpu_to_node(cpu));
771 static void __init pcpu_fc_free(void *ptr, size_t size)
773 memblock_free(__pa(ptr), size);
776 static int pcpu_cpu_distance(unsigned int from, unsigned int to)
778 if (early_cpu_to_node(from) == early_cpu_to_node(to))
779 return LOCAL_DISTANCE;
781 return REMOTE_DISTANCE;
784 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
785 EXPORT_SYMBOL(__per_cpu_offset);
787 void __init setup_per_cpu_areas(void)
789 const size_t dyn_size = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
796 * Linear mapping is one of 4K, 1M and 16M. For 4K, no need
797 * to group units. For larger mappings, use 1M atom which
798 * should be large enough to contain a number of units.
800 if (mmu_linear_psize == MMU_PAGE_4K)
801 atom_size = PAGE_SIZE;
805 rc = pcpu_embed_first_chunk(0, dyn_size, atom_size, pcpu_cpu_distance,
806 pcpu_fc_alloc, pcpu_fc_free);
808 panic("cannot initialize percpu area (err=%d)", rc);
810 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
811 for_each_possible_cpu(cpu) {
812 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
813 paca_ptrs[cpu]->data_offset = __per_cpu_offset[cpu];
818 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
819 unsigned long memory_block_size_bytes(void)
821 if (ppc_md.memory_block_size)
822 return ppc_md.memory_block_size();
824 return MIN_MEMORY_BLOCK_SIZE;
828 #if defined(CONFIG_PPC_INDIRECT_PIO) || defined(CONFIG_PPC_INDIRECT_MMIO)
829 struct ppc_pci_io ppc_pci_io;
830 EXPORT_SYMBOL(ppc_pci_io);
833 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
834 u64 hw_nmi_get_sample_period(int watchdog_thresh)
836 return ppc_proc_freq * watchdog_thresh;
841 * The perf based hardlockup detector breaks PMU event based branches, so
842 * disable it by default. Book3S has a soft-nmi hardlockup detector based
843 * on the decrementer interrupt, so it does not suffer from this problem.
845 * It is likely to get false positives in VM guests, so disable it there
848 static int __init disable_hardlockup_detector(void)
850 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
851 hardlockup_detector_disable();
853 if (firmware_has_feature(FW_FEATURE_LPAR))
854 hardlockup_detector_disable();
859 early_initcall(disable_hardlockup_detector);
861 #ifdef CONFIG_PPC_BOOK3S_64
862 static enum l1d_flush_type enabled_flush_types;
863 static void *l1d_flush_fallback_area;
864 static bool no_rfi_flush;
867 static int __init handle_no_rfi_flush(char *p)
869 pr_info("rfi-flush: disabled on command line.");
873 early_param("no_rfi_flush", handle_no_rfi_flush);
876 * The RFI flush is not KPTI, but because users will see doco that says to use
877 * nopti we hijack that option here to also disable the RFI flush.
879 static int __init handle_no_pti(char *p)
881 pr_info("rfi-flush: disabling due to 'nopti' on command line.\n");
882 handle_no_rfi_flush(NULL);
885 early_param("nopti", handle_no_pti);
887 static void do_nothing(void *unused)
890 * We don't need to do the flush explicitly, just enter+exit kernel is
891 * sufficient, the RFI exit handlers will do the right thing.
895 void rfi_flush_enable(bool enable)
898 do_rfi_flush_fixups(enabled_flush_types);
899 on_each_cpu(do_nothing, NULL, 1);
901 do_rfi_flush_fixups(L1D_FLUSH_NONE);
906 static void __ref init_fallback_flush(void)
911 /* Only allocate the fallback flush area once (at boot time). */
912 if (l1d_flush_fallback_area)
915 l1d_size = ppc64_caches.l1d.size;
918 * If there is no d-cache-size property in the device tree, l1d_size
919 * could be zero. That leads to the loop in the asm wrapping around to
920 * 2^64-1, and then walking off the end of the fallback area and
921 * eventually causing a page fault which is fatal. Just default to
922 * something vaguely sane.
925 l1d_size = (64 * 1024);
927 limit = min(ppc64_bolted_size(), ppc64_rma_size);
930 * Align to L1d size, and size it at 2x L1d size, to catch possible
931 * hardware prefetch runoff. We don't have a recipe for load patterns to
932 * reliably avoid the prefetcher.
934 l1d_flush_fallback_area = __va(memblock_alloc_base(l1d_size * 2, l1d_size, limit));
935 memset(l1d_flush_fallback_area, 0, l1d_size * 2);
937 for_each_possible_cpu(cpu) {
938 struct paca_struct *paca = paca_ptrs[cpu];
939 paca->rfi_flush_fallback_area = l1d_flush_fallback_area;
940 paca->l1d_flush_size = l1d_size;
944 void setup_rfi_flush(enum l1d_flush_type types, bool enable)
946 if (types & L1D_FLUSH_FALLBACK) {
947 pr_info("rfi-flush: fallback displacement flush available\n");
948 init_fallback_flush();
951 if (types & L1D_FLUSH_ORI)
952 pr_info("rfi-flush: ori type flush available\n");
954 if (types & L1D_FLUSH_MTTRIG)
955 pr_info("rfi-flush: mttrig type flush available\n");
957 enabled_flush_types = types;
960 rfi_flush_enable(enable);
963 #ifdef CONFIG_DEBUG_FS
964 static int rfi_flush_set(void *data, u64 val)
975 /* Only do anything if we're changing state */
976 if (enable != rfi_flush)
977 rfi_flush_enable(enable);
982 static int rfi_flush_get(void *data, u64 *val)
984 *val = rfi_flush ? 1 : 0;
988 DEFINE_SIMPLE_ATTRIBUTE(fops_rfi_flush, rfi_flush_get, rfi_flush_set, "%llu\n");
990 static __init int rfi_flush_debugfs_init(void)
992 debugfs_create_file("rfi_flush", 0600, powerpc_debugfs_root, NULL, &fops_rfi_flush);
995 device_initcall(rfi_flush_debugfs_init);
997 #endif /* CONFIG_PPC_BOOK3S_64 */