On 2017/9/12 17:42, Linus Walleij wrote:
On Fri, Sep 8, 2017 at 4:51 AM, Shawn Lin <shawn.lin@xxxxxxxxxxxxxx> wrote:
On 2017/9/8 4:02, Linus Walleij wrote:
On Thu, Sep 7, 2017 at 9:18 AM, Ulf Hansson <ulf.hansson@xxxxxxxxxx>
wrote:
Even if this fixes the problem it seems like we are papering over the
real issue, which earlier fixes also did during the release cycle for
v4.13.
I think this is the real solution to the issue.
Another unrelated issue with mmc_init_request() is that
mmc_exit_request()
is not called if mmc_init_request() fails, which means
mmc_init_request()
must free anything it allocates when it fails.
Yes, the situations it's just too fragile. We need to fix the behavior
properly, although I haven't myself been able to investigate exactly
how yet.
Adding, Linus, perhaps he has some ideas.
Maybe we should simply bite the bullet and do what was suggested
by another contributor when I refactored the bounce buffer handling:
simply delete the bounce buffer code and let any remaining (few?)
legacy devices suffer a bit (performancewise) at the gain of way
simpler code?
Are you in the same page with what Adrian pointed to?
IIUC, the issue is:
init_rq_fn will be called each time when trying to create new reuqest
from the pre-allocated request_list memeory pool, and exit_rq_fn will is
in the corresponding routine to free request from request_list
(blk_free_request) when finished. But if alloc_request_size fails, it
won't call exit_rq_fn, so you need to prevent memory leak on your own
error path of init_rq_fn.
Yes and that is what the current patch fixes, is it not?
Nope. It fixed the *out-of-bound* memory usage as request queue will
hold 4 requests by default and only use these four requests when meeting
memory pressure. But the code didn't place the correct bouncesz so that
the 4 pre-allocated requests didn't have valid memory page when
allocated. But in runtime, normally the request queue asks for new
request from request list by dynamically allocating it. So your
init_rq_fn will be called each time. However the mmc_init_request
didn't handle the error path properly。
I simply add SDHCI_QUIRK_BROKEN_ADMA for my SDHCI to force it use
SDMA so that the host->max_segs should be 1. Then add some a hack to
simulate some random failure for mmc_alloc_sg and then after some iozone
stress test, the memory is exhausted finally.
+static u64 skip = 0;
static struct scatterlist *mmc_alloc_sg(int sg_len, gfp_t gfp)
{
struct scatterlist *sg;
+ u32 random = 0;
+
+ /* Simply skip the bootup stage */
+ if (skip++ >= 0x800)
+ random = get_random_int();
+
+ if (unlikely((random & 0xf) > 0xd)) {
+ printk("mmc_alloc_sg: make fake failure, random =
0x%x\n", random);
+ return NULL;
+ }
[ 540.507195] iozone invoked oom-killer:
gfp_mask=0x16040c0(GFP_KERNEL|__GFP_COMP|__GFP_NOTRACK),
nodemask=(null), order=0, oom_score_adj=0
[ 540.508302] iozone cpuset=/ mems_allowed=0
[ 540.508727] CPU: 2 PID: 3039 Comm: iozone Not tainted
4.13.0-next-20170912-00003-g01cc0dd5-dirty #110
[ 540.509537] Hardware name: Firefly-RK3399 Board (DT)
[ 540.509977] Call trace:
[ 540.510221] [<ffff20000808b3d0>] dump_backtrace+0x0/0x480
[ 540.510707] [<ffff20000808b864>] show_stack+0x14/0x20
[ 540.511164] [<ffff200008dfbabc>] dump_stack+0xa4/0xc8
[ 540.511621] [<ffff2000081fc744>] dump_header+0xc4/0x2e8
[ 540.512091] [<ffff2000081fb888>] oom_kill_process+0x3a8/0x728
[ 540.512605] [<ffff2000081fc090>] out_of_memory+0x1a8/0x6d8
[ 540.513099] [<ffff200008203ea8>] __alloc_pages_nodemask+0xef8/0xf80
[ 540.513660] [<ffff200008275d6c>] alloc_pages_current+0x9c/0x128
[ 540.514191] [<ffff200008281f60>] new_slab+0x488/0x5d8
[ 540.514645] [<ffff200008284198>] ___slab_alloc+0x378/0x5d0
[ 540.515138] [<ffff200008284414>] __slab_alloc.isra.23+0x24/0x38
[ 540.515668] [<ffff200008284d44>] kmem_cache_alloc+0x1ec/0x218
[ 540.516183] [<ffff20000830dc30>] __blockdev_direct_IO+0x220/0x4a00
But you seem to talk about removing the bounce buffer and so finally
get rid of registering init_rq_fn/exit_rq_fn?
That is in direct response to Ulf's statement
"the situations it's just too fragile" and what can be done about
that is not make the code even more complex but make it
simpler and easier to follow.
One way to achieve that in the long term, is to delete bounce
buffer handling since it adds overhead and complexity.
That is another thing,
and what we right now need to do is to fix the pontential memory leak.
It's quite a simple action, right? :)
It is another thing.
This patch fixes a memory leak, but Ulf is asking how we may
avoid fragile behaviour.
I am a bit hesitant about that because Pierre Ossman said it was
actually a big win on the SDHC hosts that made use of it at one
point.
You had removed packed cmd support to simplify the code, so I think
this is another trade-off need to ask: What you want? and keep
consistent with the direction you insisted on.
Packed command could be removed because it was not used by
any in-tree code. See
commit 03d640ae1f9b24b1d2a11f747143a1ecc0745019
Ok, I was surprised that no any in-tree host enabled it, but if some
host did, the situation was similar with this one.
Bounce buffers on the other hand, as this patch shows, is very much
in use. So if they e.g. improve throughput on these systems
(mainly laptops I think?) it should definately be kept around.
It might be a good idea to make a patch to remove the bounce
buffers and ask people to do before/after throughput tests,
because I do not have this hardware myself. If it doesn't provide
any throughput improvements to any system in use, it should
be deleted. But I don't know that yet.
yeap, sounds good but it's up to Ulf.
Yours,
Linus Walleij
--
To unsubscribe from this list: send the line "unsubscribe linux-mmc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
