Linode Xen 下 grsecurity >= 4.3 崩溃问题

自从 Linux 4.3 开始，在 Linode 上使用 PaX/grsecurity 时，内核会在被 pv-grub 执行后不久立即崩溃。由于崩溃是在启动后极早期立刻发生的，没有任何可以用来调试的日志，同时公司也不是盖子开的，也没有办法得到母机上有意义的调试信息。这导致了盖子的 VPS 内核从去年 12 月开始被锁定在 4.2.7。由于不知什么时候产生了 Linode 东京机房会在 2016 年 6 月从 Xen 迁移到 KVM 的错觉，也没有花精力去尝试调试这个问题。

然而今年 Linode 周年庆时硬件全部翻倍，惟独东京机房除外。而根据官方最新的说法，新机房乐观估计要第四季度上线。解决内核问题就不得不提上了盖子的日程，首先是手工修复了不少 CVE 高危漏洞，随后又祭出 diff 折腾半天，内核始终会在启动后立刻死亡。而由于 grsecurity 并不提供 git 源，所以 git bisect 也是不可能的，唯一可用的工具只有 Linux 4.2.7 / 补丁文件，与 Linux 4.3.3 / 补丁文件。

在阅读代码差异时，一个很大的挑战是如何区分上游内核的修改与下游 PaX/grsecurity 补丁的修改。直接比较补丁文件会导致代码上下文丢失，让代码的意图不可理解。最后盖子打算编写一个名为 metadiff 的工具，自动比较并去除在上游中出现的代码段，以便仅仅对 PaX/grsecurity 的代码进行比较，就连名字都想好了就叫 metadiff ，但一直没有动手。

直到上个月和 Shawn 聊天时，提到了自己装个 Xen 也不是不可行；于是周六终于动手在 VirutalBox 虚拟机里撞了个 Debian + Xen，又在 Xen 里启动了一个虚拟机，果然很快就得到了内核崩溃的 traceback。

rip: ffffffff8100b2b0 pmu_msr_read+0x10
flags: 00000282 i s nz
rsp: ffffffff81aeff30
rax: 8000000000000000  rcx: 0000000000000001  rdx: ffffffff81aeffcc
rbx: 00000000c0000080  rsi: ffffffff81aeffa0  rdi: 00000000c0000080
rbp: ffffffff81aeffa0  r8: 0000000000000001  r9: 00000000ffffffff
r10: ffffffff81cf9000  r11: 0000000000000000  r12: ffffffff81aeffcc
r13: ffffffff81aeffc4  r14: ffffffff81aeffc0  r15: 6f73b764afec1c9d
 cs: e033    ss: e02b    ds: 0000    es: 0000
 fs: 0000 @ 0000000000000000
 gs: 0000 @ 0000000000000000/0000000000000000
Code (instr addr ffffffff8100b2b0)
00 00 00 00 00 41 54 49 89 d4 55 48 89 f5 53 89 fb 48 83 ec 10 <65> 48 8b 04 25 28 00 00 00 48 89 

Stack:
 0000000000000001 0000000000000000 0000000000000000 ffffffff8100b2b0
 000000010000e030 0000000000010082 ffffffff81aeff70 000000000000e02b
 0000000000000000 0000000000000000 00000000c0000080 ffffffff81aeffcc
 ffffffff81aeffc8 ffffffff810041c8 ffffffff81aeffc8 ffffffff81aeffcc

Call Trace:
 [<ffffffff8100b2b0>] pmu_msr_read+0x10 <--
 [<ffffffff8100b2b0>] pmu_msr_read+0x10
 [<ffffffff810041c8>] xen_read_msr_safe+0x18
 [<ffffffff81be93eb>] xen_start_kernel+0x1b9

哦？可见内核在 xen_start_kernel 不久就崩溃了，这是 /* First C function to be called on Xen boot */，在如此早期就崩溃，什么错误日志到看不到也就不奇怪了。来看看 xen_read_msr 和 pmu_msr_read 在 4.2 和 4.3 之间有什么改变：

--- ../../4.2.7/linux-4.2.7/arch/x86/xen/enlighten.c  2016-09-11 00:44:12.010022936 +0800
+++ arch/x86/xen/enlighten.c  2015-12-15 13:41:43.000000000 +0800

@@ -1030,6 +1034,9 @@ static u64 xen_read_msr_safe(unsigned in
 {
    u64 val;

+    if (pmu_msr_read(msr, &val, err))
+        return val;
+
    val = native_read_msr_safe(msr, err);
    switch (msr) {
    case MSR_IA32_APICBASE:
@@ -1074,9 +1081,11 @@ static int xen_write_msr_safe(unsigned i
        /* Fast syscall setup is all done in hypercalls, so
          these are all ignored. Stub them out here to stop
          Xen console noise. */
+        break;

    default:
-        ret = native_write_msr_safe(msr, low, high);
+        if (!pmu_msr_write(msr, low, high, &ret))
+            ret = native_write_msr_safe(msr, low, high);
    }

    return ret;

可见 pmu_msr_read 完全是个新东西，使用 git blame 继续追查。

xen/PMU: Initialization code for Xen PMU 65d0cf0be79feebeb19e7626fd3ed41ae73f642d
xen/PMU: Describe vendor-specific PMU registers e27b72df01109c689062caeba1defa013b759e0e
xen/PMU: Intercept PMU-related MSR and APIC accesses 6b08cd6328c58a2ae190c5ee03a2ffcab5ef828e
xen/PMU: PMU emulation code bf6dfb154d935725c9a2005033ca33017b9df439

发现 PMU 是 Xen 在 4.3 进入主线内核的新特性，于是解决方法就很简单了，把 bf6dfb 和 6b08cd 都撤销就好，接下来的事情就让 PaX Team 和 spender 去追查吧。最后的补丁是：

diff -uprN linux-4.7.3-hardened/arch/x86/xen/apic.c linux-4.7.3-hardened.good/arch/x86/xen/apic.c
  --- linux-4.7.3-hardened/arch/x86/xen/apic.c  2016-07-24 19:23:50.000000000 +0000
  +++ linux-4.7.3-hardened.good/arch/x86/xen/apic.c  2016-09-10 20:05:21.450647009 +0000
  @@ -7,7 +7,6 @@
   #include <xen/xen.h>
   #include <xen/interface/physdev.h>
   #include "xen-ops.h"
  -#include "pmu.h"
   #include "smp.h"

   static unsigned int xen_io_apic_read(unsigned apic, unsigned reg)
  @@ -73,10 +72,8 @@ static u32 xen_apic_read(u32 reg)

   static void xen_apic_write(u32 reg, u32 val)
   {
  -  if (reg == APIC_LVTPC) {
  -   (void)pmu_apic_update(reg);
  +  if (reg == APIC_LVTPC)
      return;
  -  }

    /* Warn to see if there's any stray references */
    WARN(1,"register: %x, value: %x\n", reg, val);
  diff -uprN linux-4.7.3-hardened/arch/x86/xen/enlighten.c linux-4.7.3-hardened.good/arch/x86/xen/enlighten.c
  --- linux-4.7.3-hardened/arch/x86/xen/enlighten.c  2016-09-10 19:59:29.237313676 +0000
  +++ linux-4.7.3-hardened.good/arch/x86/xen/enlighten.c  2016-09-10 20:06:49.683980342 +0000
  @@ -1031,9 +1031,6 @@ static u64 xen_read_msr_safe(unsigned in
   {
    u64 val;

  -  if (pmu_msr_read(msr, &val, err))
  -   return val;
  -
    val = native_read_msr_safe(msr, err);
    switch (msr) {
    case MSR_IA32_APICBASE:
  @@ -1081,13 +1078,17 @@ static int xen_write_msr_safe(unsigned i
      break;

    default:
  -   if (!pmu_msr_write(msr, low, high, &ret))
  -     ret = native_write_msr_safe(msr, low, high);
  +   ret = native_write_msr_safe(msr, low, high);
    }

    return ret;
   }

  +unsigned long long xen_read_pmc(int counter)
  +{
  +  return 0;
  +}
  +
   static u64 xen_read_msr(unsigned int msr)
   {
    /*
  diff -uprN linux-4.7.3-hardened/arch/x86/xen/pmu.c linux-4.7.3-hardened.good/arch/x86/xen/pmu.c
  --- linux-4.7.3-hardened/arch/x86/xen/pmu.c  2016-07-24 19:23:50.000000000 +0000
  +++ linux-4.7.3-hardened.good/arch/x86/xen/pmu.c  2016-09-10 20:05:21.450647009 +0000
  @@ -13,20 +13,11 @@
   /* x86_pmu.handle_irq definition */
   #include "../events/perf_event.h"

  -#define XENPMU_IRQ_PROCESSING  1
  -struct xenpmu {
  -  /* Shared page between hypervisor and domain */
  -  struct xen_pmu_data *xenpmu_data;

  -  uint8_t flags;
  -};
  -static DEFINE_PER_CPU(struct xenpmu, xenpmu_shared);
  -#define get_xenpmu_data()  (this_cpu_ptr(&xenpmu_shared)->xenpmu_data)
  -#define get_xenpmu_flags()  (this_cpu_ptr(&xenpmu_shared)->flags)
  -
  -/* Macro for computing address of a PMU MSR bank */
  -#define field_offset(ctxt, field) ((void *)((uintptr_t)ctxt + \
  -          (uintptr_t)ctxt->field))
  +/* Shared page between hypervisor and domain */
  +static DEFINE_PER_CPU(struct xen_pmu_data *, xenpmu_shared);
  +#define get_xenpmu_data()  per_cpu(xenpmu_shared, smp_processor_id())
  +

   /* AMD PMU */
   #define F15H_NUM_COUNTERS  6
  @@ -60,8 +51,6 @@ static __read_mostly int amd_num_counter
   /* Alias registers (0x4c1) for full-width writes to PMCs */
   #define MSR_PMC_ALIAS_MASK     (~(MSR_IA32_PERFCTR0 ^ MSR_IA32_PMC0))

  -#define INTEL_PMC_TYPE_SHIFT    30
  -
   static __read_mostly int intel_num_arch_counters, intel_num_fixed_counters;

  @@ -178,232 +167,6 @@ static int is_intel_pmu_msr(u32 msr_inde
    }
   }

  -static bool xen_intel_pmu_emulate(unsigned int msr, u64 *val, int type,
  -       int index, bool is_read)
  -{
  -  uint64_t *reg = NULL;
  -  struct xen_pmu_intel_ctxt *ctxt;
  -  uint64_t *fix_counters;
  -  struct xen_pmu_cntr_pair *arch_cntr_pair;
  -  struct xen_pmu_data *xenpmu_data = get_xenpmu_data();
  -  uint8_t xenpmu_flags = get_xenpmu_flags();
  -
  -
  -  if (!xenpmu_data || !(xenpmu_flags & XENPMU_IRQ_PROCESSING))
  -   return false;
  -
  -  ctxt = &xenpmu_data->pmu.c.intel;
  -
  -  switch (msr) {
  -  case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
  -   reg = &ctxt->global_ovf_ctrl;
  -   break;
  -  case MSR_CORE_PERF_GLOBAL_STATUS:
  -   reg = &ctxt->global_status;
  -   break;
  -  case MSR_CORE_PERF_GLOBAL_CTRL:
  -   reg = &ctxt->global_ctrl;
  -   break;
  -  case MSR_CORE_PERF_FIXED_CTR_CTRL:
  -   reg = &ctxt->fixed_ctrl;
  -   break;
  -  default:
  -   switch (type) {
  -   case MSR_TYPE_COUNTER:
  -     fix_counters = field_offset(ctxt, fixed_counters);
  -     reg = &fix_counters[index];
  -     break;
  -   case MSR_TYPE_ARCH_COUNTER:
  -     arch_cntr_pair = field_offset(ctxt, arch_counters);
  -     reg = &arch_cntr_pair[index].counter;
  -     break;
  -   case MSR_TYPE_ARCH_CTRL:
  -     arch_cntr_pair = field_offset(ctxt, arch_counters);
  -     reg = &arch_cntr_pair[index].control;
  -     break;
  -   default:
  -     return false;
  -   }
  -  }
  -
  -  if (reg) {
  -   if (is_read)
  -     *val = *reg;
  -   else {
  -     *reg = *val;
  -
  -     if (msr == MSR_CORE_PERF_GLOBAL_OVF_CTRL)
  -      ctxt->global_status &= (~(*val));
  -   }
  -   return true;
  -  }
  -
  -  return false;
  -}
  -
  -static bool xen_amd_pmu_emulate(unsigned int msr, u64 *val, bool is_read)
  -{
  -  uint64_t *reg = NULL;
  -  int i, off = 0;
  -  struct xen_pmu_amd_ctxt *ctxt;
  -  uint64_t *counter_regs, *ctrl_regs;
  -  struct xen_pmu_data *xenpmu_data = get_xenpmu_data();
  -  uint8_t xenpmu_flags = get_xenpmu_flags();
  -
  -  if (!xenpmu_data || !(xenpmu_flags & XENPMU_IRQ_PROCESSING))
  -   return false;
  -
  -  if (k7_counters_mirrored &&
  -    ((msr >= MSR_K7_EVNTSEL0) && (msr <= MSR_K7_PERFCTR3)))
  -   msr = get_fam15h_addr(msr);
  -
  -  ctxt = &xenpmu_data->pmu.c.amd;
  -  for (i = 0; i < amd_num_counters; i++) {
  -   if (msr == amd_ctrls_base + off) {
  -     ctrl_regs = field_offset(ctxt, ctrls);
  -     reg = &ctrl_regs[i];
  -     break;
  -   } else if (msr == amd_counters_base + off) {
  -     counter_regs = field_offset(ctxt, counters);
  -     reg = &counter_regs[i];
  -     break;
  -   }
  -   off += amd_msr_step;
  -  }
  -
  -  if (reg) {
  -   if (is_read)
  -     *val = *reg;
  -   else
  -     *reg = *val;
  -
  -   return true;
  -  }
  -  return false;
  -}
  -
  -bool pmu_msr_read(unsigned int msr, uint64_t *val, int *err)
  -{
  -  if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
  -   if (is_amd_pmu_msr(msr)) {
  -     if (!xen_amd_pmu_emulate(msr, val, 1))
  -      *val = native_read_msr_safe(msr, err);
  -     return true;
  -   }
  -  } else {
  -   int type, index;
  -
  -   if (is_intel_pmu_msr(msr, &type, &index)) {
  -     if (!xen_intel_pmu_emulate(msr, val, type, index, 1))
  -      *val = native_read_msr_safe(msr, err);
  -     return true;
  -   }
  -  }
  -
  -  return false;
  -}
  -
  -bool pmu_msr_write(unsigned int msr, uint32_t low, uint32_t high, int *err)
  -{
  -  uint64_t val = ((uint64_t)high << 32) | low;
  -
  -  if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
  -   if (is_amd_pmu_msr(msr)) {
  -     if (!xen_amd_pmu_emulate(msr, &val, 0))
  -      *err = native_write_msr_safe(msr, low, high);
  -     return true;
  -   }
  -  } else {
  -   int type, index;
  -
  -   if (is_intel_pmu_msr(msr, &type, &index)) {
  -     if (!xen_intel_pmu_emulate(msr, &val, type, index, 0))
  -      *err = native_write_msr_safe(msr, low, high);
  -     return true;
  -   }
  -  }
  -
  -  return false;
  -}
  -
  -static unsigned long long xen_amd_read_pmc(int counter)
  -{
  -  struct xen_pmu_amd_ctxt *ctxt;
  -  uint64_t *counter_regs;
  -  struct xen_pmu_data *xenpmu_data = get_xenpmu_data();
  -  uint8_t xenpmu_flags = get_xenpmu_flags();
  -
  -  if (!xenpmu_data || !(xenpmu_flags & XENPMU_IRQ_PROCESSING)) {
  -   uint32_t msr;
  -   int err;
  -
  -   msr = amd_counters_base + (counter * amd_msr_step);
  -   return native_read_msr_safe(msr, &err);
  -  }
  -
  -  ctxt = &xenpmu_data->pmu.c.amd;
  -  counter_regs = field_offset(ctxt, counters);
  -  return counter_regs[counter];
  -}
  -
  -static unsigned long long xen_intel_read_pmc(int counter)
  -{
  -  struct xen_pmu_intel_ctxt *ctxt;
  -  uint64_t *fixed_counters;
  -  struct xen_pmu_cntr_pair *arch_cntr_pair;
  -  struct xen_pmu_data *xenpmu_data = get_xenpmu_data();
  -  uint8_t xenpmu_flags = get_xenpmu_flags();
  -
  -  if (!xenpmu_data || !(xenpmu_flags & XENPMU_IRQ_PROCESSING)) {
  -   uint32_t msr;
  -   int err;
  -
  -   if (counter & (1 << INTEL_PMC_TYPE_SHIFT))
  -     msr = MSR_CORE_PERF_FIXED_CTR0 + (counter & 0xffff);
  -   else
  -     msr = MSR_IA32_PERFCTR0 + counter;
  -
  -   return native_read_msr_safe(msr, &err);
  -  }
  -
  -  ctxt = &xenpmu_data->pmu.c.intel;
  -  if (counter & (1 << INTEL_PMC_TYPE_SHIFT)) {
  -   fixed_counters = field_offset(ctxt, fixed_counters);
  -   return fixed_counters[counter & 0xffff];
  -  }
  -
  -  arch_cntr_pair = field_offset(ctxt, arch_counters);
  -  return arch_cntr_pair[counter].counter;
  -}
  -
  -unsigned long long xen_read_pmc(int counter)
  -{
  -  if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
  -   return xen_amd_read_pmc(counter);
  -  else
  -   return xen_intel_read_pmc(counter);
  -}
  -
  -int pmu_apic_update(uint32_t val)
  -{
  -  int ret;
  -  struct xen_pmu_data *xenpmu_data = get_xenpmu_data();
  -
  -  if (!xenpmu_data) {
  -   pr_warn_once("%s: pmudata not initialized\n", __func__);
  -   return -EINVAL;
  -  }
  -
  -  xenpmu_data->pmu.l.lapic_lvtpc = val;
  -
  -  if (get_xenpmu_flags() & XENPMU_IRQ_PROCESSING)
  -   return 0;
  -
  -  ret = HYPERVISOR_xenpmu_op(XENPMU_lvtpc_set, NULL);
  -
  -  return ret;
  -}
  -
   /* perf callbacks */
   static int xen_is_in_guest(void)
   {
  @@ -476,37 +239,26 @@ static void xen_convert_regs(const struc

   irqreturn_t xen_pmu_irq_handler(int irq, void *dev_id)
   {
  -  int err, ret = IRQ_NONE;
  +  int ret = IRQ_NONE;
    struct pt_regs regs;
    const struct xen_pmu_data *xenpmu_data = get_xenpmu_data();
  -  uint8_t xenpmu_flags = get_xenpmu_flags();

    if (!xenpmu_data) {
      pr_warn_once("%s: pmudata not initialized\n", __func__);
      return ret;
    }

  -  this_cpu_ptr(&xenpmu_shared)->flags =
  -   xenpmu_flags | XENPMU_IRQ_PROCESSING;
    xen_convert_regs(&xenpmu_data->pmu.r.regs, ®s,
        xenpmu_data->pmu.pmu_flags);
    if (x86_pmu.handle_irq(®s))
      ret = IRQ_HANDLED;

  -  /* Write out cached context to HW */
  -  err = HYPERVISOR_xenpmu_op(XENPMU_flush, NULL);
  -  this_cpu_ptr(&xenpmu_shared)->flags = xenpmu_flags;
  -  if (err) {
  -   pr_warn_once("%s: failed hypercall, err: %d\n", __func__, err);
  -   return IRQ_NONE;
  -  }
  -
    return ret;
   }

   bool is_xen_pmu(int cpu)
   {
  -  return (get_xenpmu_data() != NULL);
  +  return (per_cpu(xenpmu_shared, cpu) != NULL);
   }

   void xen_pmu_init(int cpu)
  @@ -536,8 +288,7 @@ void xen_pmu_init(int cpu)
    if (err)
      goto fail;

  -  per_cpu(xenpmu_shared, cpu).xenpmu_data = xenpmu_data;
  -  per_cpu(xenpmu_shared, cpu).flags = 0;
  +  per_cpu(xenpmu_shared, cpu) = xenpmu_data;

    if (cpu == 0) {
      perf_register_guest_info_callbacks(&xen_guest_cbs);
  @@ -565,6 +316,6 @@ void xen_pmu_finish(int cpu)

    (void)HYPERVISOR_xenpmu_op(XENPMU_finish, &xp);

  -  free_pages((unsigned long)per_cpu(xenpmu_shared, cpu).xenpmu_data, 0);
  -  per_cpu(xenpmu_shared, cpu).xenpmu_data = NULL;
  +  free_pages((unsigned long)per_cpu(xenpmu_shared, cpu), 0);
  +  per_cpu(xenpmu_shared, cpu) = NULL;
   }
  diff -uprN linux-4.7.3-hardened/arch/x86/xen/pmu.h linux-4.7.3-hardened.good/arch/x86/xen/pmu.h
  --- linux-4.7.3-hardened/arch/x86/xen/pmu.h  2016-07-24 19:23:50.000000000 +0000
  +++ linux-4.7.3-hardened.good/arch/x86/xen/pmu.h  2016-09-10 20:05:21.453980342 +0000
  @@ -7,9 +7,5 @@ irqreturn_t xen_pmu_irq_handler(int irq,
   void xen_pmu_init(int cpu);
   void xen_pmu_finish(int cpu);
   bool is_xen_pmu(int cpu);
  -bool pmu_msr_read(unsigned int msr, uint64_t *val, int *err);
  -bool pmu_msr_write(unsigned int msr, uint32_t low, uint32_t high, int *err);
  -int pmu_apic_update(uint32_t reg);
  -unsigned long long xen_read_pmc(int counter);

   #endif /* __XEN_PMU_H */

打好补丁再编译内核，被智子锁定版本的内核果然升级成功了。

$ uname -r
4.7.3-hardened

更新：官方已在 grsecurity-3.1-4.7.4-201609152234.patch 中修复问题，不再需要此 workaround。