All of lore.kernel.org
 help / color / mirror / Atom feed
From: Felix Kuehling <Felix.Kuehling@amd.com>
To: dri-devel@lists.freedesktop.org, amd-gfx@lists.freedesktop.org
Cc: Alex Sierra <alex.sierra@amd.com>
Subject: [PATCH 13/44] drm/amdkfd: add ioctl to configure and query xnack retries
Date: Mon, 22 Mar 2021 06:58:29 -0400	[thread overview]
Message-ID: <20210322105900.14068-14-Felix.Kuehling@amd.com> (raw)
In-Reply-To: <20210322105900.14068-1-Felix.Kuehling@amd.com>

From: Alex Sierra <alex.sierra@amd.com>

Xnack retries are used for page fault recovery. Some AMD chip
families support continuously retry while page table entries are invalid.
The driver must handle the page fault interrupt and fill in a valid entry
for the GPU to continue.

This ioctl allows to enable/disable XNACK retries per KFD process.

Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 28 +++++++++++++++
 include/uapi/linux/kfd_ioctl.h           | 43 +++++++++++++++++++++++-
 2 files changed, 70 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 9511826ac8ae..63eee7ef3355 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -1743,6 +1743,31 @@ static int kfd_ioctl_smi_events(struct file *filep,
 	return kfd_smi_event_open(dev, &args->anon_fd);
 }
 
+static int kfd_ioctl_set_xnack_mode(struct file *filep,
+				    struct kfd_process *p, void *data)
+{
+	struct kfd_ioctl_set_xnack_mode_args *args = data;
+	int r = 0;
+
+	mutex_lock(&p->mutex);
+	if (args->xnack_enabled >= 0) {
+		if (!list_empty(&p->pqm.queues)) {
+			pr_debug("Process has user queues running\n");
+			mutex_unlock(&p->mutex);
+			return -EBUSY;
+		}
+		if (args->xnack_enabled && !kfd_process_xnack_supported(p))
+			r = -EPERM;
+		else
+			p->xnack_enabled = args->xnack_enabled;
+	} else {
+		args->xnack_enabled = p->xnack_enabled;
+	}
+	mutex_unlock(&p->mutex);
+
+	return r;
+}
+
 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
 {
 	struct kfd_ioctl_svm_args *args = data;
@@ -1869,6 +1894,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 			kfd_ioctl_smi_events, 0),
 
 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
+
+	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
+			kfd_ioctl_set_xnack_mode, 0),
 };
 
 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
index 247b57baa94f..3cb5b5dd9f77 100644
--- a/include/uapi/linux/kfd_ioctl.h
+++ b/include/uapi/linux/kfd_ioctl.h
@@ -597,6 +597,44 @@ struct kfd_ioctl_svm_args {
 	struct kfd_ioctl_svm_attribute attrs[0];
 };
 
+/**
+ * kfd_ioctl_set_xnack_mode_args - Arguments for set_xnack_mode
+ *
+ * @xnack_enabled:       [in/out] Whether to enable XNACK mode for this process
+ *
+ * @xnack_enabled indicates whether recoverable page faults should be
+ * enabled for the current process. 0 means disabled, positive means
+ * enabled, negative means leave unchanged. If enabled, virtual address
+ * translations on GFXv9 and later AMD GPUs can return XNACK and retry
+ * the access until a valid PTE is available. This is used to implement
+ * device page faults.
+ *
+ * On output, @xnack_enabled returns the (new) current mode (0 or
+ * positive). Therefore, a negative input value can be used to query
+ * the current mode without changing it.
+ *
+ * The XNACK mode fundamentally changes the way SVM managed memory works
+ * in the driver, with subtle effects on application performance and
+ * functionality.
+ *
+ * Enabling XNACK mode requires shader programs to be compiled
+ * differently. Furthermore, not all GPUs support changing the mode
+ * per-process. Therefore changing the mode is only allowed while no
+ * user mode queues exist in the process. This ensure that no shader
+ * code is running that may be compiled for the wrong mode. And GPUs
+ * that cannot change to the requested mode will prevent the XNACK
+ * mode from occurring. All GPUs used by the process must be in the
+ * same XNACK mode.
+ *
+ * GFXv8 or older GPUs do not support 48 bit virtual addresses or SVM.
+ * Therefore those GPUs are not considered for the XNACK mode switch.
+ *
+ * Return: 0 on success, -errno on failure
+ */
+struct kfd_ioctl_set_xnack_mode_args {
+	__s32 xnack_enabled;
+};
+
 #define AMDKFD_IOCTL_BASE 'K'
 #define AMDKFD_IO(nr)			_IO(AMDKFD_IOCTL_BASE, nr)
 #define AMDKFD_IOR(nr, type)		_IOR(AMDKFD_IOCTL_BASE, nr, type)
@@ -699,7 +737,10 @@ struct kfd_ioctl_svm_args {
 
 #define AMDKFD_IOC_SVM	AMDKFD_IOWR(0x20, struct kfd_ioctl_svm_args)
 
+#define AMDKFD_IOC_SET_XNACK_MODE		\
+		AMDKFD_IOWR(0x21, struct kfd_ioctl_set_xnack_mode_args)
+
 #define AMDKFD_COMMAND_START		0x01
-#define AMDKFD_COMMAND_END		0x21
+#define AMDKFD_COMMAND_END		0x22
 
 #endif
-- 
2.31.0

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

WARNING: multiple messages have this Message-ID (diff)
From: Felix Kuehling <Felix.Kuehling@amd.com>
To: dri-devel@lists.freedesktop.org, amd-gfx@lists.freedesktop.org
Cc: Alex Sierra <alex.sierra@amd.com>
Subject: [PATCH 13/44] drm/amdkfd: add ioctl to configure and query xnack retries
Date: Mon, 22 Mar 2021 06:58:29 -0400	[thread overview]
Message-ID: <20210322105900.14068-14-Felix.Kuehling@amd.com> (raw)
In-Reply-To: <20210322105900.14068-1-Felix.Kuehling@amd.com>

From: Alex Sierra <alex.sierra@amd.com>

Xnack retries are used for page fault recovery. Some AMD chip
families support continuously retry while page table entries are invalid.
The driver must handle the page fault interrupt and fill in a valid entry
for the GPU to continue.

This ioctl allows to enable/disable XNACK retries per KFD process.

Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 28 +++++++++++++++
 include/uapi/linux/kfd_ioctl.h           | 43 +++++++++++++++++++++++-
 2 files changed, 70 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 9511826ac8ae..63eee7ef3355 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -1743,6 +1743,31 @@ static int kfd_ioctl_smi_events(struct file *filep,
 	return kfd_smi_event_open(dev, &args->anon_fd);
 }
 
+static int kfd_ioctl_set_xnack_mode(struct file *filep,
+				    struct kfd_process *p, void *data)
+{
+	struct kfd_ioctl_set_xnack_mode_args *args = data;
+	int r = 0;
+
+	mutex_lock(&p->mutex);
+	if (args->xnack_enabled >= 0) {
+		if (!list_empty(&p->pqm.queues)) {
+			pr_debug("Process has user queues running\n");
+			mutex_unlock(&p->mutex);
+			return -EBUSY;
+		}
+		if (args->xnack_enabled && !kfd_process_xnack_supported(p))
+			r = -EPERM;
+		else
+			p->xnack_enabled = args->xnack_enabled;
+	} else {
+		args->xnack_enabled = p->xnack_enabled;
+	}
+	mutex_unlock(&p->mutex);
+
+	return r;
+}
+
 static int kfd_ioctl_svm(struct file *filep, struct kfd_process *p, void *data)
 {
 	struct kfd_ioctl_svm_args *args = data;
@@ -1869,6 +1894,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 			kfd_ioctl_smi_events, 0),
 
 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SVM, kfd_ioctl_svm, 0),
+
+	AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
+			kfd_ioctl_set_xnack_mode, 0),
 };
 
 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)
diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
index 247b57baa94f..3cb5b5dd9f77 100644
--- a/include/uapi/linux/kfd_ioctl.h
+++ b/include/uapi/linux/kfd_ioctl.h
@@ -597,6 +597,44 @@ struct kfd_ioctl_svm_args {
 	struct kfd_ioctl_svm_attribute attrs[0];
 };
 
+/**
+ * kfd_ioctl_set_xnack_mode_args - Arguments for set_xnack_mode
+ *
+ * @xnack_enabled:       [in/out] Whether to enable XNACK mode for this process
+ *
+ * @xnack_enabled indicates whether recoverable page faults should be
+ * enabled for the current process. 0 means disabled, positive means
+ * enabled, negative means leave unchanged. If enabled, virtual address
+ * translations on GFXv9 and later AMD GPUs can return XNACK and retry
+ * the access until a valid PTE is available. This is used to implement
+ * device page faults.
+ *
+ * On output, @xnack_enabled returns the (new) current mode (0 or
+ * positive). Therefore, a negative input value can be used to query
+ * the current mode without changing it.
+ *
+ * The XNACK mode fundamentally changes the way SVM managed memory works
+ * in the driver, with subtle effects on application performance and
+ * functionality.
+ *
+ * Enabling XNACK mode requires shader programs to be compiled
+ * differently. Furthermore, not all GPUs support changing the mode
+ * per-process. Therefore changing the mode is only allowed while no
+ * user mode queues exist in the process. This ensure that no shader
+ * code is running that may be compiled for the wrong mode. And GPUs
+ * that cannot change to the requested mode will prevent the XNACK
+ * mode from occurring. All GPUs used by the process must be in the
+ * same XNACK mode.
+ *
+ * GFXv8 or older GPUs do not support 48 bit virtual addresses or SVM.
+ * Therefore those GPUs are not considered for the XNACK mode switch.
+ *
+ * Return: 0 on success, -errno on failure
+ */
+struct kfd_ioctl_set_xnack_mode_args {
+	__s32 xnack_enabled;
+};
+
 #define AMDKFD_IOCTL_BASE 'K'
 #define AMDKFD_IO(nr)			_IO(AMDKFD_IOCTL_BASE, nr)
 #define AMDKFD_IOR(nr, type)		_IOR(AMDKFD_IOCTL_BASE, nr, type)
@@ -699,7 +737,10 @@ struct kfd_ioctl_svm_args {
 
 #define AMDKFD_IOC_SVM	AMDKFD_IOWR(0x20, struct kfd_ioctl_svm_args)
 
+#define AMDKFD_IOC_SET_XNACK_MODE		\
+		AMDKFD_IOWR(0x21, struct kfd_ioctl_set_xnack_mode_args)
+
 #define AMDKFD_COMMAND_START		0x01
-#define AMDKFD_COMMAND_END		0x21
+#define AMDKFD_COMMAND_END		0x22
 
 #endif
-- 
2.31.0

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

  parent reply	other threads:[~2021-03-22 11:08 UTC|newest]

Thread overview: 100+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-03-22 10:58 [PATCH 00/44] Add HMM-based SVM memory manager to KFD v2 Felix Kuehling
2021-03-22 10:58 ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 01/44] drm/amdgpu: replace per_device_list by array Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-23 14:30   ` Kim, Jonathan
2021-03-23 14:30     ` Kim, Jonathan
2021-03-22 10:58 ` [PATCH 02/44] drm/amdkfd: helper to convert gpu id and idx Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 03/44] drm/amdkfd: add svm ioctl API Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 04/44] drm/amdkfd: register svm range Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 05/44] drm/amdkfd: add svm ioctl GET_ATTR op Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 06/44] drm/amdgpu: add common HMM get pages function Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 07/44] drm/amdkfd: validate svm range system memory Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 08/44] drm/amdkfd: deregister svm range Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 09/44] drm/amdgpu: export vm update mapping interface Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 10/44] drm/amdkfd: map svm range to GPUs Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 11/44] drm/amdkfd: svm range eviction and restore Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 12/44] drm/amdkfd: add xnack enabled flag to kfd_process Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` Felix Kuehling [this message]
2021-03-22 10:58   ` [PATCH 13/44] drm/amdkfd: add ioctl to configure and query xnack retries Felix Kuehling
2021-03-22 10:58 ` [PATCH 14/44] drm/amdkfd: register HMM device private zone Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 15/44] drm/amdkfd: validate vram svm range from TTM Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 16/44] drm/amdkfd: support xgmi same hive mapping Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 17/44] drm/amdkfd: copy memory through gart table Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 18/44] drm/amdkfd: HMM migrate ram to vram Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 19/44] drm/amdkfd: HMM migrate vram to ram Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 20/44] drm/amdkfd: invalidate tables on page retry fault Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 21/44] drm/amdgpu: enable 48-bit IH timestamp counter Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 22/44] drm/amdkfd: page table restore through svm API Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 23/44] drm/amdkfd: SVM API call to restore page tables Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 24/44] drm/amdkfd: add svm_bo reference for eviction fence Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 25/44] drm/amdgpu: add param bit flag to create SVM BOs Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 26/44] drm/amdkfd: add svm_bo eviction mechanism support Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 27/44] drm/amdgpu: svm bo enable_signal call condition Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 28/44] drm/amdgpu: add svm_bo eviction to enable_signal cb Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 29/44] drm/amdgpu: reserve fence slot to update page table Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 13:32   ` Christian König
2021-03-22 13:32     ` Christian König
2021-03-22 10:58 ` [PATCH 30/44] drm/amdkfd: refine migration policy with xnack on Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 31/44] drm/amdkfd: add svm range validate timestamp Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 32/44] drm/amdkfd: multiple gpu migrate vram to vram Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 33/44] drm/amdkfd: Add SVM API support capability bits Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 34/44] drm/amdkfd: Fix dma unmapping Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 35/44] drm/amdkfd: Call mutex_destroy Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 36/44] drm/amdkfd: Fix spurious restore failures Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 37/44] drm/amdkfd: Fix svm_bo_list locking in eviction worker Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 38/44] drm/amdkfd: Simplify split_by_granularity Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 39/44] drm/amdkfd: Point out several race conditions Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 40/44] drm/amdkfd: Return pdd from kfd_process_device_from_gduid Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 41/44] drm/amdkfd: Remove broken deferred mapping Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 42/44] drm/amdkfd: Allow invalid pages in migration.src Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:58 ` [PATCH 43/44] drm/amdkfd: Correct locking during migration and mapping Felix Kuehling
2021-03-22 10:58   ` Felix Kuehling
2021-03-22 10:59 ` [PATCH 44/44] drm/amdkfd: Nested locking and invalidation of child ranges Felix Kuehling
2021-03-22 10:59   ` Felix Kuehling
2021-03-22 14:15 ` [PATCH 00/44] Add HMM-based SVM memory manager to KFD v2 Daniel Vetter
2021-03-22 14:15   ` Daniel Vetter
2021-03-22 16:06   ` Felix Kuehling
2021-03-22 16:06     ` Felix Kuehling
2021-03-22 17:04     ` Daniel Vetter
2021-03-22 17:04       ` Daniel Vetter

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210322105900.14068-14-Felix.Kuehling@amd.com \
    --to=felix.kuehling@amd.com \
    --cc=alex.sierra@amd.com \
    --cc=amd-gfx@lists.freedesktop.org \
    --cc=dri-devel@lists.freedesktop.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.