On 12/1/2020 11:05 AM, Loic Poulain wrote:
On Tue, 1 Dec 2020 at 18:51, Jeffrey Hugo <jhugo@xxxxxxxxxxxxxx> wrote:
On 12/1/2020 10:52 AM, Loic Poulain wrote:
On Tue, 1 Dec 2020 at 18:37, Jeffrey Hugo <jhugo@xxxxxxxxxxxxxx> wrote:
On 12/1/2020 10:36 AM, Loic Poulain wrote:
On Tue, 1 Dec 2020 at 02:16, Hemant Kumar <hemantk@xxxxxxxxxxxxxx> wrote:
Hi Loic,
On 11/30/20 10:22 AM, Loic Poulain wrote:
On Sat, 28 Nov 2020 at 04:26, Hemant Kumar <hemantk@xxxxxxxxxxxxxx> wrote:
This MHI client driver allows userspace clients to transfer
raw data between MHI device and host using standard file operations.
Driver instantiates UCI device object which is associated to device
file node. UCI device object instantiates UCI channel object when device
file node is opened. UCI channel object is used to manage MHI channels
by calling MHI core APIs for read and write operations. MHI channels
are started as part of device open(). MHI channels remain in start
state until last release() is called on UCI device file node. Device
file node is created with format
[...]
+struct uci_chan {
+ struct uci_dev *udev;
+ wait_queue_head_t ul_wq;
+
+ /* ul channel lock to synchronize multiple writes */
+ struct mutex write_lock;
+
+ wait_queue_head_t dl_wq;
+
+ /* dl channel lock to synchronize multiple reads */
+ struct mutex read_lock;
+
+ /*
+ * protects pending list in bh context, channel release, read and
+ * poll
+ */
+ spinlock_t dl_pending_lock;
+
+ struct list_head dl_pending;
+ struct uci_buf *cur_buf;
+ size_t dl_size;
+ struct kref ref_count;
+};
[...]
+ * struct uci_dev - MHI UCI device
+ * @minor: UCI device node minor number
+ * @mhi_dev: associated mhi device object
+ * @uchan: UCI uplink and downlink channel object
+ * @mtu: max TRE buffer length
+ * @enabled: Flag to track the state of the UCI device
+ * @lock: mutex lock to manage uchan object
+ * @ref_count: uci_dev reference count
+ */
+struct uci_dev {
+ unsigned int minor;
+ struct mhi_device *mhi_dev;
+ struct uci_chan *uchan;
Why a pointer to uci_chan and not just plainly integrating the
structure here, AFAIU uci_chan describes the channels and is just a
subpart of uci_dev. That would reduce the number of dynamic
allocations you manage and the extra kref. do you even need a separate
structure for this?
This goes back to one of my patch versions i tried to address concern
from Greg. Since we need to ref count the channel as well as the uci
device i decoupled the two objects and used two reference counts for two
different objects.
What Greg complained about is the two kref in the same structure and
that you were using kref as an open() counter. But splitting your
struct in two in order to keep the two kref does not make the much
code better (and simpler). I'm still a bit puzzled about the driver
complexity, it's supposed to be just a passthrough interface to MHI
after all.
I would suggest several changes, that IMHO would simplify reviewing:
- Use only one structure representing the 'uci' context (uci_dev)
- Keep the read path simple (mhi_uci_read), do no use an intermediate
cur_buf pointer, only dequeue the buffer when it is fully consumed.
- As I commented before, take care of the dl_pending list access
concurrency, even in wait_event.
- You don't need to count the number of open() calls, AFAIK,
mhi_prepare_for_transfer() simply fails if channels are already
started...
Unless I missed something, you seem to have ignored the root issue that
Hemant needs to solve, which is when to call
mhi_unprepare_for_transfer(). You can't just call that when close() is
called because there might be multiple users, and each one is going to
trigger a close(), so you need to know how many close() instances to
expect, and only call mhi_unprepare_for_transfer() for the last one.
That one part of his problem, yes, but if you unconditionally call
mhi_prepare_for_transfer in open(), it should fail for subsequent
users, and so only one user will successfully open the device.
I'm pretty sure that falls under "trying to prevent users from opening a
device" which Greg gave a pretty strong NACK to. So, that's not a solution.
That would deserve clarification since other drivers like
virtio_console clearly prevent that (guest_connected).
Quoting Greg from the source - https://lkml.org/lkml/2020/9/17/873
"
I told you before, do not try to keep a device node from being opened
multiple times, as it will always fail (think about passing file handles
around between programs...)
If userspace wants to do this, it will do it. If your driver can't
handle that, that's fine, userspace will learn not to do that. But the
kernel can not prevent this from happening.
"
So, the complete problem is -
N users need to be able to use the device (and by proxy, the channel)
concurrently, but prepare needs to be called on the first user coming
into the picture, and unprepare needs to be called on the last user
exiting the picture.
Hemant has supported this usecase in every rev of this series I've
looked at, but apparently every solution he has proposed to handle this
has caused confusion.
Understood, but that can be done with a simple counter in that case.
Regards,
Loic
--
Jeffrey Hugo
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.
