freebsd-nq/sys/dev/mps/mps.c

1879 lines
53 KiB
C
Raw Normal View History

/*-
* Copyright (c) 2009 Yahoo! Inc.
* All rights reserved.
*
* 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, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* 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/cdefs.h>
__FBSDID("$FreeBSD$");
/* Communications core for LSI MPT2 */
#include <sys/types.h>
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/selinfo.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/module.h>
#include <sys/bus.h>
#include <sys/conf.h>
#include <sys/bio.h>
#include <sys/malloc.h>
#include <sys/uio.h>
#include <sys/sysctl.h>
Fix an event handling bug with the mps(4) driver. This bug manifested itself after repeated device arrivals and departures. The root of the problem was that the last entry in the reply array wasn't initialized/allocated. So every time we got around to that event, we had a bogus address. There were a couple more problems with the code that are also fixed: - The reply mechanism was being treated as sequential (indexed by sc->replycurindex) even though the spec says that the driver should use the ReplyFrameAddress field of the post queue descriptor to figure out where the reply is. There is no guarantee that the reply descriptors will be used in sequential order. - The second word of the reply post queue descriptor wasn't being checked in mps_intr_locked() to make sure that it wasn't 0xffffffff. So the driver could potentially come across a partially DMAed descriptor. - The number of replies allocated was one less than the actual size of the queue. Instead, it was the size of the number of replies that can be used at one time. (Which is one less than the size of the queue.) mps.c: When initializing the entries in the reply free queue, make sure we initialize the full number that we tell the chip we have (sc->fqdepth), not the number that can be used at any one time (sc->num_replies). When allocating replies, make sure we allocate the number of replies that we've told the chip exist, not just the number that can be used simultaneously. Use the ReplyFrameAddress field of the post queue descriptor to figure out which reply is being referenced. This is what the spec says to do, and the spec doesn't guarantee that the replies will be used in order. Put a check in to verify that the reply address passed back from the card is valid. (Panic if it isn't, we'll panic when we try to deference the reply pointer in any case.) In mps_intr_locked(), verify that the second word of the post queue descriptor is not 0xffffffff in addition to verifying that the unused flag is not set, so we can make sure we didn't get a partially DMAed descriptor. Remove references to sc->replycurindex, it isn't needed now. mpsvar.h: Remove replycurindex from the softc, it isn't needed now. Reviewed by: scottl
2010-12-10 21:45:10 +00:00
#include <sys/endian.h>
#include <machine/bus.h>
#include <machine/resource.h>
#include <sys/rman.h>
#include <cam/scsi/scsi_all.h>
#include <dev/mps/mpi/mpi2_type.h>
#include <dev/mps/mpi/mpi2.h>
#include <dev/mps/mpi/mpi2_ioc.h>
#include <dev/mps/mpi/mpi2_cnfg.h>
#include <dev/mps/mpsvar.h>
#include <dev/mps/mps_table.h>
static void mps_startup(void *arg);
static void mps_startup_complete(struct mps_softc *sc, struct mps_command *cm);
static int mps_send_iocinit(struct mps_softc *sc);
static int mps_attach_log(struct mps_softc *sc);
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
static __inline void mps_complete_command(struct mps_command *cm);
static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data, MPI2_EVENT_NOTIFICATION_REPLY *reply);
static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
static void mps_periodic(void *);
SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
/*
* Do a "Diagnostic Reset" aka a hard reset. This should get the chip out of
* any state and back to its initialization state machine.
*/
static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
static int
mps_hard_reset(struct mps_softc *sc)
{
uint32_t reg;
int i, error, tries = 0;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
/* Clear any pending interrupts */
mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
/* Push the magic sequence */
error = ETIMEDOUT;
while (tries++ < 20) {
for (i = 0; i < sizeof(mpt2_reset_magic); i++)
mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
mpt2_reset_magic[i]);
DELAY(100 * 1000);
reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
error = 0;
break;
}
}
if (error)
return (error);
/* Send the actual reset. XXX need to refresh the reg? */
mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET,
reg | MPI2_DIAG_RESET_ADAPTER);
/* Wait up to 300 seconds in 50ms intervals */
error = ETIMEDOUT;
for (i = 0; i < 60000; i++) {
DELAY(50000);
reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
error = 0;
break;
}
}
if (error)
return (error);
mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
return (0);
}
static int
mps_soft_reset(struct mps_softc *sc)
{
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
MPI2_DOORBELL_FUNCTION_SHIFT);
DELAY(50000);
return (0);
}
static int
mps_transition_ready(struct mps_softc *sc)
{
uint32_t reg, state;
int error, tries = 0;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
error = 0;
while (tries++ < 5) {
reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
mps_dprint(sc, MPS_INFO, "Doorbell= 0x%x\n", reg);
/*
* Ensure the IOC is ready to talk. If it's not, try
* resetting it.
*/
if (reg & MPI2_DOORBELL_USED) {
mps_hard_reset(sc);
DELAY(50000);
continue;
}
/* Is the adapter owned by another peer? */
if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
(MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
device_printf(sc->mps_dev, "IOC is under the control "
"of another peer host, aborting initialization.\n");
return (ENXIO);
}
state = reg & MPI2_IOC_STATE_MASK;
if (state == MPI2_IOC_STATE_READY) {
/* Ready to go! */
error = 0;
break;
} else if (state == MPI2_IOC_STATE_FAULT) {
mps_dprint(sc, MPS_INFO, "IOC in fault state 0x%x\n",
state & MPI2_DOORBELL_FAULT_CODE_MASK);
mps_hard_reset(sc);
} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
/* Need to take ownership */
mps_soft_reset(sc);
} else if (state == MPI2_IOC_STATE_RESET) {
/* Wait a bit, IOC might be in transition */
mps_dprint(sc, MPS_FAULT,
"IOC in unexpected reset state\n");
} else {
mps_dprint(sc, MPS_FAULT,
"IOC in unknown state 0x%x\n", state);
error = EINVAL;
break;
}
/* Wait 50ms for things to settle down. */
DELAY(50000);
}
if (error)
device_printf(sc->mps_dev, "Cannot transition IOC to ready\n");
return (error);
}
static int
mps_transition_operational(struct mps_softc *sc)
{
uint32_t reg, state;
int error;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
error = 0;
reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
mps_dprint(sc, MPS_INFO, "Doorbell= 0x%x\n", reg);
state = reg & MPI2_IOC_STATE_MASK;
if (state != MPI2_IOC_STATE_READY) {
if ((error = mps_transition_ready(sc)) != 0)
return (error);
}
error = mps_send_iocinit(sc);
return (error);
}
/* Wait for the chip to ACK a word that we've put into its FIFO */
static int
mps_wait_db_ack(struct mps_softc *sc)
{
int retry;
for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
MPI2_HIS_SYS2IOC_DB_STATUS) == 0)
return (0);
DELAY(2000);
}
return (ETIMEDOUT);
}
/* Wait for the chip to signal that the next word in its FIFO can be fetched */
static int
mps_wait_db_int(struct mps_softc *sc)
{
int retry;
for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
return (0);
DELAY(2000);
}
return (ETIMEDOUT);
}
/* Step through the synchronous command state machine, i.e. "Doorbell mode" */
static int
mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
int req_sz, int reply_sz, int timeout)
{
uint32_t *data32;
uint16_t *data16;
int i, count, ioc_sz, residual;
/* Step 1 */
mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
/* Step 2 */
if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
return (EBUSY);
/* Step 3
* Announce that a message is coming through the doorbell. Messages
* are pushed at 32bit words, so round up if needed.
*/
count = (req_sz + 3) / 4;
mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
(MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
(count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
/* Step 4 */
if (mps_wait_db_int(sc) ||
(mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
return (ENXIO);
}
mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
if (mps_wait_db_ack(sc) != 0) {
mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
return (ENXIO);
}
/* Step 5 */
/* Clock out the message data synchronously in 32-bit dwords*/
data32 = (uint32_t *)req;
for (i = 0; i < count; i++) {
mps_regwrite(sc, MPI2_DOORBELL_OFFSET, data32[i]);
if (mps_wait_db_ack(sc) != 0) {
mps_dprint(sc, MPS_FAULT,
"Timeout while writing doorbell\n");
return (ENXIO);
}
}
/* Step 6 */
/* Clock in the reply in 16-bit words. The total length of the
* message is always in the 4th byte, so clock out the first 2 words
* manually, then loop the rest.
*/
data16 = (uint16_t *)reply;
if (mps_wait_db_int(sc) != 0) {
mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
return (ENXIO);
}
data16[0] =
mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
if (mps_wait_db_int(sc) != 0) {
mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
return (ENXIO);
}
data16[1] =
mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
/* Number of 32bit words in the message */
ioc_sz = reply->MsgLength;
/*
* Figure out how many 16bit words to clock in without overrunning.
* The precision loss with dividing reply_sz can safely be
* ignored because the messages can only be multiples of 32bits.
*/
residual = 0;
count = MIN((reply_sz / 4), ioc_sz) * 2;
if (count < ioc_sz * 2) {
residual = ioc_sz * 2 - count;
mps_dprint(sc, MPS_FAULT, "Driver error, throwing away %d "
"residual message words\n", residual);
}
for (i = 2; i < count; i++) {
if (mps_wait_db_int(sc) != 0) {
mps_dprint(sc, MPS_FAULT,
"Timeout reading doorbell %d\n", i);
return (ENXIO);
}
data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
MPI2_DOORBELL_DATA_MASK;
mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
}
/*
* Pull out residual words that won't fit into the provided buffer.
* This keeps the chip from hanging due to a driver programming
* error.
*/
while (residual--) {
if (mps_wait_db_int(sc) != 0) {
mps_dprint(sc, MPS_FAULT,
"Timeout reading doorbell\n");
return (ENXIO);
}
(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
}
/* Step 7 */
if (mps_wait_db_int(sc) != 0) {
mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
return (ENXIO);
}
if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
return (0);
}
void
mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
{
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
mtx_assert(&sc->mps_mtx, MA_OWNED);
if ((cm->cm_desc.Default.SMID < 1)
|| (cm->cm_desc.Default.SMID >= sc->num_reqs)) {
mps_printf(sc, "%s: invalid SMID %d, desc %#x %#x\n",
__func__, cm->cm_desc.Default.SMID,
cm->cm_desc.Words.High, cm->cm_desc.Words.Low);
}
mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
cm->cm_desc.Words.Low);
mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
cm->cm_desc.Words.High);
}
int
mps_request_polled(struct mps_softc *sc, struct mps_command *cm)
{
int error, timeout = 0;
error = 0;
cm->cm_flags |= MPS_CM_FLAGS_POLLED;
cm->cm_complete = NULL;
mps_map_command(sc, cm);
while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
mps_intr(sc);
DELAY(50 * 1000);
if (timeout++ > 1000) {
mps_dprint(sc, MPS_FAULT, "polling failed\n");
error = ETIMEDOUT;
break;
}
}
return (error);
}
/*
* Just the FACTS, ma'am.
*/
static int
mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
{
MPI2_DEFAULT_REPLY *reply;
MPI2_IOC_FACTS_REQUEST request;
int error, req_sz, reply_sz;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
reply = (MPI2_DEFAULT_REPLY *)facts;
bzero(&request, req_sz);
request.Function = MPI2_FUNCTION_IOC_FACTS;
error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
return (error);
}
static int
mps_get_portfacts(struct mps_softc *sc, MPI2_PORT_FACTS_REPLY *facts, int port)
{
MPI2_PORT_FACTS_REQUEST *request;
MPI2_PORT_FACTS_REPLY *reply;
struct mps_command *cm;
int error;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
if ((cm = mps_alloc_command(sc)) == NULL)
return (EBUSY);
request = (MPI2_PORT_FACTS_REQUEST *)cm->cm_req;
request->Function = MPI2_FUNCTION_PORT_FACTS;
request->PortNumber = port;
cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
cm->cm_data = NULL;
error = mps_request_polled(sc, cm);
reply = (MPI2_PORT_FACTS_REPLY *)cm->cm_reply;
if ((reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
error = ENXIO;
bcopy(reply, facts, sizeof(MPI2_PORT_FACTS_REPLY));
mps_free_command(sc, cm);
return (error);
}
static int
mps_send_iocinit(struct mps_softc *sc)
{
MPI2_IOC_INIT_REQUEST init;
MPI2_DEFAULT_REPLY reply;
int req_sz, reply_sz, error;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
bzero(&init, req_sz);
bzero(&reply, reply_sz);
/*
* Fill in the init block. Note that most addresses are
* deliberately in the lower 32bits of memory. This is a micro-
* optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
*/
init.Function = MPI2_FUNCTION_IOC_INIT;
init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
init.MsgVersion = MPI2_VERSION;
init.HeaderVersion = MPI2_HEADER_VERSION;
init.SystemRequestFrameSize = sc->facts->IOCRequestFrameSize;
init.ReplyDescriptorPostQueueDepth = sc->pqdepth;
init.ReplyFreeQueueDepth = sc->fqdepth;
init.SenseBufferAddressHigh = 0;
init.SystemReplyAddressHigh = 0;
init.SystemRequestFrameBaseAddress.High = 0;
init.SystemRequestFrameBaseAddress.Low = (uint32_t)sc->req_busaddr;
init.ReplyDescriptorPostQueueAddress.High = 0;
init.ReplyDescriptorPostQueueAddress.Low = (uint32_t)sc->post_busaddr;
init.ReplyFreeQueueAddress.High = 0;
init.ReplyFreeQueueAddress.Low = (uint32_t)sc->free_busaddr;
init.TimeStamp.High = 0;
init.TimeStamp.Low = (uint32_t)time_uptime;
error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
error = ENXIO;
mps_dprint(sc, MPS_INFO, "IOCInit status= 0x%x\n", reply.IOCStatus);
return (error);
}
static int
mps_send_portenable(struct mps_softc *sc)
{
MPI2_PORT_ENABLE_REQUEST *request;
struct mps_command *cm;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
if ((cm = mps_alloc_command(sc)) == NULL)
return (EBUSY);
request = (MPI2_PORT_ENABLE_REQUEST *)cm->cm_req;
request->Function = MPI2_FUNCTION_PORT_ENABLE;
request->MsgFlags = 0;
request->VP_ID = 0;
cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
cm->cm_complete = mps_startup_complete;
mps_enqueue_request(sc, cm);
return (0);
}
static int
mps_send_mur(struct mps_softc *sc)
{
/* Placeholder */
return (0);
}
void
mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
{
bus_addr_t *addr;
addr = arg;
*addr = segs[0].ds_addr;
}
static int
mps_alloc_queues(struct mps_softc *sc)
{
bus_addr_t queues_busaddr;
uint8_t *queues;
int qsize, fqsize, pqsize;
/*
* The reply free queue contains 4 byte entries in multiples of 16 and
* aligned on a 16 byte boundary. There must always be an unused entry.
* This queue supplies fresh reply frames for the firmware to use.
*
* The reply descriptor post queue contains 8 byte entries in
* multiples of 16 and aligned on a 16 byte boundary. This queue
* contains filled-in reply frames sent from the firmware to the host.
*
* These two queues are allocated together for simplicity.
*/
sc->fqdepth = roundup2((sc->num_replies + 1), 16);
sc->pqdepth = roundup2((sc->num_replies + 1), 16);
fqsize= sc->fqdepth * 4;
pqsize = sc->pqdepth * 8;
qsize = fqsize + pqsize;
if (bus_dma_tag_create( sc->mps_parent_dmat, /* parent */
16, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
qsize, /* maxsize */
1, /* nsegments */
qsize, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
&sc->queues_dmat)) {
device_printf(sc->mps_dev, "Cannot allocate queues DMA tag\n");
return (ENOMEM);
}
if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
&sc->queues_map)) {
device_printf(sc->mps_dev, "Cannot allocate queues memory\n");
return (ENOMEM);
}
bzero(queues, qsize);
bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
mps_memaddr_cb, &queues_busaddr, 0);
sc->free_queue = (uint32_t *)queues;
sc->free_busaddr = queues_busaddr;
sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
sc->post_busaddr = queues_busaddr + fqsize;
return (0);
}
static int
mps_alloc_replies(struct mps_softc *sc)
{
Fix an event handling bug with the mps(4) driver. This bug manifested itself after repeated device arrivals and departures. The root of the problem was that the last entry in the reply array wasn't initialized/allocated. So every time we got around to that event, we had a bogus address. There were a couple more problems with the code that are also fixed: - The reply mechanism was being treated as sequential (indexed by sc->replycurindex) even though the spec says that the driver should use the ReplyFrameAddress field of the post queue descriptor to figure out where the reply is. There is no guarantee that the reply descriptors will be used in sequential order. - The second word of the reply post queue descriptor wasn't being checked in mps_intr_locked() to make sure that it wasn't 0xffffffff. So the driver could potentially come across a partially DMAed descriptor. - The number of replies allocated was one less than the actual size of the queue. Instead, it was the size of the number of replies that can be used at one time. (Which is one less than the size of the queue.) mps.c: When initializing the entries in the reply free queue, make sure we initialize the full number that we tell the chip we have (sc->fqdepth), not the number that can be used at any one time (sc->num_replies). When allocating replies, make sure we allocate the number of replies that we've told the chip exist, not just the number that can be used simultaneously. Use the ReplyFrameAddress field of the post queue descriptor to figure out which reply is being referenced. This is what the spec says to do, and the spec doesn't guarantee that the replies will be used in order. Put a check in to verify that the reply address passed back from the card is valid. (Panic if it isn't, we'll panic when we try to deference the reply pointer in any case.) In mps_intr_locked(), verify that the second word of the post queue descriptor is not 0xffffffff in addition to verifying that the unused flag is not set, so we can make sure we didn't get a partially DMAed descriptor. Remove references to sc->replycurindex, it isn't needed now. mpsvar.h: Remove replycurindex from the softc, it isn't needed now. Reviewed by: scottl
2010-12-10 21:45:10 +00:00
int rsize, num_replies;
Fix an event handling bug with the mps(4) driver. This bug manifested itself after repeated device arrivals and departures. The root of the problem was that the last entry in the reply array wasn't initialized/allocated. So every time we got around to that event, we had a bogus address. There were a couple more problems with the code that are also fixed: - The reply mechanism was being treated as sequential (indexed by sc->replycurindex) even though the spec says that the driver should use the ReplyFrameAddress field of the post queue descriptor to figure out where the reply is. There is no guarantee that the reply descriptors will be used in sequential order. - The second word of the reply post queue descriptor wasn't being checked in mps_intr_locked() to make sure that it wasn't 0xffffffff. So the driver could potentially come across a partially DMAed descriptor. - The number of replies allocated was one less than the actual size of the queue. Instead, it was the size of the number of replies that can be used at one time. (Which is one less than the size of the queue.) mps.c: When initializing the entries in the reply free queue, make sure we initialize the full number that we tell the chip we have (sc->fqdepth), not the number that can be used at any one time (sc->num_replies). When allocating replies, make sure we allocate the number of replies that we've told the chip exist, not just the number that can be used simultaneously. Use the ReplyFrameAddress field of the post queue descriptor to figure out which reply is being referenced. This is what the spec says to do, and the spec doesn't guarantee that the replies will be used in order. Put a check in to verify that the reply address passed back from the card is valid. (Panic if it isn't, we'll panic when we try to deference the reply pointer in any case.) In mps_intr_locked(), verify that the second word of the post queue descriptor is not 0xffffffff in addition to verifying that the unused flag is not set, so we can make sure we didn't get a partially DMAed descriptor. Remove references to sc->replycurindex, it isn't needed now. mpsvar.h: Remove replycurindex from the softc, it isn't needed now. Reviewed by: scottl
2010-12-10 21:45:10 +00:00
/*
* sc->num_replies should be one less than sc->fqdepth. We need to
* allocate space for sc->fqdepth replies, but only sc->num_replies
* replies can be used at once.
*/
num_replies = max(sc->fqdepth, sc->num_replies);
rsize = sc->facts->ReplyFrameSize * num_replies * 4;
if (bus_dma_tag_create( sc->mps_parent_dmat, /* parent */
4, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
rsize, /* maxsize */
1, /* nsegments */
rsize, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
&sc->reply_dmat)) {
device_printf(sc->mps_dev, "Cannot allocate replies DMA tag\n");
return (ENOMEM);
}
if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
BUS_DMA_NOWAIT, &sc->reply_map)) {
device_printf(sc->mps_dev, "Cannot allocate replies memory\n");
return (ENOMEM);
}
bzero(sc->reply_frames, rsize);
bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
mps_memaddr_cb, &sc->reply_busaddr, 0);
return (0);
}
static int
mps_alloc_requests(struct mps_softc *sc)
{
struct mps_command *cm;
struct mps_chain *chain;
int i, rsize, nsegs;
rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
if (bus_dma_tag_create( sc->mps_parent_dmat, /* parent */
16, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
rsize, /* maxsize */
1, /* nsegments */
rsize, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
&sc->req_dmat)) {
device_printf(sc->mps_dev, "Cannot allocate request DMA tag\n");
return (ENOMEM);
}
if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
BUS_DMA_NOWAIT, &sc->req_map)) {
device_printf(sc->mps_dev, "Cannot allocate request memory\n");
return (ENOMEM);
}
bzero(sc->req_frames, rsize);
bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
mps_memaddr_cb, &sc->req_busaddr, 0);
rsize = sc->facts->IOCRequestFrameSize * MPS_CHAIN_FRAMES * 4;
if (bus_dma_tag_create( sc->mps_parent_dmat, /* parent */
16, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
rsize, /* maxsize */
1, /* nsegments */
rsize, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
&sc->chain_dmat)) {
device_printf(sc->mps_dev, "Cannot allocate chain DMA tag\n");
return (ENOMEM);
}
if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
BUS_DMA_NOWAIT, &sc->chain_map)) {
device_printf(sc->mps_dev, "Cannot allocate chain memory\n");
return (ENOMEM);
}
bzero(sc->chain_frames, rsize);
bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
mps_memaddr_cb, &sc->chain_busaddr, 0);
rsize = MPS_SENSE_LEN * sc->num_reqs;
if (bus_dma_tag_create( sc->mps_parent_dmat, /* parent */
1, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
rsize, /* maxsize */
1, /* nsegments */
rsize, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
&sc->sense_dmat)) {
device_printf(sc->mps_dev, "Cannot allocate sense DMA tag\n");
return (ENOMEM);
}
if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
BUS_DMA_NOWAIT, &sc->sense_map)) {
device_printf(sc->mps_dev, "Cannot allocate sense memory\n");
return (ENOMEM);
}
bzero(sc->sense_frames, rsize);
bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
mps_memaddr_cb, &sc->sense_busaddr, 0);
sc->chains = malloc(sizeof(struct mps_chain) * MPS_CHAIN_FRAMES,
M_MPT2, M_WAITOK | M_ZERO);
for (i = 0; i < MPS_CHAIN_FRAMES; i++) {
chain = &sc->chains[i];
chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
i * sc->facts->IOCRequestFrameSize * 4);
chain->chain_busaddr = sc->chain_busaddr +
i * sc->facts->IOCRequestFrameSize * 4;
mps_free_chain(sc, chain);
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
sc->chain_free_lowwater++;
}
/* XXX Need to pick a more precise value */
nsegs = (MAXPHYS / PAGE_SIZE) + 1;
if (bus_dma_tag_create( sc->mps_parent_dmat, /* parent */
1, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR, /* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
nsegs, /* nsegments */
BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
BUS_DMA_ALLOCNOW, /* flags */
busdma_lock_mutex, /* lockfunc */
&sc->mps_mtx, /* lockarg */
&sc->buffer_dmat)) {
device_printf(sc->mps_dev, "Cannot allocate sense DMA tag\n");
return (ENOMEM);
}
/*
* SMID 0 cannot be used as a free command per the firmware spec.
* Just drop that command instead of risking accounting bugs.
*/
sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
M_MPT2, M_WAITOK | M_ZERO);
for (i = 1; i < sc->num_reqs; i++) {
cm = &sc->commands[i];
cm->cm_req = sc->req_frames +
i * sc->facts->IOCRequestFrameSize * 4;
cm->cm_req_busaddr = sc->req_busaddr +
i * sc->facts->IOCRequestFrameSize * 4;
cm->cm_sense = &sc->sense_frames[i];
cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
cm->cm_desc.Default.SMID = i;
cm->cm_sc = sc;
TAILQ_INIT(&cm->cm_chain_list);
callout_init(&cm->cm_callout, 1 /*MPSAFE*/);
/* XXX Is a failure here a critical problem? */
if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
mps_free_command(sc, cm);
else {
sc->num_reqs = i;
break;
}
}
return (0);
}
static int
mps_init_queues(struct mps_softc *sc)
{
int i;
memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
Fix an event handling bug with the mps(4) driver. This bug manifested itself after repeated device arrivals and departures. The root of the problem was that the last entry in the reply array wasn't initialized/allocated. So every time we got around to that event, we had a bogus address. There were a couple more problems with the code that are also fixed: - The reply mechanism was being treated as sequential (indexed by sc->replycurindex) even though the spec says that the driver should use the ReplyFrameAddress field of the post queue descriptor to figure out where the reply is. There is no guarantee that the reply descriptors will be used in sequential order. - The second word of the reply post queue descriptor wasn't being checked in mps_intr_locked() to make sure that it wasn't 0xffffffff. So the driver could potentially come across a partially DMAed descriptor. - The number of replies allocated was one less than the actual size of the queue. Instead, it was the size of the number of replies that can be used at one time. (Which is one less than the size of the queue.) mps.c: When initializing the entries in the reply free queue, make sure we initialize the full number that we tell the chip we have (sc->fqdepth), not the number that can be used at any one time (sc->num_replies). When allocating replies, make sure we allocate the number of replies that we've told the chip exist, not just the number that can be used simultaneously. Use the ReplyFrameAddress field of the post queue descriptor to figure out which reply is being referenced. This is what the spec says to do, and the spec doesn't guarantee that the replies will be used in order. Put a check in to verify that the reply address passed back from the card is valid. (Panic if it isn't, we'll panic when we try to deference the reply pointer in any case.) In mps_intr_locked(), verify that the second word of the post queue descriptor is not 0xffffffff in addition to verifying that the unused flag is not set, so we can make sure we didn't get a partially DMAed descriptor. Remove references to sc->replycurindex, it isn't needed now. mpsvar.h: Remove replycurindex from the softc, it isn't needed now. Reviewed by: scottl
2010-12-10 21:45:10 +00:00
/*
* According to the spec, we need to use one less reply than we
* have space for on the queue. So sc->num_replies (the number we
* use) should be less than sc->fqdepth (allocated size).
*/
if (sc->num_replies >= sc->fqdepth)
return (EINVAL);
Fix an event handling bug with the mps(4) driver. This bug manifested itself after repeated device arrivals and departures. The root of the problem was that the last entry in the reply array wasn't initialized/allocated. So every time we got around to that event, we had a bogus address. There were a couple more problems with the code that are also fixed: - The reply mechanism was being treated as sequential (indexed by sc->replycurindex) even though the spec says that the driver should use the ReplyFrameAddress field of the post queue descriptor to figure out where the reply is. There is no guarantee that the reply descriptors will be used in sequential order. - The second word of the reply post queue descriptor wasn't being checked in mps_intr_locked() to make sure that it wasn't 0xffffffff. So the driver could potentially come across a partially DMAed descriptor. - The number of replies allocated was one less than the actual size of the queue. Instead, it was the size of the number of replies that can be used at one time. (Which is one less than the size of the queue.) mps.c: When initializing the entries in the reply free queue, make sure we initialize the full number that we tell the chip we have (sc->fqdepth), not the number that can be used at any one time (sc->num_replies). When allocating replies, make sure we allocate the number of replies that we've told the chip exist, not just the number that can be used simultaneously. Use the ReplyFrameAddress field of the post queue descriptor to figure out which reply is being referenced. This is what the spec says to do, and the spec doesn't guarantee that the replies will be used in order. Put a check in to verify that the reply address passed back from the card is valid. (Panic if it isn't, we'll panic when we try to deference the reply pointer in any case.) In mps_intr_locked(), verify that the second word of the post queue descriptor is not 0xffffffff in addition to verifying that the unused flag is not set, so we can make sure we didn't get a partially DMAed descriptor. Remove references to sc->replycurindex, it isn't needed now. mpsvar.h: Remove replycurindex from the softc, it isn't needed now. Reviewed by: scottl
2010-12-10 21:45:10 +00:00
/*
* Initialize all of the free queue entries.
*/
for (i = 0; i < sc->fqdepth; i++)
sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
sc->replyfreeindex = sc->num_replies;
return (0);
}
int
mps_attach(struct mps_softc *sc)
{
int i, error;
char tmpstr[80], tmpstr2[80];
/*
* Grab any tunable-set debug level so that tracing works as early
* as possible.
*/
snprintf(tmpstr, sizeof(tmpstr), "hw.mps.%d.debug_level",
device_get_unit(sc->mps_dev));
TUNABLE_INT_FETCH(tmpstr, &sc->mps_debug);
snprintf(tmpstr, sizeof(tmpstr), "hw.mps.%d.allow_multiple_tm_cmds",
device_get_unit(sc->mps_dev));
TUNABLE_INT_FETCH(tmpstr, &sc->allow_multiple_tm_cmds);
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
TAILQ_INIT(&sc->event_list);
/*
* Setup the sysctl variable so the user can change the debug level
* on the fly.
*/
snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
device_get_unit(sc->mps_dev));
snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
sysctl_ctx_init(&sc->sysctl_ctx);
sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2, CTLFLAG_RD,
0, tmpstr);
if (sc->sysctl_tree == NULL)
return (ENOMEM);
SYSCTL_ADD_UINT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mps_debug, 0,
"mps debug level");
SYSCTL_ADD_UINT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
OID_AUTO, "allow_multiple_tm_cmds", CTLFLAG_RW,
&sc->allow_multiple_tm_cmds, 0,
"allow multiple simultaneous task management cmds");
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
SYSCTL_ADD_INT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
OID_AUTO, "io_cmds_active", CTLFLAG_RD,
&sc->io_cmds_active, 0, "number of currently active commands");
SYSCTL_ADD_INT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
&sc->io_cmds_highwater, 0, "maximum active commands seen");
SYSCTL_ADD_INT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
OID_AUTO, "chain_free", CTLFLAG_RD,
&sc->chain_free, 0, "number of free chain elements");
SYSCTL_ADD_INT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
&sc->chain_free_lowwater, 0,"lowest number of free chain elements");
SYSCTL_ADD_UQUAD(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
&sc->chain_alloc_fail, "chain allocation failures");
if ((error = mps_transition_ready(sc)) != 0)
return (error);
sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
M_ZERO|M_NOWAIT);
if ((error = mps_get_iocfacts(sc, sc->facts)) != 0)
return (error);
mps_print_iocfacts(sc, sc->facts);
mps_printf(sc, "Firmware: %02d.%02d.%02d.%02d\n",
sc->facts->FWVersion.Struct.Major,
sc->facts->FWVersion.Struct.Minor,
sc->facts->FWVersion.Struct.Unit,
sc->facts->FWVersion.Struct.Dev);
mps_printf(sc, "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
"\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
"\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
"\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
/*
* If the chip doesn't support event replay then a hard reset will be
* required to trigger a full discovery. Do the reset here then
* retransition to Ready. A hard reset might have already been done,
* but it doesn't hurt to do it again.
*/
if ((sc->facts->IOCCapabilities &
MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0) {
mps_hard_reset(sc);
if ((error = mps_transition_ready(sc)) != 0)
return (error);
}
/*
* Size the queues. Since the reply queues always need one free entry,
* we'll just deduct one reply message here.
*/
sc->num_reqs = MIN(MPS_REQ_FRAMES, sc->facts->RequestCredit);
sc->num_replies = MIN(MPS_REPLY_FRAMES + MPS_EVT_REPLY_FRAMES,
sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
mps_dprint(sc, MPS_INFO, "num_reqs %d, num_replies %d\n", sc->num_reqs,
sc->num_replies);
TAILQ_INIT(&sc->req_list);
TAILQ_INIT(&sc->chain_list);
TAILQ_INIT(&sc->tm_list);
TAILQ_INIT(&sc->io_list);
if (((error = mps_alloc_queues(sc)) != 0) ||
((error = mps_alloc_replies(sc)) != 0) ||
((error = mps_alloc_requests(sc)) != 0)) {
mps_free(sc);
return (error);
}
if (((error = mps_init_queues(sc)) != 0) ||
((error = mps_transition_operational(sc)) != 0)) {
mps_free(sc);
return (error);
}
/*
* Finish the queue initialization.
* These are set here instead of in mps_init_queues() because the
* IOC resets these values during the state transition in
* mps_transition_operational(). The free index is set to 1
* because the corresponding index in the IOC is set to 0, and the
* IOC treats the queues as full if both are set to the same value.
* Hence the reason that the queue can't hold all of the possible
* replies.
*/
sc->replypostindex = 0;
mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
sc->pfacts = malloc(sizeof(MPI2_PORT_FACTS_REPLY) *
sc->facts->NumberOfPorts, M_MPT2, M_ZERO|M_WAITOK);
for (i = 0; i < sc->facts->NumberOfPorts; i++) {
if ((error = mps_get_portfacts(sc, &sc->pfacts[i], i)) != 0) {
mps_free(sc);
return (error);
}
mps_print_portfacts(sc, &sc->pfacts[i]);
}
/* Attach the subsystems so they can prepare their event masks. */
/* XXX Should be dynamic so that IM/IR and user modules can attach */
if (((error = mps_attach_log(sc)) != 0) ||
((error = mps_attach_sas(sc)) != 0) ||
((error = mps_attach_user(sc)) != 0)) {
mps_printf(sc, "%s failed to attach all subsystems: error %d\n",
__func__, error);
mps_free(sc);
return (error);
}
if ((error = mps_pci_setup_interrupts(sc)) != 0) {
mps_free(sc);
return (error);
}
/* Start the periodic watchdog check on the IOC Doorbell */
mps_periodic(sc);
/*
* The portenable will kick off discovery events that will drive the
* rest of the initialization process. The CAM/SAS module will
* hold up the boot sequence until discovery is complete.
*/
sc->mps_ich.ich_func = mps_startup;
sc->mps_ich.ich_arg = sc;
if (config_intrhook_establish(&sc->mps_ich) != 0) {
mps_dprint(sc, MPS_FAULT, "Cannot establish MPS config hook\n");
error = EINVAL;
}
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
return (error);
}
static void
mps_startup(void *arg)
{
struct mps_softc *sc;
sc = (struct mps_softc *)arg;
mps_lock(sc);
mps_unmask_intr(sc);
mps_send_portenable(sc);
mps_unlock(sc);
}
/* Periodic watchdog. Is called with the driver lock already held. */
static void
mps_periodic(void *arg)
{
struct mps_softc *sc;
uint32_t db;
sc = (struct mps_softc *)arg;
if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
return;
db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
device_printf(sc->mps_dev, "IOC Fault 0x%08x, Resetting\n", db);
/* XXX Need to broaden this to re-initialize the chip */
mps_hard_reset(sc);
db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
device_printf(sc->mps_dev, "Second IOC Fault 0x%08x, "
"Giving up!\n", db);
return;
}
}
callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
}
static void
mps_startup_complete(struct mps_softc *sc, struct mps_command *cm)
{
MPI2_PORT_ENABLE_REPLY *reply;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
reply = (MPI2_PORT_ENABLE_REPLY *)cm->cm_reply;
if ((reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
mps_dprint(sc, MPS_FAULT, "Portenable failed\n");
mps_free_command(sc, cm);
config_intrhook_disestablish(&sc->mps_ich);
}
static void
mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
MPI2_EVENT_NOTIFICATION_REPLY *event)
{
MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
mps_print_event(sc, event);
switch (event->Event) {
case MPI2_EVENT_LOG_DATA:
device_printf(sc->mps_dev, "MPI2_EVENT_LOG_DATA:\n");
hexdump(event->EventData, event->EventDataLength, NULL, 0);
break;
case MPI2_EVENT_LOG_ENTRY_ADDED:
entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
mps_dprint(sc, MPS_INFO, "MPI2_EVENT_LOG_ENTRY_ADDED event "
"0x%x Sequence %d:\n", entry->LogEntryQualifier,
entry->LogSequence);
break;
default:
break;
}
return;
}
static int
mps_attach_log(struct mps_softc *sc)
{
uint8_t events[16];
bzero(events, 16);
setbit(events, MPI2_EVENT_LOG_DATA);
setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
mps_register_events(sc, events, mps_log_evt_handler, NULL,
&sc->mps_log_eh);
return (0);
}
static int
mps_detach_log(struct mps_softc *sc)
{
if (sc->mps_log_eh != NULL)
mps_deregister_events(sc, sc->mps_log_eh);
return (0);
}
/*
* Free all of the driver resources and detach submodules. Should be called
* without the lock held.
*/
int
mps_free(struct mps_softc *sc)
{
struct mps_command *cm;
int i, error;
/* Turn off the watchdog */
mps_lock(sc);
sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
mps_unlock(sc);
/* Lock must not be held for this */
callout_drain(&sc->periodic);
if (((error = mps_detach_log(sc)) != 0) ||
((error = mps_detach_sas(sc)) != 0))
return (error);
/* Put the IOC back in the READY state. */
mps_lock(sc);
if ((error = mps_send_mur(sc)) != 0) {
mps_unlock(sc);
return (error);
}
mps_unlock(sc);
if (sc->facts != NULL)
free(sc->facts, M_MPT2);
if (sc->pfacts != NULL)
free(sc->pfacts, M_MPT2);
if (sc->post_busaddr != 0)
bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
if (sc->post_queue != NULL)
bus_dmamem_free(sc->queues_dmat, sc->post_queue,
sc->queues_map);
if (sc->queues_dmat != NULL)
bus_dma_tag_destroy(sc->queues_dmat);
if (sc->chain_busaddr != 0)
bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
if (sc->chain_frames != NULL)
bus_dmamem_free(sc->chain_dmat, sc->chain_frames,sc->chain_map);
if (sc->chain_dmat != NULL)
bus_dma_tag_destroy(sc->chain_dmat);
if (sc->sense_busaddr != 0)
bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
if (sc->sense_frames != NULL)
bus_dmamem_free(sc->sense_dmat, sc->sense_frames,sc->sense_map);
if (sc->sense_dmat != NULL)
bus_dma_tag_destroy(sc->sense_dmat);
if (sc->reply_busaddr != 0)
bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
if (sc->reply_frames != NULL)
bus_dmamem_free(sc->reply_dmat, sc->reply_frames,sc->reply_map);
if (sc->reply_dmat != NULL)
bus_dma_tag_destroy(sc->reply_dmat);
if (sc->req_busaddr != 0)
bus_dmamap_unload(sc->req_dmat, sc->req_map);
if (sc->req_frames != NULL)
bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
if (sc->req_dmat != NULL)
bus_dma_tag_destroy(sc->req_dmat);
if (sc->chains != NULL)
free(sc->chains, M_MPT2);
if (sc->commands != NULL) {
for (i = 1; i < sc->num_reqs; i++) {
cm = &sc->commands[i];
bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
}
free(sc->commands, M_MPT2);
}
if (sc->buffer_dmat != NULL)
bus_dma_tag_destroy(sc->buffer_dmat);
if (sc->sysctl_tree != NULL)
sysctl_ctx_free(&sc->sysctl_ctx);
mtx_destroy(&sc->mps_mtx);
return (0);
}
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
static __inline void
mps_complete_command(struct mps_command *cm)
{
if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
if (cm->cm_complete != NULL)
cm->cm_complete(cm->cm_sc, cm);
if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
mps_dprint(cm->cm_sc, MPS_TRACE, "%s: waking up %p\n",
__func__, cm);
wakeup(cm);
}
}
void
mps_intr(void *data)
{
struct mps_softc *sc;
uint32_t status;
sc = (struct mps_softc *)data;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
/*
* Check interrupt status register to flush the bus. This is
* needed for both INTx interrupts and driver-driven polling
*/
status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
return;
mps_lock(sc);
mps_intr_locked(data);
mps_unlock(sc);
return;
}
/*
* In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
* chip. Hopefully this theory is correct.
*/
void
mps_intr_msi(void *data)
{
struct mps_softc *sc;
sc = (struct mps_softc *)data;
mps_lock(sc);
mps_intr_locked(data);
mps_unlock(sc);
return;
}
/*
* The locking is overly broad and simplistic, but easy to deal with for now.
*/
void
mps_intr_locked(void *data)
{
MPI2_REPLY_DESCRIPTORS_UNION *desc;
struct mps_softc *sc;
struct mps_command *cm = NULL;
uint8_t flags;
u_int pq;
sc = (struct mps_softc *)data;
pq = sc->replypostindex;
for ( ;; ) {
cm = NULL;
desc = &sc->post_queue[pq];
flags = desc->Default.ReplyFlags &
MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
Fix an event handling bug with the mps(4) driver. This bug manifested itself after repeated device arrivals and departures. The root of the problem was that the last entry in the reply array wasn't initialized/allocated. So every time we got around to that event, we had a bogus address. There were a couple more problems with the code that are also fixed: - The reply mechanism was being treated as sequential (indexed by sc->replycurindex) even though the spec says that the driver should use the ReplyFrameAddress field of the post queue descriptor to figure out where the reply is. There is no guarantee that the reply descriptors will be used in sequential order. - The second word of the reply post queue descriptor wasn't being checked in mps_intr_locked() to make sure that it wasn't 0xffffffff. So the driver could potentially come across a partially DMAed descriptor. - The number of replies allocated was one less than the actual size of the queue. Instead, it was the size of the number of replies that can be used at one time. (Which is one less than the size of the queue.) mps.c: When initializing the entries in the reply free queue, make sure we initialize the full number that we tell the chip we have (sc->fqdepth), not the number that can be used at any one time (sc->num_replies). When allocating replies, make sure we allocate the number of replies that we've told the chip exist, not just the number that can be used simultaneously. Use the ReplyFrameAddress field of the post queue descriptor to figure out which reply is being referenced. This is what the spec says to do, and the spec doesn't guarantee that the replies will be used in order. Put a check in to verify that the reply address passed back from the card is valid. (Panic if it isn't, we'll panic when we try to deference the reply pointer in any case.) In mps_intr_locked(), verify that the second word of the post queue descriptor is not 0xffffffff in addition to verifying that the unused flag is not set, so we can make sure we didn't get a partially DMAed descriptor. Remove references to sc->replycurindex, it isn't needed now. mpsvar.h: Remove replycurindex from the softc, it isn't needed now. Reviewed by: scottl
2010-12-10 21:45:10 +00:00
if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
|| (desc->Words.High == 0xffffffff))
break;
switch (flags) {
case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
cm = &sc->commands[desc->SCSIIOSuccess.SMID];
cm->cm_reply = NULL;
break;
case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
{
uint32_t baddr;
uint8_t *reply;
Fix an event handling bug with the mps(4) driver. This bug manifested itself after repeated device arrivals and departures. The root of the problem was that the last entry in the reply array wasn't initialized/allocated. So every time we got around to that event, we had a bogus address. There were a couple more problems with the code that are also fixed: - The reply mechanism was being treated as sequential (indexed by sc->replycurindex) even though the spec says that the driver should use the ReplyFrameAddress field of the post queue descriptor to figure out where the reply is. There is no guarantee that the reply descriptors will be used in sequential order. - The second word of the reply post queue descriptor wasn't being checked in mps_intr_locked() to make sure that it wasn't 0xffffffff. So the driver could potentially come across a partially DMAed descriptor. - The number of replies allocated was one less than the actual size of the queue. Instead, it was the size of the number of replies that can be used at one time. (Which is one less than the size of the queue.) mps.c: When initializing the entries in the reply free queue, make sure we initialize the full number that we tell the chip we have (sc->fqdepth), not the number that can be used at any one time (sc->num_replies). When allocating replies, make sure we allocate the number of replies that we've told the chip exist, not just the number that can be used simultaneously. Use the ReplyFrameAddress field of the post queue descriptor to figure out which reply is being referenced. This is what the spec says to do, and the spec doesn't guarantee that the replies will be used in order. Put a check in to verify that the reply address passed back from the card is valid. (Panic if it isn't, we'll panic when we try to deference the reply pointer in any case.) In mps_intr_locked(), verify that the second word of the post queue descriptor is not 0xffffffff in addition to verifying that the unused flag is not set, so we can make sure we didn't get a partially DMAed descriptor. Remove references to sc->replycurindex, it isn't needed now. mpsvar.h: Remove replycurindex from the softc, it isn't needed now. Reviewed by: scottl
2010-12-10 21:45:10 +00:00
/*
* Re-compose the reply address from the address
* sent back from the chip. The ReplyFrameAddress
* is the lower 32 bits of the physical address of
* particular reply frame. Convert that address to
* host format, and then use that to provide the
* offset against the virtual address base
* (sc->reply_frames).
*/
baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
reply = sc->reply_frames +
Fix an event handling bug with the mps(4) driver. This bug manifested itself after repeated device arrivals and departures. The root of the problem was that the last entry in the reply array wasn't initialized/allocated. So every time we got around to that event, we had a bogus address. There were a couple more problems with the code that are also fixed: - The reply mechanism was being treated as sequential (indexed by sc->replycurindex) even though the spec says that the driver should use the ReplyFrameAddress field of the post queue descriptor to figure out where the reply is. There is no guarantee that the reply descriptors will be used in sequential order. - The second word of the reply post queue descriptor wasn't being checked in mps_intr_locked() to make sure that it wasn't 0xffffffff. So the driver could potentially come across a partially DMAed descriptor. - The number of replies allocated was one less than the actual size of the queue. Instead, it was the size of the number of replies that can be used at one time. (Which is one less than the size of the queue.) mps.c: When initializing the entries in the reply free queue, make sure we initialize the full number that we tell the chip we have (sc->fqdepth), not the number that can be used at any one time (sc->num_replies). When allocating replies, make sure we allocate the number of replies that we've told the chip exist, not just the number that can be used simultaneously. Use the ReplyFrameAddress field of the post queue descriptor to figure out which reply is being referenced. This is what the spec says to do, and the spec doesn't guarantee that the replies will be used in order. Put a check in to verify that the reply address passed back from the card is valid. (Panic if it isn't, we'll panic when we try to deference the reply pointer in any case.) In mps_intr_locked(), verify that the second word of the post queue descriptor is not 0xffffffff in addition to verifying that the unused flag is not set, so we can make sure we didn't get a partially DMAed descriptor. Remove references to sc->replycurindex, it isn't needed now. mpsvar.h: Remove replycurindex from the softc, it isn't needed now. Reviewed by: scottl
2010-12-10 21:45:10 +00:00
(baddr - ((uint32_t)sc->reply_busaddr));
/*
* Make sure the reply we got back is in a valid
* range. If not, go ahead and panic here, since
* we'll probably panic as soon as we deference the
* reply pointer anyway.
*/
if ((reply < sc->reply_frames)
|| (reply > (sc->reply_frames +
(sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
printf("%s: WARNING: reply %p out of range!\n",
__func__, reply);
printf("%s: reply_frames %p, fqdepth %d, "
"frame size %d\n", __func__,
sc->reply_frames, sc->fqdepth,
sc->facts->ReplyFrameSize * 4);
printf("%s: baddr %#x,\n", __func__, baddr);
panic("Reply address out of range");
}
if (desc->AddressReply.SMID == 0) {
mps_dispatch_event(sc, baddr,
(MPI2_EVENT_NOTIFICATION_REPLY *) reply);
} else {
cm = &sc->commands[desc->AddressReply.SMID];
cm->cm_reply = reply;
cm->cm_reply_data =
desc->AddressReply.ReplyFrameAddress;
}
break;
}
case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
default:
/* Unhandled */
device_printf(sc->mps_dev, "Unhandled reply 0x%x\n",
desc->Default.ReplyFlags);
cm = NULL;
break;
}
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
if (cm != NULL)
mps_complete_command(cm);
desc->Words.Low = 0xffffffff;
desc->Words.High = 0xffffffff;
if (++pq >= sc->pqdepth)
pq = 0;
}
if (pq != sc->replypostindex) {
mps_dprint(sc, MPS_INFO, "writing postindex %d\n", pq);
mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, pq);
sc->replypostindex = pq;
}
return;
}
static void
mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
MPI2_EVENT_NOTIFICATION_REPLY *reply)
{
struct mps_event_handle *eh;
2010-12-06 10:24:06 +00:00
int event, handled = 0;
event = reply->Event;
TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
if (isset(eh->mask, event)) {
eh->callback(sc, data, reply);
handled++;
}
}
if (handled == 0)
device_printf(sc->mps_dev, "Unhandled event 0x%x\n", event);
}
/*
* For both register_events and update_events, the caller supplies a bitmap
* of events that it _wants_. These functions then turn that into a bitmask
* suitable for the controller.
*/
int
mps_register_events(struct mps_softc *sc, uint8_t *mask,
mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
{
struct mps_event_handle *eh;
int error = 0;
eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
eh->callback = cb;
eh->data = data;
TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
if (mask != NULL)
error = mps_update_events(sc, eh, mask);
*handle = eh;
return (error);
}
int
mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
uint8_t *mask)
{
MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
MPI2_EVENT_NOTIFICATION_REPLY *reply;
struct mps_command *cm;
struct mps_event_handle *eh;
int error, i;
mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
if ((mask != NULL) && (handle != NULL))
bcopy(mask, &handle->mask[0], 16);
memset(sc->event_mask, 0xff, 16);
TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
for (i = 0; i < 16; i++)
sc->event_mask[i] &= ~eh->mask[i];
}
if ((cm = mps_alloc_command(sc)) == NULL)
return (EBUSY);
evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
evtreq->MsgFlags = 0;
evtreq->SASBroadcastPrimitiveMasks = 0;
#ifdef MPS_DEBUG_ALL_EVENTS
{
u_char fullmask[16];
memset(fullmask, 0x00, 16);
bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
}
#else
bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
#endif
cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
cm->cm_data = NULL;
error = mps_request_polled(sc, cm);
reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
if ((reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
error = ENXIO;
mps_print_event(sc, reply);
mps_free_command(sc, cm);
return (error);
}
int
mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
{
TAILQ_REMOVE(&sc->event_list, handle, eh_list);
free(handle, M_MPT2);
return (mps_update_events(sc, NULL, NULL));
}
/*
* Add a chain element as the next SGE for the specified command.
* Reset cm_sge and cm_sgesize to indicate all the available space.
*/
static int
mps_add_chain(struct mps_command *cm)
{
MPI2_SGE_CHAIN32 *sgc;
struct mps_chain *chain;
int space;
if (cm->cm_sglsize < MPS_SGC_SIZE)
panic("MPS: Need SGE Error Code\n");
chain = mps_alloc_chain(cm->cm_sc);
if (chain == NULL)
return (ENOBUFS);
space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
/*
* Note: a double-linked list is used to make it easier to
* walk for debugging.
*/
TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
sgc->Length = space;
sgc->NextChainOffset = 0;
sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
sgc->Address = chain->chain_busaddr;
cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
cm->cm_sglsize = space;
return (0);
}
/*
* Add one scatter-gather element (chain, simple, transaction context)
* to the scatter-gather list for a command. Maintain cm_sglsize and
* cm_sge as the remaining size and pointer to the next SGE to fill
* in, respectively.
*/
int
mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
{
MPI2_SGE_TRANSACTION_UNION *tc = sgep;
MPI2_SGE_SIMPLE64 *sge = sgep;
int error, type;
type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
#ifdef INVARIANTS
switch (type) {
case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
if (len != tc->DetailsLength + 4)
panic("TC %p length %u or %zu?", tc,
tc->DetailsLength + 4, len);
}
break;
case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
/* Driver only uses 32-bit chain elements */
if (len != MPS_SGC_SIZE)
panic("CHAIN %p length %u or %zu?", sgep,
MPS_SGC_SIZE, len);
break;
case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
/* Driver only uses 64-bit SGE simple elements */
sge = sgep;
if (len != MPS_SGE64_SIZE)
panic("SGE simple %p length %u or %zu?", sge,
MPS_SGE64_SIZE, len);
if (((sge->FlagsLength >> MPI2_SGE_FLAGS_SHIFT) &
MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
panic("SGE simple %p flags %02x not marked 64-bit?",
sge, sge->FlagsLength >> MPI2_SGE_FLAGS_SHIFT);
break;
default:
panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
}
#endif
/*
* case 1: 1 more segment, enough room for it
* case 2: 2 more segments, enough room for both
* case 3: >=2 more segments, only enough room for 1 and a chain
* case 4: >=1 more segment, enough room for only a chain
* case 5: >=1 more segment, no room for anything (error)
*/
/*
* There should be room for at least a chain element, or this
* code is buggy. Case (5).
*/
if (cm->cm_sglsize < MPS_SGC_SIZE)
panic("MPS: Need SGE Error Code\n");
if (segsleft >= 2 &&
cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
/*
* There are 2 or more segments left to add, and only
* enough room for 1 and a chain. Case (3).
*
* Mark as last element in this chain if necessary.
*/
if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
sge->FlagsLength |=
(MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
}
/*
* Add the item then a chain. Do the chain now,
* rather than on the next iteration, to simplify
* understanding the code.
*/
cm->cm_sglsize -= len;
bcopy(sgep, cm->cm_sge, len);
cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
return (mps_add_chain(cm));
}
if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
/*
* 1 or more segment, enough room for only a chain.
* Hope the previous element wasn't a Simple entry
* that needed to be marked with
* MPI2_SGE_FLAGS_LAST_ELEMENT. Case (4).
*/
if ((error = mps_add_chain(cm)) != 0)
return (error);
}
#ifdef INVARIANTS
/* Case 1: 1 more segment, enough room for it. */
if (segsleft == 1 && cm->cm_sglsize < len)
panic("1 seg left and no room? %u versus %zu",
cm->cm_sglsize, len);
/* Case 2: 2 more segments, enough room for both */
if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
panic("2 segs left and no room? %u versus %zu",
cm->cm_sglsize, len);
#endif
if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
/*
* Last element of the last segment of the entire
* buffer.
*/
sge->FlagsLength |= ((MPI2_SGE_FLAGS_LAST_ELEMENT |
MPI2_SGE_FLAGS_END_OF_BUFFER |
MPI2_SGE_FLAGS_END_OF_LIST) << MPI2_SGE_FLAGS_SHIFT);
}
cm->cm_sglsize -= len;
bcopy(sgep, cm->cm_sge, len);
cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
return (0);
}
/*
* Add one dma segment to the scatter-gather list for a command.
*/
int
mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
int segsleft)
{
MPI2_SGE_SIMPLE64 sge;
/*
* This driver always uses 64-bit address elements for
* simplicity.
*/
flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT | MPI2_SGE_FLAGS_ADDRESS_SIZE;
sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
mps_from_u64(pa, &sge.Address);
return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
}
static void
mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
{
struct mps_softc *sc;
struct mps_command *cm;
u_int i, dir, sflags;
cm = (struct mps_command *)arg;
sc = cm->cm_sc;
/*
Add Serial Management Protocol (SMP) passthrough support to CAM. This includes support in the kernel, camcontrol(8), libcam and the mps(4) driver for SMP passthrough. The CAM SCSI probe code has been modified to fetch Inquiry VPD page 0x00 to determine supported pages, and will now fetch page 0x83 in addition to page 0x80 if supported. Add two new CAM CCBs, XPT_SMP_IO, and XPT_GDEV_ADVINFO. The SMP CCB is intended for SMP requests and responses. The ADVINFO is currently used to fetch cached VPD page 0x83 data from the transport layer, but is intended to be extensible to fetch other types of device-specific data. SMP-only devices are not currently represented in the CAM topology, and so the current semantics are that the SIM will route SMP CCBs to either the addressed device, if it contains an SMP target, or its parent, if it contains an SMP target. (This is noted in cam_ccb.h, since it will change later once we have the ability to have SMP-only devices in CAM's topology.) smp_all.c, smp_all.h: New helper routines for SMP. This includes SMP request building routines, response parsing routines, error decoding routines, and structure definitions for a number of SMP commands. libcam/Makefile: Add smp_all.c to libcam, so that SMP functionality is available to userland applications. camcontrol.8, camcontrol.c: Add smp passthrough support to camcontrol. Several new subcommands are now available: 'smpcmd' functions much like 'cmd', except that it allows the user to send generic SMP commands. 'smprg' sends the SMP report general command, and displays the decoded output. It will automatically fetch extended output if it is available. 'smppc' sends the SMP phy control command, with any number of potential options. Among other things, this allows the user to reset a phy on a SAS expander, or disable a phy on an expander. 'smpmaninfo' sends the SMP report manufacturer information and displays the decoded output. 'smpphylist' displays a list of phys on an expander, and the CAM devices attached to those phys, if any. cam.h, cam.c: Add a status value for SMP errors (CAM_SMP_STATUS_ERROR). Add a missing description for CAM_SCSI_IT_NEXUS_LOST. Add support for SMP commands to cam_error_string(). cam_ccb.h: Rename the CAM_DIR_RESV flag to CAM_DIR_BOTH. SMP commands are by nature bi-directional, and we may need to support bi-directional SCSI commands later. Add the XPT_SMP_IO CCB. Since SMP commands are bi-directional, there are pointers for both the request and response. Add a fill routine for SMP CCBs. Add the XPT_GDEV_ADVINFO CCB. This is currently used to fetch cached page 0x83 data from the transport later, but is extensible to fetch many other types of data. cam_periph.c: Add support in cam_periph_mapmem() for XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. cam_xpt.c: Add support for executing XPT_SMP_IO CCBs. cam_xpt_internal.h: Add fields for VPD pages 0x00 and 0x83 in struct cam_ed. scsi_all.c: Add scsi_get_sas_addr(), a function that parses VPD page 0x83 data and pulls out a SAS address. scsi_all.h: Add VPD page 0x00 and 0x83 structures, and a prototype for scsi_get_sas_addr(). scsi_pass.c: Add support for mapping buffers in XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. scsi_xpt.c: In the SCSI probe code, first ask the device for VPD page 0x00. If any VPD pages are supported, that page is required to be implemented. Based on the response, we may probe for the serial number (page 0x80) or device id (page 0x83). Add support for the XPT_GDEV_ADVINFO CCB. sys/conf/files: Add smp_all.c. mps.c: Add support for passing in a uio in mps_map_command(), so we can map a S/G list at once. Add support for SMP passthrough commands in mps_data_cb(). SMP is a special case, because the first buffer in the S/G list is outbound and the second buffer is inbound. Add support for warning the user if the busdma code comes back with more buffers than will work for the command. This will, for example, help the user determine why an SMP command failed if busdma comes back with three buffers. mps_pci.c: Add sys/uio.h. mps_sas.c: Add the SAS address and the parent handle to the list of fields we pull from device page 0 and cache in struct mpssas_target. These are needed for SMP passthrough. Add support for the XPT_SMP_IO CCB. For now, this CCB is routed to the addressed device if it supports SMP, or to its parent if it does not and the parent does. This is necessary because CAM does not currently support SMP-only nodes in the topology. Make SMP passthrough support conditional on __FreeBSD_version >= 900026. This will make it easier to MFC this change to the driver without MFCing the CAM changes as well. mps_user.c: Un-staticize mpi_init_sge() so we can use it for the SMP passthrough code. mpsvar.h: Add a uio and iovecs into struct mps_command for SMP passthrough commands. Add a cm_max_segs field to struct mps_command so that we can warn the user if busdma comes back with too many segments. Clear the cm_reply when a command gets freed. If it is not cleared, reply frames will eventually get freed into the pool multiple times and corrupt the pool. (This fix is from scottl.) Add a prototype for mpi_init_sge(). sys/param.h: Bump __FreeBSD_version to 900026 for the for the inclusion of the XPT_GDEV_ADVINFO and XPT_SMP_IO CAM CCBs.
2010-11-30 22:39:46 +00:00
* In this case, just print out a warning and let the chip tell the
* user they did the wrong thing.
*/
if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
mps_printf(sc, "%s: warning: busdma returned %d segments, "
"more than the %d allowed\n", __func__, nsegs,
cm->cm_max_segs);
}
/*
* Set up DMA direction flags. Note that we don't support
* bi-directional transfers, with the exception of SMP passthrough.
*/
sflags = 0;
Add Serial Management Protocol (SMP) passthrough support to CAM. This includes support in the kernel, camcontrol(8), libcam and the mps(4) driver for SMP passthrough. The CAM SCSI probe code has been modified to fetch Inquiry VPD page 0x00 to determine supported pages, and will now fetch page 0x83 in addition to page 0x80 if supported. Add two new CAM CCBs, XPT_SMP_IO, and XPT_GDEV_ADVINFO. The SMP CCB is intended for SMP requests and responses. The ADVINFO is currently used to fetch cached VPD page 0x83 data from the transport layer, but is intended to be extensible to fetch other types of device-specific data. SMP-only devices are not currently represented in the CAM topology, and so the current semantics are that the SIM will route SMP CCBs to either the addressed device, if it contains an SMP target, or its parent, if it contains an SMP target. (This is noted in cam_ccb.h, since it will change later once we have the ability to have SMP-only devices in CAM's topology.) smp_all.c, smp_all.h: New helper routines for SMP. This includes SMP request building routines, response parsing routines, error decoding routines, and structure definitions for a number of SMP commands. libcam/Makefile: Add smp_all.c to libcam, so that SMP functionality is available to userland applications. camcontrol.8, camcontrol.c: Add smp passthrough support to camcontrol. Several new subcommands are now available: 'smpcmd' functions much like 'cmd', except that it allows the user to send generic SMP commands. 'smprg' sends the SMP report general command, and displays the decoded output. It will automatically fetch extended output if it is available. 'smppc' sends the SMP phy control command, with any number of potential options. Among other things, this allows the user to reset a phy on a SAS expander, or disable a phy on an expander. 'smpmaninfo' sends the SMP report manufacturer information and displays the decoded output. 'smpphylist' displays a list of phys on an expander, and the CAM devices attached to those phys, if any. cam.h, cam.c: Add a status value for SMP errors (CAM_SMP_STATUS_ERROR). Add a missing description for CAM_SCSI_IT_NEXUS_LOST. Add support for SMP commands to cam_error_string(). cam_ccb.h: Rename the CAM_DIR_RESV flag to CAM_DIR_BOTH. SMP commands are by nature bi-directional, and we may need to support bi-directional SCSI commands later. Add the XPT_SMP_IO CCB. Since SMP commands are bi-directional, there are pointers for both the request and response. Add a fill routine for SMP CCBs. Add the XPT_GDEV_ADVINFO CCB. This is currently used to fetch cached page 0x83 data from the transport later, but is extensible to fetch many other types of data. cam_periph.c: Add support in cam_periph_mapmem() for XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. cam_xpt.c: Add support for executing XPT_SMP_IO CCBs. cam_xpt_internal.h: Add fields for VPD pages 0x00 and 0x83 in struct cam_ed. scsi_all.c: Add scsi_get_sas_addr(), a function that parses VPD page 0x83 data and pulls out a SAS address. scsi_all.h: Add VPD page 0x00 and 0x83 structures, and a prototype for scsi_get_sas_addr(). scsi_pass.c: Add support for mapping buffers in XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. scsi_xpt.c: In the SCSI probe code, first ask the device for VPD page 0x00. If any VPD pages are supported, that page is required to be implemented. Based on the response, we may probe for the serial number (page 0x80) or device id (page 0x83). Add support for the XPT_GDEV_ADVINFO CCB. sys/conf/files: Add smp_all.c. mps.c: Add support for passing in a uio in mps_map_command(), so we can map a S/G list at once. Add support for SMP passthrough commands in mps_data_cb(). SMP is a special case, because the first buffer in the S/G list is outbound and the second buffer is inbound. Add support for warning the user if the busdma code comes back with more buffers than will work for the command. This will, for example, help the user determine why an SMP command failed if busdma comes back with three buffers. mps_pci.c: Add sys/uio.h. mps_sas.c: Add the SAS address and the parent handle to the list of fields we pull from device page 0 and cache in struct mpssas_target. These are needed for SMP passthrough. Add support for the XPT_SMP_IO CCB. For now, this CCB is routed to the addressed device if it supports SMP, or to its parent if it does not and the parent does. This is necessary because CAM does not currently support SMP-only nodes in the topology. Make SMP passthrough support conditional on __FreeBSD_version >= 900026. This will make it easier to MFC this change to the driver without MFCing the CAM changes as well. mps_user.c: Un-staticize mpi_init_sge() so we can use it for the SMP passthrough code. mpsvar.h: Add a uio and iovecs into struct mps_command for SMP passthrough commands. Add a cm_max_segs field to struct mps_command so that we can warn the user if busdma comes back with too many segments. Clear the cm_reply when a command gets freed. If it is not cleared, reply frames will eventually get freed into the pool multiple times and corrupt the pool. (This fix is from scottl.) Add a prototype for mpi_init_sge(). sys/param.h: Bump __FreeBSD_version to 900026 for the for the inclusion of the XPT_GDEV_ADVINFO and XPT_SMP_IO CAM CCBs.
2010-11-30 22:39:46 +00:00
if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
/*
* We have to add a special case for SMP passthrough, there
* is no easy way to generically handle it. The first
* S/G element is used for the command (therefore the
* direction bit needs to be set). The second one is used
* for the reply. We'll leave it to the caller to make
* sure we only have two buffers.
*/
/*
* Even though the busdma man page says it doesn't make
* sense to have both direction flags, it does in this case.
* We have one s/g element being accessed in each direction.
*/
dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
/*
* Set the direction flag on the first buffer in the SMP
* passthrough request. We'll clear it for the second one.
*/
sflags |= MPI2_SGE_FLAGS_DIRECTION |
MPI2_SGE_FLAGS_END_OF_BUFFER;
} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
sflags |= MPI2_SGE_FLAGS_DIRECTION;
dir = BUS_DMASYNC_PREWRITE;
} else
dir = BUS_DMASYNC_PREREAD;
for (i = 0; i < nsegs; i++) {
Add Serial Management Protocol (SMP) passthrough support to CAM. This includes support in the kernel, camcontrol(8), libcam and the mps(4) driver for SMP passthrough. The CAM SCSI probe code has been modified to fetch Inquiry VPD page 0x00 to determine supported pages, and will now fetch page 0x83 in addition to page 0x80 if supported. Add two new CAM CCBs, XPT_SMP_IO, and XPT_GDEV_ADVINFO. The SMP CCB is intended for SMP requests and responses. The ADVINFO is currently used to fetch cached VPD page 0x83 data from the transport layer, but is intended to be extensible to fetch other types of device-specific data. SMP-only devices are not currently represented in the CAM topology, and so the current semantics are that the SIM will route SMP CCBs to either the addressed device, if it contains an SMP target, or its parent, if it contains an SMP target. (This is noted in cam_ccb.h, since it will change later once we have the ability to have SMP-only devices in CAM's topology.) smp_all.c, smp_all.h: New helper routines for SMP. This includes SMP request building routines, response parsing routines, error decoding routines, and structure definitions for a number of SMP commands. libcam/Makefile: Add smp_all.c to libcam, so that SMP functionality is available to userland applications. camcontrol.8, camcontrol.c: Add smp passthrough support to camcontrol. Several new subcommands are now available: 'smpcmd' functions much like 'cmd', except that it allows the user to send generic SMP commands. 'smprg' sends the SMP report general command, and displays the decoded output. It will automatically fetch extended output if it is available. 'smppc' sends the SMP phy control command, with any number of potential options. Among other things, this allows the user to reset a phy on a SAS expander, or disable a phy on an expander. 'smpmaninfo' sends the SMP report manufacturer information and displays the decoded output. 'smpphylist' displays a list of phys on an expander, and the CAM devices attached to those phys, if any. cam.h, cam.c: Add a status value for SMP errors (CAM_SMP_STATUS_ERROR). Add a missing description for CAM_SCSI_IT_NEXUS_LOST. Add support for SMP commands to cam_error_string(). cam_ccb.h: Rename the CAM_DIR_RESV flag to CAM_DIR_BOTH. SMP commands are by nature bi-directional, and we may need to support bi-directional SCSI commands later. Add the XPT_SMP_IO CCB. Since SMP commands are bi-directional, there are pointers for both the request and response. Add a fill routine for SMP CCBs. Add the XPT_GDEV_ADVINFO CCB. This is currently used to fetch cached page 0x83 data from the transport later, but is extensible to fetch many other types of data. cam_periph.c: Add support in cam_periph_mapmem() for XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. cam_xpt.c: Add support for executing XPT_SMP_IO CCBs. cam_xpt_internal.h: Add fields for VPD pages 0x00 and 0x83 in struct cam_ed. scsi_all.c: Add scsi_get_sas_addr(), a function that parses VPD page 0x83 data and pulls out a SAS address. scsi_all.h: Add VPD page 0x00 and 0x83 structures, and a prototype for scsi_get_sas_addr(). scsi_pass.c: Add support for mapping buffers in XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. scsi_xpt.c: In the SCSI probe code, first ask the device for VPD page 0x00. If any VPD pages are supported, that page is required to be implemented. Based on the response, we may probe for the serial number (page 0x80) or device id (page 0x83). Add support for the XPT_GDEV_ADVINFO CCB. sys/conf/files: Add smp_all.c. mps.c: Add support for passing in a uio in mps_map_command(), so we can map a S/G list at once. Add support for SMP passthrough commands in mps_data_cb(). SMP is a special case, because the first buffer in the S/G list is outbound and the second buffer is inbound. Add support for warning the user if the busdma code comes back with more buffers than will work for the command. This will, for example, help the user determine why an SMP command failed if busdma comes back with three buffers. mps_pci.c: Add sys/uio.h. mps_sas.c: Add the SAS address and the parent handle to the list of fields we pull from device page 0 and cache in struct mpssas_target. These are needed for SMP passthrough. Add support for the XPT_SMP_IO CCB. For now, this CCB is routed to the addressed device if it supports SMP, or to its parent if it does not and the parent does. This is necessary because CAM does not currently support SMP-only nodes in the topology. Make SMP passthrough support conditional on __FreeBSD_version >= 900026. This will make it easier to MFC this change to the driver without MFCing the CAM changes as well. mps_user.c: Un-staticize mpi_init_sge() so we can use it for the SMP passthrough code. mpsvar.h: Add a uio and iovecs into struct mps_command for SMP passthrough commands. Add a cm_max_segs field to struct mps_command so that we can warn the user if busdma comes back with too many segments. Clear the cm_reply when a command gets freed. If it is not cleared, reply frames will eventually get freed into the pool multiple times and corrupt the pool. (This fix is from scottl.) Add a prototype for mpi_init_sge(). sys/param.h: Bump __FreeBSD_version to 900026 for the for the inclusion of the XPT_GDEV_ADVINFO and XPT_SMP_IO CAM CCBs.
2010-11-30 22:39:46 +00:00
if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS)
&& (i != 0)) {
sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
}
error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
sflags, nsegs - i);
if (error != 0) {
/* Resource shortage, roll back! */
mps_dprint(sc, MPS_INFO, "out of chain frames\n");
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
mps_complete_command(cm);
return;
}
}
bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
mps_enqueue_request(sc, cm);
return;
}
Add Serial Management Protocol (SMP) passthrough support to CAM. This includes support in the kernel, camcontrol(8), libcam and the mps(4) driver for SMP passthrough. The CAM SCSI probe code has been modified to fetch Inquiry VPD page 0x00 to determine supported pages, and will now fetch page 0x83 in addition to page 0x80 if supported. Add two new CAM CCBs, XPT_SMP_IO, and XPT_GDEV_ADVINFO. The SMP CCB is intended for SMP requests and responses. The ADVINFO is currently used to fetch cached VPD page 0x83 data from the transport layer, but is intended to be extensible to fetch other types of device-specific data. SMP-only devices are not currently represented in the CAM topology, and so the current semantics are that the SIM will route SMP CCBs to either the addressed device, if it contains an SMP target, or its parent, if it contains an SMP target. (This is noted in cam_ccb.h, since it will change later once we have the ability to have SMP-only devices in CAM's topology.) smp_all.c, smp_all.h: New helper routines for SMP. This includes SMP request building routines, response parsing routines, error decoding routines, and structure definitions for a number of SMP commands. libcam/Makefile: Add smp_all.c to libcam, so that SMP functionality is available to userland applications. camcontrol.8, camcontrol.c: Add smp passthrough support to camcontrol. Several new subcommands are now available: 'smpcmd' functions much like 'cmd', except that it allows the user to send generic SMP commands. 'smprg' sends the SMP report general command, and displays the decoded output. It will automatically fetch extended output if it is available. 'smppc' sends the SMP phy control command, with any number of potential options. Among other things, this allows the user to reset a phy on a SAS expander, or disable a phy on an expander. 'smpmaninfo' sends the SMP report manufacturer information and displays the decoded output. 'smpphylist' displays a list of phys on an expander, and the CAM devices attached to those phys, if any. cam.h, cam.c: Add a status value for SMP errors (CAM_SMP_STATUS_ERROR). Add a missing description for CAM_SCSI_IT_NEXUS_LOST. Add support for SMP commands to cam_error_string(). cam_ccb.h: Rename the CAM_DIR_RESV flag to CAM_DIR_BOTH. SMP commands are by nature bi-directional, and we may need to support bi-directional SCSI commands later. Add the XPT_SMP_IO CCB. Since SMP commands are bi-directional, there are pointers for both the request and response. Add a fill routine for SMP CCBs. Add the XPT_GDEV_ADVINFO CCB. This is currently used to fetch cached page 0x83 data from the transport later, but is extensible to fetch many other types of data. cam_periph.c: Add support in cam_periph_mapmem() for XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. cam_xpt.c: Add support for executing XPT_SMP_IO CCBs. cam_xpt_internal.h: Add fields for VPD pages 0x00 and 0x83 in struct cam_ed. scsi_all.c: Add scsi_get_sas_addr(), a function that parses VPD page 0x83 data and pulls out a SAS address. scsi_all.h: Add VPD page 0x00 and 0x83 structures, and a prototype for scsi_get_sas_addr(). scsi_pass.c: Add support for mapping buffers in XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. scsi_xpt.c: In the SCSI probe code, first ask the device for VPD page 0x00. If any VPD pages are supported, that page is required to be implemented. Based on the response, we may probe for the serial number (page 0x80) or device id (page 0x83). Add support for the XPT_GDEV_ADVINFO CCB. sys/conf/files: Add smp_all.c. mps.c: Add support for passing in a uio in mps_map_command(), so we can map a S/G list at once. Add support for SMP passthrough commands in mps_data_cb(). SMP is a special case, because the first buffer in the S/G list is outbound and the second buffer is inbound. Add support for warning the user if the busdma code comes back with more buffers than will work for the command. This will, for example, help the user determine why an SMP command failed if busdma comes back with three buffers. mps_pci.c: Add sys/uio.h. mps_sas.c: Add the SAS address and the parent handle to the list of fields we pull from device page 0 and cache in struct mpssas_target. These are needed for SMP passthrough. Add support for the XPT_SMP_IO CCB. For now, this CCB is routed to the addressed device if it supports SMP, or to its parent if it does not and the parent does. This is necessary because CAM does not currently support SMP-only nodes in the topology. Make SMP passthrough support conditional on __FreeBSD_version >= 900026. This will make it easier to MFC this change to the driver without MFCing the CAM changes as well. mps_user.c: Un-staticize mpi_init_sge() so we can use it for the SMP passthrough code. mpsvar.h: Add a uio and iovecs into struct mps_command for SMP passthrough commands. Add a cm_max_segs field to struct mps_command so that we can warn the user if busdma comes back with too many segments. Clear the cm_reply when a command gets freed. If it is not cleared, reply frames will eventually get freed into the pool multiple times and corrupt the pool. (This fix is from scottl.) Add a prototype for mpi_init_sge(). sys/param.h: Bump __FreeBSD_version to 900026 for the for the inclusion of the XPT_GDEV_ADVINFO and XPT_SMP_IO CAM CCBs.
2010-11-30 22:39:46 +00:00
static void
mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
int error)
{
mps_data_cb(arg, segs, nsegs, error);
}
/*
* Note that the only error path here is from bus_dmamap_load(), which can
* return EINPROGRESS if it is waiting for resources.
*/
int
mps_map_command(struct mps_softc *sc, struct mps_command *cm)
{
MPI2_SGE_SIMPLE32 *sge;
int error = 0;
Add Serial Management Protocol (SMP) passthrough support to CAM. This includes support in the kernel, camcontrol(8), libcam and the mps(4) driver for SMP passthrough. The CAM SCSI probe code has been modified to fetch Inquiry VPD page 0x00 to determine supported pages, and will now fetch page 0x83 in addition to page 0x80 if supported. Add two new CAM CCBs, XPT_SMP_IO, and XPT_GDEV_ADVINFO. The SMP CCB is intended for SMP requests and responses. The ADVINFO is currently used to fetch cached VPD page 0x83 data from the transport layer, but is intended to be extensible to fetch other types of device-specific data. SMP-only devices are not currently represented in the CAM topology, and so the current semantics are that the SIM will route SMP CCBs to either the addressed device, if it contains an SMP target, or its parent, if it contains an SMP target. (This is noted in cam_ccb.h, since it will change later once we have the ability to have SMP-only devices in CAM's topology.) smp_all.c, smp_all.h: New helper routines for SMP. This includes SMP request building routines, response parsing routines, error decoding routines, and structure definitions for a number of SMP commands. libcam/Makefile: Add smp_all.c to libcam, so that SMP functionality is available to userland applications. camcontrol.8, camcontrol.c: Add smp passthrough support to camcontrol. Several new subcommands are now available: 'smpcmd' functions much like 'cmd', except that it allows the user to send generic SMP commands. 'smprg' sends the SMP report general command, and displays the decoded output. It will automatically fetch extended output if it is available. 'smppc' sends the SMP phy control command, with any number of potential options. Among other things, this allows the user to reset a phy on a SAS expander, or disable a phy on an expander. 'smpmaninfo' sends the SMP report manufacturer information and displays the decoded output. 'smpphylist' displays a list of phys on an expander, and the CAM devices attached to those phys, if any. cam.h, cam.c: Add a status value for SMP errors (CAM_SMP_STATUS_ERROR). Add a missing description for CAM_SCSI_IT_NEXUS_LOST. Add support for SMP commands to cam_error_string(). cam_ccb.h: Rename the CAM_DIR_RESV flag to CAM_DIR_BOTH. SMP commands are by nature bi-directional, and we may need to support bi-directional SCSI commands later. Add the XPT_SMP_IO CCB. Since SMP commands are bi-directional, there are pointers for both the request and response. Add a fill routine for SMP CCBs. Add the XPT_GDEV_ADVINFO CCB. This is currently used to fetch cached page 0x83 data from the transport later, but is extensible to fetch many other types of data. cam_periph.c: Add support in cam_periph_mapmem() for XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. cam_xpt.c: Add support for executing XPT_SMP_IO CCBs. cam_xpt_internal.h: Add fields for VPD pages 0x00 and 0x83 in struct cam_ed. scsi_all.c: Add scsi_get_sas_addr(), a function that parses VPD page 0x83 data and pulls out a SAS address. scsi_all.h: Add VPD page 0x00 and 0x83 structures, and a prototype for scsi_get_sas_addr(). scsi_pass.c: Add support for mapping buffers in XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. scsi_xpt.c: In the SCSI probe code, first ask the device for VPD page 0x00. If any VPD pages are supported, that page is required to be implemented. Based on the response, we may probe for the serial number (page 0x80) or device id (page 0x83). Add support for the XPT_GDEV_ADVINFO CCB. sys/conf/files: Add smp_all.c. mps.c: Add support for passing in a uio in mps_map_command(), so we can map a S/G list at once. Add support for SMP passthrough commands in mps_data_cb(). SMP is a special case, because the first buffer in the S/G list is outbound and the second buffer is inbound. Add support for warning the user if the busdma code comes back with more buffers than will work for the command. This will, for example, help the user determine why an SMP command failed if busdma comes back with three buffers. mps_pci.c: Add sys/uio.h. mps_sas.c: Add the SAS address and the parent handle to the list of fields we pull from device page 0 and cache in struct mpssas_target. These are needed for SMP passthrough. Add support for the XPT_SMP_IO CCB. For now, this CCB is routed to the addressed device if it supports SMP, or to its parent if it does not and the parent does. This is necessary because CAM does not currently support SMP-only nodes in the topology. Make SMP passthrough support conditional on __FreeBSD_version >= 900026. This will make it easier to MFC this change to the driver without MFCing the CAM changes as well. mps_user.c: Un-staticize mpi_init_sge() so we can use it for the SMP passthrough code. mpsvar.h: Add a uio and iovecs into struct mps_command for SMP passthrough commands. Add a cm_max_segs field to struct mps_command so that we can warn the user if busdma comes back with too many segments. Clear the cm_reply when a command gets freed. If it is not cleared, reply frames will eventually get freed into the pool multiple times and corrupt the pool. (This fix is from scottl.) Add a prototype for mpi_init_sge(). sys/param.h: Bump __FreeBSD_version to 900026 for the for the inclusion of the XPT_GDEV_ADVINFO and XPT_SMP_IO CAM CCBs.
2010-11-30 22:39:46 +00:00
if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
&cm->cm_uio, mps_data_cb2, cm, 0);
} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
} else {
/* Add a zero-length element as needed */
if (cm->cm_sge != NULL) {
sge = (MPI2_SGE_SIMPLE32 *)cm->cm_sge;
sge->FlagsLength = (MPI2_SGE_FLAGS_LAST_ELEMENT |
MPI2_SGE_FLAGS_END_OF_BUFFER |
MPI2_SGE_FLAGS_END_OF_LIST |
MPI2_SGE_FLAGS_SIMPLE_ELEMENT) <<
MPI2_SGE_FLAGS_SHIFT;
sge->Address = 0;
}
Add Serial Management Protocol (SMP) passthrough support to CAM. This includes support in the kernel, camcontrol(8), libcam and the mps(4) driver for SMP passthrough. The CAM SCSI probe code has been modified to fetch Inquiry VPD page 0x00 to determine supported pages, and will now fetch page 0x83 in addition to page 0x80 if supported. Add two new CAM CCBs, XPT_SMP_IO, and XPT_GDEV_ADVINFO. The SMP CCB is intended for SMP requests and responses. The ADVINFO is currently used to fetch cached VPD page 0x83 data from the transport layer, but is intended to be extensible to fetch other types of device-specific data. SMP-only devices are not currently represented in the CAM topology, and so the current semantics are that the SIM will route SMP CCBs to either the addressed device, if it contains an SMP target, or its parent, if it contains an SMP target. (This is noted in cam_ccb.h, since it will change later once we have the ability to have SMP-only devices in CAM's topology.) smp_all.c, smp_all.h: New helper routines for SMP. This includes SMP request building routines, response parsing routines, error decoding routines, and structure definitions for a number of SMP commands. libcam/Makefile: Add smp_all.c to libcam, so that SMP functionality is available to userland applications. camcontrol.8, camcontrol.c: Add smp passthrough support to camcontrol. Several new subcommands are now available: 'smpcmd' functions much like 'cmd', except that it allows the user to send generic SMP commands. 'smprg' sends the SMP report general command, and displays the decoded output. It will automatically fetch extended output if it is available. 'smppc' sends the SMP phy control command, with any number of potential options. Among other things, this allows the user to reset a phy on a SAS expander, or disable a phy on an expander. 'smpmaninfo' sends the SMP report manufacturer information and displays the decoded output. 'smpphylist' displays a list of phys on an expander, and the CAM devices attached to those phys, if any. cam.h, cam.c: Add a status value for SMP errors (CAM_SMP_STATUS_ERROR). Add a missing description for CAM_SCSI_IT_NEXUS_LOST. Add support for SMP commands to cam_error_string(). cam_ccb.h: Rename the CAM_DIR_RESV flag to CAM_DIR_BOTH. SMP commands are by nature bi-directional, and we may need to support bi-directional SCSI commands later. Add the XPT_SMP_IO CCB. Since SMP commands are bi-directional, there are pointers for both the request and response. Add a fill routine for SMP CCBs. Add the XPT_GDEV_ADVINFO CCB. This is currently used to fetch cached page 0x83 data from the transport later, but is extensible to fetch many other types of data. cam_periph.c: Add support in cam_periph_mapmem() for XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. cam_xpt.c: Add support for executing XPT_SMP_IO CCBs. cam_xpt_internal.h: Add fields for VPD pages 0x00 and 0x83 in struct cam_ed. scsi_all.c: Add scsi_get_sas_addr(), a function that parses VPD page 0x83 data and pulls out a SAS address. scsi_all.h: Add VPD page 0x00 and 0x83 structures, and a prototype for scsi_get_sas_addr(). scsi_pass.c: Add support for mapping buffers in XPT_SMP_IO and XPT_GDEV_ADVINFO CCBs. scsi_xpt.c: In the SCSI probe code, first ask the device for VPD page 0x00. If any VPD pages are supported, that page is required to be implemented. Based on the response, we may probe for the serial number (page 0x80) or device id (page 0x83). Add support for the XPT_GDEV_ADVINFO CCB. sys/conf/files: Add smp_all.c. mps.c: Add support for passing in a uio in mps_map_command(), so we can map a S/G list at once. Add support for SMP passthrough commands in mps_data_cb(). SMP is a special case, because the first buffer in the S/G list is outbound and the second buffer is inbound. Add support for warning the user if the busdma code comes back with more buffers than will work for the command. This will, for example, help the user determine why an SMP command failed if busdma comes back with three buffers. mps_pci.c: Add sys/uio.h. mps_sas.c: Add the SAS address and the parent handle to the list of fields we pull from device page 0 and cache in struct mpssas_target. These are needed for SMP passthrough. Add support for the XPT_SMP_IO CCB. For now, this CCB is routed to the addressed device if it supports SMP, or to its parent if it does not and the parent does. This is necessary because CAM does not currently support SMP-only nodes in the topology. Make SMP passthrough support conditional on __FreeBSD_version >= 900026. This will make it easier to MFC this change to the driver without MFCing the CAM changes as well. mps_user.c: Un-staticize mpi_init_sge() so we can use it for the SMP passthrough code. mpsvar.h: Add a uio and iovecs into struct mps_command for SMP passthrough commands. Add a cm_max_segs field to struct mps_command so that we can warn the user if busdma comes back with too many segments. Clear the cm_reply when a command gets freed. If it is not cleared, reply frames will eventually get freed into the pool multiple times and corrupt the pool. (This fix is from scottl.) Add a prototype for mpi_init_sge(). sys/param.h: Bump __FreeBSD_version to 900026 for the for the inclusion of the XPT_GDEV_ADVINFO and XPT_SMP_IO CAM CCBs.
2010-11-30 22:39:46 +00:00
mps_enqueue_request(sc, cm);
}
return (error);
}
/*
* The MPT driver had a verbose interface for config pages. In this driver,
* reduce it to much simplier terms, similar to the Linux driver.
*/
int
mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
{
MPI2_CONFIG_REQUEST *req;
struct mps_command *cm;
int error;
if (sc->mps_flags & MPS_FLAGS_BUSY) {
return (EBUSY);
}
cm = mps_alloc_command(sc);
if (cm == NULL) {
return (EBUSY);
}
req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
req->Function = MPI2_FUNCTION_CONFIG;
req->Action = params->action;
req->SGLFlags = 0;
req->ChainOffset = 0;
req->PageAddress = params->page_address;
if (params->hdr.Ext.ExtPageType != 0) {
MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
hdr = &params->hdr.Ext;
req->ExtPageType = hdr->ExtPageType;
req->ExtPageLength = hdr->ExtPageLength;
req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
req->Header.PageLength = 0; /* Must be set to zero */
req->Header.PageNumber = hdr->PageNumber;
req->Header.PageVersion = hdr->PageVersion;
} else {
MPI2_CONFIG_PAGE_HEADER *hdr;
hdr = &params->hdr.Struct;
req->Header.PageType = hdr->PageType;
req->Header.PageNumber = hdr->PageNumber;
req->Header.PageLength = hdr->PageLength;
req->Header.PageVersion = hdr->PageVersion;
}
cm->cm_data = params->buffer;
cm->cm_length = params->length;
cm->cm_sge = &req->PageBufferSGE;
cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
cm->cm_complete_data = params;
if (params->callback != NULL) {
cm->cm_complete = mps_config_complete;
return (mps_map_command(sc, cm));
} else {
cm->cm_complete = NULL;
cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
if ((error = mps_map_command(sc, cm)) != 0)
return (error);
msleep(cm, &sc->mps_mtx, 0, "mpswait", 0);
mps_config_complete(sc, cm);
}
return (0);
}
int
mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
{
return (EINVAL);
}
static void
mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
{
MPI2_CONFIG_REPLY *reply;
struct mps_config_params *params;
params = cm->cm_complete_data;
if (cm->cm_data != NULL) {
bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
BUS_DMASYNC_POSTREAD);
bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
}
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
/*
* XXX KDM need to do more error recovery? This results in the
* device in question not getting probed.
*/
if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
params->status = MPI2_IOCSTATUS_BUSY;
goto bailout;
}
reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
params->status = reply->IOCStatus;
if (params->hdr.Ext.ExtPageType != 0) {
params->hdr.Ext.ExtPageType = reply->ExtPageType;
params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
} else {
params->hdr.Struct.PageType = reply->Header.PageType;
params->hdr.Struct.PageNumber = reply->Header.PageNumber;
params->hdr.Struct.PageLength = reply->Header.PageLength;
params->hdr.Struct.PageVersion = reply->Header.PageVersion;
}
Fix several issues with the mps(4) driver. When the driver ran out of DMA chaining buffers, it kept the timeout for the I/O, and I/O would stall. The driver was not freezing the device queue on errors. mps.c: Pull command completion logic into a separate function, and call the callback/wakeup for commands that are never sent due to lack of chain buffers. Add a number of extra diagnostic sysctl variables. Handle pre-hardware errors for configuration I/O. This doesn't panic the system, but it will fail the configuration I/O and there is no retry mechanism. So the device probe will not succeed. This should be a very uncommon situation, however. mps_sas.c: Freeze the SIM queue when we run out of chain buffers, and unfreeze it when more commands complete. Freeze the device queue when errors occur, so that CAM can insure proper command ordering. Report pre-hardware errors for task management commands. In general, that shouldn't be possible because task management commands don't have S/G lists, and that is currently the only error path before we get to the hardware. Handle pre-hardware errors (like out of chain elements) for SMP requests. That shouldn't happen either, since we should have enough space for two S/G elements in the standard request. For commands that end with MPI2_IOCSTATUS_SCSI_IOC_TERMINATED and MPI2_IOCSTATUS_SCSI_EXT_TERMINATED, return them with CAM_REQUEUE_REQ to retry them unconditionally. These seem to be related to back end, transport related problems that are hopefully transient. We don't want to go through the retry count for something that is not a permanent error. Keep track of the number of outstanding I/Os. mpsvar.h: Track the number of free chain elements. Add variables for the number of outstanding I/Os, and I/O high water mark. Add variables to track the number of free chain buffers and the chain low water mark, as well as the number of chain allocation failures. Add I/O state flags and an attach done flag. MFC after: 3 days
2011-02-18 17:06:06 +00:00
bailout:
mps_free_command(sc, cm);
if (params->callback != NULL)
params->callback(sc, params);
return;
}