freebsd-dev/sys/dev/isp/isp_pci.c

2031 lines
53 KiB
C
Raw Normal View History

1999-08-28 01:08:13 +00:00
/* $FreeBSD$ */
/*
* PCI specific probe and attach routines for Qlogic ISP SCSI adapters.
* FreeBSD Version.
*
* Copyright (c) 1997, 1998, 1999, 2000, 2001 by Matthew Jacob
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice immediately at the beginning of the file, without modification,
* this list of conditions, and the following disclaimer.
2000-09-21 20:16:04 +00:00
* 2. The name of the author may not be used to endorse or promote products
* derived from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
* ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/module.h>
#include <sys/bus.h>
#include <pci/pcireg.h>
#include <pci/pcivar.h>
#include <machine/bus_memio.h>
#include <machine/bus_pio.h>
#include <machine/bus.h>
#include <machine/resource.h>
#include <sys/rman.h>
#include <sys/malloc.h>
#include <dev/isp/isp_freebsd.h>
static u_int16_t isp_pci_rd_reg(struct ispsoftc *, int);
static void isp_pci_wr_reg(struct ispsoftc *, int, u_int16_t);
static u_int16_t isp_pci_rd_reg_1080(struct ispsoftc *, int);
static void isp_pci_wr_reg_1080(struct ispsoftc *, int, u_int16_t);
static int
isp_pci_rd_isr(struct ispsoftc *, u_int16_t *, u_int16_t *, u_int16_t *);
static int
isp_pci_rd_isr_2300(struct ispsoftc *, u_int16_t *, u_int16_t *, u_int16_t *);
static int isp_pci_mbxdma(struct ispsoftc *);
static int
isp_pci_dmasetup(struct ispsoftc *, XS_T *, ispreq_t *, u_int16_t *, u_int16_t);
static void
isp_pci_dmateardown(struct ispsoftc *, XS_T *, u_int16_t);
static void isp_pci_reset1(struct ispsoftc *);
static void isp_pci_dumpregs(struct ispsoftc *, const char *);
#ifndef ISP_CODE_ORG
#define ISP_CODE_ORG 0x1000
#endif
static struct ispmdvec mdvec = {
isp_pci_rd_isr,
isp_pci_rd_reg,
isp_pci_wr_reg,
isp_pci_mbxdma,
isp_pci_dmasetup,
isp_pci_dmateardown,
NULL,
isp_pci_reset1,
isp_pci_dumpregs,
NULL,
BIU_BURST_ENABLE|BIU_PCI_CONF1_FIFO_64
};
static struct ispmdvec mdvec_1080 = {
isp_pci_rd_isr,
isp_pci_rd_reg_1080,
isp_pci_wr_reg_1080,
isp_pci_mbxdma,
isp_pci_dmasetup,
isp_pci_dmateardown,
NULL,
isp_pci_reset1,
isp_pci_dumpregs,
NULL,
BIU_BURST_ENABLE|BIU_PCI_CONF1_FIFO_64
};
static struct ispmdvec mdvec_12160 = {
isp_pci_rd_isr,
isp_pci_rd_reg_1080,
isp_pci_wr_reg_1080,
isp_pci_mbxdma,
isp_pci_dmasetup,
isp_pci_dmateardown,
NULL,
isp_pci_reset1,
isp_pci_dumpregs,
NULL,
BIU_BURST_ENABLE|BIU_PCI_CONF1_FIFO_64
};
static struct ispmdvec mdvec_2100 = {
isp_pci_rd_isr,
isp_pci_rd_reg,
isp_pci_wr_reg,
isp_pci_mbxdma,
isp_pci_dmasetup,
isp_pci_dmateardown,
NULL,
isp_pci_reset1,
isp_pci_dumpregs
};
static struct ispmdvec mdvec_2200 = {
isp_pci_rd_isr,
isp_pci_rd_reg,
isp_pci_wr_reg,
isp_pci_mbxdma,
isp_pci_dmasetup,
isp_pci_dmateardown,
NULL,
isp_pci_reset1,
isp_pci_dumpregs
};
static struct ispmdvec mdvec_2300 = {
isp_pci_rd_isr_2300,
isp_pci_rd_reg,
isp_pci_wr_reg,
isp_pci_mbxdma,
isp_pci_dmasetup,
isp_pci_dmateardown,
NULL,
isp_pci_reset1,
isp_pci_dumpregs
};
#ifndef PCIM_CMD_INVEN
#define PCIM_CMD_INVEN 0x10
#endif
#ifndef PCIM_CMD_BUSMASTEREN
#define PCIM_CMD_BUSMASTEREN 0x0004
#endif
#ifndef PCIM_CMD_PERRESPEN
#define PCIM_CMD_PERRESPEN 0x0040
#endif
#ifndef PCIM_CMD_SEREN
#define PCIM_CMD_SEREN 0x0100
#endif
#ifndef PCIR_COMMAND
#define PCIR_COMMAND 0x04
#endif
#ifndef PCIR_CACHELNSZ
#define PCIR_CACHELNSZ 0x0c
#endif
#ifndef PCIR_LATTIMER
#define PCIR_LATTIMER 0x0d
#endif
#ifndef PCIR_ROMADDR
#define PCIR_ROMADDR 0x30
#endif
#ifndef PCI_VENDOR_QLOGIC
#define PCI_VENDOR_QLOGIC 0x1077
#endif
#ifndef PCI_PRODUCT_QLOGIC_ISP1020
#define PCI_PRODUCT_QLOGIC_ISP1020 0x1020
#endif
#ifndef PCI_PRODUCT_QLOGIC_ISP1080
#define PCI_PRODUCT_QLOGIC_ISP1080 0x1080
#endif
#ifndef PCI_PRODUCT_QLOGIC_ISP12160
#define PCI_PRODUCT_QLOGIC_ISP12160 0x1216
#endif
#ifndef PCI_PRODUCT_QLOGIC_ISP1240
#define PCI_PRODUCT_QLOGIC_ISP1240 0x1240
#endif
1999-12-16 05:42:02 +00:00
#ifndef PCI_PRODUCT_QLOGIC_ISP1280
#define PCI_PRODUCT_QLOGIC_ISP1280 0x1280
#endif
#ifndef PCI_PRODUCT_QLOGIC_ISP2100
#define PCI_PRODUCT_QLOGIC_ISP2100 0x2100
#endif
#ifndef PCI_PRODUCT_QLOGIC_ISP2200
#define PCI_PRODUCT_QLOGIC_ISP2200 0x2200
#endif
#ifndef PCI_PRODUCT_QLOGIC_ISP2300
#define PCI_PRODUCT_QLOGIC_ISP2300 0x2300
#endif
#ifndef PCI_PRODUCT_QLOGIC_ISP2312
#define PCI_PRODUCT_QLOGIC_ISP2312 0x2312
#endif
#define PCI_QLOGIC_ISP1020 \
((PCI_PRODUCT_QLOGIC_ISP1020 << 16) | PCI_VENDOR_QLOGIC)
#define PCI_QLOGIC_ISP1080 \
((PCI_PRODUCT_QLOGIC_ISP1080 << 16) | PCI_VENDOR_QLOGIC)
#define PCI_QLOGIC_ISP12160 \
((PCI_PRODUCT_QLOGIC_ISP12160 << 16) | PCI_VENDOR_QLOGIC)
#define PCI_QLOGIC_ISP1240 \
((PCI_PRODUCT_QLOGIC_ISP1240 << 16) | PCI_VENDOR_QLOGIC)
1999-12-16 05:42:02 +00:00
#define PCI_QLOGIC_ISP1280 \
((PCI_PRODUCT_QLOGIC_ISP1280 << 16) | PCI_VENDOR_QLOGIC)
#define PCI_QLOGIC_ISP2100 \
((PCI_PRODUCT_QLOGIC_ISP2100 << 16) | PCI_VENDOR_QLOGIC)
#define PCI_QLOGIC_ISP2200 \
((PCI_PRODUCT_QLOGIC_ISP2200 << 16) | PCI_VENDOR_QLOGIC)
#define PCI_QLOGIC_ISP2300 \
((PCI_PRODUCT_QLOGIC_ISP2300 << 16) | PCI_VENDOR_QLOGIC)
#define PCI_QLOGIC_ISP2312 \
((PCI_PRODUCT_QLOGIC_ISP2312 << 16) | PCI_VENDOR_QLOGIC)
/*
* Odd case for some AMI raid cards... We need to *not* attach to this.
*/
#define AMI_RAID_SUBVENDOR_ID 0x101e
#define IO_MAP_REG 0x10
#define MEM_MAP_REG 0x14
#define PCI_DFLT_LTNCY 0x40
#define PCI_DFLT_LNSZ 0x10
static int isp_pci_probe (device_t);
static int isp_pci_attach (device_t);
struct isp_pcisoftc {
struct ispsoftc pci_isp;
device_t pci_dev;
struct resource * pci_reg;
bus_space_tag_t pci_st;
bus_space_handle_t pci_sh;
void * ih;
int16_t pci_poff[_NREG_BLKS];
bus_dma_tag_t dmat;
bus_dmamap_t *dmaps;
};
ispfwfunc *isp_get_firmware_p = NULL;
static device_method_t isp_pci_methods[] = {
/* Device interface */
DEVMETHOD(device_probe, isp_pci_probe),
DEVMETHOD(device_attach, isp_pci_attach),
{ 0, 0 }
};
static void isp_pci_intr(void *);
static driver_t isp_pci_driver = {
"isp", isp_pci_methods, sizeof (struct isp_pcisoftc)
};
static devclass_t isp_devclass;
DRIVER_MODULE(isp, pci, isp_pci_driver, isp_devclass, 0, 0);
MODULE_VERSION(isp, 1);
static int
isp_pci_probe(device_t dev)
{
switch ((pci_get_device(dev) << 16) | (pci_get_vendor(dev))) {
case PCI_QLOGIC_ISP1020:
device_set_desc(dev, "Qlogic ISP 1020/1040 PCI SCSI Adapter");
break;
case PCI_QLOGIC_ISP1080:
device_set_desc(dev, "Qlogic ISP 1080 PCI SCSI Adapter");
break;
case PCI_QLOGIC_ISP1240:
device_set_desc(dev, "Qlogic ISP 1240 PCI SCSI Adapter");
break;
1999-12-16 05:42:02 +00:00
case PCI_QLOGIC_ISP1280:
device_set_desc(dev, "Qlogic ISP 1280 PCI SCSI Adapter");
break;
case PCI_QLOGIC_ISP12160:
if (pci_get_subvendor(dev) == AMI_RAID_SUBVENDOR_ID) {
return (ENXIO);
}
device_set_desc(dev, "Qlogic ISP 12160 PCI SCSI Adapter");
1999-12-16 05:42:02 +00:00
break;
case PCI_QLOGIC_ISP2100:
device_set_desc(dev, "Qlogic ISP 2100 PCI FC-AL Adapter");
break;
case PCI_QLOGIC_ISP2200:
device_set_desc(dev, "Qlogic ISP 2200 PCI FC-AL Adapter");
break;
case PCI_QLOGIC_ISP2300:
device_set_desc(dev, "Qlogic ISP 2300 PCI FC-AL Adapter");
break;
case PCI_QLOGIC_ISP2312:
device_set_desc(dev, "Qlogic ISP 2312 PCI FC-AL Adapter");
break;
default:
return (ENXIO);
}
if (device_get_unit(dev) == 0 && bootverbose) {
printf("Qlogic ISP Driver, FreeBSD Version %d.%d, "
"Core Version %d.%d\n",
ISP_PLATFORM_VERSION_MAJOR, ISP_PLATFORM_VERSION_MINOR,
ISP_CORE_VERSION_MAJOR, ISP_CORE_VERSION_MINOR);
}
/*
* XXXX: Here is where we might load the f/w module
* XXXX: (or increase a reference count to it).
*/
return (0);
}
static int
isp_pci_attach(device_t dev)
{
struct resource *regs, *irq;
int tval, rtp, rgd, iqd, m1, m2, isp_debug, role;
u_int32_t data, cmd, linesz, psize, basetype;
struct isp_pcisoftc *pcs;
struct ispsoftc *isp = NULL;
struct ispmdvec *mdvp;
const char *sptr;
int locksetup = 0;
/*
* Figure out if we're supposed to skip this one.
* If we are, we actually go to ISP_ROLE_NONE.
*/
tval = 0;
if (resource_int_value(device_get_name(dev), device_get_unit(dev),
"disable", &tval) == 0 && tval) {
device_printf(dev, "device is disabled\n");
/* but return 0 so the !$)$)*!$*) unit isn't reused */
return (0);
}
role = 0;
if (resource_int_value(device_get_name(dev), device_get_unit(dev),
"role", &role) == 0 &&
((role & ~(ISP_ROLE_INITIATOR|ISP_ROLE_TARGET)) == 0)) {
device_printf(dev, "setting role to 0x%x\n", role);
} else {
#ifdef ISP_TARGET_MODE
role = ISP_ROLE_INITIATOR|ISP_ROLE_TARGET;
#else
role = ISP_DEFAULT_ROLES;
#endif
}
pcs = malloc(sizeof (struct isp_pcisoftc), M_DEVBUF, M_NOWAIT | M_ZERO);
if (pcs == NULL) {
device_printf(dev, "cannot allocate softc\n");
return (ENOMEM);
}
/*
* Figure out which we should try first - memory mapping or i/o mapping?
*/
#ifdef __alpha__
m1 = PCIM_CMD_MEMEN;
m2 = PCIM_CMD_PORTEN;
#else
m1 = PCIM_CMD_PORTEN;
m2 = PCIM_CMD_MEMEN;
#endif
tval = 0;
if (resource_int_value(device_get_name(dev), device_get_unit(dev),
"prefer_iomap", &tval) == 0 && tval != 0) {
m1 = PCIM_CMD_PORTEN;
m2 = PCIM_CMD_MEMEN;
}
tval = 0;
if (resource_int_value(device_get_name(dev), device_get_unit(dev),
"prefer_memmap", &tval) == 0 && tval != 0) {
m1 = PCIM_CMD_MEMEN;
m2 = PCIM_CMD_PORTEN;
}
linesz = PCI_DFLT_LNSZ;
irq = regs = NULL;
rgd = rtp = iqd = 0;
cmd = pci_read_config(dev, PCIR_COMMAND, 1);
if (cmd & m1) {
rtp = (m1 == PCIM_CMD_MEMEN)? SYS_RES_MEMORY : SYS_RES_IOPORT;
rgd = (m1 == PCIM_CMD_MEMEN)? MEM_MAP_REG : IO_MAP_REG;
regs = bus_alloc_resource(dev, rtp, &rgd, 0, ~0, 1, RF_ACTIVE);
}
if (regs == NULL && (cmd & m2)) {
rtp = (m2 == PCIM_CMD_MEMEN)? SYS_RES_MEMORY : SYS_RES_IOPORT;
rgd = (m2 == PCIM_CMD_MEMEN)? MEM_MAP_REG : IO_MAP_REG;
regs = bus_alloc_resource(dev, rtp, &rgd, 0, ~0, 1, RF_ACTIVE);
}
if (regs == NULL) {
device_printf(dev, "unable to map any ports\n");
goto bad;
}
if (bootverbose)
device_printf(dev, "using %s space register mapping\n",
(rgd == IO_MAP_REG)? "I/O" : "Memory");
pcs->pci_dev = dev;
pcs->pci_reg = regs;
pcs->pci_st = rman_get_bustag(regs);
pcs->pci_sh = rman_get_bushandle(regs);
pcs->pci_poff[BIU_BLOCK >> _BLK_REG_SHFT] = BIU_REGS_OFF;
pcs->pci_poff[MBOX_BLOCK >> _BLK_REG_SHFT] = PCI_MBOX_REGS_OFF;
pcs->pci_poff[SXP_BLOCK >> _BLK_REG_SHFT] = PCI_SXP_REGS_OFF;
pcs->pci_poff[RISC_BLOCK >> _BLK_REG_SHFT] = PCI_RISC_REGS_OFF;
pcs->pci_poff[DMA_BLOCK >> _BLK_REG_SHFT] = DMA_REGS_OFF;
mdvp = &mdvec;
basetype = ISP_HA_SCSI_UNKNOWN;
psize = sizeof (sdparam);
if (pci_get_devid(dev) == PCI_QLOGIC_ISP1020) {
mdvp = &mdvec;
basetype = ISP_HA_SCSI_UNKNOWN;
psize = sizeof (sdparam);
}
if (pci_get_devid(dev) == PCI_QLOGIC_ISP1080) {
mdvp = &mdvec_1080;
basetype = ISP_HA_SCSI_1080;
psize = sizeof (sdparam);
pcs->pci_poff[DMA_BLOCK >> _BLK_REG_SHFT] =
ISP1080_DMA_REGS_OFF;
}
if (pci_get_devid(dev) == PCI_QLOGIC_ISP1240) {
mdvp = &mdvec_1080;
1999-12-16 05:42:02 +00:00
basetype = ISP_HA_SCSI_1240;
psize = 2 * sizeof (sdparam);
pcs->pci_poff[DMA_BLOCK >> _BLK_REG_SHFT] =
ISP1080_DMA_REGS_OFF;
}
if (pci_get_devid(dev) == PCI_QLOGIC_ISP1280) {
1999-12-16 05:42:02 +00:00
mdvp = &mdvec_1080;
basetype = ISP_HA_SCSI_1280;
psize = 2 * sizeof (sdparam);
pcs->pci_poff[DMA_BLOCK >> _BLK_REG_SHFT] =
ISP1080_DMA_REGS_OFF;
}
if (pci_get_devid(dev) == PCI_QLOGIC_ISP12160) {
mdvp = &mdvec_12160;
basetype = ISP_HA_SCSI_12160;
psize = 2 * sizeof (sdparam);
pcs->pci_poff[DMA_BLOCK >> _BLK_REG_SHFT] =
ISP1080_DMA_REGS_OFF;
}
if (pci_get_devid(dev) == PCI_QLOGIC_ISP2100) {
mdvp = &mdvec_2100;
basetype = ISP_HA_FC_2100;
psize = sizeof (fcparam);
pcs->pci_poff[MBOX_BLOCK >> _BLK_REG_SHFT] =
PCI_MBOX_REGS2100_OFF;
if (pci_get_revid(dev) < 3) {
/*
* XXX: Need to get the actual revision
* XXX: number of the 2100 FB. At any rate,
* XXX: lower cache line size for early revision
* XXX; boards.
*/
linesz = 1;
}
}
if (pci_get_devid(dev) == PCI_QLOGIC_ISP2200) {
mdvp = &mdvec_2200;
basetype = ISP_HA_FC_2200;
psize = sizeof (fcparam);
pcs->pci_poff[MBOX_BLOCK >> _BLK_REG_SHFT] =
PCI_MBOX_REGS2100_OFF;
}
if (pci_get_devid(dev) == PCI_QLOGIC_ISP2300) {
mdvp = &mdvec_2300;
basetype = ISP_HA_FC_2300;
psize = sizeof (fcparam);
pcs->pci_poff[MBOX_BLOCK >> _BLK_REG_SHFT] =
PCI_MBOX_REGS2300_OFF;
}
if (pci_get_devid(dev) == PCI_QLOGIC_ISP2312) {
mdvp = &mdvec_2300;
basetype = ISP_HA_FC_2312;
psize = sizeof (fcparam);
pcs->pci_poff[MBOX_BLOCK >> _BLK_REG_SHFT] =
PCI_MBOX_REGS2300_OFF;
}
isp = &pcs->pci_isp;
isp->isp_param = malloc(psize, M_DEVBUF, M_NOWAIT | M_ZERO);
if (isp->isp_param == NULL) {
device_printf(dev, "cannot allocate parameter data\n");
goto bad;
}
isp->isp_mdvec = mdvp;
isp->isp_type = basetype;
isp->isp_revision = pci_get_revid(dev);
isp->isp_role = role;
isp->isp_dev = dev;
/*
* Try and find firmware for this device.
*/
if (isp_get_firmware_p) {
int device = (int) pci_get_device(dev);
#ifdef ISP_TARGET_MODE
(*isp_get_firmware_p)(0, 1, device, &mdvp->dv_ispfw);
#else
(*isp_get_firmware_p)(0, 0, device, &mdvp->dv_ispfw);
#endif
}
/*
* Make sure that SERR, PERR, WRITE INVALIDATE and BUSMASTER
* are set.
*/
cmd |= PCIM_CMD_SEREN | PCIM_CMD_PERRESPEN |
PCIM_CMD_BUSMASTEREN | PCIM_CMD_INVEN;
if (IS_2300(isp)) { /* per QLogic errata */
cmd &= ~PCIM_CMD_INVEN;
}
pci_write_config(dev, PCIR_COMMAND, cmd, 1);
/*
* Make sure the Cache Line Size register is set sensibly.
*/
data = pci_read_config(dev, PCIR_CACHELNSZ, 1);
if (data != linesz) {
data = PCI_DFLT_LNSZ;
isp_prt(isp, ISP_LOGCONFIG, "set PCI line size to %d", data);
pci_write_config(dev, PCIR_CACHELNSZ, data, 1);
}
/*
* Make sure the Latency Timer is sane.
*/
data = pci_read_config(dev, PCIR_LATTIMER, 1);
if (data < PCI_DFLT_LTNCY) {
data = PCI_DFLT_LTNCY;
isp_prt(isp, ISP_LOGCONFIG, "set PCI latency to %d", data);
pci_write_config(dev, PCIR_LATTIMER, data, 1);
}
/*
* Make sure we've disabled the ROM.
*/
data = pci_read_config(dev, PCIR_ROMADDR, 4);
data &= ~1;
pci_write_config(dev, PCIR_ROMADDR, data, 4);
iqd = 0;
irq = bus_alloc_resource(dev, SYS_RES_IRQ, &iqd, 0, ~0,
1, RF_ACTIVE | RF_SHAREABLE);
if (irq == NULL) {
device_printf(dev, "could not allocate interrupt\n");
goto bad;
}
tval = 0;
if (resource_int_value(device_get_name(dev), device_get_unit(dev),
"fwload_disable", &tval) == 0 && tval != 0) {
isp->isp_confopts |= ISP_CFG_NORELOAD;
}
tval = 0;
if (resource_int_value(device_get_name(dev), device_get_unit(dev),
"ignore_nvram", &tval) == 0 && tval != 0) {
isp->isp_confopts |= ISP_CFG_NONVRAM;
}
tval = 0;
if (resource_int_value(device_get_name(dev), device_get_unit(dev),
"fullduplex", &tval) == 0 && tval != 0) {
isp->isp_confopts |= ISP_CFG_FULL_DUPLEX;
}
#ifdef ISP_FW_CRASH_DUMP
tval = 0;
if (resource_int_value(device_get_name(dev), device_get_unit(dev),
"fw_dump_enable", &tval) == 0 && tval != 0) {
size_t amt = 0;
if (IS_2200(isp)) {
amt = QLA2200_RISC_IMAGE_DUMP_SIZE;
} else if (IS_23XX(isp)) {
amt = QLA2300_RISC_IMAGE_DUMP_SIZE;
}
if (amt) {
FCPARAM(isp)->isp_dump_data =
malloc(amt, M_DEVBUF, M_WAITOK | M_ZERO);
} else {
device_printf(dev,
"f/w crash dumps not supported for this model\n");
}
}
#endif
sptr = 0;
if (resource_string_value(device_get_name(dev), device_get_unit(dev),
2001-07-04 18:51:58 +00:00
"topology", (const char **) &sptr) == 0 && sptr != 0) {
if (strcmp(sptr, "lport") == 0) {
isp->isp_confopts |= ISP_CFG_LPORT;
} else if (strcmp(sptr, "nport") == 0) {
isp->isp_confopts |= ISP_CFG_NPORT;
} else if (strcmp(sptr, "lport-only") == 0) {
isp->isp_confopts |= ISP_CFG_LPORT_ONLY;
} else if (strcmp(sptr, "nport-only") == 0) {
isp->isp_confopts |= ISP_CFG_NPORT_ONLY;
}
}
/*
* Because the resource_*_value functions can neither return
* 64 bit integer values, nor can they be directly coerced
* to interpret the right hand side of the assignment as
* you want them to interpret it, we have to force WWN
* hint replacement to specify WWN strings with a leading
* 'w' (e..g w50000000aaaa0001). Sigh.
*/
sptr = 0;
tval = resource_string_value(device_get_name(dev), device_get_unit(dev),
2001-07-04 18:51:58 +00:00
"portwwn", (const char **) &sptr);
if (tval == 0 && sptr != 0 && *sptr++ == 'w') {
char *eptr = 0;
isp->isp_osinfo.default_port_wwn = strtouq(sptr, &eptr, 16);
if (eptr < sptr + 16 || isp->isp_osinfo.default_port_wwn == 0) {
device_printf(dev, "mangled portwwn hint '%s'\n", sptr);
isp->isp_osinfo.default_port_wwn = 0;
} else {
isp->isp_confopts |= ISP_CFG_OWNWWPN;
}
}
if (isp->isp_osinfo.default_port_wwn == 0) {
isp->isp_osinfo.default_port_wwn = 0x400000007F000009ull;
}
sptr = 0;
tval = resource_string_value(device_get_name(dev), device_get_unit(dev),
2001-07-04 18:51:58 +00:00
"nodewwn", (const char **) &sptr);
if (tval == 0 && sptr != 0 && *sptr++ == 'w') {
char *eptr = 0;
isp->isp_osinfo.default_node_wwn = strtouq(sptr, &eptr, 16);
if (eptr < sptr + 16 || isp->isp_osinfo.default_node_wwn == 0) {
device_printf(dev, "mangled nodewwn hint '%s'\n", sptr);
isp->isp_osinfo.default_node_wwn = 0;
} else {
isp->isp_confopts |= ISP_CFG_OWNWWNN;
}
}
if (isp->isp_osinfo.default_node_wwn == 0) {
isp->isp_osinfo.default_node_wwn = 0x400000007F000009ull;
}
isp_debug = 0;
(void) resource_int_value(device_get_name(dev), device_get_unit(dev),
"debug", &isp_debug);
/* Make sure the lock is set up. */
mtx_init(&isp->isp_osinfo.lock, "isp", NULL, MTX_DEF);
locksetup++;
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
#ifdef ISP_SMPLOCK
#define INTR_FLAGS INTR_TYPE_CAM | INTR_MPSAFE | INTR_ENTROPY
#else
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
#define INTR_FLAGS INTR_TYPE_CAM | INTR_ENTROPY
#endif
if (bus_setup_intr(dev, irq, INTR_FLAGS, isp_pci_intr, isp, &pcs->ih)) {
device_printf(dev, "could not setup interrupt\n");
goto bad;
}
/*
* Set up logging levels.
*/
if (isp_debug) {
isp->isp_dblev = isp_debug;
} else {
isp->isp_dblev = ISP_LOGWARN|ISP_LOGERR;
}
if (bootverbose)
isp->isp_dblev |= ISP_LOGCONFIG|ISP_LOGINFO;
/*
* Last minute checks...
*/
if (IS_2312(isp)) {
isp->isp_port = pci_get_function(dev);
}
/*
* Make sure we're in reset state.
*/
ISP_LOCK(isp);
isp_reset(isp);
if (isp->isp_state != ISP_RESETSTATE) {
ISP_UNLOCK(isp);
goto bad;
}
isp_init(isp);
if (isp->isp_role != ISP_ROLE_NONE && isp->isp_state != ISP_INITSTATE) {
isp_uninit(isp);
ISP_UNLOCK(isp);
goto bad;
}
isp_attach(isp);
if (isp->isp_role != ISP_ROLE_NONE && isp->isp_state != ISP_RUNSTATE) {
isp_uninit(isp);
ISP_UNLOCK(isp);
goto bad;
}
/*
* XXXX: Here is where we might unload the f/w module
* XXXX: (or decrease the reference count to it).
*/
ISP_UNLOCK(isp);
return (0);
bad:
if (pcs && pcs->ih) {
(void) bus_teardown_intr(dev, irq, pcs->ih);
}
if (locksetup && isp) {
mtx_destroy(&isp->isp_osinfo.lock);
}
if (irq) {
(void) bus_release_resource(dev, SYS_RES_IRQ, iqd, irq);
}
if (regs) {
(void) bus_release_resource(dev, rtp, rgd, regs);
}
if (pcs) {
if (pcs->pci_isp.isp_param)
free(pcs->pci_isp.isp_param, M_DEVBUF);
free(pcs, M_DEVBUF);
}
/*
* XXXX: Here is where we might unload the f/w module
* XXXX: (or decrease the reference count to it).
*/
return (ENXIO);
}
static void
isp_pci_intr(void *arg)
{
struct ispsoftc *isp = arg;
u_int16_t isr, sema, mbox;
ISP_LOCK(isp);
isp->isp_intcnt++;
if (ISP_READ_ISR(isp, &isr, &sema, &mbox) == 0) {
isp->isp_intbogus++;
} else {
int iok = isp->isp_osinfo.intsok;
isp->isp_osinfo.intsok = 0;
isp_intr(isp, isr, sema, mbox);
isp->isp_osinfo.intsok = iok;
}
ISP_UNLOCK(isp);
}
#define IspVirt2Off(a, x) \
(((struct isp_pcisoftc *)a)->pci_poff[((x) & _BLK_REG_MASK) >> \
_BLK_REG_SHFT] + ((x) & 0xff))
#define BXR2(pcs, off) \
bus_space_read_2(pcs->pci_st, pcs->pci_sh, off)
#define BXW2(pcs, off, v) \
bus_space_write_2(pcs->pci_st, pcs->pci_sh, off, v)
static INLINE int
isp_pci_rd_debounced(struct ispsoftc *isp, int off, u_int16_t *rp)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *) isp;
u_int16_t val0, val1;
int i = 0;
do {
val0 = BXR2(pcs, IspVirt2Off(isp, off));
val1 = BXR2(pcs, IspVirt2Off(isp, off));
} while (val0 != val1 && ++i < 1000);
if (val0 != val1) {
return (1);
}
*rp = val0;
return (0);
}
static int
isp_pci_rd_isr(struct ispsoftc *isp, u_int16_t *isrp,
u_int16_t *semap, u_int16_t *mbp)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *) isp;
u_int16_t isr, sema;
if (IS_2100(isp)) {
if (isp_pci_rd_debounced(isp, BIU_ISR, &isr)) {
return (0);
}
if (isp_pci_rd_debounced(isp, BIU_SEMA, &sema)) {
return (0);
}
} else {
isr = BXR2(pcs, IspVirt2Off(isp, BIU_ISR));
sema = BXR2(pcs, IspVirt2Off(isp, BIU_SEMA));
}
isp_prt(isp, ISP_LOGDEBUG3, "ISR 0x%x SEMA 0x%x", isr, sema);
isr &= INT_PENDING_MASK(isp);
sema &= BIU_SEMA_LOCK;
if (isr == 0 && sema == 0) {
return (0);
}
*isrp = isr;
if ((*semap = sema) != 0) {
if (IS_2100(isp)) {
if (isp_pci_rd_debounced(isp, OUTMAILBOX0, mbp)) {
return (0);
}
} else {
*mbp = BXR2(pcs, IspVirt2Off(isp, OUTMAILBOX0));
}
}
return (1);
}
static int
isp_pci_rd_isr_2300(struct ispsoftc *isp, u_int16_t *isrp,
u_int16_t *semap, u_int16_t *mbox0p)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *) isp;
u_int32_t r2hisr;
if (!(BXR2(pcs, IspVirt2Off(isp, BIU_ISR) & BIU2100_ISR_RISC_INT))) {
*isrp = 0;
return (0);
}
r2hisr = bus_space_read_4(pcs->pci_st, pcs->pci_sh,
IspVirt2Off(pcs, BIU_R2HSTSLO));
isp_prt(isp, ISP_LOGDEBUG3, "RISC2HOST ISR 0x%x", r2hisr);
if ((r2hisr & BIU_R2HST_INTR) == 0) {
*isrp = 0;
return (0);
}
switch (r2hisr & BIU_R2HST_ISTAT_MASK) {
case ISPR2HST_ROM_MBX_OK:
case ISPR2HST_ROM_MBX_FAIL:
case ISPR2HST_MBX_OK:
case ISPR2HST_MBX_FAIL:
case ISPR2HST_ASYNC_EVENT:
case ISPR2HST_RIO_16:
case ISPR2HST_FPOST:
case ISPR2HST_FPOST_CTIO:
*isrp = r2hisr & 0xffff;
*mbox0p = (r2hisr >> 16);
*semap = 1;
return (1);
case ISPR2HST_RSPQ_UPDATE:
*isrp = r2hisr & 0xffff;
*mbox0p = 0;
*semap = 0;
return (1);
default:
return (0);
}
}
static u_int16_t
isp_pci_rd_reg(struct ispsoftc *isp, int regoff)
{
u_int16_t rv;
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *) isp;
int oldconf = 0;
if ((regoff & _BLK_REG_MASK) == SXP_BLOCK) {
/*
* We will assume that someone has paused the RISC processor.
*/
oldconf = BXR2(pcs, IspVirt2Off(isp, BIU_CONF1));
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1),
oldconf | BIU_PCI_CONF1_SXP);
}
rv = BXR2(pcs, IspVirt2Off(isp, regoff));
if ((regoff & _BLK_REG_MASK) == SXP_BLOCK) {
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1), oldconf);
}
return (rv);
}
static void
isp_pci_wr_reg(struct ispsoftc *isp, int regoff, u_int16_t val)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *) isp;
int oldconf = 0;
if ((regoff & _BLK_REG_MASK) == SXP_BLOCK) {
/*
* We will assume that someone has paused the RISC processor.
*/
oldconf = BXR2(pcs, IspVirt2Off(isp, BIU_CONF1));
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1),
oldconf | BIU_PCI_CONF1_SXP);
}
BXW2(pcs, IspVirt2Off(isp, regoff), val);
if ((regoff & _BLK_REG_MASK) == SXP_BLOCK) {
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1), oldconf);
}
}
static u_int16_t
isp_pci_rd_reg_1080(struct ispsoftc *isp, int regoff)
{
1999-12-16 05:42:02 +00:00
u_int16_t rv, oc = 0;
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *) isp;
1999-12-16 05:42:02 +00:00
if ((regoff & _BLK_REG_MASK) == SXP_BLOCK ||
(regoff & _BLK_REG_MASK) == (SXP_BLOCK|SXP_BANK1_SELECT)) {
u_int16_t tc;
/*
* We will assume that someone has paused the RISC processor.
*/
oc = BXR2(pcs, IspVirt2Off(isp, BIU_CONF1));
1999-12-16 05:42:02 +00:00
tc = oc & ~BIU_PCI1080_CONF1_DMA;
if (regoff & SXP_BANK1_SELECT)
tc |= BIU_PCI1080_CONF1_SXP1;
else
tc |= BIU_PCI1080_CONF1_SXP0;
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1), tc);
} else if ((regoff & _BLK_REG_MASK) == DMA_BLOCK) {
oc = BXR2(pcs, IspVirt2Off(isp, BIU_CONF1));
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1),
oc | BIU_PCI1080_CONF1_DMA);
}
rv = BXR2(pcs, IspVirt2Off(isp, regoff));
1999-12-16 05:42:02 +00:00
if (oc) {
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1), oc);
}
return (rv);
}
static void
isp_pci_wr_reg_1080(struct ispsoftc *isp, int regoff, u_int16_t val)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *) isp;
int oc = 0;
1999-12-16 05:42:02 +00:00
if ((regoff & _BLK_REG_MASK) == SXP_BLOCK ||
(regoff & _BLK_REG_MASK) == (SXP_BLOCK|SXP_BANK1_SELECT)) {
u_int16_t tc;
/*
* We will assume that someone has paused the RISC processor.
*/
oc = BXR2(pcs, IspVirt2Off(isp, BIU_CONF1));
1999-12-16 05:42:02 +00:00
tc = oc & ~BIU_PCI1080_CONF1_DMA;
if (regoff & SXP_BANK1_SELECT)
tc |= BIU_PCI1080_CONF1_SXP1;
else
tc |= BIU_PCI1080_CONF1_SXP0;
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1), tc);
} else if ((regoff & _BLK_REG_MASK) == DMA_BLOCK) {
oc = BXR2(pcs, IspVirt2Off(isp, BIU_CONF1));
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1),
oc | BIU_PCI1080_CONF1_DMA);
}
BXW2(pcs, IspVirt2Off(isp, regoff), val);
1999-12-16 05:42:02 +00:00
if (oc) {
BXW2(pcs, IspVirt2Off(isp, BIU_CONF1), oc);
}
}
struct imush {
struct ispsoftc *isp;
int error;
};
static void imc(void *, bus_dma_segment_t *, int, int);
static void
imc(void *arg, bus_dma_segment_t *segs, int nseg, int error)
{
struct imush *imushp = (struct imush *) arg;
if (error) {
imushp->error = error;
} else {
struct ispsoftc *isp =imushp->isp;
bus_addr_t addr = segs->ds_addr;
isp->isp_rquest_dma = addr;
addr += ISP_QUEUE_SIZE(RQUEST_QUEUE_LEN(isp));
isp->isp_result_dma = addr;
if (IS_FC(isp)) {
addr += ISP_QUEUE_SIZE(RESULT_QUEUE_LEN(isp));
FCPARAM(isp)->isp_scdma = addr;
}
}
}
/*
* Should be BUS_SPACE_MAXSIZE, but MAXPHYS is larger than BUS_SPACE_MAXSIZE
*/
#define ISP_NSEGS ((MAXPHYS / PAGE_SIZE) + 1)
static int
1998-12-28 19:24:23 +00:00
isp_pci_mbxdma(struct ispsoftc *isp)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *)isp;
caddr_t base;
u_int32_t len;
int i, error, ns;
bus_size_t bl;
struct imush im;
/*
* Already been here? If so, leave...
*/
if (isp->isp_rquest) {
return (0);
}
if (IS_ULTRA2(isp) || IS_FC(isp) || IS_1240(isp)) {
bl = BUS_SPACE_UNRESTRICTED;
} else {
bl = BUS_SPACE_MAXADDR_24BIT;
}
if (bus_dma_tag_create(NULL, 1, 0, BUS_SPACE_MAXADDR,
BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
ISP_NSEGS, bl, 0, &pcs->dmat)) {
isp_prt(isp, ISP_LOGERR, "could not create master dma tag");
return(1);
}
len = sizeof (XS_T **) * isp->isp_maxcmds;
isp->isp_xflist = (XS_T **) malloc(len, M_DEVBUF, M_WAITOK | M_ZERO);
if (isp->isp_xflist == NULL) {
isp_prt(isp, ISP_LOGERR, "cannot alloc xflist array");
return (1);
}
len = sizeof (bus_dmamap_t) * isp->isp_maxcmds;
pcs->dmaps = (bus_dmamap_t *) malloc(len, M_DEVBUF, M_WAITOK);
if (pcs->dmaps == NULL) {
isp_prt(isp, ISP_LOGERR, "can't alloc dma map storage");
free(isp->isp_xflist, M_DEVBUF);
return (1);
}
/*
* Allocate and map the request, result queues, plus FC scratch area.
*/
len = ISP_QUEUE_SIZE(RQUEST_QUEUE_LEN(isp));
len += ISP_QUEUE_SIZE(RESULT_QUEUE_LEN(isp));
if (IS_FC(isp)) {
len += ISP2100_SCRLEN;
}
ns = (len / PAGE_SIZE) + 1;
if (bus_dma_tag_create(pcs->dmat, QENTRY_LEN, 0, BUS_SPACE_MAXADDR,
BUS_SPACE_MAXADDR, NULL, NULL, len, ns, bl, 0, &isp->isp_cdmat)) {
isp_prt(isp, ISP_LOGERR,
"cannot create a dma tag for control spaces");
free(pcs->dmaps, M_DEVBUF);
free(isp->isp_xflist, M_DEVBUF);
return (1);
}
if (bus_dmamem_alloc(isp->isp_cdmat, (void **)&base, BUS_DMA_NOWAIT,
&isp->isp_cdmap) != 0) {
isp_prt(isp, ISP_LOGERR,
"cannot allocate %d bytes of CCB memory", len);
bus_dma_tag_destroy(isp->isp_cdmat);
free(isp->isp_xflist, M_DEVBUF);
free(pcs->dmaps, M_DEVBUF);
return (1);
}
for (i = 0; i < isp->isp_maxcmds; i++) {
error = bus_dmamap_create(pcs->dmat, 0, &pcs->dmaps[i]);
if (error) {
isp_prt(isp, ISP_LOGERR,
"error %d creating per-cmd DMA maps", error);
while (--i >= 0) {
bus_dmamap_destroy(pcs->dmat, pcs->dmaps[i]);
}
goto bad;
}
}
im.isp = isp;
im.error = 0;
bus_dmamap_load(isp->isp_cdmat, isp->isp_cdmap, base, len, imc, &im, 0);
if (im.error) {
isp_prt(isp, ISP_LOGERR,
"error %d loading dma map for control areas", im.error);
goto bad;
}
isp->isp_rquest = base;
base += ISP_QUEUE_SIZE(RQUEST_QUEUE_LEN(isp));
isp->isp_result = base;
if (IS_FC(isp)) {
base += ISP_QUEUE_SIZE(RESULT_QUEUE_LEN(isp));
FCPARAM(isp)->isp_scratch = base;
1999-03-25 22:53:56 +00:00
}
return (0);
bad:
bus_dmamem_free(isp->isp_cdmat, base, isp->isp_cdmap);
bus_dma_tag_destroy(isp->isp_cdmat);
free(isp->isp_xflist, M_DEVBUF);
free(pcs->dmaps, M_DEVBUF);
isp->isp_rquest = NULL;
return (1);
}
typedef struct {
struct ispsoftc *isp;
void *cmd_token;
void *rq;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
u_int16_t *nxtip;
u_int16_t optr;
u_int error;
} mush_t;
#define MUSHERR_NOQENTRIES -2
#ifdef ISP_TARGET_MODE
/*
* We need to handle DMA for target mode differently from initiator mode.
*
* DMA mapping and construction and submission of CTIO Request Entries
* and rendevous for completion are very tightly coupled because we start
* out by knowing (per platform) how much data we have to move, but we
* don't know, up front, how many DMA mapping segments will have to be used
* cover that data, so we don't know how many CTIO Request Entries we
* will end up using. Further, for performance reasons we may want to
* (on the last CTIO for Fibre Channel), send status too (if all went well).
*
* The standard vector still goes through isp_pci_dmasetup, but the callback
* for the DMA mapping routines comes here instead with the whole transfer
* mapped and a pointer to a partially filled in already allocated request
* queue entry. We finish the job.
*/
static void tdma_mk(void *, bus_dma_segment_t *, int, int);
static void tdma_mkfc(void *, bus_dma_segment_t *, int, int);
#define STATUS_WITH_DATA 1
static void
tdma_mk(void *arg, bus_dma_segment_t *dm_segs, int nseg, int error)
{
mush_t *mp;
struct ccb_scsiio *csio;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
struct ispsoftc *isp;
struct isp_pcisoftc *pcs;
bus_dmamap_t *dp;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
ct_entry_t *cto, *qe;
u_int8_t scsi_status;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
u_int16_t curi, nxti, handle;
u_int32_t sflags;
int32_t resid;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
int nth_ctio, nctios, send_status;
mp = (mush_t *) arg;
if (error) {
mp->error = error;
return;
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp = mp->isp;
csio = mp->cmd_token;
cto = mp->rq;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
curi = isp->isp_reqidx;
qe = (ct_entry_t *) ISP_QUEUE_ENTRY(isp->isp_rquest, curi);
cto->ct_xfrlen = 0;
cto->ct_seg_count = 0;
cto->ct_header.rqs_entry_count = 1;
MEMZERO(cto->ct_dataseg, sizeof(cto->ct_dataseg));
if (nseg == 0) {
cto->ct_header.rqs_seqno = 1;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGTDEBUG1,
"CTIO[%x] lun%d iid%d tag %x flgs %x sts %x ssts %x res %d",
cto->ct_fwhandle, csio->ccb_h.target_lun, cto->ct_iid,
cto->ct_tag_val, cto->ct_flags, cto->ct_status,
cto->ct_scsi_status, cto->ct_resid);
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
ISP_TDQE(isp, "tdma_mk[no data]", curi, cto);
isp_put_ctio(isp, cto, qe);
return;
}
nctios = nseg / ISP_RQDSEG;
if (nseg % ISP_RQDSEG) {
nctios++;
}
/*
* Save syshandle, and potentially any SCSI status, which we'll
* reinsert on the last CTIO we're going to send.
*/
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
handle = cto->ct_syshandle;
cto->ct_syshandle = 0;
cto->ct_header.rqs_seqno = 0;
send_status = (cto->ct_flags & CT_SENDSTATUS) != 0;
if (send_status) {
sflags = cto->ct_flags & (CT_SENDSTATUS | CT_CCINCR);
cto->ct_flags &= ~(CT_SENDSTATUS | CT_CCINCR);
/*
* Preserve residual.
*/
resid = cto->ct_resid;
/*
* Save actual SCSI status.
*/
scsi_status = cto->ct_scsi_status;
#ifndef STATUS_WITH_DATA
sflags |= CT_NO_DATA;
/*
* We can't do a status at the same time as a data CTIO, so
* we need to synthesize an extra CTIO at this level.
*/
nctios++;
#endif
} else {
sflags = scsi_status = resid = 0;
}
cto->ct_resid = 0;
cto->ct_scsi_status = 0;
pcs = (struct isp_pcisoftc *)isp;
dp = &pcs->dmaps[isp_handle_index(handle)];
if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
bus_dmamap_sync(pcs->dmat, *dp, BUS_DMASYNC_PREREAD);
} else {
bus_dmamap_sync(pcs->dmat, *dp, BUS_DMASYNC_PREWRITE);
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
nxti = *mp->nxtip;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
for (nth_ctio = 0; nth_ctio < nctios; nth_ctio++) {
int seglim;
seglim = nseg;
if (seglim) {
int seg;
if (seglim > ISP_RQDSEG)
seglim = ISP_RQDSEG;
for (seg = 0; seg < seglim; seg++, nseg--) {
/*
* Unlike normal initiator commands, we don't
* do any swizzling here.
*/
cto->ct_dataseg[seg].ds_count = dm_segs->ds_len;
cto->ct_dataseg[seg].ds_base = dm_segs->ds_addr;
cto->ct_xfrlen += dm_segs->ds_len;
dm_segs++;
}
cto->ct_seg_count = seg;
} else {
/*
* This case should only happen when we're sending an
* extra CTIO with final status.
*/
if (send_status == 0) {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGWARN,
"tdma_mk ran out of segments");
mp->error = EINVAL;
return;
}
}
/*
* At this point, the fields ct_lun, ct_iid, ct_tagval,
* ct_tagtype, and ct_timeout have been carried over
* unchanged from what our caller had set.
*
* The dataseg fields and the seg_count fields we just got
* through setting. The data direction we've preserved all
* along and only clear it if we're now sending status.
*/
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
if (nth_ctio == nctios - 1) {
/*
* We're the last in a sequence of CTIOs, so mark
* this CTIO and save the handle to the CCB such that
* when this CTIO completes we can free dma resources
* and do whatever else we need to do to finish the
* rest of the command. We *don't* give this to the
* firmware to work on- the caller will do that.
*/
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
cto->ct_syshandle = handle;
cto->ct_header.rqs_seqno = 1;
if (send_status) {
cto->ct_scsi_status = scsi_status;
cto->ct_flags |= sflags;
cto->ct_resid = resid;
}
if (send_status) {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGTDEBUG1,
"CTIO[%x] lun%d iid %d tag %x ct_flags %x "
"scsi status %x resid %d",
cto->ct_fwhandle, csio->ccb_h.target_lun,
cto->ct_iid, cto->ct_tag_val, cto->ct_flags,
cto->ct_scsi_status, cto->ct_resid);
} else {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGTDEBUG1,
"CTIO[%x] lun%d iid%d tag %x ct_flags 0x%x",
cto->ct_fwhandle, csio->ccb_h.target_lun,
cto->ct_iid, cto->ct_tag_val,
cto->ct_flags);
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_put_ctio(isp, cto, qe);
ISP_TDQE(isp, "last tdma_mk", curi, cto);
if (nctios > 1) {
MEMORYBARRIER(isp, SYNC_REQUEST,
curi, QENTRY_LEN);
}
} else {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
ct_entry_t *oqe = qe;
/*
* Make sure syshandle fields are clean
*/
cto->ct_syshandle = 0;
cto->ct_header.rqs_seqno = 0;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGTDEBUG1,
"CTIO[%x] lun%d for ID%d ct_flags 0x%x",
cto->ct_fwhandle, csio->ccb_h.target_lun,
cto->ct_iid, cto->ct_flags);
/*
* Get a new CTIO
*/
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
qe = (ct_entry_t *)
ISP_QUEUE_ENTRY(isp->isp_rquest, nxti);
nxti = ISP_NXT_QENTRY(nxti, RQUEST_QUEUE_LEN(isp));
if (nxti == mp->optr) {
isp_prt(isp, ISP_LOGTDEBUG0,
"Queue Overflow in tdma_mk");
mp->error = MUSHERR_NOQENTRIES;
return;
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
/*
* Now that we're done with the old CTIO,
* flush it out to the request queue.
*/
ISP_TDQE(isp, "dma_tgt_fc", curi, cto);
isp_put_ctio(isp, cto, oqe);
if (nth_ctio != 0) {
MEMORYBARRIER(isp, SYNC_REQUEST, curi,
QENTRY_LEN);
}
curi = ISP_NXT_QENTRY(curi, RQUEST_QUEUE_LEN(isp));
/*
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
* Reset some fields in the CTIO so we can reuse
* for the next one we'll flush to the request
* queue.
*/
cto->ct_header.rqs_entry_type = RQSTYPE_CTIO;
cto->ct_header.rqs_entry_count = 1;
cto->ct_header.rqs_flags = 0;
cto->ct_status = 0;
cto->ct_scsi_status = 0;
cto->ct_xfrlen = 0;
cto->ct_resid = 0;
cto->ct_seg_count = 0;
MEMZERO(cto->ct_dataseg, sizeof(cto->ct_dataseg));
}
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
*mp->nxtip = nxti;
}
static void
tdma_mkfc(void *arg, bus_dma_segment_t *dm_segs, int nseg, int error)
{
mush_t *mp;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
u_int8_t sense[QLTM_SENSELEN];
struct ccb_scsiio *csio;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
struct ispsoftc *isp;
struct isp_pcisoftc *pcs;
bus_dmamap_t *dp;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
ct2_entry_t *cto, *qe;
u_int16_t scsi_status, send_status, send_sense, handle;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
u_int16_t curi, nxti;
int32_t resid;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
int nth_ctio, nctios;
mp = (mush_t *) arg;
if (error) {
mp->error = error;
return;
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp = mp->isp;
csio = mp->cmd_token;
cto = mp->rq;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
curi = isp->isp_reqidx;
qe = (ct2_entry_t *) ISP_QUEUE_ENTRY(isp->isp_rquest, curi);
if (nseg == 0) {
if ((cto->ct_flags & CT2_FLAG_MMASK) != CT2_FLAG_MODE1) {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGWARN,
"dma2_tgt_fc, a status CTIO2 without MODE1 "
"set (0x%x)", cto->ct_flags);
mp->error = EINVAL;
return;
}
cto->ct_header.rqs_entry_count = 1;
cto->ct_header.rqs_seqno = 1;
/* ct_syshandle contains the handle set by caller */
/*
* We preserve ct_lun, ct_iid, ct_rxid. We set the data
* flags to NO DATA and clear relative offset flags.
* We preserve the ct_resid and the response area.
*/
cto->ct_flags |= CT2_NO_DATA;
if (cto->ct_resid > 0)
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
cto->rsp.m1.ct_scsi_status |= CT2_DATA_UNDER;
else if (cto->ct_resid < 0)
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
cto->rsp.m1.ct_scsi_status |= CT2_DATA_OVER;
cto->ct_seg_count = 0;
cto->ct_reloff = 0;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGTDEBUG1,
"CTIO2[%x] lun %d->iid%d flgs 0x%x sts 0x%x ssts "
"0x%x res %d", cto->ct_rxid, csio->ccb_h.target_lun,
cto->ct_iid, cto->ct_flags, cto->ct_status,
cto->rsp.m1.ct_scsi_status, cto->ct_resid);
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_put_ctio2(isp, cto, qe);
ISP_TDQE(isp, "dma2_tgt_fc[no data]", curi, qe);
return;
}
if ((cto->ct_flags & CT2_FLAG_MMASK) != CT2_FLAG_MODE0) {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGWARN,
"dma2_tgt_fc, a data CTIO2 without MODE0 set "
"(0x%x)", cto->ct_flags);
mp->error = EINVAL;
return;
}
nctios = nseg / ISP_RQDSEG_T2;
if (nseg % ISP_RQDSEG_T2) {
nctios++;
}
/*
* Save the handle, status, reloff, and residual. We'll reinsert the
* handle into the last CTIO2 we're going to send, and reinsert status
* and residual (and possibly sense data) if that's to be sent as well.
*
* We preserve ct_reloff and adjust it for each data CTIO2 we send past
* the first one. This is needed so that the FCP DATA IUs being sent
* out have the correct offset (they can arrive at the other end out
* of order).
*/
handle = cto->ct_syshandle;
cto->ct_syshandle = 0;
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
send_status = (cto->ct_flags & CT2_SENDSTATUS) != 0;
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
if (send_status) {
cto->ct_flags &= ~(CT2_SENDSTATUS|CT2_CCINCR);
/*
* Preserve residual.
*/
resid = cto->ct_resid;
/*
* Save actual SCSI status. We'll reinsert the
* CT2_SNSLEN_VALID later if appropriate.
*/
scsi_status = cto->rsp.m0.ct_scsi_status & 0xff;
send_sense = cto->rsp.m0.ct_scsi_status & CT2_SNSLEN_VALID;
/*
* If we're sending status and have a CHECK CONDTION and
* have sense data, we send one more CTIO2 with just the
* status and sense data. The upper layers have stashed
* the sense data in the dataseg structure for us.
*/
if ((scsi_status & 0xf) == SCSI_STATUS_CHECK_COND &&
send_sense) {
bcopy(cto->rsp.m0.ct_dataseg, sense, QLTM_SENSELEN);
nctios++;
}
} else {
scsi_status = send_sense = resid = 0;
}
cto->ct_resid = 0;
cto->rsp.m0.ct_scsi_status = 0;
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
MEMZERO(&cto->rsp, sizeof (cto->rsp));
pcs = (struct isp_pcisoftc *)isp;
dp = &pcs->dmaps[isp_handle_index(handle)];
if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
bus_dmamap_sync(pcs->dmat, *dp, BUS_DMASYNC_PREREAD);
} else {
bus_dmamap_sync(pcs->dmat, *dp, BUS_DMASYNC_PREWRITE);
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
nxti = *mp->nxtip;
for (nth_ctio = 0; nth_ctio < nctios; nth_ctio++) {
u_int32_t oxfrlen;
int seglim;
seglim = nseg;
if (seglim) {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
int seg;
if (seglim > ISP_RQDSEG_T2)
seglim = ISP_RQDSEG_T2;
for (seg = 0; seg < seglim; seg++) {
cto->rsp.m0.ct_dataseg[seg].ds_base =
dm_segs->ds_addr;
cto->rsp.m0.ct_dataseg[seg].ds_count =
dm_segs->ds_len;
cto->rsp.m0.ct_xfrlen += dm_segs->ds_len;
dm_segs++;
}
cto->ct_seg_count = seg;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
oxfrlen = cto->rsp.m0.ct_xfrlen;
} else {
/*
* This case should only happen when we're sending a
* synthesized MODE1 final status with sense data.
*/
if (send_sense == 0) {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGWARN,
"dma2_tgt_fc ran out of segments, "
"no SENSE DATA");
mp->error = EINVAL;
return;
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
oxfrlen = 0;
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
/*
* At this point, the fields ct_lun, ct_iid, ct_rxid,
* ct_timeout have been carried over unchanged from what
* our caller had set.
*
* The field ct_reloff is either what the caller set, or
* what we've added to below.
*
* The dataseg fields and the seg_count fields we just got
* through setting. The data direction we've preserved all
* along and only clear it if we're sending a MODE1 status
* as the last CTIO.
*
*/
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
if (nth_ctio == nctios - 1) {
/*
* We're the last in a sequence of CTIO2s, so mark this
* CTIO2 and save the handle to the CCB such that when
* this CTIO2 completes we can free dma resources and
* do whatever else we need to do to finish the rest
* of the command.
*/
cto->ct_syshandle = handle;
cto->ct_header.rqs_seqno = 1;
if (send_status) {
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
/*
* Get 'real' residual and set flags based
* on it.
*/
cto->ct_resid = resid;
if (send_sense) {
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
MEMCPY(cto->rsp.m1.ct_resp, sense,
QLTM_SENSELEN);
cto->rsp.m1.ct_senselen =
QLTM_SENSELEN;
scsi_status |= CT2_SNSLEN_VALID;
cto->rsp.m1.ct_scsi_status =
scsi_status;
cto->ct_flags &= CT2_FLAG_MMASK;
cto->ct_flags |= CT2_FLAG_MODE1 |
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
CT2_NO_DATA | CT2_SENDSTATUS |
CT2_CCINCR;
if (cto->ct_resid > 0)
cto->rsp.m1.ct_scsi_status |=
CT2_DATA_UNDER;
else if (cto->ct_resid < 0)
cto->rsp.m1.ct_scsi_status |=
CT2_DATA_OVER;
} else {
cto->rsp.m0.ct_scsi_status =
scsi_status;
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
cto->ct_flags |=
CT2_SENDSTATUS | CT2_CCINCR;
if (cto->ct_resid > 0)
cto->rsp.m0.ct_scsi_status |=
CT2_DATA_UNDER;
else if (cto->ct_resid < 0)
cto->rsp.m0.ct_scsi_status |=
CT2_DATA_OVER;
}
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGTDEBUG1,
"CTIO2[%x] lun %d->iid%d flgs 0x%x sts 0x%x"
" ssts 0x%x res %d", cto->ct_rxid,
csio->ccb_h.target_lun, (int) cto->ct_iid,
cto->ct_flags, cto->ct_status,
cto->rsp.m1.ct_scsi_status, cto->ct_resid);
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_put_ctio2(isp, cto, qe);
ISP_TDQE(isp, "last dma2_tgt_fc", curi, qe);
if (nctios > 1) {
MEMORYBARRIER(isp, SYNC_REQUEST,
curi, QENTRY_LEN);
}
} else {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
ct2_entry_t *oqe = qe;
/*
* Make sure handle fields are clean
*/
cto->ct_syshandle = 0;
cto->ct_header.rqs_seqno = 0;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_prt(isp, ISP_LOGTDEBUG1,
"CTIO2[%x] lun %d->iid%d flgs 0x%x",
cto->ct_rxid, csio->ccb_h.target_lun,
(int) cto->ct_iid, cto->ct_flags);
/*
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
* Get a new CTIO2 entry from the request queue.
*/
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
qe = (ct2_entry_t *)
ISP_QUEUE_ENTRY(isp->isp_rquest, nxti);
nxti = ISP_NXT_QENTRY(nxti, RQUEST_QUEUE_LEN(isp));
if (nxti == mp->optr) {
isp_prt(isp, ISP_LOGWARN,
"Queue Overflow in dma2_tgt_fc");
mp->error = MUSHERR_NOQENTRIES;
return;
}
/*
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
* Now that we're done with the old CTIO2,
* flush it out to the request queue.
*/
ISP_TDQE(isp, "tdma_mkfc", curi, cto);
isp_put_ctio2(isp, cto, oqe);
if (nth_ctio != 0) {
MEMORYBARRIER(isp, SYNC_REQUEST, curi,
QENTRY_LEN);
}
curi = ISP_NXT_QENTRY(curi, RQUEST_QUEUE_LEN(isp));
/*
* Reset some fields in the CTIO2 so we can reuse
* for the next one we'll flush to the request
* queue.
*/
cto->ct_header.rqs_entry_type = RQSTYPE_CTIO2;
cto->ct_header.rqs_entry_count = 1;
cto->ct_header.rqs_flags = 0;
cto->ct_status = 0;
cto->ct_resid = 0;
cto->ct_seg_count = 0;
/*
* Adjust the new relative offset by the amount which
* is recorded in the data segment of the old CTIO2 we
* just finished filling out.
*/
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
cto->ct_reloff += oxfrlen;
Spring MegaChange #1. ---- Make a device for each ISP- really usable only with devfs and add an ioctl entry point (this can be used to (re)set debug levels, reset the HBA, rescan the fabric, issue lips, etc). ---- Add in a kernel thread for Fibre Channel cards. The purpose of this thread is to be woken up to clean up after Fibre Channel events block things. Basically, any FC event that casts doubt on the location or identify of FC devices blocks the queues. When, and if, we get the PORT DATABASE CHANGED or NAME SERVER DATABASE CHANGED async event, we activate the kthread which will then, in full thread context, re-evaluate the local loop and/or the fabric. When it's satisfied that things are stable, it can then release the blocked queues and let commands flow again. The prior mechanism was a lazy evaluation. That is, the next command to come down the pipe after change events would pay the full price for re-evaluation. And if this was done off of a softcall, it really could hang up the system. These changes brings the FreeBSD port more in line with the Solaris, Linux and NetBSD ports. It also, more importantly, gets us being more proactive about topology changes which could then be reflected upwards to CAM so that the periph driver can be informed sooner rather than later when things arrive or depart. --- Add in the (correct) usage of locking macros- we now have lock transition macros which allow us to transition from holding the CAM lock (Giant) and grabbing the softc lock and vice versa. Switch over to having this HBA do real locking. Some folks claim this won't be a win. They're right. But you have to start somewhere, and this will begin to teach us how to DTRT for HBAs, etc. -- Start putting in prototype 2300 support. Add back in LIP and Loop Reset as async events that each platform will handle. Add in another int_bogus instrumentation point. Do some more substantial target mode cleanups. MFC after: 8 weeks
2001-05-28 21:20:43 +00:00
MEMZERO(&cto->rsp, sizeof (cto->rsp));
}
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
*mp->nxtip = nxti;
}
#endif
static void dma2(void *, bus_dma_segment_t *, int, int);
static void
1998-12-28 19:24:23 +00:00
dma2(void *arg, bus_dma_segment_t *dm_segs, int nseg, int error)
{
mush_t *mp;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
struct ispsoftc *isp;
struct ccb_scsiio *csio;
struct isp_pcisoftc *pcs;
bus_dmamap_t *dp;
bus_dma_segment_t *eseg;
ispreq_t *rq;
int seglim, datalen;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
u_int16_t nxti;
mp = (mush_t *) arg;
if (error) {
mp->error = error;
return;
}
if (nseg < 1) {
isp_prt(mp->isp, ISP_LOGERR, "bad segment count (%d)", nseg);
mp->error = EFAULT;
return;
}
csio = mp->cmd_token;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp = mp->isp;
rq = mp->rq;
pcs = (struct isp_pcisoftc *)mp->isp;
dp = &pcs->dmaps[isp_handle_index(rq->req_handle)];
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
nxti = *mp->nxtip;
if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
bus_dmamap_sync(pcs->dmat, *dp, BUS_DMASYNC_PREREAD);
} else {
bus_dmamap_sync(pcs->dmat, *dp, BUS_DMASYNC_PREWRITE);
}
datalen = XS_XFRLEN(csio);
/*
* We're passed an initial partially filled in entry that
* has most fields filled in except for data transfer
* related values.
*
* Our job is to fill in the initial request queue entry and
* then to start allocating and filling in continuation entries
* until we've covered the entire transfer.
*/
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
if (IS_FC(isp)) {
seglim = ISP_RQDSEG_T2;
((ispreqt2_t *)rq)->req_totalcnt = datalen;
if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
((ispreqt2_t *)rq)->req_flags |= REQFLAG_DATA_IN;
} else {
((ispreqt2_t *)rq)->req_flags |= REQFLAG_DATA_OUT;
}
} else {
if (csio->cdb_len > 12) {
seglim = 0;
} else {
seglim = ISP_RQDSEG;
}
if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
rq->req_flags |= REQFLAG_DATA_IN;
} else {
rq->req_flags |= REQFLAG_DATA_OUT;
}
}
eseg = dm_segs + nseg;
while (datalen != 0 && rq->req_seg_count < seglim && dm_segs != eseg) {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
if (IS_FC(isp)) {
ispreqt2_t *rq2 = (ispreqt2_t *)rq;
rq2->req_dataseg[rq2->req_seg_count].ds_base =
dm_segs->ds_addr;
rq2->req_dataseg[rq2->req_seg_count].ds_count =
dm_segs->ds_len;
} else {
rq->req_dataseg[rq->req_seg_count].ds_base =
dm_segs->ds_addr;
rq->req_dataseg[rq->req_seg_count].ds_count =
dm_segs->ds_len;
}
datalen -= dm_segs->ds_len;
rq->req_seg_count++;
dm_segs++;
}
while (datalen > 0 && dm_segs != eseg) {
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
u_int16_t onxti;
ispcontreq_t local, *crq = &local, *cqe;
cqe = (ispcontreq_t *) ISP_QUEUE_ENTRY(isp->isp_rquest, nxti);
onxti = nxti;
nxti = ISP_NXT_QENTRY(onxti, RQUEST_QUEUE_LEN(isp));
if (nxti == mp->optr) {
isp_prt(isp, ISP_LOGDEBUG0, "Request Queue Overflow++");
mp->error = MUSHERR_NOQENTRIES;
return;
}
rq->req_header.rqs_entry_count++;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
MEMZERO((void *)crq, sizeof (*crq));
crq->req_header.rqs_entry_count = 1;
crq->req_header.rqs_entry_type = RQSTYPE_DATASEG;
seglim = 0;
while (datalen > 0 && seglim < ISP_CDSEG && dm_segs != eseg) {
crq->req_dataseg[seglim].ds_base =
dm_segs->ds_addr;
crq->req_dataseg[seglim].ds_count =
dm_segs->ds_len;
rq->req_seg_count++;
dm_segs++;
seglim++;
datalen -= dm_segs->ds_len;
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
isp_put_cont_req(isp, crq, cqe);
MEMORYBARRIER(isp, SYNC_REQUEST, onxti, QENTRY_LEN);
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
*mp->nxtip = nxti;
}
static int
isp_pci_dmasetup(struct ispsoftc *isp, struct ccb_scsiio *csio, ispreq_t *rq,
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
u_int16_t *nxtip, u_int16_t optr)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *)isp;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
ispreq_t *qep;
bus_dmamap_t *dp = NULL;
mush_t mush, *mp;
void (*eptr)(void *, bus_dma_segment_t *, int, int);
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
qep = (ispreq_t *) ISP_QUEUE_ENTRY(isp->isp_rquest, isp->isp_reqidx);
#ifdef ISP_TARGET_MODE
if (csio->ccb_h.func_code == XPT_CONT_TARGET_IO) {
if (IS_FC(isp)) {
eptr = tdma_mkfc;
} else {
eptr = tdma_mk;
}
if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE ||
(csio->dxfer_len == 0)) {
mp = &mush;
mp->isp = isp;
mp->cmd_token = csio;
mp->rq = rq; /* really a ct_entry_t or ct2_entry_t */
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
mp->nxtip = nxtip;
mp->optr = optr;
mp->error = 0;
(*eptr)(mp, NULL, 0, 0);
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
goto mbxsync;
}
} else
#endif
eptr = dma2;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE ||
(csio->dxfer_len == 0)) {
rq->req_seg_count = 1;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
goto mbxsync;
}
/*
* Do a virtual grapevine step to collect info for
* the callback dma allocation that we have to use...
*/
mp = &mush;
mp->isp = isp;
mp->cmd_token = csio;
mp->rq = rq;
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
mp->nxtip = nxtip;
mp->optr = optr;
mp->error = 0;
if ((csio->ccb_h.flags & CAM_SCATTER_VALID) == 0) {
if ((csio->ccb_h.flags & CAM_DATA_PHYS) == 0) {
int error, s;
dp = &pcs->dmaps[isp_handle_index(rq->req_handle)];
s = splsoftvm();
error = bus_dmamap_load(pcs->dmat, *dp,
csio->data_ptr, csio->dxfer_len, eptr, mp, 0);
if (error == EINPROGRESS) {
bus_dmamap_unload(pcs->dmat, *dp);
mp->error = EINVAL;
isp_prt(isp, ISP_LOGERR,
"deferred dma allocation not supported");
} else if (error && mp->error == 0) {
#ifdef DIAGNOSTIC
isp_prt(isp, ISP_LOGERR,
"error %d in dma mapping code", error);
#endif
mp->error = error;
}
splx(s);
} else {
/* Pointer to physical buffer */
struct bus_dma_segment seg;
seg.ds_addr = (bus_addr_t)csio->data_ptr;
seg.ds_len = csio->dxfer_len;
(*eptr)(mp, &seg, 1, 0);
}
} else {
struct bus_dma_segment *segs;
if ((csio->ccb_h.flags & CAM_DATA_PHYS) != 0) {
isp_prt(isp, ISP_LOGERR,
"Physical segment pointers unsupported");
mp->error = EINVAL;
} else if ((csio->ccb_h.flags & CAM_SG_LIST_PHYS) == 0) {
isp_prt(isp, ISP_LOGERR,
"Virtual segment addresses unsupported");
mp->error = EINVAL;
} else {
/* Just use the segments provided */
segs = (struct bus_dma_segment *) csio->data_ptr;
(*eptr)(mp, segs, csio->sglist_cnt, 0);
}
}
if (mp->error) {
int retval = CMD_COMPLETE;
if (mp->error == MUSHERR_NOQENTRIES) {
retval = CMD_EAGAIN;
} else if (mp->error == EFBIG) {
XS_SETERR(csio, CAM_REQ_TOO_BIG);
} else if (mp->error == EINVAL) {
XS_SETERR(csio, CAM_REQ_INVALID);
} else {
XS_SETERR(csio, CAM_UNREC_HBA_ERROR);
}
return (retval);
}
Major restructuring for swizzling to the request queue and unswizzling from the response queue. Instead of the ad hoc ISP_SWIZZLE_REQUEST, we now have a complete set of inline functions in isp_inline.h. Each platform is responsible for providing just one of a set of ISP_IOX_{GET,PUT}{8,16,32} macros. The reason this needs to be done is that we need to have a single set of functions that will work correctly on multiple architectures for both little and big endian machines. It also needs to work correctly in the case that we have the request or response queues in memory that has to be treated specially (e.g., have ddi_dma_sync called on it for Solaris after we update it or before we read from it). It also has to handle the SBus cards (for platforms that have them) which, while on a Big Endian machine, do *not* require *most* of the request/response queue entry fields to be swizzled or unswizzled. One thing that falls out of this is that we no longer build requests in the request queue itself. Instead, we build the request locally (e.g., on the stack) and then as part of the swizzling operation, copy it to the request queue entry we've allocated. I thought long and hard about whether this was too expensive a change to make as it in a lot of cases requires an extra copy. On balance, the flexbility is worth it. With any luck, the entry that we build locally stays in a processor writeback cache (after all, it's only 64 bytes) so that the cost of actually flushing it to the memory area that is the shared queue with the PCI device is not all that expensive. We may examine this again and try to get clever in the future to try and avoid copies. Another change that falls out of this is that MEMORYBARRIER should be taken a lot more seriously. The macro ISP_ADD_REQUEST does a MEMORYBARRIER on the entry being added. But there had been many other places this had been missing. It's now very important that it be done. Additional changes: Fix a longstanding buglet of sorts. When we get an entry via isp_getrqentry, the iptr value that gets returned is the value we intend to eventually plug into the ISP registers as the entry *one past* the last one we've written- *not* the current entry we're updating. All along we've been calling sync functions on the wrong index value. Argh. The 'fix' here is to rename all 'iptr' variables as 'nxti' to remember that this is the 'next' pointer- not the current pointer. Devote a single bit to mboxbsy- and set aside bits for output mbox registers that we need to pick up- we can have at least one command which does not have any defined output registers (MBOX_EXECUTE_FIRMWARE). MFC after: 2 weeks
2001-12-11 00:18:45 +00:00
mbxsync:
switch (rq->req_header.rqs_entry_type) {
case RQSTYPE_REQUEST:
isp_put_request(isp, rq, qep);
break;
case RQSTYPE_CMDONLY:
isp_put_extended_request(isp, (ispextreq_t *)rq,
(ispextreq_t *)qep);
break;
case RQSTYPE_T2RQS:
isp_put_request_t2(isp, (ispreqt2_t *) rq, (ispreqt2_t *) qep);
break;
}
return (CMD_QUEUED);
}
static void
2001-03-04 18:42:51 +00:00
isp_pci_dmateardown(struct ispsoftc *isp, XS_T *xs, u_int16_t handle)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *)isp;
bus_dmamap_t *dp = &pcs->dmaps[isp_handle_index(handle)];
if ((xs->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
bus_dmamap_sync(pcs->dmat, *dp, BUS_DMASYNC_POSTREAD);
} else {
bus_dmamap_sync(pcs->dmat, *dp, BUS_DMASYNC_POSTWRITE);
}
bus_dmamap_unload(pcs->dmat, *dp);
}
static void
1998-12-28 19:24:23 +00:00
isp_pci_reset1(struct ispsoftc *isp)
{
/* Make sure the BIOS is disabled */
isp_pci_wr_reg(isp, HCCR, PCI_HCCR_CMD_BIOS);
/* and enable interrupts */
ENABLE_INTS(isp);
}
static void
isp_pci_dumpregs(struct ispsoftc *isp, const char *msg)
{
struct isp_pcisoftc *pcs = (struct isp_pcisoftc *)isp;
if (msg)
printf("%s: %s\n", device_get_nameunit(isp->isp_dev), msg);
else
printf("%s:\n", device_get_nameunit(isp->isp_dev));
if (IS_SCSI(isp))
printf(" biu_conf1=%x", ISP_READ(isp, BIU_CONF1));
else
printf(" biu_csr=%x", ISP_READ(isp, BIU2100_CSR));
printf(" biu_icr=%x biu_isr=%x biu_sema=%x ", ISP_READ(isp, BIU_ICR),
ISP_READ(isp, BIU_ISR), ISP_READ(isp, BIU_SEMA));
printf("risc_hccr=%x\n", ISP_READ(isp, HCCR));
if (IS_SCSI(isp)) {
ISP_WRITE(isp, HCCR, HCCR_CMD_PAUSE);
printf(" cdma_conf=%x cdma_sts=%x cdma_fifostat=%x\n",
ISP_READ(isp, CDMA_CONF), ISP_READ(isp, CDMA_STATUS),
ISP_READ(isp, CDMA_FIFO_STS));
printf(" ddma_conf=%x ddma_sts=%x ddma_fifostat=%x\n",
ISP_READ(isp, DDMA_CONF), ISP_READ(isp, DDMA_STATUS),
ISP_READ(isp, DDMA_FIFO_STS));
printf(" sxp_int=%x sxp_gross=%x sxp(scsi_ctrl)=%x\n",
ISP_READ(isp, SXP_INTERRUPT),
ISP_READ(isp, SXP_GROSS_ERR),
ISP_READ(isp, SXP_PINS_CTRL));
ISP_WRITE(isp, HCCR, HCCR_CMD_RELEASE);
}
printf(" mbox regs: %x %x %x %x %x\n",
ISP_READ(isp, OUTMAILBOX0), ISP_READ(isp, OUTMAILBOX1),
ISP_READ(isp, OUTMAILBOX2), ISP_READ(isp, OUTMAILBOX3),
ISP_READ(isp, OUTMAILBOX4));
printf(" PCI Status Command/Status=%x\n",
pci_read_config(pcs->pci_dev, PCIR_COMMAND, 1));
}