[Crash-utility] [PATCH] ARM64 support for 3-level page tables with 64K pages
Dave Anderson
anderson at redhat.com
Thu Jun 16 20:14:11 UTC 2016
Hi Jim,
The problem looks to be the use of PAGEOFFSET() into the 512-byte pgd used
for 3-level/64K-pages:
static int
arm64_vtop_3level_64k(ulong pgd, ulong vaddr, physaddr_t *paddr, int verbose)
{
ulong *pgd_base, *pgd_ptr, pgd_val;
ulong *pmd_base, *pmd_ptr, pmd_val;
ulong *pte_base, *pte_ptr, pte_val;
if (verbose)
fprintf(fp, "PAGE DIRECTORY: %lx\n", pgd);
pgd_base = (ulong *)pgd;
FILL_PGD(pgd_base, KVADDR, PTRS_PER_PGD_L3_64K * sizeof(ulong));
pgd_ptr = pgd_base + (((vaddr) >> PGDIR_SHIFT_L3_64K) & (PTRS_PER_PGD_L3_64K - 1));
pgd_val = ULONG(machdep->pgd + PAGEOFFSET(pgd_ptr));
^^^^^^^^^^^^^^^^^^
For kernel vmalloc addresses, the kernel's "swapper_pg_dir" is used, which has a base
address that's aligned on a 64k page boundary. So PAGEOFFSET() works as an index.
But for user-space addresses, the mm_struct->pgd is not 64K page aligned so the
use of PAGEOFFSET() calculates an address that's goes way beyond the end of the
512-byte aligned pgd:
crash> for user vm -p | grep -e PGD -e "k "
MM PGD RSS TOTAL_VM
ffff8003d61ec240 ffff8003d5b12c00 11904k 12992k
MM PGD RSS TOTAL_VM
ffff8003d60aec40 ffff8003dcb7ca00 4864k 88448k
MM PGD RSS TOTAL_VM
ffff8000dbf1dc80 ffff8003d69bca00 8704k 92928k
MM PGD RSS TOTAL_VM
ffff8000dbf1e700 ffff8003d69b4a00 8768k 17536k
MM PGD RSS TOTAL_VM
ffff8003d65c9300 ffff8000d4055600 5696k 16576k
MM PGD RSS TOTAL_VM
ffff8003d65c9300 ffff8000d4055600 5696k 16576k
MM PGD RSS TOTAL_VM
ffff8003d65c9840 ffff8000d4051800 1664k 2560k
MM PGD RSS TOTAL_VM
ffff8003d60a8340 ffff8003dcb7fa00 10944k 303552k
MM PGD RSS TOTAL_VM
ffff8003d65c8880 ffff8000d4052c00 19456k 353600k
MM PGD RSS TOTAL_VM
ffff8003d61ab7c0 ffff8003dc86b800 4864k 14016k
MM PGD RSS TOTAL_VM
ffff8000dbf18340 ffff8003d69b1400 7424k 157632k
MM PGD RSS TOTAL_VM
ffff8003d60a8340 ffff8003dcb7fa00 10944k 303552k
MM PGD RSS TOTAL_VM
ffff8003d60a8340 ffff8003dcb7fa00 10944k 303552k
MM PGD RSS TOTAL_VM
ffff8000dbf18340 ffff8003d69b1400 7424k 157632k
MM PGD RSS TOTAL_VM
ffff8000dbf18340 ffff8003d69b1400 7424k 157632k
MM PGD RSS TOTAL_VM
ffff8000dbf18340 ffff8003d69b1400 7424k 157632k
MM PGD RSS TOTAL_VM
ffff8000dbf18340 ffff8003d69b1400 7424k 157632k
MM PGD RSS TOTAL_VM
ffff8000dbf18340 ffff8003d69b1400 7424k 157632k
MM PGD RSS TOTAL_VM
ffff8003d61ab7c0 ffff8003dc86b800 4864k 14016k
MM PGD RSS TOTAL_VM
ffff8000c403b7c0 ffff8003d6597800 4928k 5888k
MM PGD RSS TOTAL_VM
ffff8003d59db7c0 ffff8003dcd18600 4352k 5248k
MM PGD RSS TOTAL_VM
ffff8003d61ead40 ffff8003d4107400 3776k 5760k
MM PGD RSS TOTAL_VM
ffff8000c403ad40 ffff8003d6592000 4224k 5120k
MM PGD RSS TOTAL_VM
ffff8003d60a78c0 ffff8003dcb72600 2496k 109888k
MM PGD RSS TOTAL_VM
ffff8003d61a8880 ffff8003dc86b200 2176k 109888k
MM PGD RSS TOTAL_VM
ffff8003d65c8880 ffff8000d4052c00 19456k 353600k
MM PGD RSS TOTAL_VM
ffff8003d65c8880 ffff8000d4052c00 19456k 353600k
...
Dave
----- Original Message -----
>
> ----- Original Message -----
> > > Adds ARM64 support for 3-level page tables with 64K pages and 48 VA bits.
> >
> > Nicely done, Jim. Queued for crash-7.1.5:
> >
> > https://github.com/crash-utility/crash/commit/ab91852f945bfecfa0bca6a42253fbecb38723db
> >
> > Thanks,
> > Dave
> >
>
> Hi Jim,
>
> I just noticed today that your 3-level 64K patch does not work for user
> virtual address space.
> I haven't looked too deeply into it, but for example on a live system, all
> user virtual address
> vtop operations fail, all disk-backed user memory space shows the "FILE:"
> backing, and the
> anonymous space shows "(not mapped)":
>
> crash> help -m | grep VM
> flags: 10400069
> (KSYMS_START|VM_L3_64K|VMEMMAP|KDUMP_ENABLED|IRQ_STACKS|MACHDEP_BT_TEXT)
> crash> sys | grep RELEASE
> RELEASE: 4.5.0-0.38.el7.aarch64
> crash> set
> PID: 1212
> COMMAND: "crash"
> TASK: ffff8003d74f3f00 [THREAD_INFO: ffff8003d7454000]
> CPU: 1
> STATE: TASK_RUNNING (ACTIVE)
> crash> vm -p
> PID: 1212 TASK: ffff8003d74f3f00 CPU: 1 COMMAND: "crash"
> MM PGD RSS TOTAL_VM
> ffff8000c40363c0 ffff8003db6a9200 211904k 355264k
> VMA START END FLAGS FILE
> ffff8003de746d40 400000 a00000 875 /root/crash.git/crash
> VIRTUAL PHYSICAL
> 400000 FILE: /root/crash.git/crash OFFSET: 0
> 410000 FILE: /root/crash.git/crash OFFSET: 10000
> 420000 FILE: /root/crash.git/crash OFFSET: 20000
> 430000 FILE: /root/crash.git/crash OFFSET: 30000
> 440000 FILE: /root/crash.git/crash OFFSET: 40000
> 450000 FILE: /root/crash.git/crash OFFSET: 50000
> ... [ cut ] ...
> VMA START END FLAGS FILE
> ffff8003de745d70 a50000 b00000 100073
> VIRTUAL PHYSICAL
> a50000 (not mapped)
> a60000 (not mapped)
> a70000 (not mapped)
> a80000 (not mapped)
> a90000 (not mapped)
> aa0000 (not mapped)
> ab0000 (not mapped)
> ac0000 (not mapped)
> ...
>
> In all cases, the PGD value reads as 0 and therefore fails:
>
> crash> vtop 400000
> VIRTUAL PHYSICAL
> 400000 (not mapped)
>
> PAGE DIRECTORY: ffff8003db6a9200
> PGD: ffff8003db6a9200 => 0
>
> VMA START END FLAGS FILE
> ffff8003de746d40 400000 a00000 875 /root/crash.git/crash
>
> FILE: /root/crash.git/crash OFFSET: 0
>
> crash>
>
> That is the correct PGD address, and when read, it looks like a valid PTE:
>
> crash> rd ffff8003db6a9200
> ffff8003db6a9200: 00000043dee60003 ....C...
> crash>
>
> vmalloc() addresses translate just fine, and since they use the same
> function,
> I'm not sure what's going on? Did you ever check user-space translations?
>
> Thanks,
> Dave
>
>
>
> >
> >
> > > ---
> > > arm64.c | 126
> > > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++--------
> > > defs.h | 28 +++++++++++----
> > > 2 files changed, 133 insertions(+), 21 deletions(-)
> > >
> > > diff --git a/arm64.c b/arm64.c
> > > index f6ea7a1..d1c9c3e 100644
> > > --- a/arm64.c
> > > +++ b/arm64.c
> > > @@ -34,6 +34,7 @@ static void arm64_init_kernel_pgd(void);
> > > static int arm64_kvtop(struct task_context *, ulong, physaddr_t *, int);
> > > static int arm64_uvtop(struct task_context *, ulong, physaddr_t *, int);
> > > static int arm64_vtop_2level_64k(ulong, ulong, physaddr_t *, int);
> > > +static int arm64_vtop_3level_64k(ulong, ulong, physaddr_t *, int);
> > > static int arm64_vtop_3level_4k(ulong, ulong, physaddr_t *, int);
> > > static ulong arm64_get_task_pgd(ulong);
> > > static void arm64_irq_stack_init(void);
> > > @@ -188,15 +189,29 @@ arm64_init(int when)
> > > break;
> > >
> > > case 65536:
> > > - machdep->flags |= VM_L2_64K;
> > > - machdep->ptrs_per_pgd = PTRS_PER_PGD_L2_64K;
> > > - if ((machdep->pgd =
> > > - (char *)malloc(PTRS_PER_PGD_L2_64K * 8)) == NULL)
> > > - error(FATAL, "cannot malloc pgd space.");
> > > - if ((machdep->ptbl =
> > > - (char *)malloc(PTRS_PER_PTE_L2_64K * 8)) == NULL)
> > > - error(FATAL, "cannot malloc ptbl space.");
> > > - machdep->pmd = NULL; /* not used */
> > > + if (machdep->machspec->VA_BITS > PGDIR_SHIFT_L3_64K) {
> > > + machdep->flags |= VM_L3_64K;
> > > + machdep->ptrs_per_pgd = PTRS_PER_PGD_L3_64K;
> > > + if ((machdep->pgd =
> > > + (char *)malloc(PTRS_PER_PGD_L3_64K * 8)) == NULL)
> > > + error(FATAL, "cannot malloc pgd space.");
> > > + if ((machdep->pmd =
> > > + (char *)malloc(PTRS_PER_PMD_L3_64K * 8)) == NULL)
> > > + error(FATAL, "cannot malloc pmd space.");
> > > + if ((machdep->ptbl =
> > > + (char *)malloc(PTRS_PER_PTE_L3_64K * 8)) == NULL)
> > > + error(FATAL, "cannot malloc ptbl space.");
> > > + } else {
> > > + machdep->flags |= VM_L2_64K;
> > > + machdep->ptrs_per_pgd = PTRS_PER_PGD_L2_64K;
> > > + if ((machdep->pgd =
> > > + (char *)malloc(PTRS_PER_PGD_L2_64K * 8)) == NULL)
> > > + error(FATAL, "cannot malloc pgd space.");
> > > + if ((machdep->ptbl =
> > > + (char *)malloc(PTRS_PER_PTE_L2_64K * 8)) == NULL)
> > > + error(FATAL, "cannot malloc ptbl space.");
> > > + machdep->pmd = NULL; /* not used */
> > > + }
> > > machdep->pud = NULL; /* not used */
> > > break;
> > >
> > > @@ -379,6 +394,8 @@ arm64_dump_machdep_table(ulong arg)
> > > fprintf(fp, "%sPHYS_OFFSET", others++ ? "|" : "");
> > > if (machdep->flags & VM_L2_64K)
> > > fprintf(fp, "%sVM_L2_64K", others++ ? "|" : "");
> > > + if (machdep->flags & VM_L3_64K)
> > > + fprintf(fp, "%sVM_L3_64K", others++ ? "|" : "");
> > > if (machdep->flags & VM_L3_4K)
> > > fprintf(fp, "%sVM_L3_4K", others++ ? "|" : "");
> > > if (machdep->flags & VMEMMAP)
> > > @@ -410,10 +427,14 @@ arm64_dump_machdep_table(ulong arg)
> > > fprintf(fp, " processor_speed: arm64_processor_speed()\n");
> > > fprintf(fp, " uvtop: arm64_uvtop()->%s()\n",
> > > machdep->flags & VM_L3_4K ?
> > > - "arm64_vtop_3level_4k" : "arm64_vtop_2level_64k");
> > > + "arm64_vtop_3level_4k" :
> > > + machdep->flags & VM_L3_64K ?
> > > + "arm64_vtop_3level_64k" : "arm64_vtop_2level_64k");
> > > fprintf(fp, " kvtop: arm64_kvtop()->%s()\n",
> > > machdep->flags & VM_L3_4K ?
> > > - "arm64_vtop_3level_4k" : "arm64_vtop_2level_64k");
> > > + "arm64_vtop_3level_4k" :
> > > + machdep->flags & VM_L3_64K ?
> > > + "arm64_vtop_3level_64k" : "arm64_vtop_2level_64k");
> > > fprintf(fp, " get_task_pgd: arm64_get_task_pgd()\n");
> > > fprintf(fp, " dump_irq: generic_dump_irq()\n");
> > > fprintf(fp, " get_stack_frame: arm64_get_stack_frame()\n");
> > > @@ -719,10 +740,12 @@ arm64_kvtop(struct task_context *tc, ulong kvaddr,
> > > physaddr_t *paddr, int verbos
> > > kernel_pgd = vt->kernel_pgd[0];
> > > *paddr = 0;
> > >
> > > - switch (machdep->flags & (VM_L2_64K|VM_L3_4K))
> > > + switch (machdep->flags & (VM_L2_64K|VM_L3_64K|VM_L3_4K))
> > > {
> > > case VM_L2_64K:
> > > return arm64_vtop_2level_64k(kernel_pgd, kvaddr, paddr, verbose);
> > > + case VM_L3_64K:
> > > + return arm64_vtop_3level_64k(kernel_pgd, kvaddr, paddr, verbose);
> > > case VM_L3_4K:
> > > return arm64_vtop_3level_4k(kernel_pgd, kvaddr, paddr, verbose);
> > > default:
> > > @@ -740,10 +763,12 @@ arm64_uvtop(struct task_context *tc, ulong uvaddr,
> > > physaddr_t *paddr, int verbos
> > >
> > > *paddr = 0;
> > >
> > > - switch (machdep->flags & (VM_L2_64K|VM_L3_4K))
> > > + switch (machdep->flags & (VM_L2_64K|VM_L3_64K|VM_L3_4K))
> > > {
> > > case VM_L2_64K:
> > > return arm64_vtop_2level_64k(user_pgd, uvaddr, paddr, verbose);
> > > + case VM_L3_64K:
> > > + return arm64_vtop_3level_64k(user_pgd, uvaddr, paddr, verbose);
> > > case VM_L3_4K:
> > > return arm64_vtop_3level_4k(user_pgd, uvaddr, paddr, verbose);
> > > default:
> > > @@ -820,6 +845,78 @@ no_page:
> > > return FALSE;
> > > }
> > >
> > > +static int
> > > +arm64_vtop_3level_64k(ulong pgd, ulong vaddr, physaddr_t *paddr, int
> > > verbose)
> > > +{
> > > + ulong *pgd_base, *pgd_ptr, pgd_val;
> > > + ulong *pmd_base, *pmd_ptr, pmd_val;
> > > + ulong *pte_base, *pte_ptr, pte_val;
> > > +
> > > + if (verbose)
> > > + fprintf(fp, "PAGE DIRECTORY: %lx\n", pgd);
> > > +
> > > + pgd_base = (ulong *)pgd;
> > > + FILL_PGD(pgd_base, KVADDR, PTRS_PER_PGD_L3_64K * sizeof(ulong));
> > > + pgd_ptr = pgd_base + (((vaddr) >> PGDIR_SHIFT_L3_64K) &
> > > (PTRS_PER_PGD_L3_64K - 1));
> > > + pgd_val = ULONG(machdep->pgd + PAGEOFFSET(pgd_ptr));
> > > + if (verbose)
> > > + fprintf(fp, " PGD: %lx => %lx\n", (ulong)pgd_ptr,
> > > pgd_val);
> > > + if (!pgd_val)
> > > + goto no_page;
> > > +
> > > + /*
> > > + * #define __PAGETABLE_PUD_FOLDED
> > > + */
> > > +
> > > + pmd_base = (ulong *)PTOV(pgd_val & PHYS_MASK & (s32)machdep->pagemask);
> > > + FILL_PMD(pmd_base, KVADDR, PTRS_PER_PMD_L3_64K * sizeof(ulong));
> > > + pmd_ptr = pmd_base + (((vaddr) >> PMD_SHIFT_L3_64K) &
> > > (PTRS_PER_PMD_L3_64K
> > > - 1));
> > > + pmd_val = ULONG(machdep->pmd + PAGEOFFSET(pmd_ptr));
> > > + if (verbose)
> > > + fprintf(fp, " PMD: %lx => %lx\n", (ulong)pmd_ptr,
> > > pmd_val);
> > > + if (!pmd_val)
> > > + goto no_page;
> > > +
> > > + if ((pmd_val & PMD_TYPE_MASK) == PMD_TYPE_SECT) {
> > > + ulong sectionbase = (pmd_val & SECTION_PAGE_MASK_512MB) & PHYS_MASK;
> > > + if (verbose) {
> > > + fprintf(fp, " PAGE: %lx (512MB)\n\n", sectionbase);
> > > + arm64_translate_pte(pmd_val, 0, 0);
> > > + }
> > > + *paddr = sectionbase + (vaddr & ~SECTION_PAGE_MASK_512MB);
> > > + return TRUE;
> > > + }
> > > +
> > > + pte_base = (ulong *)PTOV(pmd_val & PHYS_MASK & (s32)machdep->pagemask);
> > > + FILL_PTBL(pte_base, KVADDR, PTRS_PER_PTE_L3_64K * sizeof(ulong));
> > > + pte_ptr = pte_base + (((vaddr) >> machdep->pageshift) &
> > > (PTRS_PER_PTE_L3_64K - 1));
> > > + pte_val = ULONG(machdep->ptbl + PAGEOFFSET(pte_ptr));
> > > + if (verbose)
> > > + fprintf(fp, " PTE: %lx => %lx\n", (ulong)pte_ptr,
> > > pte_val);
> > > + if (!pte_val)
> > > + goto no_page;
> > > +
> > > + if (pte_val & PTE_VALID) {
> > > + *paddr = (PAGEBASE(pte_val) & PHYS_MASK) + PAGEOFFSET(vaddr);
> > > + if (verbose) {
> > > + fprintf(fp, " PAGE: %lx\n\n", PAGEBASE(*paddr));
> > > + arm64_translate_pte(pte_val, 0, 0);
> > > + }
> > > + } else {
> > > + if (IS_UVADDR(vaddr, NULL))
> > > + *paddr = pte_val;
> > > + if (verbose) {
> > > + fprintf(fp, "\n");
> > > + arm64_translate_pte(pte_val, 0, 0);
> > > + }
> > > + goto no_page;
> > > + }
> > > +
> > > + return TRUE;
> > > +no_page:
> > > + return FALSE;
> > > +}
> > > +
> > > static int
> > > arm64_vtop_3level_4k(ulong pgd, ulong vaddr, physaddr_t *paddr, int
> > > verbose)
> > > {
> > > @@ -2348,9 +2445,10 @@ arm64_calc_virtual_memory_ranges(void)
> > >
> > > STRUCT_SIZE_INIT(page, "page");
> > >
> > > - switch (machdep->flags & (VM_L2_64K|VM_L3_4K))
> > > + switch (machdep->flags & (VM_L2_64K|VM_L3_64K|VM_L3_4K))
> > > {
> > > case VM_L2_64K:
> > > + case VM_L3_64K:
> > > PUD_SIZE = PGDIR_SIZE_L2_64K;
> > > break;
> > > case VM_L3_4K:
> > > diff --git a/defs.h b/defs.h
> > > index 56ae06c..d1b49d0 100644
> > > --- a/defs.h
> > > +++ b/defs.h
> > > @@ -2815,7 +2815,7 @@ typedef u64 pte_t;
> > >
> > > typedef signed int s32;
> > >
> > > -/*
> > > +/*
> > > * 3-levels / 4K pages
> > > */
> > > #define PTRS_PER_PGD_L3_4K (512)
> > > @@ -2823,10 +2823,23 @@ typedef signed int s32;
> > > #define PTRS_PER_PTE_L3_4K (512)
> > > #define PGDIR_SHIFT_L3_4K (30)
> > > #define PGDIR_SIZE_L3_4K ((1UL) << PGDIR_SHIFT_L3_4K)
> > > -#define PGDIR_MASK_L3 4K (~(PGDIR_SIZE_L3_4K-1))
> > > +#define PGDIR_MASK_L3_4K (~(PGDIR_SIZE_L3_4K-1))
> > > #define PMD_SHIFT_L3_4K (21)
> > > -#define PMD_SIZE_L3_4K (1UL << PMD_SHIFT_4K)
> > > -#define PMD_MASK_L3 4K (~(PMD_SIZE_4K-1))
> > > +#define PMD_SIZE_L3_4K (1UL << PMD_SHIFT_L3_4K)
> > > +#define PMD_MASK_L3_4K (~(PMD_SIZE_L3_4K-1))
> > > +
> > > +/*
> > > + * 3-levels / 64K pages
> > > + */
> > > +#define PTRS_PER_PGD_L3_64K (64)
> > > +#define PTRS_PER_PMD_L3_64K (8192)
> > > +#define PTRS_PER_PTE_L3_64K (8192)
> > > +#define PGDIR_SHIFT_L3_64K (42)
> > > +#define PGDIR_SIZE_L3_64K ((1UL) << PGDIR_SHIFT_L3_64K)
> > > +#define PGDIR_MASK_L3_64K (~(PGDIR_SIZE_L3_64K-1))
> > > +#define PMD_SHIFT_L3_64K (29)
> > > +#define PMD_SIZE_L3_64K (1UL << PMD_SHIFT_L3_64K)
> > > +#define PMD_MASK_L3_64K (~(PMD_SIZE_L3_64K-1))
> > >
> > > /*
> > > * 2-levels / 64K pages
> > > @@ -2868,9 +2881,10 @@ typedef signed int s32;
> > > #define KSYMS_START (0x1)
> > > #define PHYS_OFFSET (0x2)
> > > #define VM_L2_64K (0x4)
> > > -#define VM_L3_4K (0x8)
> > > -#define KDUMP_ENABLED (0x10)
> > > -#define IRQ_STACKS (0x20)
> > > +#define VM_L3_64K (0x8)
> > > +#define VM_L3_4K (0x10)
> > > +#define KDUMP_ENABLED (0x20)
> > > +#define IRQ_STACKS (0x40)
> > >
> > > /*
> > > * sources: Documentation/arm64/memory.txt
> > > --
> > > 2.1.4
> > >
> > > --
> > > Crash-utility mailing list
> > > Crash-utility redhat com
> > > https://www.redhat.com/mailman/listinfo/crash-utility
> > >
More information about the Crash-utility
mailing list