[PATCH] powerpc: fix large nvram access

[Arnd Bergmann]

Am Dienstag, 10. Januar 2006 23:13 schrieb Olaf Hering:
>  On Mon, Jan 09, Linux Kernel Mailing List wrote:
> > tree da8f883f72d08f9534e9eca3bba662413b9bd865
> > parent d52771fce4e774fa786097d34412a057d487c697
> > author Arnd Bergmann <arnd at arndb.de> Fri, 09 Dec 2005 19:21:44 +0100
> > committer Paul Mackerras <paulus at samba.org> Mon, 09 Jan 2006 14:53:31
> > +1100
> >
> > [PATCH] powerpc: fix large nvram access
> >
> > /dev/nvram uses the user-provided read/write size
> > for kmalloc, which fails, if a large number is passed.
> > This will always use a single page at most, which
> > can be expected to succeed.
> >
> > Signed-off-by: Arnd Bergmann <arndb at de.ibm.com>
> > Signed-off-by: Paul Mackerras <paulus at samba.org>
> >
> >  arch/powerpc/kernel/nvram_64.c |  114
> > +++++++++++++++++++----------------------
>
> this change breaks my nvsetenv binary on JS20.
>

I think this has uncovered a bug that has been in nvsetenv for a long time.

> open("/dev/nvram", O_RDONLY)            = 3
> read(3, "P\347\0\34of-config\0\0\0", 16) = 16
> lseek(3, 432, SEEK_CUR)                 = 448
> read(3, "p\375\2\0common\0\0\0\0\0\0", 16) = 16
> read(3, "ibm,fw-phandle-enable?=false\0lit"..., 8176) = 4096

This is where my patch changed the behavior of the driver. It used to return all the data in one chunk, now it returns 4096 bytes at most.

> dup(2)                                  = 4
> fcntl64(4, F_GETFL)                     = 0x2 (flags O_RDWR)
> fstat64(4, {st_mode=S_IFCHR|0600, st_rdev=makedev(5, 1), ...}) = 0
> ioctl(4, TCGETS, {B38400 opost isig icanon echo ...}) = 0
> mmap(NULL, 131072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0xf7fc0000 
> _llseek(4, 0, 0xffd63fd8, SEEK_CUR)     = -1 ESPIPE 
> (Illegal seek) 
> write(4, "Error reading /dev/nvram: Succes"..., 34Error reading  /dev/nvram: Success ) = 34
> close(4)                                = 0
> munmap(0xf7fc0000, 131072)              = 0
> exit_group(1)                           = ?
> (none):/#

And then it got here:

|    if (lseek(nvfd, NVSTART, 0) < 0
|        || read(nvfd, &nvbuf, NVSIZE) != NVSIZE) {
|        perror("Error reading /dev/nvram");
|        exit(EXIT_FAILURE);
|    }

I'm pretty sure that the failing _llseek is part of the perror 
implementation and not the real problem, as we were first thinking.

The problem is that nvsetenv bails out on a short read, while according 
to the man page, "It is not an error if this number is smaller than the
number of bytes requested". Of course, character devices often differ
in their behavior from what is in the man pages for file I/O.

I'd suggest we change both ends: nvsetenv should really be able to deal
with a short read, and we can double the size of the kmalloc buffer
for the kernel implementation. Please test the patch below.

---

Subject: powerpc: fix nvsetenv regression with fixed nvram read/write

reading from /dev/nvram was recently fixed to not try to allocate the
user-defined amount of kernel memory. This broke the nvsetenv tool
which relied on never getting short reads from /dev/nvram.

Signed-off-by: Arnd Bergmann <arnd at arndb.de>

diff --git a/arch/powerpc/kernel/nvram_64.c b/arch/powerpc/kernel/nvram_64.c
index fd7db8d..8da6e57 100644
--- a/arch/powerpc/kernel/nvram_64.c
+++ b/arch/powerpc/kernel/nvram_64.c
@@ -34,6 +34,15 @@
 
 #undef DEBUG_NVRAM
 
+/*
+ * Old nvsetenv versions expect us to be able to return 8kb
+ * of data in a single read, so use that as a limit for short
+ * reads.
+ * We can't simply use the requested size because large
+ * kmalloc allocations often fail.
+ */
+#define NVRAM_ALLOC_SIZE 8192
+
 static int nvram_scan_partitions(void);
 static int nvram_setup_partition(void);
 static int nvram_create_os_partition(void);
@@ -94,7 +103,7 @@ static ssize_t dev_nvram_read(struct fil
 		goto out;
 
 	count = min_t(size_t, count, size - *ppos);
-	count = min(count, PAGE_SIZE);
+	count = min(count, NVRAM_ALLOC_SIZE);
 
 	ret = -ENOMEM;
 	tmp = kmalloc(count, GFP_KERNEL);
@@ -131,7 +140,7 @@ static ssize_t dev_nvram_write(struct fi
 		goto out;
 
 	count = min_t(size_t, count, size - *ppos);
-	count = min(count, PAGE_SIZE);
+	count = min(count, NVRAM_ALLOC_SIZE);
 
 	ret = -ENOMEM;
 	tmp = kmalloc(count, GFP_KERNEL);