source: trunk/src/os2ahci/ahci.c@ 32

/******************************************************************************
 * ahci.c - ahci hardware access functions
 *
 * Copyright (c) 2010 Christian Mueller. Parts copied from/inspired by the
 * Linux AHCI driver; those parts are (c) Linux AHCI/ATA maintainers
 *
 *  This program is free software; you can redistribute it and/or modify
 *  it under the terms of the GNU General Public License as published by
 *  the Free Software Foundation; either version 2 of the License, or
 *  (at your option) any later version.
 *
 *  This program is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with this program; if not, write to the Free Software
 *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 */

#include "os2ahci.h"
#include "ata.h"
#include "atapi.h"

/* -------------------------- macros and constants ------------------------- */

/* produce ata/atapi function pointer with the given func name */
#define cmd_func(iorb, func)  ad_infos[iorb_unit_adapter(iorb)].     \
                                 ports[iorb_unit_port(iorb)].        \
                                  devs[iorb_unit_device(iorb)].atapi \
                              ? atapi_##func : ata_##func


/* ------------------------ typedefs and structures ------------------------ */

/* -------------------------- function prototypes -------------------------- */

static void ahci_setup_device     (AD_INFO *ai, int p, int d, u16 *id_buf);
static void _cdecl _far timeout_callback (ULONG timer_handle, ULONG p1, ULONG p2);

/* ------------------------ global/static variables ------------------------ */

/* Initial driver status flags indexed by the board_* constants in os2ahci.h
 *
 * NOTE: The Linux AHCI driver uses a combination of board-specific quirk
 *       flags and overriding certain libata service functions to handle
 *       adapter flaws. However, there were only three overrides at the time
 *       os2ahci was written, one for hard adapter resets and two for port
 *       resets, and we can easily implement those within the corresponding
 *       reset handlers. If this becomes more complex, this array of flags
 *       should be converted into a structure array which contains function
 *       pointers to all handler functions which may need to be overridden.
 */
u16 initial_flags[] = {
  0,                                        /* board_ahci */
  AHCI_HFLAG_NO_NCQ |                       /* board_ahci_vt8251 */
    AHCI_HFLAG_NO_PMP,
  AHCI_HFLAG_IGN_IRQ_IF_ERR,                /* board_ahci_ign_iferr */
  AHCI_HFLAG_IGN_SERR_INTERNAL |            /* board_ahci_sb600 */
    AHCI_HFLAG_NO_MSI |
    AHCI_HFLAG_SECT255 |
    AHCI_HFLAG_32BIT_ONLY,
  AHCI_HFLAG_NO_NCQ |                       /* board_ahci_mv */
    AHCI_HFLAG_NO_MSI |
    AHCI_HFLAG_MV_PATA |
    AHCI_HFLAG_NO_PMP,
  AHCI_HFLAG_IGN_SERR_INTERNAL,             /* board_ahci_sb700 */
  AHCI_HFLAG_YES_NCQ,                       /* board_ahci_mcp65 */
  AHCI_HFLAG_NO_PMP,                        /* board_ahci_nopmp */
  AHCI_HFLAG_YES_NCQ,                       /* board_ahci_yesncq */
  AHCI_HFLAG_NO_SNTF,                       /* board_ahci_nosntf */
};

/* IRQ levels for stub interrupt handlers. OS/2 calls interrupt handlers
 * without passing the IRQ level, yet it expects the interrupt handler to
 * know the IRQ level for EOI processing. Thus we need multiple interrupt
 * handlers, one for each IRQ, and some mapping from the interrupt handler
 * index to the corresponding IRQ.
 */
static u16      irq_map[MAX_AD];   /* IRQ level for each stub IRQ func */
static int      irq_map_cnt;       /* number of IRQ stub funcs used */

/* ----------------------------- start of code ----------------------------- */

/******************************************************************************
 * Interrupt handlers. Those are stubs which call the real interrupt handler
 * with the IRQ level as parameter. This mapping is required because OS/2
 * calls interrupt handlers without any parameters, yet expects them to know
 * which IRQ level to complete when calling DevHelp_EOI().
 *
 * This array of functions needs to be extended when increasing MAX_AD.
 */
#if MAX_AD > 8
#error must extend irq_handler_xx and irq_handlers[] when increasing MAX_AD
#endif

/* Macro to call AHCI interrupt handler and set/clear carry flag accordingly.
 * We need to set the carry flag if the interrupt was not handled. This is
 * done by shifting the return value of ahci_intr() to the right, implying
 * bit 0 will be set when the interrupt was not handled.
 */
#define call_ahci_intr(i)   return(ahci_intr(irq_map[i]) >> 1)

static USHORT _cdecl _far irq_handler_00(void)  { call_ahci_intr(0); }
static USHORT _cdecl _far irq_handler_01(void)  { call_ahci_intr(1); }
static USHORT _cdecl _far irq_handler_02(void)  { call_ahci_intr(2); }
static USHORT _cdecl _far irq_handler_03(void)  { call_ahci_intr(3); }
static USHORT _cdecl _far irq_handler_04(void)  { call_ahci_intr(4); }
static USHORT _cdecl _far irq_handler_05(void)  { call_ahci_intr(5); }
static USHORT _cdecl _far irq_handler_06(void)  { call_ahci_intr(6); }
static USHORT _cdecl _far irq_handler_07(void)  { call_ahci_intr(7); }

PFN irq_handlers[] = {
  (PFN) irq_handler_00, (PFN) irq_handler_01, (PFN) irq_handler_02,
  (PFN) irq_handler_03, (PFN) irq_handler_04, (PFN) irq_handler_05,
  (PFN) irq_handler_06, (PFN) irq_handler_07
};

/******************************************************************************
 * Save BIOS configuration of AHCI adapter. As a side effect, this also saves
 * generic configuration information which we may have to restore after an
 * adapter reset.
 *
 * NOTE: This function also saves working copies of the CAP and CAP2 registers
 *       as well as the initial port map in the AD_INFO structure after
 *       removing features which are known to cause trouble on this specific
 *       piece of hardware.
 */
int ahci_save_bios_config(AD_INFO *ai)
{
  int ports;
  int i;

  /* save BIOS configuration */
  for (i = 0; i < HOST_CAP2; i += sizeof(u32)) {
    ai->bios_config[i / sizeof(u32)] = readl(ai->mmio + i);
  }

  /* HOST_CAP2 only exists for AHCI V1.2 and later */
  if (ai->bios_config[HOST_VERSION / sizeof(u32)] >= 0x00010200L) {
    ai->bios_config[HOST_CAP2 / sizeof(u32)] = readl(ai->mmio + HOST_CAP2);
  } else {
    ai->bios_config[HOST_CAP2 / sizeof(u32)] = 0;
  }

  /* print AHCI register debug information */
  if (debug) {
    printf("AHCI global controller registers:\n");
    for (i = 0; i <= HOST_CAP2 / sizeof(u32); i++) {
      u32 val = ai->bios_config[i];
      printf(" %02x: %08lx", i, val);

      if (i == HOST_CAP) {
        printf(" -");
        if (val & HOST_CAP_64)         printf(" 64bit");
        if (val & HOST_CAP_NCQ)        printf(" ncq");
        if (val & HOST_CAP_SNTF)       printf(" sntf");
        if (val & HOST_CAP_MPS)        printf(" mps");
        if (val & HOST_CAP_SSS)        printf(" sss");
        if (val & HOST_CAP_ALPM)       printf(" alpm");
        if (val & HOST_CAP_LED)        printf(" led");
        if (val & HOST_CAP_CLO)        printf(" clo");
        if (val & HOST_CAP_ONLY)       printf(" ahci_only");
        if (val & HOST_CAP_PMP)        printf(" pmp");
        if (val & HOST_CAP_FBS)        printf(" fbs");
        if (val & HOST_CAP_PIO_MULTI)  printf(" pio_multi");
        if (val & HOST_CAP_SSC)        printf(" ssc");
        if (val & HOST_CAP_PART)       printf(" part");
        if (val & HOST_CAP_CCC)        printf(" ccc");
        if (val & HOST_CAP_EMS)        printf(" ems");
        if (val & HOST_CAP_SXS)        printf(" sxs");
        printf(" cmd_slots:%d", (u16) ((val >> 8) & 0x1f) + 1);
        printf(" ports:%d",     (u16) (val & 0x1f) + 1);

      } else if (i == HOST_CTL) {
        printf(" -");
        if (val & HOST_AHCI_EN)        printf(" ahci_enabled");
        if (val & HOST_IRQ_EN)         printf(" irq_enabled");
        if (val & HOST_RESET)          printf(" resetting");

      } else if (i == HOST_CAP2) {
        printf(" -");
        if (val & HOST_CAP2_BOH)       printf(" boh");
        if (val & HOST_CAP2_NVMHCI)    printf(" nvmhci");
        if (val & HOST_CAP2_APST)      printf(" apst");

      }
      printf("\n");
    }
  }

  /* Save working copies of CAP, CAP2 and port_map and remove broken feature
   * bits. This is largely copied from the Linux AHCI driver -- the wisdom
   * around quirks and faulty hardware is hard to come by...
   */
  ai->cap      = ai->bios_config[HOST_CAP        / sizeof(u32)];
  ai->cap2     = ai->bios_config[HOST_CAP2       / sizeof(u32)];
  ai->port_map = ai->bios_config[HOST_PORTS_IMPL / sizeof(u32)];

  if (ai->pci->board >= sizeof(initial_flags) / sizeof(*initial_flags)) {
    dprintf("error: invalid board index in PCI info\n");
    return(-1);
  }
  ai->flags = initial_flags[ai->pci->board];

  if ((ai->cap & HOST_CAP_64) && (ai->flags & AHCI_HFLAG_32BIT_ONLY)) {
    /* disable 64-bit support for faulty controllers; OS/2 can't do 64 bits at
     * this point, of course, but who knows where all this will be in a few
     * years...
     */
    ai->cap &= ~HOST_CAP_64;
  }

  if ((ai->cap & HOST_CAP_NCQ) && (ai->flags & AHCI_HFLAG_NO_NCQ)) {
    dprintf("controller can't do NCQ, turning off CAP_NCQ\n");
    ai->cap &= ~HOST_CAP_NCQ;
  }

  if (!(ai->cap & HOST_CAP_NCQ) && (ai->flags & AHCI_HFLAG_YES_NCQ)) {
    dprintf("controller can do NCQ, turning on CAP_NCQ\n");
    ai->cap |= HOST_CAP_NCQ;
  }

  if ((ai->cap & HOST_CAP_PMP) && (ai->flags & AHCI_HFLAG_NO_PMP)) {
    dprintf("controller can't do PMP, turning off CAP_PMP\n");
    ai->cap |= HOST_CAP_PMP;
  }

  if ((ai->cap & HOST_CAP_SNTF) && (ai->flags & AHCI_HFLAG_NO_SNTF)) {
    dprintf("controller can't do SNTF, turning off CAP_SNTF\n");
    ai->cap &= ~HOST_CAP_SNTF;
  }

  if (ai->pci->vendor == PCI_VENDOR_ID_JMICRON &&
      ai->pci->device == 0x2361 && ai->port_map != 1) {
    dprintf("JMB361 has only one port, port_map 0x%x -> 0x%x\n", ai->port_map, 1);
    ai->port_map = 1;
  }

  /* Correlate port map to number of ports reported in HOST_CAP
   *
   * NOTE: Port map and number of ports handling differs a bit from the
   *       Linux AHCI driver because we're storing both in AI_INFO. As in the
   *       Linux driver, the port map is the main driver for port scanning but
   *       we're also saving a maximum port number in AI_INFO to reduce the
   *       number of IORB queues to look at in trigger_engine(). This is done
   *       in ahci_scan_ports().
   */
  ports = (ai->cap & 0x1f) + 1;
  for (i = 0; i < AHCI_MAX_PORTS; i++) {
    if (ai->port_map & (1UL << i)) {
      ports--;
    }
  }
  if (ports < 0) {
    /* more ports in port_map than in HOST_CAP & 0x1f */
    ports = (ai->cap & 0x1f) + 1;
    dprintf("implemented port map (0x%lx) contains more "
            "ports than nr_ports (%d), using nr_ports\n",
            ai->port_map, ports);
    ai->port_map = (1UL << ports) - 1UL;
  }

  /* set maximum command slot number */
  ai->cmd_max = (u16) ((ai->cap >> 8) & 0x1f);

  return(0);
}

/******************************************************************************
 * Restore BIOS configuration of AHCI adapter. This is needed after scanning
 * for devices because we still need the BIOS until the initial boot sequence
 * has completed.
 */
int ahci_restore_bios_config(AD_INFO *ai)
{
  ddprintf("restoring AHCI BIOS configuration\n");

  /* restore saved BIOS configuration */
  writel(ai->mmio + HOST_CTL,       ai->bios_config[HOST_CTL / sizeof(u32)]);
  writel(ai->mmio + HOST_CCC,       ai->bios_config[HOST_CCC / sizeof(u32)]);
  writel(ai->mmio + HOST_CCC_PORTS, ai->bios_config[HOST_CCC_PORTS / sizeof(u32)]);
  writel(ai->mmio + HOST_EM_CTL,    ai->bios_config[HOST_EM_CTL / sizeof(u32)]);

  /* flush PCI MMIO delayed write buffers */
  readl(ai->mmio + HOST_EM_CTL);

  return(0);
}

/******************************************************************************
 * Restore initial configuration (e.g. after an adapter reset). This relies
 * on information saved by 'ahci_save_bios_config()'.
 */
int ahci_restore_initial_config(AD_INFO *ai)
{
  ddprintf("restoring initial configuration\n");

  /* restore saved BIOS configuration */
  writel(ai->mmio + HOST_CTL,       ai->bios_config[HOST_CTL / sizeof(u32)]);
  writel(ai->mmio + HOST_CCC,       ai->bios_config[HOST_CCC / sizeof(u32)]);
  writel(ai->mmio + HOST_CCC_PORTS, ai->bios_config[HOST_CCC_PORTS / sizeof(u32)]);
  writel(ai->mmio + HOST_EM_CTL,    ai->bios_config[HOST_EM_CTL / sizeof(u32)]);

  /* flush PCI MMIO delayed write buffers */
  readl(ai->mmio + HOST_EM_CTL);

  /* (re-)enable AHCI mode */
  ahci_enable_ahci(ai);

  return(0);
}

/******************************************************************************
 * Save port configuration. This is primarily used to save the BIOS port
 * configuration (command list and FIS buffers and the IRQ mask).
 *
 * The port configuration returned by this function is dynamically allocated
 * and automatically freed when calling ahci_restore_port_config().
 */
AHCI_PORT_CFG *ahci_save_port_config(AD_INFO *ai, int p)
{
  AHCI_PORT_CFG *pc;
  u8 _far *port_mmio = port_base(ai, p);

  if ((pc = malloc(sizeof(*pc))) == NULL) {
    return(NULL);
  }

  pc->cmd_list   = readl(port_mmio + PORT_LST_ADDR);
  pc->cmd_list_h = readl(port_mmio + PORT_LST_ADDR_HI);
  pc->fis_rx     = readl(port_mmio + PORT_FIS_ADDR);
  pc->fis_rx_h   = readl(port_mmio + PORT_FIS_ADDR_HI);
  pc->irq_mask   = readl(port_mmio + PORT_IRQ_MASK);

  return(pc);
}

/******************************************************************************
 * Restore port configuration. This is primarily used to restore the BIOS port
 * configuration (command list and FIS buffers and the IRQ mask).
 *
 * The port configuration automatically freed.
 */
void ahci_restore_port_config(AD_INFO *ai, int p, AHCI_PORT_CFG *pc)
{
  u8 _far *port_mmio = port_base(ai, p);

  writel(port_mmio + PORT_LST_ADDR,    pc->cmd_list);
  writel(port_mmio + PORT_LST_ADDR_HI, pc->cmd_list_h);
  writel(port_mmio + PORT_FIS_ADDR,    pc->fis_rx);
  writel(port_mmio + PORT_FIS_ADDR_HI, pc->fis_rx_h);
  writel(port_mmio + PORT_IRQ_MASK,    pc->irq_mask);

  readl(port_base(ai, p) + PORT_IRQ_MASK); /* flush */

  free(pc);
}

/******************************************************************************
 * Enable AHCI mode on this controller.
 */
int ahci_enable_ahci(AD_INFO *ai)
{
  u32 ctl = readl(ai->mmio + HOST_CTL);
  int i;

  if (ctl & HOST_AHCI_EN) {
    /* AHCI mode already enbled */
    return(0);
  }

  /* some controllers need AHCI_EN to be written multiple times */
  for (i = 0; i < 5; i++) {
    ctl |= HOST_AHCI_EN;
    writel(ai->mmio + HOST_CTL, ctl);
    ctl = readl(ai->mmio + HOST_CTL);   /* flush && sanity check */
    if (ctl & HOST_AHCI_EN) {
      return(0);
    }
    mdelay(10);
  }

  /* couldn't enable AHCI mode */
  dprintf("failed to enable AHCI mode on adapter #%d\n", ad_no(ai));
  return(1);
}

/******************************************************************************
 * Scan all ports for connected devices and fill in the corresponding device
 * information.
 *
 * NOTES:
 *
 *  - The adapter is temporarily configured for os2ahci but the original BIOS
 *    configuration will be restored when done. This happens only until we
 *    have received the IOCC_COMPLETE_INIT command.
 *
 *  - Subsequent calls are currently not planned but may be required for
 *    suspend/resume handling, hot swap functionality, etc.
 *
 *  - This function is expected to be called with the spinlock released but
 *    the corresponding adapter's busy flag set. It will aquire the spinlock
 *    temporarily to allocate/free memory for the ATA identify buffer.
 */
int ahci_scan_ports(AD_INFO *ai)
{
  AHCI_PORT_CFG *pc = NULL;
  u16 *id_buf;
  int rc;
  int p;

  spin_lock(drv_lock);
  id_buf = malloc(ATA_ID_WORDS * sizeof(u16));
  spin_unlock(drv_lock);
  if (id_buf == NULL) {
    return(-1);
  }

  if (ai->bios_config[0] == 0) {
    /* first call */
    ahci_save_bios_config(ai);
  }

  if (ahci_enable_ahci(ai)) {
    goto exit_port_scan;
  }

  /* perform port scan */
  dprintf("scanning ports on adapter #%d\n", ad_no(ai));
  for (p = 0; p < AHCI_MAX_PORTS; p++) {
    if (ai->port_map & (1UL << p)) {

      if (!init_complete) {
        if ((pc = ahci_save_port_config(ai, p)) == NULL) {
          goto exit_port_scan;
        }
      }

      /* start/reset port; if no device is attached, this is expected to fail */
      if (init_reset) {
        ddprintf("resetting port #%d\n", p);
        rc = ahci_reset_port(ai, p, 0);
      } else {
        ddprintf("(re)starting port #%d\n", p);
        ahci_stop_port(ai, p);
        rc = ahci_start_port(ai, p, 0);
      }
      if (rc) {
        /* no device attached to this port */
        ai->port_map &= ~(1UL << p);
        goto restore_port_config;
      }

      /* this port has a device attached and is ready to accept commands */
      ddprintf("port #%d seems to be attached to a device; probing...\n", p);
      rc = ahci_exec_polled_cmd(ai, p, 0, 500, ATA_CMD_ID_ATA,
                                AP_VADDR, (void _far *) id_buf, 512,
                                AP_END);

      if (rc != 0 || id_buf[ATA_ID_CONFIG] & (1U << 15)) {
        /* this might be an ATAPI device; run IDENTIFY_PACKET_DEVICE */
        rc = ahci_exec_polled_cmd(ai, p, 0, 500, ATA_CMD_ID_ATAPI,
                                  AP_VADDR, (void _far *) id_buf, 512,
                                  AP_END);
      }

      if (rc == 0) {
        /* we have a valid IDENTIFY or IDENTIFY_PACKET response */
        ddphex(id_buf, 512, "ATA_IDENTIFY(_PACKET) results:\n");
        ahci_setup_device(ai, p, 0, id_buf);
      } else {
        /* no device attached to this port */
        ai->port_map &= ~(1UL << p);
      }

    restore_port_config:
      if (pc != NULL) {
        ahci_restore_port_config(ai, p, pc);
      }
    }
  }

exit_port_scan:
  if (!init_complete) {
    ahci_restore_bios_config(ai);
  }
  spin_lock(drv_lock);
  free(id_buf);
  spin_unlock(drv_lock);
  return(0);
}

/******************************************************************************
 * Complete initialization of adapter. This includes restarting all active
 * ports and initializing interrupt processing. This is called when receiving
 * the IOCM_COMPLETE_INIT request.
 */
int ahci_complete_init(AD_INFO *ai)
{
  int rc;
  int p;
  int i;

  dprintf("completing initialization of adapter #%d\n", ad_no(ai));

  /* register IRQ handlers; each IRQ level is registered only once */
  for (i = 0; i < irq_map_cnt; i++) {
    if (irq_map[i] == ai->irq) {
      /* we already have this IRQ registered */
      break;
    }
  }

  if (i >= irq_map_cnt) {
    dprintf("registering interrupt #%d\n", ai->irq);

    if (DevHelp_SetIRQ(mk_NPFN(irq_handlers[irq_map_cnt]), ai->irq, 1) != 0) {
      dprintf("failed to register shared interrupt\n");

      if (DevHelp_SetIRQ(mk_NPFN(irq_handlers[irq_map_cnt]), ai->irq, 0) != 0) {
        dprintf("failed to register exclusive interrupt\n");
        return(-1);
      }
    }
    irq_map[irq_map_cnt++] = ai->irq;
  }

  /* enable AHCI mode */
  if ((rc = ahci_enable_ahci(ai)) != 0) {
    return(rc);
  }

  /* Start all ports. The main purpose is to set the command list and FIS
   * receive area addresses properly and to enable port-level interrupts; we
   * don't really care about the return status because we'll find out soon
   * enough if a previously detected device has problems.
   */
  for (p = 0; p < AHCI_MAX_PORTS; p++) {
    if (ai->port_map & (1UL << p)) {
      dprintf("restarting port #%d\n", p);
      ahci_stop_port(ai, p);
      ahci_start_port(ai, p, 1);
    }
  }

  /* enable adapter-level interrupts */
  writel(ai->mmio + HOST_CTL, HOST_IRQ_EN);
  readl(ai->mmio + HOST_CTL); /* flush */

  /* enable interrupts on PCI-level (PCI 2.3 added a feature to disable ints) */
  pci_enable_int(ai->bus, ai->dev_func);

  return(0);
}

/******************************************************************************
 * Reset specified port. This function is typically called during adapter
 * initialization and first gets the port into a defined status, then resets
 * the port by sending a COMRESET signal.
 *
 * This function is also the location of the link speed initialization (link
 * needs to be restablished after changing link speed, anyway).
 *
 * NOTE: This function uses a busy loop to wait for DMA engines to stop and
 *       the COMRESET to complete. It should only be called at task time
 *       during initialization or in a context hook.
 */
int ahci_reset_port(AD_INFO *ai, int p, int ei)
{
  u8 _far *port_mmio = port_base(ai, p);
  u32 tmp;
  int timeout = 500;

  dprintf("resetting port %d.%d\n", ad_no(ai), p);

  /* stop port engines (we don't care whether there is an error doing so) */
  ahci_stop_port(ai, p);

  /* clear SError */
  tmp = readl(port_mmio + PORT_SCR_ERR);
  ddprintf(" PORT_SCR_ERR  = 0x%lx\n", tmp);
  writel(port_mmio + PORT_SCR_ERR, tmp);

  /* clear pending port IRQs */
  tmp = readl(port_mmio + PORT_IRQ_STAT);
  if (tmp) {
    writel(port_mmio + PORT_IRQ_STAT, tmp);
  }
  ddprintf(" PORT_IRQ_STAT = 0x%lx\n", tmp);
  ddprintf(" PORT_IRQ_MASK = 0x%lx\n", readl(port_mmio + PORT_IRQ_MASK));
  ddprintf(" HOST_IRQ_STAT = 0x%lx\n", readl(ai->mmio + HOST_IRQ_STAT));
  writel(ai->mmio + HOST_IRQ_STAT, 1UL << p);

  /* set link speed */
  tmp = readl(port_mmio + PORT_SCR_CTL) & ~0x000000f0UL;
  writel(port_mmio + PORT_SCR_CTL, tmp | (link_speed[ad_no(ai)][p] << 4));

  /* issue COMRESET on the port */
  tmp = readl(port_mmio + PORT_SCR_CTL) & ~0x0000000fUL;
  writel(port_mmio + PORT_SCR_CTL, tmp | 1);
  readl(port_mmio + PORT_SCR_CTL);  /* flush */

  /* spec says "leave reset bit on for at least 1ms"; make it 2ms */
  mdelay(2);

  writel(port_mmio + PORT_SCR_CTL, tmp);
  readl(port_mmio + PORT_SCR_CTL);  /* flush */

  /* wait for communication to be re-established after port reset */
  while (((tmp = readl(port_mmio + PORT_SCR_STAT) & 3)) != 3) {
    mdelay(10);
    timeout -= 10;
    if (timeout <= 0) {
      dprintf("no device present after resetting port #%d "
              "(PORT_SCR_STAT = 0x%lx)\n", p, tmp);
      return(-1);
    }
  }

  /* clear SError again (recommended by AHCI spec) */
  tmp = readl(port_mmio + PORT_SCR_ERR);
  writel(port_mmio + PORT_SCR_ERR, tmp);

  /* start port so we can receive the COMRESET FIS */
  ahci_start_port(ai, p, ei);

  /* wait for device to be ready ((PxTFD & (BSY | DRQ | ERR)) == 0) */
  while (((tmp = readl(port_mmio + PORT_TFDATA)) & 0x89) != 0) {
    mdelay(10);
    timeout -= 10;
    if (timeout <= 0) {
      dprintf("device not ready on port #%d "
              "(PORT_TFDATA = 0x%lx)\n", p, tmp);
      ahci_stop_port(ai, p);
      return(-1);
    }
  }
  ddprintf(" PORT_TFDATA   = 0x%lx\n", readl(port_mmio + PORT_TFDATA));

  return(0);
}

/******************************************************************************
 * Start specified port.
 */
int ahci_start_port(AD_INFO *ai, int p, int ei)
{
  u8 _far *port_mmio = port_base(ai, p);
  u32 status;

  /* check whether device presence is detected and link established */
  status = readl(port_mmio + PORT_SCR_STAT);
  ddprintf(" PORT_SCR_STAT = 0x%lx\n", status);
  if ((status & 0xf) != 3) {
    return(-1);
  }

  /* clear SError, if any */
  status = readl(port_mmio + PORT_SCR_ERR);
  ddprintf(" PORT_SCR_ERR  = 0x%lx\n", status);
  writel(port_mmio + PORT_SCR_ERR, status);

  /* enable FIS reception */
  ahci_start_fis_rx(ai, p);

  /* enable DMA */
  ahci_start_engine(ai, p);

  if (ei) {
    /* clear any pending interrupts on this port */
    if ((status = readl(port_mmio + PORT_IRQ_STAT)) != 0) {
      writel(port_mmio + PORT_IRQ_STAT, status);
    }

    /* enable port interrupts */
    writel(port_mmio + PORT_IRQ_MASK, PORT_IRQ_TF_ERR |
                                      PORT_IRQ_HBUS_ERR |
                                      PORT_IRQ_HBUS_DATA_ERR |
                                      PORT_IRQ_IF_ERR |
                                      PORT_IRQ_OVERFLOW |
                                      PORT_IRQ_BAD_PMP |
                                      PORT_IRQ_UNK_FIS |
                                      PORT_IRQ_SDB_FIS |
                                      PORT_IRQ_D2H_REG_FIS);
  } else {
    writel(port_mmio + PORT_IRQ_MASK, 0);
  }
  readl(port_mmio + PORT_IRQ_MASK);  /* flush */

  return(0);
}

/******************************************************************************
 * Start port FIS reception. Copied from Linux AHCI driver and adopted to
 * OS2AHCI.
 */
void ahci_start_fis_rx(AD_INFO *ai, int p)
{
  u8 _far *port_mmio = port_base(ai, p);
  u32 port_dma = port_dma_base_phys(ai, p);
  u32 tmp;

  /* set comand header and FIS address registers */
  writel(port_mmio + PORT_LST_ADDR, port_dma + offsetof(AHCI_PORT_DMA, cmd_hdr));
  writel(port_mmio + PORT_LST_ADDR_HI, 0);
  writel(port_mmio + PORT_FIS_ADDR, port_dma + offsetof(AHCI_PORT_DMA, rx_fis));
  writel(port_mmio + PORT_FIS_ADDR_HI, 0);

  /* enable FIS reception */
  tmp = readl(port_mmio + PORT_CMD);
  tmp |= PORT_CMD_FIS_RX;
  writel(port_mmio + PORT_CMD, tmp);

  /* flush */
  readl(port_mmio + PORT_CMD);
}

/******************************************************************************
 * Start port HW engine. Copied from Linux AHCI driver and adopted to OS2AHCI.
 */
void ahci_start_engine(AD_INFO *ai, int p)
{
  u8 _far *port_mmio = port_base(ai, p);
  u32 tmp;

  /* start DMA */
  tmp = readl(port_mmio + PORT_CMD);
  tmp |= PORT_CMD_START;
  writel(port_mmio + PORT_CMD, tmp);
  readl(port_mmio + PORT_CMD); /* flush */
}

/******************************************************************************
 * Stop specified port
 */
int ahci_stop_port(AD_INFO *ai, int p)
{
  u8 _far *port_mmio = port_base(ai, p);
  int rc;

  /* disable FIS reception */
  if ((rc = ahci_stop_fis_rx(ai, p)) != 0) {
    dprintf("error: failed to stop FIS receive (%d)\n", rc);
    return(rc);
  }

  /* disable DMA */
  if ((rc = ahci_stop_engine(ai, p)) != 0) {
    dprintf("error: failed to stop port HW engine (%d)\n", rc);
    return(rc);
  }

  /* reset PxSACT register (tagged command queues, not reset by COMRESET) */
  writel(port_mmio + PORT_SCR_ACT, 0);
  readl(port_mmio + PORT_SCR_ACT);  /* flush */

  return(0);
}

/******************************************************************************
 * Stop port FIS reception. Copied from Linux AHCI driver and adopted to
 * OS2AHCI.
 *
 * NOTE: This function uses a busy loop to wait for the DMA engine to stop. It
 *       should only be called at task time during initialization or in a
 *       context hook (e.g. when resetting a port).
 */
int ahci_stop_fis_rx(AD_INFO *ai, int p)
{
  u8 _far *port_mmio = port_base(ai, p);
  int timeout = 1000;
  u32 tmp;

  /* disable FIS reception */
  tmp = readl(port_mmio + PORT_CMD);
  tmp &= ~PORT_CMD_FIS_RX;
  writel(port_mmio + PORT_CMD, tmp);

  /* wait for completion, spec says 500ms, give it 1000 */
  while (timeout > 0 && (readl(port_mmio + PORT_CMD) & PORT_CMD_FIS_ON)) {
    mdelay(10);
    timeout -= 10;
  }

  return((timeout <= 0) ? -1 : 0);
}

/******************************************************************************
 * Stop port HW engine. Copied from Linux AHCI driver and adopted to OS2AHCI.
 *
 * NOTE: This function uses a busy loop to wait for the DMA engine to stop. It
 *       should only be called at task time during initialization or in a
 *       context hook (e.g. when resetting a port).
 */
int ahci_stop_engine(AD_INFO *ai, int p)
{
  u8 _far *port_mmio = port_base(ai, p);
  int timeout = 500;
  u32 tmp;

  tmp = readl(port_mmio + PORT_CMD);

  /* check if the port is already stopped */
  if ((tmp & (PORT_CMD_START | PORT_CMD_LIST_ON)) == 0) {
    return 0;
  }

  /* set port to idle */
  tmp &= ~PORT_CMD_START;
  writel(port_mmio + PORT_CMD, tmp);

  /* wait for engine to stop. This could be as long as 500 msec */
  while (timeout > 0 && (readl(port_mmio + PORT_CMD) & PORT_CMD_LIST_ON)) {
    mdelay(10);
    timeout -= 10;
  }

  return((timeout <= 0) ? -1 : 0);
}

/******************************************************************************
 * Execute AHCI command for given IORB. This includes all steps typically
 * required by any of the ahci_*() IORB processing functions.
 *
 * NOTE: In order to prevent race conditions with port restart and reset
 *       handlers, we either need to keep the spinlock during the whole
 *       operation or set the adapter's busy flag. Since the expectation
 *       is that command preparation will be quick (it certainly doesn't
 *       involve delays), we're going with the spinlock for the time being.
 */
void ahci_exec_iorb(IORBH _far *iorb, int ncq_capable,
                    int (*func)(IORBH _far *, int))
{
  volatile u32 *cmds;
  ADD_WORKSPACE _far *aws = add_workspace(iorb);
  AD_INFO *ai = ad_infos + iorb_unit_adapter(iorb);
  P_INFO *port = ai->ports + iorb_unit_port(iorb);
  ULONG timeout = (iorb->Timeout > 0) ? iorb->Timeout : DEFAULT_TIMEOUT;
  u8 _far *port_mmio = port_base(ai, iorb_unit_port(iorb));
  u16 cmd_max = ai->cmd_max;
  int i;

  /* determine whether this will be an NCQ request */
  aws->is_ncq = 0;
  if (ncq_capable && port->devs[iorb_unit_device(iorb)].ncq_max > 1 &&
      (ai->cap & HOST_CAP_NCQ) && !aws->no_ncq && init_complete) {

    /* We can make this an NCQ request; limit command slots to the maximum
     * NCQ tag number reported by the device - 1. Why "minus one"? I seem to
     * recall an issue related to using all 32 tag numbers but can't quite
     * pinpoint it right now. One less won't make much of a difference...
     */
    aws->is_ncq = 1;
    if ((cmd_max = port->devs[iorb_unit_device(iorb)].ncq_max - 1) > ai->cmd_max) {
      cmd_max = ai->cmd_max;
    }
    ddprintf("NCQ command; cmd_max = %d->%d\n", (u16) ai->cmd_max, cmd_max);
  }

  /* make sure adapter is available */
  spin_lock(drv_lock);
  if (!ai->busy) {

    if (!init_complete) {
      ai->busy = 1;
      spin_unlock(drv_lock);
      ahci_exec_polled_iorb(iorb, func, timeout);
      ai->busy = 0;
      return;
    }

    /* make sure we don't mix NCQ and regular commands */
    if (aws->is_ncq && port->reg_cmds == 0 || !aws->is_ncq && port->ncq_cmds == 0) {

      /* Find next available command slot. We use a simple round-robin
       * algorithm for this to prevent commands with higher slot indexes
       * from stalling when new commands are coming in frequently.
       */
      cmds = (aws->is_ncq) ? &port->ncq_cmds : &port->reg_cmds;
      for (i = 0; i <= cmd_max; i++) {
        if (++(port->cmd_slot) > cmd_max) {
          port->cmd_slot = 0;
        }
        if ((*cmds & (1UL << port->cmd_slot)) == 0) {
          break;
        }
      }

      if ((*cmds & (1UL << port->cmd_slot)) == 0) {
        /* prepare command */
        if (func(iorb, port->cmd_slot)) {
          /* Command preparation failed, or no HW command required; IORB
           * will already have the error code if there was an error.
           */
          spin_unlock(drv_lock);
          iorb_done(iorb);
          return;
        }

        /* start timer for this IORB */
        ADD_StartTimerMS(&aws->timer, timeout, (PFN) timeout_callback, iorb, 0);

        /* update IORB */
        aws->queued_hw = 1;
        aws->cmd_slot = port->cmd_slot;

        /* issue command to hardware */
        ddprintf("issuing command on slot %d\n", port->cmd_slot);
        *cmds |= (1UL << port->cmd_slot);
        if (aws->is_ncq) {
          writel(port_mmio + PORT_SCR_ACT, (1UL << port->cmd_slot));
          readl(port_mmio + PORT_SCR_ACT); /* flush */
        }
        writel(port_mmio + PORT_CMD_ISSUE, (1UL << port->cmd_slot));
        readl(port_mmio + PORT_CMD_ISSUE); /* flush */

        spin_unlock(drv_lock);
        return;
      }
    }
  }

  /* requeue this IORB; it will be picked up again in trigger_engine() */
  aws->processing = 0;
  spin_unlock(drv_lock);
}

/******************************************************************************
 * Execute polled IORB command. This function is called by ahci_exec_iorb()
 * when the initialization has not yet completed. The reasons for polling until
 * initialization has completed are:
 *
 *  - We need to restore the BIOS configuration after we're done with this
 *    command because someone might still call int 13h routines; sending
 *    asynchronous commands and waiting for interrupts to indicate completion
 *    won't work in such a scenario.
 *  - Our context hooks won't work while the device managers are initializing
 *    (they can't yield at init time).
 *  - The device managers typically poll for command completion during
 *    initialization so it won't make much of a difference, anyway.
 *
 * NOTE: This function must be called with the adapter-level busy flag set but
 *       without the driver-level spinlock held.
 */
void ahci_exec_polled_iorb(IORBH _far *iorb, int (*func)(IORBH _far *, int),
                           ULONG timeout)
{
  AHCI_PORT_CFG *pc = NULL;
  AD_INFO *ai = ad_infos + iorb_unit_adapter(iorb);
  int p = iorb_unit_port(iorb);
  u8 _far *port_mmio = port_base(ai, p);

  /* enable AHCI mode */
  if (ahci_enable_ahci(ai) != 0) {
    iorb_seterr(iorb, IOERR_ADAPTER_NONSPECIFIC);
    goto restore_bios_config;
  }

  /* check whether command slot 0 is available */
  if ((readl(port_mmio + PORT_CMD_ISSUE) & 1) != 0) {
    iorb_seterr(iorb, IOERR_DEVICE_BUSY);
    goto restore_bios_config;
  }

  /* save port configuration */
  if ((pc = ahci_save_port_config(ai, p)) == NULL) {
    iorb_seterr(iorb, IOERR_CMD_SW_RESOURCE);
    goto restore_bios_config;
  }

  /* restart port (includes the necessary port configuration) */
  if (ahci_stop_port(ai, p) || ahci_start_port(ai, p, 0)) {
    iorb_seterr(iorb, IOERR_ADAPTER_NONSPECIFIC);
    goto restore_bios_config;
  }

  /* prepare command */
  if (func(iorb, 0) == 0) {
    /* successfully prepared cmd; issue cmd and wait for completion */
    ddprintf("executing polled cmd...");
    writel(port_mmio + PORT_CMD_ISSUE, 1);
    timeout /= 10;
    while (timeout > 0 && (readl(port_mmio + PORT_CMD_ISSUE) & 1)) {
      mdelay(10);
      timeout--;
    }
    ddprintf(" done (time left = %ld)\n", timeout * 10);

    if (timeout == 0) {
      dprintf("timeout for IORB %Fp\n", iorb);
      iorb_seterr(iorb, IOERR_ADAPTER_TIMEOUT);

    } else if (readl(port_mmio + PORT_SCR_ERR) != 0 ||
               readl(port_mmio + PORT_TFDATA) & 0x89) {
      dprintf("polled cmd error for IORB %Fp\n", iorb);
      iorb_seterr(iorb, IOERR_DEVICE_NONSPECIFIC);
      ahci_reset_port(ai, iorb_unit_port(iorb), 0);

    } else {
      /* successfully executed command */
      if (add_workspace(iorb)->ppfunc != NULL) {
        add_workspace(iorb)->ppfunc(iorb);
      } else {
        add_workspace(iorb)->complete = 1;
      }
    }
  }

restore_bios_config:
  /* restore BIOS configuration */
  if (pc != NULL) {
    ahci_restore_port_config(ai, p, pc);
  }
  ahci_restore_bios_config(ai);

  if (add_workspace(iorb)->complete | (iorb->Status | IORB_ERROR)) {
    aws_free(add_workspace(iorb));
    iorb_done(iorb);
  }
  return;
}

/******************************************************************************
 * Execute polled ATA/ATAPI command. This function will block until the command
 * has completed or the timeout has expired, thus it should only be used during
 * initialization. Furthermore, it will always use command slot zero.
 *
 * The difference to ahci_exec_polled_iorb() is that this function executes
 * arbitrary ATA/ATAPI commands outside the context of an IORB. It's typically
 * used when scanning for devices during initialization.
 */
int ahci_exec_polled_cmd(AD_INFO *ai, int p, int d, int timeout, int cmd, ...)
{
  va_list va;
  u8 _far *port_mmio = port_base(ai, p);
  u32 tmp;
  int rc;

  /* verify that command slot 0 is idle */
  if (readl(port_mmio + PORT_CMD_ISSUE) & 1) {
    ddprintf("port %d slot 0 is not idle; not executing polled cmd\n", p);
    return(-1);
  }

  /* fill in command slot 0 */
  va_start(va, cmd);
  if ((rc = v_ata_cmd(ai, p, d, 0, cmd, va)) != 0) {
    return(rc);
  }

  /* start command execution for slot 0 */
  ddprintf("executing polled cmd...");
  writel(port_mmio + PORT_CMD_ISSUE, 1);

  /* wait until command has completed */
  while (timeout > 0 && (readl(port_mmio + PORT_CMD_ISSUE) & 1)) {
    mdelay(10);
    timeout -= 10;
  }
  ddprintf(" done (time left = %d)\n", timeout);

  /* check error condition */
  if ((tmp = readl(port_mmio + PORT_SCR_ERR)) != 0) {
    dprintf("SERR = 0x%08lx\n", tmp);
    return(-1);
  }
  if (((tmp = readl(port_mmio + PORT_TFDATA)) & 0x89) != 0) {
    dprintf("TFDATA = 0x%08lx\n", tmp);
    return(-1);
  }

  return((timeout <= 0) ? -1 : 0);
}

/******************************************************************************
 * AHCI top-level hardware interrupt handler. This handler finds the adapters
 * and ports which have issued the interrupt and calls the corresponding
 * port interrupt handler.
 *
 * On entry, OS/2 will have processor interrupts enabled because we're using
 * shared IRQs but we won't be preempted by another interrupt on the same IRQ
 * IRQ level until we indicated EOI. We'll keep it this way, only requesting
 * the driver-level spinlock when actually changing the driver state (IORB
 * queues, ...)
 */
int ahci_intr(u16 irq)
{
  u32 irq_stat;
  int handled = 0;
  int a;
  int p;

  /* find adapter(s) with pending interrupts */
  for (a = 0; a < ad_info_cnt; a++) {
    AD_INFO *ai = ad_infos + a;

    if (ai->irq == irq && (irq_stat = readl(ai->mmio + HOST_IRQ_STAT)) != 0) {
      /* this adapter has interrupts pending */
      u32 irq_masked = irq_stat & ai->port_map;

      for (p = 0; p <= ai->port_max; p++) {
        if (irq_masked & (1UL << p)) {
          ahci_port_intr(ai, p);
        }
      }

      /* clear interrupt condition on the adapter */
      writel(ai->mmio + HOST_IRQ_STAT, irq_stat);
      readl(ai->mmio + HOST_IRQ_STAT); /* flush */
      handled = 1;
    }
  }

  if (handled) {
    /* trigger state machine to process next IORBs, if any */
    spin_lock(drv_lock);
    trigger_engine();
    spin_unlock(drv_lock);

    /* complete the interrupt */
    DevHelp_EOI(irq);
    return(0);
  } else {
    return(1);
  }
}

/******************************************************************************
 * AHCI port-level interrupt handler. As described above, processor interrupts
 * are enabled on entry thus we have to protect shared resources with a
 * spinlock.
 */
void ahci_port_intr(AD_INFO *ai, int p)
{
  IORB_QUEUE done_queue;
  IORBH _far *iorb;
  IORBH _far *next = NULL;
  u8 _far *port_mmio = port_base(ai, p);
  u32 irq_stat;
  u32 active_cmds;
  u32 done_mask;

  ddprintf("port interrupt for adapter #%d, port #%d\n", ad_no(ai), p);
  memset(&done_queue, 0x00, sizeof(done_queue));

  /* get interrupt status and clear it right away */
  irq_stat = readl(port_mmio + PORT_IRQ_STAT);
  writel(port_mmio + PORT_IRQ_STAT, irq_stat);
  readl(port_mmio + PORT_IRQ_STAT); /* flush */

  if (irq_stat & PORT_IRQ_ERROR) {
    /* this is an error interrupt */
    ahci_error_intr(ai, p, irq_stat);
    return;
  }

  spin_lock(drv_lock);

  /* Find out which command slots have completed. Since error recovery for
   * NCQ commands interfers with non-NCQ commands, the upper layers will
   * make sure there's never a mixture of NCQ and non-NCQ commands active
   * on any port at any given time. This makes it easier to find out which
   * commands have completed, too.
   */
  if (ai->ports[p].ncq_cmds != 0) {
    active_cmds = readl(port_mmio + PORT_SCR_ACT);
    done_mask = ai->ports[p].ncq_cmds ^ active_cmds;
    ddprintf("[ncq_cmds]: active_cmds = 0x%08lx, done_mask = 0x%08lx\n",
             active_cmds, done_mask);
  } else {
    active_cmds = readl(port_mmio + PORT_CMD_ISSUE);
    done_mask = ai->ports[p].reg_cmds ^ active_cmds;
    ddprintf("[reg_cmds]: active_cmds = 0x%08lx, done_mask = 0x%08lx\n",
             active_cmds, done_mask);
  }

  /* Find the IORBs related to the completed commands and complete them.
   *
   * NOTES: The spinlock must not be released while in this loop to prevent
   *        race conditions with timeout handlers or other threads in SMP
   *        systems.
   *
   *        Since we hold the spinlock when IORBs complete, we can't call the
   *        IORB notification routine right away because this routine might
   *        schedule another IORB which could cause a deadlock. Thus, we'll
   *        add all IORBs to be completed to a temporary queue which will be
   *        processed after releasing the spinlock.
   */
  for (iorb = ai->ports[p].iorb_queue.root; iorb != NULL; iorb = next) {
    ADD_WORKSPACE _far *aws = (ADD_WORKSPACE _far *) &iorb->ADDWorkSpace;
    next = iorb->pNxtIORB;
    if (aws->queued_hw && (done_mask & (1UL << aws->cmd_slot))) {
      /* this command has completed */
      if (aws->ppfunc != NULL) {
        aws->ppfunc(iorb);
      } else {
        aws->complete = 1;
      }

      if (aws->complete) {
        /* this IORB is complete */
        aws_free(aws);

        /* move IORB to our temporary done queue */
        iorb_queue_del(&ai->ports[p].iorb_queue, iorb);
        iorb_queue_add(&done_queue, iorb);
      }

      /* clear corresponding bit in issued command bitmaps */
      ai->ports[p].ncq_cmds &= ~(1UL << aws->cmd_slot);
      ai->ports[p].reg_cmds &= ~(1UL << aws->cmd_slot);
    }
  }

  spin_unlock(drv_lock);

  /* call notification routines for all IORBs in the done queue */
  for (iorb = done_queue.root; iorb != NULL; iorb = next) {
    next = iorb->pNxtIORB;
    iorb->Status = IORB_DONE;
    if (iorb->RequestControl & IORB_ASYNC_POST) {
      iorb->NotifyAddress(iorb);
    }
  }
}

/******************************************************************************
 * AHCI error interrupt handler. Errors include interface errors and device
 * errors (usually triggered by the error bit in the AHCI task file register).
 *
 * Since this involves long-running operations such as restarting or even
 * resetting a port, this function is invoked at task time via a context
 * hook.
 *
 * NOTE: AHCI controllers stop all processing when encountering an error
 *       condition in order to give the driver time to find out what exactly
 *       went wrong. This means no new commands will be processed until we
 *       clear the error register and restore the "commands issued" register.
 */
void ahci_error_intr(AD_INFO *ai, int p, u32 irq_stat)
{
  int reset_port = 0;

  /* Handle adapter and interface errors. Those typically require a port
   * reset, or worse.
   */
  if (irq_stat & PORT_IRQ_UNK_FIS) {
    u32 _far *unk = (u32 _far *) (port_dma_base(ai, p)->rx_fis + RX_FIS_UNK);
    dprintf("warning: unknown FIS %08lx %08lx %08lx %08lx\n",
            unk[0], unk[1], unk[2], unk[3]);
    reset_port = 1;
  }
  if (irq_stat & (PORT_IRQ_HBUS_ERR | PORT_IRQ_HBUS_DATA_ERR)) {
    dprintf("warning: host bus [data] error for port #%d\n", p);
    reset_port = 1;
  }
  if (irq_stat & PORT_IRQ_IF_ERR && !(ai->flags & AHCI_HFLAG_IGN_IRQ_IF_ERR)) {
    dprintf("warning: interface fatal error for port #%d\n", p);
    reset_port = 1;
  }
  if (reset_port) {
    /* need to reset the port; leave this to the reset context hook */
    ports_to_reset[ad_no(ai)] |= 1UL << p;
    DevHelp_ArmCtxHook(0, reset_ctxhook_h);

    /* no point analyzing device errors after a reset... */
    return;
  }

  /* Handle device-specific errors. Those errors typically involve restarting
   * the corresponding port to resume operations which can take some time,
   * thus we need to offload this functionality to the restart context hook.
   */
  if (irq_stat & PORT_IRQ_TF_ERR) {
    ports_to_restart[ad_no(ai)] |= 1UL << p;
    DevHelp_ArmCtxHook(0, restart_ctxhook_h);
  }
}

/******************************************************************************
 * Get device or media geometry. Device and media geometry are expected to be
 * the same for non-removable devices.
 */
void ahci_get_geometry(IORBH _far *iorb)
{
  dprintf("ahci_get_geometry(%d.%d.%d)\n", (int) iorb_unit_adapter(iorb),
          (int) iorb_unit_port(iorb), (int) iorb_unit_device(iorb));

  ahci_exec_iorb(iorb, 0, cmd_func(iorb, get_geometry));
}

/******************************************************************************
 * Test whether unit is ready.
 */
void ahci_unit_ready(IORBH _far *iorb)
{
  dprintf("ahci_unit_ready(%d.%d.%d)\n", (int) iorb_unit_adapter(iorb),
          (int) iorb_unit_port(iorb), (int) iorb_unit_device(iorb));

  ahci_exec_iorb(iorb, 0, cmd_func(iorb, unit_ready));
}

/******************************************************************************
 * Read sectors from AHCI device.
 */
void ahci_read(IORBH _far *iorb)
{
  dprintf("ahci_read(%d.%d.%d, %ld, %ld)\n", (int) iorb_unit_adapter(iorb),
          (int) iorb_unit_port(iorb), (int) iorb_unit_device(iorb),
          (long) ((IORB_EXECUTEIO _far *) iorb)->RBA,
          (long) ((IORB_EXECUTEIO _far *) iorb)->BlockCount);

  ahci_exec_iorb(iorb, 1, cmd_func(iorb, read));
}

/******************************************************************************
 * Verify readability of sectors on AHCI device.
 */
void ahci_verify(IORBH _far *iorb)
{
  dprintf("ahci_verify(%d.%d.%d, %ld, %ld)\n", (int) iorb_unit_adapter(iorb),
          (int) iorb_unit_port(iorb), (int) iorb_unit_device(iorb),
          (long) ((IORB_EXECUTEIO _far *) iorb)->RBA,
          (long) ((IORB_EXECUTEIO _far *) iorb)->BlockCount);

  ahci_exec_iorb(iorb, 0, cmd_func(iorb, verify));
}

/******************************************************************************
 * Write sectors to AHCI device.
 */
void ahci_write(IORBH _far *iorb)
{
  dprintf("ahci_write(%d.%d.%d, %ld, %ld)\n", (int) iorb_unit_adapter(iorb),
          (int) iorb_unit_port(iorb), (int) iorb_unit_device(iorb),
          (long) ((IORB_EXECUTEIO _far *) iorb)->RBA,
          (long) ((IORB_EXECUTEIO _far *) iorb)->BlockCount);

  ahci_exec_iorb(iorb, 1, cmd_func(iorb, write));
}

/******************************************************************************
 * Execute SCSI (ATAPI) command.
 */
void ahci_execute_cdb(IORBH _far *iorb)
{
  int a = iorb_unit_adapter(iorb);
  int p = iorb_unit_port(iorb);
  int d = iorb_unit_device(iorb);

  dphex(((IORB_ADAPTER_PASSTHRU _far *) iorb)->pControllerCmd,
        ((IORB_ADAPTER_PASSTHRU _far *) iorb)->ControllerCmdLen,
        "ahci_execute_cdb(%d.%d.%d)", a, p, d);

  if (ad_infos[a].ports[p].devs[d].atapi) {
    ahci_exec_iorb(iorb, 0, atapi_execute_cdb);
  } else {
    iorb_seterr(iorb, IOERR_CMD_NOT_SUPPORTED);
    iorb_done(iorb);
  }
}

/******************************************************************************
 * Execute ATA command.
 */
void ahci_execute_ata(IORBH _far *iorb)
{
  int a = iorb_unit_adapter(iorb);
  int p = iorb_unit_port(iorb);
  int d = iorb_unit_device(iorb);

  dphex(((IORB_ADAPTER_PASSTHRU _far *) iorb)->pControllerCmd,
        ((IORB_ADAPTER_PASSTHRU _far *) iorb)->ControllerCmdLen,
        "ahci_execute_cdb(%d.%d.%d)", a, p, d);

  if (ad_infos[a].ports[p].devs[d].atapi) {
    iorb_seterr(iorb, IOERR_CMD_NOT_SUPPORTED);
    iorb_done(iorb);
  } else {
    ahci_exec_iorb(iorb, 0, ata_execute_ata);
  }
}

/******************************************************************************
 * Set up device attached to the specified port based on ATA_IDENTFY_DEVICE or
 * ATA_IDENTFY_PACKET_DEVICE data.
 *
 * NOTE: Port multipliers are not supported, yet, thus the device number is
 *       expected to be 0 for the time being.
 */
static void ahci_setup_device(AD_INFO *ai, int p, int d, u16 *id_buf)
{
  DEVICESTRUCT ds;
  ADJUNCT adj;
  HDEVICE dh;

  if (ai->port_max < p) {
    ai->port_max = p;
  }
  if (ai->ports[p].dev_max < d) {
    ai->ports[p].dev_max = d;
  }
  memset(ai->ports[p].devs + d, 0x00, sizeof(*ai->ports[p].devs));

  /* set generic device information (assuming an ATA disk device for now) */
  ai->ports[p].devs[d].present   = 1;
  ai->ports[p].devs[d].removable = (id_buf[ATA_ID_CONFIG] & 0x0080U) != 0;
  ai->ports[p].devs[d].dev_type  = UIB_TYPE_DISK;

  if (id_buf[ATA_ID_CONFIG] & 0x8000U) {
    /* this is an ATAPI device; augment device information */
    ai->ports[p].devs[d].atapi     = 1;
    ai->ports[p].devs[d].atapi_16  = (id_buf[ATA_ID_CONFIG] & 0x0001U) != 0;
    ai->ports[p].devs[d].dev_type  = (id_buf[ATA_ID_CONFIG] & 0x1f00U) >> 8;

  } else {
    /* complete ATA-specific device information */
    if (!disable_ncq[ad_no(ai)][p]) {
      ai->ports[p].devs[d].ncq_max = id_buf[ATA_ID_QUEUE_DEPTH] & 0x001fU;
    }
    if (id_buf[ATA_ID_CFS_ENABLE_2] & 0x0400U) {
      ai->ports[p].devs[d].lba48   = 1;
    }
  }

  dprintf("found device %d.%d.%d: removable = %d, dev_type = %d, atapi = %d\n",
          ad_no(ai), p, d,
          ai->ports[p].devs[d].removable,
          ai->ports[p].devs[d].dev_type,
          ai->ports[p].devs[d].atapi);

  /* add device to resource manager; we don't really care about errors here */
  memset(&ds, 0x00, sizeof(ds));
  memset(&adj, 0x00, sizeof(adj));

  adj.pNextAdj           = NULL;
  adj.AdjLength          = sizeof(adj);
  adj.AdjType            = ADJ_ADD_UNIT;
  adj.Add_Unit.ADDHandle = rm_drvh;

  ds.DevDescriptName = ata_dev_name(id_buf);
  ds.DevFlags        = (ai->ports[p].devs[d].removable) ? DS_REMOVEABLE_MEDIA
                                                        : DS_FIXED_LOGICALNAME;
  ds.DevType         = ai->ports[p].devs[d].dev_type;

  RMCreateDevice(rm_drvh, &dh, &ds, ai->rm_adh, NULL);

  /* try to detect virtualbox environment to enable a hack for IRQ routing */
  if (ai == ad_infos && p == 7 &&
      ai->pci->vendor == 0x8086 && ai->pci->device == 0x2829 &&
      !memcmp(ds.DevDescriptName, "VBOX HARDDISK", 13)) {
    /* running inside virtualbox */
    pci_hack_virtualbox();
  }
}

/******************************************************************************
 * Timeout handler for I/O commands. Since timeout handling can involve
 * lengthy operations like port resets, the main code is located in a
 * separate function which is invoked via a context hook.
 */
static void _cdecl _far timeout_callback(ULONG timer_handle, ULONG p1,
                                         ULONG p2)
{
  IORBH _far *iorb = (IORBH _far *) p1;
  int a = iorb_unit_adapter(iorb);
  int p = iorb_unit_port(iorb);

  ADD_CancelTimer(timer_handle);
  dprintf("timeout for IORB %Fp\n", iorb);

  /* Move the timed-out IORB to the abort queue. Since it's possible that the
   * IORB has completed after the timeout has expired but before we got to
   * this line of code, we'll check the return code of iorb_queue_del(): If it
   * returns an error, the IORB must have completed a few microseconds ago and
   * there is no timeout.
   */
  spin_lock(drv_lock);
  if (iorb_queue_del(&ad_infos[a].ports[p].iorb_queue, iorb) == 0) {
    iorb_queue_add(&abort_queue, iorb);
    iorb->ErrorCode = IOERR_ADAPTER_TIMEOUT;
  }
  spin_unlock(drv_lock);

  /* Trigger abort processing function. We don't really care whether this
   * succeeds because the only reason why it would fail should be multiple
   * calls to DevHelp_ArmCtxHook() before the context hook had a chance to
   * start executing, which leaves two scenarios:
   *
   *  - We succeded in arming the context hook. Fine.
   *
   *  - We armed the context hook a second time before it had a chance to
   *    start executing. In this case, the already scheduled context hook
   *    will process our IORB as well.
   */
  DevHelp_ArmCtxHook(0, reset_ctxhook_h);
}