2010-02-02 19:07:50

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 00/18] Increase maximum number of Infiniband HCAs per system

This is v2 of a patch series that increases the maximum number of
IB HCAs supported per system.

The original mail thread is here:
http://lkml.org/lkml/2010/1/29/346

One note, I decided to "copy/paste" since factoring out the overflow
code in the three drivers seemed like overkill. If so desired, I could
factor those three separate functions into something provided by the
core, but that seemed like more trouble than it was worth at the time.

As before, I still don't have access to a giant system, so what I did
to test was to stick 4 cards into a small system, and then modify the
limits with debug patches similar to this:

diff --git a/drivers/infiniband/core/ucm.c b/drivers/infiniband/core/ucm.c
index 7bf0a82..8581e64 100644
--- a/drivers/infiniband/core/ucm.c
+++ b/drivers/infiniband/core/ucm.c
@@ -102,7 +102,7 @@ struct ib_ucm_event {
enum {
IB_UCM_MAJOR = 231,
IB_UCM_BASE_MINOR = 224,
- IB_UCM_MAX_DEVICES = 32
+ IB_UCM_MAX_DEVICES = 2
};


I tested all 3 drivers this way (uverbs, umad, ucm). I verified that
we're not leaking device numbers on multiple modprobe/rmmod cycles,
that there aren't any funny interactions when various combinations of
the drivers are loaded.

I did not test the rest of the OFED stack. I did write some trivial
programs to open the devices in /dev and close them again.

Here's an example of some of the testing:

dl585g2:~ # modprobe ib_uverbs
dl585g2:~ # modprobe ib_umad
dl585g2:~ # modprobe ib_ucm
dl585g2:~ # ls -l /dev/uverb*
crw-rw---- 1 root root 231, 192 Feb 2 05:55 /dev/uverbs0
crw-rw---- 1 root root 231, 193 Feb 2 05:55 /dev/uverbs1
crw-rw---- 1 root root 249, 0 Feb 2 05:55 /dev/uverbs2
crw-rw---- 1 root root 249, 1 Feb 2 05:55 /dev/uverbs3
dl585g2:~ # ls -l /dev/umad*
crw-rw---- 1 root root 231, 0 Feb 2 05:55 /dev/umad0
crw-rw---- 1 root root 231, 1 Feb 2 05:55 /dev/umad1
crw-rw---- 1 root root 231, 2 Feb 2 05:55 /dev/umad2
crw-rw---- 1 root root 231, 3 Feb 2 05:55 /dev/umad3
crw-rw---- 1 root root 248, 0 Feb 2 05:55 /dev/umad4
crw-rw---- 1 root root 248, 1 Feb 2 05:55 /dev/umad5
crw-rw---- 1 root root 248, 2 Feb 2 05:55 /dev/umad6
crw-rw---- 1 root root 248, 3 Feb 2 05:55 /dev/umad7
dl585g2:~ # ls -l /dev/issm*
crw-rw---- 1 root root 231, 4 Feb 2 05:55 /dev/issm0
crw-rw---- 1 root root 231, 5 Feb 2 05:55 /dev/issm1
crw-rw---- 1 root root 231, 6 Feb 2 05:55 /dev/issm2
crw-rw---- 1 root root 231, 7 Feb 2 05:55 /dev/issm3
crw-rw---- 1 root root 248, 4 Feb 2 05:55 /dev/issm4
crw-rw---- 1 root root 248, 5 Feb 2 05:55 /dev/issm5
crw-rw---- 1 root root 248, 6 Feb 2 05:55 /dev/issm6
crw-rw---- 1 root root 248, 7 Feb 2 05:55 /dev/issm7
dl585g2:~ # ls -l /dev/ucm*
crw-rw---- 1 root root 231, 224 Feb 2 05:55 /dev/ucm0
crw-rw---- 1 root root 231, 225 Feb 2 05:55 /dev/ucm1
crw-rw---- 1 root root 247, 0 Feb 2 05:55 /dev/ucm2
crw-rw---- 1 root root 247, 1 Feb 2 05:55 /dev/ucm3

Note that the major and minor numbers are behaving rather sanely.

dl585g2:~ # rmmod ib_ucm
dl585g2:~ # rmmod ib_uverbs
dl585g2:~ # rmmod ib_umad

Reset.

dl585g2:~ # modprobe ib_ucm
dl585g2:~ # ls -l /dev/ucm*
crw-rw---- 1 root root 231, 224 Feb 2 05:57 /dev/ucm0
crw-rw---- 1 root root 231, 225 Feb 2 05:57 /dev/ucm1
crw-rw---- 1 root root 248, 0 Feb 2 05:57 /dev/ucm2
crw-rw---- 1 root root 248, 1 Feb 2 05:57 /dev/ucm3

See that /dev/ucm* devices now have a different major number
compared to last time(248 vs 247), since we loaded that driver first.

But wait, why is it 248 and not 249? Is there a leak somewhere?

dl585g2:~ # ls -l /dev/uverb*
crw-rw---- 1 root root 231, 192 Feb 2 05:57 /dev/uverbs0
crw-rw---- 1 root root 231, 193 Feb 2 05:57 /dev/uverbs1
crw-rw---- 1 root root 249, 0 Feb 2 05:57 /dev/uverbs2
crw-rw---- 1 root root 249, 1 Feb 2 05:57 /dev/uverbs3
dl585g2:~ # rmmod ib_uverbs
ERROR: Module ib_uverbs is in use by ib_ucm

Ah, ib_ucm is dependent on ib_uverbs, so when we modprobed ib_ucm,
in reality ib_uverbs got loaded first. See how it has a higher
major number.

dl585g2:~ # rmmod ib_ucm
dl585g2:~ # rmmod ib_uverbs
dl585g2:~ # modprobe ib_umad
dl585g2:~ # ls -l /dev/umad*
crw-rw---- 1 root root 231, 0 Feb 2 05:58 /dev/umad0
crw-rw---- 1 root root 231, 1 Feb 2 05:58 /dev/umad1
crw-rw---- 1 root root 231, 2 Feb 2 05:58 /dev/umad2
crw-rw---- 1 root root 231, 3 Feb 2 05:58 /dev/umad3
crw-rw---- 1 root root 249, 0 Feb 2 05:58 /dev/umad4
crw-rw---- 1 root root 249, 1 Feb 2 05:58 /dev/umad5
crw-rw---- 1 root root 249, 2 Feb 2 05:58 /dev/umad6
crw-rw---- 1 root root 249, 3 Feb 2 05:58 /dev/umad7
dl585g2:~ # ls -l /dev/issm*
crw-rw---- 1 root root 231, 4 Feb 2 05:58 /dev/issm0
crw-rw---- 1 root root 231, 5 Feb 2 05:58 /dev/issm1
crw-rw---- 1 root root 231, 6 Feb 2 05:58 /dev/issm2
crw-rw---- 1 root root 231, 7 Feb 2 05:58 /dev/issm3
crw-rw---- 1 root root 249, 4 Feb 2 05:58 /dev/issm4
crw-rw---- 1 root root 249, 5 Feb 2 05:58 /dev/issm5
crw-rw---- 1 root root 249, 6 Feb 2 05:58 /dev/issm6
crw-rw---- 1 root root 249, 7 Feb 2 05:58 /dev/issm7

Finally, after one more reset, we see ib_umad loaded first and
obtaining the major number of 249.

v1 -> v2:
- update umad and ucm drivers too

---

Alex Chiang (18):
IB/uverbs: convert *cdev to cdev in struct ib_uverbs_device
IB/uverbs: remove dev_table
IB/uverbs: use stack variable 'devnum' in ib_uverbs_add_one
IB/uverbs: use stack variable 'base' in ib_uverbs_add_one
IB/uverbs: increase maximum devices supported
IB/uverbs: pack struct ib_uverbs_event_file tighter
IB/uverbs: whitespace cleanup
IB/umad: convert cdev pointers to embedded structs in struct ib_umad_port
IB/umad: remove port_table[]
IB/umad: use stack variable 'devnum' in ib_umad_init_port
IB/umad: use stack variable 'base' in ib_umad_init_port
IB/umad: increase maximum devices supported
IB/umad: clean whitespace
IB/ucm: use stack variable 'devnum' in ib_ucm_add_one
IB/ucm: use stack variable 'base' in ib_ucm_add_one
IB/ucm: increase maximum devices supported
IB/ucm: clean whitespace errors
IB/core: pack struct ib_device a little tighter


drivers/infiniband/core/ucm.c | 63 ++++++++++--
drivers/infiniband/core/user_mad.c | 173 +++++++++++++++++----------------
drivers/infiniband/core/uverbs.h | 11 +-
drivers/infiniband/core/uverbs_main.c | 175 +++++++++++++++++++--------------
include/rdma/ib_verbs.h | 4 -
5 files changed, 255 insertions(+), 171 deletions(-)


2010-02-02 19:08:06

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 01/18] IB/uverbs: convert *cdev to cdev in struct ib_uverbs_device

Instead of storing a pointer to a cdev, embed the entire struct cdev.

This change allows us to use the container_of() macro in
ib_uverbs_open() in a future patch.

This change increases the size of struct ib_uverbs_device to
168 bytes across 3 cachelines from 80 bytes in 2 cachelines.
However, we rearrange the members so that everything fits into
the first cacheline except for the struct cdev. Finally, we don't
touch the cdev in any fastpaths, so this change shouldn't negatively
affect performance.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/uverbs.h | 7 ++++---
drivers/infiniband/core/uverbs_main.c | 23 ++++++++++-------------
2 files changed, 14 insertions(+), 16 deletions(-)

diff --git a/drivers/infiniband/core/uverbs.h b/drivers/infiniband/core/uverbs.h
index b3ea958..e695f65 100644
--- a/drivers/infiniband/core/uverbs.h
+++ b/drivers/infiniband/core/uverbs.h
@@ -41,6 +41,7 @@
#include <linux/idr.h>
#include <linux/mutex.h>
#include <linux/completion.h>
+#include <linux/cdev.h>

#include <rdma/ib_verbs.h>
#include <rdma/ib_umem.h>
@@ -69,12 +70,12 @@

struct ib_uverbs_device {
struct kref ref;
+ int num_comp_vectors;
struct completion comp;
- int devnum;
- struct cdev *cdev;
struct device *dev;
struct ib_device *ib_dev;
- int num_comp_vectors;
+ int devnum;
+ struct cdev cdev;
};

struct ib_uverbs_event_file {
diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
index 5f284ff..5da9a73 100644
--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -43,7 +43,6 @@
#include <linux/sched.h>
#include <linux/file.h>
#include <linux/mount.h>
-#include <linux/cdev.h>

#include <asm/uaccess.h>

@@ -761,17 +760,15 @@ static void ib_uverbs_add_one(struct ib_device *device)
uverbs_dev->ib_dev = device;
uverbs_dev->num_comp_vectors = device->num_comp_vectors;

- uverbs_dev->cdev = cdev_alloc();
- if (!uverbs_dev->cdev)
- goto err;
- uverbs_dev->cdev->owner = THIS_MODULE;
- uverbs_dev->cdev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
- kobject_set_name(&uverbs_dev->cdev->kobj, "uverbs%d", uverbs_dev->devnum);
- if (cdev_add(uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
+ cdev_init(&uverbs_dev->cdev, NULL);
+ uverbs_dev->cdev.owner = THIS_MODULE;
+ uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
+ kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
+ if (cdev_add(&uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
goto err_cdev;

uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
- uverbs_dev->cdev->dev, uverbs_dev,
+ uverbs_dev->cdev.dev, uverbs_dev,
"uverbs%d", uverbs_dev->devnum);
if (IS_ERR(uverbs_dev->dev))
goto err_cdev;
@@ -790,10 +787,10 @@ static void ib_uverbs_add_one(struct ib_device *device)
return;

err_class:
- device_destroy(uverbs_class, uverbs_dev->cdev->dev);
+ device_destroy(uverbs_class, uverbs_dev->cdev.dev);

err_cdev:
- cdev_del(uverbs_dev->cdev);
+ cdev_del(&uverbs_dev->cdev);
clear_bit(uverbs_dev->devnum, dev_map);

err:
@@ -811,8 +808,8 @@ static void ib_uverbs_remove_one(struct ib_device *device)
return;

dev_set_drvdata(uverbs_dev->dev, NULL);
- device_destroy(uverbs_class, uverbs_dev->cdev->dev);
- cdev_del(uverbs_dev->cdev);
+ device_destroy(uverbs_class, uverbs_dev->cdev.dev);
+ cdev_del(&uverbs_dev->cdev);

spin_lock(&map_lock);
dev_table[uverbs_dev->devnum] = NULL;

2010-02-02 19:08:17

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 02/18] IB/uverbs: remove dev_table

dev_table's raison d'etre was to associate an inode back to a
struct ib_uverbs_device.

However, now that we've converted ib_uverbs_device to contain
an embedded cdev (instead of a *cdev), we can use the container_of()
macro and cast back to the containing device.

There's no longer any need for dev_table, so get rid of it.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/uverbs_main.c | 24 +++++-------------------
1 files changed, 5 insertions(+), 19 deletions(-)

diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
index 5da9a73..3f11292 100644
--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -74,7 +74,6 @@ DEFINE_IDR(ib_uverbs_qp_idr);
DEFINE_IDR(ib_uverbs_srq_idr);

static DEFINE_SPINLOCK(map_lock);
-static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);

static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
@@ -616,14 +615,12 @@ static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
/*
* ib_uverbs_open() does not need the BKL:
*
- * - dev_table[] accesses are protected by map_lock, the
- * ib_uverbs_device structures are properly reference counted, and
+ * - the ib_uverbs_device structures are properly reference counted and
* everything else is purely local to the file being created, so
* races against other open calls are not a problem;
* - there is no ioctl method to race against;
- * - the device is added to dev_table[] as the last part of module
- * initialization, the open method will either immediately run
- * -ENXIO, or all required initialization will be done.
+ * - the open method will either immediately run -ENXIO, or all
+ * required initialization will be done.
*/
static int ib_uverbs_open(struct inode *inode, struct file *filp)
{
@@ -631,13 +628,10 @@ static int ib_uverbs_open(struct inode *inode, struct file *filp)
struct ib_uverbs_file *file;
int ret;

- spin_lock(&map_lock);
- dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
+ dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
if (dev)
kref_get(&dev->ref);
- spin_unlock(&map_lock);
-
- if (!dev)
+ else
return -ENXIO;

if (!try_module_get(dev->ib_dev->owner)) {
@@ -778,10 +772,6 @@ static void ib_uverbs_add_one(struct ib_device *device)
if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
goto err_class;

- spin_lock(&map_lock);
- dev_table[uverbs_dev->devnum] = uverbs_dev;
- spin_unlock(&map_lock);
-
ib_set_client_data(device, &uverbs_client, uverbs_dev);

return;
@@ -811,10 +801,6 @@ static void ib_uverbs_remove_one(struct ib_device *device)
device_destroy(uverbs_class, uverbs_dev->cdev.dev);
cdev_del(&uverbs_dev->cdev);

- spin_lock(&map_lock);
- dev_table[uverbs_dev->devnum] = NULL;
- spin_unlock(&map_lock);
-
clear_bit(uverbs_dev->devnum, dev_map);

kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);

2010-02-02 19:08:20

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 03/18] IB/uverbs: use stack variable 'devnum' in ib_uverbs_add_one

This change is not useful by itself, but it sets us up for a future
change that allows us to dynamically allocate device numbers in case
we have more than IB_UVERBS_MAX_DEVICES in the system.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/uverbs_main.c | 12 +++++++-----
1 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
index 3f11292..acae9ed 100644
--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -730,6 +730,7 @@ static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);

static void ib_uverbs_add_one(struct ib_device *device)
{
+ int devnum;
struct ib_uverbs_device *uverbs_dev;

if (!device->alloc_ucontext)
@@ -743,12 +744,13 @@ static void ib_uverbs_add_one(struct ib_device *device)
init_completion(&uverbs_dev->comp);

spin_lock(&map_lock);
- uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
- if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
+ devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
+ if (devnum >= IB_UVERBS_MAX_DEVICES) {
spin_unlock(&map_lock);
goto err;
}
- set_bit(uverbs_dev->devnum, dev_map);
+ uverbs_dev->devnum = devnum;
+ set_bit(devnum, dev_map);
spin_unlock(&map_lock);

uverbs_dev->ib_dev = device;
@@ -758,7 +760,7 @@ static void ib_uverbs_add_one(struct ib_device *device)
uverbs_dev->cdev.owner = THIS_MODULE;
uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
- if (cdev_add(&uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
+ if (cdev_add(&uverbs_dev->cdev, IB_UVERBS_BASE_DEV + devnum, 1))
goto err_cdev;

uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
@@ -781,7 +783,7 @@ err_class:

err_cdev:
cdev_del(&uverbs_dev->cdev);
- clear_bit(uverbs_dev->devnum, dev_map);
+ clear_bit(devnum, dev_map);

err:
kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);

2010-02-02 19:08:37

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 08/18] IB/umad: convert cdev pointers to embedded structs in struct ib_umad_port

Instead of storing pointers to cdev and sm_cdev, embed the full
structures instead.

This change allows us to use the container_of() macro in ib_umad_open()
and ib_umad_sm_open() in a future patch.

This change increases the size of struct ib_umad_port to 320 bytes
from 128.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/user_mad.c | 46 ++++++++++++++++--------------------
1 files changed, 20 insertions(+), 26 deletions(-)

diff --git a/drivers/infiniband/core/user_mad.c b/drivers/infiniband/core/user_mad.c
index 7de0296..40440ef 100644
--- a/drivers/infiniband/core/user_mad.c
+++ b/drivers/infiniband/core/user_mad.c
@@ -87,10 +87,10 @@ enum {
*/

struct ib_umad_port {
- struct cdev *cdev;
+ struct cdev cdev;
struct device *dev;

- struct cdev *sm_cdev;
+ struct cdev sm_cdev;
struct device *sm_dev;
struct semaphore sm_sem;

@@ -1008,17 +1008,14 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
mutex_init(&port->file_mutex);
INIT_LIST_HEAD(&port->file_list);

- port->cdev = cdev_alloc();
- if (!port->cdev)
- return -1;
- port->cdev->owner = THIS_MODULE;
- port->cdev->ops = &umad_fops;
- kobject_set_name(&port->cdev->kobj, "umad%d", port->dev_num);
- if (cdev_add(port->cdev, base_dev + port->dev_num, 1))
+ cdev_init(&port->cdev, &umad_fops);
+ port->cdev.owner = THIS_MODULE;
+ kobject_set_name(&port->cdev.kobj, "umad%d", port->dev_num);
+ if (cdev_add(&port->cdev, base_dev + port->dev_num, 1))
goto err_cdev;

port->dev = device_create(umad_class, device->dma_device,
- port->cdev->dev, port,
+ port->cdev.dev, port,
"umad%d", port->dev_num);
if (IS_ERR(port->dev))
goto err_cdev;
@@ -1028,17 +1025,14 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
if (device_create_file(port->dev, &dev_attr_port))
goto err_dev;

- port->sm_cdev = cdev_alloc();
- if (!port->sm_cdev)
- goto err_dev;
- port->sm_cdev->owner = THIS_MODULE;
- port->sm_cdev->ops = &umad_sm_fops;
- kobject_set_name(&port->sm_cdev->kobj, "issm%d", port->dev_num);
- if (cdev_add(port->sm_cdev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
+ cdev_init(&port->sm_cdev, &umad_sm_fops);
+ port->sm_cdev.owner = THIS_MODULE;
+ kobject_set_name(&port->sm_cdev.kobj, "issm%d", port->dev_num);
+ if (cdev_add(&port->sm_cdev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
goto err_sm_cdev;

port->sm_dev = device_create(umad_class, device->dma_device,
- port->sm_cdev->dev, port,
+ port->sm_cdev.dev, port,
"issm%d", port->dev_num);
if (IS_ERR(port->sm_dev))
goto err_sm_cdev;
@@ -1055,16 +1049,16 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
return 0;

err_sm_dev:
- device_destroy(umad_class, port->sm_cdev->dev);
+ device_destroy(umad_class, port->sm_cdev.dev);

err_sm_cdev:
- cdev_del(port->sm_cdev);
+ cdev_del(&port->sm_cdev);

err_dev:
- device_destroy(umad_class, port->cdev->dev);
+ device_destroy(umad_class, port->cdev.dev);

err_cdev:
- cdev_del(port->cdev);
+ cdev_del(&port->cdev);
clear_bit(port->dev_num, dev_map);

return -1;
@@ -1079,11 +1073,11 @@ static void ib_umad_kill_port(struct ib_umad_port *port)
dev_set_drvdata(port->dev, NULL);
dev_set_drvdata(port->sm_dev, NULL);

- device_destroy(umad_class, port->cdev->dev);
- device_destroy(umad_class, port->sm_cdev->dev);
+ device_destroy(umad_class, port->cdev.dev);
+ device_destroy(umad_class, port->sm_cdev.dev);

- cdev_del(port->cdev);
- cdev_del(port->sm_cdev);
+ cdev_del(&port->cdev);
+ cdev_del(&port->sm_cdev);

spin_lock(&port_lock);
umad_port[port->dev_num] = NULL;

2010-02-02 19:08:54

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 12/18] IB/umad: increase maximum devices supported

Some large systems may support more than IB_UMAD_MAX_PORTS
(currently 64).

This change allows us to support more ports in a backwards-compatible
manner. The first IB_UMAD_MAX_PORTS keep the same major/minor device
numbers they've always had.

If there are more than IB_UMAD_MAX_PORTS, we then dynamically request
a new major device number (new minors start at 0).

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/user_mad.c | 50 ++++++++++++++++++++++++++++++++----
1 files changed, 44 insertions(+), 6 deletions(-)

diff --git a/drivers/infiniband/core/user_mad.c b/drivers/infiniband/core/user_mad.c
index 5a66bd0..7fc1e51 100644
--- a/drivers/infiniband/core/user_mad.c
+++ b/drivers/infiniband/core/user_mad.c
@@ -971,6 +971,28 @@ static ssize_t show_abi_version(struct class *class, char *buf)
}
static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);

+static dev_t oflo_maj;
+static DECLARE_BITMAP(oflo_map, IB_UMAD_MAX_PORTS);
+static int find_oflo_devnum(void)
+{
+ int ret;
+
+ if (!oflo_maj) {
+ ret = alloc_chrdev_region(&oflo_maj, 0, IB_UMAD_MAX_PORTS * 2,
+ "infiniband_mad");
+ if (ret) {
+ printk(KERN_ERR "user_mad: couldn't register dynamic device number\n");
+ return ret;
+ }
+ }
+
+ ret = find_first_zero_bit(oflo_map, IB_UMAD_MAX_PORTS);
+ if (ret >= IB_UMAD_MAX_PORTS)
+ return -1;
+
+ return ret;
+}
+
static int ib_umad_init_port(struct ib_device *device, int port_num,
struct ib_umad_port *port)
{
@@ -981,11 +1003,19 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
devnum = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
if (devnum >= IB_UMAD_MAX_PORTS) {
spin_unlock(&port_lock);
- return -1;
+ devnum = find_oflo_devnum();
+ if (devnum < 0)
+ return -1;
+
+ spin_lock(&port_lock);
+ port->dev_num = devnum + IB_UMAD_MAX_PORTS;
+ base = devnum + oflo_maj;
+ set_bit(devnum, oflo_map);
+ } else {
+ port->dev_num = devnum;
+ base = devnum + base_dev;
+ set_bit(devnum, dev_map);
}
- port->dev_num = devnum;
- base = devnum + base_dev;
- set_bit(devnum, dev_map);
spin_unlock(&port_lock);

port->ib_dev = device;
@@ -1042,7 +1072,10 @@ err_dev:

err_cdev:
cdev_del(&port->cdev);
- clear_bit(devnum, dev_map);
+ if (port->dev_num < IB_UMAD_MAX_PORTS)
+ clear_bit(devnum, dev_map);
+ else
+ clear_bit(devnum, oflo_map);

return -1;
}
@@ -1079,7 +1112,10 @@ static void ib_umad_kill_port(struct ib_umad_port *port)

mutex_unlock(&port->file_mutex);

- clear_bit(port->dev_num, dev_map);
+ if (port->dev_num < IB_UMAD_MAX_PORTS)
+ clear_bit(port->dev_num, dev_map);
+ else
+ clear_bit(port->dev_num - IB_UMAD_MAX_PORTS, oflo_map);
}

static void ib_umad_add_one(struct ib_device *device)
@@ -1187,6 +1223,8 @@ static void __exit ib_umad_cleanup(void)
ib_unregister_client(&umad_client);
class_destroy(umad_class);
unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
+ if (oflo_maj)
+ unregister_chrdev_region(oflo_maj, IB_UMAD_MAX_PORTS * 2);
}

module_init(ib_umad_init);

2010-02-02 19:09:09

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 13/18] IB/umad: clean whitespace

Clean errors as shown when 'let c_space_errors=1' is set in vim.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/user_mad.c | 26 +++++++++++++-------------
1 files changed, 13 insertions(+), 13 deletions(-)

diff --git a/drivers/infiniband/core/user_mad.c b/drivers/infiniband/core/user_mad.c
index 7fc1e51..3595bd5 100644
--- a/drivers/infiniband/core/user_mad.c
+++ b/drivers/infiniband/core/user_mad.c
@@ -487,8 +487,8 @@ static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
ah_attr.ah_flags = IB_AH_GRH;
memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
ah_attr.grh.sgid_index = packet->mad.hdr.gid_index;
- ah_attr.grh.flow_label = be32_to_cpu(packet->mad.hdr.flow_label);
- ah_attr.grh.hop_limit = packet->mad.hdr.hop_limit;
+ ah_attr.grh.flow_label = be32_to_cpu(packet->mad.hdr.flow_label);
+ ah_attr.grh.hop_limit = packet->mad.hdr.hop_limit;
ah_attr.grh.traffic_class = packet->mad.hdr.traffic_class;
}

@@ -519,9 +519,9 @@ static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
goto err_ah;
}

- packet->msg->ah = ah;
+ packet->msg->ah = ah;
packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
- packet->msg->retries = packet->mad.hdr.retries;
+ packet->msg->retries = packet->mad.hdr.retries;
packet->msg->context[0] = packet;

/* Copy MAD header. Any RMPP header is already in place. */
@@ -856,16 +856,16 @@ static int ib_umad_close(struct inode *inode, struct file *filp)
}

static const struct file_operations umad_fops = {
- .owner = THIS_MODULE,
- .read = ib_umad_read,
- .write = ib_umad_write,
- .poll = ib_umad_poll,
+ .owner = THIS_MODULE,
+ .read = ib_umad_read,
+ .write = ib_umad_write,
+ .poll = ib_umad_poll,
.unlocked_ioctl = ib_umad_ioctl,
#ifdef CONFIG_COMPAT
- .compat_ioctl = ib_umad_compat_ioctl,
+ .compat_ioctl = ib_umad_compat_ioctl,
#endif
- .open = ib_umad_open,
- .release = ib_umad_close
+ .open = ib_umad_open,
+ .release = ib_umad_close
};

static int ib_umad_sm_open(struct inode *inode, struct file *filp)
@@ -930,8 +930,8 @@ static int ib_umad_sm_close(struct inode *inode, struct file *filp)
}

static const struct file_operations umad_sm_fops = {
- .owner = THIS_MODULE,
- .open = ib_umad_sm_open,
+ .owner = THIS_MODULE,
+ .open = ib_umad_sm_open,
.release = ib_umad_sm_close
};

2010-02-02 19:09:17

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 15/18] IB/ucm: use stack variable 'base' in ib_ucm_add_one

This change is not useful by itself, but sets us up for a future change
that allows us to support more than IB_UCM_MAX_DEVICES.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/ucm.c | 4 +++-
1 files changed, 3 insertions(+), 1 deletions(-)

diff --git a/drivers/infiniband/core/ucm.c b/drivers/infiniband/core/ucm.c
index 7ff3300..06c50d8 100644
--- a/drivers/infiniband/core/ucm.c
+++ b/drivers/infiniband/core/ucm.c
@@ -1240,6 +1240,7 @@ static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
static void ib_ucm_add_one(struct ib_device *device)
{
int devnum;
+ dev_t base;
struct ib_ucm_device *ucm_dev;

if (!device->alloc_ucontext ||
@@ -1257,12 +1258,13 @@ static void ib_ucm_add_one(struct ib_device *device)
goto err;

ucm_dev->devnum = devnum;
+ base = devnum + IB_UCM_BASE_DEV;
set_bit(devnum, dev_map);

cdev_init(&ucm_dev->cdev, &ucm_fops);
ucm_dev->cdev.owner = THIS_MODULE;
kobject_set_name(&ucm_dev->cdev.kobj, "ucm%d", ucm_dev->devnum);
- if (cdev_add(&ucm_dev->cdev, IB_UCM_BASE_DEV + devnum, 1))
+ if (cdev_add(&ucm_dev->cdev, base, 1))
goto err;

ucm_dev->dev.class = &cm_class;

2010-02-02 19:09:30

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 18/18] IB/core: pack struct ib_device a little tighter

A small change to reduce the size of ib_device to 1112 bytes
(from 1128).

Signed-off-by: Alex Chiang <[email protected]>
---

include/rdma/ib_verbs.h | 4 ++--
1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h
index 09509ed..a585e0f 100644
--- a/include/rdma/ib_verbs.h
+++ b/include/rdma/ib_verbs.h
@@ -984,9 +984,9 @@ struct ib_device {
struct list_head event_handler_list;
spinlock_t event_handler_lock;

+ spinlock_t client_data_lock;
struct list_head core_list;
struct list_head client_data_list;
- spinlock_t client_data_lock;

struct ib_cache cache;
int *pkey_tbl_len;
@@ -1144,8 +1144,8 @@ struct ib_device {
IB_DEV_UNREGISTERED
} reg_state;

- u64 uverbs_cmd_mask;
int uverbs_abi_ver;
+ u64 uverbs_cmd_mask;

char node_desc[64];
__be64 node_guid;

2010-02-02 19:09:22

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 17/18] IB/ucm: clean whitespace errors

As shown when 'let c_space_errors=1' is set in vim.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/ucm.c | 6 +++---
1 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/infiniband/core/ucm.c b/drivers/infiniband/core/ucm.c
index 1203ca4..7bf0a82 100644
--- a/drivers/infiniband/core/ucm.c
+++ b/drivers/infiniband/core/ucm.c
@@ -1223,10 +1223,10 @@ static void ib_ucm_release_dev(struct device *dev)
}

static const struct file_operations ucm_fops = {
- .owner = THIS_MODULE,
- .open = ib_ucm_open,
+ .owner = THIS_MODULE,
+ .open = ib_ucm_open,
.release = ib_ucm_close,
- .write = ib_ucm_write,
+ .write = ib_ucm_write,
.poll = ib_ucm_poll,
};

2010-02-02 19:08:27

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 05/18] IB/uverbs: increase maximum devices supported

Some large systems may support more than IB_UVERBS_MAX_DEVICES
(currently 32).

This change allows us to support more devices in a backwards-compatible
manner. The first IB_UVERBS_MAX_DEVICES keep the same major/minor
device numbers that they've always had.

If there are more than IB_UVERBS_MAX_DEVICES, we then dynamically
request a new major device number (new minors start at 0).

This change increases the maximum number of HCAs to 64 (from 32).

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/uverbs_main.c | 56 +++++++++++++++++++++++++++++----
1 files changed, 50 insertions(+), 6 deletions(-)

diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
index 8c594cd..38b9085 100644
--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -728,6 +728,34 @@ static ssize_t show_abi_version(struct class *class, char *buf)
}
static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);

+static dev_t oflo_maj;
+static DECLARE_BITMAP(oflo_map, IB_UVERBS_MAX_DEVICES);
+
+/*
+ * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
+ * requesting a new major number and doubling the number of max devices we
+ * support. It's stupid, but simple.
+ */
+static int find_oflo_devnum(void)
+{
+ int ret;
+
+ if (!oflo_maj) {
+ ret = alloc_chrdev_region(&oflo_maj, 0, IB_UVERBS_MAX_DEVICES,
+ "infiniband_verbs");
+ if (ret) {
+ printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
+ return ret;
+ }
+ }
+
+ ret = find_first_zero_bit(oflo_map, IB_UVERBS_MAX_DEVICES);
+ if (ret >= IB_UVERBS_MAX_DEVICES)
+ return -1;
+
+ return ret;
+}
+
static void ib_uverbs_add_one(struct ib_device *device)
{
int devnum;
@@ -748,11 +776,19 @@ static void ib_uverbs_add_one(struct ib_device *device)
devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
if (devnum >= IB_UVERBS_MAX_DEVICES) {
spin_unlock(&map_lock);
- goto err;
+ devnum = find_oflo_devnum();
+ if (devnum < 0)
+ goto err;
+
+ spin_lock(&map_lock);
+ uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
+ base = devnum + oflo_maj;
+ set_bit(devnum, oflo_map);
+ } else {
+ uverbs_dev->devnum = devnum;
+ base = devnum + IB_UVERBS_BASE_DEV;
+ set_bit(devnum, dev_map);
}
- uverbs_dev->devnum = devnum;
- base = devnum + IB_UVERBS_BASE_DEV;
- set_bit(devnum, dev_map);
spin_unlock(&map_lock);

uverbs_dev->ib_dev = device;
@@ -785,7 +821,10 @@ err_class:

err_cdev:
cdev_del(&uverbs_dev->cdev);
- clear_bit(devnum, dev_map);
+ if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
+ clear_bit(devnum, dev_map);
+ else
+ clear_bit(devnum, oflo_map);

err:
kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
@@ -805,7 +844,10 @@ static void ib_uverbs_remove_one(struct ib_device *device)
device_destroy(uverbs_class, uverbs_dev->cdev.dev);
cdev_del(&uverbs_dev->cdev);

- clear_bit(uverbs_dev->devnum, dev_map);
+ if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
+ clear_bit(uverbs_dev->devnum, dev_map);
+ else
+ clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, oflo_map);

kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
wait_for_completion(&uverbs_dev->comp);
@@ -895,6 +937,8 @@ static void __exit ib_uverbs_cleanup(void)
unregister_filesystem(&uverbs_event_fs);
class_destroy(uverbs_class);
unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
+ if (oflo_maj)
+ unregister_chrdev_region(oflo_maj, IB_UVERBS_MAX_DEVICES);
idr_destroy(&ib_uverbs_pd_idr);
idr_destroy(&ib_uverbs_mr_idr);
idr_destroy(&ib_uverbs_mw_idr);

2010-02-02 19:09:05

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 14/18] IB/ucm: use stack variable 'devnum' in ib_ucm_add_one

This change is not useful by itself, but sets us up for a future
change that allows us to dynamically allocate device numbers in case
we have more than IB_UCM_MAX_DEVICES in the system.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/ucm.c | 12 +++++++-----
1 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/drivers/infiniband/core/ucm.c b/drivers/infiniband/core/ucm.c
index f504c9b..7ff3300 100644
--- a/drivers/infiniband/core/ucm.c
+++ b/drivers/infiniband/core/ucm.c
@@ -1239,6 +1239,7 @@ static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);

static void ib_ucm_add_one(struct ib_device *device)
{
+ int devnum;
struct ib_ucm_device *ucm_dev;

if (!device->alloc_ucontext ||
@@ -1251,16 +1252,17 @@ static void ib_ucm_add_one(struct ib_device *device)

ucm_dev->ib_dev = device;

- ucm_dev->devnum = find_first_zero_bit(dev_map, IB_UCM_MAX_DEVICES);
- if (ucm_dev->devnum >= IB_UCM_MAX_DEVICES)
+ devnum = find_first_zero_bit(dev_map, IB_UCM_MAX_DEVICES);
+ if (devnum >= IB_UCM_MAX_DEVICES)
goto err;

- set_bit(ucm_dev->devnum, dev_map);
+ ucm_dev->devnum = devnum;
+ set_bit(devnum, dev_map);

cdev_init(&ucm_dev->cdev, &ucm_fops);
ucm_dev->cdev.owner = THIS_MODULE;
kobject_set_name(&ucm_dev->cdev.kobj, "ucm%d", ucm_dev->devnum);
- if (cdev_add(&ucm_dev->cdev, IB_UCM_BASE_DEV + ucm_dev->devnum, 1))
+ if (cdev_add(&ucm_dev->cdev, IB_UCM_BASE_DEV + devnum, 1))
goto err;

ucm_dev->dev.class = &cm_class;
@@ -1281,7 +1283,7 @@ err_dev:
device_unregister(&ucm_dev->dev);
err_cdev:
cdev_del(&ucm_dev->cdev);
- clear_bit(ucm_dev->devnum, dev_map);
+ clear_bit(devnum, dev_map);
err:
kfree(ucm_dev);
return;

2010-02-02 19:12:31

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 06/18] IB/uverbs: pack struct ib_uverbs_event_file tighter

Eliminate some padding in the structure by rearranging the members.

sizeof(struct ib_uverbs_event_file) is now 72 bytes (from 80) and
more members now fit in the first cacheline.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/uverbs.h | 4 ++--
1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/core/uverbs.h b/drivers/infiniband/core/uverbs.h
index e695f65..e54d9ac 100644
--- a/drivers/infiniband/core/uverbs.h
+++ b/drivers/infiniband/core/uverbs.h
@@ -80,13 +80,13 @@ struct ib_uverbs_device {

struct ib_uverbs_event_file {
struct kref ref;
+ int is_async;
struct ib_uverbs_file *uverbs_file;
spinlock_t lock;
+ int is_closed;
wait_queue_head_t poll_wait;
struct fasync_struct *async_queue;
struct list_head event_list;
- int is_async;
- int is_closed;
};

struct ib_uverbs_file {

2010-02-02 19:08:46

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 11/18] IB/umad: use stack variable 'base' in ib_umad_init_port

This change is not useful by itself, but sets us up for a future change
that allows us to support more than IB_UMAD_MAX_PORTS in a system.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/user_mad.c | 7 +++++--
1 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/core/user_mad.c b/drivers/infiniband/core/user_mad.c
index df403ec..5a66bd0 100644
--- a/drivers/infiniband/core/user_mad.c
+++ b/drivers/infiniband/core/user_mad.c
@@ -975,6 +975,7 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
struct ib_umad_port *port)
{
int devnum;
+ dev_t base;

spin_lock(&port_lock);
devnum = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
@@ -983,6 +984,7 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
return -1;
}
port->dev_num = devnum;
+ base = devnum + base_dev;
set_bit(devnum, dev_map);
spin_unlock(&port_lock);

@@ -995,7 +997,7 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
cdev_init(&port->cdev, &umad_fops);
port->cdev.owner = THIS_MODULE;
kobject_set_name(&port->cdev.kobj, "umad%d", port->dev_num);
- if (cdev_add(&port->cdev, base_dev + devnum, 1))
+ if (cdev_add(&port->cdev, base, 1))
goto err_cdev;

port->dev = device_create(umad_class, device->dma_device,
@@ -1009,10 +1011,11 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
if (device_create_file(port->dev, &dev_attr_port))
goto err_dev;

+ base += IB_UMAD_MAX_PORTS;
cdev_init(&port->sm_cdev, &umad_sm_fops);
port->sm_cdev.owner = THIS_MODULE;
kobject_set_name(&port->sm_cdev.kobj, "issm%d", port->dev_num);
- if (cdev_add(&port->sm_cdev, base_dev + devnum + IB_UMAD_MAX_PORTS, 1))
+ if (cdev_add(&port->sm_cdev, base, 1))
goto err_sm_cdev;

port->sm_dev = device_create(umad_class, device->dma_device,

2010-02-02 19:08:32

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 07/18] IB/uverbs: whitespace cleanup

Clean up the errors as shown when 'let c_space_errors=1' is set in vim.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/uverbs_main.c | 68 +++++++++++++++++----------------
1 files changed, 34 insertions(+), 34 deletions(-)

diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
index 38b9085..aff706e 100644
--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -79,34 +79,34 @@ static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
const char __user *buf, int in_len,
int out_len) = {
- [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
- [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
- [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
- [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
- [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
- [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
- [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
+ [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
+ [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
+ [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
+ [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
+ [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
+ [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
+ [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
[IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
- [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
- [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
- [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
- [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
- [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
- [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
- [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
- [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
- [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
- [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
- [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
- [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
- [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
- [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
- [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
- [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
- [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
- [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
- [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
- [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
+ [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
+ [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
+ [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
+ [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
+ [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
+ [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
+ [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
+ [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
+ [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
+ [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
+ [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
+ [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
+ [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
+ [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
+ [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
+ [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
+ [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
+ [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
+ [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
+ [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
};

static struct vfsmount *uverbs_event_mnt;
@@ -368,7 +368,7 @@ static int ib_uverbs_event_close(struct inode *inode, struct file *filp)

static const struct file_operations uverbs_event_fops = {
.owner = THIS_MODULE,
- .read = ib_uverbs_event_read,
+ .read = ib_uverbs_event_read,
.poll = ib_uverbs_event_poll,
.release = ib_uverbs_event_close,
.fasync = ib_uverbs_event_fasync
@@ -678,17 +678,17 @@ static int ib_uverbs_close(struct inode *inode, struct file *filp)
}

static const struct file_operations uverbs_fops = {
- .owner = THIS_MODULE,
- .write = ib_uverbs_write,
- .open = ib_uverbs_open,
+ .owner = THIS_MODULE,
+ .write = ib_uverbs_write,
+ .open = ib_uverbs_open,
.release = ib_uverbs_close
};

static const struct file_operations uverbs_mmap_fops = {
- .owner = THIS_MODULE,
- .write = ib_uverbs_write,
+ .owner = THIS_MODULE,
+ .write = ib_uverbs_write,
.mmap = ib_uverbs_mmap,
- .open = ib_uverbs_open,
+ .open = ib_uverbs_open,
.release = ib_uverbs_close
};

2010-02-02 19:10:33

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 16/18] IB/ucm: increase maximum devices supported

Some large systems may support more than IB_UCM_MAX_DEVICES
(currently 32).

This change allows us to support more devices in a backwards-compatible
manner. the first IB_UCM_MAX_DEVICES keep the same major/minor device
numbers they've always had.

If there are more than IB_UCM_MAX_DEVICES, then we dynamically request
a new major device number (new minors start at 0).

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/ucm.c | 53 +++++++++++++++++++++++++++++++++++------
1 files changed, 45 insertions(+), 8 deletions(-)

diff --git a/drivers/infiniband/core/ucm.c b/drivers/infiniband/core/ucm.c
index 06c50d8..1203ca4 100644
--- a/drivers/infiniband/core/ucm.c
+++ b/drivers/infiniband/core/ucm.c
@@ -1215,7 +1215,10 @@ static void ib_ucm_release_dev(struct device *dev)

ucm_dev = container_of(dev, struct ib_ucm_device, dev);
cdev_del(&ucm_dev->cdev);
- clear_bit(ucm_dev->devnum, dev_map);
+ if (ucm_dev->devnum < IB_UCM_MAX_DEVICES)
+ clear_bit(ucm_dev->devnum, dev_map);
+ else
+ clear_bit(ucm_dev->devnum - IB_UCM_MAX_DEVICES, dev_map);
kfree(ucm_dev);
}

@@ -1237,6 +1240,28 @@ static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
}
static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);

+static dev_t oflo_maj;
+static DECLARE_BITMAP(oflo_map, IB_UCM_MAX_DEVICES);
+static int find_oflo_devnum(void)
+{
+ int ret;
+
+ if (!oflo_maj) {
+ ret = alloc_chrdev_region(&oflo_maj, 0, IB_UCM_MAX_DEVICES,
+ "infiniband_cm");
+ if (ret) {
+ printk(KERN_ERR "ucm: couldn't register dynamic device number\n");
+ return ret;
+ }
+ }
+
+ ret = find_first_zero_bit(oflo_map, IB_UCM_MAX_DEVICES);
+ if (ret >= IB_UCM_MAX_DEVICES)
+ return -1;
+
+ return ret;
+}
+
static void ib_ucm_add_one(struct ib_device *device)
{
int devnum;
@@ -1254,12 +1279,19 @@ static void ib_ucm_add_one(struct ib_device *device)
ucm_dev->ib_dev = device;

devnum = find_first_zero_bit(dev_map, IB_UCM_MAX_DEVICES);
- if (devnum >= IB_UCM_MAX_DEVICES)
- goto err;
-
- ucm_dev->devnum = devnum;
- base = devnum + IB_UCM_BASE_DEV;
- set_bit(devnum, dev_map);
+ if (devnum >= IB_UCM_MAX_DEVICES) {
+ devnum = find_oflo_devnum();
+ if (devnum < 0)
+ goto err;
+
+ ucm_dev->devnum = devnum + IB_UCM_MAX_DEVICES;
+ base = devnum + oflo_maj;
+ set_bit(devnum, oflo_map);
+ } else {
+ ucm_dev->devnum = devnum;
+ base = devnum + IB_UCM_BASE_DEV;
+ set_bit(devnum, dev_map);
+ }

cdev_init(&ucm_dev->cdev, &ucm_fops);
ucm_dev->cdev.owner = THIS_MODULE;
@@ -1285,7 +1317,10 @@ err_dev:
device_unregister(&ucm_dev->dev);
err_cdev:
cdev_del(&ucm_dev->cdev);
- clear_bit(devnum, dev_map);
+ if (ucm_dev->devnum < IB_UCM_MAX_DEVICES)
+ clear_bit(devnum, dev_map);
+ else
+ clear_bit(devnum, oflo_map);
err:
kfree(ucm_dev);
return;
@@ -1344,6 +1379,8 @@ static void __exit ib_ucm_cleanup(void)
ib_unregister_client(&ucm_client);
class_remove_file(&cm_class, &class_attr_abi_version);
unregister_chrdev_region(IB_UCM_BASE_DEV, IB_UCM_MAX_DEVICES);
+ if (oflo_maj)
+ unregister_chrdev_region(oflo_maj, IB_UCM_MAX_DEVICES);
idr_destroy(&ctx_id_table);
}

2010-02-02 19:11:22

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 09/18] IB/umad: remove port_table[]

We no longer need this data structure, as it was used to associate
an inode back to a struct ib_umad_port during ->open().

But now that we're embedding a struct cdev in struct ib_umad_port,
we can use the container_of() macro to go from the inode back to
the device instead.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/user_mad.c | 45 +++++++-----------------------------
1 files changed, 9 insertions(+), 36 deletions(-)

diff --git a/drivers/infiniband/core/user_mad.c b/drivers/infiniband/core/user_mad.c
index 40440ef..46c8879 100644
--- a/drivers/infiniband/core/user_mad.c
+++ b/drivers/infiniband/core/user_mad.c
@@ -65,12 +65,9 @@ enum {
};

/*
- * Our lifetime rules for these structs are the following: each time a
- * device special file is opened, we look up the corresponding struct
- * ib_umad_port by minor in the umad_port[] table while holding the
- * port_lock. If this lookup succeeds, we take a reference on the
- * ib_umad_port's struct ib_umad_device while still holding the
- * port_lock; if the lookup fails, we fail the open(). We drop these
+ * Our lifetime rules for these structs are the following:
+ * device special file is opened, we take a reference on the
+ * ib_umad_port's struct ib_umad_device. We drop these
* references in the corresponding close().
*
* In addition to references coming from open character devices, there
@@ -78,12 +75,7 @@ enum {
* module's reference taken when allocating the ib_umad_device in
* ib_umad_add_one().
*
- * When destroying an ib_umad_device, we clear all of its
- * ib_umad_ports from umad_port[] while holding port_lock before
- * dropping the module's reference to the ib_umad_device. This is
- * always safe because any open() calls will either succeed and obtain
- * a reference before we clear the umad_port[] entries, or fail after
- * we clear the umad_port[] entries.
+ * When destroying an ib_umad_device, we drop the module's reference.
*/

struct ib_umad_port {
@@ -136,7 +128,6 @@ static struct class *umad_class;
static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);

static DEFINE_SPINLOCK(port_lock);
-static struct ib_umad_port *umad_port[IB_UMAD_MAX_PORTS];
static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS);

static void ib_umad_add_one(struct ib_device *device);
@@ -779,15 +770,11 @@ static long ib_umad_compat_ioctl(struct file *filp, unsigned int cmd,
/*
* ib_umad_open() does not need the BKL:
*
- * - umad_port[] accesses are protected by port_lock, the
- * ib_umad_port structures are properly reference counted, and
+ * - the ib_umad_port structures are properly reference counted, and
* everything else is purely local to the file being created, so
* races against other open calls are not a problem;
* - the ioctl method does not affect any global state outside of the
* file structure being operated on;
- * - the port is added to umad_port[] as the last part of module
- * initialization so the open method will either immediately run
- * -ENXIO, or all required initialization will be done.
*/
static int ib_umad_open(struct inode *inode, struct file *filp)
{
@@ -795,13 +782,10 @@ static int ib_umad_open(struct inode *inode, struct file *filp)
struct ib_umad_file *file;
int ret = 0;

- spin_lock(&port_lock);
- port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE];
+ port = container_of(inode->i_cdev, struct ib_umad_port, cdev);
if (port)
kref_get(&port->umad_dev->ref);
- spin_unlock(&port_lock);
-
- if (!port)
+ else
return -ENXIO;

mutex_lock(&port->file_mutex);
@@ -892,13 +876,10 @@ static int ib_umad_sm_open(struct inode *inode, struct file *filp)
};
int ret;

- spin_lock(&port_lock);
- port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE - IB_UMAD_MAX_PORTS];
+ port = container_of(inode->i_cdev, struct ib_umad_port, sm_cdev);
if (port)
kref_get(&port->umad_dev->ref);
- spin_unlock(&port_lock);
-
- if (!port)
+ else
return -ENXIO;

if (filp->f_flags & O_NONBLOCK) {
@@ -1042,10 +1023,6 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
if (device_create_file(port->sm_dev, &dev_attr_port))
goto err_sm_dev;

- spin_lock(&port_lock);
- umad_port[port->dev_num] = port;
- spin_unlock(&port_lock);
-
return 0;

err_sm_dev:
@@ -1079,10 +1056,6 @@ static void ib_umad_kill_port(struct ib_umad_port *port)
cdev_del(&port->cdev);
cdev_del(&port->sm_cdev);

- spin_lock(&port_lock);
- umad_port[port->dev_num] = NULL;
- spin_unlock(&port_lock);
-
mutex_lock(&port->file_mutex);

port->ib_dev = NULL;

2010-02-02 19:08:50

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 10/18] IB/umad: use stack variable 'devnum' in ib_umad_init_port

This change is not useful by itself, but sets us up for a future
change that allows us to dynamically allocate device numbers
in case we have more than IB_UMAD_MAX_PORTS in the system.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/user_mad.c | 15 +++++++++------
1 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/drivers/infiniband/core/user_mad.c b/drivers/infiniband/core/user_mad.c
index 46c8879..df403ec 100644
--- a/drivers/infiniband/core/user_mad.c
+++ b/drivers/infiniband/core/user_mad.c
@@ -974,13 +974,16 @@ static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
static int ib_umad_init_port(struct ib_device *device, int port_num,
struct ib_umad_port *port)
{
+ int devnum;
+
spin_lock(&port_lock);
- port->dev_num = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
- if (port->dev_num >= IB_UMAD_MAX_PORTS) {
+ devnum = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
+ if (devnum >= IB_UMAD_MAX_PORTS) {
spin_unlock(&port_lock);
return -1;
}
- set_bit(port->dev_num, dev_map);
+ port->dev_num = devnum;
+ set_bit(devnum, dev_map);
spin_unlock(&port_lock);

port->ib_dev = device;
@@ -992,7 +995,7 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
cdev_init(&port->cdev, &umad_fops);
port->cdev.owner = THIS_MODULE;
kobject_set_name(&port->cdev.kobj, "umad%d", port->dev_num);
- if (cdev_add(&port->cdev, base_dev + port->dev_num, 1))
+ if (cdev_add(&port->cdev, base_dev + devnum, 1))
goto err_cdev;

port->dev = device_create(umad_class, device->dma_device,
@@ -1009,7 +1012,7 @@ static int ib_umad_init_port(struct ib_device *device, int port_num,
cdev_init(&port->sm_cdev, &umad_sm_fops);
port->sm_cdev.owner = THIS_MODULE;
kobject_set_name(&port->sm_cdev.kobj, "issm%d", port->dev_num);
- if (cdev_add(&port->sm_cdev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
+ if (cdev_add(&port->sm_cdev, base_dev + devnum + IB_UMAD_MAX_PORTS, 1))
goto err_sm_cdev;

port->sm_dev = device_create(umad_class, device->dma_device,
@@ -1036,7 +1039,7 @@ err_dev:

err_cdev:
cdev_del(&port->cdev);
- clear_bit(port->dev_num, dev_map);
+ clear_bit(devnum, dev_map);

return -1;
}

2010-02-02 19:12:04

by Alex Chiang

[permalink] [raw]
Subject: [PATCH v2 04/18] IB/uverbs: use stack variable 'base' in ib_uverbs_add_one

This change is not useful by itself, but sets us up for a future change
that allows us to support more than IB_UVERBS_MAX_DEVICES in a system.

Signed-off-by: Alex Chiang <[email protected]>
---

drivers/infiniband/core/uverbs_main.c | 4 +++-
1 files changed, 3 insertions(+), 1 deletions(-)

diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
index acae9ed..8c594cd 100644
--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -731,6 +731,7 @@ static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
static void ib_uverbs_add_one(struct ib_device *device)
{
int devnum;
+ dev_t base;
struct ib_uverbs_device *uverbs_dev;

if (!device->alloc_ucontext)
@@ -750,6 +751,7 @@ static void ib_uverbs_add_one(struct ib_device *device)
goto err;
}
uverbs_dev->devnum = devnum;
+ base = devnum + IB_UVERBS_BASE_DEV;
set_bit(devnum, dev_map);
spin_unlock(&map_lock);

@@ -760,7 +762,7 @@ static void ib_uverbs_add_one(struct ib_device *device)
uverbs_dev->cdev.owner = THIS_MODULE;
uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
- if (cdev_add(&uverbs_dev->cdev, IB_UVERBS_BASE_DEV + devnum, 1))
+ if (cdev_add(&uverbs_dev->cdev, base, 1))
goto err_cdev;

uverbs_dev->dev = device_create(uverbs_class, device->dma_device,

2010-02-08 21:38:45

by Alex Chiang

[permalink] [raw]
Subject: Re: [PATCH v2 00/18] Increase maximum number of Infiniband HCAs per system

Hi Roland,

Any thoughts on this patch series?

Thanks,
/ac

* Alex Chiang <[email protected]>:
> This is v2 of a patch series that increases the maximum number of
> IB HCAs supported per system.
>
> The original mail thread is here:
> http://lkml.org/lkml/2010/1/29/346
>
> One note, I decided to "copy/paste" since factoring out the overflow
> code in the three drivers seemed like overkill. If so desired, I could
> factor those three separate functions into something provided by the
> core, but that seemed like more trouble than it was worth at the time.
>
> As before, I still don't have access to a giant system, so what I did
> to test was to stick 4 cards into a small system, and then modify the
> limits with debug patches similar to this:
>
> diff --git a/drivers/infiniband/core/ucm.c b/drivers/infiniband/core/ucm.c
> index 7bf0a82..8581e64 100644
> --- a/drivers/infiniband/core/ucm.c
> +++ b/drivers/infiniband/core/ucm.c
> @@ -102,7 +102,7 @@ struct ib_ucm_event {
> enum {
> IB_UCM_MAJOR = 231,
> IB_UCM_BASE_MINOR = 224,
> - IB_UCM_MAX_DEVICES = 32
> + IB_UCM_MAX_DEVICES = 2
> };
>
>
> I tested all 3 drivers this way (uverbs, umad, ucm). I verified that
> we're not leaking device numbers on multiple modprobe/rmmod cycles,
> that there aren't any funny interactions when various combinations of
> the drivers are loaded.
>
> I did not test the rest of the OFED stack. I did write some trivial
> programs to open the devices in /dev and close them again.
>
> Here's an example of some of the testing:
>
> dl585g2:~ # modprobe ib_uverbs
> dl585g2:~ # modprobe ib_umad
> dl585g2:~ # modprobe ib_ucm
> dl585g2:~ # ls -l /dev/uverb*
> crw-rw---- 1 root root 231, 192 Feb 2 05:55 /dev/uverbs0
> crw-rw---- 1 root root 231, 193 Feb 2 05:55 /dev/uverbs1
> crw-rw---- 1 root root 249, 0 Feb 2 05:55 /dev/uverbs2
> crw-rw---- 1 root root 249, 1 Feb 2 05:55 /dev/uverbs3
> dl585g2:~ # ls -l /dev/umad*
> crw-rw---- 1 root root 231, 0 Feb 2 05:55 /dev/umad0
> crw-rw---- 1 root root 231, 1 Feb 2 05:55 /dev/umad1
> crw-rw---- 1 root root 231, 2 Feb 2 05:55 /dev/umad2
> crw-rw---- 1 root root 231, 3 Feb 2 05:55 /dev/umad3
> crw-rw---- 1 root root 248, 0 Feb 2 05:55 /dev/umad4
> crw-rw---- 1 root root 248, 1 Feb 2 05:55 /dev/umad5
> crw-rw---- 1 root root 248, 2 Feb 2 05:55 /dev/umad6
> crw-rw---- 1 root root 248, 3 Feb 2 05:55 /dev/umad7
> dl585g2:~ # ls -l /dev/issm*
> crw-rw---- 1 root root 231, 4 Feb 2 05:55 /dev/issm0
> crw-rw---- 1 root root 231, 5 Feb 2 05:55 /dev/issm1
> crw-rw---- 1 root root 231, 6 Feb 2 05:55 /dev/issm2
> crw-rw---- 1 root root 231, 7 Feb 2 05:55 /dev/issm3
> crw-rw---- 1 root root 248, 4 Feb 2 05:55 /dev/issm4
> crw-rw---- 1 root root 248, 5 Feb 2 05:55 /dev/issm5
> crw-rw---- 1 root root 248, 6 Feb 2 05:55 /dev/issm6
> crw-rw---- 1 root root 248, 7 Feb 2 05:55 /dev/issm7
> dl585g2:~ # ls -l /dev/ucm*
> crw-rw---- 1 root root 231, 224 Feb 2 05:55 /dev/ucm0
> crw-rw---- 1 root root 231, 225 Feb 2 05:55 /dev/ucm1
> crw-rw---- 1 root root 247, 0 Feb 2 05:55 /dev/ucm2
> crw-rw---- 1 root root 247, 1 Feb 2 05:55 /dev/ucm3
>
> Note that the major and minor numbers are behaving rather sanely.
>
> dl585g2:~ # rmmod ib_ucm
> dl585g2:~ # rmmod ib_uverbs
> dl585g2:~ # rmmod ib_umad
>
> Reset.
>
> dl585g2:~ # modprobe ib_ucm
> dl585g2:~ # ls -l /dev/ucm*
> crw-rw---- 1 root root 231, 224 Feb 2 05:57 /dev/ucm0
> crw-rw---- 1 root root 231, 225 Feb 2 05:57 /dev/ucm1
> crw-rw---- 1 root root 248, 0 Feb 2 05:57 /dev/ucm2
> crw-rw---- 1 root root 248, 1 Feb 2 05:57 /dev/ucm3
>
> See that /dev/ucm* devices now have a different major number
> compared to last time(248 vs 247), since we loaded that driver first.
>
> But wait, why is it 248 and not 249? Is there a leak somewhere?
>
> dl585g2:~ # ls -l /dev/uverb*
> crw-rw---- 1 root root 231, 192 Feb 2 05:57 /dev/uverbs0
> crw-rw---- 1 root root 231, 193 Feb 2 05:57 /dev/uverbs1
> crw-rw---- 1 root root 249, 0 Feb 2 05:57 /dev/uverbs2
> crw-rw---- 1 root root 249, 1 Feb 2 05:57 /dev/uverbs3
> dl585g2:~ # rmmod ib_uverbs
> ERROR: Module ib_uverbs is in use by ib_ucm
>
> Ah, ib_ucm is dependent on ib_uverbs, so when we modprobed ib_ucm,
> in reality ib_uverbs got loaded first. See how it has a higher
> major number.
>
> dl585g2:~ # rmmod ib_ucm
> dl585g2:~ # rmmod ib_uverbs
> dl585g2:~ # modprobe ib_umad
> dl585g2:~ # ls -l /dev/umad*
> crw-rw---- 1 root root 231, 0 Feb 2 05:58 /dev/umad0
> crw-rw---- 1 root root 231, 1 Feb 2 05:58 /dev/umad1
> crw-rw---- 1 root root 231, 2 Feb 2 05:58 /dev/umad2
> crw-rw---- 1 root root 231, 3 Feb 2 05:58 /dev/umad3
> crw-rw---- 1 root root 249, 0 Feb 2 05:58 /dev/umad4
> crw-rw---- 1 root root 249, 1 Feb 2 05:58 /dev/umad5
> crw-rw---- 1 root root 249, 2 Feb 2 05:58 /dev/umad6
> crw-rw---- 1 root root 249, 3 Feb 2 05:58 /dev/umad7
> dl585g2:~ # ls -l /dev/issm*
> crw-rw---- 1 root root 231, 4 Feb 2 05:58 /dev/issm0
> crw-rw---- 1 root root 231, 5 Feb 2 05:58 /dev/issm1
> crw-rw---- 1 root root 231, 6 Feb 2 05:58 /dev/issm2
> crw-rw---- 1 root root 231, 7 Feb 2 05:58 /dev/issm3
> crw-rw---- 1 root root 249, 4 Feb 2 05:58 /dev/issm4
> crw-rw---- 1 root root 249, 5 Feb 2 05:58 /dev/issm5
> crw-rw---- 1 root root 249, 6 Feb 2 05:58 /dev/issm6
> crw-rw---- 1 root root 249, 7 Feb 2 05:58 /dev/issm7
>
> Finally, after one more reset, we see ib_umad loaded first and
> obtaining the major number of 249.
>
> v1 -> v2:
> - update umad and ucm drivers too
>
> ---
>
> Alex Chiang (18):
> IB/uverbs: convert *cdev to cdev in struct ib_uverbs_device
> IB/uverbs: remove dev_table
> IB/uverbs: use stack variable 'devnum' in ib_uverbs_add_one
> IB/uverbs: use stack variable 'base' in ib_uverbs_add_one
> IB/uverbs: increase maximum devices supported
> IB/uverbs: pack struct ib_uverbs_event_file tighter
> IB/uverbs: whitespace cleanup
> IB/umad: convert cdev pointers to embedded structs in struct ib_umad_port
> IB/umad: remove port_table[]
> IB/umad: use stack variable 'devnum' in ib_umad_init_port
> IB/umad: use stack variable 'base' in ib_umad_init_port
> IB/umad: increase maximum devices supported
> IB/umad: clean whitespace
> IB/ucm: use stack variable 'devnum' in ib_ucm_add_one
> IB/ucm: use stack variable 'base' in ib_ucm_add_one
> IB/ucm: increase maximum devices supported
> IB/ucm: clean whitespace errors
> IB/core: pack struct ib_device a little tighter
>
>
> drivers/infiniband/core/ucm.c | 63 ++++++++++--
> drivers/infiniband/core/user_mad.c | 173 +++++++++++++++++----------------
> drivers/infiniband/core/uverbs.h | 11 +-
> drivers/infiniband/core/uverbs_main.c | 175 +++++++++++++++++++--------------
> include/rdma/ib_verbs.h | 4 -
> 5 files changed, 255 insertions(+), 171 deletions(-)
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/

2010-02-24 18:24:22

by Roland Dreier

[permalink] [raw]
Subject: Re: [PATCH v2 18/18] IB/core: pack struct ib_device a little tighter

Applied all 18, thanks a lot Alex. You really made the review easy by
splitting this up nicely...

- R.
--
Roland Dreier <[email protected]>
For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html