MDEV-34705: Binlog-in-engine: Implement page checksum

Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>

mysql-test/suite/binlog_in_engine/binlog_flush_purge.result

@@ -21,6 +21,12 @@ Log_name	File_size
 binlog-000000.ibb	49152
 binlog-000001.ibb	262144
 binlog-000002.ibb	262144
+SET STATEMENT sql_log_bin=0 FOR
+CALL mtr.add_suppression("InnoDB: Page corruption in binlog tablespace file page number 0");
+FLUSH BINARY LOGS;
+FLUSH BINARY LOGS;
+SHOW BINLOG EVENTS IN 'binlog-000000.ibb' LIMIT 1;
+ERROR HY000: Error when executing command SHOW BINLOG EVENTS: error reading event data
 RESET MASTER;
 SHOW BINARY LOGS;
 Log_name	File_size

mysql-test/suite/binlog_in_engine/binlog_flush_purge.test

@@ -26,6 +26,30 @@ FLUSH BINARY LOGS;
 --source include/wait_for_engine_binlog.inc
 SHOW BINARY LOGS;
 
+# Flush couple logs so we are sure the first file is on disk.
+# Corrupt one bit in the first page of the first file to test that crc32
+# mismatch is caught.
+SET STATEMENT sql_log_bin=0 FOR
+  CALL mtr.add_suppression("InnoDB: Page corruption in binlog tablespace file page number 0");
+FLUSH BINARY LOGS;
+FLUSH BINARY LOGS;
+--let $file= binlog-000000.ibb
+--let $datadir= `SELECT @@datadir`
+--let BINLOG_FILE= $datadir/$file
+perl;
+open F, '+<', $ENV{BINLOG_FILE} or die $!;
+sysseek F, 50, 0 or die $!;
+my $x;
+sysread F, $x, 1 or die $!;
+$x= chr(ord($x) ^ (1 <<3));
+sysseek F, 50, 0 or die $!;
+syswrite F, $x, 1 or die $!;
+EOF
+
+--error ER_ERROR_WHEN_EXECUTING_COMMAND
+eval SHOW BINLOG EVENTS IN '$file' LIMIT 1;
+
+
 RESET MASTER;
 --let $binlog_name= binlog-000001.ibb
 --let $binlog_size= 262144

storage/innobase/fsp/fsp_binlog.cc

@@ -253,9 +253,7 @@ fsp_binlog_page_fifo::flush_one_page(uint64_t file_no, bool force)
   {
     File fh= get_fh(file_no);
     ut_a(pl->fh >= (File)0);
-    size_t res= my_pwrite(fh, e->page_buf, srv_page_size,
-                          (uint64_t)page_no << srv_page_size_shift,
-                          MYF(MY_WME));
+    size_t res= crc32_pwrite_page(fh, e->page_buf, page_no, MYF(MY_WME));
     ut_a(res == srv_page_size);
     e->flushed_clean= true;
   }
@@ -576,6 +574,42 @@ fsp_binlog_page_fifo::flush_thread_run()
   mysql_mutex_unlock(&m_mutex);
 }
 
+
+size_t
+crc32_pwrite_page(File fd, byte *buf, uint32_t page_no, myf MyFlags) noexcept
+{
+  const uint32_t payload= (uint32_t)srv_page_size - BINLOG_PAGE_CHECKSUM;
+  mach_write_to_4(buf + payload, my_crc32c(0, buf, payload));
+  return my_pwrite(fd, (const uchar *)buf, srv_page_size,
+                   (my_off_t)page_no << srv_page_size_shift, MyFlags);
+}
+
+
+size_t
+crc32_pread_page(File fd, byte *buf, uint32_t page_no, myf MyFlags) noexcept
+{
+  size_t res= my_pread(fd, buf, srv_page_size,
+                       (my_off_t)page_no << srv_page_size_shift, MyFlags);
+  if (UNIV_LIKELY(res == srv_page_size))
+  {
+    const uint32_t payload= (uint32_t)srv_page_size - BINLOG_PAGE_CHECKSUM;
+    uint32_t crc32= mach_read_from_4(buf + payload);
+    /* Allow a completely zero (empty) page as well. */
+    if (UNIV_UNLIKELY(crc32 != my_crc32c(0, buf, payload)) &&
+        (buf[0] != 0 || 0 != memcmp(buf, buf+1, srv_page_size - 1)))
+    {
+      res= -1;
+      my_errno= EIO;
+      if (MyFlags & MY_WME)
+        sql_print_error("InnoDB: Page corruption in binlog tablespace file "
+                        "page number %u (invalid crc32 checksum 0x%08X)",
+                        page_no, crc32);
+    }
+  }
+  return res;
+}
+
+
 void
 binlog_write_up_to_now() noexcept
 {
@@ -1208,8 +1242,8 @@ binlog_chunk_reader::fetch_current_page()
       continue;
     }
 
-    size_t res= my_pread(cur_file_handle, page_buffer, srv_page_size,
-                         s.page_no << srv_page_size_shift, MYF(MY_WME));
+    size_t res= crc32_pread_page(cur_file_handle, page_buffer, s.page_no,
+                                 MYF(MY_WME));
     if (res == (size_t)-1)
       return CHUNK_READER_ERROR;
     if (res == 0 && my_errno == HA_ERR_FILE_TOO_SHORT)

storage/innobase/handler/innodb_binlog.cc

@@ -518,7 +518,7 @@ static int read_gtid_state_from_page(rpl_binlog_state_base *state,
 
   Returns:
    -1  error
-    0  File is missing (ENOENT)
+    0  File is missing (ENOENT) or has bad checksum on first page.
     1  File found (but may be empty according to out_empty).
 */
 int
@@ -542,6 +542,15 @@ binlog_recovery::get_header(uint64_t file_no, lsn_t &out_lsn, bool &out_empty)
     return -1;
   if (read == 0)
     return 0;
+  /*
+    If the crc32 does not match, the page was not written properly, so treat
+    it as an empty file.
+  */
+  const uint32_t payload= (uint32_t)srv_page_size - BINLOG_PAGE_CHECKSUM;
+  uint32_t crc32= mach_read_from_4(page_buf + payload);
+  if (UNIV_UNLIKELY(crc32 != my_crc32c(0, page_buf, payload)))
+    return 0;
+
   dummy_state.init();
   int res= read_gtid_state_from_page(&dummy_state, page_buf, 0, &header);
   if (res <= 0)
@@ -826,9 +835,8 @@ binlog_recovery::flush_page() noexcept
   if (cur_file_fh < (File)0 &&
       open_cur_file())
     return true;
-  size_t res= my_pwrite(cur_file_fh, page_buf, srv_page_size,
-                          (uint64_t)cur_page_no << srv_page_size_shift,
-                          MYF(MY_WME));
+  size_t res=
+    crc32_pwrite_page(cur_file_fh, page_buf, cur_page_no, MYF(MY_WME));
   if (res != srv_page_size)
     return true;
   cur_page_offset= 0;
@@ -844,7 +852,7 @@ binlog_recovery::zero_out_cur_file()
     return;
 
   /* Recover the original size from the current file. */
-  size_t read= my_pread(cur_file_fh, page_buf, srv_page_size, 0, MYF(0));
+  size_t read= crc32_pread_page(cur_file_fh, page_buf, 0, MYF(0));
   if (read != (size_t)srv_page_size)
   {
     sql_print_warning("InnoDB: Could not read last binlog file during recovery");
@@ -1900,9 +1908,8 @@ read_gtid_state(rpl_binlog_state_base *state, File file, uint32_t page_no,
   if (UNIV_UNLIKELY(!page_buf))
     return -1;
 
-  /* ToDo: Verify checksum, and handle encryption. */
-  size_t res= my_pread(file, page_buf.get(), srv_page_size,
-                       (uint64_t)page_no << srv_page_size_shift, MYF(MY_WME));
+  /* ToDo: Handle encryption. */
+  size_t res= crc32_pread_page(file, page_buf.get(), page_no, MYF(MY_WME));
   if (UNIV_UNLIKELY(res == (size_t)-1))
     return -1;
 

storage/innobase/include/fsp_binlog.h

@@ -34,6 +34,10 @@ InnoDB implementation of binlog.
 struct chunk_data_base;
 
 
+/** Store crc32 checksum at the end of the page */
+#define BINLOG_PAGE_CHECKSUM 4
+
+
 enum fsp_binlog_chunk_types {
   /* Zero means no data, effectively EOF. */
   FSP_BINLOG_TYPE_EMPTY= 0,
@@ -302,6 +306,10 @@ fsp_binlog_release(fsp_binlog_page_entry *page)
   binlog_page_fifo->release_page(page);
 }
 
+extern size_t crc32_pwrite_page(File fd, byte *buf, uint32_t page_no,
+                                myf MyFlags) noexcept;
+extern size_t crc32_pread_page(File fd, byte *buf, uint32_t page_no,
+                               myf MyFlags) noexcept;
 extern void binlog_write_up_to_now() noexcept;
 extern void fsp_log_binlog_write(mtr_t *mtr, fsp_binlog_page_entry *page,
                                  uint32_t page_offset, uint32_t len);