JFFS2 & SMP

[Artem B. Bityuckiy]

Hello Estelle.

It seems I now what is the problem. Thank you for your message.

The wbuf is protected by the alloc_sem because of all the writes go 
through the space reservation (or in case of GC, it also holds the 
alloc_sem).

When JFFS2 reads the flash, it also looks to the write buffer and if the 
NAND page which should be read is currently in the wbuf, it reads some 
data from the wbuf too. But there is no any protection there.

I've introduced the new read/write semaphore (wbuf_sem). It seems the 
problem is fixed, but I'm not sure yet.

Please, could you try the attached patch? The patch was made against the 
MTD snapshot of date 20041008 
(ftp://ftp.uk.linux.org/pub/people/dwmw2/mtd/cvs/mtd-snapshot-20041008.tar.bz2).

David, could you comment this?

Estelle HAMMACHE wrote:
> Hello Artem,
> 
> Have you made any progress with this issue ?
> 
> I believe this problem could be reproduced more easily
> by adding a long sleep() at the end of __jffs2_flush_wbuf
> between memset() and the affectation of c->wbuf_ofs.
> Then write and fsync a file and try to read it back during
> the sleep.
> I've done something similar on my non-linux setup (actually
> I use a longjmp to exit from __jffs2_flush_wbuf) and I obtain 
> the following when I try to read back the file:
> <7>Node read from 00080160: node_crc ffffffff, calculated CRC f09e7845. dsize ffffffff, csize ffffffff, offset ffffffff,
> buf 1100a48
> <4>Node CRC ffffffff != calculated CRC f09e7845 for node at 00080160
> 
> I think the wbuf variables are well protected against
> cuncurrent writes, however if the wbuf is being flushed 
> (fsync or GC) and you try to _read_ the file which has a 
> node in wbuf, there may be some inconsistency between 
> wbuf_ofs and wbuf contents.
> I suggest to modify the end of __jffs2_flush_wbuf this way:
> 
> 	c->wbuf_len = 0; /* do this FIRST so that reads won't attempt to use wbuf */
> 	memset(c->wbuf,0xff,c->wbuf_pagesize);
> 	/* adjust write buffer offset, else we get a non contiguous write bug */
> 	c->wbuf_ofs += c->wbuf_pagesize;
> 
> Does this seem useful ?
> Estelle
> 
> "Artem B. Bityuckiy" wrote:
> 
>>Hello,
>>
>>how the the wbuf and related variables in the jffs2_sb_info are protected?
>>
>>--
>>Best Regards,
>>Artem B. Bityuckiy,
>>St.-Petersburg, Russia.
>>
>>______________________________________________________
>>Linux MTD discussion mailing list
>>http://lists.infradead.org/mailman/listinfo/linux-mtd/
> 
> 

-- 
Best Regards,
Artem B. Bityuckiy,
St.-Petersburg, Russia.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: preempt-patch.diff
Type: text/x-patch
Size: 16621 bytes
Desc: not available
Url : http://lists.infradead.org/pipermail/linux-mtd/attachments/20041103/1d7bd682/attachment.bin