lkml.org 
[lkml]   [2018]   [Jul]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH v2 1/4] arm64: export memblock_reserve()d regions via /proc/iomem
    On Tue, Jun 19, 2018 at 10:22:46AM -0500, Dave Kleikamp wrote:
    > On 06/19/2018 10:00 AM, James Morse wrote:
    > > Hi Dave,
    > >
    > > On 19/06/18 14:37, Dave Kleikamp wrote:
    > >> On 06/19/2018 01:44 AM, AKASHI Takahiro wrote:
    > >>> +static int __init reserve_memblock_reserved_regions(void)
    > >>> +{
    > >>> + phys_addr_t start, end, roundup_end = 0;
    > >>> + struct resource *mem, *res;
    > >>> + u64 i;
    > >>> +
    > >>> + for_each_reserved_mem_region(i, &start, &end) {
    > >>> + if (end <= roundup_end)
    > >>> + continue; /* done already */
    > >>> +
    > >>> + start = __pfn_to_phys(PFN_DOWN(start));
    > >>> + end = __pfn_to_phys(PFN_UP(end)) - 1;
    > >>> + roundup_end = end;
    > >>> +
    > >>> + res = kzalloc(sizeof(*res), GFP_ATOMIC);
    > >>> + if (WARN_ON(!res))
    > >>> + return -ENOMEM;
    > >>> + res->start = start;
    > >>> + res->end = end;
    > >>> + res->name = "reserved";
    > >>> + res->flags = IORESOURCE_MEM;
    > >>> +
    > >>> + mem = request_resource_conflict(&iomem_resource, res);
    > >>> + /*
    > >>> + * We expected memblock_reserve() regions to conflict with
    > >>> + * memory created by request_standard_resources().
    > >>> + */
    > >>> + if (WARN_ON_ONCE(!mem))
    > >>> + continue;
    > >>> + kfree(res);
    > >>
    > >> Why is kfree() after the conditional continue? This is a memory leak.
    > >
    > > request_resource_conflict() inserts res into the iomem_resource tree, or returns
    > > the conflict if there is already something at this address.
    > >
    > > We expect something at this address: a 'System RAM' section added by
    > > request_resource(). This code wants the conflicting 'System RAM' entry so that
    > > reserve_region_with_split() can fill in the gaps below it with 'reserved'. See
    > > the commit-message for an example.
    > >
    > > If there was no conflict, it means the memory map doesn't look like we expect,
    > > we can't pass NULL to reserve_region_with_split(), and we just inserted the
    > > 'reserved' at the top level. Freeing res at this point would be a use-after-free
    > > hanging from /proc/iomem.
    > > This code generates a WARN_ON_ONCE() and leaves the 'reserved' description where
    > > it is.
    >
    > Okay. I get it.
    >
    > > Trying to cleanup here is pointless, we can remove the resource entry and free
    > > it ... but we still want to report it as reserved, which is what just happened,
    > > just not quite how we we wanted it.
    > >
    > > If you ever see this warning, it means some RAM stopped being nomap between
    > > request_resources() and reserve_memblock_reserved_regions(). I can't find any
    > > case where that ever happens.
    > >
    > >
    > > If all that makes sense: how can I improve the comment above the WARN_ON_ONCE()
    > > to make it clearer?
    >
    > I guess something like you described above.
    >
    > /*
    > * We expect a 'System RAM' section at this address. In the unexpected
    > * case where one is not found, request_resource_conflict() will insert
    > * res into the iomem_resource tree.
    > */
    >
    > Do you think this is clearer?

    If this is the only change needed in my patchset, I'd like the maintainers
    to take care of it instead of my posting another version.

    Thanks,
    -Takahiro AKASHI

    > >
    > >
    > > Thanks,
    > >
    > > James
    > >
    > >
    > >>> +
    > >>> + reserve_region_with_split(mem, start, end, "reserved");
    > >>> + }
    > >>> +
    > >>> + return 0;
    > >>> +}
    > >>> +arch_initcall(reserve_memblock_reserved_regions);
    > >>> +
    > >>> u64 __cpu_logical_map[NR_CPUS] = { [0 ... NR_CPUS-1] = INVALID_HWID };
    > >>>
    > >>> void __init setup_arch(char **cmdline_p)
    > >>>
    > >

    \
     
     \ /
      Last update: 2018-07-03 08:46    [W:2.281 / U:0.108 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site