arch/x86/xen/setup.c

   1 /*
   2  * Machine specific setup for xen
   3  *
   4  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
   5  */
   6
   7 #include <linux/module.h>
   8 #include <linux/sched.h>
   9 #include <linux/mm.h>
  10 #include <linux/pm.h>
  11
  12 #include <asm/elf.h>
  13 #include <asm/vdso.h>
  14 #include <asm/e820.h>
  15 #include <asm/setup.h>
  16 #include <asm/acpi.h>
  17 #include <asm/xen/hypervisor.h>
  18 #include <asm/xen/hypercall.h>
  19
  20 #include <xen/xen.h>
  21 #include <xen/page.h>
  22 #include <xen/interface/callback.h>
  23 #include <xen/interface/memory.h>
  24 #include <xen/interface/physdev.h>
  25 #include <xen/interface/memory.h>
  26 #include <xen/features.h>
  27
  28 #include "xen-ops.h"
  29 #include "vdso.h"
  30
  31 /* These are code, but not functions.  Defined in entry.S */
  32 extern const char xen_hypervisor_callback[];
  33 extern const char xen_failsafe_callback[];
  34 extern void xen_sysenter_target(void);
  35 extern void xen_syscall_target(void);
  36 extern void xen_syscall32_target(void);
  37
  38 /* Amount of extra memory space we add to the e820 ranges */
  39 phys_addr_t xen_extra_mem_start, xen_extra_mem_size;
  40
  41 /*
  42  * The maximum amount of extra memory compared to the base size.  The
  43  * main scaling factor is the size of struct page.  At extreme ratios
  44  * of base:extra, all the base memory can be filled with page
  45  * structures for the extra memory, leaving no space for anything
  46  * else.
  47  *
  48  * 10x seems like a reasonable balance between scaling flexibility and
  49  * leaving a practically usable system.
  50  */
  51 #define EXTRA_MEM_RATIO         (10)
  52
  53 static __init void xen_add_extra_mem(unsigned long pages)
  54 {
  55         u64 size = (u64)pages * PAGE_SIZE;
  56         u64 extra_start = xen_extra_mem_start + xen_extra_mem_size;
  57
  58         if (!pages)
  59                 return;
  60
  61         e820_add_region(extra_start, size, E820_RAM);
  62         sanitize_e820_map(e820.map, ARRAY_SIZE(e820.map), &e820.nr_map);
  63
  64         reserve_early(extra_start, extra_start + size, "XEN EXTRA");
  65
  66         xen_extra_mem_size += size;
  67
  68         xen_max_p2m_pfn = PFN_DOWN(extra_start + size);
  69 }
  70
  71 static unsigned long __init xen_release_chunk(phys_addr_t start_addr,
  72                                               phys_addr_t end_addr)
  73 {
  74         struct xen_memory_reservation reservation = {
  75                 .address_bits = 0,
  76                 .extent_order = 0,
  77                 .domid        = DOMID_SELF
  78         };
  79         unsigned long start, end;
  80         unsigned long len = 0;
  81         unsigned long pfn;
  82         int ret;
  83
  84         start = PFN_UP(start_addr);
  85         end = PFN_DOWN(end_addr);
  86
  87         if (end <= start)
  88                 return 0;
  89
  90         printk(KERN_INFO "xen_release_chunk: looking at area pfn %lx-%lx: ",
  91                start, end);
  92         for(pfn = start; pfn < end; pfn++) {
  93                 unsigned long mfn = pfn_to_mfn(pfn);
  94
  95                 /* Make sure pfn exists to start with */
  96                 if (mfn == INVALID_P2M_ENTRY || mfn_to_pfn(mfn) != pfn)
  97                         continue;
  98
  99                 set_xen_guest_handle(reservation.extent_start, &mfn);
 100                 reservation.nr_extents = 1;
 101
 102                 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation,
 103                                            &reservation);
 104                 WARN(ret != 1, "Failed to release memory %lx-%lx err=%d\n",
 105                      start, end, ret);
 106                 if (ret == 1) {
 107                         set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 108                         len++;
 109                 }
 110         }
 111         printk(KERN_CONT "%ld pages freed\n", len);
 112
 113         return len;
 114 }
 115
 116 static unsigned long __init xen_return_unused_memory(unsigned long max_pfn,
 117                                                      const struct e820map *e820)
 118 {
 119         phys_addr_t max_addr = PFN_PHYS(max_pfn);
 120         phys_addr_t last_end = 0;
 121         unsigned long released = 0;
 122         int i;
 123
 124         for (i = 0; i < e820->nr_map && last_end < max_addr; i++) {
 125                 phys_addr_t end = e820->map[i].addr;
 126                 end = min(max_addr, end);
 127
 128                 released += xen_release_chunk(last_end, end);
 129                 last_end = e820->map[i].addr + e820->map[i].size;
 130         }
 131
 132         if (last_end < max_addr)
 133                 released += xen_release_chunk(last_end, max_addr);
 134
 135         printk(KERN_INFO "released %ld pages of unused memory\n", released);
 136         return released;
 137 }
 138
 139 /**
 140  * machine_specific_memory_setup - Hook for machine specific memory setup.
 141  **/
 142 char * __init xen_memory_setup(void)
 143 {
 144         static struct e820entry map[E820MAX] __initdata;
 145
 146         unsigned long max_pfn = xen_start_info->nr_pages;
 147         unsigned long long mem_end;
 148         int rc;
 149         struct xen_memory_map memmap;
 150         unsigned long extra_pages = 0;
 151         unsigned long extra_limit;
 152         int i;
 153         int op;
 154
 155         max_pfn = min(MAX_DOMAIN_PAGES, max_pfn);
 156         mem_end = PFN_PHYS(max_pfn);
 157
 158         memmap.nr_entries = E820MAX;
 159         set_xen_guest_handle(memmap.buffer, map);
 160
 161         op = xen_initial_domain() ?
 162                 XENMEM_machine_memory_map :
 163                 XENMEM_memory_map;
 164         rc = HYPERVISOR_memory_op(op, &memmap);
 165         if (rc == -ENOSYS) {
 166                 memmap.nr_entries = 1;
 167                 map[0].addr = 0ULL;
 168                 map[0].size = mem_end;
 169                 /* 8MB slack (to balance backend allocations). */
 170                 map[0].size += 8ULL << 20;
 171                 map[0].type = E820_RAM;
 172                 rc = 0;
 173         }
 174         BUG_ON(rc);
 175
 176         e820.nr_map = 0;
 177         xen_extra_mem_start = mem_end;
 178         for (i = 0; i < memmap.nr_entries; i++) {
 179                 unsigned long long end = map[i].addr + map[i].size;
 180
 181                 if (map[i].type == E820_RAM) {
 182                         if (map[i].addr < mem_end && end > mem_end) {
 183                                 /* Truncate region to max_mem. */
 184                                 u64 delta = end - mem_end;
 185
 186                                 map[i].size -= delta;
 187                                 extra_pages += PFN_DOWN(delta);
 188
 189                                 end = mem_end;
 190                         }
 191                 }
 192
 193                 if (end > xen_extra_mem_start)
 194                         xen_extra_mem_start = end;
 195
 196                 /* If region is non-RAM or below mem_end, add what remains */
 197                 if ((map[i].type != E820_RAM || map[i].addr < mem_end) &&
 198                     map[i].size > 0)
 199                         e820_add_region(map[i].addr, map[i].size, map[i].type);
 200         }
 201
 202         /*
 203          * Even though this is normal, usable memory under Xen, reserve
 204          * ISA memory anyway because too many things think they can poke
 205          * about in there.
 206          */
 207         e820_add_region(ISA_START_ADDRESS, ISA_END_ADDRESS - ISA_START_ADDRESS,
 208                         E820_RESERVED);
 209
 210         /*
 211          * Reserve Xen bits:
 212          *  - mfn_list
 213          *  - xen_start_info
 214          * See comment above "struct start_info" in <xen/interface/xen.h>
 215          */
 216         reserve_early(__pa(xen_start_info->mfn_list),
 217                       __pa(xen_start_info->pt_base),
 218                         "XEN START INFO");
 219
 220         sanitize_e820_map(e820.map, ARRAY_SIZE(e820.map), &e820.nr_map);
 221
 222         extra_pages += xen_return_unused_memory(xen_start_info->nr_pages, &e820);
 223
 224         /*
 225          * Clamp the amount of extra memory to a EXTRA_MEM_RATIO
 226          * factor the base size.  On non-highmem systems, the base
 227          * size is the full initial memory allocation; on highmem it
 228          * is limited to the max size of lowmem, so that it doesn't
 229          * get completely filled.
 230          *
 231          * In principle there could be a problem in lowmem systems if
 232          * the initial memory is also very large with respect to
 233          * lowmem, but we won't try to deal with that here.
 234          */
 235         extra_limit = min(EXTRA_MEM_RATIO * min(max_pfn, PFN_DOWN(MAXMEM)),
 236                           max_pfn + extra_pages);
 237
 238         if (extra_limit >= max_pfn)
 239                 extra_pages = extra_limit - max_pfn;
 240         else
 241                 extra_pages = 0;
 242
 243         if (!xen_initial_domain())
 244                 xen_add_extra_mem(extra_pages);
 245
 246         return "Xen";
 247 }
 248
 249 static void xen_idle(void)
 250 {
 251         local_irq_disable();
 252
 253         if (need_resched())
 254                 local_irq_enable();
 255         else {
 256                 current_thread_info()->status &= ~TS_POLLING;
 257                 smp_mb__after_clear_bit();
 258                 safe_halt();
 259                 current_thread_info()->status |= TS_POLLING;
 260         }
 261 }
 262
 263 /*
 264  * Set the bit indicating "nosegneg" library variants should be used.
 265  * We only need to bother in pure 32-bit mode; compat 32-bit processes
 266  * can have un-truncated segments, so wrapping around is allowed.
 267  */
 268 static void __init fiddle_vdso(void)
 269 {
 270 #ifdef CONFIG_X86_32
 271         u32 *mask;
 272         mask = VDSO32_SYMBOL(&vdso32_int80_start, NOTE_MASK);
 273         *mask |= 1 << VDSO_NOTE_NONEGSEG_BIT;
 274         mask = VDSO32_SYMBOL(&vdso32_sysenter_start, NOTE_MASK);
 275         *mask |= 1 << VDSO_NOTE_NONEGSEG_BIT;
 276 #endif
 277 }
 278
 279 static __cpuinit int register_callback(unsigned type, const void *func)
 280 {
 281         struct callback_register callback = {
 282                 .type = type,
 283                 .address = XEN_CALLBACK(__KERNEL_CS, func),
 284                 .flags = CALLBACKF_mask_events,
 285         };
 286
 287         return HYPERVISOR_callback_op(CALLBACKOP_register, &callback);
 288 }
 289
 290 void __cpuinit xen_enable_sysenter(void)
 291 {
 292         int ret;
 293         unsigned sysenter_feature;
 294
 295 #ifdef CONFIG_X86_32
 296         sysenter_feature = X86_FEATURE_SEP;
 297 #else
 298         sysenter_feature = X86_FEATURE_SYSENTER32;
 299 #endif
 300
 301         if (!boot_cpu_has(sysenter_feature))
 302                 return;
 303
 304         ret = register_callback(CALLBACKTYPE_sysenter, xen_sysenter_target);
 305         if(ret != 0)
 306                 setup_clear_cpu_cap(sysenter_feature);
 307 }
 308
 309 void __cpuinit xen_enable_syscall(void)
 310 {
 311 #ifdef CONFIG_X86_64
 312         int ret;
 313
 314         ret = register_callback(CALLBACKTYPE_syscall, xen_syscall_target);
 315         if (ret != 0) {
 316                 printk(KERN_ERR "Failed to set syscall callback: %d\n", ret);
 317                 /* Pretty fatal; 64-bit userspace has no other
 318                    mechanism for syscalls. */
 319         }
 320
 321         if (boot_cpu_has(X86_FEATURE_SYSCALL32)) {
 322                 ret = register_callback(CALLBACKTYPE_syscall32,
 323                                         xen_syscall32_target);
 324                 if (ret != 0)
 325                         setup_clear_cpu_cap(X86_FEATURE_SYSCALL32);
 326         }
 327 #endif /* CONFIG_X86_64 */
 328 }
 329
 330 void __init xen_arch_setup(void)
 331 {
 332         struct physdev_set_iopl set_iopl;
 333         int rc;
 334
 335         xen_panic_handler_init();
 336
 337         HYPERVISOR_vm_assist(VMASST_CMD_enable, VMASST_TYPE_4gb_segments);
 338         HYPERVISOR_vm_assist(VMASST_CMD_enable, VMASST_TYPE_writable_pagetables);
 339
 340         if (!xen_feature(XENFEAT_auto_translated_physmap))
 341                 HYPERVISOR_vm_assist(VMASST_CMD_enable,
 342                                      VMASST_TYPE_pae_extended_cr3);
 343
 344         if (register_callback(CALLBACKTYPE_event, xen_hypervisor_callback) ||
 345             register_callback(CALLBACKTYPE_failsafe, xen_failsafe_callback))
 346                 BUG();
 347
 348         xen_enable_sysenter();
 349         xen_enable_syscall();
 350
 351         set_iopl.iopl = 1;
 352         rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 353         if (rc != 0)
 354                 printk(KERN_INFO "physdev_op failed %d\n", rc);
 355
 356 #ifdef CONFIG_ACPI
 357         if (!(xen_start_info->flags & SIF_INITDOMAIN)) {
 358                 printk(KERN_INFO "ACPI in unprivileged domain disabled\n");
 359                 disable_acpi();
 360         }
 361 #endif
 362
 363         memcpy(boot_command_line, xen_start_info->cmd_line,
 364                MAX_GUEST_CMDLINE > COMMAND_LINE_SIZE ?
 365                COMMAND_LINE_SIZE : MAX_GUEST_CMDLINE);
 366
 367         pm_idle = xen_idle;
 368
 369         paravirt_disable_iospace();
 370
 371         fiddle_vdso();
 372 }