2.6.10.rc1.bk6 /lib/kobject_uevent.c buffer issues
Subsystems (2)
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 10 lines · Andrew Duggan
Hi,
In the lastest code I've seen 2.6.10 rc1 bk6 for
lib/kobject_uevent.cthe incrementing of the sequence number is held off until the call tohotplug_ops->hotplug(). That function can consume addtional buffer and use addtional pointers in the envp array. The problem is that the new value for i and the new logical end of buffer (held in scratch) do not get updated. When the sequnce number is then processed the first item generated byhotplug_ops->hotplug() gets overwriten, and can potentially corrupt any env vars created byhotplug_ops->hotplug.
Ah, good catch. Sorry about adding that bug.
Quoted 6 lines
Here is a patch to handle that. An element in envp is reserved for the seq num and a different buffer is sent to
hotplug_ops->hotplugaltogether. Using a different buffer there is OK sincesend_ueventwasn't seeing the part of the buffer filled in byhotplug_ops->hotpluganyway, and the oldercall_usermodehelpergets passed envp and not buffer.
Why not just use the same buffer? We should be able to do that.
Quoted 5 lines
I used the same size for the
buffer_pt2as buffer for simplicity, since I didn't know if saving the memory was THAT important. If need be the kmalloc forbuffer_pt2could be held off until just before the call tohotplug_ops->hotplugso that it could be allocated only (BUFFERSIZE - (scratch - buffer)).
I'd prefer we use the same buffer. Care to respin your patch?
thanks,
greg k-h
Quoted 7 lines · Greg KH
Why not just use the same buffer? We should be able to do that.
I'd prefer we use the same buffer. Care to respin your patch?
Sure, I can only see two ways of achieving that however.
1) Change the API of kset_hotplug_ops.hotplug() to return the amount of consumed buffer (and possibly an updated value for i (num_envp) and then changing every real function that implements that interface or
2) Spin through the envp[] starting at i to NUM_ENVP looking for a NULL pointer dropping back 1 (last_used) then do a
scratch += strlen(envp[last_used]) + 1
I don't know if you would rather keep the kset_hotplug_ops.hotplug() API unchanged and have a "find the NULL" or vice-versa? -- or are both too ugly? Ultimately the first option would generate the "cleanest" solution, but I would be wary to change an API (especially one I didn't create myself)
I can also see variation of 2), where buffer is prefilled with something like '0xff' then the find the null would run in reverse from buffer+BUFFER_SIZE-1. Either way with 2), the reservation in envp[] for the seq_num would have to stay, but its locaton in buffer would be at the end. As a bonus of course then the call to send_uevent could see the entire buffer.
I'm looking now for how many changes would be required to do option 1.
Andrew
Quoted 5 lines
thanks,
greg k-h
Quoted 10 lines · Andrew
Quoted 7 lines · Greg KH
Why not just use the same buffer? We should be able to do that.
I'd prefer we use the same buffer. Care to respin your patch?
Quoted 8 lines
Sure, I can only see two ways of achieving that however.
1) Change the API ofkset_hotplug_ops.hotplug()to return the amount of consumed buffer (and possibly an updated value for i (num_envp) and then changing every real function that implements that interface or
2) Spin through the envp[] starting at i toNUM_ENVPlooking for a NULL pointer dropping back 1 (last_used) then do a
scratch +=strlen(envp[last_used])+ 1
Ick, ok, let's stick with 2 buffers. How about the patch below? It's a bit smaller than yours.
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
I'll hold off on applying this patch until we figure this out...
thanks,
greg k-h
kobject_uevent.c renamed from lib/kobject_uevent.c +9 -3
@@ -182,6 +182,7 @@char *argv [3 ];char **envp =NULL ;char *buffer =NULL ; +char *seq_buff =NULL ;char *scratch;int i =0 ;int retval;
@@ -229,6 +230,9 @@ buffer =kmalloc (BUFFER_SIZE ,GFP_KERNEL );if (!buffer) goto exit; + seq_buff =kmalloc (BUFFER_SIZE ,GFP_KERNEL ); +if (!seq_buff) + goto exit;if (hotplug_ops->name) name = hotplug_ops->name (kset, kobj);
@@ -258,6 +262,9 @@ envp [i++] = scratch; scratch +=sprintf (scratch,"SUBSYSTEM=%s" , name) +1 ; +/* point to the buffer, but don't fill it until after the hotplug call */ + envp [i++] = seq_buff; +if (hotplug_ops->hotplug) {/* have the kset specific function add its stuff */ retval = hotplug_ops->hotplug (kset, kobj,
@@ -273,9 +280,7 @@spin_lock (&sequence_lock); seq = ++hotplug_seqnum;spin_unlock (&sequence_lock); - - envp [i++] = scratch; - scratch +=sprintf (scratch,"SEQNUM=%lld" , (long long )seq) +1 ; +sprintf (seq_buff,"SEQNUM=%lld" , (long long )seq); pr_debug ("%s: %s %s seq=%lld %s %s %s %s %s\n", __FUNCTION__, argv[0], argv[1], (long long)seq,
@@ -293,6 +298,7 @@ exit:kfree (kobj_path); +kfree (seq_buff);kfree (buffer);kfree (envp);return ;
Quoted 20 lines · Greg KH
Quoted 10 lines · Andrew
Quoted 7 lines · Greg KH
Why not just use the same buffer? We should be able to do that.
I'd prefer we use the same buffer. Care to respin your patch?
Quoted 8 lines
Sure, I can only see two ways of achieving that however.
1) Change the API ofkset_hotplug_ops.hotplug()to return the amount of consumed buffer (and possibly an updated value for i (num_envp) and then changing every real function that implements that interface or
2) Spin through the envp[] starting at i toNUM_ENVPlooking for a NULL pointer dropping back 1 (last_used) then do a
scratch +=strlen(envp[last_used])+ 1
Quoted 6 lines
Ick, ok, let's stick with 2 buffers. How about the patch below? It's a bit smaller than yours.
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
Right, especially not the events with the same DEVPATH, like "remove" beating an "add". But I'm not sure if the number isn't useful. Whatever we may do with the hotplug over netlink in the future, we will only have /sbin/hotplug for the early boot and it may be nice to know, what events we have already handled...
Quoted 1 line
I'll hold off on applying this patch until we figure this out...
How about just reserving 20 bytes for the number (u64 will never be more than that), save the pointer to that field, and fill the number in later?
Thanks,
Kay
Quoted 10 lines · Kay Sievers
Quoted 3 lines
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
Right, especially not the events with the same DEVPATH, like "remove" beating an "add". But I'm not sure if the number isn't useful. Whatever we may do with the hotplug over netlink in the future, we will only have
/sbin/hotplugfor the early boot and it may be nice to know, what events we have already handled...
Quoted 5 lines
Quoted 1 line
I'll hold off on applying this patch until we figure this out...
How about just reserving 20 bytes for the number (u64 will never be more than that), save the pointer to that field, and fill the number in later?
Ah, something like this instead? I like it, it's even smaller than the previous patch. Compile tested only...
thanks,
greg k-h
--------
kobject_uevent.c renamed from lib/kobject_uevent.c +7 -3
@@ -182,6 +182,7 @@char *argv [3 ];char **envp =NULL ;char *buffer =NULL ; +char *seq_buff;char *scratch;int i =0 ;int retval;
@@ -258,6 +259,11 @@ envp [i++] = scratch; scratch +=sprintf (scratch,"SUBSYSTEM=%s" , name) +1 ; +/* reserve space for the sequence, +* put the real one in after the hotplug call */ + envp[i++] = seq_buff = scratch; + scratch +=sprintf (scratch,"SEQNUM=12345678901234567890" ) +1 ; +if (hotplug_ops->hotplug) {/* have the kset specific function add its stuff */ retval = hotplug_ops->hotplug (kset, kobj,
@@ -273,9 +279,7 @@spin_lock (&sequence_lock); seq = ++hotplug_seqnum;spin_unlock (&sequence_lock); - - envp [i++] = scratch; - scratch +=sprintf (scratch,"SEQNUM=%lld" , (long long )seq) +1 ; +sprintf (seq_buff,"SEQNUM=%lld" , (long long )seq); pr_debug ("%s: %s %s seq=%lld %s %s %s %s %s\n", __FUNCTION__, argv[0], argv[1], (long long)seq,
Quoted 11 lines · Greg KH
Quoted 10 lines · Kay Sievers
Quoted 3 lines
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
Right, especially not the events with the same DEVPATH, like "remove" beating an "add". But I'm not sure if the number isn't useful. Whatever we may do with the hotplug over netlink in the future, we will only have
/sbin/hotplugfor the early boot and it may be nice to know, what events we have already handled...
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
I'll hold off on applying this patch until we figure this out...
How about just reserving 20 bytes for the number (u64 will never be more than that), save the pointer to that field, and fill the number in later?
Ah, something like this instead? I like it, it's even smaller than the previous patch. Compile tested only...
I like that. How about the following. It will keep the buffer clean from random chars, cause the kevent does not have the vector and relies on the '\0' to separate the strings from each other.
I've tested it. The netlink-hotplug message looks like this:
recv(3, "remove@/class/input/mouse2\0ACTION=remove\0DEVPATH=/class/input/mouse2\0SUBSYSTEM=input\0SEQNUM=961 \0", 1024, 0) = 113
Quoted 13 lines
--------
kobject_uevent.c renamed from lib/kobject_uevent.c +1
-0@@ -182,6 +182,7 @@char *argv [3 ];char **envp =NULL ;char *buffer =NULL ; +char *seq_buff;char *scratch;int i =0 ;int retval;
kobject_uevent.c line 265
Quoted 7 lines
kobject_uevent.c renamed from lib/kobject_uevent.c +4
-0@@ -258,6 +259,11 @@ envp [i++] = scratch; scratch +=sprintf (scratch,"SUBSYSTEM=%s" , name) +1 ; +/* reserve space for the sequence, +* put the real one in after the hotplug call */ + envp[i++] = seq_buff = scratch; + scratch +=sprintf (scratch,"SEQNUM=12345678901234567890" ) +1 ;
+ scratch += strlen("SEQNUM=12345678901234567890") + 1;
kobject_uevent.c line 282
Quoted 12 lines
kobject_uevent.c renamed from lib/kobject_uevent.c +2
-3+if (hotplug_ops->hotplug) {/* have the kset specific function add its stuff */ retval = hotplug_ops->hotplug (kset, kobj,@@ -273,9 +279,7 @@spin_lock (&sequence_lock); seq = ++hotplug_seqnum;spin_unlock (&sequence_lock); - - envp [i++] = scratch; - scratch +=sprintf (scratch,"SEQNUM=%lld" , (long long )seq) +1 ; +sprintf (seq_buff,"SEQNUM=%lld" , (long long )seq);
+ sprintf(seq_buff, "SEQNUM=%-20lld", (long long)seq);
Quoted 3 lines
kobject_uevent.c renamed from lib/kobject_uevent.c +0
-0pr_debug ("%s: %s %s seq=%lld %s %s %s %s %s\n", __FUNCTION__, argv[0], argv[1], (long long)seq,
Quoted 21 lines · Kay Sievers
Quoted 11 lines · Greg KH
Quoted 10 lines · Kay Sievers
Quoted 3 lines
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
Right, especially not the events with the same DEVPATH, like "remove" beating an "add". But I'm not sure if the number isn't useful. Whatever we may do with the hotplug over netlink in the future, we will only have
/sbin/hotplugfor the early boot and it may be nice to know, what events we have already handled...Quoted 8 lines
Quoted 5 lines
Quoted 1 line
I'll hold off on applying this patch until we figure this out...
How about just reserving 20 bytes for the number (u64 will never be more than that), save the pointer to that field, and fill the number in later?
Ah, something like this instead? I like it, it's even smaller than the previous patch. Compile tested only...
Quoted 6 lines
I like that. How about the following. It will keep the buffer clean from random chars, cause the kevent does not have the vector and relies on the '\0' to separate the strings from each other.
I've tested it. The netlink-hotplug message looks like this:recv(3, "remove@/class/input/mouse2\0ACTION=remove\0DEVPATH=/class/input/mouse2\0SUBSYSTEM=input\0SEQNUM=961 \0", 1024, 0) = 113
Hmm, these trailing spaces are just bad, sorry. I'll better pass the envp array over to send_uevent() and clean up the keys while copying the env values into the skb buffer. This will make the event payload more safe too. So your first version looks better.
Thanks,
Kay
Quoted 27 lines · Kay Sievers
Quoted 26 lines · Kay Sievers
Quoted 11 lines · Greg KH
Quoted 10 lines · Kay Sievers
Quoted 3 lines
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
Right, especially not the events with the same DEVPATH, like "remove" beating an "add". But I'm not sure if the number isn't useful. Whatever we may do with the hotplug over netlink in the future, we will only have
/sbin/hotplugfor the early boot and it may be nice to know, what events we have already handled...Quoted 8 lines
Quoted 5 lines
Quoted 1 line
I'll hold off on applying this patch until we figure this out...
How about just reserving 20 bytes for the number (u64 will never be more than that), save the pointer to that field, and fill the number in later?
Ah, something like this instead? I like it, it's even smaller than the previous patch. Compile tested only...
I like that. How about the following. It will keep the buffer clean from random chars, cause the kevent does not have the vector and relies on the '\0' to separate the strings from each other.
I've tested it. The netlink-hotplug message looks like this:
Quoted 6 lines
Quoted 1 line
recv(3, "remove@/class/input/mouse2\0ACTION=remove\0DEVPATH=/class/input/mouse2\0SUBSYSTEM=input\0SEQNUM=961 \0", 1024, 0) = 113
Hmm, these trailing spaces are just bad, sorry. I'll better pass the envp array over to
send_uevent()and clean up the keys while copying the env values into the skb buffer. This will make the event payload more safe too. So your first version looks better.
How about this? We copy over key by key into the skb buffer and the netlink message can get the envp array without depending on a single continuous buffer.
The netlink message looks nice like this now:
recv(3, "
add@/devices/pci0000:00/0000:00:1d.1/usb3/3-2/3-2:1.0\0 HOME=/\0
PATH=/sbin:/bin:/usr/sbin:/usr/bin\0
ACTION=add\0 DEVPATH=/devices/pci0000:00/0000:00:1d.1/usb3/3-2/3-2:1.0\0 SUBSYSTEM=usb\0
SEQNUM=991\0
DEVICE=/proc/bus/usb/003/008\0 PRODUCT=46d/c03e/2000\0
TYPE=0/0/0\0
INTERFACE=3/1/2\0
", 1024, 0) = 268
===== lib/kobject_uevent.c 1.10 vs edited =====
lib/kobject_uevent.c +34 -17
@@ -23,6 +23,9 @@#include <linux/kobject.h> #include <net/sock.h> +#define BUFFER_SIZE 1024/* buffer for the hotplug env */ +#define NUM_ENVP 32/* number of env pointers */ +#if defined(CONFIG_KOBJECT_UEVENT ) || defined(CONFIG_HOTPLUG )static char *action_to_string (enum kobject_action action) {
@@ -57,28 +60,40 @@ static struct sock *uevent_sock;* @buflen: * gfp_mask: */ -static int send_uevent (const char *signal,const char *obj,const void *buf, -int buflen,int gfp_mask) +static int send_uevent (const char *signal,const char *obj, +char **envp,int gfp_mask) {struct sk_buff *skb;char *pos; -int len; +int i; +int skblen, linelen;if (!uevent_sock)return -EIO ; - len =strlen (signal) +1 ; - len +=strlen (obj) +1 ; - len += buflen; + skblen =strlen (signal) +1 ; + skblen +=strlen (obj) +1 ; - skb =alloc_skb (len, gfp_mask); +/* allocate buffer with maximum message size */ + skb =alloc_skb (skblen +BUFFER_SIZE , gfp_mask);if (!skb)return -ENOMEM ; - pos =skb_put (skb, len); + pos =skb_put (skb, skblen +BUFFER_SIZE ); pos +=sprintf (pos,"%s@%s" , signal, obj) +1 ; -memcpy (pos, buf, buflen); + +/* copy the environment key by key to our '\0' delimited buffer */ +if (envp) { +for (i =0 ; envp[i]; i++) { + linelen =strlcpy (pos, envp[i],BUFFER_SIZE ) +1 ; + skblen += linelen; + pos += linelen; + } + } + +/* trim skb buffer to the actual used size */ +skb_trim (skb, skblen);return netlink_broadcast (uevent_sock, skb,0 ,1 , gfp_mask); }
@@ -107,10 +122,10 @@ static int do_kobject_uevent(struct kobjif (!attrpath) goto exit;sprintf (attrpath,"%s/%s" , path, attr->name); - rc =send_uevent (signal, attrpath,NULL ,0 , gfp_mask); + rc =send_uevent (signal, attrpath,NULL , gfp_mask);kfree (attrpath); }else { - rc =send_uevent (signal, path,NULL ,0 , gfp_mask); + rc =send_uevent (signal, path,NULL , gfp_mask); } exit:
@@ -169,8 +184,6 @@ static inline int send_uevent(const charu64 hotplug_seqnum;static spinlock_t sequence_lock =SPIN_LOCK_UNLOCKED ; -#define BUFFER_SIZE 1024/* should be enough memory for the env */ -#define NUM_ENVP 32/* number of env pointers */ /** * kobject_hotplug - notify userspace by executing /sbin/hotplug *
@@ -182,6 +195,7 @@ void kobject_hotplug(struct kobject *kobchar *argv [3 ];char **envp =NULL ;char *buffer =NULL ; +char *seq_buff;char *scratch;int i =0 ;int retval;
@@ -258,6 +272,11 @@ void kobject_hotplug(struct kobject *kob envp [i++] = scratch; scratch +=sprintf (scratch,"SUBSYSTEM=%s" , name) +1 ; +/* reserve space for the sequence, +* put the real one in after the hotplug call */ + envp[i++] = seq_buff = scratch; + scratch +=strlen ("SEQNUM=18446744073709551616" ) +1 ; +if (hotplug_ops->hotplug) {/* have the kset specific function add its stuff */ retval = hotplug_ops->hotplug (kset, kobj,
@@ -273,15 +292,13 @@ void kobject_hotplug(struct kobject *kobspin_lock (&sequence_lock); seq = ++hotplug_seqnum;spin_unlock (&sequence_lock); - - envp [i++] = scratch; - scratch +=sprintf (scratch,"SEQNUM=%lld" , (long long )seq) +1 ; +sprintf (seq_buff,"SEQNUM=%lld" , (long long )seq);pr_debug ("%s: %s %s seq=%lld %s %s %s %s %s\n" , __FUNCTION__, argv[0 ], argv[1 ], (long long )seq, envp[0 ], envp[1 ], envp[2 ], envp[3 ], envp[4 ]); -send_uevent (action_string, kobj_path, buffer, scratch - buffer,GFP_KERNEL ); +send_uevent (action_string, kobj_path, envp,GFP_KERNEL );if (!hotplug_path[0 ]) goto exit;
Quoted 47 lines · Kay Sievers
Quoted 31 lines · Kay Sievers
Quoted 29 lines · Kay Sievers
Quoted 13 lines · Greg KH
Quoted 11 lines · Kay Sievers
Quoted 3 lines
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
Right, especially not the events with the same DEVPATH, like "remove" beating an "add". But I'm not sure if the number isn't useful. Whatever we may do with the hotplug over netlink in the future, we will only have
/sbin/hotplugfor the early boot and it may be nice to know, what events we have already handled...Quoted 8 lines
Quoted 5 lines
Quoted 1 line
I'll hold off on applying this patch until we figure this out...
How about just reserving 20 bytes for the number (u64 will never be more than that), save the pointer to that field, and fill the number in later?
Ah, something like this instead? I like it, it's even smaller than the previous patch. Compile tested only...
I like that. How about the following. It will keep the buffer clean from random chars, cause the kevent does not have the vector and relies on the '\0' to separate the strings from each other.
I've tested it. The netlink-hotplug message looks like this:Quoted 6 lines
Quoted 1 line
recv(3, "remove@/class/input/mouse2\0ACTION=remove\0DEVPATH=/class/input/mouse2\0SUBSYSTEM=input\0SEQNUM=961 \0", 1024, 0) = 113
Hmm, these trailing spaces are just bad, sorry. I'll better pass the envp array over to
send_uevent()and clean up the keys while copying the env values into the skb buffer. This will make the event payload more safe too. So your first version looks better.How about this? We copy over key by key into the skb buffer and the netlink message can get the envp array without depending on a single continuous buffer.
The netlink message looks nice like this now:
Quoted 14 lines
recv(3, "
add@/devices/pci0000:00/0000:00:1d.1/usb3/3-2/3-2:1.0\0 HOME=/\0
PATH=/sbin:/bin:/usr/sbin:/usr/bin\0
ACTION=add\0 DEVPATH=/devices/pci0000:00/0000:00:1d.1/usb3/3-2/3-2:1.0\0 SUBSYSTEM=usb\0
SEQNUM=991\0
DEVICE=/proc/bus/usb/003/008\0 PRODUCT=46d/c03e/2000\0
TYPE=0/0/0\0
INTERFACE=3/1/2\0
", 1024, 0) = 268
Yeah, sending the envp array instead of buffer to send_uevent() handles all the cases. Great stuff - Thanks
Andrew
Quoted 41 lines · Kay Sievers
Quoted 27 lines · Kay Sievers
Quoted 26 lines · Kay Sievers
Quoted 11 lines · Greg KH
Quoted 10 lines · Kay Sievers
Quoted 3 lines
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
Right, especially not the events with the same DEVPATH, like "remove" beating an "add". But I'm not sure if the number isn't useful. Whatever we may do with the hotplug over netlink in the future, we will only have
/sbin/hotplugfor the early boot and it may be nice to know, what events we have already handled...Quoted 8 lines
Quoted 5 lines
Quoted 1 line
I'll hold off on applying this patch until we figure this out...
How about just reserving 20 bytes for the number (u64 will never be more than that), save the pointer to that field, and fill the number in later?
Ah, something like this instead? I like it, it's even smaller than the previous patch. Compile tested only...
I like that. How about the following. It will keep the buffer clean from random chars, cause the kevent does not have the vector and relies on the '\0' to separate the strings from each other.
I've tested it. The netlink-hotplug message looks like this:Quoted 6 lines
Quoted 1 line
recv(3, "remove@/class/input/mouse2\0ACTION=remove\0DEVPATH=/class/input/mouse2\0SUBSYSTEM=input\0SEQNUM=961 \0", 1024, 0) = 113
Hmm, these trailing spaces are just bad, sorry. I'll better pass the envp array over to
send_uevent()and clean up the keys while copying the env values into the skb buffer. This will make the event payload more safe too. So your first version looks better.How about this? We copy over key by key into the skb buffer and the netlink message can get the envp array without depending on a single continuous buffer.
The netlink message looks nice like this now:
Quoted 13 lines
recv(3, "
add@/devices/pci0000:00/0000:00:1d.1/usb3/3-2/3-2:1.0\0 HOME=/\0
PATH=/sbin:/bin:/usr/sbin:/usr/bin\0
ACTION=add\0 DEVPATH=/devices/pci0000:00/0000:00:1d.1/usb3/3-2/3-2:1.0\0 SUBSYSTEM=usb\0
SEQNUM=991\0
DEVICE=/proc/bus/usb/003/008\0 PRODUCT=46d/c03e/2000\0
TYPE=0/0/0\0
INTERFACE=3/1/2\0
", 1024, 0) = 268
Here is an improved version that uses skb_put() to fill the skb buffer, instead of trimming the buffer to the final size after we've copied over all keys.
Thanks,
Kay
===== lib/kobject_uevent.c 1.10 vs edited =====
lib/kobject_uevent.c +30 -17
@@ -23,6 +23,9 @@#include <linux/kobject.h> #include <net/sock.h> +#define BUFFER_SIZE 1024/* buffer for the hotplug env */ +#define NUM_ENVP 32/* number of env pointers */ +#if defined(CONFIG_KOBJECT_UEVENT ) || defined(CONFIG_HOTPLUG )static char *action_to_string (enum kobject_action action) {
@@ -53,12 +56,11 @@ static struct sock *uevent_sock;* * @signal: signal name * @obj: object path (kobject) -* @buf: buffer used to pass auxiliary data like the hotplug environment -* @buflen: -* gfp_mask: +* @envp: possible hotplug environment to pass with the message +* @gfp_mask: */ -static int send_uevent (const char *signal,const char *obj,const void *buf, -int buflen,int gfp_mask) +static int send_uevent (const char *signal,const char *obj, +char **envp,int gfp_mask) {struct sk_buff *skb;char *pos;
@@ -69,16 +71,25 @@ static int send_uevent(const char *signa len =strlen (signal) +1 ; len +=strlen (obj) +1 ; - len += buflen; - skb =alloc_skb (len, gfp_mask); +/* allocate buffer with the maximum possible message size */ + skb =alloc_skb (len +BUFFER_SIZE , gfp_mask);if (!skb)return -ENOMEM ; pos =skb_put (skb, len); +sprintf (pos,"%s@%s" , signal, obj); - pos +=sprintf (pos,"%s@%s" , signal, obj) +1 ; -memcpy (pos, buf, buflen); +/* copy the environment key by key to our continuous buffer */ +if (envp) { +int i; + +for (i =2 ; envp[i]; i++) { + len =strlen (envp[i]) +1 ; + pos =skb_put (skb, len); +strcpy (pos, envp[i]); + } + }return netlink_broadcast (uevent_sock, skb,0 ,1 , gfp_mask); }
@@ -107,10 +118,10 @@ static int do_kobject_uevent(struct kobjif (!attrpath) goto exit;sprintf (attrpath,"%s/%s" , path, attr->name); - rc =send_uevent (signal, attrpath,NULL ,0 , gfp_mask); + rc =send_uevent (signal, attrpath,NULL , gfp_mask);kfree (attrpath); }else { - rc =send_uevent (signal, path,NULL ,0 , gfp_mask); + rc =send_uevent (signal, path,NULL , gfp_mask); } exit:
@@ -169,8 +180,6 @@ static inline int send_uevent(const charu64 hotplug_seqnum;static spinlock_t sequence_lock =SPIN_LOCK_UNLOCKED ; -#define BUFFER_SIZE 1024/* should be enough memory for the env */ -#define NUM_ENVP 32/* number of env pointers */ /** * kobject_hotplug - notify userspace by executing /sbin/hotplug *
@@ -182,6 +191,7 @@ void kobject_hotplug(struct kobject *kobchar *argv [3 ];char **envp =NULL ;char *buffer =NULL ; +char *seq_buff;char *scratch;int i =0 ;int retval;
@@ -258,6 +268,11 @@ void kobject_hotplug(struct kobject *kob envp [i++] = scratch; scratch +=sprintf (scratch,"SUBSYSTEM=%s" , name) +1 ; +/* reserve space for the sequence, +* put the real one in after the hotplug call */ + envp[i++] = seq_buff = scratch; + scratch +=strlen ("SEQNUM=18446744073709551616" ) +1 ; +if (hotplug_ops->hotplug) {/* have the kset specific function add its stuff */ retval = hotplug_ops->hotplug (kset, kobj,
@@ -273,15 +288,13 @@ void kobject_hotplug(struct kobject *kobspin_lock (&sequence_lock); seq = ++hotplug_seqnum;spin_unlock (&sequence_lock); - - envp [i++] = scratch; - scratch +=sprintf (scratch,"SEQNUM=%lld" , (long long )seq) +1 ; +sprintf (seq_buff,"SEQNUM=%lld" , (long long )seq);pr_debug ("%s: %s %s seq=%lld %s %s %s %s %s\n" , __FUNCTION__, argv[0 ], argv[1 ], (long long )seq, envp[0 ], envp[1 ], envp[2 ], envp[3 ], envp[4 ]); -send_uevent (action_string, kobj_path, buffer, scratch - buffer,GFP_KERNEL ); +send_uevent (action_string, kobj_path, envp,GFP_KERNEL );if (!hotplug_path[0 ]) goto exit;
Quoted 56 lines · Kay Sievers
Quoted 41 lines · Kay Sievers
Quoted 27 lines · Kay Sievers
Quoted 26 lines · Kay Sievers
Quoted 11 lines · Greg KH
Quoted 10 lines · Kay Sievers
Quoted 3 lines
But there might still be a problem. With this change, the sequence number is not sent out the kevent message. Kay, do you think this is an issue? I don't think we can get netlink messages out of order, right?
Right, especially not the events with the same DEVPATH, like "remove" beating an "add". But I'm not sure if the number isn't useful. Whatever we may do with the hotplug over netlink in the future, we will only have
/sbin/hotplugfor the early boot and it may be nice to know, what events we have already handled...Quoted 8 lines
Quoted 5 lines
Quoted 1 line
I'll hold off on applying this patch until we figure this out...
How about just reserving 20 bytes for the number (u64 will never be more than that), save the pointer to that field, and fill the number in later?
Ah, something like this instead? I like it, it's even smaller than the previous patch. Compile tested only...
I like that. How about the following. It will keep the buffer clean from random chars, cause the kevent does not have the vector and relies on the '\0' to separate the strings from each other.
I've tested it. The netlink-hotplug message looks like this:Quoted 6 lines
Quoted 1 line
recv(3, "remove@/class/input/mouse2\0ACTION=remove\0DEVPATH=/class/input/mouse2\0SUBSYSTEM=input\0SEQNUM=961 \0", 1024, 0) = 113
Hmm, these trailing spaces are just bad, sorry. I'll better pass the envp array over to
send_uevent()and clean up the keys while copying the env values into the skb buffer. This will make the event payload more safe too. So your first version looks better.How about this? We copy over key by key into the skb buffer and the netlink message can get the envp array without depending on a single continuous buffer.
The netlink message looks nice like this now:
Quoted 13 lines
recv(3, "
add@/devices/pci0000:00/0000:00:1d.1/usb3/3-2/3-2:1.0\0 HOME=/\0
PATH=/sbin:/bin:/usr/sbin:/usr/bin\0
ACTION=add\0 DEVPATH=/devices/pci0000:00/0000:00:1d.1/usb3/3-2/3-2:1.0\0 SUBSYSTEM=usb\0
SEQNUM=991\0
DEVICE=/proc/bus/usb/003/008\0 PRODUCT=46d/c03e/2000\0
TYPE=0/0/0\0
INTERFACE=3/1/2\0
", 1024, 0) = 268
Quoted 3 lines
Here is an improved version that uses
skb_put()to fill the skb buffer, instead of trimming the buffer to the final size after we've copied over all keys.
Thanks, I've applied this version. Hopefully I'll get it out to Linus later today.
greg k-h