[PATCH] SCSI tape direct transfers for 2.5.31

This adds direct write and read from/to the user space to/from the SCSI
adapter.  If something (e.g., too high address) prevents direct
transfer, the existing driver buffer code is used.  The patch contains
functions to map and pin the user buffer pages so that this patch is not
dependent on kiobufs.  Similar functions are already used in sg.c.

drivers/scsi/README.st

@@ -2,7 +2,7 @@ This file contains brief information about the SCSI tape driver.
 The driver is currently maintained by Kai M{kisara (email
 Kai.Makisara@metla.fi)
 
-Last modified: Mon Jul 15 16:30:40 2002 by makisara
+Last modified: Fri Jul 26 16:01:39 2002 by makisara
 
 
 BASICS
@@ -105,22 +105,31 @@ The default is BSD semantics.
 
 BUFFERING
 
-The driver uses tape buffers allocated at run-time when needed and it
-is freed when the device file is closed. One buffer is used for each
-open tape device. 
-
-The size of the buffers is always at least one tape block. In fixed
+The driver tries to do tranfers directly to/from user space. If this
+is not possible, a driver buffer allocated at run-time is used. If
+direct i/o is not possible for the whole transfer, the driver buffer
+is used (i.e., bounce buffers for individual pages are not
+used). Direct i/o can be impossible because of several reasons, e.g.:
+- one or more pages are at addresses not reachable by the HBA
+- the number of pages in the transfer exceeds the number of
+  scatter/gather segments permitted by the HBA
+- one or more pages can't be locked into memory (should not happen in
+  any reasonable situation)
+
+The size of the driver buffers is always at least one tape block. In fixed
 block mode, the minimum buffer size is defined (in 1024 byte units) by
 ST_FIXED_BUFFER_BLOCKS. With small block size this allows buffering of
 several blocks and using one SCSI read or write to transfer all of the
 blocks. Buffering of data across write calls in fixed block mode is
-allowed if ST_BUFFER_WRITES is non-zero. Buffer allocation uses chunks of
-memory having sizes 2^n * (page size). Because of this the actual
-buffer size may be larger than the minimum allowable buffer size.
+allowed if ST_BUFFER_WRITES is non-zero and direct i/o is not used.
+Buffer allocation uses chunks of memory having sizes 2^n * (page
+size). Because of this the actual buffer size may be larger than the
+minimum allowable buffer size.
 
 Asynchronous writing. Writing the buffer contents to the tape is
 started and the write call returns immediately. The status is checked
-at the next tape operation.
+at the next tape operation. Asynchronous writes are not done with
+direct i/o.
 
 Buffered writes and asynchronous writes may in some rare cases cause
 problems in multivolume operations if there is not enough space on the
@@ -181,9 +190,10 @@ are configurable when the driver is loaded as a module. The keywords are:
 buffer_kbs=xxx             the buffer size for fixed block mode is set
 			   to xxx kilobytes
 write_threshold_kbs=xxx    the write threshold in kilobytes set to xxx
-max_buffers=xxx            the maximum number of tape buffer set to xxx
 max_sg_segs=xxx		   the maximum number of scatter/gather
 			   segments
+try_direct_io=x		   try direct transfer between user buffer and
+			   tape drive if this is non-zero
 
 Note that if the buffer size is changed but the write threshold is not
 set, the write threshold is set to the new buffer size - 2 kB.

drivers/scsi/st.h

@@ -12,6 +12,7 @@
 typedef struct {
 	unsigned char in_use;
 	unsigned char dma;	/* DMA-able buffer */
+	unsigned char do_dio;
 	int buffer_size;
 	int buffer_blocks;
 	int buffer_bytes;
@@ -21,7 +22,7 @@ typedef struct {
 	int syscall_result;
 	Scsi_Request *last_SRpnt;
 	unsigned char *b_data;
-	unsigned short use_sg;	/* zero or maximum number of s/g segments for this adapter */
+	unsigned short use_sg;	/* zero or max number of s/g segments for this adapter */
 	unsigned short sg_segs;		/* number of segments in s/g list */
 	unsigned short orig_frp_segs;	/* number of segments allocated at first try */
 	unsigned short frp_segs;	/* number of buffer segments */
@@ -91,12 +92,15 @@ typedef struct {
 	unsigned char cln_sense_value;
 	unsigned char cln_sense_mask;
 	unsigned char use_pf;			/* Set Page Format bit in all mode selects? */
+	unsigned char try_dio;			/* try direct i/o? */
 	unsigned char c_algo;			/* compression algorithm */
 	int tape_type;
 	int write_threshold;
 	int timeout;		/* timeout for normal commands */
 	int long_timeout;	/* timeout for commands known to take long time */
 
+	unsigned long max_pfn;	/* the maximum page number reachable by the HBA */
+
 	/* Mode characteristics */
 	ST_mode modes[ST_NBR_MODES];
 	int current_mode;
@@ -135,6 +139,10 @@ typedef struct {
 	unsigned char write_pending;
 	int nbr_finished;
 	int nbr_waits;
+	int nbr_requests;
+	int nbr_dio;
+	int nbr_pages;
+	int nbr_combinable;
 	unsigned char last_cmnd[6];
 	unsigned char last_sense[16];
 #endif

drivers/scsi/st_options.h

 /*
    The compile-time configurable defaults for the Linux SCSI tape driver.
 
-   Copyright 1995-2000 Kai Makisara.
+   Copyright 1995-2002 Kai Makisara.
 
-   Last modified: Sun May  5 15:09:56 2002 by makisara
+   Last modified: Fri Jul 26 15:54:31 2002 by makisara
 */
 
 #ifndef _ST_OPTIONS_H
 #define _ST_OPTIONS_H
 
+/* If TRY_DIRECT_IO is non-zero, the driver tries to transfer data directly
+   between the user buffer and tape drive. If this is not possible, driver
+   buffer is used. If TRY_DIRECT_IO is zero, driver buffer is always used. */
+#define TRY_DIRECT_IO 1
+
 /* The driver does not wait for some operations to finish before returning
    to the user program if ST_NOWAIT is non-zero. This helps if the SCSI
    adapter does not support multiple outstanding commands. However, the user
@@ -40,7 +45,7 @@
 #define ST_WRITE_THRESHOLD_BLOCKS 30
 
 /* Maximum number of scatter/gather segments */
-#define ST_MAX_SG      64
+#define ST_MAX_SG      256
 
 /* The number of scatter/gather segments to allocate at first try (must be
    smaller or equal to the maximum). */