firewire: interface to remote memory (mem1394)
cover firewire: interface to remote memory (mem1394)
The following series of four patches adds the possiblity for creating device nodes for each node attached to the firewire bus and accessing them (currently read-only) which yields access to the node's RAM, if supported. This is useful for peeking into machines during debugging (obviously only if you can attach a firewire cable). This is already supported via the raw1394 interface, but surfacing this like /dev/mem is advantageous because then user-space tools can work without modification.
mem1394 itself is currently a bit limited and lacking on the error-checking, it will be improved but I wanted to get some comments on the current patches too. Write support will also be added.
The node interface and dynamic character device allocation are useful on their own if another layer is ever introduced -- which has been under discussion once a while to make something similar to raw1394 that gives access only to certain nodes to separate users.
If you want to test this add a udev rule like the following:
KERNEL=="fwmem-[0-9]*", NAME="fwmem-%s{device/guid}"
and then use
dd if=/dev/fwmem-0x<GUID> bs=1K count=1024 of=/tmp/first-megabyte
to, for example, read the targets first megabyte of memory.
The patches (will be posted as follow-ups):
1/4: node interface
2/4: dynamic cdev allocation below firewire major 3/4: unconditionally export hpsb_send_packet_and_wait 4/4: add mem1394
Comments would be appreciated.
Thanks,
Johannes
RFC 1/4 firewire: node interface
This patch adds just two small helper functions to allow other code to register a struct class_interface to interface node entries.
drivers/ieee1394/nodemgr.c +8 -0
@@ -1796,3 +1796,11 @@ void cleanup_ieee1394_nodemgr(void)class_unregister (&nodemgr_ud_class);class_unregister (&nodemgr_ne_class); } + +int hpsb_register_node_interface (struct class_interface *intf) +{ + intf->class = &nodemgr_ne_class; + +return class_interface_register (intf); +} +EXPORT_SYMBOL_GPL (hpsb_register_node_interface);
drivers/ieee1394/nodemgr.h +8 -0
@@ -170,6 +170,14 @@ int hpsb_node_write(struct node_entry *nint hpsb_node_lock (struct node_entry *ne,u64 addr,int extcode,quadlet_t *data,quadlet_t arg); +/* +* things like mem1394 are interfaces to nodes, thus +* allow them to register and unregister one. +*/ +int hpsb_register_node_interface (struct class_interface *intf); +static inline void hpsb_unregister_node_interface (struct class_interface *intf) { +class_interface_unregister (intf); +}/* Iterate the hosts, calling a given function with supplied data for each * host. */
RFC 2/4 firewire: dynamic cdev allocation below firewire major
In reply to Johannes BergThis patch implements dynamic minor number allocation below the 171 major allocated for ieee1394. Since on today's systems one doesn't need to have fixed device numbers any more we could just use any, but it's probably still useful to use the ieee1394 major number for any firewire related devices (like mem1394).
drivers/ieee1394/ieee1394_core.c +48 -0
@@ -29,10 +29,12 @@#include <linux/interrupt.h> #include <linux/module.h> #include <linux/moduleparam.h> +#include <linux/bitmap.h> #include <linux/bitops.h> #include <linux/kdev_t.h> #include <linux/skbuff.h> #include <linux/suspend.h> +#include <linux/cdev.h> #include <asm/byteorder.h> #include <asm/semaphore.h>
@@ -1053,9 +1055,14 @@ static int hpsbpkt_thread(void *__hi)complete_and_exit (&khpsbpkt_complete,0 ); } +/* used further below, but needs to be here for initialisation */ +static spinlock_t used_minors_lock; +static int __init ieee1394_init (void ) {int i, ret; + +spin_lock_init (&used_minors_lock);skb_queue_head_init (&hpsbpkt_queue);
@@ -1197,6 +1204,47 @@ static void __exit ieee1394_cleanup(voidmodule_init (ieee1394_init);module_exit (ieee1394_cleanup); +/* dynamic minor allocation functions */ +static DECLARE_BITMAP(used_minors, IEEE1394_MINOR_DYNAMIC_COUNT); + +int hpsb_cdev_add (struct cdev *chardev) +{ +int minor, ret; + +spin_lock (&used_minors_lock); + minor =find_first_zero_bit (used_minors,IEEE1394_MINOR_DYNAMIC_COUNT ); +if (minor >=IEEE1394_MINOR_DYNAMIC_COUNT ) { +spin_unlock (&used_minors_lock); +return -ENODEV ; + } +set_bit (minor, used_minors); +spin_unlock (&used_minors_lock); + + minor +=IEEE1394_MINOR_DYNAMIC_FIRST ; + ret =cdev_add (chardev,MKDEV (IEEE1394_MAJOR , minor),1 ); +if (unlikely (ret)) { +spin_lock (&used_minors_lock); +clear_bit (minor-IEEE1394_MINOR_DYNAMIC_FIRST , used_minors); +spin_unlock (&used_minors_lock); + } +return ret; +} +EXPORT_SYMBOL_GPL (hpsb_cdev_add); + +void hpsb_cdev_del (struct cdev *chardev) +{ +dev_t dev; + +BUG_ON (MAJOR (chardev->dev) !=IEEE1394_MAJOR ); + dev = chardev->dev; +cdev_del (chardev); + +spin_lock (&used_minors_lock); +clear_bit (MINOR (dev) -IEEE1394_MINOR_DYNAMIC_FIRST , used_minors); +spin_unlock (&used_minors_lock); +} +EXPORT_SYMBOL_GPL (hpsb_cdev_del); +/* Exported symbols */ /** hosts.c **/
drivers/ieee1394/ieee1394_core.h +25 -1
@@ -186,19 +186,38 @@ void hpsb_packet_received(struct hpsb_ho* 171:0-255, the various drivers must then cdev_add() their cdev * objects to handle their respective sub-regions. * +* Alternatively, drivers may use a dynamic minor number character +* device by using the functions hpsb_cdev_add and hpsb_cdev_del. +* hpsb_cdev_add requires an initialised struct cdev and will add +* it with cdev_add() automatically, reserving a new minor number +* for the new device (unless cdev_add() fails). It returns the +* status of cdev_add(), or -ENODEV if no minor could be allocated. +* +* Currently 64 minor numbers are reserved for that, if necessary +* this number can be increased by simply adjusting the constant +* IEEE1394_MINOR_DYNAMIC_FIRST. +* * Minor device number block allocations: * * Block 0 ( 0- 15) raw1394 * Block 1 ( 16- 31) video1394 * Block 2 ( 32- 47) dv1394 * -* Blocks 3-14 free for future allocation +* Blocks 3-10 free for future allocation * +* Block 11 (176-191) dynamic allocation region +* Block 12 (192-207) dynamic allocation region +* Block 13 (208-223) dynamic allocation region +* Block 14 (224-239) dynamic allocation region * Block 15 (240-255) reserved for drivers under development, etc. */ #define IEEE1394_MAJOR 171 +#define IEEE1394_MINOR_DYNAMIC_FIRST 176 +#define IEEE1394_MINOR_DYNAMIC_LAST 239 +#define IEEE1394_MINOR_DYNAMIC_COUNT (IEEE1394_MINOR_DYNAMIC_LAST-IEEE1394_MINOR_DYNAMIC_FIRST+1) +#define IEEE1394_MINOR_BLOCK_RAW1394 0#define IEEE1394_MINOR_BLOCK_VIDEO1394 1#define IEEE1394_MINOR_BLOCK_DV1394 2
@@ -218,6 +237,11 @@ static inline unsigned char ieee1394_filreturn file->f_dentry->d_inode->i_cindex; } +/* add a dynamic ieee1394 device */ +int hpsb_cdev_add (struct cdev *chardev); +/* remove a dynamic ieee1394 device */ +void hpsb_cdev_del (struct cdev *chardev); +extern int hpsb_disable_irm;/* Our sysfs bus entry */
RFC 3/4 firewire: unconditionally export hpsb_send_packet_and_wait
In reply to Johannes Bergmem1394 will need hpsb_send_packet_and_wait so it needs to be exported unconditionally.
drivers/ieee1394/ieee1394_core.c +1 -1
@@ -589,6 +589,7 @@ int hpsb_send_packet_and_wait(struct hpsreturn retval; } +EXPORT_SYMBOL (hpsb_send_packet_and_wait);static void send_packet_nocare (struct hpsb_packet *packet) {
@@ -1269,7 +1270,6 @@ EXPORT_SYMBOL(hpsb_packet_received);EXPORT_SYMBOL_GPL (hpsb_disable_irm);#ifdef CONFIG_IEEE1394_EXPORT_FULL_API EXPORT_SYMBOL (hpsb_send_phy_config); -EXPORT_SYMBOL (hpsb_send_packet_and_wait);#endif /** ieee1394_transactions.c **/
RFC 4/4 firewire: add mem1394
In reply to Johannes BergWhile the previous patches were purely infrastructure, this patch actually adds the code using it: mem1394.
There are some open questions on a few things, maybe someone can help out there.
drivers/ieee1394/Kconfig +17 -0
@@ -169,4 +169,21 @@ config IEEE1394_RAWIO To compile this driver as a module, say M here: the module will be called raw1394. +config IEEE1394_MEMDEV + tristate "IEEE1394 memory device support" + depends on IEEE1394 && EXPERIMENTAL + help + Say Y here if you want support for the ieee1394 memory device. + This is useful for debugging systems attached via firewire + since it usually allows you to read from and write to their memory, + depending on the controller and machine setup. + + It differs from raw access (which allows the same usage) in that + it provides devices nodes (usually called /dev/fwmem-<guid>) that can + be read and written with any tool, as opposed to specialised tools + required for raw1394. + + To compile this driver as a module, say M here: the + module will be called mem1394. + endmenu
drivers/ieee1394/Makefile +1 -0
@@ -14,6 +14,7 @@ obj-$(CONFIG_IEEE1394_RAWIO) += raw1394. obj-$(CONFIG_IEEE1394_SBP2) += sbp2.o obj-$(CONFIG_IEEE1394_DV1394) += dv1394.o obj-$(CONFIG_IEEE1394_ETH1394) += eth1394.o +obj-$(CONFIG_IEEE1394_MEMDEV) += mem1394.o quiet_cmd_oui2c = OUI2C $@ cmd_oui2c = $(CONFIG_SHELL) $(src)/oui2c.sh < $< > $@
drivers/ieee1394/mem1394.c new file +277 -0
277 lines not on this page · raw message
drivers/ieee1394/mem1394.h new file +20 -0
@@ -0,0 +1,20 @@ +#ifndef IEEE1394_MEM1394_H +#define IEEE1394_MEM1394_H + +#include <asm/types.h> +#include <linux/cdev.h> +#include <linux/list.h> +#include "nodemgr.h" + +struct mem1394_dev { +struct device *dev; +struct node_entry *ne; +struct cdev cdev; +struct list_head list; +}; + +struct mem1394_file_info { +struct mem1394_dev *memdev; +}; + +#endif /* IEEE1394_MEM1394_H */
Re: [RFC 4/4] firewire: add mem1394
Hi Johannes,
Review
Quoted 1 line · Johannes Berg
(quoted diff) +1
-0+ spin_lock(&dev_list_lock);
Stupid question: are you sure that something coming from an interrupt handler won't try to grab this lock? For example from a cable unplug?
Regards,
Signature · 2 lines
Andy Wingo http://wingolog.org/
Re: [RFC 4/4] firewire: add mem1394
Quoted 5 lines · Andy Wingo
Review
Quoted 1 line · Johannes Berg
(quoted diff) +1
-0+ spin_lock(&dev_list_lock);Stupid question: are you sure that something coming from an interrupt handler won't try to grab this lock? For example from a cable unplug?
Yes, I'm pretty sure (but I hope some of the firewire experts will chime in) -- but if you unplug or anything the node only goes into 'limbo' and afaict if it is ever cleaned up then that comes from a thread context.
johannes
Re: [RFC 4/4] firewire: add mem1394
In reply to Johannes Bergdrivers/ieee1394/Kconfig config IEEE1394_RAWIO · line 179
Quoted 9 lines · Johannes Berg
drivers/ieee1394/Kconfig +8
-0+config IEEE1394_MEMDEV + tristate "IEEE1394 memory device support" + depends on IEEE1394 && EXPERIMENTAL + help + Say Y here if you want support for the ieee1394 memory device. + This is useful for debugging systems attached via firewire + since it usually allows you to read from and write to their memory, + depending on the controller and machine setup.
1394 is evil. Does this mean that if a machine is completely dead-and-crashed, we can still suck all its memory out over 1394 with no cooperation from the dead machine's kernel? If not, what limitations are there?
Triviata:
drivers/ieee1394/mem1394.c line 49
Quoted 4 lines
drivers/ieee1394/mem1394.c new file +4
-0+static int mem1394_read (struct file *file,char __user * buffer, +size_t count,loff_t *offset) +{ +struct mem1394_file_info *fi = (struct mem1394_file_info *)file->private_data;
Unneeded cast.
drivers/ieee1394/mem1394.c line 91
Quoted 1 line
drivers/ieee1394/mem1394.c new file +1
-0+ packet =hpsb_make_readpacket (fi->memdev->ne->host, fi->memdev->ne->nodeid, *offset, submitcount);
xterm too big!
drivers/ieee1394/mem1394.c line 126
Quoted 3 lines
drivers/ieee1394/mem1394.c new file +3
-0+static int mem1394_release (struct inode *inode,struct file *file) +{ +struct mem1394_file_info *fi = (struct mem1394_file_info *)file->private_data;
Unneeded cast.
Review
Quoted 1 line
(quoted diff) +1
-0+
Adds trailing whitespace ;)
drivers/ieee1394/mem1394.c line 150
Quoted 6 lines
drivers/ieee1394/mem1394.c new file +6
-0+static struct mem1394_dev *alloc_mem1394_dev (struct device *dev) +{ +struct mem1394_dev *result; + struct node_entry *ne = container_of(dev, struct node_entry, device); +int ret; +struct class_device * mem1394_class_member;
Inconsistent space-after-asterisk policy (no-space is preferred).
drivers/ieee1394/mem1394.c line 168
Quoted 2 lines
drivers/ieee1394/mem1394.c new file +2
-0+ mem1394_class_member =class_device_create (mem1394_sysfs_class,NULL , result->cdev.dev, + dev,"fwmem-%d" ,atomic_read (&mem1394_dev_ctr));
My eyes!
drivers/ieee1394/mem1394.c line 173
Quoted 5 lines
drivers/ieee1394/mem1394.c new file +5
-0+if (IS_ERR (mem1394_class_member)) { +printk (KERN_WARNING "mem1394: class_device_create failed\n" ); + }else { +class_set_devdata (mem1394_class_member, result); + }
Unneeded braces.
drivers/ieee1394/mem1394.c line 256
Quoted 3 lines
drivers/ieee1394/mem1394.c new file +3
-0+if (IS_ERR (mem1394_sysfs_class)) { +return PTR_ERR (mem1394_sysfs_class); + }
Ditto.
Re: [RFC 4/4] firewire: add mem1394
Quoted 13 lines · Andrew Morton
drivers/ieee1394/Kconfig config IEEE1394_RAWIO · line 179
Quoted 10 lines · Johannes Berg
drivers/ieee1394/Kconfig +7
-0+config IEEE1394_MEMDEV + tristate "IEEE1394 memory device support" + depends on IEEE1394 && EXPERIMENTAL + help + Say Y here if you want support for the ieee1394 memory device. + This is useful for debugging systems attached via firewire + since it usually allows you to read from and write to theirmemory,
drivers/ieee1394/Kconfig +1
-0+ depending on the controller and machine setup.1394 is evil. Does this mean that if a machine is completely dead-and-crashed, we can still suck all its memory out over 1394 with no cooperation from the dead machine's kernel? If not, what limitations are there?
I think you snipped too much of the description. This was after the part you quoted:
Quoted 1 line
It differs from raw access (which allows the same usage) in that it provides devices nodes (usually called /dev/fwmem-<guid>) that can be read and written with any tool, as opposed to specialised tools required for raw1394.
This seems to indicate that this is a _client_ for a IEEE1394 memory device, as opposed to a server. Perhaps the description should be clarified, but I don't see any security issues (the kernel does not expose its own memory, it accesses the memory that another device is exposing).
Cheers,
Kyle Moffett
--
There is no way to make Linux robust with unreliable memory subsystems, sorry. It would be like trying to make a human more robust with an unreliable O2 supply. Memory just has to work.
-- Andi Kleen
Re: [RFC 4/4] firewire: add mem1394
In reply to Johannes BergQuoted 11 lines · Johannes Berg
Quoted 6 lines · Andy Wingo
Review
Quoted 1 line · Johannes Berg
(quoted diff) +1
-0+ spin_lock(&dev_list_lock);Stupid question: are you sure that something coming from an interrupt handler won't try to grab this lock? For example from a cable unplug?
Yes, I'm pretty sure (but I hope some of the firewire experts will chime in) -- but if you unplug or anything the node only goes into 'limbo' and afaict if it is ever cleaned up then that comes from a thread context.
The lock will only be taken in non-atomic context. In particular, if a mem1394 device is to be removed after cable unplug, the code will be run by knodemgrd.
What is more recommendable for mutual exclusion in non-atomic context (here also with very low probality of lock contention, given the current implementation of ieee1394) --- a mutex or a spinlock?
Signature · 3 lines
Stefan Richter -=====-=-==- --=- --=-= http://arcgraph.de/sr/
Re: [RFC 2/4] firewire: dynamic cdev allocation below firewire major
In reply to Johannes BergQuoted 5 lines · Johannes Berg
This patch implements dynamic minor number allocation below the 171 major allocated for ieee1394. Since on today's systems one doesn't need to have fixed device numbers any more we could just use any, but it's probably still useful to use the ieee1394 major number for any firewire related devices (like mem1394).
[...]
drivers/ieee1394/ieee1394_core.h void hpsb_packet_received(struct hpsb_ho · line 194
Quoted 12 lines
drivers/ieee1394/ieee1394_core.h +6
-0@@ -186,19 +186,38 @@ void hpsb_packet_received(struct hpsb_ho* 171:0-255, the various drivers must then cdev_add() their cdev * objects to handle their respective sub-regions. * +* Alternatively, drivers may use a dynamic minor number character +* device by using the functions hpsb_cdev_add and hpsb_cdev_del. +* hpsb_cdev_add requires an initialised struct cdev and will add +* it with cdev_add() automatically, reserving a new minor number +* for the new device (unless cdev_add() fails). It returns the +* status of cdev_add(), or -ENODEV if no minor could be allocated.
This comment should be moved to the implementations of respective functions and be formatted as described by Documentation/kernel-doc-nano-HOWTO.txt. (We should eventually check all exported ieee1394 symbols if they are documented that way.)
Quoted 30 lines
drivers/ieee1394/ieee1394_core.h +13
-1+* Currently 64 minor numbers are reserved for that, if necessary +* this number can be increased by simply adjusting the constant +* IEEE1394_MINOR_DYNAMIC_FIRST. +* * Minor device number block allocations: * * Block 0 ( 0- 15) raw1394 * Block 1 ( 16- 31) video1394 * Block 2 ( 32- 47) dv1394 * -* Blocks 3-14 free for future allocation +* Blocks 3-10 free for future allocation * +* Block 11 (176-191) dynamic allocation region +* Block 12 (192-207) dynamic allocation region +* Block 13 (208-223) dynamic allocation region +* Block 14 (224-239) dynamic allocation region * Block 15 (240-255) reserved for drivers under development, etc. */ #define IEEE1394_MAJOR 171 +#define IEEE1394_MINOR_DYNAMIC_FIRST 176 +#define IEEE1394_MINOR_DYNAMIC_LAST 239 +#define IEEE1394_MINOR_DYNAMIC_COUNT (IEEE1394_MINOR_DYNAMIC_LAST-IEEE1394_MINOR_DYNAMIC_FIRST+1) +#define IEEE1394_MINOR_BLOCK_RAW1394 0#define IEEE1394_MINOR_BLOCK_VIDEO1394 1#define IEEE1394_MINOR_BLOCK_DV1394 2
drivers/ieee1394/ieee1394_core.h static inline unsigned char ieee1394_fil · line 243
Quoted 7 lines
drivers/ieee1394/ieee1394_core.h +4
-0@@ -218,6 +237,11 @@ static inline unsigned char ieee1394_filreturn file->f_dentry->d_inode->i_cindex; } +/* add a dynamic ieee1394 device */ +int hpsb_cdev_add (struct cdev *chardev); +/* remove a dynamic ieee1394 device */ +void hpsb_cdev_del (struct cdev *chardev);
[...]
Again, better no comment here than these two which are not very enlightening.
Signature · 3 lines
Stefan Richter -=====-=-==- --=- --=-= http://arcgraph.de/sr/
Re: [RFC 3/4] firewire: unconditionally export hpsb_send_packet_and_wait
In reply to Johannes BergQuoted 16 lines · Johannes Berg
mem1394 will need
hpsb_send_packet_and_waitso it needs to be exported unconditionally.drivers/ieee1394/ieee1394_core.c +1
-0@@ -589,6 +589,7 @@ int hpsb_send_packet_and_wait(struct hpsreturn retval; } +EXPORT_SYMBOL (hpsb_send_packet_and_wait);static void send_packet_nocare (struct hpsb_packet *packet) {
drivers/ieee1394/ieee1394_core.c EXPORT_SYMBOL(hpsb_packet_received); · line 1275
Quoted 7 lines
drivers/ieee1394/ieee1394_core.c +0
-1@@ -1269,7 +1270,6 @@ EXPORT_SYMBOL(hpsb_packet_received);EXPORT_SYMBOL_GPL (hpsb_disable_irm);#ifdef CONFIG_IEEE1394_EXPORT_FULL_API EXPORT_SYMBOL (hpsb_send_phy_config); -EXPORT_SYMBOL (hpsb_send_packet_and_wait);#endif /** ieee1394_transactions.c **/
Leave the export down at the end of ieee1394_core.c among all other exports of ieee1394. Just move the export above the #ifdef.
Same for the two new exports by "[RFC 2/4] firewire: dynamic cdev allocation below firewire major": Place them at the end of ieee1394_core.c.
Signature · 3 lines
Stefan Richter -=====-=-==- --=- --=-= http://arcgraph.de/sr/
Re: [RFC 4/4] firewire: add mem1394
In reply to Johannes BergQuoted 5 lines · Johannes Berg
While the previous patches were purely infrastructure, this patch actually adds the code using it: mem1394.
There are some open questions on a few things, maybe someone can help out there.
[...]
drivers/ieee1394/mem1394.c line 65
Quoted 11 lines
drivers/ieee1394/mem1394.c new file +11
-0+/* this is a bit icky. I think I'll want to create a +* "struct hpsb_node_class_interface" that you register +* with nodemgr.c instead of registering the "struct class_interface" +* directly. It would wrap around the "struct class_interface" +* and handle things like this. +* +* This means it would call the node_class_interface's +* - "add" method whenever the device is fully there, and an +* - "update" method when it survived a bus reset, and the +* - "remove" method when it went away, also taking care of +* debouncing, which the mem1394 interface currently doesn't handle.
I currently think so too.
drivers/ieee1394/mem1394.c line 71
Quoted 5 lines
drivers/ieee1394/mem1394.c new file +5
-0+* But I need advice on this. It'll probably works this way +* but most likely not once this interface stuff gets more +* use; I can imagine using it for scanners instead of raw1394 +* so that the kernel can validate that a user can only +* access a certain scanner and not all 1394 devices on the bus.
Probably not. All devices (except perhaps custom embedded devices) which implement one or another high level protocol will always be accessed either by a protocol driver in kernelspace (like sbp2, eth1394, video1394) on top of a struct unit_directory, or by a driver or library in userspace on top of libraw1394/ raw1394. This is because such devices and protocols all implement the ISO/IEC 13213 CSR architecture.
drivers/ieee1394/mem1394.c line 83
Quoted 12 lines
drivers/ieee1394/mem1394.c new file +12
-0+* In other words some 'raw1394intf' instead of 'raw1394' which +* creates one character device per ieee1394 node for finer +* grained access control. +* That would definitely want to have debouncing etc. +* +* However, I don't fully understand the states node_entries go +* through yet, so I'm not sure this should even be here! +* Maybe it should be in open? But then the device could go +* into limbo when it is already opened... +* +* Similarly, what happens if a node is suspended? +*/
[...]
When a node represented by a node_entry leaves the bus, the node_entry is "suspended" and "put into limbo", which is both the same for the 1394 stack. The node_entry is only "removed" when forced by userspace through ieee1394's sysfs interface or when the ieee1394 driver module is unloaded. A unit_directory is either "suspended" or "removed", depending on what the protocol driver bound to the unit_directory implements. This behaviour of ieee1394 is currently not extensively used, but I plan to implement capability of sbp2 to survive transient disconnection on top of it.
I still haven't tested your driver yet and won't be able to do so during the next days...
Signature · 3 lines
Stefan Richter -=====-=-==- --=- --=-= http://arcgraph.de/sr/
Re: [RFC 4/4] firewire: add mem1394
In reply to a message not in the archiveI wrote:
(ohci1394's programming of physical DMA filters)
Quoted 1 line
The else clause is BTW bogus [...], but this is only important if an IEEE 1394.1 bus bridge was present.
I had a closer look. Physical DMA _is_ reliably disabled for all nodes (including nodes on bridged buses) if ohci1394 was loaded with phys_dma=0.
Signature · 3 lines
Stefan Richter -=====-=-==- --=- --=-= http://arcgraph.de/sr/
Re: [RFC 4/4] firewire: add mem1394
In reply to Andrew MortonQuoted 15 lines · Andrew Morton
drivers/ieee1394/Kconfig config IEEE1394_RAWIO · line 179
Quoted 9 lines · Johannes Berg
drivers/ieee1394/Kconfig +8
-0+config IEEE1394_MEMDEV + tristate "IEEE1394 memory device support" + depends on IEEE1394 && EXPERIMENTAL + help + Say Y here if you want support for the ieee1394 memory device. + This is useful for debugging systems attached via firewire + since it usually allows you to read from and write to their memory, + depending on the controller and machine setup.1394 is evil. Does this mean that if a machine is completely dead-and-crashed, we can still suck all its memory out over 1394 with no cooperation from the dead machine's kernel? If not, what limitations are there?
Yes it can. BenH's firescope tool does this already using raw1394 (I have it working now on x86-64 too). I dont quite see the point of adding another kernel driver for it though. This can be all done fine in userspace.
-Andi
Re: [RFC 4/4] firewire: add mem1394
Quoted 4 lines · Andi Kleen
BenH's firescope tool does this already using raw1394 (I have it working now on x86-64 too). I dont quite see the point of adding another kernel driver for it though. This can be all done fine in userspace.
The point is to provide an interface like /dev/mem in order to use a wider range of debug/ forensics/ hacker tools than specialized libraw1394 clients. Sure enough, the benefit is small, but so is this driver's code footprint.
Signature · 3 lines
Stefan Richter -=====-=-==- --=- --=-= http://arcgraph.de/sr/
Re: [RFC 4/4] firewire: add mem1394
Quoted 9 lines · Stefan Richter
Quoted 4 lines · Andi Kleen
BenH's firescope tool does this already using raw1394 (I have it working now on x86-64 too). I dont quite see the point of adding another kernel driver for it though. This can be all done fine in userspace.
The point is to provide an interface like
/dev/memin order to use a wider range of debug/ forensics/ hacker tools than specialized libraw1394 clients.
I don't see the benefit really. It can be as well provided by a userspace library
Many of the debug tools don't even work on /dev/mem, but use different interfaces (/proc/kcore, gdb remote protocol etc.)
Also raw1394 could possibly be used to cause interrupts on the target and also stop the target CPU this way.
-Andi
Re: [RFC 4/4] firewire: add mem1394
In reply to Stefan RichterQuoted 4 lines · Stefan Richter
Quoted 4 lines
drivers/ieee1394/mem1394.c new file +4
-0+* But I need advice on this. It'll probably works this way +* but most likely not once this interface stuff gets more +* use; I can imagine using it for scanners instead of raw1394 +* so that the kernel can validate that a user can only
Quoted 8 lines
drivers/ieee1394/mem1394.c line 71
Quoted 1 line
drivers/ieee1394/mem1394.c new file +1
-0+* access a certain scanner and not all 1394 devices on the bus. Probably not. All devices (except perhaps custom embedded devices) which implement one or another high level protocol will always be accessed either by a protocol driver in kernelspace (like sbp2, eth1394, video1394) on top of a struct
unit_directory, or by a driver or library in userspace on top of libraw1394/ raw1394. This is because such devices and protocols all implement the ISO/IEC 13213 CSR architecture.
You snipped too much :) At this point I was thinking of the raw1394 replacement that has finer grained access control which we talked about in other threads too.
Quoted 7 lines
drivers/ieee1394/mem1394.c line 75
Quoted 4 lines
drivers/ieee1394/mem1394.c new file +4
-0+* In other words some 'raw1394intf' instead of 'raw1394' which +* creates one character device per ieee1394 node for finer +* grained access control. +* That would definitely want to have debouncing etc. When a node represented by a
node_entryleaves the bus, thenode_entry
Quoted 8 lines
is "suspended" and "put into limbo", which is both the same for the 1394 stack. The
node_entryis only "removed" when forced by userspace through ieee1394's sysfs interface or when the ieee1394 driver module is unloaded. Aunit_directoryis either "suspended" or "removed", depending on what the protocol driver bound to theunit_directoryimplements. This behaviour of ieee1394 is currently not extensively used, but I plan to implement capability of sbp2 to survive transient disconnection on top of it.
Thanks for the explanation. I'll have to test my driver under the light of this.
johannes
Re: [RFC 3/4] firewire: unconditionally export hpsb_send_packet_and_wait
In reply to Stefan RichterQuoted 5 lines · Stefan Richter
Leave the export down at the end of
ieee1394_core.camong all other exports of ieee1394. Just move the export above the #ifdef.Same for the two new exports by "[RFC 2/4] firewire: dynamic cdev allocation below firewire major": Place them at the end of
ieee1394_core.c.
Somehow I thought we were supposed to now put the EXPORT_SYMBOL{,_GPL} right after the declaration. I can post a patch to move all of them instead just the few if desired, or change this patch.
johannes
Re: [RFC 2/4] firewire: dynamic cdev allocation below firewire major
In reply to Johannes BergQuoted 5 lines · Johannes Berg
This patch implements dynamic minor number allocation below the 171 major allocated for ieee1394. Since on today's systems one doesn't need to have fixed device numbers any more we could just use any, but it's probably still useful to use the ieee1394 major number for any firewire related devices (like mem1394).
I don't really like this. There's no benefit to using the 1394 major number. I'd rather see an improved alloc_chrdev_region() that does something like this but for the whole kernel (currently it "wastes" an entire major even if you only want 1 minor, and for what you're doing, grabbing 1 minor at a time makes the most sense.)
I'll get to that someday unless someone beats me to it (tm) because it will be necessary for the raw1394 security improvements I'm (slowly) working on. Unless someone convinces me that dynamic allocation under 171 is really the way to go.
Cheers,
Jody
Quoted 144 lines
drivers/ieee1394/ieee1394_core.c +48
-0@@ -29,10 +29,12 @@#include <linux/interrupt.h> #include <linux/module.h> #include <linux/moduleparam.h> +#include <linux/bitmap.h> #include <linux/bitops.h> #include <linux/kdev_t.h> #include <linux/skbuff.h> #include <linux/suspend.h> +#include <linux/cdev.h> #include <asm/byteorder.h> #include <asm/semaphore.h> @@ -1053,9 +1055,14 @@ static int hpsbpkt_thread(void *__hi)complete_and_exit (&khpsbpkt_complete,0 ); } +/* used further below, but needs to be here for initialisation */ +static spinlock_t used_minors_lock; +static int __init ieee1394_init (void ) {int i, ret; + +spin_lock_init (&used_minors_lock);skb_queue_head_init (&hpsbpkt_queue);@@ -1197,6 +1204,47 @@ static void __exit ieee1394_cleanup(voidmodule_init (ieee1394_init);module_exit (ieee1394_cleanup); +/* dynamic minor allocation functions */ +static DECLARE_BITMAP(used_minors, IEEE1394_MINOR_DYNAMIC_COUNT); + +int hpsb_cdev_add (struct cdev *chardev) +{ +int minor, ret; + +spin_lock (&used_minors_lock); + minor =find_first_zero_bit (used_minors,IEEE1394_MINOR_DYNAMIC_COUNT ); +if (minor >=IEEE1394_MINOR_DYNAMIC_COUNT ) { +spin_unlock (&used_minors_lock); +return -ENODEV ; + } +set_bit (minor, used_minors); +spin_unlock (&used_minors_lock); + + minor +=IEEE1394_MINOR_DYNAMIC_FIRST ; + ret =cdev_add (chardev,MKDEV (IEEE1394_MAJOR , minor),1 ); +if (unlikely (ret)) { +spin_lock (&used_minors_lock); +clear_bit (minor-IEEE1394_MINOR_DYNAMIC_FIRST , used_minors); +spin_unlock (&used_minors_lock); + } +return ret; +} +EXPORT_SYMBOL_GPL (hpsb_cdev_add); + +void hpsb_cdev_del (struct cdev *chardev) +{ +dev_t dev; + +BUG_ON (MAJOR (chardev->dev) !=IEEE1394_MAJOR ); + dev = chardev->dev; +cdev_del (chardev); + +spin_lock (&used_minors_lock); +clear_bit (MINOR (dev) -IEEE1394_MINOR_DYNAMIC_FIRST , used_minors); +spin_unlock (&used_minors_lock); +} +EXPORT_SYMBOL_GPL (hpsb_cdev_del); +/* Exported symbols */ /** hosts.c **/ drivers/ieee1394/ieee1394_core.h +25
-2@@ -186,19 +186,38 @@ void hpsb_packet_received(struct hpsb_ho* 171:0-255, the various drivers must then cdev_add() their cdev * objects to handle their respective sub-regions. * +* Alternatively, drivers may use a dynamic minor number character +* device by using the functions hpsb_cdev_add and hpsb_cdev_del. +* hpsb_cdev_add requires an initialised struct cdev and will add +* it with cdev_add() automatically, reserving a new minor number +* for the new device (unless cdev_add() fails). It returns the +* status of cdev_add(), or -ENODEV if no minor could be allocated. +* +* Currently 64 minor numbers are reserved for that, if necessary +* this number can be increased by simply adjusting the constant +* IEEE1394_MINOR_DYNAMIC_FIRST. +* * Minor device number block allocations: * * Block 0 ( 0- 15) raw1394 * Block 1 ( 16- 31) video1394 * Block 2 ( 32- 47) dv1394 * -* Blocks 3-14 free for future allocation +* Blocks 3-10 free for future allocation * +* Block 11 (176-191) dynamic allocation region +* Block 12 (192-207) dynamic allocation region +* Block 13 (208-223) dynamic allocation region +* Block 14 (224-239) dynamic allocation region * Block 15 (240-255) reserved for drivers under development, etc. */ #define IEEE1394_MAJOR 171 +#define IEEE1394_MINOR_DYNAMIC_FIRST 176 +#define IEEE1394_MINOR_DYNAMIC_LAST 239 +#define IEEE1394_MINOR_DYNAMIC_COUNT (IEEE1394_MINOR_DYNAMIC_LAST-IEEE1394_MINOR_DYNAMIC_FIRST+1) +#define IEEE1394_MINOR_BLOCK_RAW1394 0#define IEEE1394_MINOR_BLOCK_VIDEO1394 1#define IEEE1394_MINOR_BLOCK_DV1394 2@@ -218,6 +237,11 @@ static inline unsigned char ieee1394_filreturn file->f_dentry->d_inode->i_cindex; } +/* add a dynamic ieee1394 device */ +int hpsb_cdev_add (struct cdev *chardev); +/* remove a dynamic ieee1394 device */ +void hpsb_cdev_del (struct cdev *chardev); +extern int hpsb_disable_irm;/* Our sysfs bus entry */ -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/
Re: [RFC 2/4] firewire: dynamic cdev allocation below firewire major
Quoted 7 lines · Jody McIntyre
Quoted 5 lines · Johannes Berg
This patch implements dynamic minor number allocation below the 171 major allocated for ieee1394. Since on today's systems one doesn't need to have fixed device numbers any more we could just use any, but it's probably still useful to use the ieee1394 major number for any firewire related devices (like mem1394).
Quoted 5 lines
I don't really like this. There's no benefit to using the 1394 major number. I'd rather see an improved
alloc_chrdev_region()that does something like this but for the whole kernel (currently it "wastes" an entire major even if you only want 1 minor, and for what you're doing, grabbing 1 minor at a time makes the most sense.)
why bother? There's a LOT of majors nowadays (12 bits) so... what's the problem with keeping the kernel side simple?
(it's not as if userspace needs to care about the exact numbers anyway for almost everything)
Re: [RFC 2/4] firewire: dynamic cdev allocation below firewire major
Quoted 10 lines · Arjan van de Ven
Quoted 5 lines
I don't really like this. There's no benefit to using the 1394 major number. I'd rather see an improved
alloc_chrdev_region()that does something like this but for the whole kernel (currently it "wastes" an entire major even if you only want 1 minor, and for what you're doing, grabbing 1 minor at a time makes the most sense.)why bother? There's a LOT of majors nowadays (12 bits) so... what's the problem with keeping the kernel side simple?
(it's not as if userspace needs to care about the exact numbers anyway for almost everything)
Uh, ok. Seems pretty weird to effectively allocate 256 device numbers for just a single device, but ok :)
I'll drop the patch and make it allocate a new major for every device plugged in.
johannes
Re: [RFC 2/4] firewire: dynamic cdev allocation below firewire major
Quoted 4 lines · Johannes Berg
Seems pretty weird to effectively allocate 256 device numbers for just a single device, but ok :)
I'll drop the patch and make it allocate a new major for every device plugged in.
Your driver could internally dispatch 256 minor device numbers after it got itself its own dynamic major number. This should even be acceptable as a hard limit of mem1394 devices. One would need quite a lot of 1394 cards (or 1394.1 bridges) to get access to 256 FireWire nodes at once.
Signature · 3 lines
Stefan Richter -=====-=-==- --=- -==-= http://arcgraph.de/sr/
Re: [RFC 2/4] firewire: dynamic cdev allocation below firewire major
In reply to Johannes BergQuoted 7 lines · Johannes Berg
Quoted 6 lines · Arjan van de Ven
Quoted 5 lines
I don't really like this. There's no benefit to using the 1394 major number. I'd rather see an improved
alloc_chrdev_region()that does something like this but for the whole kernel (currently it "wastes" an entire major even if you only want 1 minor, and for what you're doing, grabbing 1 minor at a time makes the most sense.)
Quoted 7 lines
Quoted 4 lines
why bother? There's a LOT of majors nowadays (12 bits) so... what's the problem with keeping the kernel side simple?
(it's not as if userspace needs to care about the exact numbers anyway for almost everything)Uh, ok. Seems pretty weird to effectively allocate 256 device numbers for just a single device, but ok :)
it's not 256 it's 2^20.... but still :)
(eg there are 20 bits to a minor, 12 to a major)
Re: [RFC 2/4] firewire: dynamic cdev allocation below firewire major
Quoted 2 lines · Arjan van de Ven
it's not 256 it's 2^20.... but still :)
(eg there are 20 bits to a minor, 12 to a major)
Umm, right. But I guess Stefan's right too, I should instead allocate a single major and use it's minors.
johannes