| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <158387214232.149997.3935472981193001512@swboyd.mtv.corp.google.com>
Date: Tue, 10 Mar 2020 13:29:02 -0700
From: Stephen Boyd <swboyd@...omium.org>
To: Andy Gross <agross@...nel.org>,
Bjorn Andersson <bjorn.andersson@...aro.org>,
Ohad Ben-Cohen <ohad@...ery.com>,
Rob Herring <robh+dt@...nel.org>
Cc: linux-arm-msm@...r.kernel.org, linux-remoteproc@...r.kernel.org,
devicetree@...r.kernel.org, linux-kernel@...r.kernel.org,
Mathieu Poirier <mathieu.poirier@...aro.org>
Subject: Re: [PATCH v4 2/5] remoteproc: qcom: Introduce driver to store pil info in IMEM
Quoting Bjorn Andersson (2020-03-09 23:33:35)
> A region in IMEM is used to communicate load addresses of remoteproc to
> post mortem debug tools. Implement a driver that can be used to store
> this information in order to enable these tools to process collected
> ramdumps.
>
> Signed-off-by: Bjorn Andersson <bjorn.andersson@...aro.org>
> ---
>
> Changes since v3:
> - Don't carry shadow of all entries
> - Reworked indexing of entries in qcom_pil_info_store()
> - Marked global __read_mostly
>
> Stephen requested, in v3, that the driver should be turned into a library that,
> in the context of the remoteproc drivers, would resolve the pil info region and
> update an appropriate entry, preferably a statically assigned one.
>
> Unfortunately the entries must be packed, so it's not possible to pre-assign
> entries for each remoteproc, in case some of them aren't booted. Further more,
> it turns out that the IMEM region must be zero-initialized in order to have a
> reliable way to determining which entries are available.
>
> We therefor have the need for global mutual exclusion and a mechanism that is
> guaranteed to run before any clients attempt to update the content - so the
> previously proposed design is maintained.
The library could have a static variable that tracks when it's been
called once, holds a lock to protect concurrent access and then zero
initializes on the first call.
> diff --git a/drivers/remoteproc/qcom_pil_info.c b/drivers/remoteproc/qcom_pil_info.c
> new file mode 100644
> index 000000000000..0ddfb2639b7f
> --- /dev/null
> +++ b/drivers/remoteproc/qcom_pil_info.c
> @@ -0,0 +1,180 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (c) 2019-2020 Linaro Ltd.
> + */
> +#include <linux/kernel.h>
> +#include <linux/mfd/syscon.h>
> +#include <linux/module.h>
> +#include <linux/mutex.h>
> +#include <linux/of.h>
> +#include <linux/of_address.h>
> +#include <linux/platform_device.h>
> +#include <linux/regmap.h>
> +#include <linux/slab.h>
> +
> +#define PIL_RELOC_NAME_LEN 8
> +
> +struct pil_reloc_entry {
> + char name[PIL_RELOC_NAME_LEN];
> + __le64 base;
> + __le32 size;
Why are these little endian? Isn't regmap doing endian swaps?
> +} __packed;
Is __packed necessary? It looks like things are naturally aligned
already.
> +
> +struct pil_reloc {
> + struct device *dev;
> + struct regmap *map;
> + size_t offset;
> + size_t num_entries;
> + size_t entry_size;
> +};
> +
> +static struct pil_reloc *_reloc __read_mostly;
> +static DEFINE_MUTEX(reloc_mutex);
> +
> +/**
> + * qcom_pil_info_store() - store PIL information of image in IMEM
> + * @image: name of the image
> + * @base: base address of the loaded image
> + * @size: size of the loaded image
> + *
> + * Return: 0 on success, negative errno on failure
> + */
> +int qcom_pil_info_store(const char *image, phys_addr_t base, size_t size)
> +{
> + struct pil_reloc_entry entry;
> + unsigned int offset;
> + int selected = -1;
> + int ret;
> + int i;
> +
> + offset = _reloc->offset;
> +
> + mutex_lock(&reloc_mutex);
> + for (i = 0; i < _reloc->num_entries; i++) {
> + ret = regmap_bulk_read(_reloc->map, offset, &entry, _reloc->entry_size);
> + if (ret < 0)
> + continue;
> +
> + if (!entry.name[0]) {
> + if (selected == -1)
> + selected = offset;
> + continue;
Why not goto found?
> + }
> +
> + if (!strncmp(entry.name, image, sizeof(entry.name))) {
> + selected = offset;
> + goto found;
Why not goto found_again? And then jump over the strscpy() in this case?
> + }
> +
> + offset += sizeof(entry);
> + }
> +
> + if (selected == -1) {
Do we need this check? It should have been jumped over if it found it?
> + dev_warn(_reloc->dev, "insufficient PIL info slots\n");
> + ret = -ENOMEM;
> + goto unlock;
> + }
> +
> +found:
> + strscpy(entry.name, image, ARRAY_SIZE(entry.name));
> + entry.base = base;
> + entry.size = size;
Sparse should complain here.
> +
> + ret = regmap_bulk_write(_reloc->map, selected, &entry, _reloc->entry_size);
It would make a lot more sense to me if this was written with plain
readl/writel logic. Then I don't have to think about structs being
stored in I/O memory regions, and instead I can concentrate on there
being two 64-bit registers for name and base and one 32-bit register for
the size.
> +unlock:
> + mutex_unlock(&reloc_mutex);
> +
> + return ret;
> +}
> +EXPORT_SYMBOL_GPL(qcom_pil_info_store);
> +
> +/**
> + * qcom_pil_info_available() - query if the pil info is probed
> + *
> + * Return: boolean indicating if the pil info device is probed
> + */
> +bool qcom_pil_info_available(void)
> +{
> + return !!_reloc;
> +}
> +EXPORT_SYMBOL_GPL(qcom_pil_info_available);
> +
> +static int pil_reloc_probe(struct platform_device *pdev)
> +{
> + struct pil_reloc_entry entry = {0};
> + struct pil_reloc *reloc;
> + struct resource *res;
> + struct resource imem;
> + unsigned int offset;
> + int ret;
> + int i;
> +
> + reloc = devm_kzalloc(&pdev->dev, sizeof(*reloc), GFP_KERNEL);
> + if (!reloc)
> + return -ENOMEM;
> +
> + reloc->dev = &pdev->dev;
> +
> + res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
> + if (!res)
> + return -EINVAL;
> +
> + /* reloc->offset is relative to parent (IMEM) base address */
> + ret = of_address_to_resource(pdev->dev.of_node->parent, 0, &imem);
> + if (ret < 0)
> + return ret;
> +
> + reloc->offset = res->start - imem.start;
> + reloc->num_entries = resource_size(res) / sizeof(entry);
> +
> + reloc->map = syscon_node_to_regmap(pdev->dev.parent->of_node);
> + if (IS_ERR(reloc->map))
> + return PTR_ERR(reloc->map);
> +
> + ret = regmap_get_val_bytes(reloc->map);
> + if (ret < 0)
> + return -EINVAL;
> + reloc->entry_size = sizeof(entry) / ret;
> +
> + offset = reloc->offset;
> + for (i = 0; i < reloc->num_entries; i++, offset += sizeof(entry)) {
> + ret = regmap_bulk_write(reloc->map, offset, &entry,
> + reloc->entry_size);
> + if (ret < 0)
> + return ret;
> + }
This would be a lot easier to read with a memset_io().
> +
> + mutex_lock(&reloc_mutex);
> + _reloc = reloc;
> + mutex_unlock(&reloc_mutex);
> +
> + return 0;
> +}
> +
> +static int pil_reloc_remove(struct platform_device *pdev)
> +{
> + mutex_lock(&reloc_mutex);
> + _reloc = NULL;
> + mutex_unlock(&reloc_mutex);
> +
> + return 0;
> +}
> +
> +static const struct of_device_id pil_reloc_of_match[] = {
> + { .compatible = "qcom,pil-reloc-info" },
> + {}
> +};
> +MODULE_DEVICE_TABLE(of, pil_reloc_of_match);
> +
> +static struct platform_driver pil_reloc_driver = {
I still don't understand the need for a platform device driver and probe
ordering. It's not a device in the usual ways, it's just some memory
that's lying around and always there! Why can't we search in DT for the
imem node and then find the pil reloc info compatible string on the
first call to this library? Then we don't need an API to see if the
device has probed yet (qcom_pil_info_available) and we can just ioremap
some region of memory that's carved out for this reason. Forcing
everything through the regmap is mostly adding pain.
> + .probe = pil_reloc_probe,
Powered by blists - more mailing lists