linux/arch/x86/mm/mmio-mod.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 *
   4 * Copyright (C) IBM Corporation, 2005
   5 *               Jeff Muizelaar, 2006, 2007
   6 *               Pekka Paalanen, 2008 <pq@iki.fi>
   7 *
   8 * Derived from the read-mod example from relay-examples by Tom Zanussi.
   9 */
  10
  11#define pr_fmt(fmt) "mmiotrace: " fmt
  12
  13#include <linux/moduleparam.h>
  14#include <linux/debugfs.h>
  15#include <linux/slab.h>
  16#include <linux/uaccess.h>
  17#include <linux/io.h>
  18#include <linux/mmiotrace.h>
  19#include <linux/pgtable.h>
  20#include <asm/e820/api.h> /* for ISA_START_ADDRESS */
  21#include <linux/atomic.h>
  22#include <linux/percpu.h>
  23#include <linux/cpu.h>
  24
  25#include "pf_in.h"
  26
  27struct trap_reason {
  28        unsigned long addr;
  29        unsigned long ip;
  30        enum reason_type type;
  31        int active_traces;
  32};
  33
  34struct remap_trace {
  35        struct list_head list;
  36        struct kmmio_probe probe;
  37        resource_size_t phys;
  38        unsigned long id;
  39};
  40
  41/* Accessed per-cpu. */
  42static DEFINE_PER_CPU(struct trap_reason, pf_reason);
  43static DEFINE_PER_CPU(struct mmiotrace_rw, cpu_trace);
  44
  45static DEFINE_MUTEX(mmiotrace_mutex);
  46static DEFINE_SPINLOCK(trace_lock);
  47static atomic_t mmiotrace_enabled;
  48static LIST_HEAD(trace_list);           /* struct remap_trace */
  49
  50/*
  51 * Locking in this file:
  52 * - mmiotrace_mutex enforces enable/disable_mmiotrace() critical sections.
  53 * - mmiotrace_enabled may be modified only when holding mmiotrace_mutex
  54 *   and trace_lock.
  55 * - Routines depending on is_enabled() must take trace_lock.
  56 * - trace_list users must hold trace_lock.
  57 * - is_enabled() guarantees that mmio_trace_{rw,mapping} are allowed.
  58 * - pre/post callbacks assume the effect of is_enabled() being true.
  59 */
  60
  61/* module parameters */
  62static unsigned long    filter_offset;
  63static bool             nommiotrace;
  64static bool             trace_pc;
  65
  66module_param(filter_offset, ulong, 0);
  67module_param(nommiotrace, bool, 0);
  68module_param(trace_pc, bool, 0);
  69
  70MODULE_PARM_DESC(filter_offset, "Start address of traced mappings.");
  71MODULE_PARM_DESC(nommiotrace, "Disable actual MMIO tracing.");
  72MODULE_PARM_DESC(trace_pc, "Record address of faulting instructions.");
  73
  74static bool is_enabled(void)
  75{
  76        return atomic_read(&mmiotrace_enabled);
  77}
  78
  79static void print_pte(unsigned long address)
  80{
  81        unsigned int level;
  82        pte_t *pte = lookup_address(address, &level);
  83
  84        if (!pte) {
  85                pr_err("Error in %s: no pte for page 0x%08lx\n",
  86                       __func__, address);
  87                return;
  88        }
  89
  90        if (level == PG_LEVEL_2M) {
  91                pr_emerg("4MB pages are not currently supported: 0x%08lx\n",
  92                         address);
  93                BUG();
  94        }
  95        pr_info("pte for 0x%lx: 0x%llx 0x%llx\n",
  96                address,
  97                (unsigned long long)pte_val(*pte),
  98                (unsigned long long)pte_val(*pte) & _PAGE_PRESENT);
  99}
 100
 101/*
 102 * For some reason the pre/post pairs have been called in an
 103 * unmatched order. Report and die.
 104 */
 105static void die_kmmio_nesting_error(struct pt_regs *regs, unsigned long addr)
 106{
 107        const struct trap_reason *my_reason = &get_cpu_var(pf_reason);
 108        pr_emerg("unexpected fault for address: 0x%08lx, last fault for address: 0x%08lx\n",
 109                 addr, my_reason->addr);
 110        print_pte(addr);
 111        pr_emerg("faulting IP is at %pS\n", (void *)regs->ip);
 112        pr_emerg("last faulting IP was at %pS\n", (void *)my_reason->ip);
 113#ifdef __i386__
 114        pr_emerg("eax: %08lx   ebx: %08lx   ecx: %08lx   edx: %08lx\n",
 115                 regs->ax, regs->bx, regs->cx, regs->dx);
 116        pr_emerg("esi: %08lx   edi: %08lx   ebp: %08lx   esp: %08lx\n",
 117                 regs->si, regs->di, regs->bp, regs->sp);
 118#else
 119        pr_emerg("rax: %016lx   rcx: %016lx   rdx: %016lx\n",
 120                 regs->ax, regs->cx, regs->dx);
 121        pr_emerg("rsi: %016lx   rdi: %016lx   rbp: %016lx   rsp: %016lx\n",
 122                 regs->si, regs->di, regs->bp, regs->sp);
 123#endif
 124        put_cpu_var(pf_reason);
 125        BUG();
 126}
 127
 128static void pre(struct kmmio_probe *p, struct pt_regs *regs,
 129                                                unsigned long addr)
 130{
 131        struct trap_reason *my_reason = &get_cpu_var(pf_reason);
 132        struct mmiotrace_rw *my_trace = &get_cpu_var(cpu_trace);
 133        const unsigned long instptr = instruction_pointer(regs);
 134        const enum reason_type type = get_ins_type(instptr);
 135        struct remap_trace *trace = p->private;
 136
 137        /* it doesn't make sense to have more than one active trace per cpu */
 138        if (my_reason->active_traces)
 139                die_kmmio_nesting_error(regs, addr);
 140        else
 141                my_reason->active_traces++;
 142
 143        my_reason->type = type;
 144        my_reason->addr = addr;
 145        my_reason->ip = instptr;
 146
 147        my_trace->phys = addr - trace->probe.addr + trace->phys;
 148        my_trace->map_id = trace->id;
 149
 150        /*
 151         * Only record the program counter when requested.
 152         * It may taint clean-room reverse engineering.
 153         */
 154        if (trace_pc)
 155                my_trace->pc = instptr;
 156        else
 157                my_trace->pc = 0;
 158
 159        /*
 160         * XXX: the timestamp recorded will be *after* the tracing has been
 161         * done, not at the time we hit the instruction. SMP implications
 162         * on event ordering?
 163         */
 164
 165        switch (type) {
 166        case REG_READ:
 167                my_trace->opcode = MMIO_READ;
 168                my_trace->width = get_ins_mem_width(instptr);
 169                break;
 170        case REG_WRITE:
 171                my_trace->opcode = MMIO_WRITE;
 172                my_trace->width = get_ins_mem_width(instptr);
 173                my_trace->value = get_ins_reg_val(instptr, regs);
 174                break;
 175        case IMM_WRITE:
 176                my_trace->opcode = MMIO_WRITE;
 177                my_trace->width = get_ins_mem_width(instptr);
 178                my_trace->value = get_ins_imm_val(instptr);
 179                break;
 180        default:
 181                {
 182                        unsigned char *ip = (unsigned char *)instptr;
 183                        my_trace->opcode = MMIO_UNKNOWN_OP;
 184                        my_trace->width = 0;
 185                        my_trace->value = (*ip) << 16 | *(ip + 1) << 8 |
 186                                                                *(ip + 2);
 187                }
 188        }
 189        put_cpu_var(cpu_trace);
 190        put_cpu_var(pf_reason);
 191}
 192
 193static void post(struct kmmio_probe *p, unsigned long condition,
 194                                                        struct pt_regs *regs)
 195{
 196        struct trap_reason *my_reason = &get_cpu_var(pf_reason);
 197        struct mmiotrace_rw *my_trace = &get_cpu_var(cpu_trace);
 198
 199        /* this should always return the active_trace count to 0 */
 200        my_reason->active_traces--;
 201        if (my_reason->active_traces) {
 202                pr_emerg("unexpected post handler");
 203                BUG();
 204        }
 205
 206        switch (my_reason->type) {
 207        case REG_READ:
 208                my_trace->value = get_ins_reg_val(my_reason->ip, regs);
 209                break;
 210        default:
 211                break;
 212        }
 213
 214        mmio_trace_rw(my_trace);
 215        put_cpu_var(cpu_trace);
 216        put_cpu_var(pf_reason);
 217}
 218
 219static void ioremap_trace_core(resource_size_t offset, unsigned long size,
 220                                                        void __iomem *addr)
 221{
 222        static atomic_t next_id;
 223        struct remap_trace *trace = kmalloc(sizeof(*trace), GFP_KERNEL);
 224        /* These are page-unaligned. */
 225        struct mmiotrace_map map = {
 226                .phys = offset,
 227                .virt = (unsigned long)addr,
 228                .len = size,
 229                .opcode = MMIO_PROBE
 230        };
 231
 232        if (!trace) {
 233                pr_err("kmalloc failed in ioremap\n");
 234                return;
 235        }
 236
 237        *trace = (struct remap_trace) {
 238                .probe = {
 239                        .addr = (unsigned long)addr,
 240                        .len = size,
 241                        .pre_handler = pre,
 242                        .post_handler = post,
 243                        .private = trace
 244                },
 245                .phys = offset,
 246                .id = atomic_inc_return(&next_id)
 247        };
 248        map.map_id = trace->id;
 249
 250        spin_lock_irq(&trace_lock);
 251        if (!is_enabled()) {
 252                kfree(trace);
 253                goto not_enabled;
 254        }
 255
 256        mmio_trace_mapping(&map);
 257        list_add_tail(&trace->list, &trace_list);
 258        if (!nommiotrace)
 259                register_kmmio_probe(&trace->probe);
 260
 261not_enabled:
 262        spin_unlock_irq(&trace_lock);
 263}
 264
 265void mmiotrace_ioremap(resource_size_t offset, unsigned long size,
 266                                                void __iomem *addr)
 267{
 268        if (!is_enabled()) /* recheck and proper locking in *_core() */
 269                return;
 270
 271        pr_debug("ioremap_*(0x%llx, 0x%lx) = %p\n",
 272                 (unsigned long long)offset, size, addr);
 273        if ((filter_offset) && (offset != filter_offset))
 274                return;
 275        ioremap_trace_core(offset, size, addr);
 276}
 277
 278static void iounmap_trace_core(volatile void __iomem *addr)
 279{
 280        struct mmiotrace_map map = {
 281                .phys = 0,
 282                .virt = (unsigned long)addr,
 283                .len = 0,
 284                .opcode = MMIO_UNPROBE
 285        };
 286        struct remap_trace *trace;
 287        struct remap_trace *tmp;
 288        struct remap_trace *found_trace = NULL;
 289
 290        pr_debug("Unmapping %p.\n", addr);
 291
 292        spin_lock_irq(&trace_lock);
 293        if (!is_enabled())
 294                goto not_enabled;
 295
 296        list_for_each_entry_safe(trace, tmp, &trace_list, list) {
 297                if ((unsigned long)addr == trace->probe.addr) {
 298                        if (!nommiotrace)
 299                                unregister_kmmio_probe(&trace->probe);
 300                        list_del(&trace->list);
 301                        found_trace = trace;
 302                        break;
 303                }
 304        }
 305        map.map_id = (found_trace) ? found_trace->id : -1;
 306        mmio_trace_mapping(&map);
 307
 308not_enabled:
 309        spin_unlock_irq(&trace_lock);
 310        if (found_trace) {
 311                synchronize_rcu(); /* unregister_kmmio_probe() requirement */
 312                kfree(found_trace);
 313        }
 314}
 315
 316void mmiotrace_iounmap(volatile void __iomem *addr)
 317{
 318        might_sleep();
 319        if (is_enabled()) /* recheck and proper locking in *_core() */
 320                iounmap_trace_core(addr);
 321}
 322
 323int mmiotrace_printk(const char *fmt, ...)
 324{
 325        int ret = 0;
 326        va_list args;
 327        unsigned long flags;
 328        va_start(args, fmt);
 329
 330        spin_lock_irqsave(&trace_lock, flags);
 331        if (is_enabled())
 332                ret = mmio_trace_printk(fmt, args);
 333        spin_unlock_irqrestore(&trace_lock, flags);
 334
 335        va_end(args);
 336        return ret;
 337}
 338EXPORT_SYMBOL(mmiotrace_printk);
 339
 340static void clear_trace_list(void)
 341{
 342        struct remap_trace *trace;
 343        struct remap_trace *tmp;
 344
 345        /*
 346         * No locking required, because the caller ensures we are in a
 347         * critical section via mutex, and is_enabled() is false,
 348         * i.e. nothing can traverse or modify this list.
 349         * Caller also ensures is_enabled() cannot change.
 350         */
 351        list_for_each_entry(trace, &trace_list, list) {
 352                pr_notice("purging non-iounmapped trace @0x%08lx, size 0x%lx.\n",
 353                          trace->probe.addr, trace->probe.len);
 354                if (!nommiotrace)
 355                        unregister_kmmio_probe(&trace->probe);
 356        }
 357        synchronize_rcu(); /* unregister_kmmio_probe() requirement */
 358
 359        list_for_each_entry_safe(trace, tmp, &trace_list, list) {
 360                list_del(&trace->list);
 361                kfree(trace);
 362        }
 363}
 364
 365#ifdef CONFIG_HOTPLUG_CPU
 366static cpumask_var_t downed_cpus;
 367
 368static void enter_uniprocessor(void)
 369{
 370        int cpu;
 371        int err;
 372
 373        if (!cpumask_available(downed_cpus) &&
 374            !alloc_cpumask_var(&downed_cpus, GFP_KERNEL)) {
 375                pr_notice("Failed to allocate mask\n");
 376                goto out;
 377        }
 378
 379        get_online_cpus();
 380        cpumask_copy(downed_cpus, cpu_online_mask);
 381        cpumask_clear_cpu(cpumask_first(cpu_online_mask), downed_cpus);
 382        if (num_online_cpus() > 1)
 383                pr_notice("Disabling non-boot CPUs...\n");
 384        put_online_cpus();
 385
 386        for_each_cpu(cpu, downed_cpus) {
 387                err = remove_cpu(cpu);
 388                if (!err)
 389                        pr_info("CPU%d is down.\n", cpu);
 390                else
 391                        pr_err("Error taking CPU%d down: %d\n", cpu, err);
 392        }
 393out:
 394        if (num_online_cpus() > 1)
 395                pr_warn("multiple CPUs still online, may miss events.\n");
 396}
 397
 398static void leave_uniprocessor(void)
 399{
 400        int cpu;
 401        int err;
 402
 403        if (!cpumask_available(downed_cpus) || cpumask_weight(downed_cpus) == 0)
 404                return;
 405        pr_notice("Re-enabling CPUs...\n");
 406        for_each_cpu(cpu, downed_cpus) {
 407                err = add_cpu(cpu);
 408                if (!err)
 409                        pr_info("enabled CPU%d.\n", cpu);
 410                else
 411                        pr_err("cannot re-enable CPU%d: %d\n", cpu, err);
 412        }
 413}
 414
 415#else /* !CONFIG_HOTPLUG_CPU */
 416static void enter_uniprocessor(void)
 417{
 418        if (num_online_cpus() > 1)
 419                pr_warn("multiple CPUs are online, may miss events. "
 420                        "Suggest booting with maxcpus=1 kernel argument.\n");
 421}
 422
 423static void leave_uniprocessor(void)
 424{
 425}
 426#endif
 427
 428void enable_mmiotrace(void)
 429{
 430        mutex_lock(&mmiotrace_mutex);
 431        if (is_enabled())
 432                goto out;
 433
 434        if (nommiotrace)
 435                pr_info("MMIO tracing disabled.\n");
 436        kmmio_init();
 437        enter_uniprocessor();
 438        spin_lock_irq(&trace_lock);
 439        atomic_inc(&mmiotrace_enabled);
 440        spin_unlock_irq(&trace_lock);
 441        pr_info("enabled.\n");
 442out:
 443        mutex_unlock(&mmiotrace_mutex);
 444}
 445
 446void disable_mmiotrace(void)
 447{
 448        mutex_lock(&mmiotrace_mutex);
 449        if (!is_enabled())
 450                goto out;
 451
 452        spin_lock_irq(&trace_lock);
 453        atomic_dec(&mmiotrace_enabled);
 454        BUG_ON(is_enabled());
 455        spin_unlock_irq(&trace_lock);
 456
 457        clear_trace_list(); /* guarantees: no more kmmio callbacks */
 458        leave_uniprocessor();
 459        kmmio_cleanup();
 460        pr_info("disabled.\n");
 461out:
 462        mutex_unlock(&mmiotrace_mutex);
 463}
 464