- 07 Jan, 2016 40 commits
-
-
Finn Thain authored
Because of the rudimentary design of the chip, it is necessary to poll the SCSI bus signals during PIO and this tends to hog the CPU. The driver will accept new commands while others execute, and this causes a soft lockup because the workqueue item will not terminate until the issue queue is emptied. When exercising dmx3191d using sequential IO from dd, the driver is sent 512 KiB WRITE commands and 128 KiB READs. For a PIO transfer, the rate is is only about 300 KiB/s, so these are long-running commands. And although PDMA may run at several MiB/s, interrupts are disabled for the duration of the transfer. Fix the unresponsiveness and soft lockup issues by calling cond_resched() after each command is completed and by limiting max_sectors for drivers that don't implement real DMA. Signed-off-by:
Finn Thain <fthain@telegraphics.com.au> Reviewed-by:
Hannes Reinecke <hare@suse.com> Tested-by:
Ondrej Zary <linux@rainbow-software.org> Tested-by:
Michael Schmitz <schmitzmic@gmail.com> Signed-off-by:
Martin K. Petersen <martin.petersen@oracle.com>
-
Finn Thain authored
This refactoring removes two global Scsi_Host pointers. This improves consistency with other ncr5380 drivers. Adopting the same conventions as the other drivers makes them easier to read. Signed-off-by:
-
Finn Thain authored
Keep the two core driver forks in sync. Signed-off-by:
-
Finn Thain authored
Remove the HOSTNO macro that is peculiar to atari_NCR5380.c and contributes to the problem of divergence of the NCR5380 core drivers. Keep NCR5380.c in sync. Signed-off-by:
-
Finn Thain authored
NCR5380.c lacks a sane eh_bus_reset_handler. The atari_NCR5380.c code is much better but it should not throw out the issue queue (that would be a host reset) and it neglects to set the result code for commands that it throws out. Fix these bugs and keep the two core drivers in sync. Signed-off-by:
-
Finn Thain authored
During arbitration and selection, the relevant command is invisible to exception handlers and can be found only in a pointer on the stack of a different thread. When eh_abort_handler can't find a given command, it can't decide whether that command was completed already or is still in arbitration or selection phase. But it must return either SUCCESS (e.g. command completed earlier) or FAILED (could not abort the nexus, try bus reset). The solution is to make sure all commands belonging to the LLD are always visible to exception handlers. Add another scsi_cmnd pointer to the hostdata struct to track the command in arbitration or selection phase. Replace 'retain_dma_irq' with the new 'selecting' pointer, to bring atari_NCR5380.c into line with NCR5380.c. Signed-off-by:
-
Finn Thain authored
Introduce a new eh_abort_handler implementation. This one attempts to follow all of the rules relating to EH handlers. There is still a known bug: during selection, a command becomes invisible to the EH handlers because it only appears in a pointer on the stack of a different thread. This bug is addressed in a subsequent patch. Signed-off-by:
-
Finn Thain authored
NCR5380_information_transfer() may re-queue a command for autosense, after calling scsi_eh_prep_cmnd(). This creates several possibilities: 1. Reselection may intervene before the re-queued command gets processed. If the reconnected command then undergoes autosense, this causes the scsi_eh_save data from the previous command to be overwritten. 2. After NCR5380_information_transfer() calls scsi_eh_prep_cmnd(), a new REQUEST SENSE command may arrive. This would be queued ahead of any command already undergoing autosense, which means the scsi_eh_save data might be restored to the wrong command. 3. After NCR5380_information_transfer() calls scsi_eh_prep_cmnd(), eh_abort_handler() may abort the command. But the scsi_eh_save data is not discarded, which means the scsi_eh_save data might be incorrectly restored to the next REQUEST SENSE command issued. This patch adds a new autosense list so that commands that are re-queued because of a CHECK CONDITION result can be kept apart from the REQUEST SENSE commands that arrive via queuecommand. This patch also adds a function dedicated to dequeueing and preparing the next command for processing. By refactoring the main loop in this way, scsi_eh_save takes place when an autosense command is dequeued rather than when re-queued. Signed-off-by:
-
Finn Thain authored
Implement a 'complete_cmd' function to complete commands. This is needed by the following patch; the new function provides a site for the logic needed to correctly handle REQUEST SENSE commands. Signed-off-by:
-
Finn Thain authored
The NCR5380 drivers have a home-spun linked list implementation for scsi_cmnd structs that uses cmd->host_scribble as a 'next' pointer. Adopt the standard list_head data structure and list operations instead. Remove the eh_abort_handler rather than convert it. Doing the conversion would only be churn because the existing EH handlers don't work and get replaced in a subsequent patch. Signed-off-by:
-
Finn Thain authored
The hostdata struct is now protected by a spin lock so the volatile qualifiers are redundant. Remove them. Signed-off-by:
-
Finn Thain authored
Printing command pointers can be useful when debugging queues. Other than that, the LIST and REMOVE macros are just clutter. These macros are redundant now that NDEBUG_QUEUES causes pointers to be printed, so remove them. Signed-off-by:
-
Finn Thain authored
Print the command pointers in the log messages for debugging queue data structures. The LIST and REMOVE macros can then be removed. Signed-off-by:
-
Finn Thain authored
Make use of the shost_priv() helper. Remove HOSTDATA and SETUP_HOSTDATA macros because they harm readability. Signed-off-by:
-
Finn Thain authored
Replace all H_NO and some HOSTNO macros (both peculiar to atari_NCR5380.c) with a new dsprintk macro that's more useful and more consistent. The new macro avoids a lot of boilerplate in new code in subsequent patches. Keep NCR5380.c in sync. Remaining HOSTNO macros are removed as side-effects of subsequent patches. Signed-off-by:
-
Finn Thain authored
Some NCR5380 hosts offer a .show_info method to access the contents of the various command list data structures from a procfs file. When NDEBUG is set, the same information is sent to the console during EH. The two core drivers, atari_NCR5380.c and NCR5380.c differ here. Because it is just for debugging, the easiest way to fix the discrepancy is simply remove this code. The only remaining users of NCR5380_show_info() and NCR5380_write_info() are drivers that define PSEUDO_DMA. The others have no use for the .show_info method, so don't initialize it. Signed-off-by:
-
Finn Thain authored
NCR5380.c presently uses the instance->host_lock spin lock. Convert this to a new spin lock that protects the NCR5380_hostdata struct. atari_NCR5380.c previously used local_irq_save/restore() rather than a spin lock. Convert this to hostdata->lock in irq mode. For SMP platforms, the interrupt handler now also acquires the spin lock. This brings all locking in the two core drivers into agreement. Adding this locking also means that a bunch of volatile qualifiers can be removed from the members of the NCR5380_hostdata struct. This is done in a subsequent patch. Proper locking will allow the abort handler to locate a command being aborted. This is presently impossible if the abort handler is invoked when the command has been moved from a queue to a pointer on the stack. (If eh_abort_handler can't determine whether a command has been completed or is still being processed then it can't decide whether to return success or failure.) The hostdata spin lock is now held when calling NCR5380_select() and NCR5380_information_transfer(). Where possible, the lock is dropped for polling and PIO transfers. Clean up the now-redundant SELECT_ENABLE_REG writes, that used to provide limited mutual exclusion between information_transfer() and reselect(). Accessing hostdata->connected without data races means taking the lock; cleanup these accesses. The new spin lock falls away for m68k and other UP builds, so this should have little impact there. In the SMP case the new lock should be uncontested even when the SCSI bus is contested. Signed-off-by:
-
Finn Thain authored
Remove FLAG_DTC3181E. It was used to suppress a final Arbitration Lost (SEL asserted) test that isn't actually needed. The test was suppressed because it causes problems for DTC436 and DTC536 chips. It takes place after the host wins arbitration, so SEL has been asserted. These chips can't seem to tell whether it was the host or another bus device that did so. This questionable final test appears in a flow chart in an early NCR5380 datasheet. It was removed from later documents like the DP5380 datasheet. By the time this final test takes place, the driver has already tested the Arbitration Lost bit several times. The first test happens 3 us after BUS FREE (or longer due to register access delays). The protocol requires that a device stop signalling within 1.8 us after BUS FREE unless it won arbitration, in which case it must assert SEL, which is detected 1.2 us later by the first Arbitration Lost test. Signed-off-by:
-
Finn Thain authored
When a target reports a QUEUE_FULL condition it causes the driver to update the 'queue_size' limit with the number of currently allocated tags. At least, that's what's supposed to happen, according to the comments. Unfortunately the terms in the assignment are swapped. Fix this and cleanup some obsolete comments. Signed-off-by:
-
Finn Thain authored
Add missing .module initializer. Use distinct .proc_name values for the g_NCR5380 and g_NCR5380_mmio modules. Remove pointless CAN_QUEUE and CMD_PER_LUN override macros. Cleanup whitespace and code style. Signed-off-by:
-
Finn Thain authored
NDEBUG_NO_DATAOUT should not disable DATA IN phases too. Fix this. (This bug has long been fixed in atari_NCR5380.c.) Signed-off-by:
-
Finn Thain authored
Remove unused includes (stat.h, signal.h, proc_fs.h) and move includes needed by the core drivers into the common header (delay.h etc). Signed-off-by:
-
Finn Thain authored
Fix the array bounds check when transferring an extended message from the target. Signed-off-by:
-
Finn Thain authored
Bring the two NCR5380_reselect() implementations into agreement. Replace infinite loops in atari_NCR5380.c with timeouts, as per NCR5380.c. Remove 'abort' flag in NCR5380.c as per atari_NCR5380.c -- if reselection fails, there may be no MESSAGE IN phase so don't attempt data transfer. During selection, don't interfere with the chip registers after a reselection interrupt intervenes. Clean up some trivial issues with code style, comments and printk. Signed-off-by:
-
Finn Thain authored
The workarounds for chip errata appear twice, in slightly different forms. One is used when defined(REAL_DMA) || defined(REAL_DMA_POLL), the other when defined(PSEUDO_DMA). In the PDMA case, the workarounds have been made conditional on FLAG_NO_DMA_FIXUPS. Do the same for the DMA case, to eliminate the READ_OVERRUNS macro. Signed-off-by:
-
Finn Thain authored
The flags DMA_WORKS_RIGHT, FLAG_NCR53C400 and FLAG_HAS_LAST_BYTE_SENT all mean the same thing, i.e. the chip is not a 538[01]. (More recent devices such as the 53C80 have a 'Last Byte Sent' bit in the Target Command Register as well as other fixes for End-of-DMA errata.) These flags have no additional meanings since previous cleanup patches eliminated the NCR53C400 macro, moved g_NCR5380-specific code out of the core driver and standardized interrupt handling. Use the FLAG_NO_DMA_FIXUP flag to suppress End-of-DMA errata workarounds, for those cards and drivers that make use of the TCR_LAST_BYTE_SENT bit. Remove the old flags. Signed-off-by:
-
Finn Thain authored
SCSI bus protocol sometimes requires monitoring two related conditions simultaneously. Enhance NCR5380_poll_politely() for this purpose, and put it to use in the arbitration algorithm. It will also find use in pseudo DMA. Signed-off-by:
-
Finn Thain authored
Because interrupt handling is crucial to the core driver(s), all wrapper drivers need to agree on this code. This patch removes discrepancies. NCR5380_intr() in NCR5380.c has the following pointless loop that differs from the code in atari_NCR5380.c. done = 1; do { /* ... */ } while (!done); The 'done' flag gets cleared when a reconnected command is to be processed from the work queue. But in NCR5380.c, the flag is also used to cause the interrupt conditions to be re-examined. Perhaps this was because NCR5380_reselect() was expected to cause another interrupt, or perhaps the remaining present interrupt conditions need to be handled after the NCR5380_reselect() call? Actually, both possibilities are bogus, as is the loop itself. It seems have been overlooked in the hit-and-miss removal of scsi host instance list iteration many years ago; see history/history.git commit 491447e1 ("[PATCH] next NCR5380 updates") and commit 69e1a948 ("[PATCH] fix up NCR5380 private data"). See also my earlier patch, "Always retry arbitration and selection". The datasheet says, "IRQ can be reset simply by reading the Reset Parity/Interrupt Register". So don't treat the chip IRQ like a level-triggered interrupt. Of the conditions that set the IRQ flag, some are level-triggered and some are edge-triggered, which means IRQ itself must be edge-triggered. Some interrupt conditions are latched and some are not. Before clearing the chip IRQ flag, clear all state that may cause it to be raised. That means clearing the DMA Mode and Busy Monitor bits in the Mode Register and clearing the host ID in the Select Enable register. Also clean up some printk's and some comments. Keep atari_NCR5380.c and NCR5380.c in agreement. Signed-off-by: -
Finn Thain authored
Configuring core drivers using macros like this one prevents re-unifying the core driver forks, and prevents implementing the core driver as a library or a platform driver. The UNSAFE macro in particular is a poor workaround for the problem of interrupt latency. Releasing the locks complicates things because then we would have to handle the possibility of EH handler invocation during a PDMA transfer. The comments say that instead of using this macro, "you're going to be better off twiddling with transfersize". I agree. Remove this stuff. Signed-off-by:
-
Finn Thain authored
The complex main_running/queue_main mechanism is peculiar to atari_NCR5380.c. It isn't SMP safe and offers little value given that the work queue already offers concurrency management. Remove this complexity to bring atari_NCR5380.c closer to NCR5380.c. It is not a good idea to call the information transfer state machine from queuecommand because, according to Documentation/scsi/scsi_mid_low_api.txt that could happen in soft irq context. Fix this. Signed-off-by:
-
Finn Thain authored
Each host instance now has it's own work queue so the main() work item can sleep when necessary. That means we can use a simple work item rather than a delayed work item. This brings NCR5380.c closer to atari_NCR5380.c. Signed-off-by:
-
Finn Thain authored
When there is a queued command and no connected command, NCR5380_select() is called and arbitration begins. The chip waits for BUS FREE once the MR_ARBITRATE bit in the mode register is enabled. That means there is no need to wait for BUS FREE after disconnecting. There is presently no polling for BUS FREE after sending an ABORT or other message that might lead to disconnection. It only happens after COMMAND COMPLETE or DISCONNECT messages, which seems inconsistent. Remove the polling for !BSY in the COMMAND COMPLETE and DISCONNECT cases. BTW, the comments say "avoid nasty timeouts" and perhaps BUS FREE polling was somehow helpful back in Linux v0.99.14u, when it was introduced. The relevant timeout is presently 1 second (for bus arbitration). Signed-off-by:
-
Finn Thain authored
Allow target selection to fail with a timeout instead of waiting in infinite loops. This gets rid of the unused NCR_TIMEOUT macro, it is more defensive and has proved helpful in debugging. Signed-off-by:
-
Finn Thain authored
Use timeouts in do_abort() in atari_NCR5380.c instead of infinite loops. Also fix the kernel-doc comment. Keep the two core driver forks in sync. Signed-off-by:
-
Finn Thain authored
NCR5380_poll_politely() returns either 0 (success) or -ETIMEDOUT. However, in do_abort(), the return value is incorrectly taken to be the status register value. This means that the bus is put into DATA OUT phase instead of MESSAGE OUT. Fix this. Signed-off-by:
-
Finn Thain authored
NCR5380_poll_politely() never returns -1. That means do_abort() can fail to handle a timeout after waiting for the target to negate REQ. Fix this and cleanup other NCR5380_poll_politely() call sites. Signed-off-by:
-
Finn Thain authored
MESSAGE REJECT does not imply DISCONNECT: the target is about to enter MESSAGE IN or MESSAGE OUT phase. This bug fix comes from atari_NCR5380.c. Unfortunately it never made it into the original NCR5380.c core driver. Signed-off-by:
-
Hannes Reinecke authored
Some old drivers partially implemented support for linked commands using a "proposed" next_link pointer in struct scsi_cmnd that never actually existed. Remove this code. Signed-off-by:
-
Finn Thain authored
Remove the DEF_SCSI_QCMD macro (already removed from atari_NCR5380.c). The lock provided by DEF_SCSI_QCMD is only needed for queue data structures. Signed-off-by:
-
Finn Thain authored
The host spin lock needs to be acquired by NCR5380_abort() before it calls NCR5380_select(). This patch doesn't actually fix the EH issues in this driver but it does avoid this: BUG: spinlock already unlocked on CPU#0, kworker/u4:1/14 lock: 0xc0c0f834, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1 CPU: 0 PID: 14 Comm: kworker/u4:1 Not tainted 3.15.5 #5 Workqueue: scsi_tmf_4 scmd_eh_abort_handler Call Trace: [ef885d70] [c0008acc] show_stack+0x70/0x1bc (unreliable) [ef885db0] [c0492a00] dump_stack+0x84/0x684 [ef885dc0] [c006f314] spin_dump+0xd0/0xe8 [ef885dd0] [c006f460] do_raw_spin_unlock+0xd4/0xd8 [ef885df0] [c0491c8c] _raw_spin_unlock_irq+0x10/0x3c [ef885e00] [f381fe3c] NCR5380_select+0x3e4/0x6e8 [dmx3191d] [ef885e40] [f382026c] NCR5380_abort+0x12c/0x190 [dmx3191d] [ef885e60] [c02fec9c] scmd_eh_abort_handler+0x100/0x460 [ef885e80] [c0046470] process_one_work+0x16c/0x420 [ef885ea0] [c0046870] worker_thread+0x14c/0x430 [ef885ed0] [c004e4f4] kthread+0xd8/0xec [ef885f40] [c00124d4] ret_from_kernel_thread+0x5c/0x64 Signed-off-by:
-
