Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > linux.kernel > #1545754 > unrolled thread
| Started by | Sunil Kovvuri <sunil.kovvuri@gmail.com> |
|---|---|
| First post | 2016-12-21 14:10 +0100 |
| Last post | 2016-12-27 05:30 +0100 |
| Articles | 3 — 2 participants |
Back to article view | Back to linux.kernel
This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by
below is the oldest one visible, not the original post.
Re: [RFC PATCH 1/7] PF driver modified to enable HW filter support, changes works in backward compatibility mode Enable required things in Makefile Enable LZ4 dependecy inside config file Sunil Kovvuri <sunil.kovvuri@gmail.com> - 2016-12-21 14:10 +0100
RE: [RFC PATCH 1/7] PF driver modified to enable HW filter support, changes works in backward compatibility mode Enable required things in Makefile Enable LZ4 dependecy inside config file "Koteshwar Rao, Satha" <Satha.Rao@cavium.com> - 2016-12-26 15:30 +0100
Re: [RFC PATCH 1/7] PF driver modified to enable HW filter support, changes works in backward compatibility mode Enable required things in Makefile Enable LZ4 dependecy inside config file Sunil Kovvuri <sunil.kovvuri@gmail.com> - 2016-12-27 05:30 +0100
| From | Sunil Kovvuri <sunil.kovvuri@gmail.com> |
|---|---|
| Date | 2016-12-21 14:10 +0100 |
| Subject | Re: [RFC PATCH 1/7] PF driver modified to enable HW filter support, changes works in backward compatibility mode Enable required things in Makefile Enable LZ4 dependecy inside config file |
| Message-ID | <sQOOE-3SW-89@gated-at.bofh.it> |
>
> #define NIC_MAX_RSS_HASH_BITS 8
> #define NIC_MAX_RSS_IDR_TBL_SIZE (1 << NIC_MAX_RSS_HASH_BITS)
> +#define NIC_TNS_RSS_IDR_TBL_SIZE 5
So you want to use only 5 queues per VF when TNS is enabled, is it ??
There are 4096 RSS indices in total, for each VF you can use max 32.
I guess you wanted to set no of hash bits to 5 instead of table size.
> #define RSS_HASH_KEY_SIZE 5 /* 320 bit key */
>
> struct nicvf_rss_info {
> @@ -255,74 +258,6 @@ struct nicvf_drv_stats {
> struct u64_stats_sync syncp;
> };
>
> -struct nicvf {
> - struct nicvf *pnicvf;
> - struct net_device *netdev;
> - struct pci_dev *pdev;
> - void __iomem *reg_base;
Didn't get why you moved this structure to the end of file.
Looks like an unnecessary modification.
> +static unsigned int num_vfs;
> +module_param(num_vfs, uint, 0644);
> +MODULE_PARM_DESC(num_vfs, "Non zero positive value, specifies number of VF's per physical port");
So what if driver is built-in instead of module, I can't use TNS is it ?
>
> +/* Set RBDR Backpressure (RBDR_BP) and CQ backpressure (CQ_BP) of vnic queues
> + * to 129 each
Why 129 ??
RBDR minimum size is 8K buffers, why you want to assert BP when still
~4K buffers
are available. Isn't 4K a huge number to start asserting backpressure ?
[toc] | [next] | [standalone]
| From | "Koteshwar Rao, Satha" <Satha.Rao@cavium.com> |
|---|---|
| Date | 2016-12-26 15:30 +0100 |
| Message-ID | <sSErL-5SM-13@gated-at.bofh.it> |
| In reply to | #1545754 |
Responses inline
Thanks,
Satha
-----Original Message-----
From: Sunil Kovvuri [mailto:sunil.kovvuri@gmail.com]
Sent: Wednesday, December 21, 2016 5:05 AM
To: Koteshwar Rao, Satha
Cc: LKML; Goutham, Sunil; Robert Richter; David S. Miller; Daney, David; Vatsavayi, Raghu; Chickles, Derek; Romanov, Philip; Linux Netdev List; LAKML
Subject: Re: [RFC PATCH 1/7] PF driver modified to enable HW filter support, changes works in backward compatibility mode Enable required things in Makefile Enable LZ4 dependecy inside config file
>
> #define NIC_MAX_RSS_HASH_BITS 8
> #define NIC_MAX_RSS_IDR_TBL_SIZE (1 << NIC_MAX_RSS_HASH_BITS)
> +#define NIC_TNS_RSS_IDR_TBL_SIZE 5
So you want to use only 5 queues per VF when TNS is enabled, is it ??
There are 4096 RSS indices in total, for each VF you can use max 32.
I guess you wanted to set no of hash bits to 5 instead of table size.
SATHA>>> We enabled 8 queues for VF. Yes Macro name misleads it has to be hash bits, will change this in next version
> #define RSS_HASH_KEY_SIZE 5 /* 320 bit key */
>
> struct nicvf_rss_info {
> @@ -255,74 +258,6 @@ struct nicvf_drv_stats {
> struct u64_stats_sync syncp;
> };
>
> -struct nicvf {
> - struct nicvf *pnicvf;
> - struct net_device *netdev;
> - struct pci_dev *pdev;
> - void __iomem *reg_base;
Didn't get why you moved this structure to the end of file.
Looks like an unnecessary modification.
SATHA>>> Previously we have some dependency, we look into this, and address in next verison
> +static unsigned int num_vfs;
> +module_param(num_vfs, uint, 0644);
> +MODULE_PARM_DESC(num_vfs, "Non zero positive value, specifies number
> +of VF's per physical port");
So what if driver is built-in instead of module, I can't use TNS is it ?
SATHA>>> Still you can enable this special features by passing boot argument "nicpf.num_vfs=X"
>
> +/* Set RBDR Backpressure (RBDR_BP) and CQ backpressure (CQ_BP) of
> +vnic queues
> + * to 129 each
Why 129 ??
RBDR minimum size is 8K buffers, why you want to assert BP when still ~4K buffers are available. Isn't 4K a huge number to start asserting backpressure ?
SATHA>>> As CQ count was 4K entries, I used same BP value for both, will address this in next version
[toc] | [prev] | [next] | [standalone]
| From | Sunil Kovvuri <sunil.kovvuri@gmail.com> |
|---|---|
| Date | 2016-12-27 05:30 +0100 |
| Message-ID | <sSRyG-5IB-9@gated-at.bofh.it> |
| In reply to | #1547325 |
>> #define NIC_MAX_RSS_HASH_BITS 8 >> #define NIC_MAX_RSS_IDR_TBL_SIZE (1 << NIC_MAX_RSS_HASH_BITS) >> +#define NIC_TNS_RSS_IDR_TBL_SIZE 5 > > So you want to use only 5 queues per VF when TNS is enabled, is it ?? > There are 4096 RSS indices in total, for each VF you can use max 32. > I guess you wanted to set no of hash bits to 5 instead of table size. > > SATHA>>> We enabled 8 queues for VF. Yes Macro name misleads it has to be hash bits, will change this in next version No, I am not referring to any discrepancy in naming the macro. If you check your code - hw->rss_ind_tbl_size = NIC_MAX_RSS_IDR_TBL_SIZE; + hw->rss_ind_tbl_size = veb_enabled ? NIC_TNS_RSS_IDR_TBL_SIZE : + NIC_MAX_RSS_IDR_TBL_SIZE; You are setting RSS table size to 5, i.e RSS hash bits will be set to 2. Hence only 4 queues (not even 5 as i mentioned earlier). Please check 'nicvf_rss_init' you will understand what i am saying. Have you tested and observed pkts in all 8 queues ?
[toc] | [prev] | [standalone]
Back to top | Article view | linux.kernel
csiph-web