qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Qemu-devel] [PULL 3/3] vfio-pci: Reset workaround for AMD Bonaire and H


From: Alex Williamson
Subject: [Qemu-devel] [PULL 3/3] vfio-pci: Reset workaround for AMD Bonaire and Hawaii GPUs
Date: Tue, 28 Apr 2015 11:52:47 -0600
User-agent: StGit/0.17.1-dirty

Somehow these GPUs manage not to respond to a PCI bus reset, removing
our primary mechanism for resetting graphics cards.  The result is
that these devices typically work well for a single VM boot.  If the
VM is rebooted or restarted, the guest driver is not able to init the
card from the dirty state, resulting in a blue screen for Windows
guests.

The workaround is to use a device specific reset.  This is not 100%
reliable though since it depends on the incoming state of the device,
but it substantially improves the usability of these devices in a VM.

Credit to Alex Deucher <address@hidden> for his guidance.

Signed-off-by: Alex Williamson <address@hidden>
---
 hw/vfio/pci.c |  162 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 162 insertions(+)

diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index 576c677..e0e339a 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -154,6 +154,7 @@ typedef struct VFIOPCIDevice {
     PCIHostDeviceAddress host;
     EventNotifier err_notifier;
     EventNotifier req_notifier;
+    int (*resetfn)(struct VFIOPCIDevice *);
     uint32_t features;
 #define VFIO_FEATURE_ENABLE_VGA_BIT 0
 #define VFIO_FEATURE_ENABLE_VGA (1 << VFIO_FEATURE_ENABLE_VGA_BIT)
@@ -3325,6 +3326,162 @@ static void vfio_unregister_req_notifier(VFIOPCIDevice 
*vdev)
     vdev->req_enabled = false;
 }
 
+/*
+ * AMD Radeon PCI config reset, based on Linux:
+ *   drivers/gpu/drm/radeon/ci_smc.c:ci_is_smc_running()
+ *   drivers/gpu/drm/radeon/radeon_device.c:radeon_pci_config_reset
+ *   drivers/gpu/drm/radeon/ci_smc.c:ci_reset_smc()
+ *   drivers/gpu/drm/radeon/ci_smc.c:ci_stop_smc_clock()
+ * IDs: include/drm/drm_pciids.h
+ * Registers: http://cgit.freedesktop.org/~agd5f/linux/commit/?id=4e2aa447f6f0
+ *
+ * Bonaire and Hawaii GPUs do not respond to a bus reset.  This is a bug in the
+ * hardware that should be fixed on future ASICs.  The symptom of this is that
+ * once the accerlated driver loads, Windows guests will bsod on subsequent
+ * attmpts to load the driver, such as after VM reset or shutdown/restart.  To
+ * work around this, we do an AMD specific PCI config reset, followed by an SMC
+ * reset.  The PCI config reset only works if SMC firmware is running, so we
+ * have a dependency on the state of the device as to whether this reset will
+ * be effective.  There are still cases where we won't be able to kick the
+ * device into working, but this greatly improves the usability overall.  The
+ * config reset magic is relatively common on AMD GPUs, but the setup and SMC
+ * poking is largely ASIC specific.
+ */
+static bool vfio_radeon_smc_is_running(VFIOPCIDevice *vdev)
+{
+    uint32_t clk, pc_c;
+
+    /*
+     * Registers 200h and 204h are index and data registers for acessing
+     * indirect configuration registers within the device.
+     */
+    vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000004, 4);
+    clk = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
+    vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000370, 4);
+    pc_c = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
+
+    return (!(clk & 1) && (0x20100 <= pc_c));
+}
+
+/*
+ * The scope of a config reset is controlled by a mode bit in the misc register
+ * and a fuse, exposed as a bit in another register.  The fuse is the default
+ * (0 = GFX, 1 = whole GPU), the misc bit is a toggle, with the forumula
+ * scope = !(misc ^ fuse), where the resulting scope is defined the same as
+ * the fuse.  A truth table therefore tells us that if misc == fuse, we need
+ * to flip the value of the bit in the misc register.
+ */
+static void vfio_radeon_set_gfx_only_reset(VFIOPCIDevice *vdev)
+{
+    uint32_t misc, fuse;
+    bool a, b;
+
+    vfio_region_write(&vdev->bars[5].region, 0x200, 0xc00c0000, 4);
+    fuse = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
+    b = fuse & 64;
+
+    vfio_region_write(&vdev->bars[5].region, 0x200, 0xc0000010, 4);
+    misc = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
+    a = misc & 2;
+
+    if (a == b) {
+        vfio_region_write(&vdev->bars[5].region, 0x204, misc ^ 2, 4);
+        vfio_region_read(&vdev->bars[5].region, 0x204, 4); /* flush */
+    }
+}
+
+static int vfio_radeon_reset(VFIOPCIDevice *vdev)
+{
+    PCIDevice *pdev = &vdev->pdev;
+    int i, ret = 0;
+    uint32_t data;
+
+    /* Defer to a kernel implemented reset */
+    if (vdev->vbasedev.reset_works) {
+        return -ENODEV;
+    }
+
+    /* Enable only memory BAR access */
+    vfio_pci_write_config(pdev, PCI_COMMAND, PCI_COMMAND_MEMORY, 2);
+
+    /* Reset only works if SMC firmware is loaded and running */
+    if (!vfio_radeon_smc_is_running(vdev)) {
+        ret = -EINVAL;
+        goto out;
+    }
+
+    /* Make sure only the GFX function is reset */
+    vfio_radeon_set_gfx_only_reset(vdev);
+
+    /* AMD PCI config reset */
+    vfio_pci_write_config(pdev, 0x7c, 0x39d5e86b, 4);
+    usleep(100);
+
+    /* Read back the memory size to make sure we're out of reset */
+    for (i = 0; i < 100000; i++) {
+        if (vfio_region_read(&vdev->bars[5].region, 0x5428, 4) != 0xffffffff) {
+            break;
+        }
+        usleep(1);
+    }
+
+    /* Reset SMC */
+    vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000000, 4);
+    data = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
+    data |= 1;
+    vfio_region_write(&vdev->bars[5].region, 0x204, data, 4);
+
+    /* Disable SMC clock */
+    vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000004, 4);
+    data = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
+    data |= 1;
+    vfio_region_write(&vdev->bars[5].region, 0x204, data, 4);
+
+out:
+    /* Restore PCI command register */
+    vfio_pci_write_config(pdev, PCI_COMMAND, 0, 2);
+
+    return ret;
+}
+
+static void vfio_setup_resetfn(VFIOPCIDevice *vdev)
+{
+    PCIDevice *pdev = &vdev->pdev;
+    uint16_t vendor, device;
+
+    vendor = pci_get_word(pdev->config + PCI_VENDOR_ID);
+    device = pci_get_word(pdev->config + PCI_DEVICE_ID);
+
+    switch (vendor) {
+    case 0x1002:
+        switch (device) {
+        /* Bonaire */
+        case 0x6649: /* Bonaire [FirePro W5100] */
+        case 0x6650:
+        case 0x6651:
+        case 0x6658: /* Bonaire XTX [Radeon R7 260X] */
+        case 0x665c: /* Bonaire XT [Radeon HD 7790/8770 / R9 260 OEM] */
+        case 0x665d: /* Bonaire [Radeon R7 200 Series] */
+        /* Hawaii */
+        case 0x67A0: /* Hawaii XT GL [FirePro W9100] */
+        case 0x67A1: /* Hawaii PRO GL [FirePro W8100] */
+        case 0x67A2:
+        case 0x67A8:
+        case 0x67A9:
+        case 0x67AA:
+        case 0x67B0: /* Hawaii XT [Radeon R9 290X] */
+        case 0x67B1: /* Hawaii PRO [Radeon R9 290] */
+        case 0x67B8:
+        case 0x67B9:
+        case 0x67BA:
+        case 0x67BE:
+            vdev->resetfn = vfio_radeon_reset;
+            break;
+        }
+        break;
+    }
+}
+
 static int vfio_initfn(PCIDevice *pdev)
 {
     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
@@ -3473,6 +3630,7 @@ static int vfio_initfn(PCIDevice *pdev)
 
     vfio_register_err_notifier(vdev);
     vfio_register_req_notifier(vdev);
+    vfio_setup_resetfn(vdev);
 
     return 0;
 
@@ -3520,6 +3678,10 @@ static void vfio_pci_reset(DeviceState *dev)
 
     vfio_pci_pre_reset(vdev);
 
+    if (vdev->resetfn && !vdev->resetfn(vdev)) {
+        goto post_reset;
+    }
+
     if (vdev->vbasedev.reset_works &&
         (vdev->has_flr || !vdev->has_pm_reset) &&
         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {




reply via email to

[Prev in Thread] Current Thread [Next in Thread]