1aa04b4ccSPaul Mackerras /* 2aa04b4ccSPaul Mackerras * Copyright 2011 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> 3aa04b4ccSPaul Mackerras * 4aa04b4ccSPaul Mackerras * This program is free software; you can redistribute it and/or modify 5aa04b4ccSPaul Mackerras * it under the terms of the GNU General Public License, version 2, as 6aa04b4ccSPaul Mackerras * published by the Free Software Foundation. 7aa04b4ccSPaul Mackerras */ 8aa04b4ccSPaul Mackerras 9441c19c8SMichael Ellerman #include <linux/cpu.h> 10aa04b4ccSPaul Mackerras #include <linux/kvm_host.h> 11aa04b4ccSPaul Mackerras #include <linux/preempt.h> 1266b15db6SPaul Gortmaker #include <linux/export.h> 13aa04b4ccSPaul Mackerras #include <linux/sched.h> 14aa04b4ccSPaul Mackerras #include <linux/spinlock.h> 15aa04b4ccSPaul Mackerras #include <linux/bootmem.h> 16aa04b4ccSPaul Mackerras #include <linux/init.h> 17fa61a4e3SAneesh Kumar K.V #include <linux/memblock.h> 18fa61a4e3SAneesh Kumar K.V #include <linux/sizes.h> 19aa04b4ccSPaul Mackerras 20aa04b4ccSPaul Mackerras #include <asm/cputable.h> 21aa04b4ccSPaul Mackerras #include <asm/kvm_ppc.h> 22aa04b4ccSPaul Mackerras #include <asm/kvm_book3s.h> 23aa04b4ccSPaul Mackerras 24fa61a4e3SAneesh Kumar K.V #include "book3s_hv_cma.h" 25fa61a4e3SAneesh Kumar K.V /* 26fa61a4e3SAneesh Kumar K.V * Hash page table alignment on newer cpus(CPU_FTR_ARCH_206) 27fa61a4e3SAneesh Kumar K.V * should be power of 2. 28fa61a4e3SAneesh Kumar K.V */ 29fa61a4e3SAneesh Kumar K.V #define HPT_ALIGN_PAGES ((1 << 18) >> PAGE_SHIFT) /* 256k */ 30fa61a4e3SAneesh Kumar K.V /* 31fa61a4e3SAneesh Kumar K.V * By default we reserve 5% of memory for hash pagetable allocation. 32fa61a4e3SAneesh Kumar K.V */ 33fa61a4e3SAneesh Kumar K.V static unsigned long kvm_cma_resv_ratio = 5; 34aa04b4ccSPaul Mackerras /* 356c45b810SAneesh Kumar K.V * We allocate RMAs (real mode areas) for KVM guests from the KVM CMA area. 36aa04b4ccSPaul Mackerras * Each RMA has to be physically contiguous and of a size that the 37aa04b4ccSPaul Mackerras * hardware supports. PPC970 and POWER7 support 64MB, 128MB and 256MB, 38aa04b4ccSPaul Mackerras * and other larger sizes. Since we are unlikely to be allocate that 39aa04b4ccSPaul Mackerras * much physically contiguous memory after the system is up and running, 406c45b810SAneesh Kumar K.V * we preallocate a set of RMAs in early boot using CMA. 416c45b810SAneesh Kumar K.V * should be power of 2. 42aa04b4ccSPaul Mackerras */ 436c45b810SAneesh Kumar K.V unsigned long kvm_rma_pages = (1 << 27) >> PAGE_SHIFT; /* 128MB */ 446c45b810SAneesh Kumar K.V EXPORT_SYMBOL_GPL(kvm_rma_pages); 45aa04b4ccSPaul Mackerras 46aa04b4ccSPaul Mackerras /* Work out RMLS (real mode limit selector) field value for a given RMA size. 479e368f29SPaul Mackerras Assumes POWER7 or PPC970. */ 48aa04b4ccSPaul Mackerras static inline int lpcr_rmls(unsigned long rma_size) 49aa04b4ccSPaul Mackerras { 50aa04b4ccSPaul Mackerras switch (rma_size) { 51aa04b4ccSPaul Mackerras case 32ul << 20: /* 32 MB */ 529e368f29SPaul Mackerras if (cpu_has_feature(CPU_FTR_ARCH_206)) 539e368f29SPaul Mackerras return 8; /* only supported on POWER7 */ 549e368f29SPaul Mackerras return -1; 55aa04b4ccSPaul Mackerras case 64ul << 20: /* 64 MB */ 56aa04b4ccSPaul Mackerras return 3; 57aa04b4ccSPaul Mackerras case 128ul << 20: /* 128 MB */ 58aa04b4ccSPaul Mackerras return 7; 59aa04b4ccSPaul Mackerras case 256ul << 20: /* 256 MB */ 60aa04b4ccSPaul Mackerras return 4; 61aa04b4ccSPaul Mackerras case 1ul << 30: /* 1 GB */ 62aa04b4ccSPaul Mackerras return 2; 63aa04b4ccSPaul Mackerras case 16ul << 30: /* 16 GB */ 64aa04b4ccSPaul Mackerras return 1; 65aa04b4ccSPaul Mackerras case 256ul << 30: /* 256 GB */ 66aa04b4ccSPaul Mackerras return 0; 67aa04b4ccSPaul Mackerras default: 68aa04b4ccSPaul Mackerras return -1; 69aa04b4ccSPaul Mackerras } 70aa04b4ccSPaul Mackerras } 71aa04b4ccSPaul Mackerras 72b4e70611SAlexander Graf static int __init early_parse_rma_size(char *p) 73b4e70611SAlexander Graf { 746c45b810SAneesh Kumar K.V unsigned long kvm_rma_size; 756c45b810SAneesh Kumar K.V 766c45b810SAneesh Kumar K.V pr_debug("%s(%s)\n", __func__, p); 77b4e70611SAlexander Graf if (!p) 786c45b810SAneesh Kumar K.V return -EINVAL; 79b4e70611SAlexander Graf kvm_rma_size = memparse(p, &p); 806c45b810SAneesh Kumar K.V /* 816c45b810SAneesh Kumar K.V * Check that the requested size is one supported in hardware 826c45b810SAneesh Kumar K.V */ 836c45b810SAneesh Kumar K.V if (lpcr_rmls(kvm_rma_size) < 0) { 846c45b810SAneesh Kumar K.V pr_err("RMA size of 0x%lx not supported\n", kvm_rma_size); 856c45b810SAneesh Kumar K.V return -EINVAL; 866c45b810SAneesh Kumar K.V } 876c45b810SAneesh Kumar K.V kvm_rma_pages = kvm_rma_size >> PAGE_SHIFT; 88b4e70611SAlexander Graf return 0; 89b4e70611SAlexander Graf } 90b4e70611SAlexander Graf early_param("kvm_rma_size", early_parse_rma_size); 91b4e70611SAlexander Graf 926c45b810SAneesh Kumar K.V struct kvm_rma_info *kvm_alloc_rma() 93b4e70611SAlexander Graf { 946c45b810SAneesh Kumar K.V struct page *page; 956c45b810SAneesh Kumar K.V struct kvm_rma_info *ri; 96b4e70611SAlexander Graf 976c45b810SAneesh Kumar K.V ri = kmalloc(sizeof(struct kvm_rma_info), GFP_KERNEL); 986c45b810SAneesh Kumar K.V if (!ri) 996c45b810SAneesh Kumar K.V return NULL; 1006c45b810SAneesh Kumar K.V page = kvm_alloc_cma(kvm_rma_pages, kvm_rma_pages); 1016c45b810SAneesh Kumar K.V if (!page) 1026c45b810SAneesh Kumar K.V goto err_out; 1036c45b810SAneesh Kumar K.V atomic_set(&ri->use_count, 1); 1046c45b810SAneesh Kumar K.V ri->base_pfn = page_to_pfn(page); 1056c45b810SAneesh Kumar K.V return ri; 1066c45b810SAneesh Kumar K.V err_out: 1076c45b810SAneesh Kumar K.V kfree(ri); 1086c45b810SAneesh Kumar K.V return NULL; 109b4e70611SAlexander Graf } 110b4e70611SAlexander Graf EXPORT_SYMBOL_GPL(kvm_alloc_rma); 111b4e70611SAlexander Graf 1126c45b810SAneesh Kumar K.V void kvm_release_rma(struct kvm_rma_info *ri) 113b4e70611SAlexander Graf { 1146c45b810SAneesh Kumar K.V if (atomic_dec_and_test(&ri->use_count)) { 1156c45b810SAneesh Kumar K.V kvm_release_cma(pfn_to_page(ri->base_pfn), kvm_rma_pages); 1166c45b810SAneesh Kumar K.V kfree(ri); 1176c45b810SAneesh Kumar K.V } 118b4e70611SAlexander Graf } 119b4e70611SAlexander Graf EXPORT_SYMBOL_GPL(kvm_release_rma); 120b4e70611SAlexander Graf 121fa61a4e3SAneesh Kumar K.V static int __init early_parse_kvm_cma_resv(char *p) 122d2a1b483SAlexander Graf { 123fa61a4e3SAneesh Kumar K.V pr_debug("%s(%s)\n", __func__, p); 124d2a1b483SAlexander Graf if (!p) 125fa61a4e3SAneesh Kumar K.V return -EINVAL; 126fa61a4e3SAneesh Kumar K.V return kstrtoul(p, 0, &kvm_cma_resv_ratio); 127d2a1b483SAlexander Graf } 128fa61a4e3SAneesh Kumar K.V early_param("kvm_cma_resv_ratio", early_parse_kvm_cma_resv); 129d2a1b483SAlexander Graf 130fa61a4e3SAneesh Kumar K.V struct page *kvm_alloc_hpt(unsigned long nr_pages) 131d2a1b483SAlexander Graf { 132fa61a4e3SAneesh Kumar K.V unsigned long align_pages = HPT_ALIGN_PAGES; 133fa61a4e3SAneesh Kumar K.V 134fa61a4e3SAneesh Kumar K.V /* Old CPUs require HPT aligned on a multiple of its size */ 135fa61a4e3SAneesh Kumar K.V if (!cpu_has_feature(CPU_FTR_ARCH_206)) 136fa61a4e3SAneesh Kumar K.V align_pages = nr_pages; 137fa61a4e3SAneesh Kumar K.V return kvm_alloc_cma(nr_pages, align_pages); 138d2a1b483SAlexander Graf } 139d2a1b483SAlexander Graf EXPORT_SYMBOL_GPL(kvm_alloc_hpt); 140d2a1b483SAlexander Graf 141fa61a4e3SAneesh Kumar K.V void kvm_release_hpt(struct page *page, unsigned long nr_pages) 142d2a1b483SAlexander Graf { 143fa61a4e3SAneesh Kumar K.V kvm_release_cma(page, nr_pages); 144d2a1b483SAlexander Graf } 145d2a1b483SAlexander Graf EXPORT_SYMBOL_GPL(kvm_release_hpt); 146d2a1b483SAlexander Graf 147fa61a4e3SAneesh Kumar K.V /** 148fa61a4e3SAneesh Kumar K.V * kvm_cma_reserve() - reserve area for kvm hash pagetable 149fa61a4e3SAneesh Kumar K.V * 150fa61a4e3SAneesh Kumar K.V * This function reserves memory from early allocator. It should be 151fa61a4e3SAneesh Kumar K.V * called by arch specific code once the early allocator (memblock or bootmem) 152fa61a4e3SAneesh Kumar K.V * has been activated and all other subsystems have already allocated/reserved 153fa61a4e3SAneesh Kumar K.V * memory. 154fa61a4e3SAneesh Kumar K.V */ 155fa61a4e3SAneesh Kumar K.V void __init kvm_cma_reserve(void) 156fa61a4e3SAneesh Kumar K.V { 157fa61a4e3SAneesh Kumar K.V unsigned long align_size; 158fa61a4e3SAneesh Kumar K.V struct memblock_region *reg; 159fa61a4e3SAneesh Kumar K.V phys_addr_t selected_size = 0; 160fa61a4e3SAneesh Kumar K.V /* 161fa61a4e3SAneesh Kumar K.V * We cannot use memblock_phys_mem_size() here, because 162fa61a4e3SAneesh Kumar K.V * memblock_analyze() has not been called yet. 163fa61a4e3SAneesh Kumar K.V */ 164fa61a4e3SAneesh Kumar K.V for_each_memblock(memory, reg) 165fa61a4e3SAneesh Kumar K.V selected_size += memblock_region_memory_end_pfn(reg) - 166fa61a4e3SAneesh Kumar K.V memblock_region_memory_base_pfn(reg); 167fa61a4e3SAneesh Kumar K.V 168fa61a4e3SAneesh Kumar K.V selected_size = (selected_size * kvm_cma_resv_ratio / 100) << PAGE_SHIFT; 169fa61a4e3SAneesh Kumar K.V if (selected_size) { 170fa61a4e3SAneesh Kumar K.V pr_debug("%s: reserving %ld MiB for global area\n", __func__, 171fa61a4e3SAneesh Kumar K.V (unsigned long)selected_size / SZ_1M); 172fa61a4e3SAneesh Kumar K.V /* 173fa61a4e3SAneesh Kumar K.V * Old CPUs require HPT aligned on a multiple of its size. So for them 174fa61a4e3SAneesh Kumar K.V * make the alignment as max size we could request. 175fa61a4e3SAneesh Kumar K.V */ 176fa61a4e3SAneesh Kumar K.V if (!cpu_has_feature(CPU_FTR_ARCH_206)) 177fa61a4e3SAneesh Kumar K.V align_size = __rounddown_pow_of_two(selected_size); 178fa61a4e3SAneesh Kumar K.V else 179fa61a4e3SAneesh Kumar K.V align_size = HPT_ALIGN_PAGES << PAGE_SHIFT; 1806c45b810SAneesh Kumar K.V 1816c45b810SAneesh Kumar K.V align_size = max(kvm_rma_pages << PAGE_SHIFT, align_size); 182fa61a4e3SAneesh Kumar K.V kvm_cma_declare_contiguous(selected_size, align_size); 183fa61a4e3SAneesh Kumar K.V } 184fa61a4e3SAneesh Kumar K.V } 185441c19c8SMichael Ellerman 186441c19c8SMichael Ellerman /* 187441c19c8SMichael Ellerman * When running HV mode KVM we need to block certain operations while KVM VMs 188441c19c8SMichael Ellerman * exist in the system. We use a counter of VMs to track this. 189441c19c8SMichael Ellerman * 190441c19c8SMichael Ellerman * One of the operations we need to block is onlining of secondaries, so we 191441c19c8SMichael Ellerman * protect hv_vm_count with get/put_online_cpus(). 192441c19c8SMichael Ellerman */ 193441c19c8SMichael Ellerman static atomic_t hv_vm_count; 194441c19c8SMichael Ellerman 195441c19c8SMichael Ellerman void kvm_hv_vm_activated(void) 196441c19c8SMichael Ellerman { 197441c19c8SMichael Ellerman get_online_cpus(); 198441c19c8SMichael Ellerman atomic_inc(&hv_vm_count); 199441c19c8SMichael Ellerman put_online_cpus(); 200441c19c8SMichael Ellerman } 201441c19c8SMichael Ellerman EXPORT_SYMBOL_GPL(kvm_hv_vm_activated); 202441c19c8SMichael Ellerman 203441c19c8SMichael Ellerman void kvm_hv_vm_deactivated(void) 204441c19c8SMichael Ellerman { 205441c19c8SMichael Ellerman get_online_cpus(); 206441c19c8SMichael Ellerman atomic_dec(&hv_vm_count); 207441c19c8SMichael Ellerman put_online_cpus(); 208441c19c8SMichael Ellerman } 209441c19c8SMichael Ellerman EXPORT_SYMBOL_GPL(kvm_hv_vm_deactivated); 210441c19c8SMichael Ellerman 211441c19c8SMichael Ellerman bool kvm_hv_mode_active(void) 212441c19c8SMichael Ellerman { 213441c19c8SMichael Ellerman return atomic_read(&hv_vm_count) != 0; 214441c19c8SMichael Ellerman } 215