On 10/02/2022 11:42, Damien Le Moal wrote:
In pm80xx_send_abort_all(), the n_elem field of the ccb used is not
initialized to 0. This missing initialization sometimes lead to the
task completion path seeing the ccb with a non-zero n_elem resulting in
the execution of invalid dma_unmap_sg() calls in pm8001_ccb_task_free(),
causing a crash such as:
[ 197.676341] RIP: 0010:iommu_dma_unmap_sg+0x6d/0x280
[ 197.700204] RSP: 0018:ffff889bbcf89c88 EFLAGS: 00010012
[ 197.705485] RAX: dffffc0000000000 RBX: 0000000000000000 RCX: ffffffff83d0bda0
[ 197.712687] RDX: 0000000000000002 RSI: 0000000000000000 RDI: ffff88810dffc0d0
[ 197.719887] RBP: 0000000000000000 R08: 0000000000000000 R09: ffff8881c790098b
[ 197.727089] R10: ffffed1038f20131 R11: 0000000000000001 R12: 0000000000000000
[ 197.734296] R13: ffff88810dffc0d0 R14: 0000000000000010 R15: 0000000000000000
[ 197.741493] FS: 0000000000000000(0000) GS:ffff889bbcf80000(0000) knlGS:0000000000000000
[ 197.749659] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 197.755459] CR2: 00007f16c1b42734 CR3: 0000000004814000 CR4: 0000000000350ee0
[ 197.762656] Call Trace:
[ 197.765127] <IRQ>
[ 197.767162] pm8001_ccb_task_free+0x5f1/0x820 [pm80xx]
[ 197.772364] ? do_raw_spin_unlock+0x54/0x220
[ 197.776680] pm8001_mpi_task_abort_resp+0x2ce/0x4f0 [pm80xx]
[ 197.782406] process_oq+0xe85/0x7890 [pm80xx]
[ 197.786817] ? lock_acquire+0x194/0x490
[ 197.790697] ? handle_irq_event+0x10e/0x1b0
[ 197.794920] ? mpi_sata_completion+0x2d70/0x2d70 [pm80xx]
[ 197.800378] ? __wake_up_bit+0x100/0x100
[ 197.804340] ? lock_is_held_type+0x98/0x110
[ 197.808565] pm80xx_chip_isr+0x94/0x130 [pm80xx]
[ 197.813243] tasklet_action_common.constprop.0+0x24b/0x2f0
[ 197.818785] __do_softirq+0x1b5/0x82d
[ 197.822485] ? do_raw_spin_unlock+0x54/0x220
[ 197.826799] __irq_exit_rcu+0x17e/0x1e0
[ 197.830678] irq_exit_rcu+0xa/0x20
[ 197.834114] common_interrupt+0x78/0x90
[ 197.840051] </IRQ>
[ 197.844236] <TASK>
[ 197.848397] asm_common_interrupt+0x1e/0x40
That's nasty.
Avoid this issue by always initializing the ccb n_elem field to 0 in
pm8001_send_abort_all(), pm8001_send_read_log() and
pm80xx_send_abort_all().
Fixes: c6b9ef5779c3 ("[SCSI] pm80xx: NCQ error handling changes")
Signed-off-by: Damien Le Moal <damien.lemoal@xxxxxxxxxxxxxxxxxx>
---
drivers/scsi/pm8001/pm8001_hwi.c | 2 ++
drivers/scsi/pm8001/pm80xx_hwi.c | 1 +
2 files changed, 3 insertions(+)
diff --git a/drivers/scsi/pm8001/pm8001_hwi.c b/drivers/scsi/pm8001/pm8001_hwi.c
index 8095eb0b04f7..d853e8d0195a 100644
--- a/drivers/scsi/pm8001/pm8001_hwi.c
+++ b/drivers/scsi/pm8001/pm8001_hwi.c
@@ -1788,6 +1788,7 @@ static void pm8001_send_abort_all(struct pm8001_hba_info *pm8001_ha,
ccb->device = pm8001_ha_dev;
ccb->ccb_tag = ccb_tag;
ccb->task = task;
+ ccb->n_elem = 0;
Do you think that it would be better to clear this field when we free
the tag/ccb in pm8001_ccb_task_free()? I will note that there are error
paths whch only free the tag (not ccb), so need to be careful there.
BTW, I see that this never landed:
https://lore.kernel.org/lkml/20211214090337.29156-1-niejianglei2021@xxxxxxx/
Though alloc'ing a domain_device in pm8001_send_read_log() is questionable.
Thanks,
John
circularQ = &pm8001_ha->inbnd_q_tbl[0];
@@ -1849,6 +1850,7 @@ static void pm8001_send_read_log(struct pm8001_hba_info *pm8001_ha,
ccb->device = pm8001_ha_dev;
ccb->ccb_tag = ccb_tag;
ccb->task = task;
+ ccb->n_elem = 0;
pm8001_ha_dev->id |= NCQ_READ_LOG_FLAG;
pm8001_ha_dev->id |= NCQ_2ND_RLE_FLAG;
diff --git a/drivers/scsi/pm8001/pm80xx_hwi.c b/drivers/scsi/pm8001/pm80xx_hwi.c
index 4d88c0dbcefc..902af4eefa26 100644
--- a/drivers/scsi/pm8001/pm80xx_hwi.c
+++ b/drivers/scsi/pm8001/pm80xx_hwi.c
@@ -1801,6 +1801,7 @@ static void pm80xx_send_abort_all(struct pm8001_hba_info *pm8001_ha,
ccb->device = pm8001_ha_dev;
ccb->ccb_tag = ccb_tag;
ccb->task = task;
+ ccb->n_elem = 0;
circularQ = &pm8001_ha->inbnd_q_tbl[0];
