Would the father of init_mem_lth please stand up
Really someone needs slapping across. What kind of code is that (in 2.5.1-pre6):
drivers/scsi/sd.c:sd_init()
/* allocate memory */
#define init_mem_lth(x,n) x = kmalloc((n) * sizeof(*x), GFP_ATOMIC)
#define zero_mem_lth(x,n) memset(x, 0, (n) * sizeof(*x))
init_mem_lth(rscsi_disks, sd_template.dev_max);
init_mem_lth(sd_sizes, maxparts);
init_mem_lth(sd_blocksizes, maxparts);
init_mem_lth(sd, maxparts);
init_mem_lth(sd_gendisks, N_USED_SD_MAJORS);
init_mem_lth(sd_max_sectors, sd_template.dev_max << 4);
if (!rscsi_disks || !sd_sizes || !sd_blocksizes || !sd || !sd_gendisks)
goto cleanup_mem;#undef init_mem_lth
#undef zero_mem_lth
.....................cleanup_mem:
kfree(sd_gendisks);
kfree(sd);
kfree(sd_blocksizes);
kfree(sd_sizes);
kfree(rscsi_disks);However, it's not only about the puking and keyboard cleanups. The code is buggy as well. Scenario:
0. User inserts a large number of FC-AL adapters with 56 disks each
1. modprobe sd_mod
No SCSI hosts, sd_init() is NOT called.
2. modprobe qla_something
sd_init is called and fails on sd_gendisks. modprobe fails. sd_sizes, sd_blocksizes, etc. are LEFT DANGLING
3. modprobe qla_something
sd_init is called and fails on sd_sizes.
kfree is called with a bunch of dangling pointers
I stringly urge Linus to drop this so-called "cleanup" from 2.5.1.
No doubt, the existing code was bad. I fixed it somewhat for 2.4, and am feeding it to Marcelo. I can forward-port that to 2.5 if anyone is interested.
-- Pete
Quoted 2 lines · Pete Zaitcev
Really someone needs slapping across. What kind of code is that (in 2.5.1-pre6):
It went in with the bio stuff, without being too much of a finger-pointer I think it is Andries.
Quoted 3 lines
No doubt, the existing code was bad. I fixed it somewhat for 2.4, and am feeding it to Marcelo. I can forward-port that to 2.5 if anyone is interested.
Please, just fix it in 2.5 too.
Signature · 1 line
Jens Axboe
Quoted 2 lines
Really someone needs slapping across. What kind of code is that (in 2.5.1-pre6):
Whoever merged that crap also wants a good kicking. First we have the joke ps/2 merge, now this. Is Linus trying to force someone to take over or is small children combined with a pending christmas really the doom that some folks claim 8)
Quoted 5 lines
I stringly urge Linus to drop this so-called "cleanup" from 2.5.1.
No doubt, the existing code was bad. I fixed it somewhat for 2.4, and am feeding it to Marcelo. I can forward-port that to 2.5 if anyone is interested.
Did you clean it up as far as making the disks an array of pointers not of objects ?
Alan
Quoted 29 lines · Pete Zaitcev
drivers/scsi/sd.c:sd_init()
/* allocate memory */#define
init_mem_lth(x,n)x =kmalloc((n) * sizeof(*x), GFP_ATOMIC)
#definezero_mem_lth(x,n)memset(x, 0, (n) * sizeof(*x))init_mem_lth(rscsi_disks, sd_template.dev_max); init_mem_lth(sd_sizes, maxparts); init_mem_lth(sd_blocksizes, maxparts); init_mem_lth(sd, maxparts); init_mem_lth(sd_gendisks, N_USED_SD_MAJORS); init_mem_lth(sd_max_sectors, sd_template.dev_max << 4); if (!rscsi_disks || !sd_sizes || !sd_blocksizes || !sd || !sd_gendisks) goto cleanup_mem;#undef
init_mem_lth
#undefzero_mem_lth
.....................cleanup_mem:kfree(sd_gendisks); kfree(sd); kfree(sd_blocksizes); kfree(sd_sizes); kfree(rscsi_disks);However, it's not only about the puking and keyboard cleanups. The code is buggy as well. Scenario:
Quoted 9 lines
0. User inserts a large number of FC-AL adapters with 56 disks each
1. modprobesd_mod
No SCSI hosts,sd_init()is NOT called.
2. modprobeqla_something
sd_initis called and fails onsd_gendisks. modprobe fails.sd_sizes,sd_blocksizes, etc. are LEFT DANGLING
3. modprobeqla_something
sd_initis called and fails onsd_sizes.
kfree is called with a bunch of dangling pointers
I agree it's ugly, but why would kfree be called on dangling pointers? All those pointers are initialized again on every sd_init() call. And there are no "goto cleanup_mem"s between the kmallocs.
Eric
Quoted 5 lines · Eric Lammerts
I agree it's ugly, but why would kfree be called on dangling pointers? All those pointers are initialized again on every
sd_init()call. And there are no "goto cleanup_mem"s between the kmallocs.Eric
I was too rush - indeed there are no gotos between kmallocs. So the scenario that I presented cannot happen.
-- Pete
Quoted 8 lines
Quoted 3 lines
No doubt, the existing code was bad. I fixed it somewhat for 2.4, and am feeding it to Marcelo. I can forward-port that to 2.5 if anyone is interested.
Did you clean it up as far as making the disks an array of pointers not of objects ?
Alan
I was going to do it, but have not done yet.
BTW, it was pointed out that my analyzis of the init_mem_lth was completely wrong - there is no bug that I thought is there. So I better be careful with 2.4 changes.
-- Pete
Quoted 3 lines
BTW, it was pointed out that my analyzis of the
init_mem_lthwas completely wrong - there is no bug that I thought is there. So I better be careful with 2.4 changes.
Its merely hideously ugly and without taste. I didn't think there was anything else wrong with it 8)