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