Patch 02: Add a hotswap infrastructure to libata
As described in the archives, this patch adds a framework and API to
libata to allow for hotswapping.
This version comes with a debounce timer... exciting!
Luke Kosewski
01.08.05 Luke Kosewski <lkosewsk@xxxxxx>
* A patch to add a general-purpose hotswap framework to libata. From
the point of view of the driver writer, we only have one new API
function; 'ata_hotplug_op', which is called to schedule a plug/unplug
event. We also need to prepare our driver to tell this function, when
executed, whether we are doing a plug/unplug; for this we also have
to fill in the ap->ops->hotplug_action callback in our driver.
* A few misc changes are necessary to properly reset flags which are
set by devices when they are present to make sure that new devices
(for instance, disks with different max UDMA transfer rates, not using
LBA48, etc.) being swapped in do not assume values for the old
devices.
* Gratuitous comments here that can probably be removed, so that anyone
looking at this can understand this and poke holes in my logic.
* This is a resend of the original patch which in particular:
- uses a 'debounce handler', done with a timer, to take care of cases
where we get spurious interrupts, usually caused by the metal pins
on a connector wobbling irregularly as they're pushed in.
Signed-off-by: Luke Kosewski <lkosewsk@xxxxxx>
--- linux-2.6.13-rc3/drivers/scsi/libata-core.c.old 2005-08-01 00:21:55.000000000 -0700
+++ linux-2.6.13-rc3/drivers/scsi/libata-core.c 2005-08-01 02:13:36.000000000 -0700
@@ -1134,6 +1135,11 @@ static void ata_dev_identify(struct ata_
return;
}
+ /* Necessary if we had an LBA48 drive in, we pulled it out, and put in
+ * a non-LBA48 drive to replace it.
+ */
+ dev->flags &= ~ATA_DFLAG_LBA48;
+
if (ap->flags & (ATA_FLAG_SRST | ATA_FLAG_SATA_RESET))
using_edd = 0;
else
@@ -3635,6 +3641,93 @@ idle_irq:
return 0; /* irq not handled */
}
+void ata_check_kill_qc(struct ata_port *ap)
+{
+ struct ata_queued_cmd *qc = ata_qc_from_tag(ap, ap->active_tag);
+
+ if (unlikely(qc)) {
+ /* This is SO bad. But we can't just run
+ * ata_qc_complete without doing this, because
+ * ata_scsi_qc_complete wants to talk to the device,
+ * and we can't let it do that since it doesn't exist
+ * anymore.
+ */
+ ata_scsi_prepare_qc_abort(qc);
+ ata_qc_complete(qc, ATA_ERR);
+ }
+}
+
+static inline void ata_hotplug_unplug_func(struct ata_port *ap)
+{
+ DPRINTK("Got an unplug request on port %d\n", ap->id);
+
+ ata_port_disable(ap); //Gone gone gone!
+ ata_scsi_handle_unplug(ap);
+}
+
+static inline void ata_hotplug_plug_func(struct ata_port *ap)
+{
+ DPRINTK("Got a plug request on port %d\n", ap->id);
+
+ /* Pure evil. Suppose that you have an 'unplug' waiting on your
+ * queue, and this function executes while it's there (because
+ * you unplugged/plugged in a disk on an SMP system VERY FAST).
+ * REALLY bad news, because when you unplugged your disk, you
+ * might have had a pending qc which will now sit there and time
+ * out like the mofo it is. Check to see if we have one sitting
+ * around and KILL IT if this is so.
+ */
+ ata_check_kill_qc(ap);
+ ap->flags |= ATA_FLAG_SATA_RESET; // Necessary on some Seagate drives.
+ ap->udma_mask = ap->orig_udma_mask;
+
+ if (ata_bus_probe(ap))
+ ata_scsi_handle_unplug(ap); //might be necessary on SMP
+ else
+ ata_scsi_handle_plug(ap);
+}
+
+static void ata_hotplug_timer_func(unsigned long _data)
+{
+ struct ata_port *ap = (struct ata_port *)_data;
+ int action;
+
+ //A driver that wants to play with hotplugging HAS to define this func.
+ BUG_ON(!ap->ops->hotplug_action);
+
+ spin_lock(&ap->host_set->lock);
+ /* I would have thought that we can just set a 'plug' or 'unplug'
+ * function up as the timer caller. However, some info from the Sil
+ * guys indicates that we should do another check to see whether we want
+ * to plug or unplug. Hence, we use this function callback instead.
+ * This has the advantage of only having one timer per ap; we can't use
+ * only one timer to make callbacks for both plug/unplug because on an
+ * SMP non-x86 (?) system, we might be overriding the memory location
+ * storing a pointer to the callback at the same time as a CPU grabs
+ * that timer and tries to run it == possible kernel panic.
+ * That COULD have been fixed by running del_timer_sync rather than
+ * del_timer in the ata_hotplug_(un)plug functions, except we can't run
+ * del_timer_sync from an interrupt context, so we have this
+ * hotplug_action solution instead.
+ */
+ action = ap->ops->hotplug_action(ap);
+ spin_unlock(&ap->host_set->lock);
+
+ if (action)
+ ata_hotplug_plug_func(ap);
+ else
+ ata_hotplug_unplug_func(ap);
+}
+
+/* Should be protected by host_set->lock */
+void ata_hotplug_op(struct ata_port *ap)
+{
+ del_timer(&ap->hotplug_timer);
+
+ ap->hotplug_timer.expires = jiffies + ATA_TMOUT_HOTSWAP;
+ add_timer(&ap->hotplug_timer);
+}
+
/**
* ata_interrupt - Default ATA host interrupt handler
* @irq: irq line (unused)
@@ -3860,7 +3953,11 @@ static void ata_host_init(struct ata_por
ap->cbl = ATA_CBL_NONE;
ap->active_tag = ATA_TAG_POISON;
ap->last_ctl = 0xFF;
+ ap->orig_udma_mask = ent->udma_mask;
+ ap->hotplug_timer.function = ata_hotplug_timer_func;
+ ap->hotplug_timer.data = (unsigned int)ap;
+ init_timer(&ap->hotplug_timer);
INIT_WORK(&ap->packet_task, atapi_packet_task, ap);
INIT_WORK(&ap->pio_task, ata_pio_task, ap);
@@ -4531,6 +4628,7 @@ EXPORT_SYMBOL_GPL(ata_dev_classify);
EXPORT_SYMBOL_GPL(ata_dev_id_string);
EXPORT_SYMBOL_GPL(ata_dev_config);
EXPORT_SYMBOL_GPL(ata_scsi_simulate);
+EXPORT_SYMBOL_GPL(ata_hotplug_op);
#ifdef CONFIG_PCI
EXPORT_SYMBOL_GPL(pci_test_config_bits);
--- linux-2.6.13-rc3/drivers/scsi/libata-scsi.c.old 2005-08-01 00:22:01.000000000 -0700
+++ linux-2.6.13-rc3/drivers/scsi/libata-scsi.c 2005-08-01 00:22:26.000000000 -0700
@@ -1011,6 +1011,53 @@ static int ata_scsi_qc_complete(struct a
return 0;
}
+static int ata_scsi_qc_abort(struct ata_queued_cmd *qc, u8 drv_stat)
+{
+ struct scsi_cmnd *cmd = qc->scsicmd;
+
+ cmd->result = SAM_STAT_TASK_ABORTED; //FIXME: Is this what we want?
+
+ qc->scsidone(cmd);
+
+ return 0;
+}
+
+void ata_scsi_prepare_qc_abort(struct ata_queued_cmd *qc)
+{
+ // For some reason or another, we can't allow a normal scsi_qc_complete
+ if (qc->complete_fn == ata_scsi_qc_complete);
+ qc->complete_fn = ata_scsi_qc_abort;
+}
+
+void ata_scsi_handle_plug(struct ata_port *ap)
+{
+ //Currently SATA only
+ scsi_add_device(ap->host, 0, 0, 0);
+}
+
+void ata_scsi_handle_unplug(struct ata_port *ap)
+{
+ //SATA only, no PATA
+ struct scsi_device *scd = scsi_device_lookup(ap->host, 0, 0, 0);
+ /* scd might not exist; someone did 'echo "scsi remove-single-device
+ * ... " > /proc/scsi/scsi' or somebody was turning the key in the
+ * hotswap bay between on and off really really fast.
+ */
+ if (scd) {
+ scsi_device_set_state(scd, SDEV_CANCEL);
+ /* We might have a pending qc on I/O to a removed device,
+ * however, I argue it's impossible unless we have an 'scd'
+ * because it means we never completed a 'plug' into the system
+ * (or no device was present on bootup), so either we have no
+ * possible I/O, or a qc which 'ata_hotplug_plug_func' took
+ * care of
+ */
+ ata_check_kill_qc(ap);
+ scsi_remove_device(scd);
+ scsi_device_put(scd);
+ }
+}
+
/**
* ata_scsi_translate - Translate then issue SCSI command to ATA device
* @ap: ATA port to which the command is addressed
--- linux-2.6.13-rc3/drivers/scsi/libata.h.old 2005-08-01 00:22:05.000000000 -0700
+++ linux-2.6.13-rc3/drivers/scsi/libata.h 2005-08-01 00:22:26.000000000 -0700
@@ -40,6 +40,7 @@ extern struct ata_queued_cmd *ata_qc_new
extern void ata_qc_free(struct ata_queued_cmd *qc);
extern int ata_qc_issue(struct ata_queued_cmd *qc);
extern int ata_check_atapi_dma(struct ata_queued_cmd *qc);
+extern void ata_check_kill_qc(struct ata_port *ap);
extern void ata_dev_select(struct ata_port *ap, unsigned int device,
unsigned int wait, unsigned int can_sleep);
extern void ata_tf_to_host_nolock(struct ata_port *ap, struct ata_taskfile *tf);
@@ -76,6 +77,9 @@ extern void ata_scsi_badcmd(struct scsi_
extern void ata_scsi_rbuf_fill(struct ata_scsi_args *args,
unsigned int (*actor) (struct ata_scsi_args *args,
u8 *rbuf, unsigned int buflen));
+extern void ata_scsi_prepare_qc_abort(struct ata_queued_cmd *qc);
+extern void ata_scsi_handle_plug(struct ata_port *ap);
+extern void ata_scsi_handle_unplug(struct ata_port *ap);
static inline void ata_bad_scsiop(struct scsi_cmnd *cmd, void (*done)(struct scsi_cmnd *))
{
--- linux-2.6.13-rc3/include/linux/libata.h.old 2005-08-01 00:22:13.000000000 -0700
+++ linux-2.6.13-rc3/include/linux/libata.h 2005-08-01 02:13:57.000000000 -0700
@@ -29,6 +29,7 @@
#include <asm/io.h>
#include <linux/ata.h>
#include <linux/workqueue.h>
+#include <linux/timer.h>
/*
* compile-time options
@@ -127,6 +128,7 @@ enum {
ATA_TMOUT_BOOT_QUICK = 7 * HZ, /* hueristic */
ATA_TMOUT_CDB = 30 * HZ,
ATA_TMOUT_CDB_QUICK = 5 * HZ,
+ ATA_TMOUT_HOTSWAP = HZ / 2, /* Heuristic suggested by Sil */
/* ATA bus states */
BUS_UNKNOWN = 0,
@@ -319,6 +321,7 @@ struct ata_port {
struct ata_host_stats stats;
struct ata_host_set *host_set;
+ struct timer_list hotplug_timer;
struct work_struct packet_task;
struct work_struct pio_task;
@@ -326,6 +329,8 @@ struct ata_port {
unsigned long pio_task_timeout;
void *private_data;
+
+ unsigned int orig_udma_mask;
};
struct ata_port_operations {
@@ -372,6 +377,7 @@ struct ata_port_operations {
void (*bmdma_stop) (struct ata_port *ap);
u8 (*bmdma_status) (struct ata_port *ap);
+ int (*hotplug_action) (struct ata_port *ap);
};
struct ata_port_info {
@@ -402,6 +408,7 @@ extern int ata_scsi_queuecmd(struct scsi
extern int ata_scsi_error(struct Scsi_Host *host);
extern int ata_scsi_release(struct Scsi_Host *host);
extern unsigned int ata_host_intr(struct ata_port *ap, struct ata_queued_cmd *qc);
+extern void ata_hotplug_op(struct ata_port *ap);
/*
* Default driver ops implementations
*/
