dmabuf: Add the capability to expose DMA-BUF stats in sysfs
authorHridya Valsaraju <hridya@google.com>
Thu, 3 Jun 2021 21:47:51 +0000 (14:47 -0700)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Tue, 15 Jun 2021 09:50:24 +0000 (11:50 +0200)
Overview
========
The patch adds DMA-BUF statistics to /sys/kernel/dmabuf/buffers. It
allows statistics to be enabled for each DMA-BUF in sysfs by enabling
the config CONFIG_DMABUF_SYSFS_STATS.

The following stats will be exposed by the interface:

/sys/kernel/dmabuf/buffers/<inode_number>/exporter_name
/sys/kernel/dmabuf/buffers/<inode_number>/size
/sys/kernel/dmabuf/buffers/<inode_number>/attachments/<attach_uid>/device
/sys/kernel/dmabuf/buffers/<inode_number>/attachments/<attach_uid>/map_counter

The inode_number is unique for each DMA-BUF and was added earlier [1]
in order to allow userspace to track DMA-BUF usage across different
processes.

Use Cases
=========
The interface provides a way to gather DMA-BUF per-buffer statistics
from production devices. These statistics will be used to derive DMA-BUF
per-exporter stats and per-device usage stats for Android Bug reports.
The corresponding userspace changes can be found at [2].
Telemetry tools will also capture this information(along with other
memory metrics) periodically as well as on important events like a
foreground app kill (which might have been triggered by Low Memory
Killer). It will also contribute to provide a snapshot of the system
memory usage on other events such as OOM kills and Application Not
Responding events.

Background
==========
Currently, there are two existing interfaces that provide information
about DMA-BUFs.
1) /sys/kernel/debug/dma_buf/bufinfo
debugfs is however unsuitable to be mounted in production systems and
cannot be considered as an alternative to the sysfs interface being
proposed.
2) proc/<pid>/fdinfo/<fd>
The proc/<pid>/fdinfo/<fd> files expose information about DMA-BUF fds.
However, the existing procfs interfaces can only provide information
about the buffers for which processes hold fds or have the buffers
mmapped into their address space. Since the procfs interfaces alone
cannot provide a full picture of all DMA-BUFs in the system, there is
the need for an alternate interface to provide this information on
production systems.

The patch contains the following major improvements over v1:
1) Each attachment is represented by its own directory to allow creating
a symlink to the importing device and to also provide room for future
expansion.
2) The number of distinct mappings of each attachment is exposed in a
separate file.
3) The per-buffer statistics are now in /sys/kernel/dmabuf/buffers
inorder to make the interface expandable in future.

All of the improvements above are based on suggestions/feedback from
Daniel Vetter and Christian König.

A shell script that can be run on a classic Linux environment to read
out the DMA-BUF statistics can be found at [3](suggested by John
Stultz).

[1]: https://lore.kernel.org/patchwork/patch/1088791/
[2]: https://android-review.googlesource.com/q/topic:%22dmabuf-sysfs%22+(status:open%20OR%20status:merged)
[3]: https://android-review.googlesource.com/c/platform/system/memory/libmeminfo/+/1549734

Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Hridya Valsaraju <hridya@google.com>
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20210603214758.2955251-1-hridya@google.com
Documentation/ABI/testing/sysfs-kernel-dmabuf-buffers [new file with mode: 0644]
Documentation/driver-api/dma-buf.rst
drivers/dma-buf/Kconfig
drivers/dma-buf/Makefile
drivers/dma-buf/dma-buf-sysfs-stats.c [new file with mode: 0644]
drivers/dma-buf/dma-buf-sysfs-stats.h [new file with mode: 0644]
drivers/dma-buf/dma-buf.c
include/linux/dma-buf.h

diff --git a/Documentation/ABI/testing/sysfs-kernel-dmabuf-buffers b/Documentation/ABI/testing/sysfs-kernel-dmabuf-buffers
new file mode 100644 (file)
index 0000000..a243984
--- /dev/null
@@ -0,0 +1,52 @@
+What:          /sys/kernel/dmabuf/buffers
+Date:          May 2021
+KernelVersion: v5.13
+Contact:       Hridya Valsaraju <hridya@google.com>
+Description:   The /sys/kernel/dmabuf/buffers directory contains a
+               snapshot of the internal state of every DMA-BUF.
+               /sys/kernel/dmabuf/buffers/<inode_number> will contain the
+               statistics for the DMA-BUF with the unique inode number
+               <inode_number>
+Users:         kernel memory tuning/debugging tools
+
+What:          /sys/kernel/dmabuf/buffers/<inode_number>/exporter_name
+Date:          May 2021
+KernelVersion: v5.13
+Contact:       Hridya Valsaraju <hridya@google.com>
+Description:   This file is read-only and contains the name of the exporter of
+               the DMA-BUF.
+
+What:          /sys/kernel/dmabuf/buffers/<inode_number>/size
+Date:          May 2021
+KernelVersion: v5.13
+Contact:       Hridya Valsaraju <hridya@google.com>
+Description:   This file is read-only and specifies the size of the DMA-BUF in
+               bytes.
+
+What:          /sys/kernel/dmabuf/buffers/<inode_number>/attachments
+Date:          May 2021
+KernelVersion: v5.13
+Contact:       Hridya Valsaraju <hridya@google.com>
+Description:   This directory will contain subdirectories representing every
+               attachment of the DMA-BUF.
+
+What:          /sys/kernel/dmabuf/buffers/<inode_number>/attachments/<attachment_uid>
+Date:          May 2021
+KernelVersion: v5.13
+Contact:       Hridya Valsaraju <hridya@google.com>
+Description:   This directory will contain information on the attached device
+               and the number of current distinct device mappings.
+
+What:          /sys/kernel/dmabuf/buffers/<inode_number>/attachments/<attachment_uid>/device
+Date:          May 2021
+KernelVersion: v5.13
+Contact:       Hridya Valsaraju <hridya@google.com>
+Description:   This file is read-only and is a symlink to the attached device's
+               sysfs entry.
+
+What:          /sys/kernel/dmabuf/buffers/<inode_number>/attachments/<attachment_uid>/map_counter
+Date:          May 2021
+KernelVersion: v5.13
+Contact:       Hridya Valsaraju <hridya@google.com>
+Description:   This file is read-only and contains a map_counter indicating the
+               number of distinct device mappings of the attachment.
index 7f21425..84aab65 100644 (file)
@@ -106,6 +106,11 @@ Implicit Fence Poll Support
 .. kernel-doc:: drivers/dma-buf/dma-buf.c
    :doc: implicit fence polling
 
+DMA-BUF statistics
+~~~~~~~~~~~~~~~~~~
+.. kernel-doc:: drivers/dma-buf/dma-buf-sysfs-stats.c
+   :doc: overview
+
 Kernel Functions and Structures Reference
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
index 4e16c71..9561e3d 100644 (file)
@@ -72,6 +72,17 @@ menuconfig DMABUF_HEAPS
          allows userspace to allocate dma-bufs that can be shared
          between drivers.
 
+menuconfig DMABUF_SYSFS_STATS
+       bool "DMA-BUF sysfs statistics"
+       select DMA_SHARED_BUFFER
+       help
+          Choose this option to enable DMA-BUF sysfs statistics
+          in location /sys/kernel/dmabuf/buffers.
+
+          /sys/kernel/dmabuf/buffers/<inode_number> will contain
+          statistics for the DMA-BUF with the unique inode number
+          <inode_number>.
+
 source "drivers/dma-buf/heaps/Kconfig"
 
 endmenu
index 995e05f..40d81f2 100644 (file)
@@ -6,6 +6,7 @@ obj-$(CONFIG_DMABUF_HEAPS)      += heaps/
 obj-$(CONFIG_SYNC_FILE)                += sync_file.o
 obj-$(CONFIG_SW_SYNC)          += sw_sync.o sync_debug.o
 obj-$(CONFIG_UDMABUF)          += udmabuf.o
+obj-$(CONFIG_DMABUF_SYSFS_STATS) += dma-buf-sysfs-stats.o
 
 dmabuf_selftests-y := \
        selftest.o \
diff --git a/drivers/dma-buf/dma-buf-sysfs-stats.c b/drivers/dma-buf/dma-buf-sysfs-stats.c
new file mode 100644 (file)
index 0000000..a2638e8
--- /dev/null
@@ -0,0 +1,337 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * DMA-BUF sysfs statistics.
+ *
+ * Copyright (C) 2021 Google LLC.
+ */
+
+#include <linux/dma-buf.h>
+#include <linux/dma-resv.h>
+#include <linux/kobject.h>
+#include <linux/printk.h>
+#include <linux/slab.h>
+#include <linux/sysfs.h>
+
+#include "dma-buf-sysfs-stats.h"
+
+#define to_dma_buf_entry_from_kobj(x) container_of(x, struct dma_buf_sysfs_entry, kobj)
+
+/**
+ * DOC: overview
+ *
+ * ``/sys/kernel/debug/dma_buf/bufinfo`` provides an overview of every DMA-BUF
+ * in the system. However, since debugfs is not safe to be mounted in
+ * production, procfs and sysfs can be used to gather DMA-BUF statistics on
+ * production systems.
+ *
+ * The ``/proc/<pid>/fdinfo/<fd>`` files in procfs can be used to gather
+ * information about DMA-BUF fds. Detailed documentation about the interface
+ * is present in Documentation/filesystems/proc.rst.
+ *
+ * Unfortunately, the existing procfs interfaces can only provide information
+ * about the DMA-BUFs for which processes hold fds or have the buffers mmapped
+ * into their address space. This necessitated the creation of the DMA-BUF sysfs
+ * statistics interface to provide per-buffer information on production systems.
+ *
+ * The interface at ``/sys/kernel/dma-buf/buffers`` exposes information about
+ * every DMA-BUF when ``CONFIG_DMABUF_SYSFS_STATS`` is enabled.
+ *
+ * The following stats are exposed by the interface:
+ *
+ * * ``/sys/kernel/dmabuf/buffers/<inode_number>/exporter_name``
+ * * ``/sys/kernel/dmabuf/buffers/<inode_number>/size``
+ * * ``/sys/kernel/dmabuf/buffers/<inode_number>/attachments/<attach_uid>/device``
+ * * ``/sys/kernel/dmabuf/buffers/<inode_number>/attachments/<attach_uid>/map_counter``
+ *
+ * The information in the interface can also be used to derive per-exporter and
+ * per-device usage statistics. The data from the interface can be gathered
+ * on error conditions or other important events to provide a snapshot of
+ * DMA-BUF usage. It can also be collected periodically by telemetry to monitor
+ * various metrics.
+ *
+ * Detailed documentation about the interface is present in
+ * Documentation/ABI/testing/sysfs-kernel-dmabuf-buffers.
+ */
+
+struct dma_buf_stats_attribute {
+       struct attribute attr;
+       ssize_t (*show)(struct dma_buf *dmabuf,
+                       struct dma_buf_stats_attribute *attr, char *buf);
+};
+#define to_dma_buf_stats_attr(x) container_of(x, struct dma_buf_stats_attribute, attr)
+
+static ssize_t dma_buf_stats_attribute_show(struct kobject *kobj,
+                                           struct attribute *attr,
+                                           char *buf)
+{
+       struct dma_buf_stats_attribute *attribute;
+       struct dma_buf_sysfs_entry *sysfs_entry;
+       struct dma_buf *dmabuf;
+
+       attribute = to_dma_buf_stats_attr(attr);
+       sysfs_entry = to_dma_buf_entry_from_kobj(kobj);
+       dmabuf = sysfs_entry->dmabuf;
+
+       if (!dmabuf || !attribute->show)
+               return -EIO;
+
+       return attribute->show(dmabuf, attribute, buf);
+}
+
+static const struct sysfs_ops dma_buf_stats_sysfs_ops = {
+       .show = dma_buf_stats_attribute_show,
+};
+
+static ssize_t exporter_name_show(struct dma_buf *dmabuf,
+                                 struct dma_buf_stats_attribute *attr,
+                                 char *buf)
+{
+       return sysfs_emit(buf, "%s\n", dmabuf->exp_name);
+}
+
+static ssize_t size_show(struct dma_buf *dmabuf,
+                        struct dma_buf_stats_attribute *attr,
+                        char *buf)
+{
+       return sysfs_emit(buf, "%zu\n", dmabuf->size);
+}
+
+static struct dma_buf_stats_attribute exporter_name_attribute =
+       __ATTR_RO(exporter_name);
+static struct dma_buf_stats_attribute size_attribute = __ATTR_RO(size);
+
+static struct attribute *dma_buf_stats_default_attrs[] = {
+       &exporter_name_attribute.attr,
+       &size_attribute.attr,
+       NULL,
+};
+ATTRIBUTE_GROUPS(dma_buf_stats_default);
+
+static void dma_buf_sysfs_release(struct kobject *kobj)
+{
+       struct dma_buf_sysfs_entry *sysfs_entry;
+
+       sysfs_entry = to_dma_buf_entry_from_kobj(kobj);
+       kfree(sysfs_entry);
+}
+
+static struct kobj_type dma_buf_ktype = {
+       .sysfs_ops = &dma_buf_stats_sysfs_ops,
+       .release = dma_buf_sysfs_release,
+       .default_groups = dma_buf_stats_default_groups,
+};
+
+#define to_dma_buf_attach_entry_from_kobj(x) container_of(x, struct dma_buf_attach_sysfs_entry, kobj)
+
+struct dma_buf_attach_stats_attribute {
+       struct attribute attr;
+       ssize_t (*show)(struct dma_buf_attach_sysfs_entry *sysfs_entry,
+                       struct dma_buf_attach_stats_attribute *attr, char *buf);
+};
+#define to_dma_buf_attach_stats_attr(x) container_of(x, struct dma_buf_attach_stats_attribute, attr)
+
+static ssize_t dma_buf_attach_stats_attribute_show(struct kobject *kobj,
+                                                  struct attribute *attr,
+                                                  char *buf)
+{
+       struct dma_buf_attach_stats_attribute *attribute;
+       struct dma_buf_attach_sysfs_entry *sysfs_entry;
+
+       attribute = to_dma_buf_attach_stats_attr(attr);
+       sysfs_entry = to_dma_buf_attach_entry_from_kobj(kobj);
+
+       if (!attribute->show)
+               return -EIO;
+
+       return attribute->show(sysfs_entry, attribute, buf);
+}
+
+static const struct sysfs_ops dma_buf_attach_stats_sysfs_ops = {
+       .show = dma_buf_attach_stats_attribute_show,
+};
+
+static ssize_t map_counter_show(struct dma_buf_attach_sysfs_entry *sysfs_entry,
+                               struct dma_buf_attach_stats_attribute *attr,
+                               char *buf)
+{
+       return sysfs_emit(buf, "%u\n", sysfs_entry->map_counter);
+}
+
+static struct dma_buf_attach_stats_attribute map_counter_attribute =
+       __ATTR_RO(map_counter);
+
+static struct attribute *dma_buf_attach_stats_default_attrs[] = {
+       &map_counter_attribute.attr,
+       NULL,
+};
+ATTRIBUTE_GROUPS(dma_buf_attach_stats_default);
+
+static void dma_buf_attach_sysfs_release(struct kobject *kobj)
+{
+       struct dma_buf_attach_sysfs_entry *sysfs_entry;
+
+       sysfs_entry = to_dma_buf_attach_entry_from_kobj(kobj);
+       kfree(sysfs_entry);
+}
+
+static struct kobj_type dma_buf_attach_ktype = {
+       .sysfs_ops = &dma_buf_attach_stats_sysfs_ops,
+       .release = dma_buf_attach_sysfs_release,
+       .default_groups = dma_buf_attach_stats_default_groups,
+};
+
+void dma_buf_attach_stats_teardown(struct dma_buf_attachment *attach)
+{
+       struct dma_buf_attach_sysfs_entry *sysfs_entry;
+
+       sysfs_entry = attach->sysfs_entry;
+       if (!sysfs_entry)
+               return;
+
+       sysfs_delete_link(&sysfs_entry->kobj, &attach->dev->kobj, "device");
+
+       kobject_del(&sysfs_entry->kobj);
+       kobject_put(&sysfs_entry->kobj);
+}
+
+int dma_buf_attach_stats_setup(struct dma_buf_attachment *attach,
+                              unsigned int uid)
+{
+       struct dma_buf_attach_sysfs_entry *sysfs_entry;
+       int ret;
+       struct dma_buf *dmabuf;
+
+       if (!attach)
+               return -EINVAL;
+
+       dmabuf = attach->dmabuf;
+
+       sysfs_entry = kzalloc(sizeof(struct dma_buf_attach_sysfs_entry),
+                             GFP_KERNEL);
+       if (!sysfs_entry)
+               return -ENOMEM;
+
+       sysfs_entry->kobj.kset = dmabuf->sysfs_entry->attach_stats_kset;
+
+       attach->sysfs_entry = sysfs_entry;
+
+       ret = kobject_init_and_add(&sysfs_entry->kobj, &dma_buf_attach_ktype,
+                                  NULL, "%u", uid);
+       if (ret)
+               goto kobj_err;
+
+       ret = sysfs_create_link(&sysfs_entry->kobj, &attach->dev->kobj,
+                               "device");
+       if (ret)
+               goto link_err;
+
+       return 0;
+
+link_err:
+       kobject_del(&sysfs_entry->kobj);
+kobj_err:
+       kobject_put(&sysfs_entry->kobj);
+       attach->sysfs_entry = NULL;
+
+       return ret;
+}
+void dma_buf_stats_teardown(struct dma_buf *dmabuf)
+{
+       struct dma_buf_sysfs_entry *sysfs_entry;
+
+       sysfs_entry = dmabuf->sysfs_entry;
+       if (!sysfs_entry)
+               return;
+
+       kset_unregister(sysfs_entry->attach_stats_kset);
+       kobject_del(&sysfs_entry->kobj);
+       kobject_put(&sysfs_entry->kobj);
+}
+
+
+/* Statistics files do not need to send uevents. */
+static int dmabuf_sysfs_uevent_filter(struct kset *kset, struct kobject *kobj)
+{
+       return 0;
+}
+
+static const struct kset_uevent_ops dmabuf_sysfs_no_uevent_ops = {
+       .filter = dmabuf_sysfs_uevent_filter,
+};
+
+static struct kset *dma_buf_stats_kset;
+static struct kset *dma_buf_per_buffer_stats_kset;
+int dma_buf_init_sysfs_statistics(void)
+{
+       dma_buf_stats_kset = kset_create_and_add("dmabuf",
+                                                &dmabuf_sysfs_no_uevent_ops,
+                                                kernel_kobj);
+       if (!dma_buf_stats_kset)
+               return -ENOMEM;
+
+       dma_buf_per_buffer_stats_kset = kset_create_and_add("buffers",
+                                                           &dmabuf_sysfs_no_uevent_ops,
+                                                           &dma_buf_stats_kset->kobj);
+       if (!dma_buf_per_buffer_stats_kset) {
+               kset_unregister(dma_buf_stats_kset);
+               return -ENOMEM;
+       }
+
+       return 0;
+}
+
+void dma_buf_uninit_sysfs_statistics(void)
+{
+       kset_unregister(dma_buf_per_buffer_stats_kset);
+       kset_unregister(dma_buf_stats_kset);
+}
+
+int dma_buf_stats_setup(struct dma_buf *dmabuf)
+{
+       struct dma_buf_sysfs_entry *sysfs_entry;
+       int ret;
+       struct kset *attach_stats_kset;
+
+       if (!dmabuf || !dmabuf->file)
+               return -EINVAL;
+
+       if (!dmabuf->exp_name) {
+               pr_err("exporter name must not be empty if stats needed\n");
+               return -EINVAL;
+       }
+
+       sysfs_entry = kzalloc(sizeof(struct dma_buf_sysfs_entry), GFP_KERNEL);
+       if (!sysfs_entry)
+               return -ENOMEM;
+
+       sysfs_entry->kobj.kset = dma_buf_per_buffer_stats_kset;
+       sysfs_entry->dmabuf = dmabuf;
+
+       dmabuf->sysfs_entry = sysfs_entry;
+
+       /* create the directory for buffer stats */
+       ret = kobject_init_and_add(&sysfs_entry->kobj, &dma_buf_ktype, NULL,
+                                  "%lu", file_inode(dmabuf->file)->i_ino);
+       if (ret)
+               goto err_sysfs_dmabuf;
+
+       /* create the directory for attachment stats */
+       attach_stats_kset = kset_create_and_add("attachments",
+                                               &dmabuf_sysfs_no_uevent_ops,
+                                               &sysfs_entry->kobj);
+       if (!attach_stats_kset) {
+               ret = -ENOMEM;
+               goto err_sysfs_attach;
+       }
+
+       sysfs_entry->attach_stats_kset = attach_stats_kset;
+
+       return 0;
+
+err_sysfs_attach:
+       kobject_del(&sysfs_entry->kobj);
+err_sysfs_dmabuf:
+       kobject_put(&sysfs_entry->kobj);
+       dmabuf->sysfs_entry = NULL;
+       return ret;
+}
diff --git a/drivers/dma-buf/dma-buf-sysfs-stats.h b/drivers/dma-buf/dma-buf-sysfs-stats.h
new file mode 100644 (file)
index 0000000..5f47032
--- /dev/null
@@ -0,0 +1,62 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * DMA-BUF sysfs statistics.
+ *
+ * Copyright (C) 2021 Google LLC.
+ */
+
+#ifndef _DMA_BUF_SYSFS_STATS_H
+#define _DMA_BUF_SYSFS_STATS_H
+
+#ifdef CONFIG_DMABUF_SYSFS_STATS
+
+int dma_buf_init_sysfs_statistics(void);
+void dma_buf_uninit_sysfs_statistics(void);
+
+int dma_buf_stats_setup(struct dma_buf *dmabuf);
+int dma_buf_attach_stats_setup(struct dma_buf_attachment *attach,
+                              unsigned int uid);
+static inline void dma_buf_update_attachment_map_count(struct dma_buf_attachment *attach,
+                                                      int delta)
+{
+       struct dma_buf_attach_sysfs_entry *entry = attach->sysfs_entry;
+
+       entry->map_counter += delta;
+}
+void dma_buf_stats_teardown(struct dma_buf *dmabuf);
+void dma_buf_attach_stats_teardown(struct dma_buf_attachment *attach);
+static inline unsigned int dma_buf_update_attach_uid(struct dma_buf *dmabuf)
+{
+       struct dma_buf_sysfs_entry *entry = dmabuf->sysfs_entry;
+
+       return entry->attachment_uid++;
+}
+#else
+
+static inline int dma_buf_init_sysfs_statistics(void)
+{
+       return 0;
+}
+
+static inline void dma_buf_uninit_sysfs_statistics(void) {}
+
+static inline int dma_buf_stats_setup(struct dma_buf *dmabuf)
+{
+       return 0;
+}
+static inline int dma_buf_attach_stats_setup(struct dma_buf_attachment *attach,
+                                            unsigned int uid)
+{
+       return 0;
+}
+
+static inline void dma_buf_stats_teardown(struct dma_buf *dmabuf) {}
+static inline void dma_buf_attach_stats_teardown(struct dma_buf_attachment *attach) {}
+static inline void dma_buf_update_attachment_map_count(struct dma_buf_attachment *attach,
+                                                      int delta) {}
+static inline unsigned int dma_buf_update_attach_uid(struct dma_buf *dmabuf)
+{
+       return 0;
+}
+#endif
+#endif // _DMA_BUF_SYSFS_STATS_H
index 511fe0d..d012140 100644 (file)
@@ -29,6 +29,8 @@
 #include <uapi/linux/dma-buf.h>
 #include <uapi/linux/magic.h>
 
+#include "dma-buf-sysfs-stats.h"
+
 static inline int is_dma_buf_file(struct file *);
 
 struct dma_buf_list {
@@ -79,6 +81,7 @@ static void dma_buf_release(struct dentry *dentry)
        if (dmabuf->resv == (struct dma_resv *)&dmabuf[1])
                dma_resv_fini(dmabuf->resv);
 
+       dma_buf_stats_teardown(dmabuf);
        module_put(dmabuf->owner);
        kfree(dmabuf->name);
        kfree(dmabuf);
@@ -580,6 +583,10 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info)
        file->f_mode |= FMODE_LSEEK;
        dmabuf->file = file;
 
+       ret = dma_buf_stats_setup(dmabuf);
+       if (ret)
+               goto err_sysfs;
+
        mutex_init(&dmabuf->lock);
        INIT_LIST_HEAD(&dmabuf->attachments);
 
@@ -589,6 +596,14 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info)
 
        return dmabuf;
 
+err_sysfs:
+       /*
+        * Set file->f_path.dentry->d_fsdata to NULL so that when
+        * dma_buf_release() gets invoked by dentry_ops, it exits
+        * early before calling the release() dma_buf op.
+        */
+       file->f_path.dentry->d_fsdata = NULL;
+       fput(file);
 err_dmabuf:
        kfree(dmabuf);
 err_module:
@@ -723,6 +738,7 @@ dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
 {
        struct dma_buf_attachment *attach;
        int ret;
+       unsigned int attach_uid;
 
        if (WARN_ON(!dmabuf || !dev))
                return ERR_PTR(-EINVAL);
@@ -748,8 +764,13 @@ dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
        }
        dma_resv_lock(dmabuf->resv, NULL);
        list_add(&attach->node, &dmabuf->attachments);
+       attach_uid = dma_buf_update_attach_uid(dmabuf);
        dma_resv_unlock(dmabuf->resv);
 
+       ret = dma_buf_attach_stats_setup(attach, attach_uid);
+       if (ret)
+               goto err_sysfs;
+
        /* When either the importer or the exporter can't handle dynamic
         * mappings we cache the mapping here to avoid issues with the
         * reservation object lock.
@@ -776,6 +797,7 @@ dma_buf_dynamic_attach(struct dma_buf *dmabuf, struct device *dev,
                        dma_resv_unlock(attach->dmabuf->resv);
                attach->sgt = sgt;
                attach->dir = DMA_BIDIRECTIONAL;
+               dma_buf_update_attachment_map_count(attach, 1 /* delta */);
        }
 
        return attach;
@@ -792,6 +814,7 @@ err_unlock:
        if (dma_buf_is_dynamic(attach->dmabuf))
                dma_resv_unlock(attach->dmabuf->resv);
 
+err_sysfs:
        dma_buf_detach(dmabuf, attach);
        return ERR_PTR(ret);
 }
@@ -841,6 +864,7 @@ void dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach)
                        dma_resv_lock(attach->dmabuf->resv, NULL);
 
                __unmap_dma_buf(attach, attach->sgt, attach->dir);
+               dma_buf_update_attachment_map_count(attach, -1 /* delta */);
 
                if (dma_buf_is_dynamic(attach->dmabuf)) {
                        dmabuf->ops->unpin(attach);
@@ -854,6 +878,7 @@ void dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach)
        if (dmabuf->ops->detach)
                dmabuf->ops->detach(dmabuf, attach);
 
+       dma_buf_attach_stats_teardown(attach);
        kfree(attach);
 }
 EXPORT_SYMBOL_GPL(dma_buf_detach);
@@ -993,6 +1018,9 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach,
        }
 #endif /* CONFIG_DMA_API_DEBUG */
 
+       if (!IS_ERR(sg_table))
+               dma_buf_update_attachment_map_count(attach, 1 /* delta */);
+
        return sg_table;
 }
 EXPORT_SYMBOL_GPL(dma_buf_map_attachment);
@@ -1030,6 +1058,8 @@ void dma_buf_unmap_attachment(struct dma_buf_attachment *attach,
        if (dma_buf_is_dynamic(attach->dmabuf) &&
            !IS_ENABLED(CONFIG_DMABUF_MOVE_NOTIFY))
                dma_buf_unpin(attach);
+
+       dma_buf_update_attachment_map_count(attach, -1 /* delta */);
 }
 EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment);
 
@@ -1469,6 +1499,12 @@ static inline void dma_buf_uninit_debugfs(void)
 
 static int __init dma_buf_init(void)
 {
+       int ret;
+
+       ret = dma_buf_init_sysfs_statistics();
+       if (ret)
+               return ret;
+
        dma_buf_mnt = kern_mount(&dma_buf_fs_type);
        if (IS_ERR(dma_buf_mnt))
                return PTR_ERR(dma_buf_mnt);
@@ -1484,5 +1520,6 @@ static void __exit dma_buf_deinit(void)
 {
        dma_buf_uninit_debugfs();
        kern_unmount(dma_buf_mnt);
+       dma_buf_uninit_sysfs_statistics();
 }
 __exitcall(dma_buf_deinit);
index efdc56b..342585b 100644 (file)
@@ -295,6 +295,9 @@ struct dma_buf_ops {
  * @poll: for userspace poll support
  * @cb_excl: for userspace poll support
  * @cb_shared: for userspace poll support
+ * @sysfs_entry: for exposing information about this buffer in sysfs.
+ * The attachment_uid member of @sysfs_entry is protected by dma_resv lock
+ * and is incremented on each attach.
  *
  * This represents a shared buffer, created by calling dma_buf_export(). The
  * userspace representation is a normal file descriptor, which can be created by
@@ -330,6 +333,15 @@ struct dma_buf {
 
                __poll_t active;
        } cb_excl, cb_shared;
+#ifdef CONFIG_DMABUF_SYSFS_STATS
+       /* for sysfs stats */
+       struct dma_buf_sysfs_entry {
+               struct kobject kobj;
+               struct dma_buf *dmabuf;
+               unsigned int attachment_uid;
+               struct kset *attach_stats_kset;
+       } *sysfs_entry;
+#endif
 };
 
 /**
@@ -379,6 +391,7 @@ struct dma_buf_attach_ops {
  * @importer_ops: importer operations for this attachment, if provided
  * dma_buf_map/unmap_attachment() must be called with the dma_resv lock held.
  * @importer_priv: importer specific attachment data.
+ * @sysfs_entry: For exposing information about this attachment in sysfs.
  *
  * This structure holds the attachment information between the dma_buf buffer
  * and its user device(s). The list contains one attachment struct per device
@@ -399,6 +412,13 @@ struct dma_buf_attachment {
        const struct dma_buf_attach_ops *importer_ops;
        void *importer_priv;
        void *priv;
+#ifdef CONFIG_DMABUF_SYSFS_STATS
+       /* for sysfs stats */
+       struct dma_buf_attach_sysfs_entry {
+               struct kobject kobj;
+               unsigned int map_counter;
+       } *sysfs_entry;
+#endif
 };
 
 /**