This bug was discovered by @simonmichal, and can be reproduced by applying the change below on the client, then trying to copy a file with xrdcp:

--- a/src/XrdCl/XrdClFileStateHandler.cc
+++ b/src/XrdCl/XrdClFileStateHandler.cc
@@ -1431,6 +1431,8 @@ namespace XrdCl
           }
         } );
 
+    cksums.front() = 0;
+
     auto st = PgWriteImpl( self, offset, size, buffer, cksums, 0, h, timeout );
     if( !st.IsOK() )
     {

The change above just sets the checksum of the first page to 0, but the server reports all pages having a bad checksum. After some debugging, we reached the conclusion that there is a bug in XrdOucPgrwUtils::csVer (shown below):

bool XrdOucPgrwUtils::csVer(dataInfo &dInfo, off_t &bado, int &badc)
{
   int pgOff = dInfo.offs & pgPageMask;

// Make sure we have something to do
//
   fprintf(stderr, "dInfo: %p, %ld %d\n", dInfo.data, dInfo.offs, dInfo.count);

   if (dInfo.count <= 0) return true;

// If this is unaligned, the we must verify the checksum of the leading bytes
// to align them to the next page boundary if one exists.
//
   if (pgOff)
      {off_t tempsave;
       int chkLen = pgPageSize - pgOff;
       if (dInfo.count < chkLen) {chkLen = dInfo.count; dInfo.count = 0;}
          else dInfo.count -= chkLen;

       bool aOK = XrdOucCRC::Ver32C((void *)dInfo.data, chkLen, dInfo.csval[0]);

       dInfo.data += chkLen;
       tempsave    = dInfo.offs;
       dInfo.offs += chkLen;
       dInfo.csval++;

       if (!aOK)
          {bado = tempsave;
           badc = chkLen;
           return false;
          }
      }

// Verify the remaining checksums, if any are left (offset is page aligned)
//
   if (dInfo.count > 0)
      {uint32_t valcs;
       int pgNum = XrdOucCRC::Ver32C((void *)dInfo.data,  dInfo.count,
                                             dInfo.csval, valcs);
       if (pgNum >= 0)
          {bado = dInfo.offs + (pgPageSize * pgNum);
           int xlen = (bado - dInfo.offs);
           dInfo.offs  += xlen;
           dInfo.count -= xlen;
           badc = (dInfo.count <= pgPageSize ? dInfo.count : pgPageSize);
           dInfo.offs  += badc;
           dInfo.count -= badc;
           dInfo.csval += (pgNum+1);
           return false;
          }
      }

// All sent well
//
   return true;
}

When if (pgNum >=0) is entered, dInfo.csval is updated, but the pointer dInfo.data remains the same, so subsequent pages are checked against the wrong checksum, which results in bad offsets being reported for all pages.

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you are subscribed to this thread.