[PATCH] irqchip/gic-v3-its: add allocation max order limitation for lpi_id_bits

On Mon, 27 Aug 2018 02:41:04 +0100,
Jia He [off-list ref] wrote:

Hi Marc
Thanks for the comments

On 8/27/2018 3:01 AM, Marc Zyngier Wrote:

quoted

[I'm travelling, so expect some major delays in responding to email]

Hi Jia,

On Sun, 26 Aug 2018 10:00:51 +0100,
Jia He [off-list ref] wrote:

quoted

There is a WARN_ON when my QDF2400 server boots up (pagesize is 4k)

[snip]

quoted

In its_alloc_lpi_tables, lpi_id_bits is 24, without this patch,
its_allocate_prop_table will try to allocate 16M(order 12 if
pagesize=4k). Thus it causes the WARN_ON.

Gah! QDF and its 24bit INTIDs... Making life hell for everyone ;-)

Sorry for breaking it.

np, maybe QDF2400 is a little bit special

quoted

quoted

This patch fixes it by limiting the lpi_id_bits.

Signed-off-by: Jia He <redacted>
---
 drivers/irqchip/irq-gic-v3-its.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index 316a575..79e6993 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c

@@ -1624,8 +1624,11 @@ static void its_free_prop_table(struct page *prop_page)
 static int __init its_alloc_lpi_tables(void)
 {
 	phys_addr_t paddr;
+	u32 max_bits;	 /*max order limitation in alloc_page*/
 
-	lpi_id_bits = GICD_TYPER_ID_BITS(gic_rdists->gicd_typer);
+	max_bits = PAGE_SHIFT + MAX_ORDER - 1;
+	lpi_id_bits = min_t(u32, max_bits,
+			GICD_TYPER_ID_BITS(gic_rdists->gicd_typer));
 	gic_rdists->prop_page = its_allocate_prop_table(GFP_NOWAIT);
 	if (!gic_rdists->prop_page) {
 		pr_err("Failed to allocate PROPBASE\n");

-- 
1.8.3.1

I find it rather odd that we end-up with different interrupt ranges
depending on the CPU page size. Also, allocating that much memory for
LPIs is rather pointless, as we actually have a pretty low limit of
interrupts the system can deal with (see IRQ_BITMAP_BITS, which is
slightly more than 8k). I've so far seen *one* request to push it up,
but I doubt that it is a real use case.

Capping lpi_id_bits at 16 (which is what we had before) is plenty,
will save a some memory, and gives some margin before we need to push
it up again.

Do you want me to revert commit fe8e93504 to cap the lpi_id_bits
to no greater than ITS_MAX_LPI_NRBITS(16) instead this patch?

Yes, this is probably the best course of action at this time. When I
dropped this, I was mostly thinking of the LPI allocator (which now
works with an almost constant memory footprint), and didn't pay enough
attention to the property table allocation (which is still a function
of the number of LPIs).

Thanks,

	M.

-- 
Jazz is not dead, it just smell funny.