1 // SPDX-License-Identifier: GPL-2.0-or-later
4 * Copyright (C) 2001 Todd Inglett, IBM Corporation
6 * pSeries LPAR support.
9 /* Enables debugging of low-level hash table routines - careful! */
11 #define pr_fmt(fmt) "lpar: " fmt
13 #include <linux/kernel.h>
14 #include <linux/dma-mapping.h>
15 #include <linux/console.h>
16 #include <linux/export.h>
17 #include <linux/jump_label.h>
18 #include <linux/delay.h>
19 #include <linux/stop_machine.h>
20 #include <linux/spinlock.h>
21 #include <linux/cpuhotplug.h>
22 #include <linux/workqueue.h>
23 #include <linux/proc_fs.h>
24 #include <linux/pgtable.h>
25 #include <linux/debugfs.h>
27 #include <asm/processor.h>
30 #include <asm/setup.h>
31 #include <asm/mmu_context.h>
32 #include <asm/iommu.h>
34 #include <asm/cputable.h>
35 #include <asm/papr-sysparm.h>
38 #include <asm/trace.h>
39 #include <asm/firmware.h>
40 #include <asm/plpar_wrappers.h>
41 #include <asm/kexec.h>
42 #include <asm/fadump.h>
48 /* Flag bits for H_BULK_REMOVE */
49 #define HBR_REQUEST 0x4000000000000000UL
50 #define HBR_RESPONSE 0x8000000000000000UL
51 #define HBR_END 0xc000000000000000UL
52 #define HBR_AVPN 0x0200000000000000UL
53 #define HBR_ANDCOND 0x0100000000000000UL
57 EXPORT_SYMBOL(plpar_hcall);
58 EXPORT_SYMBOL(plpar_hcall9);
59 EXPORT_SYMBOL(plpar_hcall_norets);
61 #ifdef CONFIG_PPC_64S_HASH_MMU
63 * H_BLOCK_REMOVE supported block size for this page size in segment who's base
64 * page size is that page size.
66 * The first index is the segment base page size, the second one is the actual
69 static int hblkrm_size[MMU_PAGE_COUNT][MMU_PAGE_COUNT] __ro_after_init;
73 * Due to the involved complexity, and that the current hypervisor is only
74 * returning this value or 0, we are limiting the support of the H_BLOCK_REMOVE
75 * buffer size to 8 size block.
77 #define HBLKRM_SUPPORTED_BLOCK_SIZE 8
79 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
80 static u8 dtl_mask = DTL_LOG_PREEMPT;
85 void alloc_dtl_buffers(unsigned long *time_limit)
88 struct paca_struct *pp;
89 struct dtl_entry *dtl;
91 for_each_possible_cpu(cpu) {
95 dtl = kmem_cache_alloc(dtl_cache, GFP_KERNEL);
97 pr_warn("Failed to allocate dispatch trace log for cpu %d\n",
99 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
100 pr_warn("Stolen time statistics will be unreliable\n");
106 pp->dispatch_log = dtl;
107 pp->dispatch_log_end = dtl + N_DISPATCH_LOG;
110 if (time_limit && time_after(jiffies, *time_limit)) {
112 *time_limit = jiffies + HZ;
117 void register_dtl_buffer(int cpu)
120 struct paca_struct *pp;
121 struct dtl_entry *dtl;
122 int hwcpu = get_hard_smp_processor_id(cpu);
125 dtl = pp->dispatch_log;
126 if (dtl && dtl_mask) {
129 lppaca_of(cpu).dtl_idx = 0;
131 /* hypervisor reads buffer length from this field */
132 dtl->enqueue_to_dispatch_time = cpu_to_be32(DISPATCH_LOG_BYTES);
133 ret = register_dtl(hwcpu, __pa(dtl));
135 pr_err("WARNING: DTL registration of cpu %d (hw %d) failed with %ld\n",
138 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
142 #ifdef CONFIG_PPC_SPLPAR
144 struct delayed_work work;
148 struct vcpu_dispatch_data {
159 int numa_remote_disp;
164 * This represents the number of cpus in the hypervisor. Since there is no
165 * architected way to discover the number of processors in the host, we
166 * provision for dealing with NR_CPUS. This is currently 2048 by default, and
167 * is sufficient for our purposes. This will need to be tweaked if
168 * CONFIG_NR_CPUS is changed.
170 #define NR_CPUS_H NR_CPUS
172 DEFINE_RWLOCK(dtl_access_lock);
173 static DEFINE_PER_CPU(struct vcpu_dispatch_data, vcpu_disp_data);
174 static DEFINE_PER_CPU(u64, dtl_entry_ridx);
175 static DEFINE_PER_CPU(struct dtl_worker, dtl_workers);
176 static enum cpuhp_state dtl_worker_state;
177 static DEFINE_MUTEX(dtl_enable_mutex);
178 static int vcpudispatch_stats_on __read_mostly;
179 static int vcpudispatch_stats_freq = 50;
180 static __be32 *vcpu_associativity, *pcpu_associativity;
183 static void free_dtl_buffers(unsigned long *time_limit)
185 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
187 struct paca_struct *pp;
189 for_each_possible_cpu(cpu) {
191 if (!pp->dispatch_log)
193 kmem_cache_free(dtl_cache, pp->dispatch_log);
195 pp->dispatch_log = NULL;
196 pp->dispatch_log_end = NULL;
199 if (time_limit && time_after(jiffies, *time_limit)) {
201 *time_limit = jiffies + HZ;
207 static int init_cpu_associativity(void)
209 vcpu_associativity = kcalloc(num_possible_cpus() / threads_per_core,
210 VPHN_ASSOC_BUFSIZE * sizeof(__be32), GFP_KERNEL);
211 pcpu_associativity = kcalloc(NR_CPUS_H / threads_per_core,
212 VPHN_ASSOC_BUFSIZE * sizeof(__be32), GFP_KERNEL);
214 if (!vcpu_associativity || !pcpu_associativity) {
215 pr_err("error allocating memory for associativity information\n");
222 static void destroy_cpu_associativity(void)
224 kfree(vcpu_associativity);
225 kfree(pcpu_associativity);
226 vcpu_associativity = pcpu_associativity = NULL;
229 static __be32 *__get_cpu_associativity(int cpu, __be32 *cpu_assoc, int flag)
234 assoc = &cpu_assoc[(int)(cpu / threads_per_core) * VPHN_ASSOC_BUFSIZE];
236 rc = hcall_vphn(cpu, flag, &assoc[0]);
244 static __be32 *get_pcpu_associativity(int cpu)
246 return __get_cpu_associativity(cpu, pcpu_associativity, VPHN_FLAG_PCPU);
249 static __be32 *get_vcpu_associativity(int cpu)
251 return __get_cpu_associativity(cpu, vcpu_associativity, VPHN_FLAG_VCPU);
254 static int cpu_relative_dispatch_distance(int last_disp_cpu, int cur_disp_cpu)
256 __be32 *last_disp_cpu_assoc, *cur_disp_cpu_assoc;
258 if (last_disp_cpu >= NR_CPUS_H || cur_disp_cpu >= NR_CPUS_H)
261 last_disp_cpu_assoc = get_pcpu_associativity(last_disp_cpu);
262 cur_disp_cpu_assoc = get_pcpu_associativity(cur_disp_cpu);
264 if (!last_disp_cpu_assoc || !cur_disp_cpu_assoc)
267 return cpu_relative_distance(last_disp_cpu_assoc, cur_disp_cpu_assoc);
270 static int cpu_home_node_dispatch_distance(int disp_cpu)
272 __be32 *disp_cpu_assoc, *vcpu_assoc;
273 int vcpu_id = smp_processor_id();
275 if (disp_cpu >= NR_CPUS_H) {
276 pr_debug_ratelimited("vcpu dispatch cpu %d > %d\n",
277 disp_cpu, NR_CPUS_H);
281 disp_cpu_assoc = get_pcpu_associativity(disp_cpu);
282 vcpu_assoc = get_vcpu_associativity(vcpu_id);
284 if (!disp_cpu_assoc || !vcpu_assoc)
287 return cpu_relative_distance(disp_cpu_assoc, vcpu_assoc);
290 static void update_vcpu_disp_stat(int disp_cpu)
292 struct vcpu_dispatch_data *disp;
295 disp = this_cpu_ptr(&vcpu_disp_data);
296 if (disp->last_disp_cpu == -1) {
297 disp->last_disp_cpu = disp_cpu;
303 if (disp->last_disp_cpu == disp_cpu ||
304 (cpu_first_thread_sibling(disp->last_disp_cpu) ==
305 cpu_first_thread_sibling(disp_cpu)))
306 disp->same_cpu_disp++;
308 distance = cpu_relative_dispatch_distance(disp->last_disp_cpu,
311 pr_debug_ratelimited("vcpudispatch_stats: cpu %d: error determining associativity\n",
316 disp->same_chip_disp++;
319 disp->diff_chip_disp++;
322 disp->far_chip_disp++;
325 pr_debug_ratelimited("vcpudispatch_stats: cpu %d (%d -> %d): unexpected relative dispatch distance %d\n",
334 distance = cpu_home_node_dispatch_distance(disp_cpu);
336 pr_debug_ratelimited("vcpudispatch_stats: cpu %d: error determining associativity\n",
341 disp->numa_home_disp++;
344 disp->numa_remote_disp++;
347 disp->numa_far_disp++;
350 pr_debug_ratelimited("vcpudispatch_stats: cpu %d on %d: unexpected numa dispatch distance %d\n",
357 disp->last_disp_cpu = disp_cpu;
360 static void process_dtl_buffer(struct work_struct *work)
362 struct dtl_entry dtle;
363 u64 i = __this_cpu_read(dtl_entry_ridx);
364 struct dtl_entry *dtl = local_paca->dispatch_log + (i % N_DISPATCH_LOG);
365 struct dtl_entry *dtl_end = local_paca->dispatch_log_end;
366 struct lppaca *vpa = local_paca->lppaca_ptr;
367 struct dtl_worker *d = container_of(work, struct dtl_worker, work.work);
369 if (!local_paca->dispatch_log)
372 /* if we have been migrated away, we cancel ourself */
373 if (d->cpu != smp_processor_id()) {
374 pr_debug("vcpudispatch_stats: cpu %d worker migrated -- canceling worker\n",
379 if (i == be64_to_cpu(vpa->dtl_idx))
382 while (i < be64_to_cpu(vpa->dtl_idx)) {
385 if (i + N_DISPATCH_LOG < be64_to_cpu(vpa->dtl_idx)) {
386 /* buffer has overflowed */
387 pr_debug_ratelimited("vcpudispatch_stats: cpu %d lost %lld DTL samples\n",
389 be64_to_cpu(vpa->dtl_idx) - N_DISPATCH_LOG - i);
390 i = be64_to_cpu(vpa->dtl_idx) - N_DISPATCH_LOG;
391 dtl = local_paca->dispatch_log + (i % N_DISPATCH_LOG);
394 update_vcpu_disp_stat(be16_to_cpu(dtle.processor_id));
398 dtl = local_paca->dispatch_log;
401 __this_cpu_write(dtl_entry_ridx, i);
404 schedule_delayed_work_on(d->cpu, to_delayed_work(work),
405 HZ / vcpudispatch_stats_freq);
408 static int dtl_worker_online(unsigned int cpu)
410 struct dtl_worker *d = &per_cpu(dtl_workers, cpu);
412 memset(d, 0, sizeof(*d));
413 INIT_DELAYED_WORK(&d->work, process_dtl_buffer);
416 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
417 per_cpu(dtl_entry_ridx, cpu) = 0;
418 register_dtl_buffer(cpu);
420 per_cpu(dtl_entry_ridx, cpu) = be64_to_cpu(lppaca_of(cpu).dtl_idx);
423 schedule_delayed_work_on(cpu, &d->work, HZ / vcpudispatch_stats_freq);
427 static int dtl_worker_offline(unsigned int cpu)
429 struct dtl_worker *d = &per_cpu(dtl_workers, cpu);
431 cancel_delayed_work_sync(&d->work);
433 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
434 unregister_dtl(get_hard_smp_processor_id(cpu));
440 static void set_global_dtl_mask(u8 mask)
445 for_each_present_cpu(cpu)
446 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
449 static void reset_global_dtl_mask(void)
453 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
454 dtl_mask = DTL_LOG_PREEMPT;
458 for_each_present_cpu(cpu)
459 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
462 static int dtl_worker_enable(unsigned long *time_limit)
466 if (!write_trylock(&dtl_access_lock)) {
471 set_global_dtl_mask(DTL_LOG_ALL);
473 /* Setup dtl buffers and register those */
474 alloc_dtl_buffers(time_limit);
476 state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "powerpc/dtl:online",
477 dtl_worker_online, dtl_worker_offline);
479 pr_err("vcpudispatch_stats: unable to setup workqueue for DTL processing\n");
480 free_dtl_buffers(time_limit);
481 reset_global_dtl_mask();
482 write_unlock(&dtl_access_lock);
486 dtl_worker_state = state;
492 static void dtl_worker_disable(unsigned long *time_limit)
494 cpuhp_remove_state(dtl_worker_state);
495 free_dtl_buffers(time_limit);
496 reset_global_dtl_mask();
497 write_unlock(&dtl_access_lock);
500 static ssize_t vcpudispatch_stats_write(struct file *file, const char __user *p,
501 size_t count, loff_t *ppos)
503 unsigned long time_limit = jiffies + HZ;
504 struct vcpu_dispatch_data *disp;
511 if (copy_from_user(buf, p, count))
515 rc = kstrtoint(buf, 0, &cmd);
516 if (rc || cmd < 0 || cmd > 1) {
517 pr_err("vcpudispatch_stats: please use 0 to disable or 1 to enable dispatch statistics\n");
518 return rc ? rc : -EINVAL;
521 mutex_lock(&dtl_enable_mutex);
523 if ((cmd == 0 && !vcpudispatch_stats_on) ||
524 (cmd == 1 && vcpudispatch_stats_on))
528 rc = init_cpu_associativity();
530 destroy_cpu_associativity();
534 for_each_possible_cpu(cpu) {
535 disp = per_cpu_ptr(&vcpu_disp_data, cpu);
536 memset(disp, 0, sizeof(*disp));
537 disp->last_disp_cpu = -1;
540 rc = dtl_worker_enable(&time_limit);
542 destroy_cpu_associativity();
546 dtl_worker_disable(&time_limit);
547 destroy_cpu_associativity();
550 vcpudispatch_stats_on = cmd;
553 mutex_unlock(&dtl_enable_mutex);
559 static int vcpudispatch_stats_display(struct seq_file *p, void *v)
562 struct vcpu_dispatch_data *disp;
564 if (!vcpudispatch_stats_on) {
565 seq_puts(p, "off\n");
569 for_each_online_cpu(cpu) {
570 disp = per_cpu_ptr(&vcpu_disp_data, cpu);
571 seq_printf(p, "cpu%d", cpu);
572 seq_put_decimal_ull(p, " ", disp->total_disp);
573 seq_put_decimal_ull(p, " ", disp->same_cpu_disp);
574 seq_put_decimal_ull(p, " ", disp->same_chip_disp);
575 seq_put_decimal_ull(p, " ", disp->diff_chip_disp);
576 seq_put_decimal_ull(p, " ", disp->far_chip_disp);
577 seq_put_decimal_ull(p, " ", disp->numa_home_disp);
578 seq_put_decimal_ull(p, " ", disp->numa_remote_disp);
579 seq_put_decimal_ull(p, " ", disp->numa_far_disp);
586 static int vcpudispatch_stats_open(struct inode *inode, struct file *file)
588 return single_open(file, vcpudispatch_stats_display, NULL);
591 static const struct proc_ops vcpudispatch_stats_proc_ops = {
592 .proc_open = vcpudispatch_stats_open,
593 .proc_read = seq_read,
594 .proc_write = vcpudispatch_stats_write,
595 .proc_lseek = seq_lseek,
596 .proc_release = single_release,
599 static ssize_t vcpudispatch_stats_freq_write(struct file *file,
600 const char __user *p, size_t count, loff_t *ppos)
608 if (copy_from_user(buf, p, count))
612 rc = kstrtoint(buf, 0, &freq);
613 if (rc || freq < 1 || freq > HZ) {
614 pr_err("vcpudispatch_stats_freq: please specify a frequency between 1 and %d\n",
616 return rc ? rc : -EINVAL;
619 vcpudispatch_stats_freq = freq;
624 static int vcpudispatch_stats_freq_display(struct seq_file *p, void *v)
626 seq_printf(p, "%d\n", vcpudispatch_stats_freq);
630 static int vcpudispatch_stats_freq_open(struct inode *inode, struct file *file)
632 return single_open(file, vcpudispatch_stats_freq_display, NULL);
635 static const struct proc_ops vcpudispatch_stats_freq_proc_ops = {
636 .proc_open = vcpudispatch_stats_freq_open,
637 .proc_read = seq_read,
638 .proc_write = vcpudispatch_stats_freq_write,
639 .proc_lseek = seq_lseek,
640 .proc_release = single_release,
643 static int __init vcpudispatch_stats_procfs_init(void)
645 if (!lppaca_shared_proc())
648 if (!proc_create("powerpc/vcpudispatch_stats", 0600, NULL,
649 &vcpudispatch_stats_proc_ops))
650 pr_err("vcpudispatch_stats: error creating procfs file\n");
651 else if (!proc_create("powerpc/vcpudispatch_stats_freq", 0600, NULL,
652 &vcpudispatch_stats_freq_proc_ops))
653 pr_err("vcpudispatch_stats_freq: error creating procfs file\n");
658 machine_device_initcall(pseries, vcpudispatch_stats_procfs_init);
660 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
661 u64 pseries_paravirt_steal_clock(int cpu)
663 struct lppaca *lppaca = &lppaca_of(cpu);
665 return be64_to_cpu(READ_ONCE(lppaca->enqueue_dispatch_tb)) +
666 be64_to_cpu(READ_ONCE(lppaca->ready_enqueue_tb));
670 #endif /* CONFIG_PPC_SPLPAR */
672 void vpa_init(int cpu)
674 int hwcpu = get_hard_smp_processor_id(cpu);
679 * The spec says it "may be problematic" if CPU x registers the VPA of
680 * CPU y. We should never do that, but wail if we ever do.
682 WARN_ON(cpu != smp_processor_id());
684 if (cpu_has_feature(CPU_FTR_ALTIVEC))
685 lppaca_of(cpu).vmxregs_in_use = 1;
687 if (cpu_has_feature(CPU_FTR_ARCH_207S))
688 lppaca_of(cpu).ebb_regs_in_use = 1;
690 addr = __pa(&lppaca_of(cpu));
691 ret = register_vpa(hwcpu, addr);
694 pr_err("WARNING: VPA registration for cpu %d (hw %d) of area "
695 "%lx failed with %ld\n", cpu, hwcpu, addr, ret);
699 #ifdef CONFIG_PPC_64S_HASH_MMU
701 * PAPR says this feature is SLB-Buffer but firmware never
702 * reports that. All SPLPAR support SLB shadow buffer.
704 if (!radix_enabled() && firmware_has_feature(FW_FEATURE_SPLPAR)) {
705 addr = __pa(paca_ptrs[cpu]->slb_shadow_ptr);
706 ret = register_slb_shadow(hwcpu, addr);
708 pr_err("WARNING: SLB shadow buffer registration for "
709 "cpu %d (hw %d) of area %lx failed with %ld\n",
710 cpu, hwcpu, addr, ret);
712 #endif /* CONFIG_PPC_64S_HASH_MMU */
715 * Register dispatch trace log, if one has been allocated.
717 register_dtl_buffer(cpu);
720 #ifdef CONFIG_PPC_BOOK3S_64
722 static int __init pseries_lpar_register_process_table(unsigned long base,
723 unsigned long page_size, unsigned long table_size)
726 unsigned long flags = 0;
729 flags |= PROC_TABLE_NEW;
730 if (radix_enabled()) {
731 flags |= PROC_TABLE_RADIX;
732 if (mmu_has_feature(MMU_FTR_GTSE))
733 flags |= PROC_TABLE_GTSE;
735 flags |= PROC_TABLE_HPT_SLB;
737 rc = plpar_hcall_norets(H_REGISTER_PROC_TBL, flags, base,
738 page_size, table_size);
739 if (!H_IS_LONG_BUSY(rc))
741 mdelay(get_longbusy_msecs(rc));
743 if (rc != H_SUCCESS) {
744 pr_err("Failed to register process table (rc=%ld)\n", rc);
750 #ifdef CONFIG_PPC_64S_HASH_MMU
752 static long pSeries_lpar_hpte_insert(unsigned long hpte_group,
753 unsigned long vpn, unsigned long pa,
754 unsigned long rflags, unsigned long vflags,
755 int psize, int apsize, int ssize)
757 unsigned long lpar_rc;
760 unsigned long hpte_v, hpte_r;
762 if (!(vflags & HPTE_V_BOLTED))
763 pr_devel("hpte_insert(group=%lx, vpn=%016lx, "
764 "pa=%016lx, rflags=%lx, vflags=%lx, psize=%d)\n",
765 hpte_group, vpn, pa, rflags, vflags, psize);
767 hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID;
768 hpte_r = hpte_encode_r(pa, psize, apsize) | rflags;
770 if (!(vflags & HPTE_V_BOLTED))
771 pr_devel(" hpte_v=%016lx, hpte_r=%016lx\n", hpte_v, hpte_r);
773 /* Now fill in the actual HPTE */
774 /* Set CEC cookie to 0 */
776 /* I-cache Invalidate = 0 */
777 /* I-cache synchronize = 0 */
781 if (firmware_has_feature(FW_FEATURE_XCMO) && !(hpte_r & HPTE_R_N))
782 flags |= H_COALESCE_CAND;
784 lpar_rc = plpar_pte_enter(flags, hpte_group, hpte_v, hpte_r, &slot);
785 if (unlikely(lpar_rc == H_PTEG_FULL)) {
786 pr_devel("Hash table group is full\n");
791 * Since we try and ioremap PHBs we don't own, the pte insert
792 * will fail. However we must catch the failure in hash_page
793 * or we will loop forever, so return -2 in this case.
795 if (unlikely(lpar_rc != H_SUCCESS)) {
796 pr_err("Failed hash pte insert with error %ld\n", lpar_rc);
799 if (!(vflags & HPTE_V_BOLTED))
800 pr_devel(" -> slot: %lu\n", slot & 7);
802 /* Because of iSeries, we have to pass down the secondary
803 * bucket bit here as well
805 return (slot & 7) | (!!(vflags & HPTE_V_SECONDARY) << 3);
808 static DEFINE_SPINLOCK(pSeries_lpar_tlbie_lock);
810 static long pSeries_lpar_hpte_remove(unsigned long hpte_group)
812 unsigned long slot_offset;
813 unsigned long lpar_rc;
815 unsigned long dummy1, dummy2;
817 /* pick a random slot to start at */
818 slot_offset = mftb() & 0x7;
820 for (i = 0; i < HPTES_PER_GROUP; i++) {
822 /* don't remove a bolted entry */
823 lpar_rc = plpar_pte_remove(H_ANDCOND, hpte_group + slot_offset,
824 HPTE_V_BOLTED, &dummy1, &dummy2);
825 if (lpar_rc == H_SUCCESS)
829 * The test for adjunct partition is performed before the
830 * ANDCOND test. H_RESOURCE may be returned, so we need to
831 * check for that as well.
833 BUG_ON(lpar_rc != H_NOT_FOUND && lpar_rc != H_RESOURCE);
842 /* Called during kexec sequence with MMU off */
843 static notrace void manual_hpte_clear_all(void)
845 unsigned long size_bytes = 1UL << ppc64_pft_size;
846 unsigned long hpte_count = size_bytes >> 4;
854 /* Read in batches of 4,
855 * invalidate only valid entries not in the VRMA
856 * hpte_count will be a multiple of 4
858 for (i = 0; i < hpte_count; i += 4) {
859 lpar_rc = plpar_pte_read_4_raw(0, i, (void *)ptes);
860 if (lpar_rc != H_SUCCESS) {
861 pr_info("Failed to read hash page table at %ld err %ld\n",
865 for (j = 0; j < 4; j++){
866 if ((ptes[j].pteh & HPTE_V_VRMA_MASK) ==
869 if (ptes[j].pteh & HPTE_V_VALID)
870 plpar_pte_remove_raw(0, i + j, 0,
871 &(ptes[j].pteh), &(ptes[j].ptel));
876 /* Called during kexec sequence with MMU off */
877 static notrace int hcall_hpte_clear_all(void)
882 rc = plpar_hcall_norets(H_CLEAR_HPT);
883 } while (rc == H_CONTINUE);
888 /* Called during kexec sequence with MMU off */
889 static notrace void pseries_hpte_clear_all(void)
893 rc = hcall_hpte_clear_all();
895 manual_hpte_clear_all();
897 #ifdef __LITTLE_ENDIAN__
899 * Reset exceptions to big endian.
901 * FIXME this is a hack for kexec, we need to reset the exception
902 * endian before starting the new kernel and this is a convenient place
905 * This is also called on boot when a fadump happens. In that case we
906 * must not change the exception endian mode.
908 if (firmware_has_feature(FW_FEATURE_SET_MODE) && !is_fadump_active())
909 pseries_big_endian_exceptions();
914 * NOTE: for updatepp ops we are fortunate that the linux "newpp" bits and
915 * the low 3 bits of flags happen to line up. So no transform is needed.
916 * We can probably optimize here and assume the high bits of newpp are
917 * already zero. For now I am paranoid.
919 static long pSeries_lpar_hpte_updatepp(unsigned long slot,
922 int psize, int apsize,
923 int ssize, unsigned long inv_flags)
925 unsigned long lpar_rc;
927 unsigned long want_v;
929 want_v = hpte_encode_avpn(vpn, psize, ssize);
931 flags = (newpp & (HPTE_R_PP | HPTE_R_N | HPTE_R_KEY_LO)) | H_AVPN;
932 flags |= (newpp & HPTE_R_KEY_HI) >> 48;
933 if (mmu_has_feature(MMU_FTR_KERNEL_RO))
934 /* Move pp0 into bit 8 (IBM 55) */
935 flags |= (newpp & HPTE_R_PP0) >> 55;
937 pr_devel(" update: avpnv=%016lx, hash=%016lx, f=%lx, psize: %d ...",
938 want_v, slot, flags, psize);
940 lpar_rc = plpar_pte_protect(flags, slot, want_v);
942 if (lpar_rc == H_NOT_FOUND) {
943 pr_devel("not found !\n");
949 BUG_ON(lpar_rc != H_SUCCESS);
954 static long __pSeries_lpar_hpte_find(unsigned long want_v, unsigned long hpte_group)
963 for (i = 0; i < HPTES_PER_GROUP; i += 4, hpte_group += 4) {
965 lpar_rc = plpar_pte_read_4(0, hpte_group, (void *)ptes);
966 if (lpar_rc != H_SUCCESS) {
967 pr_info("Failed to read hash page table at %ld err %ld\n",
968 hpte_group, lpar_rc);
972 for (j = 0; j < 4; j++) {
973 if (HPTE_V_COMPARE(ptes[j].pteh, want_v) &&
974 (ptes[j].pteh & HPTE_V_VALID))
982 static long pSeries_lpar_hpte_find(unsigned long vpn, int psize, int ssize)
986 unsigned long want_v;
987 unsigned long hpte_group;
989 hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize);
990 want_v = hpte_encode_avpn(vpn, psize, ssize);
993 * We try to keep bolted entries always in primary hash
994 * But in some case we can find them in secondary too.
996 hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
997 slot = __pSeries_lpar_hpte_find(want_v, hpte_group);
999 /* Try in secondary */
1000 hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
1001 slot = __pSeries_lpar_hpte_find(want_v, hpte_group);
1005 return hpte_group + slot;
1008 static void pSeries_lpar_hpte_updateboltedpp(unsigned long newpp,
1010 int psize, int ssize)
1013 unsigned long lpar_rc, slot, vsid, flags;
1015 vsid = get_kernel_vsid(ea, ssize);
1016 vpn = hpt_vpn(ea, vsid, ssize);
1018 slot = pSeries_lpar_hpte_find(vpn, psize, ssize);
1021 flags = newpp & (HPTE_R_PP | HPTE_R_N);
1022 if (mmu_has_feature(MMU_FTR_KERNEL_RO))
1023 /* Move pp0 into bit 8 (IBM 55) */
1024 flags |= (newpp & HPTE_R_PP0) >> 55;
1026 flags |= ((newpp & HPTE_R_KEY_HI) >> 48) | (newpp & HPTE_R_KEY_LO);
1028 lpar_rc = plpar_pte_protect(flags, slot, 0);
1030 BUG_ON(lpar_rc != H_SUCCESS);
1033 static void pSeries_lpar_hpte_invalidate(unsigned long slot, unsigned long vpn,
1034 int psize, int apsize,
1035 int ssize, int local)
1037 unsigned long want_v;
1038 unsigned long lpar_rc;
1039 unsigned long dummy1, dummy2;
1041 pr_devel(" inval : slot=%lx, vpn=%016lx, psize: %d, local: %d\n",
1042 slot, vpn, psize, local);
1044 want_v = hpte_encode_avpn(vpn, psize, ssize);
1045 lpar_rc = plpar_pte_remove(H_AVPN, slot, want_v, &dummy1, &dummy2);
1046 if (lpar_rc == H_NOT_FOUND)
1049 BUG_ON(lpar_rc != H_SUCCESS);
1054 * As defined in the PAPR's section 14.5.4.1.8
1055 * The control mask doesn't include the returned reference and change bit from
1056 * the processed PTE.
1058 #define HBLKR_AVPN 0x0100000000000000UL
1059 #define HBLKR_CTRL_MASK 0xf800000000000000UL
1060 #define HBLKR_CTRL_SUCCESS 0x8000000000000000UL
1061 #define HBLKR_CTRL_ERRNOTFOUND 0x8800000000000000UL
1062 #define HBLKR_CTRL_ERRBUSY 0xa000000000000000UL
1065 * Returned true if we are supporting this block size for the specified segment
1066 * base page size and actual page size.
1068 * Currently, we only support 8 size block.
1070 static inline bool is_supported_hlbkrm(int bpsize, int psize)
1072 return (hblkrm_size[bpsize][psize] == HBLKRM_SUPPORTED_BLOCK_SIZE);
1076 * H_BLOCK_REMOVE caller.
1077 * @idx should point to the latest @param entry set with a PTEX.
1078 * If PTE cannot be processed because another CPUs has already locked that
1079 * group, those entries are put back in @param starting at index 1.
1080 * If entries has to be retried and @retry_busy is set to true, these entries
1081 * are retried until success. If @retry_busy is set to false, the returned
1082 * is the number of entries yet to process.
1084 static unsigned long call_block_remove(unsigned long idx, unsigned long *param,
1087 unsigned long i, rc, new_idx;
1088 unsigned long retbuf[PLPAR_HCALL9_BUFSIZE];
1091 pr_warn("Unexpected empty call to H_BLOCK_REMOVE");
1096 if (idx > PLPAR_HCALL9_BUFSIZE) {
1097 pr_err("Too many PTEs (%lu) for H_BLOCK_REMOVE", idx);
1098 idx = PLPAR_HCALL9_BUFSIZE;
1099 } else if (idx < PLPAR_HCALL9_BUFSIZE)
1100 param[idx] = HBR_END;
1102 rc = plpar_hcall9(H_BLOCK_REMOVE, retbuf,
1104 param[1], param[2], param[3], param[4], /* TS0-7 */
1105 param[5], param[6], param[7], param[8]);
1106 if (rc == H_SUCCESS)
1109 BUG_ON(rc != H_PARTIAL);
1111 /* Check that the unprocessed entries were 'not found' or 'busy' */
1112 for (i = 0; i < idx-1; i++) {
1113 unsigned long ctrl = retbuf[i] & HBLKR_CTRL_MASK;
1115 if (ctrl == HBLKR_CTRL_ERRBUSY) {
1116 param[++new_idx] = param[i+1];
1120 BUG_ON(ctrl != HBLKR_CTRL_SUCCESS
1121 && ctrl != HBLKR_CTRL_ERRNOTFOUND);
1125 * If there were entries found busy, retry these entries if requested,
1126 * of if all the entries have to be retried.
1128 if (new_idx && (retry_busy || new_idx == (PLPAR_HCALL9_BUFSIZE-1))) {
1136 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1138 * Limit iterations holding pSeries_lpar_tlbie_lock to 3. We also need
1139 * to make sure that we avoid bouncing the hypervisor tlbie lock.
1141 #define PPC64_HUGE_HPTE_BATCH 12
1143 static void hugepage_block_invalidate(unsigned long *slot, unsigned long *vpn,
1144 int count, int psize, int ssize)
1146 unsigned long param[PLPAR_HCALL9_BUFSIZE];
1147 unsigned long shift, current_vpgb, vpgb;
1150 shift = mmu_psize_defs[psize].shift;
1152 for (i = 0; i < count; i++) {
1154 * Shifting 3 bits more on the right to get a
1155 * 8 pages aligned virtual addresse.
1157 vpgb = (vpn[i] >> (shift - VPN_SHIFT + 3));
1158 if (!pix || vpgb != current_vpgb) {
1160 * Need to start a new 8 pages block, flush
1161 * the current one if needed.
1164 (void)call_block_remove(pix, param, true);
1165 current_vpgb = vpgb;
1166 param[0] = hpte_encode_avpn(vpn[i], psize, ssize);
1170 param[pix++] = HBR_REQUEST | HBLKR_AVPN | slot[i];
1171 if (pix == PLPAR_HCALL9_BUFSIZE) {
1172 pix = call_block_remove(pix, param, false);
1174 * pix = 0 means that all the entries were
1175 * removed, we can start a new block.
1176 * Otherwise, this means that there are entries
1177 * to retry, and pix points to latest one, so
1178 * we should increment it and try to continue
1186 (void)call_block_remove(pix, param, true);
1189 static void hugepage_bulk_invalidate(unsigned long *slot, unsigned long *vpn,
1190 int count, int psize, int ssize)
1192 unsigned long param[PLPAR_HCALL9_BUFSIZE];
1193 int i = 0, pix = 0, rc;
1195 for (i = 0; i < count; i++) {
1197 if (!firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1198 pSeries_lpar_hpte_invalidate(slot[i], vpn[i], psize, 0,
1201 param[pix] = HBR_REQUEST | HBR_AVPN | slot[i];
1202 param[pix+1] = hpte_encode_avpn(vpn[i], psize, ssize);
1205 rc = plpar_hcall9(H_BULK_REMOVE, param,
1206 param[0], param[1], param[2],
1207 param[3], param[4], param[5],
1208 param[6], param[7]);
1209 BUG_ON(rc != H_SUCCESS);
1215 param[pix] = HBR_END;
1216 rc = plpar_hcall9(H_BULK_REMOVE, param, param[0], param[1],
1217 param[2], param[3], param[4], param[5],
1218 param[6], param[7]);
1219 BUG_ON(rc != H_SUCCESS);
1223 static inline void __pSeries_lpar_hugepage_invalidate(unsigned long *slot,
1225 int count, int psize,
1228 unsigned long flags = 0;
1229 int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
1232 spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
1234 /* Assuming THP size is 16M */
1235 if (is_supported_hlbkrm(psize, MMU_PAGE_16M))
1236 hugepage_block_invalidate(slot, vpn, count, psize, ssize);
1238 hugepage_bulk_invalidate(slot, vpn, count, psize, ssize);
1241 spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
1244 static void pSeries_lpar_hugepage_invalidate(unsigned long vsid,
1246 unsigned char *hpte_slot_array,
1247 int psize, int ssize, int local)
1250 unsigned long s_addr = addr;
1251 unsigned int max_hpte_count, valid;
1252 unsigned long vpn_array[PPC64_HUGE_HPTE_BATCH];
1253 unsigned long slot_array[PPC64_HUGE_HPTE_BATCH];
1254 unsigned long shift, hidx, vpn = 0, hash, slot;
1256 shift = mmu_psize_defs[psize].shift;
1257 max_hpte_count = 1U << (PMD_SHIFT - shift);
1259 for (i = 0; i < max_hpte_count; i++) {
1260 valid = hpte_valid(hpte_slot_array, i);
1263 hidx = hpte_hash_index(hpte_slot_array, i);
1266 addr = s_addr + (i * (1ul << shift));
1267 vpn = hpt_vpn(addr, vsid, ssize);
1268 hash = hpt_hash(vpn, shift, ssize);
1269 if (hidx & _PTEIDX_SECONDARY)
1272 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1273 slot += hidx & _PTEIDX_GROUP_IX;
1275 slot_array[index] = slot;
1276 vpn_array[index] = vpn;
1277 if (index == PPC64_HUGE_HPTE_BATCH - 1) {
1279 * Now do a bluk invalidate
1281 __pSeries_lpar_hugepage_invalidate(slot_array,
1283 PPC64_HUGE_HPTE_BATCH,
1290 __pSeries_lpar_hugepage_invalidate(slot_array, vpn_array,
1291 index, psize, ssize);
1294 static void pSeries_lpar_hugepage_invalidate(unsigned long vsid,
1296 unsigned char *hpte_slot_array,
1297 int psize, int ssize, int local)
1299 WARN(1, "%s called without THP support\n", __func__);
1303 static int pSeries_lpar_hpte_removebolted(unsigned long ea,
1304 int psize, int ssize)
1307 unsigned long slot, vsid;
1309 vsid = get_kernel_vsid(ea, ssize);
1310 vpn = hpt_vpn(ea, vsid, ssize);
1312 slot = pSeries_lpar_hpte_find(vpn, psize, ssize);
1317 * lpar doesn't use the passed actual page size
1319 pSeries_lpar_hpte_invalidate(slot, vpn, psize, 0, ssize, 0);
1324 static inline unsigned long compute_slot(real_pte_t pte,
1326 unsigned long index,
1327 unsigned long shift,
1330 unsigned long slot, hash, hidx;
1332 hash = hpt_hash(vpn, shift, ssize);
1333 hidx = __rpte_to_hidx(pte, index);
1334 if (hidx & _PTEIDX_SECONDARY)
1336 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1337 slot += hidx & _PTEIDX_GROUP_IX;
1342 * The hcall H_BLOCK_REMOVE implies that the virtual pages to processed are
1343 * "all within the same naturally aligned 8 page virtual address block".
1345 static void do_block_remove(unsigned long number, struct ppc64_tlb_batch *batch,
1346 unsigned long *param)
1349 unsigned long i, pix = 0;
1350 unsigned long index, shift, slot, current_vpgb, vpgb;
1354 psize = batch->psize;
1355 ssize = batch->ssize;
1357 for (i = 0; i < number; i++) {
1358 vpn = batch->vpn[i];
1359 pte = batch->pte[i];
1360 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1362 * Shifting 3 bits more on the right to get a
1363 * 8 pages aligned virtual addresse.
1365 vpgb = (vpn >> (shift - VPN_SHIFT + 3));
1366 if (!pix || vpgb != current_vpgb) {
1368 * Need to start a new 8 pages block, flush
1369 * the current one if needed.
1372 (void)call_block_remove(pix, param,
1374 current_vpgb = vpgb;
1375 param[0] = hpte_encode_avpn(vpn, psize,
1380 slot = compute_slot(pte, vpn, index, shift, ssize);
1381 param[pix++] = HBR_REQUEST | HBLKR_AVPN | slot;
1383 if (pix == PLPAR_HCALL9_BUFSIZE) {
1384 pix = call_block_remove(pix, param, false);
1386 * pix = 0 means that all the entries were
1387 * removed, we can start a new block.
1388 * Otherwise, this means that there are entries
1389 * to retry, and pix points to latest one, so
1390 * we should increment it and try to continue
1396 } pte_iterate_hashed_end();
1400 (void)call_block_remove(pix, param, true);
1404 * TLB Block Invalidate Characteristics
1406 * These characteristics define the size of the block the hcall H_BLOCK_REMOVE
1407 * is able to process for each couple segment base page size, actual page size.
1409 * The ibm,get-system-parameter properties is returning a buffer with the
1412 * [ 2 bytes size of the RTAS buffer (excluding these 2 bytes) ]
1414 * TLB Block Invalidate Specifiers:
1415 * [ 1 byte LOG base 2 of the TLB invalidate block size being specified ]
1416 * [ 1 byte Number of page sizes (N) that are supported for the specified
1417 * TLB invalidate block size ]
1418 * [ 1 byte Encoded segment base page size and actual page size
1419 * MSB=0 means 4k segment base page size and actual page size
1420 * MSB=1 the penc value in mmu_psize_def ]
1423 * Next TLB Block Invalidate Specifiers...
1427 static inline void set_hblkrm_bloc_size(int bpsize, int psize,
1428 unsigned int block_size)
1430 if (block_size > hblkrm_size[bpsize][psize])
1431 hblkrm_size[bpsize][psize] = block_size;
1435 * Decode the Encoded segment base page size and actual page size.
1437 * - bit 7 is the L bit
1438 * - bits 0-5 are the penc value
1439 * If the L bit is 0, this means 4K segment base page size and actual page size
1440 * otherwise the penc value should be read.
1442 #define HBLKRM_L_MASK 0x80
1443 #define HBLKRM_PENC_MASK 0x3f
1444 static inline void __init check_lp_set_hblkrm(unsigned int lp,
1445 unsigned int block_size)
1447 unsigned int bpsize, psize;
1449 /* First, check the L bit, if not set, this means 4K */
1450 if ((lp & HBLKRM_L_MASK) == 0) {
1451 set_hblkrm_bloc_size(MMU_PAGE_4K, MMU_PAGE_4K, block_size);
1455 lp &= HBLKRM_PENC_MASK;
1456 for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++) {
1457 struct mmu_psize_def *def = &mmu_psize_defs[bpsize];
1459 for (psize = 0; psize < MMU_PAGE_COUNT; psize++) {
1460 if (def->penc[psize] == lp) {
1461 set_hblkrm_bloc_size(bpsize, psize, block_size);
1469 * The size of the TLB Block Invalidate Characteristics is variable. But at the
1470 * maximum it will be the number of possible page sizes *2 + 10 bytes.
1471 * Currently MMU_PAGE_COUNT is 16, which means 42 bytes. Use a cache line size
1472 * (128 bytes) for the buffer to get plenty of space.
1474 #define SPLPAR_TLB_BIC_MAXLENGTH 128
1476 void __init pseries_lpar_read_hblkrm_characteristics(void)
1478 static struct papr_sysparm_buf buf __initdata;
1479 int len, idx, bpsize;
1481 if (!firmware_has_feature(FW_FEATURE_BLOCK_REMOVE))
1484 if (papr_sysparm_get(PAPR_SYSPARM_TLB_BLOCK_INVALIDATE_ATTRS, &buf))
1487 len = be16_to_cpu(buf.len);
1488 if (len > SPLPAR_TLB_BIC_MAXLENGTH) {
1489 pr_warn("%s too large returned buffer %d", __func__, len);
1495 u8 block_shift = buf.val[idx++];
1497 unsigned int npsize;
1502 block_size = 1 << block_shift;
1504 for (npsize = buf.val[idx++];
1505 npsize > 0 && idx < len; npsize--)
1506 check_lp_set_hblkrm((unsigned int)buf.val[idx++],
1510 for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
1511 for (idx = 0; idx < MMU_PAGE_COUNT; idx++)
1512 if (hblkrm_size[bpsize][idx])
1513 pr_info("H_BLOCK_REMOVE supports base psize:%d psize:%d block size:%d",
1514 bpsize, idx, hblkrm_size[bpsize][idx]);
1518 * Take a spinlock around flushes to avoid bouncing the hypervisor tlbie
1521 static void pSeries_lpar_flush_hash_range(unsigned long number, int local)
1524 unsigned long i, pix, rc;
1525 unsigned long flags = 0;
1526 struct ppc64_tlb_batch *batch = this_cpu_ptr(&ppc64_tlb_batch);
1527 int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
1528 unsigned long param[PLPAR_HCALL9_BUFSIZE];
1529 unsigned long index, shift, slot;
1534 spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
1536 if (is_supported_hlbkrm(batch->psize, batch->psize)) {
1537 do_block_remove(number, batch, param);
1541 psize = batch->psize;
1542 ssize = batch->ssize;
1544 for (i = 0; i < number; i++) {
1545 vpn = batch->vpn[i];
1546 pte = batch->pte[i];
1547 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1548 slot = compute_slot(pte, vpn, index, shift, ssize);
1549 if (!firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1551 * lpar doesn't use the passed actual page size
1553 pSeries_lpar_hpte_invalidate(slot, vpn, psize,
1556 param[pix] = HBR_REQUEST | HBR_AVPN | slot;
1557 param[pix+1] = hpte_encode_avpn(vpn, psize,
1561 rc = plpar_hcall9(H_BULK_REMOVE, param,
1562 param[0], param[1], param[2],
1563 param[3], param[4], param[5],
1564 param[6], param[7]);
1565 BUG_ON(rc != H_SUCCESS);
1569 } pte_iterate_hashed_end();
1572 param[pix] = HBR_END;
1573 rc = plpar_hcall9(H_BULK_REMOVE, param, param[0], param[1],
1574 param[2], param[3], param[4], param[5],
1575 param[6], param[7]);
1576 BUG_ON(rc != H_SUCCESS);
1581 spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
1584 static int __init disable_bulk_remove(char *str)
1586 if (strcmp(str, "off") == 0 &&
1587 firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1588 pr_info("Disabling BULK_REMOVE firmware feature");
1589 powerpc_firmware_features &= ~FW_FEATURE_BULK_REMOVE;
1594 __setup("bulk_remove=", disable_bulk_remove);
1596 #define HPT_RESIZE_TIMEOUT 10000 /* ms */
1598 struct hpt_resize_state {
1599 unsigned long shift;
1603 static int pseries_lpar_resize_hpt_commit(void *data)
1605 struct hpt_resize_state *state = data;
1607 state->commit_rc = plpar_resize_hpt_commit(0, state->shift);
1608 if (state->commit_rc != H_SUCCESS)
1611 /* Hypervisor has transitioned the HTAB, update our globals */
1612 ppc64_pft_size = state->shift;
1613 htab_size_bytes = 1UL << ppc64_pft_size;
1614 htab_hash_mask = (htab_size_bytes >> 7) - 1;
1620 * Must be called in process context. The caller must hold the
1623 static int pseries_lpar_resize_hpt(unsigned long shift)
1625 struct hpt_resize_state state = {
1627 .commit_rc = H_FUNCTION,
1629 unsigned int delay, total_delay = 0;
1635 if (!firmware_has_feature(FW_FEATURE_HPT_RESIZE))
1638 pr_info("Attempting to resize HPT to shift %lu\n", shift);
1642 rc = plpar_resize_hpt_prepare(0, shift);
1643 while (H_IS_LONG_BUSY(rc)) {
1644 delay = get_longbusy_msecs(rc);
1645 total_delay += delay;
1646 if (total_delay > HPT_RESIZE_TIMEOUT) {
1647 /* prepare with shift==0 cancels an in-progress resize */
1648 rc = plpar_resize_hpt_prepare(0, 0);
1649 if (rc != H_SUCCESS)
1650 pr_warn("Unexpected error %d cancelling timed out HPT resize\n",
1655 rc = plpar_resize_hpt_prepare(0, shift);
1664 pr_warn("Invalid argument from H_RESIZE_HPT_PREPARE\n");
1667 pr_warn("Operation not permitted from H_RESIZE_HPT_PREPARE\n");
1670 pr_warn("Unexpected error %d from H_RESIZE_HPT_PREPARE\n", rc);
1676 rc = stop_machine_cpuslocked(pseries_lpar_resize_hpt_commit,
1682 switch (state.commit_rc) {
1687 pr_warn("Unexpected error %d from H_RESIZE_HPT_COMMIT\n",
1693 pr_info("HPT resize to shift %lu complete (%lld ms / %lld ms)\n",
1694 shift, (long long) ktime_ms_delta(t1, t0),
1695 (long long) ktime_ms_delta(t2, t1));
1700 void __init hpte_init_pseries(void)
1702 mmu_hash_ops.hpte_invalidate = pSeries_lpar_hpte_invalidate;
1703 mmu_hash_ops.hpte_updatepp = pSeries_lpar_hpte_updatepp;
1704 mmu_hash_ops.hpte_updateboltedpp = pSeries_lpar_hpte_updateboltedpp;
1705 mmu_hash_ops.hpte_insert = pSeries_lpar_hpte_insert;
1706 mmu_hash_ops.hpte_remove = pSeries_lpar_hpte_remove;
1707 mmu_hash_ops.hpte_removebolted = pSeries_lpar_hpte_removebolted;
1708 mmu_hash_ops.flush_hash_range = pSeries_lpar_flush_hash_range;
1709 mmu_hash_ops.hpte_clear_all = pseries_hpte_clear_all;
1710 mmu_hash_ops.hugepage_invalidate = pSeries_lpar_hugepage_invalidate;
1712 if (firmware_has_feature(FW_FEATURE_HPT_RESIZE))
1713 mmu_hash_ops.resize_hpt = pseries_lpar_resize_hpt;
1716 * On POWER9, we need to do a H_REGISTER_PROC_TBL hcall
1717 * to inform the hypervisor that we wish to use the HPT.
1719 if (cpu_has_feature(CPU_FTR_ARCH_300))
1720 pseries_lpar_register_process_table(0, 0, 0);
1722 #endif /* CONFIG_PPC_64S_HASH_MMU */
1724 #ifdef CONFIG_PPC_RADIX_MMU
1725 void __init radix_init_pseries(void)
1727 pr_info("Using radix MMU under hypervisor\n");
1729 pseries_lpar_register_process_table(__pa(process_tb),
1730 0, PRTB_SIZE_SHIFT - 12);
1734 #ifdef CONFIG_PPC_SMLPAR
1735 #define CMO_FREE_HINT_DEFAULT 1
1736 static int cmo_free_hint_flag = CMO_FREE_HINT_DEFAULT;
1738 static int __init cmo_free_hint(char *str)
1741 parm = strstrip(str);
1743 if (strcasecmp(parm, "no") == 0 || strcasecmp(parm, "off") == 0) {
1744 pr_info("%s: CMO free page hinting is not active.\n", __func__);
1745 cmo_free_hint_flag = 0;
1749 cmo_free_hint_flag = 1;
1750 pr_info("%s: CMO free page hinting is active.\n", __func__);
1752 if (strcasecmp(parm, "yes") == 0 || strcasecmp(parm, "on") == 0)
1758 __setup("cmo_free_hint=", cmo_free_hint);
1760 static void pSeries_set_page_state(struct page *page, int order,
1761 unsigned long state)
1764 unsigned long cmo_page_sz, addr;
1766 cmo_page_sz = cmo_get_page_size();
1767 addr = __pa((unsigned long)page_address(page));
1769 for (i = 0; i < (1 << order); i++, addr += PAGE_SIZE) {
1770 for (j = 0; j < PAGE_SIZE; j += cmo_page_sz)
1771 plpar_hcall_norets(H_PAGE_INIT, state, addr + j, 0);
1775 void arch_free_page(struct page *page, int order)
1777 if (radix_enabled())
1779 if (!cmo_free_hint_flag || !firmware_has_feature(FW_FEATURE_CMO))
1782 pSeries_set_page_state(page, order, H_PAGE_SET_UNUSED);
1784 EXPORT_SYMBOL(arch_free_page);
1786 #endif /* CONFIG_PPC_SMLPAR */
1787 #endif /* CONFIG_PPC_BOOK3S_64 */
1789 #ifdef CONFIG_TRACEPOINTS
1790 #ifdef CONFIG_JUMP_LABEL
1791 struct static_key hcall_tracepoint_key = STATIC_KEY_INIT;
1793 int hcall_tracepoint_regfunc(void)
1795 static_key_slow_inc(&hcall_tracepoint_key);
1799 void hcall_tracepoint_unregfunc(void)
1801 static_key_slow_dec(&hcall_tracepoint_key);
1805 * We optimise our hcall path by placing hcall_tracepoint_refcount
1806 * directly in the TOC so we can check if the hcall tracepoints are
1807 * enabled via a single load.
1810 /* NB: reg/unreg are called while guarded with the tracepoints_mutex */
1811 extern long hcall_tracepoint_refcount;
1813 int hcall_tracepoint_regfunc(void)
1815 hcall_tracepoint_refcount++;
1819 void hcall_tracepoint_unregfunc(void)
1821 hcall_tracepoint_refcount--;
1826 * Keep track of hcall tracing depth and prevent recursion. Warn if any is
1827 * detected because it may indicate a problem. This will not catch all
1828 * problems with tracing code making hcalls, because the tracing might have
1829 * been invoked from a non-hcall, so the first hcall could recurse into it
1830 * without warning here, but this better than nothing.
1832 * Hcalls with specific problems being traced should use the _notrace
1833 * plpar_hcall variants.
1835 static DEFINE_PER_CPU(unsigned int, hcall_trace_depth);
1838 notrace void __trace_hcall_entry(unsigned long opcode, unsigned long *args)
1840 unsigned long flags;
1841 unsigned int *depth;
1843 local_irq_save(flags);
1845 depth = this_cpu_ptr(&hcall_trace_depth);
1847 if (WARN_ON_ONCE(*depth))
1852 trace_hcall_entry(opcode, args);
1856 local_irq_restore(flags);
1859 notrace void __trace_hcall_exit(long opcode, long retval, unsigned long *retbuf)
1861 unsigned long flags;
1862 unsigned int *depth;
1864 local_irq_save(flags);
1866 depth = this_cpu_ptr(&hcall_trace_depth);
1868 if (*depth) /* Don't warn again on the way out */
1872 trace_hcall_exit(opcode, retval, retbuf);
1877 local_irq_restore(flags);
1883 * H_GET_MPP hcall returns info in 7 parms
1885 int h_get_mpp(struct hvcall_mpp_data *mpp_data)
1888 unsigned long retbuf[PLPAR_HCALL9_BUFSIZE];
1890 rc = plpar_hcall9(H_GET_MPP, retbuf);
1892 mpp_data->entitled_mem = retbuf[0];
1893 mpp_data->mapped_mem = retbuf[1];
1895 mpp_data->group_num = (retbuf[2] >> 2 * 8) & 0xffff;
1896 mpp_data->pool_num = retbuf[2] & 0xffff;
1898 mpp_data->mem_weight = (retbuf[3] >> 7 * 8) & 0xff;
1899 mpp_data->unallocated_mem_weight = (retbuf[3] >> 6 * 8) & 0xff;
1900 mpp_data->unallocated_entitlement = retbuf[3] & 0xffffffffffffUL;
1902 mpp_data->pool_size = retbuf[4];
1903 mpp_data->loan_request = retbuf[5];
1904 mpp_data->backing_mem = retbuf[6];
1908 EXPORT_SYMBOL(h_get_mpp);
1910 int h_get_mpp_x(struct hvcall_mpp_x_data *mpp_x_data)
1913 unsigned long retbuf[PLPAR_HCALL9_BUFSIZE] = { 0 };
1915 rc = plpar_hcall9(H_GET_MPP_X, retbuf);
1917 mpp_x_data->coalesced_bytes = retbuf[0];
1918 mpp_x_data->pool_coalesced_bytes = retbuf[1];
1919 mpp_x_data->pool_purr_cycles = retbuf[2];
1920 mpp_x_data->pool_spurr_cycles = retbuf[3];
1925 #ifdef CONFIG_PPC_64S_HASH_MMU
1926 static unsigned long __init vsid_unscramble(unsigned long vsid, int ssize)
1928 unsigned long protovsid;
1929 unsigned long va_bits = VA_BITS;
1930 unsigned long modinv, vsid_modulus;
1931 unsigned long max_mod_inv, tmp_modinv;
1933 if (!mmu_has_feature(MMU_FTR_68_BIT_VA))
1936 if (ssize == MMU_SEGSIZE_256M) {
1937 modinv = VSID_MULINV_256M;
1938 vsid_modulus = ((1UL << (va_bits - SID_SHIFT)) - 1);
1940 modinv = VSID_MULINV_1T;
1941 vsid_modulus = ((1UL << (va_bits - SID_SHIFT_1T)) - 1);
1945 * vsid outside our range.
1947 if (vsid >= vsid_modulus)
1951 * If modinv is the modular multiplicate inverse of (x % vsid_modulus)
1952 * and vsid = (protovsid * x) % vsid_modulus, then we say:
1953 * protovsid = (vsid * modinv) % vsid_modulus
1956 /* Check if (vsid * modinv) overflow (63 bits) */
1957 max_mod_inv = 0x7fffffffffffffffull / vsid;
1958 if (modinv < max_mod_inv)
1959 return (vsid * modinv) % vsid_modulus;
1961 tmp_modinv = modinv/max_mod_inv;
1962 modinv %= max_mod_inv;
1964 protovsid = (((vsid * max_mod_inv) % vsid_modulus) * tmp_modinv) % vsid_modulus;
1965 protovsid = (protovsid + vsid * modinv) % vsid_modulus;
1970 static int __init reserve_vrma_context_id(void)
1972 unsigned long protovsid;
1975 * Reserve context ids which map to reserved virtual addresses. For now
1976 * we only reserve the context id which maps to the VRMA VSID. We ignore
1977 * the addresses in "ibm,adjunct-virtual-addresses" because we don't
1978 * enable adjunct support via the "ibm,client-architecture-support"
1981 protovsid = vsid_unscramble(VRMA_VSID, MMU_SEGSIZE_1T);
1982 hash__reserve_context_id(protovsid >> ESID_BITS_1T);
1985 machine_device_initcall(pseries, reserve_vrma_context_id);
1988 #ifdef CONFIG_DEBUG_FS
1989 /* debugfs file interface for vpa data */
1990 static ssize_t vpa_file_read(struct file *filp, char __user *buf, size_t len,
1993 int cpu = (long)filp->private_data;
1994 struct lppaca *lppaca = &lppaca_of(cpu);
1996 return simple_read_from_buffer(buf, len, pos, lppaca,
1997 sizeof(struct lppaca));
2000 static const struct file_operations vpa_fops = {
2001 .open = simple_open,
2002 .read = vpa_file_read,
2003 .llseek = default_llseek,
2006 static int __init vpa_debugfs_init(void)
2010 struct dentry *vpa_dir;
2012 if (!firmware_has_feature(FW_FEATURE_SPLPAR))
2015 vpa_dir = debugfs_create_dir("vpa", arch_debugfs_dir);
2017 /* set up the per-cpu vpa file*/
2018 for_each_possible_cpu(i) {
2019 sprintf(name, "cpu-%ld", i);
2020 debugfs_create_file(name, 0400, vpa_dir, (void *)i, &vpa_fops);
2025 machine_arch_initcall(pseries, vpa_debugfs_init);
2026 #endif /* CONFIG_DEBUG_FS */