Changeset 80 for trunk/src/os2ahci/os2ahci.c

Timestamp:

Mar 1, 2011, 10:10:11 AM (14 years ago)

Author:

chris

Message:

Version 1.06
============

• Finally came across a BIOS which accesses the ICH7/8 controller via SATA registers (i.e. not AHCI mode). This required a few changes to the code at boot time because it turned out that COMRESETs are required whenever switching to/from AHCI mode to allow the AHCI or SATA controller to re-discover the attached devices:

• 'init_reset' will now be forced on when finding a controller in non-AHCI mode at boot time.
• A COMRESET is initiated for each implemented port after turning off AHCI mode when restoring the BIOS configuration; this is done only for Intel controllers at this point because they map the AHCI port SCR MMIO registers even when not in AHCI mode.
• apm_suspend() has been adjusted to restore the BIOS configuration to prevent needless timeouts when the BIOS takes over during suspend or power-off operations.
• Small changes to the functions which save/restore BIOS/port settings to avoid pitfalls; among others, the port save/restore code now also saves and restores the port's engine status.

• Improvements to debug logging around port resets.

• Moved code to clear pending interrupts from ahci_reset_port() to ahci_stop_port() because both need it and resetting a port involves stopping it, first.

• NCQ mode has found to cause problems on a Dell D630. This may be related to the hard disk used for the test but since I've never seen more than one queued command regardless of the I/O load (even during simulaneous xcopy operations), NCQ mode is now off by default and needs to be turned on via the /N switch (i.e. the the /N switch now has a reversed meaning).

• Removed the code which attempts to establish another MMIO base address in case the one assigned by the BIOS can't be reserved via resource manager; if there's a conflict, it's extremely unlikely we would ever be able to restore the BIOS MMIO address at boot time without the BIOS clashing with whatever conflicts with the MMIO address, thus there's no point trying to do any of this.

• Implemented a reset context hook watchdog; in the early boot phase, some components apparently don't yield the CPU so the context hook will never execute without the watchdog. Now we'll give the context hook 10 seconds to execute, otherwise the watchdog will expire and we'll call the context hook directly from the corresponding timer callback.

File:

• trunk/src/os2ahci/os2ahci.c (modified) (6 diffs)

trunk/src/os2ahci/os2ahci.c

@@ -101 +101 @@
 /* apapter/port-specific options saved when parsing the command line */
 u8              emulate_scsi[MAX_AD][AHCI_MAX_PORTS];
-u8              disable_ncq[MAX_AD][AHCI_MAX_PORTS];
+u8              enable_ncq[MAX_AD][AHCI_MAX_PORTS];
 u8              link_speed[MAX_AD][AHCI_MAX_PORTS];
 u8              link_power[MAX_AD][AHCI_MAX_PORTS];
@@ -261 +261 @@
       case 'n':
         /* disable NCQ */
-        set_port_option(disable_ncq, !invert_option);
+        set_port_option(enable_ncq, !invert_option);
         break;
 
@@ -789 +789 @@
  *        requests.
  *
- *        In order to support applications that can't deal with ATAPI devies
+ *        In order to support applications that can't deal with ATAPI devices
  *        (i.e. need a SCSI adapter) os2ahci will optionally report ATAPI
  *        dvices as SCSI devices. The corresponding SCSI adapter doesn't
- *        really  exist and is only reported here for the IOCM_GET_DEVICETABLE
+ *        really exist and is only reported here for the IOCM_GET_DEVICETABLE
  *        request. The units attached to this adapter will use the real HW
  *        unit IDs, thus we'll never receive a command specific to the
@@ -1164 +1164 @@
 
 /******************************************************************************
- * Complete an IORB. It should be called without the adapter-level spinlock
+ * Complete an IORB. This should be called without the adapter-level spinlock
  * to allow the IORB completion routine to perform whatever processing it
  * requires. This implies that the IORB should no longer be in any global
@@ -1220 +1220 @@
 /******************************************************************************
  * Lock the adapter, waiting for availability if necessary. This is expected
- * to be called without the driver-level spinlock aquired.
+ * to be called at task/request time without the driver-level spinlock
+ * aquired. Don't call at interrupt time.
  */
 void lock_adapter(AD_INFO *ai)
@@ -1282 +1283 @@
    */
   DevHelp_ArmCtxHook(0, reset_ctxhook_h);
+
+  /* Set up a watchdog timer which calls the context hook manually in case
+   * some kernel thread is looping around the IORB_COMPLETE status bit
+   * without yielding the CPU (kernel threads don't preempt). This shouldn't
+   * happen per design because kernel threads are supposed to yield but it
+   * does in the early boot phase.
+   */
+  ADD_StartTimerMS(&th_reset_watchdog, 5000, (PFN) reset_watchdog, 0, 0);
+}
+
+/******************************************************************************
+ * Reset handler watchdog. If a timeout occurs, a context hook is armed which
+ * will execute as soon as a kernel thread yields the CPU. However, some
+ * kernel components won't yield the CPU during the early boot phase and the
+ * only way to kick some sense into those components is to run the context
+ * hook right inside this timer callback. Not exactly pretty, especially
+ * considering the fact that context hooks were implemented to prevent running
+ * lengthy operations like a port reset at task time, but without this
+ * watchdog mechanism we run the risk of getting completely stalled by device
+ * problems during the early boot phase.
+ */
+void _cdecl _far reset_watchdog(ULONG timer_handle, ULONG p1,
+                                ULONG p2)
+{
+  /* reset watchdog timer */
+  ADD_CancelTimer(timer_handle);
+  dprintf("reset watchdog invoked\n");
+
+  /* call context hook manually */
+  reset_ctxhook(0);
 }