lkml.org 
[lkml]   [2009]   [Apr]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [RFC][PATCH] fix sign extension with 1.5TB usb-storage LBD=y
    Matthew Wilcox wrote:
    > On Tue, Apr 21, 2009 at 01:52:54PM -0700, Dave Hansen wrote:
    >> This is with current git as of this morning, which is at v2.6.30-rc2.
    >>
    >> I have a 1.5TB USB device which gets a bit angry when I plug it in. It
    >> ends up with a scsi_disk->capacity of ffffffffaea87b30. I tracked it
    >> down to the lba calculation in read_capacity_10():
    >>
    >> lba = (buffer[0] << 24) | (buffer[1] << 16) |
    >> (buffer[2] << 8) | buffer[3];
    >>
    >> lba is getting all 0xf's in its high 32 bits. It seems odd that this
    >> would happen since 'buffer' is an 'unsigned char', but that is
    >> apparently what is going on. Note that this isn't an issue 32-bit
    >> kernels compiled with CONFIG_LBD=n since there's no more bits into which
    >> the sign could be extended.
    >
    > I think I know ... unsigned char gets promoted to signed int since it will
    > fit. then signed int gets cast to unsigned long long, sign-extending. C's
    > promotion rules have always felt a bit wacky to me.
    >
    >> diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
    >> index 3fcb64b..db60e96 100644
    >> --- a/drivers/scsi/sd.c
    >> +++ b/drivers/scsi/sd.c
    >> @@ -1402,7 +1402,7 @@ static int read_capacity_10(struct scsi_disk *sdkp, struct scsi_device *sdp,
    >>
    >> sector_size = (buffer[4] << 24) | (buffer[5] << 16) |
    >> (buffer[6] << 8) | buffer[7];
    >> - lba = (buffer[0] << 24) | (buffer[1] << 16) |
    >> + lba = ((sector_t)buffer[0] << 24) | (buffer[1] << 16) |
    >> (buffer[2] << 8) | buffer[3];
    >
    > this certainly fixes your problem. I'd prefer this patch instead, just
    > because I find the cast unaesthetic ...
    >
    > ----
    >
    > Fix READ CAPACITY 10 with drives over 1TB
    >
    > Shifting an unsigned char implicitly casts it to a signed int. This
    > caused 'lba' to sign-extend and Linux would then try READ CAPACITY 16
    > which was not supported by at least one drive. Making 'lba' an unsigned
    > int ensures that sign extension will not occur.
    >
    > Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
    >
    > diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
    > index 3fcb64b..c856b1b 100644
    > --- a/drivers/scsi/sd.c
    > +++ b/drivers/scsi/sd.c
    > @@ -1373,7 +1373,7 @@ static int read_capacity_10(struct scsi_disk *sdkp, struct scsi_device *sdp,
    > int sense_valid = 0;
    > int the_result;
    > int retries = 3;
    > - sector_t lba;
    > + unsigned lba;

    sector_t is either unsigned long or u64, depending on CONFIG_LBD.
    Are you saying (implying) that the higher-order bits of it don't matter here?
    If so, I'd just like for that to be clear(er).


    > unsigned sector_size;
    >
    > do {
    > @@ -1413,7 +1413,7 @@ static int read_capacity_10(struct scsi_disk *sdkp, struct scsi_device *sdp,
    > return -EOVERFLOW;
    > }
    >
    > - sdkp->capacity = lba + 1;
    > + sdkp->capacity = (sector_t)lba + 1;
    > return sector_size;
    > }
    >
    >



    \
     
     \ /
      Last update: 2009-04-21 23:33    [W:0.029 / U:33.672 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site