On 6/17/22 06:57, Arnd Bergmann wrote:
From: Arnd Bergmann <a...@arndb.de>
The BusLogic driver is the last remaining driver that relies on the
deprecated bus_to_virt() function, which in turn only works on a few
architectures, and is incompatible with both swiotlb and iommu support.
Before commit 391e2f25601e ("[SCSI] BusLogic: Port driver to 64-bit."),
the driver had a dependency on x86-32, presumably because of this
problem. However, the change introduced another bug that made it still
impossible to use the driver on any 64-bit machine.
This was in turn fixed in commit 56f396146af2 ("scsi: BusLogic: Fix
64-bit system enumeration error for Buslogic"), 8 years later, which
shows that there are not a lot of users.
Maciej is still using the driver on 32-bit hardware, and Khalid mentioned
that the driver works with the device emulation used in VirtualBox
and VMware. Both of those only emulate it for Windows 2000 and older
operating systems that did not ship with the better LSI logic driver.
Do a minimum fix that searches through the list of descriptors to find
one that matches the bus address. This is clearly as inefficient as
was indicated in the code comment about the lack of a bus_to_virt()
replacement. A better fix would likely involve changing out the entire
descriptor allocation for a simpler one, but that would be much
more invasive.
Cc: Maciej W. Rozycki <ma...@orcam.me.uk>
Cc: Matt Wang <wwen...@vmware.com>
Cc: Khalid Aziz <kha...@gonehiking.org>
Signed-off-by: Arnd Bergmann <a...@arndb.de>
---
drivers/scsi/BusLogic.c | 27 ++++++++++++++++-----------
drivers/scsi/Kconfig | 2 +-
2 files changed, 17 insertions(+), 12 deletions(-)
diff --git a/drivers/scsi/BusLogic.c b/drivers/scsi/BusLogic.c
index a897c8f914cf..d057abfcdd5c 100644
--- a/drivers/scsi/BusLogic.c
+++ b/drivers/scsi/BusLogic.c
@@ -2515,12 +2515,26 @@ static int blogic_resultcode(struct blogic_adapter
*adapter,
return (hoststatus << 16) | tgt_status;
}
+/*
+ * turn the dma address from an inbox into a ccb pointer
+ * This is rather inefficient.
+ */
+static struct blogic_ccb *
+blogic_inbox_to_ccb(struct blogic_adapter *adapter, struct blogic_inbox *inbox)
+{
+ struct blogic_ccb *ccb;
+
+ for (ccb = adapter->all_ccbs; ccb; ccb = ccb->next_all)
+ if (inbox->ccb == ccb->dma_handle)
+ break;
+
+ return ccb;
+}
/*
blogic_scan_inbox scans the Incoming Mailboxes saving any
Incoming Mailbox entries for completion processing.
*/
-
static void blogic_scan_inbox(struct blogic_adapter *adapter)
{
/*
@@ -2540,16 +2554,7 @@ static void blogic_scan_inbox(struct blogic_adapter
*adapter)
enum blogic_cmplt_code comp_code;
while ((comp_code = next_inbox->comp_code) != BLOGIC_INBOX_FREE) {
- /*
- We are only allowed to do this because we limit our
- architectures we run on to machines where bus_to_virt(
- actually works. There *needs* to be a dma_addr_to_virt()
- in the new PCI DMA mapping interface to replace
- bus_to_virt() or else this code is going to become very
- innefficient.
- */
- struct blogic_ccb *ccb =
- (struct blogic_ccb *) bus_to_virt(next_inbox->ccb);
+ struct blogic_ccb *ccb = blogic_inbox_to_ccb(adapter,
adapter->next_inbox);
This change looks good enough as workaround to not use bus_to_virt() for
now. There are two problems I see though. One, I do worry about
blogic_inbox_to_ccb() returning NULL for ccb which should not happen
unless the mailbox pointer was corrupted which would indicate a bigger
problem. Nevertheless a NULL pointer causing kernel panic concerns me.
How about adding a check before we dereference ccb?
Second, with this patch applied, I am seeing errors from the driver:
=====================
[ 1623.902685] sdb: sdb1 sdb2
[ 1623.903245] sd 2:0:0:0: [sdb] Attached SCSI disk
[ 1623.911000] scsi2: Illegal CCB #76 status 2 in Incoming Mailbox
[ 1623.911005] scsi2: Illegal CCB #76 status 2 in Incoming Mailbox
[ 1623.911070] scsi2: Illegal CCB #79 status 2 in Incoming Mailbox
[ 1651.458008] scsi2: Warning: Partition Table appears to have Geometry
256/63 which is
[ 1651.458013] scsi2: not compatible with current BusLogic Host Adapter
Geometry 255/63
[ 1658.797609] scsi2: Resetting BusLogic BT-958D Failed
[ 1659.533208] sd 2:0:0:0: Device offlined - not ready after error recovery
[ 1659.533331] sd 2:0:0:0: Device offlined - not ready after error recovery
[ 1659.533333] sd 2:0:0:0: Device offlined - not ready after error recovery
[ 1659.533342] sd 2:0:0:0: [sdb] tag#101 FAILED Result:
hostbyte=DID_TIME_OUT driverbyte=DRIVER_OK cmd_age=35s
[ 1659.533345] sd 2:0:0:0: [sdb] tag#101 CDB: Read(10) 28 00 00 00 00 28
00 00 10 00
[ 1659.533346] I/O error, dev sdb, sector 40 op 0x0:(READ) flags 0x80700
phys_seg 1 prio class 0
=================
This is on VirtualBox using emulated BusLogic adapter.
This patch needs more refinement.
Thanks,
Khalid
if (comp_code != BLOGIC_Cn erroneousMD_NOTFOUND) {
if (ccb->status == BLOGIC_CCB_ACTIVE ||
ccb->status == BLOGIC_CCB_RESET) {
diff --git a/drivers/scsi/Kconfig b/drivers/scsi/Kconfig
index cf75588a2587..56bdc08d0b77 100644
--- a/drivers/scsi/Kconfig
+++ b/drivers/scsi/Kconfig
@@ -513,7 +513,7 @@ config SCSI_HPTIOP
config SCSI_BUSLOGIC
tristate "BusLogic SCSI support"
- depends on PCI && SCSI && VIRT_TO_BUS
+ depends on PCI && SCSI
help
This is support for BusLogic MultiMaster and FlashPoint SCSI Host
Adapters. Consult the SCSI-HOWTO, available from
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu