uio_hv_generic: Fix ring buffer sysfs creation path

Hi,
This patch series aims to address the sysfs creation issue for the ring buffer by reorganizing the code. Additionally, it updates the ring sysfs size to accurately reflect the actual ring buffer size, rather than a fixed static value.

PFB change logs:

Changes since v4:
https://lore.kernel.org/all/[email protected]/
* Added Reviewed-By and Tested-By tags from Michael.
* Fixed syntax by removing extra space (addressed Greg's comment)
* Rebased to latest linux-next tip

Changes since v3:
https://lore.kernel.org/all/[email protected]/
* Addressed Michael's comments regarding handling of return value of sysfs_update_group in uio_hv_generic.

Changes since v2:
https://lore.kernel.org/all/[email protected]/
Addressed Greg's comments:
* Split the original patch into two.
* Updated the commit message to explain the problem scenario.
* Added comments for new APIs in the kerneldoc format.
* Highlighted potential race conditions and explained why sysfs should not be created in the driver probe.

* Made minor changes to how the sysfs_update_group return value is handled.

Changes since v1:
https://lore.kernel.org/all/[email protected]/
* Fixed race condition in setting channel->mmap_ring_buffer by
introducing a new variable for visibility of sysfs (addressed Greg's comments)
* Used binary attribute fields instead of regular ones for initializing attribute_group.
* Make size of ring sysfs dynamic based on actual ring buffer's size.
* Preferred to keep mmap function in uio_hv_generic to give more control over ring's mmap functionality, since this is specific to uio_hv_generic driver.
* Remove spurious warning during sysfs creation in uio_hv_generic probe.
* Added comments in a couple of places.

Changes since RFC patch:
https://lore.kernel.org/all/[email protected]/
* Different approach to solve the problem is proposed (credits to Michael Kelley).
* Core logic for sysfs creation moved out of uio_hv_generic, to VMBus drivers where rest of the sysfs attributes for a VMBus channel are defined. (addressed Greg's comments)
* Used attribute groups instead of sysfs_create* functions, and bundled ring attribute with other attributes for the channel sysfs.

Error logs:

[   35.574120] ------------[ cut here ]------------
[   35.574122] WARNING: CPU: 0 PID: 10 at fs/sysfs/file.c:591 sysfs_create_bin_file+0x81/0x90
[   35.574168] Workqueue: hv_pri_chan vmbus_add_channel_work
[   35.574172] RIP: 0010:sysfs_create_bin_file+0x81/0x90
[   35.574197] Call Trace:
[   35.574199]  <TASK>
[   35.574200]  ? show_regs+0x69/0x80
[   35.574217]  ? __warn+0x8d/0x130
[   35.574220]  ? sysfs_create_bin_file+0x81/0x90
[   35.574222]  ? report_bug+0x182/0x190
[   35.574225]  ? handle_bug+0x5b/0x90
[   35.574244]  ? exc_invalid_op+0x19/0x70
[   35.574247]  ? asm_exc_invalid_op+0x1b/0x20
[   35.574252]  ? sysfs_create_bin_file+0x81/0x90
[   35.574255]  hv_uio_probe+0x1e7/0x410 [uio_hv_generic]
[   35.574271]  vmbus_probe+0x3b/0x90
[   35.574275]  really_probe+0xf4/0x3b0
[   35.574279]  __driver_probe_device+0x8a/0x170
[   35.574282]  driver_probe_device+0x23/0xc0
[   35.574285]  __device_attach_driver+0xb5/0x140
[   35.574288]  ? __pfx___device_attach_driver+0x10/0x10
[   35.574291]  bus_for_each_drv+0x86/0xe0
[   35.574294]  __device_attach+0xc1/0x200
[   35.574297]  device_initial_probe+0x13/0x20
[   35.574315]  bus_probe_device+0x99/0xa0
[   35.574318]  device_add+0x647/0x870
[   35.574320]  ? hrtimer_init+0x28/0x70
[   35.574323]  device_register+0x1b/0x30
[   35.574326]  vmbus_device_register+0x83/0x130
[   35.574328]  vmbus_add_channel_work+0x135/0x1a0
[   35.574331]  process_one_work+0x177/0x340
[   35.574348]  worker_thread+0x2b2/0x3c0
[   35.574350]  kthread+0xe3/0x1f0
[   35.574353]  ? __pfx_worker_thread+0x10/0x10
[   35.574356]  ? __pfx_kthread+0x10/0x10

Naman Jain (2):
uio_hv_generic: Fix sysfs creation path for ring buffer
Drivers: hv: Make the sysfs node size for the ring buffer dynamic

 drivers/hv/hyperv_vmbus.h    |   6 ++
 drivers/hv/vmbus_drv.c       | 119 ++++++++++++++++++++++++++++++++++-
 drivers/uio/uio_hv_generic.c |  39 +++++-------
 include/linux/hyperv.h       |   6 ++
 4 files changed, 147 insertions(+), 23 deletions(-)

base-commit: 5b37f7bfff3b1582c34be8fb23968b226db71ebd

2.34.1

On regular bootup, devices get registered to VMBus first, so when uio_hv_generic driver for a particular device type is probed, the device is already initialized and added, so sysfs creation in uio_hv_generic probe works fine. However, when device is removed and brought back, the channel rescinds and device again gets registered to VMBus. However this time, the uio_hv_generic driver is already registered to probe for that device and in this case sysfs creation is tried before the device's kobject gets initialized completely.

Fix this by moving the core logic of sysfs creation for ring buffer, from uio_hv_generic to HyperV's VMBus driver, where rest of the sysfs attributes for the channels are defined. While doing that, make use of attribute groups and macros, instead of creating sysfs directly, to ensure better error handling and code flow.

Problem path:
vmbus_process_offer (new offer comes for the VMBus device)
vmbus_add_channel_work

    vmbus_device_register
      |-> device_register
      |     |...
      |     |-> hv_uio_probe
      |           |...
      |           |-> sysfs_create_bin_file (leads to a warning as
      |                 primary channel's kobject, which is used to
      |                 create sysfs is not yet initialized)
      |-> kset_create_and_add
      |-> vmbus_add_channel_kobj (initialization of primary channel's
                                  kobject happens later)

Above code flow is sequential and the warning is always reproducible in this path.

---
 drivers/hv/hyperv_vmbus.h    |   6 ++
 drivers/hv/vmbus_drv.c       | 110 ++++++++++++++++++++++++++++++++++-
 drivers/uio/uio_hv_generic.c |  39 ++++++-------
 include/linux/hyperv.h       |   6 ++
 4 files changed, 138 insertions(+), 23 deletions(-)

index 29780f3a7478..0b450e53161e 100644
@@ -477,4 +477,10 @@ static inline int hv_debug_add_dev_dir(struct hv_device *dev)
 
 #endif /* CONFIG_HYPERV_TESTING */
 
+/* Create and remove sysfs entry for memory mapped ring buffers for a channel */
+int hv_create_ring_sysfs(struct vmbus_channel *channel,
+			 int (*hv_mmap_ring_buffer)(struct vmbus_channel *channel,
+						    struct vm_area_struct *vma));
+int hv_remove_ring_sysfs(struct vmbus_channel *channel);
+
 #endif /* _HYPERV_VMBUS_H */

index 1b19b5647495..69c1df0f4ca5 100644
@@ -131,15 +131,12 @@ static void hv_uio_rescind(struct vmbus_channel *channel)
 	vmbus_device_unregister(channel->device_obj);
 }
 
-/* Sysfs API to allow mmap of the ring buffers
+/* Function used for mmap of ring buffer sysfs interface.
  * The ring buffer is allocated as contiguous memory by vmbus_open
  */
-static int hv_uio_ring_mmap(struct file *filp, struct kobject *kobj,
-			    const struct bin_attribute *attr,
-			    struct vm_area_struct *vma)
+static int
+hv_uio_ring_mmap(struct vmbus_channel *channel, struct vm_area_struct *vma)
 {
-	struct vmbus_channel *channel
-		= container_of(kobj, struct vmbus_channel, kobj);
 	void *ring_buffer = page_address(channel->ringbuffer_page);
 
 	if (channel->state != CHANNEL_OPENED_STATE)

@@ -149,15 +146,6 @@ static int hv_uio_ring_mmap(struct file *filp, struct kobject *kobj,
 			       channel->ringbuffer_pagecount << PAGE_SHIFT);
 }
 
-static const struct bin_attribute ring_buffer_bin_attr = {
-	.attr = {
-		.name = "ring",
-		.mode = 0600,
-	},
-	.size = 2 * SZ_2M,
-	.mmap = hv_uio_ring_mmap,
-};
-
 /* Callback from VMBUS subsystem when new channel created. */
 static void
 hv_uio_new_channel(struct vmbus_channel *new_sc)

@@ -178,8 +166,7 @@ hv_uio_new_channel(struct vmbus_channel *new_sc)
 	/* Disable interrupts on sub channel */
 	new_sc->inbound.ring_buffer->interrupt_mask = 1;
 	set_channel_read_mode(new_sc, HV_CALL_ISR);
-
-	ret = sysfs_create_bin_file(&new_sc->kobj, &ring_buffer_bin_attr);
+	ret = hv_create_ring_sysfs(new_sc, hv_uio_ring_mmap);
 	if (ret) {
 		dev_err(device, "sysfs create ring bin file failed; %d\n", ret);
 		vmbus_close(new_sc);

@@ -350,10 +337,18 @@ hv_uio_probe(struct hv_device *dev,
 		goto fail_close;
 	}
 
-	ret = sysfs_create_bin_file(&channel->kobj, &ring_buffer_bin_attr);
-	if (ret)
-		dev_notice(&dev->device,
-			   "sysfs create ring bin file failed; %d\n", ret);
+	/*
+	 * This internally calls sysfs_update_group, which returns a non-zero value if it executes
+	 * before sysfs_create_group. This is expected as the 'ring' will be created later in
+	 * vmbus_device_register() -> vmbus_add_channel_kobj(). Thus, no need to check the return
+	 * value and print warning.
+	 *
+	 * Creating/exposing sysfs in driver probe is not encouraged as it can lead to race
+	 * conditions with userspace. For backward compatibility, "ring" sysfs could not be removed
+	 * or decoupled from uio_hv_generic probe. Userspace programs can make use of inotify
+	 * APIs to make sure that ring is created.
+	 */
+	hv_create_ring_sysfs(channel, hv_uio_ring_mmap);
 
 	hv_set_drvdata(dev, pdata);
 

@@ -375,7 +370,7 @@ hv_uio_remove(struct hv_device *dev)
 	if (!pdata)
 		return;
 
-	sysfs_remove_bin_file(&dev->channel->kobj, &ring_buffer_bin_attr);
+	hv_remove_ring_sysfs(dev->channel);
 	uio_unregister_device(&pdata->info);
 	hv_uio_cleanup(dev, pdata);
 

index 675959fb97ba..d6ffe01962c2 100644
@@ -1002,6 +1002,12 @@ struct vmbus_channel {
 
 	/* The max size of a packet on this channel */
 	u32 max_pkt_size;
+
+	/* function to mmap ring buffer memory to the channel's sysfs ring attribute */
+	int (*mmap_ring_buffer)(struct vmbus_channel *channel, struct vm_area_struct *vma);
+
+	/* boolean to control visibility of sysfs for ring buffer */
+	bool ring_sysfs_visible;
 };
 
 #define lock_requestor(channel, flags)					\

2.34.1

Sorry, I'd like to nitpick :-) I guess the maintainer(s) can fix these for you so v6 might not be necessary, if there is no comment from others.

s/uio_hv_generic probe/hv_uio_probe()/
s/device/the device/

s/rescinds and device/gets rescinded and the device/

s/for/of/

s/rest/the rest/

s/Problem/Problematic/

s/new/A new/

s/primary/the primary/

s/sysfs/the sysfs file, /

s/primary/the primary/

Thanks Dexuan. I'll make the required changes and send next patch.

Regards,
Naman

The ring buffer size varies across VMBus channels. The size of sysfs node for the ring buffer is currently hardcoded to 4 MB. Userspace clients either use fstat() or hardcode this size for doing mmap(). To address this, make the sysfs node size dynamic to reflect the actual ring buffer size for each channel. This will ensure that fstat() on ring sysfs node always return the correct size of ring buffer.

---
 drivers/hv/vmbus_drv.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

index cbca403be2ab..88240b021034 100644
@@ -1820,7 +1820,6 @@ static struct bin_attribute chan_attr_ring_buffer = {
 		.name = "ring",
 		.mode = 0600,
 	},
-	.size = 2 * SZ_2M,
 	.mmap = hv_mmap_ring_buffer_wrapper,
 };
 static struct attribute *vmbus_chan_attrs[] = {

@@ -1880,11 +1879,21 @@ static umode_t vmbus_chan_bin_attr_is_visible(struct kobject *kobj,
 	return attr->attr.mode;
 }
 
+static size_t vmbus_chan_bin_size(struct kobject *kobj,
+				  const struct bin_attribute *bin_attr, int a)
+{
+	const struct vmbus_channel *channel =
+		container_of(kobj, struct vmbus_channel, kobj);
+
+	return channel->ringbuffer_pagecount << PAGE_SHIFT;
+}
+
 static const struct attribute_group vmbus_chan_group = {
 	.attrs = vmbus_chan_attrs,
 	.bin_attrs = vmbus_chan_bin_attrs,
 	.is_visible = vmbus_chan_attr_is_visible,
 	.is_bin_visible = vmbus_chan_bin_attr_is_visible,
+	.bin_size = vmbus_chan_bin_size,
 };
 
 static const struct kobj_type vmbus_chan_ktype = {

2.34.1

s/return/returns/

Thanks. I'll make the required changes.

Regards,
Naman