Date:   Sun,  4 Dec 2022 09:30:39 +0800
From:   Baoquan He <bhe@...hat.com>
To:     linux-kernel@...r.kernel.org
Cc:     linux-mm@...ck.org, stephen.s.brennan@...cle.com, urezki@...il.com,
        willy@...radead.org, akpm@...ux-foundation.org, hch@...radead.org,
        Baoquan He <bhe@...hat.com>
Subject: [PATCH v1 0/7] mm/vmalloc.c: allow vread() to read out vm_map_ram areas

Problem:
***
Stephen reported vread() will skip vm_map_ram areas when reading out
/proc/kcore with drgn utility. Please see below link to get more about
it:

  /proc/kcore reads 0's for vmap_block
  https://lore.kernel.org/all/87ilk6gos2.fsf@oracle.com/T/#u

Root cause:
***
The normal vmalloc API uses struct vmap_area to manage the virtual
kernel area allocated and associate a vm_struct to store more information
and passed out. However, area reserved through vm_map_ram() interface
doesn't allocate vm_struct to bind with. So the current code in vread()
will skip the vm_map_ram area through 'if (!va->vm)' conditional checking.

Solution:
***
There are two types of vm_map_ram area. One is the whole vmap_area being
reserved and mapped at one time; the other is the whole vmap_area with
VMAP_BLOCK_SIZE size being reserved at one time, while mapped into split
regions with smaller size several times via vb_alloc(). I will call the
2nd type vb region.

In patch 1 and 2, add flags into struct vmap_area to mark these two types
of vm_map_ram area, meanwhile add bitmap field used_map into struct
vmap_block to mark those vb regions being used to differentiate with dirty
and free regions in vmap_block.

With the help of above vmap_area->flags and vmap_block->used_map, we can
recognize them in vread() and handle them respectively in patch 3.

Now since we can identify vm_map_ram area explicitly, change the
ambiguous checking 'if (!va->vm) with clear "if (!va->vm && (va->flags & VMAP_RAM))"
in s_show() to pick out vm_map_ram area clearly. This avoids the normal
vmalloc area being unmapped to be recognized as vm_map_ram area. This is
done in pach 4.

Besides,
***
In patch 5, let's ignore vmap area with VM_UNINITIALIZED set in
vm->flags, because this kind of area is created by calling
__vmalloc_node_range(), VM_UNINITIALIZED set indicating it has
vm_struct associated with, but is still during the page allocating and
mapping process.

In patch 6 and 7, change area flag from VM_ALLOC to VM_IOREMAP in two
places. This will show them as 'ioremap' in /proc/vmallocinfo, and
exclude them from /proc/kcore.

Testing
***
Stephen helped test RFC and the draft fix patch. Now I only put this v1
patchset on a bare metal to compile and run for basic functionality
testing. It still needs Stephen to help test the vm_map_ram issue.

Changelog
***
RFC->v1:
- Add a new field flags in vmap_area to mark vm_map_ram area. It cold be
  risky reusing the vm union in vmap_area in RFC. I will consider
  reusing the union in vmap_area to save memory later. Now just take the
  simpler way to let's focus on resolving the main problem.
- Add patch 4~7 for optimization.

Baoquan He (7):
  mm/vmalloc.c: add used_map into vmap_block to track space of
    vmap_block
  mm/vmalloc.c: add flags to mark vm_map_ram area
  mm/vmalloc.c: allow vread() to read out vm_map_ram areas
  mm/vmalloc: explicitly identify vm_map_ram area when shown in
    /proc/vmcoreinfo
  mm/vmalloc: skip the uninitilized vmalloc areas
  powerpc: mm: add VM_IOREMAP flag to the vmalloc area
  sh: mm: set VM_IOREMAP flag to the vmalloc area

 arch/powerpc/kernel/pci_64.c |  2 +-
 arch/sh/kernel/cpu/sh4/sq.c  |  2 +-
 include/linux/vmalloc.h      |  1 +
 mm/vmalloc.c                 | 97 +++++++++++++++++++++++++++++++-----
 4 files changed, 87 insertions(+), 15 deletions(-)

-- 
2.34.1

Powered by blists - more mailing lists