< 1 > The background. The current crash code is still based at kernel v4.20, but the kernel is v5.17-rc4(now). The MODULE/VMALLOC/VMEMMAP ranges are not be updated since v4.20. I list all the changes from kernel v4.20 to v5.17: 1.) The current crash code is based at kernel v4.20. The virtual memory layout looks like this: +--------------------------------------------------------------------+ | KASAN | MODULE | VMALLOC | .... | VMEMMAP | +--------------------------------------------------------------------+ The macros are: #define MODULES_VADDR (VA_START + KASAN_SHADOW_SIZE) #define MODULES_END (MODULES_VADDR + MODULES_VSIZE) #define VMALLOC_START (MODULES_END) #define VMALLOC_END (PAGE_OFFSET - PUD_SIZE - VMEMMAP_SIZE - SZ_64K) #define VMEMMAP_START (PAGE_OFFSET - VMEMMAP_SIZE) 2.) In the kernel v5.0, the patch will add a new BFP JIT region: "91fc957c9b1d arm64/bpf: don't allocate BPF JIT programs in module memory" The virtual memory layout looks like this: +--------------------------------------------------------------------+ | KASAN | BPF_JIT | MODULE | VMALLOC | .... | VMEMMAP | +--------------------------------------------------------------------+ The macros are: #define MODULES_VADDR (BPF_JIT_REGION_END) #define MODULES_END (MODULES_VADDR + MODULES_VSIZE) #define VMALLOC_START (MODULES_END) #define VMALLOC_END (PAGE_OFFSET - PUD_SIZE - VMEMMAP_SIZE - SZ_64K) #define VMEMMAP_START (PAGE_OFFSET - VMEMMAP_SIZE) The layout does not changed until v5.4. 3.) In the kernel v5.4, several patches changes the layout, such as: "ce3aaed87344 arm64: mm: Modify calculation of VMEMMAP_SIZE" "14c127c957c1 arm64: mm: Flip kernel VA space" and the virtual memory layout looks like this: +--------------------------------------------------------------------+ | KASAN | BPF_JIT | MODULE | VMALLOC | .... | VMEMMAP | +--------------------------------------------------------------------+ The macros are: #define MODULES_VADDR (BPF_JIT_REGION_END) #define MODULES_END (MODULES_VADDR + MODULES_VSIZE) #define VMALLOC_START (MODULES_END) #define VMALLOC_END (- PUD_SIZE - VMEMMAP_SIZE - SZ_64K) #define VMEMMAP_START (-VMEMMAP_SIZE - SZ_2M) In the v5.7, the patch: "bbd6ec605c arm64/mm: Enable memory hot remove" adds the VMEMMAP_END. 4.) In the kernel v5.11, several patches changes the layout, such as: "9ad7c6d5e75b arm64: mm: tidy up top of kernel VA space" "f4693c2716b3 arm64: mm: extend linear region for 52-bit VA configurations" and the virtual memory layout looks like this: +--------------------------------------------------------------------+ | BPF_JIT | MODULE | VMALLOC | .... | VMEMMAP | +--------------------------------------------------------------------+ The macros are: #define MODULES_VADDR (BPF_JIT_REGION_END) #define MODULES_END (MODULES_VADDR + MODULES_VSIZE) #define VMALLOC_START (MODULES_END) #define VMALLOC_END (VMEMMAP_START - SZ_256M) #define VMEMMAP_START (-(UL(1) << (VA_BITS - VMEMMAP_SHIFT))) #define VMEMMAP_END (VMEMMAP_START + VMEMMAP_SIZE) 5.) In the kernel v5.17-rc1, after the patch "b89ddf4cca43 arm64/bpf: Remove 128MB limit for BPF JIT programs" the virtual memory layout looks like this: +--------------------------------------------------------------------+ | MODULE | VMALLOC | .... | VMEMMAP | +--------------------------------------------------------------------+ The macros are: #define MODULES_VADDR (_PAGE_END(VA_BITS_MIN)) #define MODULES_END (MODULES_VADDR + MODULES_VSIZE) #define VMALLOC_START (MODULES_END) #define VMALLOC_END (VMEMMAP_START - SZ_256M) #define VMEMMAP_START (-(UL(1) << (VA_BITS - VMEMMAP_SHIFT))) #define VMEMMAP_END (VMEMMAP_START + VMEMMAP_SIZE) < 2 > What does this patch do? 1.) Use arm64_get_struct_page_size() to get the size of struct page{} in the PRE_GDB. 2.) If we can succeed in above step, we will try to call arm64_get_va_range() to get the proper kernel virtual ranges. In the arm64_get_va_range(), we calculate the ranges by the hooks of different kernel versions: get_range: arm64_get_range_v5_17, get_range: arm64_get_range_v5_11, get_range: arm64_get_range_v5_4, get_range: arm64_get_range_v5_0, 3.) If we can succeed in above steps, the arm64_calc_virtual_memory_ranges() will be ignored. If we failed in above steps, the arm64_calc_virtual_memory_ranges() will continue to do its work. < 3 > Test this patch. Tested this patch with a vmcore produced by a 5.4.119 kernel panic. (The CONFIG_KASAN is NOT set for this kernel.) Before this patch, we get the wrong output from "help -m": ---------------------------------------------------------- vmalloc_start_addr: ffff800048000000 vmalloc_end: fffffdffbffeffff modules_vaddr: ffff800040000000 modules_end: ffff800047ffffff vmemmap_vaddr: fffffdffffe00000 vmemmap_end: ffffffffffffffff ---------------------------------------------------------- After this patch, we can get the correct output from "help -m": ---------------------------------------------------------- vmalloc_start_addr: ffff800010000000 vmalloc_end: fffffdffbfff0000 modules_vaddr: ffff800008000000 modules_end: ffff800010000000 vmemmap_vaddr: fffffdffffe00000 vmemmap_end: ffffffffffffffff ---------------------------------------------------------- Signed-off-by: Huang Shijie <shijie(a)os.amperecomputing.com&gt; --- v3 --> v4: 1.) Add struct_page_size to @ms. Change some functions, such as arm64_get_struct_page_size(). (Do not use the ASSIGN_SIZE, use the ms->struct_page_size instead.) 2.) Tested again with the new code. v2 --> v3: Fount two bugs in arm64_get_range_v5_17/arm64_get_range_v5_11: We should use the ms->CONFIG_ARM64_VA_BITS to calculate the vmemmep_vaddr, not use the ms->VA_BITS. v1 --> v2: The Crash code is based on v4.20 not v4.9. Changed the commit message about it. --- arm64.c | 367 ++++++++++++++++++++++++++++++++++++++++++++++++++++++-- defs.h | 1 + 2 files changed, 357 insertions(+), 11 deletions(-) diff --git a/arm64.c b/arm64.c index 3ab8489..2f5b1c0 100644 --- a/arm64.c +++ b/arm64.c @@ -92,6 +92,14 @@ static void arm64_calc_VA_BITS(void); static int arm64_is_uvaddr(ulong, struct task_context *); static void arm64_calc_KERNELPACMASK(void); +struct kernel_range { + unsigned long modules_vaddr, modules_end; + unsigned long vmalloc_start_addr, vmalloc_end; + unsigned long vmemmap_vaddr, vmemmap_end; +}; +static struct kernel_range *arm64_get_va_range(struct machine_specific *ms); +static void arm64_get_struct_page_size(struct machine_specific *ms); + static void arm64_calc_kernel_start(void) { struct machine_specific *ms = machdep->machspec; @@ -233,9 +241,10 @@ arm64_init(int when) machdep->pageoffset = machdep->pagesize - 1; machdep->pagemask = ~((ulonglong)machdep->pageoffset); + ms = machdep->machspec; + arm64_get_struct_page_size(ms); arm64_calc_VA_BITS(); arm64_calc_KERNELPACMASK(); - ms = machdep->machspec; /* vabits_actual introduced after mm flip, so it should be flipped layout */ if (ms->VA_BITS_ACTUAL) { @@ -252,8 +261,15 @@ arm64_init(int when) } machdep->is_kvaddr = generic_is_kvaddr; machdep->kvtop = arm64_kvtop; + + /* The defaults */ + ms->vmalloc_end = ARM64_VMALLOC_END; + ms->vmemmap_vaddr = ARM64_VMEMMAP_VADDR; + ms->vmemmap_end = ARM64_VMEMMAP_END; + if (machdep->flags & NEW_VMEMMAP) { struct syment *sp; + struct kernel_range *r; /* It is finally decided in arm64_calc_kernel_start() */ sp = kernel_symbol_search("_text"); @@ -261,27 +277,32 @@ arm64_init(int when) sp = kernel_symbol_search("_end"); ms->kimage_end = (sp ? sp->value : 0); - if (ms->VA_BITS_ACTUAL) { + if (ms->struct_page_size && (r = arm64_get_va_range(ms))) { + /* We can get all the MODULES/VMALLOC/VMEMMAP ranges now.*/ + ms->modules_vaddr = r->modules_vaddr; + ms->modules_end = r->modules_end; + ms->vmalloc_start_addr = r->vmalloc_start_addr; + ms->vmalloc_end = r->vmalloc_end; + ms->vmemmap_vaddr = r->vmemmap_vaddr; + ms->vmemmap_end = r->vmemmap_end; + } else if (ms->VA_BITS_ACTUAL) { ms->modules_vaddr = (st->_stext_vmlinux & TEXT_OFFSET_MASK) - ARM64_MODULES_VSIZE; ms->modules_end = ms->modules_vaddr + ARM64_MODULES_VSIZE -1; + ms->vmalloc_start_addr = ms->modules_end + 1; } else { ms->modules_vaddr = ARM64_VA_START; if (kernel_symbol_exists("kasan_init")) ms->modules_vaddr += ARM64_KASAN_SHADOW_SIZE; ms->modules_end = ms->modules_vaddr + ARM64_MODULES_VSIZE -1; + ms->vmalloc_start_addr = ms->modules_end + 1; } - ms->vmalloc_start_addr = ms->modules_end + 1; - arm64_calc_kimage_voffset(); } else { ms->modules_vaddr = ARM64_PAGE_OFFSET - MEGABYTES(64); ms->modules_end = ARM64_PAGE_OFFSET - 1; ms->vmalloc_start_addr = ARM64_VA_START; } - ms->vmalloc_end = ARM64_VMALLOC_END; - ms->vmemmap_vaddr = ARM64_VMEMMAP_VADDR; - ms->vmemmap_end = ARM64_VMEMMAP_END; switch (machdep->pagesize) { @@ -404,7 +425,12 @@ arm64_init(int when) case POST_GDB: /* Rely on kernel version to decide the kernel start address */ arm64_calc_kernel_start(); - arm64_calc_virtual_memory_ranges(); + + /* Can we get the size of struct page before POST_GDB */ + ms = machdep->machspec; + if (!ms->struct_page_size) + arm64_calc_virtual_memory_ranges(); + arm64_get_section_size_bits(); if (!machdep->max_physmem_bits) { @@ -419,8 +445,6 @@ arm64_init(int when) machdep->max_physmem_bits = _MAX_PHYSMEM_BITS; } - ms = machdep->machspec; - if (CRASHDEBUG(1)) { if (ms->VA_BITS_ACTUAL) { fprintf(fp, "CONFIG_ARM64_VA_BITS: %ld\n", ms->CONFIG_ARM64_VA_BITS); @@ -511,6 +535,328 @@ arm64_init(int when) } } +struct kernel_va_range_handler { + unsigned long kernel_versions_start; /* include */ + unsigned long kernel_versions_end; /* exclude */ + struct kernel_range *(*get_range)(struct machine_specific *); +}; + +static struct kernel_range tmp_range; +#define _PAGE_END(va) (-(1UL << ((va) - 1))) +#define SZ_64K 0x00010000 +#define SZ_2M 0x00200000 + +/* + * Get the max shift of the size of struct page. + * Most of the time, it is 64 bytes, but not sure. + */ +static int arm64_get_struct_page_max_shift(struct machine_specific *ms) +{ + unsigned long v = ms->struct_page_size; + + if (16 < v && v <= 32) + return 5; + if (32 < v && v <= 64) + return 6; + if (64 < v && v <= 128) + return 7; + + error(FATAL, "We should not have such struct page size:%d!\n", v); + return 0; +} + +/* + * The change is caused by the kernel patch since v5.17-rc1: + * "b89ddf4cca43 arm64/bpf: Remove 128MB limit for BPF JIT programs" + */ +static struct kernel_range *arm64_get_range_v5_17(struct machine_specific *ms) +{ + struct kernel_range *r = &tmp_range; + unsigned long v = ms->CONFIG_ARM64_VA_BITS; + unsigned long vmem_shift, vmemmap_size; + + /* Not initialized yet */ + if (v == 0) + return NULL; + + if (v > 48) + v = 48; + + /* Get the MODULES_VADDR ~ MODULES_END */ + r->modules_vaddr = _PAGE_END(v); + r->modules_end = r->modules_vaddr + MEGABYTES(128); + + /* Get the VMEMMAP_START ~ VMEMMAP_END */ + vmem_shift = machdep->pageshift - arm64_get_struct_page_max_shift(ms); + vmemmap_size = (_PAGE_END(v) - PAGE_OFFSET) >> vmem_shift; + + r->vmemmap_vaddr = (-(1UL << (ms->CONFIG_ARM64_VA_BITS - vmem_shift))); + r->vmemmap_end = r->vmemmap_vaddr + vmemmap_size; + + /* Get the VMALLOC_START ~ VMALLOC_END */ + r->vmalloc_start_addr = r->modules_end; + r->vmalloc_end = r->vmemmap_vaddr - MEGABYTES(256); + return r; +} + +/* + * The change is caused by the kernel patch since v5.11: + * "9ad7c6d5e75b arm64: mm: tidy up top of kernel VA space" + */ +static struct kernel_range *arm64_get_range_v5_11(struct machine_specific *ms) +{ + struct kernel_range *r = &tmp_range; + unsigned long v = ms->CONFIG_ARM64_VA_BITS; + unsigned long vmem_shift, vmemmap_size, bpf_jit_size = MEGABYTES(128); + + /* Not initialized yet */ + if (v == 0) + return NULL; + + if (v > 48) + v = 48; + + /* Get the MODULES_VADDR ~ MODULES_END */ + r->modules_vaddr = _PAGE_END(v) + bpf_jit_size; + r->modules_end = r->modules_vaddr + MEGABYTES(128); + + /* Get the VMEMMAP_START ~ VMEMMAP_END */ + vmem_shift = machdep->pageshift - arm64_get_struct_page_max_shift(ms); + vmemmap_size = (_PAGE_END(v) - PAGE_OFFSET) >> vmem_shift; + + r->vmemmap_vaddr = (-(1UL << (ms->CONFIG_ARM64_VA_BITS - vmem_shift))); + r->vmemmap_end = r->vmemmap_vaddr + vmemmap_size; + + /* Get the VMALLOC_START ~ VMALLOC_END */ + r->vmalloc_start_addr = r->modules_end; + r->vmalloc_end = r->vmemmap_vaddr - MEGABYTES(256); + return r; +} + +static unsigned long arm64_get_pud_size(void) +{ + unsigned long PUD_SIZE = 0; + + switch (machdep->pagesize) { + case 4096: + if (machdep->machspec->VA_BITS > PGDIR_SHIFT_L4_4K) { + PUD_SIZE = PUD_SIZE_L4_4K; + } else { + PUD_SIZE = PGDIR_SIZE_L3_4K; + } + break; + + case 65536: + PUD_SIZE = PGDIR_SIZE_L2_64K; + default: + break; + } + return PUD_SIZE; +} + +/* + * The change is caused by the kernel patches since v5.4, such as: + * "ce3aaed87344 arm64: mm: Modify calculation of VMEMMAP_SIZE" + * "14c127c957c1 arm64: mm: Flip kernel VA space" + */ +static struct kernel_range *arm64_get_range_v5_4(struct machine_specific *ms) +{ + struct kernel_range *r = &tmp_range; + unsigned long v = ms->CONFIG_ARM64_VA_BITS; + unsigned long kasan_shadow_shift, kasan_shadow_offset, PUD_SIZE; + unsigned long vmem_shift, vmemmap_size, bpf_jit_size = MEGABYTES(128); + char *string; + int ret; + + /* Not initialized yet */ + if (v == 0) + return NULL; + + if (v > 48) + v = 48; + + /* Get the MODULES_VADDR ~ MODULES_END */ + if (kernel_symbol_exists("kasan_init")) { + /* See the arch/arm64/Makefile */ + ret = get_kernel_config("CONFIG_KASAN_SW_TAGS", NULL); + if (ret == IKCONFIG_N) + return NULL; + kasan_shadow_shift = (ret == IKCONFIG_Y) ? 4: 3; + + /* See the arch/arm64/Kconfig*/ + ret = get_kernel_config("CONFIG_KASAN_SHADOW_OFFSET", &string); + if (ret != IKCONFIG_STR) + return NULL; + kasan_shadow_offset = atol(string); + + r->modules_vaddr = (1UL << (64 - kasan_shadow_shift)) + kasan_shadow_offset + + bpf_jit_size; + } else { + r->modules_vaddr = _PAGE_END(v) + bpf_jit_size; + } + + r->modules_end = r->modules_vaddr + MEGABYTES(128); + + /* Get the VMEMMAP_START ~ VMEMMAP_END */ + vmem_shift = machdep->pageshift - arm64_get_struct_page_max_shift(ms); + vmemmap_size = (_PAGE_END(v) - PAGE_OFFSET) >> vmem_shift; + + r->vmemmap_vaddr = (-vmemmap_size - SZ_2M); + if (THIS_KERNEL_VERSION >= LINUX(5, 7, 0)) { + /* + * In the v5.7, the patch: "bbd6ec605c arm64/mm: Enable memory hot remove" + * adds the VMEMMAP_END. + */ + r->vmemmap_end = r->vmemmap_vaddr + vmemmap_size; + } else { + r->vmemmap_end = 0xffffffffffffffffUL; + } + + /* Get the VMALLOC_START ~ VMALLOC_END */ + PUD_SIZE = arm64_get_pud_size(); + r->vmalloc_start_addr = r->modules_end; + r->vmalloc_end = (-PUD_SIZE - vmemmap_size - SZ_64K); + return r; +} + +/* + * The change is caused by the kernel patches since v5.0, such as: + * "91fc957c9b1d arm64/bpf: don't allocate BPF JIT programs in module memory" + */ +static struct kernel_range *arm64_get_range_v5_0(struct machine_specific *ms) +{ + struct kernel_range *r = &tmp_range; + unsigned long v = ms->CONFIG_ARM64_VA_BITS; + unsigned long kasan_shadow_shift, PUD_SIZE; + unsigned long vmemmap_size, bpf_jit_size = MEGABYTES(128); + unsigned long va_start, page_offset; + int ret; + + /* Not initialized yet */ + if (v == 0) + return NULL; + + va_start = (0xffffffffffffffffUL - (1UL << v) + 1); + page_offset = (0xffffffffffffffffUL - (1UL << (v - 1)) + 1); + + /* Get the MODULES_VADDR ~ MODULES_END */ + if (kernel_symbol_exists("kasan_init")) { + /* See the arch/arm64/Makefile */ + ret = get_kernel_config("CONFIG_KASAN_SW_TAGS", NULL); + if (ret == IKCONFIG_N) + return NULL; + kasan_shadow_shift = (ret == IKCONFIG_Y) ? 4: 3; + + r->modules_vaddr = va_start + (1UL << (v - kasan_shadow_shift)) + bpf_jit_size; + } else { + r->modules_vaddr = va_start + bpf_jit_size; + } + + r->modules_end = r->modules_vaddr + MEGABYTES(128); + + /* Get the VMEMMAP_START ~ VMEMMAP_END */ + vmemmap_size = (1UL << (v - machdep->pageshift - 1 + arm64_get_struct_page_max_shift(ms))); + + r->vmemmap_vaddr = page_offset - vmemmap_size; + r->vmemmap_end = 0xffffffffffffffffUL; /* this kernel does not have VMEMMAP_END */ + + /* Get the VMALLOC_START ~ VMALLOC_END */ + PUD_SIZE = arm64_get_pud_size(); + + r->vmalloc_start_addr = r->modules_end; + r->vmalloc_end = page_offset - PUD_SIZE - vmemmap_size - SZ_64K; + return r; +} + +static struct kernel_va_range_handler kernel_va_range_handlers[] = { + { + LINUX(5,17,0), + LINUX(6,0,0), /* Just a boundary, Change it later */ + get_range: arm64_get_range_v5_17, + }, { + LINUX(5,11,0), LINUX(5,17,0), + get_range: arm64_get_range_v5_11, + }, { + LINUX(5,4,0), LINUX(5,11,0), + get_range: arm64_get_range_v5_4, + }, { + LINUX(5,0,0), LINUX(5,4,0), + get_range: arm64_get_range_v5_0, + }, +}; + +#define ARRAY_SIZE(a) (sizeof (a) / sizeof ((a)[0])) + +static unsigned long arm64_get_kernel_version(void) +{ + char *string; + char buf[BUFSIZE]; + char *p1, *p2; + + if (THIS_KERNEL_VERSION) + return THIS_KERNEL_VERSION; + + string = pc->read_vmcoreinfo("OSRELEASE"); + if (string) { + strcpy(buf, string); + + p1 = p2 = buf; + while (*p2 != '.') + p2++; + *p2 = NULLCHAR; + kt->kernel_version[0] = atoi(p1); + + p1 = ++p2; + while (*p2 != '.') + p2++; + *p2 = NULLCHAR; + kt->kernel_version[1] = atoi(p1); + + p1 = ++p2; + while ((*p2 >= '0') && (*p2 <= '9')) + p2++; + *p2 = NULLCHAR; + kt->kernel_version[2] = atoi(p1); + } + free(string); + return THIS_KERNEL_VERSION; +} + +/* Return NULL if we fail. */ +static struct kernel_range *arm64_get_va_range(struct machine_specific *ms) +{ + struct kernel_va_range_handler *h; + unsigned long kernel_version = arm64_get_kernel_version(); + int i; + + if (!kernel_version) + return NULL; + + for (i = 0; i < ARRAY_SIZE(kernel_va_range_handlers); i++) { + h = kernel_va_range_handlers + i; + + /* Get the right hook for this kernel version */ + if (h->kernel_versions_start <= kernel_version && + kernel_version < h->kernel_versions_end) { + + /* Get the correct virtual address ranges */ + return h->get_range(ms); + } + } + return NULL; +} + +/* Get the size of struct page {} */ +static void arm64_get_struct_page_size(struct machine_specific *ms) +{ + char *string; + + string = pc->read_vmcoreinfo("SIZE(page)"); + if (string) + ms->struct_page_size = atol(string); + free(string); +} + /* * Accept or reject a symbol from the kernel namelist. */ @@ -4272,7 +4618,6 @@ arm64_calc_VA_BITS(void) #define ALIGN(x, a) __ALIGN_KERNEL((x), (a)) #define __ALIGN_KERNEL(x, a) __ALIGN_KERNEL_MASK(x, (typeof(x))(a) - 1) #define __ALIGN_KERNEL_MASK(x, mask) (((x) + (mask)) & ~(mask)) -#define SZ_64K 0x00010000 static void arm64_calc_virtual_memory_ranges(void) diff --git a/defs.h b/defs.h index bf2c59b..81ac049 100644 --- a/defs.h +++ b/defs.h @@ -3386,6 +3386,7 @@ struct machine_specific { ulong VA_START; ulong CONFIG_ARM64_KERNELPACMASK; ulong physvirt_offset; + ulong struct_page_size; }; struct arm64_stackframe { -- 2.30.2