Re: [PATCH] ipvs: change ip_vs_conn_tab_bits range to [8,31]

[Julian Anastasov]

	Hello,

On Thu, 13 Apr 2023, Abhijeet Rastogi wrote:

> Hi Simon, Andrea and Julian,
> 
> I really appreciate you taking the time to respond to my patch. Some follow up
> questions that I'll appreciate a response for.
> 
> @Simon Horman
> >In any case, I think this patch is an improvement on the current situation.
> 
> +1 to this. I wanted to add that, we're not changing the defaults
> here, the default still stays at 2^12. If a kernel user changes the
> default, they probably already know what the limitations are, so I
> personally don't think it is a big concern.
> 
> @Andrea Claudi
> >for the record, RHEL ships with CONFIG_IP_VS_TAB_BITS set to 12 as
> default.
> 
> Sorry, I should have been clearer. RHEL ships with the same default,
> yes, but it doesn't have the range check, at least, on the version I'm
> using right now (3.10.0-1160.62.1.el7.x86_64).
> 
> On this version, I'm able to load with bit size 30, 31 gives me error
> regarding allocating memory (64GB host) and anything beyond 31 is
> mysteriously switched to a lower number. The following dmesg on my
> host confirms that the bitsize 30 worked, which is not possible
> without a patch on the current kernel version.
> 
> "[Fri Apr 14 01:14:51 2023] IPVS: Connection hash table configured (size=1073741
> 824, memory=16777216Kbytes)"
> 
> @Julian Anastasov,
> >This is not a limit of number of connections. I prefer
> not to allow value above 24 without adding checks for the
> available memory,
> 
> Interesting that you brought up that number 24, that is exactly what
> we use in production today. One IPVS node is able to handle spikes of
> 10M active connections without issues. This patch idea originated as
> my company is migrating from the ancient RHEL version to a somewhat
> newer CentOS (5.* kernel) and noticed that we were unable to load the
> ip_vs kernel module with anything greater than 20 bits. Another
> motivation for kernel upgrade is utilizing maglev to reduce table size
> but that's out of context in this discussion.
> 
> My request is, can we increase the range from 20 to something larger?
> If 31 seems a bit excessive, maybe, we can settle for something like
> [8,30] or even lower. With conn_tab_bits=30, it allocates 16GB at
> initialization time, it is not entirely absurd by today's standards.
> 
> I can revise my patch to a lower range as you guys see fit.

	Some 32-bit platforms have a 120MB limit for
vmalloc. 24-bit table on 32-bit box will allocate 64MB.

	One way to solve the problem is to use in Kconfig:

range 8 20 if !64BIT
range 8 27 if 64BIT

	Why 30 and above do not work? Because we store the
size, mask in 'int' which is 32 bits. But also some places do not
allow allocations above INT_MAX, for example, kvmalloc_node().
So, even 28 may not work for 8-byte array items on 64-bit.

	It would be good to check if the provided
value does not exceed some real limits. Here is an example
that assumes IPVS will allocate up to 1/8 of the memory,
8 conns average in a hash row. Such checks should not
exceed the small vmalloc area for 32-bit boxes and also
kvmalloc allows vmalloc with huge pages. This idea is
entirely untested/compiled. These checks apply some
sane thresholds. If you need something above, you are
probably allocating more than needed.

/* This will match the Kconfig range: */
int min = 8;
#if __BITS_PER_LONG > 32
int max = 27;
#else
int max = 20;
#endif

	We can safely use 27 in Kconfig even for 32-bit
due to the below checks, they will clamp it to lower value.

	/* Order of the available memory */
	int max_avail = order_base_2(totalram_pages()) + PAGE_SHIFT;

	We can remove this 'if' check:
	if (ip_vs_conn_tab_bits < 8 || ip_vs_conn_tab_bits > 20) {
		pr_info("conn_tab_bits not in [8, 20]. Using default value\n");
		ip_vs_conn_tab_bits = CONFIG_IP_VS_TAB_BITS;
	}

	max_avail -= 3;				/* ~8 in hash row */
	max_avail -= 3;				/* IPVS up to 1/8 of mem */
	/* The hash table links allocated memory for IPVS conns */
	max_avail -= order_base_2(sizeof(struct ip_vs_conn));
	/* Range should not exceed the available memory */
	max = clamp(max, min, max_avail);
	/* Clamp configured value silently */
	ip_vs_conn_tab_bits = clamp_val(ip_vs_conn_tab_bits, min, max);
	ip_vs_conn_tab_size = 1 << ip_vs_conn_tab_bits;
	ip_vs_conn_tab_mask = ip_vs_conn_tab_size - 1;

	/* Switch to kvmalloc */
	ip_vs_conn_tab = kvmalloc_array(ip_vs_conn_tab_size,
					sizeof(*ip_vs_conn_tab), GFP_KERNEL);

	and use everywhere kvfree(ip_vs_conn_tab);

	For 64GB box the calcs should be:

max_avail = 36 - 3 - 3 - 9 => ip_vs_conn_tab_bits = 21
Allocated hash table: (2^21)*8=16MB
Allocated for IPVS conns (8 cols per row): (2^21)*8*(400..512)=6..8GB
which is ~1/8 of 64GB. All memory will be allocated with
~64 conns per row. May be the above calcs can be changed
to ~4 cols and 1/2 mem to use 128MB (24 bits instead of 21)
for our example: 36 - 2 - 1 - 9 => 24.

	Possible problems if using large table that is
not loaded with enough conns:

- walking the table will cost more cycles, for example,
ip_vs_random_dropentry() wants to walk part of the table
every second. Even normal netns cleanup has to walk it.

- cat /proc/net/ip_vs_conn will be slower

Regards

--
Julian Anastasov <ja@xxxxxx>