| Message ID | 20240320172008.2989693-1-enachman@marvell.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Fix prestera driver fail to probe twice | expand |
On Wed, Mar 20, 2024 at 07:20:03PM +0200, Elad Nachman wrote: > From: Elad Nachman <enachman@marvell.com> > > Fix issues resulting from insmod, rmmod and insmod of the > prestera driver: > > 1. Call of firmware switch HW reset was missing, and is required > in order to make the firmware loader shift to the correct state > needed for loading the next firmware. > 2. Time-out for waiting for firmware loader to be ready was too small. > 3. memory referencing after freeing > 4. MAC addresses wrapping > 5. Missing SFP unbind (phylink release) of a port during the port release. I don't see any problems with 3-5. Maybe post them independent of the first 2 and they can be merged. Are you really saying it is impossible to determine if the hardware is in the boot loader waiting for firmware, or is running the firmware? Andrew
On Wed, 20 Mar 2024 19:20:03 +0200 Elad Nachman <enachman@marvell.com> wrote: > From: Elad Nachman <enachman@marvell.com> > > Fix issues resulting from insmod, rmmod and insmod of the > prestera driver: Please add "net" prefixes to all your patches subject, like that: [PATCH net v2 x/5] I think the maintainers bots won't works if you don't. Regards,
On Thu, 21 Mar 2024 10:06:00 +0100 Kory Maincent wrote: > > Fix issues resulting from insmod, rmmod and insmod of the > > prestera driver: > > Please add "net" prefixes to all your patches subject, like that: > [PATCH net v2 x/5] > > I think the maintainers bots won't works if you don't. They will default to net-next, which right now is perfectly fine.
> -----Original Message----- > From: Andrew Lunn <andrew@lunn.ch> > Sent: Thursday, March 21, 2024 2:18 AM > To: Elad Nachman <enachman@marvell.com> > Cc: Taras Chornyi <taras.chornyi@plvision.eu>; davem@davemloft.net; > edumazet@google.com; kuba@kernel.org; pabeni@redhat.com; > kory.maincent@bootlin.com; thomas.petazzoni@bootlin.com; > miquel.raynal@bootlin.com; przemyslaw.kitszel@intel.com; > dkirjanov@suse.de; netdev@vger.kernel.org; linux-kernel@vger.kernel.org > Subject: [EXTERNAL] Re: [PATCH v2 0/5] Fix prestera driver fail to probe twice > > Prioritize security for external emails: Confirm sender and content safety > before clicking links or opening attachments > > ---------------------------------------------------------------------- > On Wed, Mar 20, 2024 at 07:20:03PM +0200, Elad Nachman wrote: > > From: Elad Nachman <enachman@marvell.com> > > > > Fix issues resulting from insmod, rmmod and insmod of the prestera > > driver: > > > > 1. Call of firmware switch HW reset was missing, and is required > > in order to make the firmware loader shift to the correct state > > needed for loading the next firmware. > > 2. Time-out for waiting for firmware loader to be ready was too small. > > 3. memory referencing after freeing > > 4. MAC addresses wrapping > > 5. Missing SFP unbind (phylink release) of a port during the port release. > > I don't see any problems with 3-5. Maybe post them independent of the first > 2 and they can be merged. > > Are you really saying it is impossible to determine if the hardware is in the > boot loader waiting for firmware, or is running the firmware? > > Andrew Originally, the pain point for Kory was the rmmod + insmod re-probing failure, Which is only fixed by the first two commits, so I see little point in submitting 3-5 alone, Without fixing Kory's problem. The problem is not with the hardware, but with the existing firmware code on the Firmware cpu, most probably secure-boot protected, which lacks the ABIs to report to The kernel what is loaded, what version, what state, etc. I agree that with better original design, we could have made a better work of avoiding This lengthy reload, but at this point, I believe my options are quite limited here, unfortunately. Elad.
> Originally, the pain point for Kory was the rmmod + insmod re-probing failure, > Which is only fixed by the first two commits, so I see little point in submitting 3-5 alone, > Without fixing Kory's problem. I thought Kory's problem was actually EPROBE_DEFER? The resources needed for the PoE are not available, so probing the switch needs to happen again later, when PoE can get the resources it needs. But if that is going to take 30 seconds, i'm not sure we can call EPROBE_DEFER solved. The later patches are pretty simple, don't need discussion, so could be merged. However, i think we need to explore different possible solutions for firmware {re}loading. > The problem is not with the hardware, but with the existing firmware code on the > Firmware cpu, most probably secure-boot protected, which lacks the ABIs to report to > The kernel what is loaded, what version, what state, etc. Can you at least tell if it is running firmware? Can you explain the boot in a bit more detail. Are you saying it could be running an old firmware when the driver first loads? So you need to hit it with a reset in order to load the firmware for /lib/firmware, which might be newer than what it is already running? That would imply the device has FLASH and has a copy of firmware in it? And if that is true, i think that also implies you have no way to upgrade the image in FLASH? Otherwise you would implement "devlink flash" to allow it to be upgraded. You then would not need to load the firmware on driver probe.... Andrew
> -----Original Message----- > From: Andrew Lunn <andrew@lunn.ch> > Sent: Thursday, March 21, 2024 9:22 PM > To: Elad Nachman <enachman@marvell.com> > Cc: Taras Chornyi <taras.chornyi@plvision.eu>; davem@davemloft.net; > edumazet@google.com; kuba@kernel.org; pabeni@redhat.com; > kory.maincent@bootlin.com; thomas.petazzoni@bootlin.com; > miquel.raynal@bootlin.com; przemyslaw.kitszel@intel.com; > dkirjanov@suse.de; netdev@vger.kernel.org; linux-kernel@vger.kernel.org > Subject: Re: [EXTERNAL] Re: [PATCH v2 0/5] Fix prestera driver fail to probe > twice > > > Originally, the pain point for Kory was the rmmod + insmod re-probing > > failure, Which is only fixed by the first two commits, so I see little > > point in submitting 3-5 alone, Without fixing Kory's problem. > > I thought Kory's problem was actually EPROBE_DEFER? The resources needed > for the PoE are not available, so probing the switch needs to happen again > later, when PoE can get the resources it needs. No, the PoE is the general high level application where he noted the problem. There is no PoE code nor special PoE resources in the Prestera driver. The problem was caused because the module exit was lacking the so called "switch HW reset" API call which would cause the firmware to exit to the firmware loader on the firmware CPU, and move to the state in the state machine when it can receive new firmware from the host CPU (running the Prestera switchDev driver). > > But if that is going to take 30 seconds, i'm not sure we can call EPROBE_DEFER > solved. > > The later patches are pretty simple, don't need discussion, so could be > merged. However, i think we need to explore different possible solutions for > firmware {re}loading. > > > The problem is not with the hardware, but with the existing firmware > > code on the Firmware cpu, most probably secure-boot protected, which > > lacks the ABIs to report to The kernel what is loaded, what version, what > state, etc. > > Can you at least tell if it is running firmware? There is no existing API/ABI for that. > > Can you explain the boot in a bit more detail. Are you saying it could be > running an old firmware when the driver first loads? So you need to hit it with Exactly. > a reset in order to load the firmware for /lib/firmware, which might be newer > than what it is already running? Right. And there is also the configuration. There is no telling what kind of Configuration the existing firmware is running. Just using the existing firmware Will lead to the situation where Linux kernel side will report certain configuration (via ip link / ip addr / tc , etc.) but the firmware configuration is completely different. Loading the firmware again will configure the switch to the default setting, making Sure that the Linux kernel switchDev side is synchronized with the firmware side And the actual switch configuration. Unfortunately, there is currently no API/ABI for warm-boot synchronization from The firmware side to the Kernel switchdev side. > > That would imply the device has FLASH and has a copy of firmware in it? And Not of the firmware, the flash holds the firmware loader code. This is a limited functionality code which has only the minimal API/ABI to load The next firmware. > if that is true, i think that also implies you have no way to upgrade the image > in FLASH? Otherwise you would implement "devlink flash" to allow it to be > upgraded. You then would not need to load the firmware on driver probe.... Right. This is a limitation of the design made. There is no option to upgrade the Firmware loader binary on the flash, and many boards have it in secure boot, Which means it cannot be upgraded without bricking the firmware CPU loader binary... > > Andrew > >
> > > Originally, the pain point for Kory was the rmmod + insmod re-probing > > > failure, Which is only fixed by the first two commits, so I see little > > > point in submitting 3-5 alone, Without fixing Kory's problem. > > > > I thought Kory's problem was actually EPROBE_DEFER? The resources needed > > for the PoE are not available, so probing the switch needs to happen again > > later, when PoE can get the resources it needs. > > No, the PoE is the general high level application where he noted the problem. > There is no PoE code nor special PoE resources in the Prestera driver. So here is Köry email: https://lore.kernel.org/netdev/20240208101005.29e8c7f3@kmaincent-XPS-13-7390/T/#mb898bb2a4bf07776d79f1a19b6a8420716ecb4a3 I don't see why the prestera needs to be involved in PoE itself. It is just a MAC. PoE happens much lower down in the network stack. Same as Prestera uses phylink, it does not need to know about the PHYs or the SFP modules, phylink manages them, not prestera. > The problem was caused because the module exit was lacking the so called > "switch HW reset" API call which would cause the firmware to exit to the firmware > loader on the firmware CPU, and move to the state in the state machine when > it can receive new firmware from the host CPU (running the Prestera switchDev > driver). > > > > > But if that is going to take 30 seconds, i'm not sure we can call EPROBE_DEFER > > solved. > > > > The later patches are pretty simple, don't need discussion, so could be > > merged. However, i think we need to explore different possible solutions for > > firmware {re}loading. > > > > > The problem is not with the hardware, but with the existing firmware > > > code on the Firmware cpu, most probably secure-boot protected, which > > > lacks the ABIs to report to The kernel what is loaded, what version, what > > state, etc. > > > > Can you at least tell if it is running firmware? > > There is no existing API/ABI for that. Do you at least have the ability to determine if an API call exists or not? It sounds like your firmware needs extending to support returning the version. If the API is missing, you know it is 4.1 or older. If it does exist, it will return 4.2 or higher. > > Can you explain the boot in a bit more detail. Are you saying it could be > > running an old firmware when the driver first loads? So you need to hit it with > > Exactly. > > > a reset in order to load the firmware for /lib/firmware, which might be newer > > than what it is already running? > > Right. And there is also the configuration. There is no telling what kind of > Configuration the existing firmware is running. Just using the existing firmware > Will lead to the situation where Linux kernel side will report certain configuration > (via ip link / ip addr / tc , etc.) but the firmware configuration is completely different. Well, during probe and -EPRODE_DEFER, linux has no configuration, since the driver failed to probe. However, for a rmmod/modprobe, the firmware could have stale configuration. However pretty much every device i've come across has the concept of a software reset which clears out the configuration. Seems to be something else your firmware is missing. Andrew
On Sun, 24 Mar 2024 16:25:28 +0100 Andrew Lunn <andrew@lunn.ch> wrote: > > > > Originally, the pain point for Kory was the rmmod + insmod re-probing > > > > failure, Which is only fixed by the first two commits, so I see little > > > > point in submitting 3-5 alone, Without fixing Kory's problem. > > > > > > I thought Kory's problem was actually EPROBE_DEFER? The resources needed > > > for the PoE are not available, so probing the switch needs to happen again > > > later, when PoE can get the resources it needs. > > > > No, the PoE is the general high level application where he noted the > > problem. There is no PoE code nor special PoE resources in the Prestera > > driver. > > So here is Köry email: > > https://lore.kernel.org/netdev/20240208101005.29e8c7f3@kmaincent-XPS-13-7390/T/#mb898bb2a4bf07776d79f1a19b6a8420716ecb4a3 > > I don't see why the prestera needs to be involved in PoE itself. It is > just a MAC. PoE happens much lower down in the network stack. Same as > Prestera uses phylink, it does not need to know about the PHYs or the > SFP modules, phylink manages them, not prestera. Prestera is indeed not directly involved in PoE. I wrote a hack to be able to get the PoE ports control, for testing my PoE patch series. The aim in the future will be to add RJ45 port abstraction. The Prestera will get the port abstraction which will get the PoE ports control. The prestera driver then might receive an EPROBE_DEFER from it. Regards,
> Prestera is indeed not directly involved in PoE. I wrote a hack to be able to > get the PoE ports control, for testing my PoE patch series. > > The aim in the future will be to add RJ45 port abstraction. > The Prestera will get the port abstraction which will get the PoE ports control. > The prestera driver then might receive an EPROBE_DEFER from it. O.K, so EPRODE_DEFER has to work in a meaningful way. I'm not sure we can call a 30 second delay meaningful. And it is not just PoE. phylink_create() can return EPROBE_DEFER, and there are probably others. Andrew
Hi Andrew, We have made internal technical review of the issues you have raised (return version API, try to get version API before starting to initialize and load the firmware, clear configuration API) versus the delay saved (almost 30 seconds minus several seconds to perform and complete the API calls) - around 20 seconds or so. Existing customers we have talked to seem to be able to cope with the existing delay. Unfortunately, the amount of coding and testing involved with saving these 20 seconds or so is beyond our available development manpower at this specific point in time. Unfortunately, we will have to defer making the development you have requested to a later period in time. Elad. > -----Original Message----- > From: Andrew Lunn <andrew@lunn.ch> > Sent: Sunday, March 24, 2024 5:25 PM > To: Elad Nachman <enachman@marvell.com> > Cc: Taras Chornyi <taras.chornyi@plvision.eu>; davem@davemloft.net; > edumazet@google.com; kuba@kernel.org; pabeni@redhat.com; > kory.maincent@bootlin.com; thomas.petazzoni@bootlin.com; > miquel.raynal@bootlin.com; przemyslaw.kitszel@intel.com; > dkirjanov@suse.de; netdev@vger.kernel.org; linux-kernel@vger.kernel.org > Subject: Re: [EXTERNAL] Re: [PATCH v2 0/5] Fix prestera driver fail to probe > twice > > > > > Originally, the pain point for Kory was the rmmod + insmod > > > > re-probing failure, Which is only fixed by the first two commits, > > > > so I see little point in submitting 3-5 alone, Without fixing Kory's > problem. > > > > > > I thought Kory's problem was actually EPROBE_DEFER? The resources > > > needed for the PoE are not available, so probing the switch needs to > > > happen again later, when PoE can get the resources it needs. > > > > No, the PoE is the general high level application where he noted the > problem. > > There is no PoE code nor special PoE resources in the Prestera driver. > > So here is Köry email: > > https://urldefense.proofpoint.com/v2/url?u=https- > 3A__lore.kernel.org_netdev_20240208101005.29e8c7f3-40kmaincent-2DXPS- > 2D13-2D7390_T_- > 23mb898bb2a4bf07776d79f1a19b6a8420716ecb4a3&d=DwIDAw&c=nKjWec2 > b6R0mOyPaz7xtfQ&r=eTeNTLEK5- > TxXczjOcKPhANIFtlB9pP4lq9qhdlFrwQ&m=SD1MhKC11sFmp4Q8l76N_DgGdac > 4aMCTdPsa7Pofb73HEqAGtJ-1p0- > etIyyldC7&s=VWat9LPub52H3nUez4itmkpuMipnYD3Ngn-paFC9wd4&e= > > I don't see why the prestera needs to be involved in PoE itself. It is just a MAC. > PoE happens much lower down in the network stack. Same as Prestera uses > phylink, it does not need to know about the PHYs or the SFP modules, phylink > manages them, not prestera. > > > The problem was caused because the module exit was lacking the so > > called "switch HW reset" API call which would cause the firmware to > > exit to the firmware loader on the firmware CPU, and move to the state > > in the state machine when it can receive new firmware from the host > > CPU (running the Prestera switchDev driver). > > > > > > > > But if that is going to take 30 seconds, i'm not sure we can call > > > EPROBE_DEFER solved. > > > > > > The later patches are pretty simple, don't need discussion, so could > > > be merged. However, i think we need to explore different possible > > > solutions for firmware {re}loading. > > > > > > > The problem is not with the hardware, but with the existing > > > > firmware code on the Firmware cpu, most probably secure-boot > > > > protected, which lacks the ABIs to report to The kernel what is > > > > loaded, what version, what > > > state, etc. > > > > > > Can you at least tell if it is running firmware? > > > > There is no existing API/ABI for that. > > Do you at least have the ability to determine if an API call exists or not? It > sounds like your firmware needs extending to support returning the version. > If the API is missing, you know it is 4.1 or older. If it does exist, it will return > 4.2 or higher. > > > > Can you explain the boot in a bit more detail. Are you saying it > > > could be running an old firmware when the driver first loads? So you > > > need to hit it with > > > > Exactly. > > > > > a reset in order to load the firmware for /lib/firmware, which might > > > be newer than what it is already running? > > > > Right. And there is also the configuration. There is no telling what > > kind of Configuration the existing firmware is running. Just using the > > existing firmware Will lead to the situation where Linux kernel side > > will report certain configuration (via ip link / ip addr / tc , etc.) but the > firmware configuration is completely different. > > Well, during probe and -EPRODE_DEFER, linux has no configuration, since the > driver failed to probe. However, for a rmmod/modprobe, the firmware could > have stale configuration. However pretty much every device i've come across > has the concept of a software reset which clears out the configuration. Seems > to be something else your firmware is missing. > > Andrew
From: Elad Nachman <enachman@marvell.com> Fix issues resulting from insmod, rmmod and insmod of the prestera driver: 1. Call of firmware switch HW reset was missing, and is required in order to make the firmware loader shift to the correct state needed for loading the next firmware. 2. Time-out for waiting for firmware loader to be ready was too small. 3. memory referencing after freeing 4. MAC addresses wrapping 5. Missing SFP unbind (phylink release) of a port during the port release. v2: 1) Split first patch to firmware call for switch HW reset and to increasing of firmware loader wait to be ready timeout 2) Explain why is switch HW reset call to the firmware required before shutdown in commit message, and fix wording 3) Use a simpler bitwise-and on the last byte of the base MAC address instead of randomizing again the entire MAC address. reflect that change in the commit message, and explain why it is needed (switch HW implementation requirement). 4) Add Fixes Tags to all commits. 5) For timeout enlargement commit, fix wording in comment and move timeout units to milliseconds. 6) Add Tested-By tags. 7) Add patch to call prestera_port_sfp_unbind() from prestera_destroy_ports() in order to release phylink. Elad Nachman (5): net: marvell: prestera: fix driver reload net: marvell: prestera: enlarge fw restart time net: marvell: prestera: fix memory use after free net: marvell: prestera: force good base mac net: marvell: prestera: unbind sfp port on exit drivers/net/ethernet/marvell/prestera/prestera_hw.c | 8 ++++++++ drivers/net/ethernet/marvell/prestera/prestera_hw.h | 1 + drivers/net/ethernet/marvell/prestera/prestera_main.c | 5 ++++- drivers/net/ethernet/marvell/prestera/prestera_pci.c | 7 ++++++- drivers/net/ethernet/marvell/prestera/prestera_router.c | 1 - .../net/ethernet/marvell/prestera/prestera_router_hw.c | 1 - 6 files changed, 19 insertions(+), 4 deletions(-)