static STATUS
DI_galloc(
DI_IO *f,
i4 n,
DI_OP *diop,
i4 *end_of_file,
CL_ERR_DESC *err_code)
{
STATUS big_status = OK, small_status =OK;
STATUS intern_status = OK;
register DI_SLAVE_CB *disl;
i4 last_page;
OFFSET_TYPE lseek_ret;
do
{
# ifdef OS_THREADS_USED
/* Seek/write must be semaphore protected */
if ((f->io_fprop & FPROP_PRIVATE) == 0)
CS_synch_lock( &f->io_sem );
# endif /* OS_THREADS_USED */
if (Di_slave)
{
disl = diop->di_evcb;
disl->file_op = DI_SL_ZALLOC;
disl->length = n * f->io_bytes_per_page;
/* Pass file properties to slave */
FPROP_COPY(f->io_fprop,disl->io_fprop);
DI_slave_send( disl->dest_slave_no, diop,
&big_status, &small_status, &intern_status );
if (( big_status != OK ) || ( small_status != OK ))
break;
if ( disl->status != OK )
{
STRUCT_ASSIGN_MACRO(disl->errcode, *err_code);
small_status = DI_BADEXTEND;
break;
}
else
{
lseek_ret = disl->length;
}
}
else
{
/*
** Running without slaves
*/
OFFSET_TYPE lseek_offset;
i8 reservation;
i4 buf_size;
i4 bytes_written;
i4 pages_remaining = n;
i4 pages_at_a_time = Di_zero_bufsize /
f->io_bytes_per_page;
/* find current end-of-file */
lseek_ret = IIdio_get_file_eof(diop->di_fd, f->io_fprop);
if ( lseek_ret == (OFFSET_TYPE)-1L )
{
SETCLERR(err_code, 0, ER_lseek);
small_status = DI_BADINFO;
break;
}
else
{
lseek_offset = lseek_ret;
/* If this filesystem can do reservations, see if we
** should reserve more space.
** Even though we have to write the zeros anyway, the
** reservation may well be larger than the zeroing
** buffer, and this way helps maintain contiguity.
** Not worth it for tiny writes.
*/
if (pages_remaining > 2
&& FPROP_ALLOCSTRATEGY_GET(f->io_fprop) == FPROP_ALLOCSTRATEGY_RESV)
{
reservation = lseek_offset + (pages_remaining * f->io_bytes_per_page);
if (reservation > f->io_reserved_bytes)
{
/* Re-check in case some other server reserved */
small_status = IIdio_get_reserved(diop->di_fd,
&f->io_reserved_bytes, err_code);
if (small_status == OK && reservation > f->io_reserved_bytes)
{
small_status = IIdio_reserve(diop->di_fd,
f->io_reserved_bytes,
reservation - f->io_reserved_bytes,
err_code);
if (small_status == OK)
{
f->io_reserved_bytes = reservation;
}
else
{
if (small_status != DI_BADFILE)
break;
/* Fallocate not supported, turn off
** "reserve" strategy, continue without.
*/
small_status = OK;
FPROP_ALLOCSTRATEGY_SET(f->io_fprop, FPROP_ALLOCSTRATEGY_VIRT);
}
}
}
} /* end reservations */
while ( pages_remaining > 0 )
{
if ( pages_remaining < pages_at_a_time )
buf_size = pages_remaining *
f->io_bytes_per_page;
else
buf_size = Di_zero_bufsize;
# if defined(OS_THREADS_USED) && !defined(xCL_NO_ATOMIC_READ_WRITE_IO)
bytes_written =
#ifdef LARGEFILE64
pwrite64( diop->di_fd, Di_zero_buffer,
buf_size, lseek_offset );
#else /* LARGEFILE64 */
pwrite( diop->di_fd, Di_zero_buffer,
buf_size, lseek_offset );
#endif /* LARGEFILE64 */
# else /* OS_THREADS_USED !xCL_NO_ATOMIC_READ_WRITE_IO */
bytes_written =
IIdio_write( diop->di_fd, Di_zero_buffer,
buf_size,
lseek_offset,
&lseek_offset,
f->io_fprop,
err_code );
# endif /* OS_THREADS_USED */
if ( bytes_written != buf_size )
{
SETCLERR(err_code, 0, ER_write);
small_status = DI_BADEXTEND;
break;
}
lseek_offset += buf_size;
pages_remaining -= pages_at_a_time;
}
if ( small_status != OK )
break;
}
}
*end_of_file = ( lseek_ret / f->io_bytes_per_page) - 1;
} while (FALSE);
if (big_status == OK && small_status == OK)
{
/*
** Update the current allocated end-of-file under mutex protection
*/
last_page = *end_of_file + n;
if (last_page > f->io_alloc_eof)
f->io_alloc_eof = last_page;
}
# ifdef OS_THREADS_USED
if ((f->io_fprop & FPROP_PRIVATE) == 0)
CS_synch_unlock( &f->io_sem );
# endif /* OS_THREADS_USED */
if ( big_status != OK )
small_status = big_status;
if ( small_status != OK )
DIlru_set_di_error( &small_status, err_code, intern_status,
DI_GENERAL_ERR);
return(small_status);
}
/*{
** Name: DI_inproc_write - writes page(s) to a file on disk.
**
** Description:
** This routine was created to make DIwrite more readable once
** error checking had been added. See DIwrite for comments.
**
** Inputs:
** f Pointer to the DI file
** context needed to do I/O.
** diop Pointer to dilru file context.
** buf Pointer to page(s) to write.
** page Value indicating page(s) to write.
** num_of_pages number of pages to write
**
** Outputs:
** err_code Pointer to a variable used
** to return operating system
** errors.
** Returns:
** OK
** other errors.
** Exceptions:
** none
**
** Side Effects:
** none
**
** History:
** 30-nov-1992 (rmuth)
** Created.
** 03-jun-1996 (canor01)
** Note in the scb that this is a DI wait.
** 05-May-1997 (merja01)
** Changed preprocessor stmt for pwrite. Not all platforms
** using OS_THREADS have a pwrite function. This function
** seems to only be available on Solaris 2.4 where async IO
** is not yet supported.
** 14-July-1997 (schte01)
** For those platforms that do direct i/o (where the
** seek and the write are separate functions), do not release and
** reaquire the semaphore on the DI_IO block. This will protect
** against i/o being done by a different thread in between the
** lseek and the write.
** 14-Aug-1997 (schte01)
** Add xCL_DIRECT_IO as a condition to the 14-July-1997 change
** instead of the test for !xCL_ASYNCH_IO.
** 22-Dec-1998 (jenjo02)
** If DI_FD_PER_THREAD is defined, call IIdio_write() instead of
** pwrite().
** 01-oct-1998 (somsa01)
** Return DI_NODISKSPACE when we are out of disk space.
** 01-Apr-2004 (fanch01)
** Add O_DIRECT support on Linux depending on the filesystem
** properties, pagesize. Fixups for misaligned buffers on read()
** and write() operations.
** 13-apr-04 (toumi01)
** Move stack variable declaration to support "standard" C compilers.
** 29-Jan-2005 (schka24)
** Ditch attempt to gather diow timing stats, not useful in
** the real world and generates excess syscalls on some platforms.
** 15-Mar-2006 (jenjo02)
** io_sem is not needed with thread affinity.
** 6-Nov-2009 (kschendel) SIR 122757
** Make io-sem a SYNCH, avoid entirely if PRIVATE.
** Delete copy-to-align, caller is supposed to do it now.
** Don't attempt SCB updating if not backend.
*/
static STATUS
DI_inproc_write(
DI_IO *f,
DI_OP *diop,
char *buf,
i4 page,
i4 num_of_pages,
CL_ERR_DESC *err_code )
{
STATUS status = OK;
CS_SCB *scb;
i4 saved_state;
/* unix variables */
int bytes_written;
int bytes_to_write;
OFFSET_TYPE lseek_offset;
/*
** seek to place to write
*/
lseek_offset =
(OFFSET_TYPE)(f->io_bytes_per_page) * (OFFSET_TYPE)page;
bytes_to_write = (f->io_bytes_per_page * (num_of_pages));
if (Di_backend)
{
CSget_scb(&scb);
if ( scb )
{
saved_state = scb->cs_state;
scb->cs_state = CS_EVENT_WAIT;
if (f->io_open_flags & DI_O_LOG_FILE_MASK)
{
scb->cs_memory = CS_LIOW_MASK;
scb->cs_liow++;
Cs_srv_block.cs_wtstatistics.cs_liow_done++;
Cs_srv_block.cs_wtstatistics.cs_liow_waits++;
Cs_srv_block.cs_wtstatistics.cs_liow_kbytes
+= bytes_to_write / 1024;
}
else
{
scb->cs_memory = CS_DIOW_MASK;
scb->cs_diow++;
Cs_srv_block.cs_wtstatistics.cs_diow_done++;
Cs_srv_block.cs_wtstatistics.cs_diow_waits++;
Cs_srv_block.cs_wtstatistics.cs_diow_kbytes
+= bytes_to_write / 1024;
}
}
}
# if defined(OS_THREADS_USED) && defined(xCL_NO_ATOMIC_READ_WRITE_IO)
if ( !Di_thread_affinity && (f->io_fprop & FPROP_PRIVATE) == 0)
CS_synch_lock( &f->io_sem );
# endif /* OS_THREADS_USED && !xCL_NO_ATOMIC_READ_WRITE_IO */
# if defined(OS_THREADS_USED) && !defined(xCL_NO_ATOMIC_READ_WRITE_IO)
bytes_written =
#ifdef LARGEFILE64
pwrite64( (int)diop->di_fd, buf, bytes_to_write, lseek_offset );
#else /* LARGEFILE64 */
pwrite( (int)diop->di_fd, buf, bytes_to_write, lseek_offset );
#endif /* LARGEFILE64 */
# else /* OS_THREADS_USED !xCL_NO_ATOMIC_READ_WRITE_IO */
bytes_written =
IIdio_write( (int)diop->di_fd, buf, bytes_to_write,
lseek_offset, 0,
f->io_fprop,
err_code );
# endif /* OS_THREADS_USED */
if ( bytes_written != bytes_to_write )
{
SETCLERR(err_code, 0, ER_write);
switch( err_code->errnum )
{
case EFBIG:
status = DI_BADEXTEND;
break;
case ENOSPC:
status = DI_NODISKSPACE;
break;
#ifdef EDQUOTA
case EDQUOT:
status = DI_EXCEED_LIMIT;
break;
#endif
default:
if (err_code->errnum == 0)
status = DI_ENDFILE;
else
status = DI_BADWRITE;
break;
}
}
# if defined(OS_THREADS_USED) && defined(xCL_NO_ATOMIC_READ_WRITE_IO)
if ( !Di_thread_affinity && (f->io_fprop & FPROP_PRIVATE) == 0)
CS_synch_unlock( &f->io_sem );
# endif /* OS_THREADS_USED && xCL_NO_ATOMIC_READ_WRITE_IO */
if ( Di_backend && scb )
{
scb->cs_memory &= ~(CS_DIOW_MASK | CS_LIOW_MASK);
scb->cs_state = saved_state;
}
return( status );
}
/*{
** Name: do_writev - Perform writev() call.
**
** Description:
** This function collects the queued write requests,
** chooses the optimum function to perform the write(s),
** and invokes the completion handler for each request.
**
** Inputs:
** DI_TGIO * tgio - Control block for current thread.
**
** Outputs:
** None.
**
** Returns:
** OK
** FAIL - One of more of the write requests failed.
**
** Exceptions:
** none
**
** Side Effects:
** The completion handler for each I/O request is invoked.
**
** History:
** 19-May-1999 (jenjo02)
** Created.
** 09-Jul-1999 (jenjo02)
** If queued list is ordered, skip the quicksort.
** 09-Apr-2001 (jenjo02)
** Increment first gio's io_count stat for each physical I/O,
** gw_pages for multi-page writes.
** 05-Nov-2002 (jenjo02)
** Cleaned up use of io_sem: only write() and writev() need
** the mutex to protect the (separate) seek. pwrite(64)
** atomically seeks and does not need the mutex.
** 25-Aug-2005 (schka24)
** Don't bother with IO timing, too slow on some platforms (Linux)
** and the results aren't all that interesting.
** 14-Oct-2005 (jenjo02)
** Chris's file descriptor properties now cached in io_fprop
** (file properties) and established on the first open,
** not every open.
** 24-Jan-2006 (jenjo02)
** Break on change in file ("f"), then lru-open to get an
** FD, do the write(v), and lru_release the FD. This keeps
** gather_write from hogging FDs while waiting for the
** signal to actually do something.
** 15-Mar-2006 (jenjo02)
** f->io_sem not needed if running with thread affinity,
** the fd is not shared by multiple threads.
*/
static STATUS
do_writev( DI_TGIO * tgio, CL_ERR_DESC *err_code )
{
CL_ERR_DESC lerr_code;
STATUS big_status = OK, small_status = OK;
i4 i, j, k;
DI_GIO *gio, *first_gio;
DI_IO *f;
DI_OP *diop;
OFFSET_TYPE next_offset, lseek_offset;
i4 bytes_to_write, bytes_written;
i4 saved_state;
i4 num_writev = 0, num_write = 0;
i4 num_writev_gio = 0, num_write_gio = 0;
#if defined(sgi_us5)
if( iov_max == 0 )
{
iov_max = sysconf(_SC_IOV_MAX);
if( iov_max <= 0 )
{
iov_max = 16; /* arbitrary minimum value */
#ifdef DEBUG_THIS_PUPPY
TRdisplay("%@ %x do_writev: %t ERROR sysconf failed with %d\n",
tgio->tgio_scb->cs_self,
f->io_l_filename, f->io_filename,
iov_max);
#endif /* DEBUG_THIS_PUPPY */
}
else if( iov_max > 2048 )
{
iov_max = 2048; /* arbitrary maximum value */
}
}
#else
iov_max = IOV_MAX;
#endif
/* If unordered, sort the queued list into file,offset order */
if ( tgio->tgio_state & TGIO_UNORDERED )
{
gio_sort( tgio->tgio_queue, 0, tgio->tgio_queued-1 );
tgio->tgio_state &= ~(TGIO_UNORDERED);
}
/*
** Method:
**
** Collect requests by file/offset into an iovec until
** the next file offset becomes discontiguous. Additionally, if
** the buffer addresses are contiguous, colaesce those requests.
**
** Up to IOV_MAX iovecs can be written by a single writev().
**
** If but a single iovec results, the probably-more-efficient
** function (p)write() is called instead of writev().
*/
k = 0;
while ( (j = k) < tgio->tgio_queued )
{
#if defined(sgi_us5)
struct iovec iov[iov_max];
#else
struct iovec iov[IOV_MAX];
#endif
/*
** "i" indexes the current iovec element
** "j" is the first GIO used in this iovec array
** "k" is the current GIO in the queue
*/
i = 0;
gio = first_gio = tgio->tgio_queue[j];
f = gio->gio_f;
lseek_offset = next_offset = gio->gio_offset;
small_status = OK;
iov[0].iov_base = gio->gio_buf;
iov[0].iov_len = 0;
do
{
/* If this buffer is contiguous with previous, coalesce it */
if ( (char *)iov[i].iov_base + iov[i].iov_len == gio->gio_buf )
{
iov[i].iov_len += gio->gio_len;
}
/* Initialize next iovec if any remain */
else if ( i < iov_max - 1 )
{
i++;
iov[i].iov_base = gio->gio_buf;
iov[i].iov_len = gio->gio_len;
}
else
break;
next_offset += gio->gio_len;
} while ( ++k < tgio->tgio_queued
&& (gio = tgio->tgio_queue[k])
&& gio->gio_f == f
&& gio->gio_offset == next_offset );
/* "k" indexes the next, unprocessed GIO */
bytes_to_write = next_offset - lseek_offset;
saved_state = tgio->tgio_scb->cs_state;
tgio->tgio_scb->cs_state = CS_EVENT_WAIT;
tgio->tgio_scb->cs_memory = CS_DIOW_MASK;
/* Accumulate multi-page write stats */
if ( k - j > 1 )
{
/*
** Using the first gio, count
** the number of multi-pages written (k-j)
** and a single I/O.
*/
if ( first_gio->gio_io_count )
++*first_gio->gio_io_count;
if ( first_gio->gio_gw_pages )
*first_gio->gio_gw_pages += k - j;
}
/* Count a single I/O write for server */
tgio->tgio_scb->cs_diow++;
Cs_srv_block.cs_wtstatistics.cs_diow_done++;
/* Count a single I/O wait for server */
Cs_srv_block.cs_wtstatistics.cs_diow_waits++;
/* Accumulate number of KB written by this I/O */
Cs_srv_block.cs_wtstatistics.cs_diow_kbytes
+= bytes_to_write / 1024;
/* Now get an FD to do the write(v) */
diop = (DI_OP*)&first_gio->gio_diop;
if ( big_status = DIlru_open(f, FALSE, diop, err_code) )
return(big_status);
#ifdef DEBUG_THIS_PUPPY
{
i4 x;
i8 offset = lseek_offset;
TRdisplay("%@ %p do_writev: %~t doing %d todo %d fd %d lseek from %ld\n",
tgio->tgio_scb->cs_self,
f->io_l_filename, f->io_filename,
i+1, tgio->tgio_queued - j,
diop->di_fd, offset);
for (x = 0; x <= i; x++)
{
TRdisplay("%@ do_writev: iovec[%d] base %p bytes %d (page %d for %d)\n",
x,
iov[x].iov_base, iov[x].iov_len,
(i4)(offset/f->io_bytes_per_page),
iov[x].iov_len/f->io_bytes_per_page);
offset += iov[x].iov_len;
}
}
#endif /* DEBUG_THIS_PUPPY */
/* If more than one iovec, seek and use writev */
if ( i++ )
{
/* writev needs seek mutex protection */
if ( !Di_thread_affinity )
CS_synch_lock( &f->io_sem );
num_writev++;
num_writev_gio += k - j;
bytes_written =
IIdio_writev( diop->di_fd,
(char *)iov,
i,
lseek_offset, 0,
f->io_fprop,
err_code);
if ( !Di_thread_affinity )
CS_synch_unlock( &f->io_sem );
}
else
{
num_write++;
num_write_gio += k - j;
# if !defined(xCL_NO_ATOMIC_READ_WRITE_IO)
/* pwrite(64) needs no seek mutex protection */
bytes_written =
#ifdef LARGEFILE64
pwrite64( diop->di_fd,
iov[0].iov_base,
bytes_to_write,
lseek_offset );
#else /* LARGEFILE64 */
pwrite( diop->di_fd,
iov[0].iov_base,
bytes_to_write,
lseek_offset );
#endif /* LARGEFILE64 */
if (bytes_written != bytes_to_write)
SETCLERR(err_code, 0, ER_write);
# else /* !xCL_NO_ATOMIC_READ_WRITE_IO */
/* write() needs seek mutex protection */
if ( !Di_thread_affinity )
CS_synch_lock( &f->io_sem );
bytes_written =
IIdio_write( diop->di_fd,
iov[0].iov_base,
bytes_to_write,
lseek_offset, 0,
f->io_fprop,
err_code );
if ( !Di_thread_affinity )
CS_synch_unlock( &f->io_sem );
# endif /* !xCL_NO_ATOMIC_READ_WRITE_IO */
}
/* Release the FD */
(VOID)DIlru_release( diop, &lerr_code );
tgio->tgio_scb->cs_memory &= ~(CS_DIOW_MASK);
tgio->tgio_scb->cs_state = saved_state;
if (bytes_written != bytes_to_write)
{
switch ( err_code->errnum )
{
case EFBIG:
small_status = DI_BADEXTEND;
break;
case ENOSPC:
small_status = DI_EXCEED_LIMIT;
break;
#ifdef EDQUOTA
case EDQUOT:
small_status = DI_EXCEED_LIMIT;
break;
#endif
default:
if (err_code->errnum == 0)
small_status = DI_ENDFILE;
else
small_status = DI_BADWRITE;
break;
}
/* Preserve the worst status from all the writes */
big_status = (big_status) ? big_status : small_status;
}
/* Invoke completion handler for each GIO written */
do
{
gio = tgio->tgio_queue[j];
(gio->gio_evcomp)( gio->gio_data, small_status, err_code );
} while ( ++j < k );
}
#ifdef DEBUG_THIS_PUPPY
TRdisplay("%@ %p do_writev: %d write requests completed using %d(%d) writev, %d(%d) write\n",
tgio->tgio_scb->cs_self,
tgio->tgio_queued,
num_writev, num_writev_gio,
num_write, num_write_gio);
#endif /* DEBUG_THIS_PUPPY */
/* Clear the queued count(s) */
tgio->tgio_queued = *tgio->tgio_uqueued = 0;
return( big_status );
}